Remove useless files in thirdparty directoy and move to simdjson

1 files changed, 65333 insertions, 0 deletions

diff --git a/thirdparty/simdjson.cpp b/thirdparty/simdjson.cpp
new file mode 100644
index 0000000..64b42aa
--- /dev/null
+++ b/thirdparty/simdjson.cpp
@@ -0,0 +1,65333 @@
+/* auto-generated on 2026-04-17 11:03:09 -0400. version 4.6.2 Do not edit! */
+/* including simdjson.cpp:  */
+/* begin file simdjson.cpp */
+#define SIMDJSON_SRC_SIMDJSON_CPP
+
+/* including base.h: #include <base.h> */
+/* begin file base.h */
+#ifndef SIMDJSON_SRC_BASE_H
+#define SIMDJSON_SRC_BASE_H
+
+/* including simdjson/base.h: #include <simdjson/base.h> */
+/* begin file simdjson/base.h */
+/**
+ * @file Base declarations for all simdjson headers
+ * @private
+ */
+#ifndef SIMDJSON_BASE_H
+#define SIMDJSON_BASE_H
+
+/* including simdjson/common_defs.h: #include "simdjson/common_defs.h" */
+/* begin file simdjson/common_defs.h */
+#ifndef SIMDJSON_COMMON_DEFS_H
+#define SIMDJSON_COMMON_DEFS_H
+
+#include <cassert>
+/* including simdjson/compiler_check.h: #include "simdjson/compiler_check.h" */
+/* begin file simdjson/compiler_check.h */
+#ifndef SIMDJSON_COMPILER_CHECK_H
+#define SIMDJSON_COMPILER_CHECK_H
+
+#ifndef __cplusplus
+#error simdjson requires a C++ compiler
+#endif
+
+#ifndef SIMDJSON_CPLUSPLUS
+#if defined(_MSVC_LANG) && !defined(__clang__)
+#define SIMDJSON_CPLUSPLUS (_MSC_VER == 1900 ? 201103L : _MSVC_LANG)
+#else
+#define SIMDJSON_CPLUSPLUS __cplusplus
+#endif
+#endif
+
+// C++ 26
+#if !defined(SIMDJSON_CPLUSPLUS26) && (SIMDJSON_CPLUSPLUS >= 202402L) // update when the standard is finalized
+#define SIMDJSON_CPLUSPLUS26 1
+#endif
+
+// C++ 23
+#if !defined(SIMDJSON_CPLUSPLUS23) && (SIMDJSON_CPLUSPLUS >= 202302L)
+#define SIMDJSON_CPLUSPLUS23 1
+#endif
+
+// C++ 20
+#if !defined(SIMDJSON_CPLUSPLUS20) && (SIMDJSON_CPLUSPLUS >= 202002L)
+#define SIMDJSON_CPLUSPLUS20 1
+#endif
+
+// C++ 17
+#if !defined(SIMDJSON_CPLUSPLUS17) && (SIMDJSON_CPLUSPLUS >= 201703L)
+#define SIMDJSON_CPLUSPLUS17 1
+#endif
+
+// C++ 14
+#if !defined(SIMDJSON_CPLUSPLUS14) && (SIMDJSON_CPLUSPLUS >= 201402L)
+#define SIMDJSON_CPLUSPLUS14 1
+#endif
+
+// C++ 11
+#if !defined(SIMDJSON_CPLUSPLUS11) && (SIMDJSON_CPLUSPLUS >= 201103L)
+#define SIMDJSON_CPLUSPLUS11 1
+#endif
+
+#ifndef SIMDJSON_CPLUSPLUS11
+#error simdjson requires a compiler compliant with the C++11 standard
+#endif
+
+#ifndef SIMDJSON_IF_CONSTEXPR
+#if SIMDJSON_CPLUSPLUS17
+#define SIMDJSON_IF_CONSTEXPR if constexpr
+#else
+#define SIMDJSON_IF_CONSTEXPR if
+#endif
+#endif
+
+#ifndef SIMDJSON_CONSTEXPR_LAMBDA
+#if SIMDJSON_CPLUSPLUS17
+#define SIMDJSON_CONSTEXPR_LAMBDA constexpr
+#else
+#define SIMDJSON_CONSTEXPR_LAMBDA
+#endif
+#endif
+
+
+
+#ifdef __has_include
+#if __has_include(<version>)
+#include <version>
+#endif
+#endif
+
+// The current specification is unclear on how we detect
+// static reflection, both __cpp_lib_reflection and
+// __cpp_impl_reflection are proposed in the draft specification.
+// For now, we disable static reflect by default. It must be
+// specified at compiler time.
+#ifndef SIMDJSON_STATIC_REFLECTION
+#define SIMDJSON_STATIC_REFLECTION 0 // disabled by default.
+#endif
+
+#if defined(__apple_build_version__)
+#if __apple_build_version__ < 14000000
+#define SIMDJSON_CONCEPT_DISABLED 1 // apple-clang/13 doesn't support std::convertible_to
+#endif
+#endif
+
+#if defined(__cpp_lib_ranges) && __cpp_lib_ranges >= 201911L
+#include <ranges>
+#define SIMDJSON_SUPPORTS_RANGES 1
+#else
+#define SIMDJSON_SUPPORTS_RANGES 0
+#endif
+
+#if defined(__cpp_concepts) && !defined(SIMDJSON_CONCEPT_DISABLED)
+#if __cpp_concepts >= 201907L
+#include <utility>
+#define SIMDJSON_SUPPORTS_CONCEPTS 1
+#else
+#define SIMDJSON_SUPPORTS_CONCEPTS 0
+#endif
+#else // defined(__cpp_concepts) && !defined(SIMDJSON_CONCEPT_DISABLED)
+#define SIMDJSON_SUPPORTS_CONCEPTS 0
+#endif // defined(__cpp_concepts) && !defined(SIMDJSON_CONCEPT_DISABLED)
+
+// copy SIMDJSON_SUPPORTS_CONCEPTS to SIMDJSON_SUPPORTS_DESERIALIZATION.
+#if SIMDJSON_SUPPORTS_CONCEPTS
+#define SIMDJSON_SUPPORTS_DESERIALIZATION 1
+#else
+#define SIMDJSON_SUPPORTS_DESERIALIZATION 0
+#endif
+
+
+#if !defined(SIMDJSON_CONSTEVAL)
+#if defined(__cpp_consteval) && __cpp_consteval >= 201811L && defined(__cpp_lib_constexpr_string) && __cpp_lib_constexpr_string >= 201907L
+#define SIMDJSON_CONSTEVAL 1
+#else
+#define SIMDJSON_CONSTEVAL 0
+#endif // defined(__cpp_consteval) && __cpp_consteval >= 201811L && defined(__cpp_lib_constexpr_string) && __cpp_lib_constexpr_string >= 201907L
+#endif // !defined(SIMDJSON_CONSTEVAL)
+#endif // SIMDJSON_COMPILER_CHECK_H
+/* end file simdjson/compiler_check.h */
+/* including simdjson/portability.h: #include "simdjson/portability.h" */
+/* begin file simdjson/portability.h */
+#ifndef SIMDJSON_PORTABILITY_H
+#define SIMDJSON_PORTABILITY_H
+
+#include <cstddef>
+#include <cstdint>
+#include <cstdlib>
+#include <cfloat>
+#include <cassert>
+#include <climits>
+#ifndef _WIN32
+// strcasecmp, strncasecmp
+#include <strings.h>
+#endif
+
+static_assert(CHAR_BIT == 8, "simdjson requires 8-bit bytes");
+
+
+// We are using size_t without namespace std:: throughout the project
+using std::size_t;
+
+#ifdef _MSC_VER
+#define SIMDJSON_VISUAL_STUDIO 1
+/**
+ * We want to differentiate carefully between
+ * clang under visual studio and regular visual
+ * studio.
+ *
+ * Under clang for Windows, we enable:
+ *  * target pragmas so that part and only part of the
+ *     code gets compiled for advanced instructions.
+ *
+ */
+#ifdef __clang__
+// clang under visual studio
+#define SIMDJSON_CLANG_VISUAL_STUDIO 1
+#else
+// just regular visual studio (best guess)
+#define SIMDJSON_REGULAR_VISUAL_STUDIO 1
+#endif // __clang__
+#endif // _MSC_VER
+
+#if (defined(__x86_64__) || defined(_M_AMD64)) && !defined(_M_ARM64EC)
+#define SIMDJSON_IS_X86_64 1
+#elif defined(__aarch64__) || defined(_M_ARM64) || defined(_M_ARM64EC)
+#define SIMDJSON_IS_ARM64 1
+#elif defined(__riscv) && __riscv_xlen == 64
+#define SIMDJSON_IS_RISCV64 1
+
+  #if __riscv_v_intrinsic >= 11000
+    #define SIMDJSON_HAS_RVV_INTRINSICS 1
+  #endif
+
+  #if SIMDJSON_HAS_RVV_INTRINSICS && __riscv_vector && __riscv_v_min_vlen >= 128 && __riscv_v_elen >= 64
+    #define SIMDJSON_IS_RVV 1 // RISC-V V extension
+  #endif
+
+  // current toolchains don't support fixed-size SIMD types that don't match VLEN directly
+  #if __riscv_v_fixed_vlen >= 128 && __riscv_v_fixed_vlen <= 512
+    #define SIMDJSON_IS_RVV_VLS 1
+  #endif
+
+#elif defined(__loongarch_lp64)
+#define SIMDJSON_IS_LOONGARCH64 1
+#if defined(__loongarch_sx) && defined(__loongarch_asx)
+  #define SIMDJSON_IS_LSX 1
+  #define SIMDJSON_IS_LASX 1 // We can always run both
+#elif defined(__loongarch_sx)
+  #define SIMDJSON_IS_LSX 1
+
+// Adjust for runtime dispatching support.
+#if defined(__GNUC__) && !defined(__clang__) && !defined(__INTEL_COMPILER) && !defined(__NVCOMPILER)
+#if __GNUC__ > 15 || (__GNUC__ == 15 && __GNUC_MINOR__ >= 0)
+  // We are ok, we will support runtime dispatch for LASX.
+#else
+  // We disable runtime dispatch for LASX, which means that we will not be able to use LASX
+  // even if it is supported by the hardware.
+  // Loongson users should update to GCC 15 or better.
+  #define SIMDJSON_IMPLEMENTATION_LASX 0
+#endif
+#else
+  // We are not using GCC, so we assume that we can support runtime dispatch for LASX.
+  // https://godbolt.org/z/jcMnrjYhs
+  #define SIMDJSON_IMPLEMENTATION_LASX 0
+#endif
+
+
+
+#endif
+#elif defined(__PPC64__) || defined(_M_PPC64)
+#define SIMDJSON_IS_PPC64 1
+#if defined(__ALTIVEC__)
+#define SIMDJSON_IS_PPC64_VMX 1
+#endif // defined(__ALTIVEC__)
+#else
+#define SIMDJSON_IS_32BITS 1
+
+#if defined(_M_IX86) || defined(__i386__)
+#define SIMDJSON_IS_X86_32BITS 1
+#elif defined(__arm__) || defined(_M_ARM)
+#define SIMDJSON_IS_ARM_32BITS 1
+#elif defined(__PPC__) || defined(_M_PPC)
+#define SIMDJSON_IS_PPC_32BITS 1
+#endif
+
+#endif // defined(__x86_64__) || defined(_M_AMD64)
+#ifndef SIMDJSON_IS_32BITS
+#define SIMDJSON_IS_32BITS 0
+#endif
+
+#if SIMDJSON_IS_32BITS
+#ifndef SIMDJSON_NO_PORTABILITY_WARNING
+// In the future, we should allow programmers
+// to get warning.
+#endif // SIMDJSON_NO_PORTABILITY_WARNING
+#endif // SIMDJSON_IS_32BITS
+
+#define SIMDJSON_CAT_IMPLEMENTATION_(a,...) a ## __VA_ARGS__
+#define SIMDJSON_CAT(a,...) SIMDJSON_CAT_IMPLEMENTATION_(a, __VA_ARGS__)
+
+#define SIMDJSON_STRINGIFY_IMPLEMENTATION_(a,...) #a SIMDJSON_STRINGIFY(__VA_ARGS__)
+#define SIMDJSON_STRINGIFY(a,...) SIMDJSON_CAT_IMPLEMENTATION_(a, __VA_ARGS__)
+
+// this is almost standard?
+#undef SIMDJSON_STRINGIFY_IMPLEMENTATION_
+#undef SIMDJSON_STRINGIFY
+#define SIMDJSON_STRINGIFY_IMPLEMENTATION_(a) #a
+#define SIMDJSON_STRINGIFY(a) SIMDJSON_STRINGIFY_IMPLEMENTATION_(a)
+
+// Our fast kernels require 64-bit systems.
+//
+// On 32-bit x86, we lack 64-bit popcnt, lzcnt, blsr instructions.
+// Furthermore, the number of SIMD registers is reduced.
+//
+// On 32-bit ARM, we would have smaller registers.
+//
+// The simdjson users should still have the fallback kernel. It is
+// slower, but it should run everywhere.
+
+//
+// Enable valid runtime implementations, and select SIMDJSON_BUILTIN_IMPLEMENTATION
+//
+
+// We are going to use runtime dispatch.
+#if defined(SIMDJSON_IS_X86_64) || defined(SIMDJSON_IS_LSX)
+#ifdef __clang__
+// clang does not have GCC push pop
+// warning: clang attribute push can't be used within a namespace in clang up
+// til 8.0 so SIMDJSON_TARGET_REGION and SIMDJSON_UNTARGET_REGION must be *outside* of a
+// namespace.
+#define SIMDJSON_TARGET_REGION(T)                                                       \
+  _Pragma(SIMDJSON_STRINGIFY(                                                           \
+      clang attribute push(__attribute__((target(T))), apply_to = function)))
+#define SIMDJSON_UNTARGET_REGION _Pragma("clang attribute pop")
+#elif defined(__GNUC__)
+// GCC is easier
+#define SIMDJSON_TARGET_REGION(T)                                                       \
+  _Pragma("GCC push_options") _Pragma(SIMDJSON_STRINGIFY(GCC target(T)))
+#define SIMDJSON_UNTARGET_REGION _Pragma("GCC pop_options")
+#endif // clang then gcc
+
+#endif // defined(SIMDJSON_IS_X86_64) || defined(SIMDJSON_IS_LSX)
+
+// Default target region macros don't do anything.
+#ifndef SIMDJSON_TARGET_REGION
+#define SIMDJSON_TARGET_REGION(T)
+#define SIMDJSON_UNTARGET_REGION
+#endif
+
+// Is threading enabled?
+#if defined(_REENTRANT) || defined(_MT)
+#ifndef SIMDJSON_THREADS_ENABLED
+#define SIMDJSON_THREADS_ENABLED
+#endif
+#endif
+
+// workaround for large stack sizes under -O0.
+// https://github.com/simdjson/simdjson/issues/691
+#ifdef __APPLE__
+#ifndef __OPTIMIZE__
+// Apple systems have small stack sizes in secondary threads.
+// Lack of compiler optimization may generate high stack usage.
+// Users may want to disable threads for safety, but only when
+// in debug mode which we detect by the fact that the __OPTIMIZE__
+// macro is not defined.
+#undef SIMDJSON_THREADS_ENABLED
+#endif
+#endif
+
+
+#if defined(__clang__)
+#define SIMDJSON_NO_SANITIZE_UNDEFINED __attribute__((no_sanitize("undefined")))
+#elif defined(__GNUC__)
+#define SIMDJSON_NO_SANITIZE_UNDEFINED __attribute__((no_sanitize_undefined))
+#else
+#define SIMDJSON_NO_SANITIZE_UNDEFINED
+#endif
+
+#if defined(__clang__) || defined(__GNUC__)
+#define simdjson_pure [[gnu::pure]]
+#else
+#define simdjson_pure
+#endif
+
+#if defined(__clang__) || defined(__GNUC__)
+#if defined(__has_feature)
+#  if __has_feature(memory_sanitizer)
+#define SIMDJSON_NO_SANITIZE_MEMORY __attribute__((no_sanitize("memory")))
+#  endif // if __has_feature(memory_sanitizer)
+#endif // defined(__has_feature)
+#endif
+// make sure it is defined as 'nothing' if it is unapplicable.
+#ifndef SIMDJSON_NO_SANITIZE_MEMORY
+#define SIMDJSON_NO_SANITIZE_MEMORY
+#endif
+
+#if SIMDJSON_VISUAL_STUDIO
+// This is one case where we do not distinguish between
+// regular visual studio and clang under visual studio.
+// clang under Windows has _stricmp (like visual studio) but not strcasecmp (as clang normally has)
+#define simdjson_strcasecmp _stricmp
+#define simdjson_strncasecmp _strnicmp
+#else
+// The strcasecmp, strncasecmp, and strcasestr functions do not work with multibyte strings (e.g. UTF-8).
+// So they are only useful for ASCII in our context.
+// https://www.gnu.org/software/libunistring/manual/libunistring.html#char-_002a-strings
+#define simdjson_strcasecmp strcasecmp
+#define simdjson_strncasecmp strncasecmp
+#endif
+
+#if (defined(NDEBUG) || defined(__OPTIMIZE__) || (defined(_MSC_VER) && !defined(_DEBUG))) && !SIMDJSON_DEVELOPMENT_CHECKS
+// If SIMDJSON_DEVELOPMENT_CHECKS is undefined or 0, we consider that we are in release mode.
+// If NDEBUG is set, or __OPTIMIZE__ is set, or we are under MSVC in release mode,
+// then do away with asserts and use __assume.
+// We still recommend that our users set NDEBUG in release mode.
+#if SIMDJSON_VISUAL_STUDIO
+#define SIMDJSON_UNREACHABLE() __assume(0)
+#define SIMDJSON_ASSUME(COND) __assume(COND)
+#else
+#define SIMDJSON_UNREACHABLE() __builtin_unreachable();
+#define SIMDJSON_ASSUME(COND) do { if (!(COND)) __builtin_unreachable(); } while (0)
+#endif
+
+#else // defined(NDEBUG) || defined(__OPTIMIZE__) || (defined(_MSC_VER) && !defined(_DEBUG)) && !SIMDJSON_DEVELOPMENT_CHECKS
+// This should only ever be enabled in debug mode.
+#define SIMDJSON_UNREACHABLE() assert(0);
+#define SIMDJSON_ASSUME(COND) assert(COND)
+
+#endif
+
+
+
+#if defined __BYTE_ORDER__ && defined __ORDER_BIG_ENDIAN__
+#define SIMDJSON_IS_BIG_ENDIAN (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__)
+#elif defined _WIN32
+#define SIMDJSON_IS_BIG_ENDIAN 0
+#else
+#if defined(__APPLE__) || defined(__FreeBSD__)
+#include <machine/endian.h>
+#elif defined(sun) || defined(__sun)
+#include <sys/byteorder.h>
+#elif defined(__MVS__)
+#include <sys/endian.h>
+#else
+#ifdef __has_include
+#if __has_include(<endian.h>)
+#include <endian.h>
+#endif //__has_include(<endian.h>)
+#endif //__has_include
+#endif
+#
+#ifndef __BYTE_ORDER__
+// safe choice
+#define SIMDJSON_IS_BIG_ENDIAN 0
+#endif
+#
+#ifndef __ORDER_LITTLE_ENDIAN__
+// safe choice
+#define SIMDJSON_IS_BIG_ENDIAN 0
+#endif
+#
+#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
+#define SIMDJSON_IS_BIG_ENDIAN 0
+#else
+#define SIMDJSON_IS_BIG_ENDIAN 1
+#endif
+#endif
+
+
+#endif // SIMDJSON_PORTABILITY_H
+/* end file simdjson/portability.h */
+
+namespace simdjson {
+namespace internal {
+/**
+ * @private
+ * Our own implementation of the C++17 to_chars function.
+ * Defined in src/to_chars
+ */
+char *to_chars(char *first, const char *last, double value);
+/**
+ * @private
+ * A number parsing routine.
+ * Defined in src/from_chars
+ */
+double from_chars(const char *first) noexcept;
+double from_chars(const char *first, const char* end) noexcept;
+}
+
+#ifndef SIMDJSON_EXCEPTIONS
+#if defined(__cpp_exceptions) || defined(_CPPUNWIND)
+#define SIMDJSON_EXCEPTIONS 1
+#else
+#define SIMDJSON_EXCEPTIONS 0
+#endif
+#endif
+
+} // namespace simdjson
+
+#if defined(__GNUC__)
+  // Marks a block with a name so that MCA analysis can see it.
+  #define SIMDJSON_BEGIN_DEBUG_BLOCK(name) __asm volatile("# LLVM-MCA-BEGIN " #name);
+  #define SIMDJSON_END_DEBUG_BLOCK(name) __asm volatile("# LLVM-MCA-END " #name);
+  #define SIMDJSON_DEBUG_BLOCK(name, block) BEGIN_DEBUG_BLOCK(name); block; END_DEBUG_BLOCK(name);
+#else
+  #define SIMDJSON_BEGIN_DEBUG_BLOCK(name)
+  #define SIMDJSON_END_DEBUG_BLOCK(name)
+  #define SIMDJSON_DEBUG_BLOCK(name, block)
+#endif
+
+// Align to N-byte boundary
+#define SIMDJSON_ROUNDUP_N(a, n) (((a) + ((n)-1)) & ~((n)-1))
+#define SIMDJSON_ROUNDDOWN_N(a, n) ((a) & ~((n)-1))
+
+#define SIMDJSON_ISALIGNED_N(ptr, n) (((uintptr_t)(ptr) & ((n)-1)) == 0)
+
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  // We could use [[deprecated]] but it requires C++14
+  #define simdjson_deprecated __declspec(deprecated)
+
+  #define simdjson_really_inline __forceinline
+  #define simdjson_never_inline __declspec(noinline)
+
+  #define simdjson_unused
+  #define simdjson_warn_unused
+
+  #ifndef simdjson_likely
+  #define simdjson_likely(x) x
+  #endif
+  #ifndef simdjson_unlikely
+  #define simdjson_unlikely(x) x
+  #endif
+
+  #define SIMDJSON_PUSH_DISABLE_WARNINGS __pragma(warning( push ))
+  #define SIMDJSON_PUSH_DISABLE_ALL_WARNINGS __pragma(warning( push, 0 ))
+  #define SIMDJSON_DISABLE_VS_WARNING(WARNING_NUMBER) __pragma(warning( disable : WARNING_NUMBER ))
+  // Get rid of Intellisense-only warnings (Code Analysis)
+  // Though __has_include is C++17, it is supported in Visual Studio 2017 or better (_MSC_VER>=1910).
+  #ifdef __has_include
+  #if __has_include(<CppCoreCheck\Warnings.h>)
+  #include <CppCoreCheck\Warnings.h>
+  #define SIMDJSON_DISABLE_UNDESIRED_WARNINGS SIMDJSON_DISABLE_VS_WARNING(ALL_CPPCORECHECK_WARNINGS)
+  #endif
+  #endif
+
+  #ifndef SIMDJSON_DISABLE_UNDESIRED_WARNINGS
+  #define SIMDJSON_DISABLE_UNDESIRED_WARNINGS
+  #endif
+
+  #define SIMDJSON_DISABLE_DEPRECATED_WARNING SIMDJSON_DISABLE_VS_WARNING(4996)
+  #define SIMDJSON_DISABLE_STRICT_OVERFLOW_WARNING
+  #define SIMDJSON_POP_DISABLE_WARNINGS __pragma(warning( pop ))
+
+  #define SIMDJSON_PUSH_DISABLE_UNUSED_WARNINGS
+  #define SIMDJSON_POP_DISABLE_UNUSED_WARNINGS
+
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO
+  // We could use [[deprecated]] but it requires C++14
+  #define simdjson_deprecated __attribute__((deprecated))
+
+  #define simdjson_really_inline inline __attribute__((always_inline))
+  #define simdjson_never_inline inline __attribute__((noinline))
+
+  #define simdjson_unused __attribute__((unused))
+  #define simdjson_warn_unused __attribute__((warn_unused_result))
+
+  #ifndef simdjson_likely
+  #define simdjson_likely(x) __builtin_expect(!!(x), 1)
+  #endif
+  #ifndef simdjson_unlikely
+  #define simdjson_unlikely(x) __builtin_expect(!!(x), 0)
+  #endif
+
+  #define SIMDJSON_PUSH_DISABLE_WARNINGS _Pragma("GCC diagnostic push")
+  // gcc doesn't seem to disable all warnings with all and extra, add warnings here as necessary
+  // We do it separately for clang since it has different warnings.
+  #ifdef __clang__
+  // clang is missing -Wmaybe-uninitialized.
+  #define SIMDJSON_PUSH_DISABLE_ALL_WARNINGS SIMDJSON_PUSH_DISABLE_WARNINGS \
+    SIMDJSON_DISABLE_GCC_WARNING(-Weffc++) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wall) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wconversion) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wextra) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wattributes) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wimplicit-fallthrough) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wnon-virtual-dtor) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wreturn-type) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wshadow) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wunused-parameter) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wunused-variable)
+  #else // __clang__
+  #define SIMDJSON_PUSH_DISABLE_ALL_WARNINGS SIMDJSON_PUSH_DISABLE_WARNINGS \
+    SIMDJSON_DISABLE_GCC_WARNING(-Weffc++) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wall) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wconversion) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wextra) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wattributes) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wimplicit-fallthrough) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wnon-virtual-dtor) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wreturn-type) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wshadow) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wunused-parameter) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wunused-variable) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wmaybe-uninitialized) \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wformat-security)
+  #endif // __clang__
+
+  #define SIMDJSON_PRAGMA(P) _Pragma(#P)
+  #define SIMDJSON_DISABLE_GCC_WARNING(WARNING) SIMDJSON_PRAGMA(GCC diagnostic ignored #WARNING)
+  #if SIMDJSON_CLANG_VISUAL_STUDIO
+  #define SIMDJSON_DISABLE_UNDESIRED_WARNINGS SIMDJSON_DISABLE_GCC_WARNING(-Wmicrosoft-include)
+  #else
+  #define SIMDJSON_DISABLE_UNDESIRED_WARNINGS
+  #endif
+  #define SIMDJSON_DISABLE_DEPRECATED_WARNING SIMDJSON_DISABLE_GCC_WARNING(-Wdeprecated-declarations)
+  #define SIMDJSON_DISABLE_STRICT_OVERFLOW_WARNING SIMDJSON_DISABLE_GCC_WARNING(-Wstrict-overflow)
+  #define SIMDJSON_POP_DISABLE_WARNINGS _Pragma("GCC diagnostic pop")
+
+  #define SIMDJSON_PUSH_DISABLE_UNUSED_WARNINGS SIMDJSON_PUSH_DISABLE_WARNINGS \
+    SIMDJSON_DISABLE_GCC_WARNING(-Wunused)
+  #define SIMDJSON_POP_DISABLE_UNUSED_WARNINGS SIMDJSON_POP_DISABLE_WARNINGS
+
+
+
+#endif // MSC_VER
+
+#if defined(simdjson_inline)
+  // Prefer the user's definition of simdjson_inline; don't define it ourselves.
+#elif defined(__GNUC__) && !defined(__OPTIMIZE__)
+  // If optimizations are disabled, forcing inlining can lead to significant
+  // code bloat and high compile times. Don't use simdjson_really_inline for
+  // unoptimized builds.
+  #define simdjson_inline inline
+#else
+  // Force inlining for most simdjson functions.
+  #define simdjson_inline simdjson_really_inline
+#endif
+
+#if SIMDJSON_VISUAL_STUDIO
+    /**
+     * Windows users need to do some extra work when building
+     * or using a dynamic library (DLL). When building, we need
+     * to set SIMDJSON_DLLIMPORTEXPORT to __declspec(dllexport).
+     * When *using* the DLL, the user needs to set
+     * SIMDJSON_DLLIMPORTEXPORT __declspec(dllimport).
+     *
+     * Static libraries not need require such work.
+     *
+     * It does not matter here whether you are using
+     * the regular visual studio or clang under visual
+     * studio, you still need to handle these issues.
+     *
+     * Non-Windows systems do not have this complexity.
+     */
+    #if SIMDJSON_BUILDING_WINDOWS_DYNAMIC_LIBRARY
+    // We set SIMDJSON_BUILDING_WINDOWS_DYNAMIC_LIBRARY when we build a DLL under Windows.
+    // It should never happen that both SIMDJSON_BUILDING_WINDOWS_DYNAMIC_LIBRARY and
+    // SIMDJSON_USING_WINDOWS_DYNAMIC_LIBRARY are set.
+    #define SIMDJSON_DLLIMPORTEXPORT __declspec(dllexport)
+    #elif SIMDJSON_USING_WINDOWS_DYNAMIC_LIBRARY
+    // Windows user who call a dynamic library should set SIMDJSON_USING_WINDOWS_DYNAMIC_LIBRARY to 1.
+    #define SIMDJSON_DLLIMPORTEXPORT __declspec(dllimport)
+    #else
+    // We assume by default static linkage
+    #define SIMDJSON_DLLIMPORTEXPORT
+    #endif
+#else
+    #define SIMDJSON_DLLIMPORTEXPORT
+#endif
+
+// C++17 requires string_view.
+#if SIMDJSON_CPLUSPLUS17
+#define SIMDJSON_HAS_STRING_VIEW
+#include <string_view> // by the standard, this has to be safe.
+#endif
+
+// This macro (__cpp_lib_string_view) has to be defined
+// for C++17 and better, but if it is otherwise defined,
+// we are going to assume that string_view is available
+// even if we do not have C++17 support.
+#ifdef __cpp_lib_string_view
+#define SIMDJSON_HAS_STRING_VIEW
+#include <string_view>
+#endif
+
+// Some systems have string_view even if we do not have C++17 support,
+// and even if __cpp_lib_string_view is undefined, it is the case
+// with Apple clang version 11.
+// We must handle it. *This is important.*
+#ifndef _MSC_VER
+#ifndef SIMDJSON_HAS_STRING_VIEW
+#if defined __has_include
+// do not combine the next #if with the previous one (unsafe)
+#if __has_include (<string_view>)
+// now it is safe to trigger the include
+#include <string_view> // though the file is there, it does not follow that we got the implementation
+#if defined(_LIBCPP_STRING_VIEW)
+// Ah! So we under libc++ which under its Library Fundamentals Technical Specification, which preceded C++17,
+// included string_view.
+// This means that we have string_view *even though* we may not have C++17.
+#define SIMDJSON_HAS_STRING_VIEW
+#endif // _LIBCPP_STRING_VIEW
+#endif // __has_include (<string_view>)
+#endif // defined __has_include
+#endif // def SIMDJSON_HAS_STRING_VIEW
+#endif // def _MSC_VER
+// end of complicated but important routine to try to detect string_view.
+
+//
+// Backfill std::string_view using nonstd::string_view on systems where
+// we expect that string_view is missing. Important: if we get this wrong,
+// we will end up with two string_view definitions and potential trouble.
+// That is why we work so hard above to avoid it.
+//
+#ifndef SIMDJSON_HAS_STRING_VIEW
+SIMDJSON_PUSH_DISABLE_ALL_WARNINGS
+/* including simdjson/nonstd/string_view.hpp: #include "simdjson/nonstd/string_view.hpp" */
+/* begin file simdjson/nonstd/string_view.hpp */
+// Copyright 2017-2020 by Martin Moene
+//
+// string-view lite, a C++17-like string_view for C++98 and later.
+// For more information see https://github.com/martinmoene/string-view-lite
+//
+// Distributed under the Boost Software License, Version 1.0.
+// (See accompanying file LICENSE.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
+
+
+#ifndef NONSTD_SV_LITE_H_INCLUDED
+#define NONSTD_SV_LITE_H_INCLUDED
+
+#define string_view_lite_MAJOR  1
+#define string_view_lite_MINOR  8
+#define string_view_lite_PATCH  0
+
+#define string_view_lite_VERSION  nssv_STRINGIFY(string_view_lite_MAJOR) "." nssv_STRINGIFY(string_view_lite_MINOR) "." nssv_STRINGIFY(string_view_lite_PATCH)
+
+#define nssv_STRINGIFY(  x )  nssv_STRINGIFY_( x )
+#define nssv_STRINGIFY_( x )  #x
+
+// string-view lite configuration:
+
+#define nssv_STRING_VIEW_DEFAULT  0
+#define nssv_STRING_VIEW_NONSTD   1
+#define nssv_STRING_VIEW_STD      2
+
+// tweak header support:
+
+#ifdef __has_include
+# if __has_include(<nonstd/string_view.tweak.hpp>)
+#  include <nonstd/string_view.tweak.hpp>
+# endif
+#define nssv_HAVE_TWEAK_HEADER  1
+#else
+#define nssv_HAVE_TWEAK_HEADER  0
+//# pragma message("string_view.hpp: Note: Tweak header not supported.")
+#endif
+
+// string_view selection and configuration:
+
+#if !defined( nssv_CONFIG_SELECT_STRING_VIEW )
+# define nssv_CONFIG_SELECT_STRING_VIEW  ( nssv_HAVE_STD_STRING_VIEW ? nssv_STRING_VIEW_STD : nssv_STRING_VIEW_NONSTD )
+#endif
+
+#ifndef  nssv_CONFIG_STD_SV_OPERATOR
+# define nssv_CONFIG_STD_SV_OPERATOR  0
+#endif
+
+#ifndef  nssv_CONFIG_USR_SV_OPERATOR
+# define nssv_CONFIG_USR_SV_OPERATOR  1
+#endif
+
+#ifdef   nssv_CONFIG_CONVERSION_STD_STRING
+# define nssv_CONFIG_CONVERSION_STD_STRING_CLASS_METHODS   nssv_CONFIG_CONVERSION_STD_STRING
+# define nssv_CONFIG_CONVERSION_STD_STRING_FREE_FUNCTIONS  nssv_CONFIG_CONVERSION_STD_STRING
+#endif
+
+#ifndef  nssv_CONFIG_CONVERSION_STD_STRING_CLASS_METHODS
+# define nssv_CONFIG_CONVERSION_STD_STRING_CLASS_METHODS  1
+#endif
+
+#ifndef  nssv_CONFIG_CONVERSION_STD_STRING_FREE_FUNCTIONS
+# define nssv_CONFIG_CONVERSION_STD_STRING_FREE_FUNCTIONS  1
+#endif
+
+#ifndef  nssv_CONFIG_NO_STREAM_INSERTION
+# define nssv_CONFIG_NO_STREAM_INSERTION  0
+#endif
+
+#ifndef  nssv_CONFIG_CONSTEXPR11_STD_SEARCH
+# define nssv_CONFIG_CONSTEXPR11_STD_SEARCH  1
+#endif
+
+// Control presence of exception handling (try and auto discover):
+
+#ifndef nssv_CONFIG_NO_EXCEPTIONS
+# if defined(_MSC_VER)
+#  include <cstddef>    // for _HAS_EXCEPTIONS
+# endif
+# if defined(__cpp_exceptions) || defined(__EXCEPTIONS) || (_HAS_EXCEPTIONS)
+#  define nssv_CONFIG_NO_EXCEPTIONS  0
+# else
+#  define nssv_CONFIG_NO_EXCEPTIONS  1
+# endif
+#endif
+
+// C++ language version detection (C++23 is speculative):
+// Note: VC14.0/1900 (VS2015) lacks too much from C++14.
+
+#ifndef   nssv_CPLUSPLUS
+# if defined(_MSVC_LANG ) && !defined(__clang__)
+#  define nssv_CPLUSPLUS  (_MSC_VER == 1900 ? 201103L : _MSVC_LANG )
+# else
+#  define nssv_CPLUSPLUS  __cplusplus
+# endif
+#endif
+
+#define nssv_CPP98_OR_GREATER  ( nssv_CPLUSPLUS >= 199711L )
+#define nssv_CPP11_OR_GREATER  ( nssv_CPLUSPLUS >= 201103L )
+#define nssv_CPP11_OR_GREATER_ ( nssv_CPLUSPLUS >= 201103L )
+#define nssv_CPP14_OR_GREATER  ( nssv_CPLUSPLUS >= 201402L )
+#define nssv_CPP17_OR_GREATER  ( nssv_CPLUSPLUS >= 201703L )
+#define nssv_CPP20_OR_GREATER  ( nssv_CPLUSPLUS >= 202002L )
+#define nssv_CPP23_OR_GREATER  ( nssv_CPLUSPLUS >= 202300L )
+
+// use C++17 std::string_view if available and requested:
+
+#if nssv_CPP17_OR_GREATER && defined(__has_include )
+# if __has_include( <string_view> )
+#  define nssv_HAVE_STD_STRING_VIEW  1
+# else
+#  define nssv_HAVE_STD_STRING_VIEW  0
+# endif
+#else
+# define  nssv_HAVE_STD_STRING_VIEW  0
+#endif
+
+#define  nssv_USES_STD_STRING_VIEW  ( (nssv_CONFIG_SELECT_STRING_VIEW == nssv_STRING_VIEW_STD) || ((nssv_CONFIG_SELECT_STRING_VIEW == nssv_STRING_VIEW_DEFAULT) && nssv_HAVE_STD_STRING_VIEW) )
+
+#define nssv_HAVE_STARTS_WITH ( nssv_CPP20_OR_GREATER || !nssv_USES_STD_STRING_VIEW )
+#define nssv_HAVE_ENDS_WITH     nssv_HAVE_STARTS_WITH
+
+//
+// Use C++17 std::string_view:
+//
+
+#if nssv_USES_STD_STRING_VIEW
+
+#include <string_view>
+
+// Extensions for std::string:
+
+#if nssv_CONFIG_CONVERSION_STD_STRING_FREE_FUNCTIONS
+
+#include <string>
+
+namespace nonstd {
+
+template< class CharT, class Traits, class Allocator = std::allocator<CharT> >
+std::basic_string<CharT, Traits, Allocator>
+to_string( std::basic_string_view<CharT, Traits> v, Allocator const & a = Allocator() )
+{
+    return std::basic_string<CharT,Traits, Allocator>( v.begin(), v.end(), a );
+}
+
+template< class CharT, class Traits, class Allocator >
+std::basic_string_view<CharT, Traits>
+to_string_view( std::basic_string<CharT, Traits, Allocator> const & s )
+{
+    return std::basic_string_view<CharT, Traits>( s.data(), s.size() );
+}
+
+// Literal operators sv and _sv:
+
+#if nssv_CONFIG_STD_SV_OPERATOR
+
+using namespace std::literals::string_view_literals;
+
+#endif
+
+#if nssv_CONFIG_USR_SV_OPERATOR
+
+inline namespace literals {
+inline namespace string_view_literals {
+
+
+constexpr std::string_view operator ""_sv( const char* str, size_t len ) noexcept  // (1)
+{
+    return std::string_view{ str, len };
+}
+
+constexpr std::u16string_view operator ""_sv( const char16_t* str, size_t len ) noexcept  // (2)
+{
+    return std::u16string_view{ str, len };
+}
+
+constexpr std::u32string_view operator ""_sv( const char32_t* str, size_t len ) noexcept  // (3)
+{
+    return std::u32string_view{ str, len };
+}
+
+constexpr std::wstring_view operator ""_sv( const wchar_t* str, size_t len ) noexcept  // (4)
+{
+    return std::wstring_view{ str, len };
+}
+
+}} // namespace literals::string_view_literals
+
+#endif // nssv_CONFIG_USR_SV_OPERATOR
+
+} // namespace nonstd
+
+#endif // nssv_CONFIG_CONVERSION_STD_STRING_FREE_FUNCTIONS
+
+namespace nonstd {
+
+using std::string_view;
+using std::wstring_view;
+using std::u16string_view;
+using std::u32string_view;
+using std::basic_string_view;
+
+// literal "sv" and "_sv", see above
+
+using std::operator==;
+using std::operator!=;
+using std::operator<;
+using std::operator<=;
+using std::operator>;
+using std::operator>=;
+
+using std::operator<<;
+
+} // namespace nonstd
+
+#else // nssv_HAVE_STD_STRING_VIEW
+
+//
+// Before C++17: use string_view lite:
+//
+
+// Compiler versions:
+//
+// MSVC++  6.0  _MSC_VER == 1200  nssv_COMPILER_MSVC_VERSION ==  60  (Visual Studio 6.0)
+// MSVC++  7.0  _MSC_VER == 1300  nssv_COMPILER_MSVC_VERSION ==  70  (Visual Studio .NET 2002)
+// MSVC++  7.1  _MSC_VER == 1310  nssv_COMPILER_MSVC_VERSION ==  71  (Visual Studio .NET 2003)
+// MSVC++  8.0  _MSC_VER == 1400  nssv_COMPILER_MSVC_VERSION ==  80  (Visual Studio 2005)
+// MSVC++  9.0  _MSC_VER == 1500  nssv_COMPILER_MSVC_VERSION ==  90  (Visual Studio 2008)
+// MSVC++ 10.0  _MSC_VER == 1600  nssv_COMPILER_MSVC_VERSION == 100  (Visual Studio 2010)
+// MSVC++ 11.0  _MSC_VER == 1700  nssv_COMPILER_MSVC_VERSION == 110  (Visual Studio 2012)
+// MSVC++ 12.0  _MSC_VER == 1800  nssv_COMPILER_MSVC_VERSION == 120  (Visual Studio 2013)
+// MSVC++ 14.0  _MSC_VER == 1900  nssv_COMPILER_MSVC_VERSION == 140  (Visual Studio 2015)
+// MSVC++ 14.1  _MSC_VER >= 1910  nssv_COMPILER_MSVC_VERSION == 141  (Visual Studio 2017)
+// MSVC++ 14.2  _MSC_VER >= 1920  nssv_COMPILER_MSVC_VERSION == 142  (Visual Studio 2019)
+
+#if defined(_MSC_VER ) && !defined(__clang__)
+# define nssv_COMPILER_MSVC_VER      (_MSC_VER )
+# define nssv_COMPILER_MSVC_VERSION  (_MSC_VER / 10 - 10 * ( 5 + (_MSC_VER < 1900 ) ) )
+#else
+# define nssv_COMPILER_MSVC_VER      0
+# define nssv_COMPILER_MSVC_VERSION  0
+#endif
+
+#define nssv_COMPILER_VERSION( major, minor, patch )  ( 10 * ( 10 * (major) + (minor) ) + (patch) )
+
+#if defined( __apple_build_version__ )
+# define nssv_COMPILER_APPLECLANG_VERSION  nssv_COMPILER_VERSION(__clang_major__, __clang_minor__, __clang_patchlevel__)
+# define nssv_COMPILER_CLANG_VERSION       0
+#elif defined( __clang__ )
+# define nssv_COMPILER_APPLECLANG_VERSION  0
+# define nssv_COMPILER_CLANG_VERSION       nssv_COMPILER_VERSION(__clang_major__, __clang_minor__, __clang_patchlevel__)
+#else
+# define nssv_COMPILER_APPLECLANG_VERSION  0
+# define nssv_COMPILER_CLANG_VERSION       0
+#endif
+
+#if defined(__GNUC__) && !defined(__clang__)
+# define nssv_COMPILER_GNUC_VERSION  nssv_COMPILER_VERSION(__GNUC__, __GNUC_MINOR__, __GNUC_PATCHLEVEL__)
+#else
+# define nssv_COMPILER_GNUC_VERSION  0
+#endif
+
+// half-open range [lo..hi):
+#define nssv_BETWEEN( v, lo, hi ) ( (lo) <= (v) && (v) < (hi) )
+
+// Presence of language and library features:
+
+#ifdef _HAS_CPP0X
+# define nssv_HAS_CPP0X  _HAS_CPP0X
+#else
+# define nssv_HAS_CPP0X  0
+#endif
+
+// Unless defined otherwise below, consider VC14 as C++11 for string-view-lite:
+
+#if nssv_COMPILER_MSVC_VER >= 1900
+# undef  nssv_CPP11_OR_GREATER
+# define nssv_CPP11_OR_GREATER  1
+#endif
+
+#define nssv_CPP11_90   (nssv_CPP11_OR_GREATER_ || nssv_COMPILER_MSVC_VER >= 1500)
+#define nssv_CPP11_100  (nssv_CPP11_OR_GREATER_ || nssv_COMPILER_MSVC_VER >= 1600)
+#define nssv_CPP11_110  (nssv_CPP11_OR_GREATER_ || nssv_COMPILER_MSVC_VER >= 1700)
+#define nssv_CPP11_120  (nssv_CPP11_OR_GREATER_ || nssv_COMPILER_MSVC_VER >= 1800)
+#define nssv_CPP11_140  (nssv_CPP11_OR_GREATER_ || nssv_COMPILER_MSVC_VER >= 1900)
+#define nssv_CPP11_141  (nssv_CPP11_OR_GREATER_ || nssv_COMPILER_MSVC_VER >= 1910)
+
+#define nssv_CPP14_000  (nssv_CPP14_OR_GREATER)
+#define nssv_CPP17_000  (nssv_CPP17_OR_GREATER)
+
+// Presence of C++11 language features:
+
+#define nssv_HAVE_CONSTEXPR_11          nssv_CPP11_140
+#define nssv_HAVE_EXPLICIT_CONVERSION   nssv_CPP11_140
+#define nssv_HAVE_INLINE_NAMESPACE      nssv_CPP11_140
+#define nssv_HAVE_IS_DEFAULT            nssv_CPP11_140
+#define nssv_HAVE_IS_DELETE             nssv_CPP11_140
+#define nssv_HAVE_NOEXCEPT              nssv_CPP11_140
+#define nssv_HAVE_NULLPTR               nssv_CPP11_100
+#define nssv_HAVE_REF_QUALIFIER         nssv_CPP11_140
+#define nssv_HAVE_UNICODE_LITERALS      nssv_CPP11_140
+#define nssv_HAVE_USER_DEFINED_LITERALS nssv_CPP11_140
+#define nssv_HAVE_WCHAR16_T             nssv_CPP11_100
+#define nssv_HAVE_WCHAR32_T             nssv_CPP11_100
+
+#if ! ( ( nssv_CPP11_OR_GREATER && nssv_COMPILER_CLANG_VERSION ) || nssv_BETWEEN( nssv_COMPILER_CLANG_VERSION, 300, 400 ) )
+# define nssv_HAVE_STD_DEFINED_LITERALS  nssv_CPP11_140
+#else
+# define nssv_HAVE_STD_DEFINED_LITERALS  0
+#endif
+
+// Presence of C++14 language features:
+
+#define nssv_HAVE_CONSTEXPR_14          nssv_CPP14_000
+
+// Presence of C++17 language features:
+
+#define nssv_HAVE_NODISCARD             nssv_CPP17_000
+
+// Presence of C++ library features:
+
+#define nssv_HAVE_STD_HASH              nssv_CPP11_120
+
+// Presence of compiler intrinsics:
+
+// Providing char-type specializations for compare() and length() that
+// use compiler intrinsics can improve compile- and run-time performance.
+//
+// The challenge is in using the right combinations of builtin availability
+// and its constexpr-ness.
+//
+// | compiler | __builtin_memcmp (constexpr) | memcmp  (constexpr) |
+// |----------|------------------------------|---------------------|
+// | clang    | 4.0              (>= 4.0   ) | any     (?        ) |
+// | clang-a  | 9.0              (>= 9.0   ) | any     (?        ) |
+// | gcc      | any              (constexpr) | any     (?        ) |
+// | msvc     | >= 14.2 C++17    (>= 14.2  ) | any     (?        ) |
+
+#define nssv_HAVE_BUILTIN_VER     ( (nssv_CPP17_000 && nssv_COMPILER_MSVC_VERSION >= 142) || nssv_COMPILER_GNUC_VERSION > 0 || nssv_COMPILER_CLANG_VERSION >= 400 || nssv_COMPILER_APPLECLANG_VERSION >= 900 )
+#define nssv_HAVE_BUILTIN_CE      (  nssv_HAVE_BUILTIN_VER )
+
+#define nssv_HAVE_BUILTIN_MEMCMP  ( (nssv_HAVE_CONSTEXPR_14 && nssv_HAVE_BUILTIN_CE) || !nssv_HAVE_CONSTEXPR_14 )
+#define nssv_HAVE_BUILTIN_STRLEN  ( (nssv_HAVE_CONSTEXPR_11 && nssv_HAVE_BUILTIN_CE) || !nssv_HAVE_CONSTEXPR_11 )
+
+#ifdef __has_builtin
+# define nssv_HAVE_BUILTIN( x )  __has_builtin( x )
+#else
+# define nssv_HAVE_BUILTIN( x )  0
+#endif
+
+#if nssv_HAVE_BUILTIN(__builtin_memcmp) || nssv_HAVE_BUILTIN_VER
+# define nssv_BUILTIN_MEMCMP  __builtin_memcmp
+#else
+# define nssv_BUILTIN_MEMCMP  memcmp
+#endif
+
+#if nssv_HAVE_BUILTIN(__builtin_strlen) || nssv_HAVE_BUILTIN_VER
+# define nssv_BUILTIN_STRLEN  __builtin_strlen
+#else
+# define nssv_BUILTIN_STRLEN  strlen
+#endif
+
+// C++ feature usage:
+
+#if nssv_HAVE_CONSTEXPR_11
+# define nssv_constexpr  constexpr
+#else
+# define nssv_constexpr  /*constexpr*/
+#endif
+
+#if  nssv_HAVE_CONSTEXPR_14
+# define nssv_constexpr14  constexpr
+#else
+# define nssv_constexpr14  /*constexpr*/
+#endif
+
+#if nssv_HAVE_EXPLICIT_CONVERSION
+# define nssv_explicit  explicit
+#else
+# define nssv_explicit  /*explicit*/
+#endif
+
+#if nssv_HAVE_INLINE_NAMESPACE
+# define nssv_inline_ns  inline
+#else
+# define nssv_inline_ns  /*inline*/
+#endif
+
+#if nssv_HAVE_NOEXCEPT
+# define nssv_noexcept  noexcept
+#else
+# define nssv_noexcept  /*noexcept*/
+#endif
+
+//#if nssv_HAVE_REF_QUALIFIER
+//# define nssv_ref_qual  &
+//# define nssv_refref_qual  &&
+//#else
+//# define nssv_ref_qual  /*&*/
+//# define nssv_refref_qual  /*&&*/
+//#endif
+
+#if nssv_HAVE_NULLPTR
+# define nssv_nullptr  nullptr
+#else
+# define nssv_nullptr  NULL
+#endif
+
+#if nssv_HAVE_NODISCARD
+# define nssv_nodiscard  simdjson_warn_unused
+#else
+# define nssv_nodiscard  /*simdjson_warn_unused*/
+#endif
+
+// Additional includes:
+
+#include <algorithm>
+#include <cassert>
+#include <iterator>
+#include <limits>
+#include <string>   // std::char_traits<>
+
+#if ! nssv_CONFIG_NO_STREAM_INSERTION
+# include <ostream>
+#endif
+
+#if ! nssv_CONFIG_NO_EXCEPTIONS
+# include <stdexcept>
+#endif
+
+#if nssv_CPP11_OR_GREATER
+# include <type_traits>
+#endif
+
+// Clang, GNUC, MSVC warning suppression macros:
+
+#if defined(__clang__)
+# pragma clang diagnostic ignored "-Wreserved-user-defined-literal"
+# pragma clang diagnostic push
+# pragma clang diagnostic ignored "-Wuser-defined-literals"
+#elif nssv_COMPILER_GNUC_VERSION >= 480
+#  pragma  GCC  diagnostic push
+#  pragma  GCC  diagnostic ignored "-Wliteral-suffix"
+#endif // __clang__
+
+#if nssv_COMPILER_MSVC_VERSION >= 140
+# define nssv_SUPPRESS_MSGSL_WARNING(expr)        [[gsl::suppress(expr)]]
+# define nssv_SUPPRESS_MSVC_WARNING(code, descr)  __pragma(warning(suppress: code) )
+# define nssv_DISABLE_MSVC_WARNINGS(codes)        __pragma(warning(push))  __pragma(warning(disable: codes))
+#else
+# define nssv_SUPPRESS_MSGSL_WARNING(expr)
+# define nssv_SUPPRESS_MSVC_WARNING(code, descr)
+# define nssv_DISABLE_MSVC_WARNINGS(codes)
+#endif
+
+#if defined(__clang__)
+# define nssv_RESTORE_WARNINGS()  _Pragma("clang diagnostic pop")
+#elif nssv_COMPILER_GNUC_VERSION >= 480
+#  define nssv_RESTORE_WARNINGS()  _Pragma("GCC diagnostic pop")
+#elif nssv_COMPILER_MSVC_VERSION >= 140
+# define nssv_RESTORE_WARNINGS()  __pragma(warning(pop ))
+#else
+# define nssv_RESTORE_WARNINGS()
+#endif
+
+// Suppress the following MSVC (GSL) warnings:
+// - C4455, non-gsl   : 'operator ""sv': literal suffix identifiers that do not
+//                      start with an underscore are reserved
+// - C26472, gsl::t.1 : don't use a static_cast for arithmetic conversions;
+//                      use brace initialization, gsl::narrow_cast or gsl::narow
+// - C26481: gsl::b.1 : don't use pointer arithmetic. Use span instead
+
+nssv_DISABLE_MSVC_WARNINGS( 4455 26481 26472 )
+//nssv_DISABLE_CLANG_WARNINGS( "-Wuser-defined-literals" )
+//nssv_DISABLE_GNUC_WARNINGS( -Wliteral-suffix )
+
+namespace nonstd { namespace sv_lite {
+
+//
+// basic_string_view declaration:
+//
+
+template
+<
+    class CharT,
+    class Traits = std::char_traits<CharT>
+>
+class basic_string_view;
+
+namespace detail {
+
+// support constexpr comparison in C++14;
+// for C++17 and later, use provided traits:
+
+template< typename CharT >
+inline nssv_constexpr14 int compare( CharT const * s1, CharT const * s2, std::size_t count )
+{
+    while ( count-- != 0 )
+    {
+        if ( *s1 < *s2 ) return -1;
+        if ( *s1 > *s2 ) return +1;
+        ++s1; ++s2;
+    }
+    return 0;
+}
+
+#if nssv_HAVE_BUILTIN_MEMCMP
+
+// specialization of compare() for char, see also generic compare() above:
+
+inline nssv_constexpr14 int compare( char const * s1, char const * s2, std::size_t count )
+{
+    return nssv_BUILTIN_MEMCMP( s1, s2, count );
+}
+
+#endif
+
+#if nssv_HAVE_BUILTIN_STRLEN
+
+// specialization of length() for char, see also generic length() further below:
+
+inline nssv_constexpr std::size_t length( char const * s )
+{
+    return nssv_BUILTIN_STRLEN( s );
+}
+
+#endif
+
+#if defined(__OPTIMIZE__)
+
+// gcc, clang provide __OPTIMIZE__
+// Expect tail call optimization to make length() non-recursive:
+
+template< typename CharT >
+inline nssv_constexpr std::size_t length( CharT * s, std::size_t result = 0 )
+{
+    return *s == '\0' ? result : length( s + 1, result + 1 );
+}
+
+#else // OPTIMIZE
+
+// non-recursive:
+
+template< typename CharT >
+inline nssv_constexpr14 std::size_t length( CharT * s )
+{
+    std::size_t result = 0;
+    while ( *s++ != '\0' )
+    {
+       ++result;
+    }
+    return result;
+}
+
+#endif // OPTIMIZE
+
+#if nssv_CPP11_OR_GREATER && ! nssv_CPP17_OR_GREATER
+#if defined(__OPTIMIZE__)
+
+// gcc, clang provide __OPTIMIZE__
+// Expect tail call optimization to make search() non-recursive:
+
+template< class CharT, class Traits = std::char_traits<CharT> >
+constexpr const CharT* search( basic_string_view<CharT, Traits> haystack, basic_string_view<CharT, Traits> needle )
+{
+    return haystack.starts_with( needle ) ? haystack.begin() :
+        haystack.empty() ? haystack.end() : search( haystack.substr(1), needle );
+}
+
+#else // OPTIMIZE
+
+// non-recursive:
+
+#if nssv_CONFIG_CONSTEXPR11_STD_SEARCH
+
+template< class CharT, class Traits = std::char_traits<CharT> >
+constexpr const CharT* search( basic_string_view<CharT, Traits> haystack, basic_string_view<CharT, Traits> needle )
+{
+    return std::search( haystack.begin(), haystack.end(), needle.begin(), needle.end() );
+}
+
+#else // nssv_CONFIG_CONSTEXPR11_STD_SEARCH
+
+template< class CharT, class Traits = std::char_traits<CharT> >
+nssv_constexpr14 const CharT* search( basic_string_view<CharT, Traits> haystack, basic_string_view<CharT, Traits> needle )
+{
+    while ( needle.size() <= haystack.size() )
+    {
+        if  ( haystack.starts_with(needle) )
+        {
+            return haystack.cbegin();
+        }
+        haystack = basic_string_view<CharT, Traits>{ haystack.begin() + 1, haystack.size() - 1U };
+    }
+    return haystack.cend();
+}
+#endif // nssv_CONFIG_CONSTEXPR11_STD_SEARCH
+
+#endif // OPTIMIZE
+#endif // nssv_CPP11_OR_GREATER && ! nssv_CPP17_OR_GREATER
+
+} // namespace detail
+
+//
+// basic_string_view:
+//
+
+template
+<
+    class CharT,
+    class Traits /* = std::char_traits<CharT> */
+>
+class basic_string_view
+{
+public:
+    // Member types:
+
+    typedef Traits traits_type;
+    typedef CharT  value_type;
+
+    typedef CharT       * pointer;
+    typedef CharT const * const_pointer;
+    typedef CharT       & reference;
+    typedef CharT const & const_reference;
+
+    typedef const_pointer iterator;
+    typedef const_pointer const_iterator;
+    typedef std::reverse_iterator< const_iterator > reverse_iterator;
+    typedef std::reverse_iterator< const_iterator > const_reverse_iterator;
+
+    typedef std::size_t     size_type;
+    typedef std::ptrdiff_t  difference_type;
+
+    // 24.4.2.1 Construction and assignment:
+
+    nssv_constexpr basic_string_view() nssv_noexcept
+        : data_( nssv_nullptr )
+        , size_( 0 )
+    {}
+
+#if nssv_CPP11_OR_GREATER
+    nssv_constexpr basic_string_view( basic_string_view const & other ) nssv_noexcept = default;
+#else
+    nssv_constexpr basic_string_view( basic_string_view const & other ) nssv_noexcept
+        : data_( other.data_)
+        , size_( other.size_)
+    {}
+#endif
+
+    nssv_constexpr basic_string_view( CharT const * s, size_type count ) nssv_noexcept // non-standard noexcept
+        : data_( s )
+        , size_( count )
+    {}
+
+    nssv_constexpr basic_string_view( CharT const * s) nssv_noexcept // non-standard noexcept
+        : data_( s )
+#if nssv_CPP17_OR_GREATER
+        , size_( Traits::length(s) )
+#elif nssv_CPP11_OR_GREATER
+        , size_( detail::length(s) )
+#else
+        , size_( Traits::length(s) )
+#endif
+    {}
+
+#if  nssv_HAVE_NULLPTR
+# if nssv_HAVE_IS_DELETE
+    nssv_constexpr basic_string_view( std::nullptr_t ) nssv_noexcept = delete;
+# else
+    private: nssv_constexpr basic_string_view( std::nullptr_t ) nssv_noexcept; public:
+# endif
+#endif
+
+    // Assignment:
+
+#if nssv_CPP11_OR_GREATER
+    nssv_constexpr14 basic_string_view & operator=( basic_string_view const & other ) nssv_noexcept = default;
+#else
+    nssv_constexpr14 basic_string_view & operator=( basic_string_view const & other ) nssv_noexcept
+    {
+        data_ = other.data_;
+        size_ = other.size_;
+        return *this;
+    }
+#endif
+
+    // 24.4.2.2 Iterator support:
+
+    nssv_constexpr const_iterator begin()  const nssv_noexcept { return data_;         }
+    nssv_constexpr const_iterator end()    const nssv_noexcept { return data_ + size_; }
+
+    nssv_constexpr const_iterator cbegin() const nssv_noexcept { return begin(); }
+    nssv_constexpr const_iterator cend()   const nssv_noexcept { return end();   }
+
+    nssv_constexpr const_reverse_iterator rbegin()  const nssv_noexcept { return const_reverse_iterator( end() );   }
+    nssv_constexpr const_reverse_iterator rend()    const nssv_noexcept { return const_reverse_iterator( begin() ); }
+
+    nssv_constexpr const_reverse_iterator crbegin() const nssv_noexcept { return rbegin(); }
+    nssv_constexpr const_reverse_iterator crend()   const nssv_noexcept { return rend();   }
+
+    // 24.4.2.3 Capacity:
+
+    nssv_constexpr size_type size()     const nssv_noexcept { return size_; }
+    nssv_constexpr size_type length()   const nssv_noexcept { return size_; }
+    nssv_constexpr size_type max_size() const nssv_noexcept { return (std::numeric_limits< size_type >::max)(); }
+
+    // since C++20
+    nssv_nodiscard nssv_constexpr bool empty() const nssv_noexcept
+    {
+        return 0 == size_;
+    }
+
+    // 24.4.2.4 Element access:
+
+    nssv_constexpr const_reference operator[]( size_type pos ) const
+    {
+        return data_at( pos );
+    }
+
+    nssv_constexpr14 const_reference at( size_type pos ) const
+    {
+#if nssv_CONFIG_NO_EXCEPTIONS
+        assert( pos < size() );
+#else
+        if ( pos >= size() )
+        {
+            throw std::out_of_range("nonstd::string_view::at()");
+        }
+#endif
+        return data_at( pos );
+    }
+
+    nssv_constexpr const_reference front() const { return data_at( 0 );          }
+    nssv_constexpr const_reference back()  const { return data_at( size() - 1 ); }
+
+    nssv_constexpr const_pointer   data()  const nssv_noexcept { return data_; }
+
+    // 24.4.2.5 Modifiers:
+
+    nssv_constexpr14 void remove_prefix( size_type n )
+    {
+        assert( n <= size() );
+        data_ += n;
+        size_ -= n;
+    }
+
+    nssv_constexpr14 void remove_suffix( size_type n )
+    {
+        assert( n <= size() );
+        size_ -= n;
+    }
+
+    nssv_constexpr14 void swap( basic_string_view & other ) nssv_noexcept
+    {
+        const basic_string_view tmp(other);
+        other = *this;
+        *this = tmp;
+    }
+
+    // 24.4.2.6 String operations:
+
+    size_type copy( CharT * dest, size_type n, size_type pos = 0 ) const
+    {
+#if nssv_CONFIG_NO_EXCEPTIONS
+        assert( pos <= size() );
+#else
+        if ( pos > size() )
+        {
+            throw std::out_of_range("nonstd::string_view::copy()");
+        }
+#endif
+        const size_type rlen = (std::min)( n, size() - pos );
+
+        (void) Traits::copy( dest, data() + pos, rlen );
+
+        return rlen;
+    }
+
+    nssv_constexpr14 basic_string_view substr( size_type pos = 0, size_type n = npos ) const
+    {
+#if nssv_CONFIG_NO_EXCEPTIONS
+        assert( pos <= size() );
+#else
+        if ( pos > size() )
+        {
+            throw std::out_of_range("nonstd::string_view::substr()");
+        }
+#endif
+        return basic_string_view( data() + pos, (std::min)( n, size() - pos ) );
+    }
+
+    // compare(), 6x:
+
+    nssv_constexpr14 int compare( basic_string_view other ) const nssv_noexcept // (1)
+    {
+#if nssv_CPP17_OR_GREATER
+        if ( const int result = Traits::compare( data(), other.data(), (std::min)( size(), other.size() ) ) )
+#else
+        if ( const int result = detail::compare( data(), other.data(), (std::min)( size(), other.size() ) ) )
+#endif
+        {
+            return result;
+        }
+
+        return size() == other.size() ? 0 : size() < other.size() ? -1 : 1;
+    }
+
+    nssv_constexpr int compare( size_type pos1, size_type n1, basic_string_view other ) const // (2)
+    {
+        return substr( pos1, n1 ).compare( other );
+    }
+
+    nssv_constexpr int compare( size_type pos1, size_type n1, basic_string_view other, size_type pos2, size_type n2 ) const // (3)
+    {
+        return substr( pos1, n1 ).compare( other.substr( pos2, n2 ) );
+    }
+
+    nssv_constexpr int compare( CharT const * s ) const // (4)
+    {
+        return compare( basic_string_view( s ) );
+    }
+
+    nssv_constexpr int compare( size_type pos1, size_type n1, CharT const * s ) const // (5)
+    {
+        return substr( pos1, n1 ).compare( basic_string_view( s ) );
+    }
+
+    nssv_constexpr int compare( size_type pos1, size_type n1, CharT const * s, size_type n2 ) const // (6)
+    {
+        return substr( pos1, n1 ).compare( basic_string_view( s, n2 ) );
+    }
+
+    // 24.4.2.7 Searching:
+
+    // starts_with(), 3x, since C++20:
+
+    nssv_constexpr bool starts_with( basic_string_view v ) const nssv_noexcept  // (1)
+    {
+        return size() >= v.size() && compare( 0, v.size(), v ) == 0;
+    }
+
+    nssv_constexpr bool starts_with( CharT c ) const nssv_noexcept  // (2)
+    {
+        return starts_with( basic_string_view( &c, 1 ) );
+    }
+
+    nssv_constexpr bool starts_with( CharT const * s ) const  // (3)
+    {
+        return starts_with( basic_string_view( s ) );
+    }
+
+    // ends_with(), 3x, since C++20:
+
+    nssv_constexpr bool ends_with( basic_string_view v ) const nssv_noexcept  // (1)
+    {
+        return size() >= v.size() && compare( size() - v.size(), npos, v ) == 0;
+    }
+
+    nssv_constexpr bool ends_with( CharT c ) const nssv_noexcept  // (2)
+    {
+        return ends_with( basic_string_view( &c, 1 ) );
+    }
+
+    nssv_constexpr bool ends_with( CharT const * s ) const  // (3)
+    {
+        return ends_with( basic_string_view( s ) );
+    }
+
+    // find(), 4x:
+
+    nssv_constexpr14 size_type find( basic_string_view v, size_type pos = 0 ) const nssv_noexcept  // (1)
+    {
+        return assert( v.size() == 0 || v.data() != nssv_nullptr )
+            , pos >= size()
+            ? npos : to_pos(
+#if nssv_CPP11_OR_GREATER && ! nssv_CPP17_OR_GREATER
+                detail::search( substr(pos), v )
+#else
+                std::search( cbegin() + pos, cend(), v.cbegin(), v.cend(), Traits::eq )
+#endif
+            );
+    }
+
+    nssv_constexpr size_type find( CharT c, size_type pos = 0 ) const nssv_noexcept  // (2)
+    {
+        return find( basic_string_view( &c, 1 ), pos );
+    }
+
+    nssv_constexpr size_type find( CharT const * s, size_type pos, size_type n ) const  // (3)
+    {
+        return find( basic_string_view( s, n ), pos );
+    }
+
+    nssv_constexpr size_type find( CharT const * s, size_type pos = 0 ) const  // (4)
+    {
+        return find( basic_string_view( s ), pos );
+    }
+
+    // rfind(), 4x:
+
+    nssv_constexpr14 size_type rfind( basic_string_view v, size_type pos = npos ) const nssv_noexcept  // (1)
+    {
+        if ( size() < v.size() )
+        {
+            return npos;
+        }
+
+        if ( v.empty() )
+        {
+            return (std::min)( size(), pos );
+        }
+
+        const_iterator last   = cbegin() + (std::min)( size() - v.size(), pos ) + v.size();
+        const_iterator result = std::find_end( cbegin(), last, v.cbegin(), v.cend(), Traits::eq );
+
+        return result != last ? size_type( result - cbegin() ) : npos;
+    }
+
+    nssv_constexpr14 size_type rfind( CharT c, size_type pos = npos ) const nssv_noexcept  // (2)
+    {
+        return rfind( basic_string_view( &c, 1 ), pos );
+    }
+
+    nssv_constexpr14 size_type rfind( CharT const * s, size_type pos, size_type n ) const  // (3)
+    {
+        return rfind( basic_string_view( s, n ), pos );
+    }
+
+    nssv_constexpr14 size_type rfind( CharT const * s, size_type pos = npos ) const  // (4)
+    {
+        return rfind( basic_string_view( s ), pos );
+    }
+
+    // find_first_of(), 4x:
+
+    nssv_constexpr size_type find_first_of( basic_string_view v, size_type pos = 0 ) const nssv_noexcept  // (1)
+    {
+        return pos >= size()
+            ? npos
+            : to_pos( std::find_first_of( cbegin() + pos, cend(), v.cbegin(), v.cend(), Traits::eq ) );
+    }
+
+    nssv_constexpr size_type find_first_of( CharT c, size_type pos = 0 ) const nssv_noexcept  // (2)
+    {
+        return find_first_of( basic_string_view( &c, 1 ), pos );
+    }
+
+    nssv_constexpr size_type find_first_of( CharT const * s, size_type pos, size_type n ) const  // (3)
+    {
+        return find_first_of( basic_string_view( s, n ), pos );
+    }
+
+    nssv_constexpr size_type find_first_of(  CharT const * s, size_type pos = 0 ) const  // (4)
+    {
+        return find_first_of( basic_string_view( s ), pos );
+    }
+
+    // find_last_of(), 4x:
+
+    nssv_constexpr size_type find_last_of( basic_string_view v, size_type pos = npos ) const nssv_noexcept  // (1)
+    {
+        return empty()
+            ? npos
+            : pos >= size()
+            ? find_last_of( v, size() - 1 )
+            : to_pos( std::find_first_of( const_reverse_iterator( cbegin() + pos + 1 ), crend(), v.cbegin(), v.cend(), Traits::eq ) );
+    }
+
+    nssv_constexpr size_type find_last_of( CharT c, size_type pos = npos ) const nssv_noexcept  // (2)
+    {
+        return find_last_of( basic_string_view( &c, 1 ), pos );
+    }
+
+    nssv_constexpr size_type find_last_of( CharT const * s, size_type pos, size_type count ) const  // (3)
+    {
+        return find_last_of( basic_string_view( s, count ), pos );
+    }
+
+    nssv_constexpr size_type find_last_of( CharT const * s, size_type pos = npos ) const  // (4)
+    {
+        return find_last_of( basic_string_view( s ), pos );
+    }
+
+    // find_first_not_of(), 4x:
+
+    nssv_constexpr size_type find_first_not_of( basic_string_view v, size_type pos = 0 ) const nssv_noexcept  // (1)
+    {
+        return pos >= size()
+            ? npos
+            : to_pos( std::find_if( cbegin() + pos, cend(), not_in_view( v ) ) );
+    }
+
+    nssv_constexpr size_type find_first_not_of( CharT c, size_type pos = 0 ) const nssv_noexcept  // (2)
+    {
+        return find_first_not_of( basic_string_view( &c, 1 ), pos );
+    }
+
+    nssv_constexpr size_type find_first_not_of( CharT const * s, size_type pos, size_type count ) const  // (3)
+    {
+        return find_first_not_of( basic_string_view( s, count ), pos );
+    }
+
+    nssv_constexpr size_type find_first_not_of( CharT const * s, size_type pos = 0 ) const  // (4)
+    {
+        return find_first_not_of( basic_string_view( s ), pos );
+    }
+
+    // find_last_not_of(), 4x:
+
+    nssv_constexpr size_type find_last_not_of( basic_string_view v, size_type pos = npos ) const nssv_noexcept  // (1)
+    {
+        return empty()
+            ? npos
+            : pos >= size()
+            ? find_last_not_of( v, size() - 1 )
+            : to_pos( std::find_if( const_reverse_iterator( cbegin() + pos + 1 ), crend(), not_in_view( v ) ) );
+    }
+
+    nssv_constexpr size_type find_last_not_of( CharT c, size_type pos = npos ) const nssv_noexcept  // (2)
+    {
+        return find_last_not_of( basic_string_view( &c, 1 ), pos );
+    }
+
+    nssv_constexpr size_type find_last_not_of( CharT const * s, size_type pos, size_type count ) const  // (3)
+    {
+        return find_last_not_of( basic_string_view( s, count ), pos );
+    }
+
+    nssv_constexpr size_type find_last_not_of( CharT const * s, size_type pos = npos ) const  // (4)
+    {
+        return find_last_not_of( basic_string_view( s ), pos );
+    }
+
+    // Constants:
+
+#if nssv_CPP17_OR_GREATER
+    static nssv_constexpr size_type npos = size_type(-1);
+#elif nssv_CPP11_OR_GREATER
+    enum : size_type { npos = size_type(-1) };
+#else
+    enum { npos = size_type(-1) };
+#endif
+
+private:
+    struct not_in_view
+    {
+        const basic_string_view v;
+
+        nssv_constexpr explicit not_in_view( basic_string_view v_ ) : v( v_ ) {}
+
+        nssv_constexpr bool operator()( CharT c ) const
+        {
+            return npos == v.find_first_of( c );
+        }
+    };
+
+    nssv_constexpr size_type to_pos( const_iterator it ) const
+    {
+        return it == cend() ? npos : size_type( it - cbegin() );
+    }
+
+    nssv_constexpr size_type to_pos( const_reverse_iterator it ) const
+    {
+        return it == crend() ? npos : size_type( crend() - it - 1 );
+    }
+
+    nssv_constexpr const_reference data_at( size_type pos ) const
+    {
+#if nssv_BETWEEN( nssv_COMPILER_GNUC_VERSION, 1, 500 )
+        return data_[pos];
+#else
+        return assert( pos < size() ), data_[pos];
+#endif
+    }
+
+private:
+    const_pointer data_;
+    size_type     size_;
+
+public:
+#if nssv_CONFIG_CONVERSION_STD_STRING_CLASS_METHODS
+
+    template< class Allocator >
+    basic_string_view( std::basic_string<CharT, Traits, Allocator> const & s ) nssv_noexcept
+        : data_( s.data() )
+        , size_( s.size() )
+    {}
+
+#if nssv_HAVE_EXPLICIT_CONVERSION
+
+    template< class Allocator >
+    explicit operator std::basic_string<CharT, Traits, Allocator>() const
+    {
+        return to_string( Allocator() );
+    }
+
+#endif // nssv_HAVE_EXPLICIT_CONVERSION
+
+#if nssv_CPP11_OR_GREATER
+
+    template< class Allocator = std::allocator<CharT> >
+    std::basic_string<CharT, Traits, Allocator>
+    to_string( Allocator const & a = Allocator() ) const
+    {
+        return std::basic_string<CharT, Traits, Allocator>( begin(), end(), a );
+    }
+
+#else
+
+    std::basic_string<CharT, Traits>
+    to_string() const
+    {
+        return std::basic_string<CharT, Traits>( begin(), end() );
+    }
+
+    template< class Allocator >
+    std::basic_string<CharT, Traits, Allocator>
+    to_string( Allocator const & a ) const
+    {
+        return std::basic_string<CharT, Traits, Allocator>( begin(), end(), a );
+    }
+
+#endif // nssv_CPP11_OR_GREATER
+
+#endif // nssv_CONFIG_CONVERSION_STD_STRING_CLASS_METHODS
+};
+
+//
+// Non-member functions:
+//
+
+// 24.4.3 Non-member comparison functions:
+// lexicographically compare two string views (function template):
+
+template< class CharT, class Traits >
+nssv_constexpr bool operator== (
+    basic_string_view <CharT, Traits> lhs,
+    basic_string_view <CharT, Traits> rhs ) nssv_noexcept
+{ return lhs.size() == rhs.size() && lhs.compare( rhs ) == 0; }
+
+template< class CharT, class Traits >
+nssv_constexpr bool operator!= (
+    basic_string_view <CharT, Traits> lhs,
+    basic_string_view <CharT, Traits> rhs ) nssv_noexcept
+{ return !( lhs == rhs ); }
+
+template< class CharT, class Traits >
+nssv_constexpr bool operator< (
+    basic_string_view <CharT, Traits> lhs,
+    basic_string_view <CharT, Traits> rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) < 0; }
+
+template< class CharT, class Traits >
+nssv_constexpr bool operator<= (
+    basic_string_view <CharT, Traits> lhs,
+    basic_string_view <CharT, Traits> rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) <= 0; }
+
+template< class CharT, class Traits >
+nssv_constexpr bool operator> (
+    basic_string_view <CharT, Traits> lhs,
+    basic_string_view <CharT, Traits> rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) > 0; }
+
+template< class CharT, class Traits >
+nssv_constexpr bool operator>= (
+    basic_string_view <CharT, Traits> lhs,
+    basic_string_view <CharT, Traits> rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) >= 0; }
+
+// Let S be basic_string_view<CharT, Traits>, and sv be an instance of S.
+// Implementations shall provide sufficient additional overloads marked
+// constexpr and noexcept so that an object t with an implicit conversion
+// to S can be compared according to Table 67.
+
+#if ! nssv_CPP11_OR_GREATER || nssv_BETWEEN( nssv_COMPILER_MSVC_VERSION, 100, 141 )
+
+// accommodate for older compilers:
+
+// ==
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator==(
+    basic_string_view<CharT, Traits> lhs,
+    CharT const * rhs ) nssv_noexcept
+{ return lhs.size() == detail::length( rhs ) && lhs.compare( rhs ) == 0; }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator==(
+    CharT const * lhs,
+    basic_string_view<CharT, Traits> rhs ) nssv_noexcept
+{ return detail::length( lhs ) == rhs.size() && rhs.compare( lhs ) == 0; }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator==(
+    basic_string_view<CharT, Traits> lhs,
+    std::basic_string<CharT, Traits> rhs ) nssv_noexcept
+{ return lhs.size() == rhs.size() && lhs.compare( rhs ) == 0; }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator==(
+    std::basic_string<CharT, Traits> rhs,
+    basic_string_view<CharT, Traits> lhs ) nssv_noexcept
+{ return lhs.size() == rhs.size() && lhs.compare( rhs ) == 0; }
+
+// !=
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator!=(
+    basic_string_view<CharT, Traits> lhs,
+    CharT const * rhs ) nssv_noexcept
+{ return !( lhs == rhs ); }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator!=(
+    CharT const * lhs,
+    basic_string_view<CharT, Traits> rhs ) nssv_noexcept
+{ return !( lhs == rhs ); }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator!=(
+    basic_string_view<CharT, Traits> lhs,
+    std::basic_string<CharT, Traits> rhs ) nssv_noexcept
+{ return !( lhs == rhs ); }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator!=(
+    std::basic_string<CharT, Traits> rhs,
+    basic_string_view<CharT, Traits> lhs ) nssv_noexcept
+{ return !( lhs == rhs ); }
+
+// <
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator<(
+    basic_string_view<CharT, Traits> lhs,
+    CharT const * rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) < 0; }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator<(
+    CharT const * lhs,
+    basic_string_view<CharT, Traits> rhs ) nssv_noexcept
+{ return rhs.compare( lhs ) > 0; }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator<(
+    basic_string_view<CharT, Traits> lhs,
+    std::basic_string<CharT, Traits> rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) < 0; }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator<(
+    std::basic_string<CharT, Traits> rhs,
+    basic_string_view<CharT, Traits> lhs ) nssv_noexcept
+{ return rhs.compare( lhs ) > 0; }
+
+// <=
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator<=(
+    basic_string_view<CharT, Traits> lhs,
+    CharT const * rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) <= 0; }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator<=(
+    CharT const * lhs,
+    basic_string_view<CharT, Traits> rhs ) nssv_noexcept
+{ return rhs.compare( lhs ) >= 0; }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator<=(
+    basic_string_view<CharT, Traits> lhs,
+    std::basic_string<CharT, Traits> rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) <= 0; }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator<=(
+    std::basic_string<CharT, Traits> rhs,
+    basic_string_view<CharT, Traits> lhs ) nssv_noexcept
+{ return rhs.compare( lhs ) >= 0; }
+
+// >
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator>(
+    basic_string_view<CharT, Traits> lhs,
+    CharT const * rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) > 0; }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator>(
+    CharT const * lhs,
+    basic_string_view<CharT, Traits> rhs ) nssv_noexcept
+{ return rhs.compare( lhs ) < 0; }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator>(
+    basic_string_view<CharT, Traits> lhs,
+    std::basic_string<CharT, Traits> rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) > 0; }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator>(
+    std::basic_string<CharT, Traits> rhs,
+    basic_string_view<CharT, Traits> lhs ) nssv_noexcept
+{ return rhs.compare( lhs ) < 0; }
+
+// >=
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator>=(
+    basic_string_view<CharT, Traits> lhs,
+    CharT const * rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) >= 0; }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator>=(
+    CharT const * lhs,
+    basic_string_view<CharT, Traits> rhs ) nssv_noexcept
+{ return rhs.compare( lhs ) <= 0; }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator>=(
+    basic_string_view<CharT, Traits> lhs,
+    std::basic_string<CharT, Traits> rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) >= 0; }
+
+template< class CharT, class Traits>
+nssv_constexpr bool operator>=(
+    std::basic_string<CharT, Traits> rhs,
+    basic_string_view<CharT, Traits> lhs ) nssv_noexcept
+{ return rhs.compare( lhs ) <= 0; }
+
+#else // newer compilers:
+
+#define nssv_BASIC_STRING_VIEW_I(T,U)  typename std::decay< basic_string_view<T,U> >::type
+
+#if defined(_MSC_VER)       // issue 40
+# define nssv_MSVC_ORDER(x)  , int=x
+#else
+# define nssv_MSVC_ORDER(x)  /*, int=x*/
+#endif
+
+// ==
+
+template< class CharT, class Traits  nssv_MSVC_ORDER(1) >
+nssv_constexpr bool operator==(
+         basic_string_view  <CharT, Traits> lhs,
+    nssv_BASIC_STRING_VIEW_I(CharT, Traits) rhs ) nssv_noexcept
+{ return lhs.size() == rhs.size() && lhs.compare( rhs ) == 0; }
+
+template< class CharT, class Traits  nssv_MSVC_ORDER(2) >
+nssv_constexpr bool operator==(
+    nssv_BASIC_STRING_VIEW_I(CharT, Traits) lhs,
+         basic_string_view  <CharT, Traits> rhs ) nssv_noexcept
+{ return lhs.size() == rhs.size() && lhs.compare( rhs ) == 0; }
+
+// !=
+
+template< class CharT, class Traits  nssv_MSVC_ORDER(1) >
+nssv_constexpr bool operator!= (
+         basic_string_view  < CharT, Traits > lhs,
+    nssv_BASIC_STRING_VIEW_I( CharT, Traits ) rhs ) nssv_noexcept
+{ return !( lhs == rhs ); }
+
+template< class CharT, class Traits  nssv_MSVC_ORDER(2) >
+nssv_constexpr bool operator!= (
+    nssv_BASIC_STRING_VIEW_I( CharT, Traits ) lhs,
+         basic_string_view  < CharT, Traits > rhs ) nssv_noexcept
+{ return !( lhs == rhs ); }
+
+// <
+
+template< class CharT, class Traits  nssv_MSVC_ORDER(1) >
+nssv_constexpr bool operator< (
+         basic_string_view  < CharT, Traits > lhs,
+    nssv_BASIC_STRING_VIEW_I( CharT, Traits ) rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) < 0; }
+
+template< class CharT, class Traits  nssv_MSVC_ORDER(2) >
+nssv_constexpr bool operator< (
+    nssv_BASIC_STRING_VIEW_I( CharT, Traits ) lhs,
+         basic_string_view  < CharT, Traits > rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) < 0; }
+
+// <=
+
+template< class CharT, class Traits  nssv_MSVC_ORDER(1) >
+nssv_constexpr bool operator<= (
+         basic_string_view  < CharT, Traits > lhs,
+    nssv_BASIC_STRING_VIEW_I( CharT, Traits ) rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) <= 0; }
+
+template< class CharT, class Traits  nssv_MSVC_ORDER(2) >
+nssv_constexpr bool operator<= (
+    nssv_BASIC_STRING_VIEW_I( CharT, Traits ) lhs,
+         basic_string_view  < CharT, Traits > rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) <= 0; }
+
+// >
+
+template< class CharT, class Traits  nssv_MSVC_ORDER(1) >
+nssv_constexpr bool operator> (
+         basic_string_view  < CharT, Traits > lhs,
+    nssv_BASIC_STRING_VIEW_I( CharT, Traits ) rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) > 0; }
+
+template< class CharT, class Traits  nssv_MSVC_ORDER(2) >
+nssv_constexpr bool operator> (
+    nssv_BASIC_STRING_VIEW_I( CharT, Traits ) lhs,
+         basic_string_view  < CharT, Traits > rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) > 0; }
+
+// >=
+
+template< class CharT, class Traits  nssv_MSVC_ORDER(1) >
+nssv_constexpr bool operator>= (
+         basic_string_view  < CharT, Traits > lhs,
+    nssv_BASIC_STRING_VIEW_I( CharT, Traits ) rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) >= 0; }
+
+template< class CharT, class Traits  nssv_MSVC_ORDER(2) >
+nssv_constexpr bool operator>= (
+    nssv_BASIC_STRING_VIEW_I( CharT, Traits ) lhs,
+         basic_string_view  < CharT, Traits > rhs ) nssv_noexcept
+{ return lhs.compare( rhs ) >= 0; }
+
+#undef nssv_MSVC_ORDER
+#undef nssv_BASIC_STRING_VIEW_I
+
+#endif // compiler-dependent approach to comparisons
+
+// 24.4.4 Inserters and extractors:
+
+#if ! nssv_CONFIG_NO_STREAM_INSERTION
+
+namespace detail {
+
+template< class Stream >
+void write_padding( Stream & os, std::streamsize n )
+{
+    for ( std::streamsize i = 0; i < n; ++i )
+        os.rdbuf()->sputc( os.fill() );
+}
+
+template< class Stream, class View >
+Stream & write_to_stream( Stream & os, View const & sv )
+{
+    typename Stream::sentry sentry( os );
+
+    if ( !sentry )
+        return os;
+
+    const std::streamsize length = static_cast<std::streamsize>( sv.length() );
+
+    // Whether, and how, to pad:
+    const bool      pad = ( length < os.width() );
+    const bool left_pad = pad && ( os.flags() & std::ios_base::adjustfield ) == std::ios_base::right;
+
+    if ( left_pad )
+        write_padding( os, os.width() - length );
+
+    // Write span characters:
+    os.rdbuf()->sputn( sv.begin(), length );
+
+    if ( pad && !left_pad )
+        write_padding( os, os.width() - length );
+
+    // Reset output stream width:
+    os.width( 0 );
+
+    return os;
+}
+
+} // namespace detail
+
+template< class CharT, class Traits >
+std::basic_ostream<CharT, Traits> &
+operator<<(
+    std::basic_ostream<CharT, Traits>& os,
+    basic_string_view <CharT, Traits> sv )
+{
+    return detail::write_to_stream( os, sv );
+}
+
+#endif // nssv_CONFIG_NO_STREAM_INSERTION
+
+// Several typedefs for common character types are provided:
+
+typedef basic_string_view<char>      string_view;
+typedef basic_string_view<wchar_t>   wstring_view;
+#if nssv_HAVE_WCHAR16_T
+typedef basic_string_view<char16_t>  u16string_view;
+typedef basic_string_view<char32_t>  u32string_view;
+#endif
+
+}} // namespace nonstd::sv_lite
+
+//
+// 24.4.6 Suffix for basic_string_view literals:
+//
+
+#if nssv_HAVE_USER_DEFINED_LITERALS
+
+namespace nonstd {
+nssv_inline_ns namespace literals {
+nssv_inline_ns namespace string_view_literals {
+
+#if nssv_CONFIG_STD_SV_OPERATOR && nssv_HAVE_STD_DEFINED_LITERALS
+
+nssv_constexpr nonstd::sv_lite::string_view operator ""sv( const char* str, size_t len ) nssv_noexcept  // (1)
+{
+    return nonstd::sv_lite::string_view{ str, len };
+}
+
+nssv_constexpr nonstd::sv_lite::u16string_view operator ""sv( const char16_t* str, size_t len ) nssv_noexcept  // (2)
+{
+    return nonstd::sv_lite::u16string_view{ str, len };
+}
+
+nssv_constexpr nonstd::sv_lite::u32string_view operator ""sv( const char32_t* str, size_t len ) nssv_noexcept  // (3)
+{
+    return nonstd::sv_lite::u32string_view{ str, len };
+}
+
+nssv_constexpr nonstd::sv_lite::wstring_view operator ""sv( const wchar_t* str, size_t len ) nssv_noexcept  // (4)
+{
+    return nonstd::sv_lite::wstring_view{ str, len };
+}
+
+#endif // nssv_CONFIG_STD_SV_OPERATOR && nssv_HAVE_STD_DEFINED_LITERALS
+
+#if nssv_CONFIG_USR_SV_OPERATOR
+
+nssv_constexpr nonstd::sv_lite::string_view operator ""_sv( const char* str, size_t len ) nssv_noexcept  // (1)
+{
+    return nonstd::sv_lite::string_view{ str, len };
+}
+
+nssv_constexpr nonstd::sv_lite::u16string_view operator ""_sv( const char16_t* str, size_t len ) nssv_noexcept  // (2)
+{
+    return nonstd::sv_lite::u16string_view{ str, len };
+}
+
+nssv_constexpr nonstd::sv_lite::u32string_view operator ""_sv( const char32_t* str, size_t len ) nssv_noexcept  // (3)
+{
+    return nonstd::sv_lite::u32string_view{ str, len };
+}
+
+nssv_constexpr nonstd::sv_lite::wstring_view operator ""_sv( const wchar_t* str, size_t len ) nssv_noexcept  // (4)
+{
+    return nonstd::sv_lite::wstring_view{ str, len };
+}
+
+#endif // nssv_CONFIG_USR_SV_OPERATOR
+
+}}} // namespace nonstd::literals::string_view_literals
+
+#endif
+
+//
+// Extensions for std::string:
+//
+
+#if nssv_CONFIG_CONVERSION_STD_STRING_FREE_FUNCTIONS
+
+namespace nonstd {
+namespace sv_lite {
+
+// Exclude MSVC 14 (19.00): it yields ambiguous to_string():
+
+#if nssv_CPP11_OR_GREATER && nssv_COMPILER_MSVC_VERSION != 140
+
+template< class CharT, class Traits, class Allocator = std::allocator<CharT> >
+std::basic_string<CharT, Traits, Allocator>
+to_string( basic_string_view<CharT, Traits> v, Allocator const & a = Allocator() )
+{
+    return std::basic_string<CharT,Traits, Allocator>( v.begin(), v.end(), a );
+}
+
+#else
+
+template< class CharT, class Traits >
+std::basic_string<CharT, Traits>
+to_string( basic_string_view<CharT, Traits> v )
+{
+    return std::basic_string<CharT, Traits>( v.begin(), v.end() );
+}
+
+template< class CharT, class Traits, class Allocator >
+std::basic_string<CharT, Traits, Allocator>
+to_string( basic_string_view<CharT, Traits> v, Allocator const & a )
+{
+    return std::basic_string<CharT, Traits, Allocator>( v.begin(), v.end(), a );
+}
+
+#endif // nssv_CPP11_OR_GREATER
+
+template< class CharT, class Traits, class Allocator >
+basic_string_view<CharT, Traits>
+to_string_view( std::basic_string<CharT, Traits, Allocator> const & s )
+{
+    return basic_string_view<CharT, Traits>( s.data(), s.size() );
+}
+
+}} // namespace nonstd::sv_lite
+
+#endif // nssv_CONFIG_CONVERSION_STD_STRING_FREE_FUNCTIONS
+
+//
+// make types and algorithms available in namespace nonstd:
+//
+
+namespace nonstd {
+
+using sv_lite::basic_string_view;
+using sv_lite::string_view;
+using sv_lite::wstring_view;
+
+#if nssv_HAVE_WCHAR16_T
+using sv_lite::u16string_view;
+#endif
+#if nssv_HAVE_WCHAR32_T
+using sv_lite::u32string_view;
+#endif
+
+// literal "sv"
+
+using sv_lite::operator==;
+using sv_lite::operator!=;
+using sv_lite::operator<;
+using sv_lite::operator<=;
+using sv_lite::operator>;
+using sv_lite::operator>=;
+
+#if ! nssv_CONFIG_NO_STREAM_INSERTION
+using sv_lite::operator<<;
+#endif
+
+#if nssv_CONFIG_CONVERSION_STD_STRING_FREE_FUNCTIONS
+using sv_lite::to_string;
+using sv_lite::to_string_view;
+#endif
+
+} // namespace nonstd
+
+// 24.4.5 Hash support (C++11):
+
+// Note: The hash value of a string view object is equal to the hash value of
+// the corresponding string object.
+
+#if nssv_HAVE_STD_HASH
+
+#include <functional>
+
+namespace std {
+
+template<>
+struct hash< nonstd::string_view >
+{
+public:
+    std::size_t operator()( nonstd::string_view v ) const nssv_noexcept
+    {
+        return std::hash<std::string>()( std::string( v.data(), v.size() ) );
+    }
+};
+
+template<>
+struct hash< nonstd::wstring_view >
+{
+public:
+    std::size_t operator()( nonstd::wstring_view v ) const nssv_noexcept
+    {
+        return std::hash<std::wstring>()( std::wstring( v.data(), v.size() ) );
+    }
+};
+
+template<>
+struct hash< nonstd::u16string_view >
+{
+public:
+    std::size_t operator()( nonstd::u16string_view v ) const nssv_noexcept
+    {
+        return std::hash<std::u16string>()( std::u16string( v.data(), v.size() ) );
+    }
+};
+
+template<>
+struct hash< nonstd::u32string_view >
+{
+public:
+    std::size_t operator()( nonstd::u32string_view v ) const nssv_noexcept
+    {
+        return std::hash<std::u32string>()( std::u32string( v.data(), v.size() ) );
+    }
+};
+
+} // namespace std
+
+#endif // nssv_HAVE_STD_HASH
+
+nssv_RESTORE_WARNINGS()
+
+#endif // nssv_HAVE_STD_STRING_VIEW
+#endif // NONSTD_SV_LITE_H_INCLUDED
+/* end file simdjson/nonstd/string_view.hpp */
+SIMDJSON_POP_DISABLE_WARNINGS
+
+namespace std {
+  using string_view = nonstd::string_view;
+}
+#endif // SIMDJSON_HAS_STRING_VIEW
+#undef SIMDJSON_HAS_STRING_VIEW // We are not going to need this macro anymore.
+
+/// If EXPR is an error, returns it.
+#define SIMDJSON_TRY(EXPR) { auto _err = (EXPR); if (_err) { return _err; } }
+
+// Unless the programmer has already set SIMDJSON_DEVELOPMENT_CHECKS,
+// we want to set it under debug builds. We detect a debug build
+// under Visual Studio when the _DEBUG macro is set. Under the other
+// compilers, we use the fact that they define __OPTIMIZE__ whenever
+// they allow optimizations.
+// It is possible that this could miss some cases where SIMDJSON_DEVELOPMENT_CHECKS
+// is helpful, but the programmer can set the macro SIMDJSON_DEVELOPMENT_CHECKS.
+// It could also wrongly set SIMDJSON_DEVELOPMENT_CHECKS (e.g., if the programmer
+// sets _DEBUG in a release build under Visual Studio, or if some compiler fails to
+// set the __OPTIMIZE__ macro).
+// We make it so that if NDEBUG is defined, then SIMDJSON_DEVELOPMENT_CHECKS
+// is not defined, irrespective of the compiler.
+// We recommend that users set NDEBUG in release builds, so that
+// SIMDJSON_DEVELOPMENT_CHECKS is not defined in release builds by default,
+// irrespective of the compiler.
+#ifndef SIMDJSON_DEVELOPMENT_CHECKS
+#ifdef _MSC_VER
+// Visual Studio seems to set _DEBUG for debug builds.
+// We set SIMDJSON_DEVELOPMENT_CHECKS to 1 if _DEBUG is defined
+// and NDEBUG is not defined.
+#if defined(_DEBUG) && !defined(NDEBUG)
+#define SIMDJSON_DEVELOPMENT_CHECKS 1
+#endif // _DEBUG
+#else // _MSC_VER
+// All other compilers appear to set __OPTIMIZE__ to a positive integer
+// when the compiler is optimizing.
+// We only set SIMDJSON_DEVELOPMENT_CHECKS if both __OPTIMIZE__
+// and NDEBUG are not defined.
+// We recognize _DEBUG as overriding __OPTIMIZE__ so that if both
+// __OPTIMIZE__ and _DEBUG are defined, we still set SIMDJSON_DEVELOPMENT_CHECKS.
+#if ((!defined(__OPTIMIZE__) || defined(_DEBUG)) && !defined(NDEBUG))
+#define SIMDJSON_DEVELOPMENT_CHECKS 1
+#endif // __OPTIMIZE__
+#endif // _MSC_VER
+#endif // SIMDJSON_DEVELOPMENT_CHECKS
+
+// The SIMDJSON_CHECK_EOF macro is a feature flag for the "don't require padding"
+// feature.
+
+#if SIMDJSON_CPLUSPLUS17
+// if we have C++, then fallthrough is a default attribute
+# define simdjson_fallthrough [[fallthrough]]
+// check if we have __attribute__ support
+#elif defined(__has_attribute)
+// check if we have the __fallthrough__ attribute
+#if __has_attribute(__fallthrough__)
+// we are good to go:
+# define simdjson_fallthrough                    __attribute__((__fallthrough__))
+#endif // __has_attribute(__fallthrough__)
+#endif // SIMDJSON_CPLUSPLUS17
+// on some systems, we simply do not have support for fallthrough, so use a default:
+#ifndef simdjson_fallthrough
+# define simdjson_fallthrough do {} while (0)  /* fallthrough */
+#endif // simdjson_fallthrough
+
+#if SIMDJSON_DEVELOPMENT_CHECKS
+#define SIMDJSON_DEVELOPMENT_ASSERT(expr) do { assert ((expr)); } while (0)
+#else
+#define SIMDJSON_DEVELOPMENT_ASSERT(expr) do { } while (0)
+#endif
+
+#ifndef SIMDJSON_UTF8VALIDATION
+#define SIMDJSON_UTF8VALIDATION 1
+#endif
+
+#ifdef __has_include
+// How do we detect that a compiler supports vbmi2?
+// For sure if the following header is found, we are ok?
+#if __has_include(<avx512vbmi2intrin.h>)
+#define SIMDJSON_COMPILER_SUPPORTS_VBMI2 1
+#endif
+#endif
+
+#ifdef _MSC_VER
+#if _MSC_VER >= 1920
+// Visual Studio 2019 and up support VBMI2 under x64 even if the header
+// avx512vbmi2intrin.h is not found.
+#define SIMDJSON_COMPILER_SUPPORTS_VBMI2 1
+#endif
+#endif
+
+// By default, we allow AVX512.
+#ifndef SIMDJSON_AVX512_ALLOWED
+#define SIMDJSON_AVX512_ALLOWED 1
+#endif
+
+
+#ifndef __has_cpp_attribute
+#define simdjson_lifetime_bound
+#elif __has_cpp_attribute(msvc::lifetimebound)
+#define simdjson_lifetime_bound [[msvc::lifetimebound]]
+#elif __has_cpp_attribute(clang::lifetimebound)
+#define simdjson_lifetime_bound [[clang::lifetimebound]]
+#elif __has_cpp_attribute(lifetimebound)
+#define simdjson_lifetime_bound [[lifetimebound]]
+#else
+#define simdjson_lifetime_bound
+#endif
+#endif // SIMDJSON_COMMON_DEFS_H
+/* end file simdjson/common_defs.h */
+/* skipped duplicate #include "simdjson/compiler_check.h" */
+/* including simdjson/error.h: #include "simdjson/error.h" */
+/* begin file simdjson/error.h */
+#ifndef SIMDJSON_ERROR_H
+#define SIMDJSON_ERROR_H
+
+/* skipped duplicate #include "simdjson/base.h" */
+
+#include <string>
+#include <ostream>
+
+namespace simdjson {
+
+/**
+ * All possible errors returned by simdjson. These error codes are subject to change
+ * and not all simdjson kernel returns the same error code given the same input: it is not
+ * well defined which error a given input should produce.
+ *
+ * Only SUCCESS evaluates to false as a Boolean. All other error codes will evaluate
+ * to true as a Boolean.
+ */
+enum error_code {
+  SUCCESS = 0,                ///< No error
+  CAPACITY,                   ///< This parser can't support a document that big
+  MEMALLOC,                   ///< Error allocating memory, most likely out of memory
+  TAPE_ERROR,                 ///< Something went wrong, this is a generic error. Fatal/unrecoverable error.
+  DEPTH_ERROR,                ///< Your document exceeds the user-specified depth limitation
+  STRING_ERROR,               ///< Problem while parsing a string
+  T_ATOM_ERROR,               ///< Problem while parsing an atom starting with the letter 't'
+  F_ATOM_ERROR,               ///< Problem while parsing an atom starting with the letter 'f'
+  N_ATOM_ERROR,               ///< Problem while parsing an atom starting with the letter 'n'
+  NUMBER_ERROR,               ///< Problem while parsing a number
+  BIGINT_ERROR,               ///< The integer value exceeds 64 bits
+  UTF8_ERROR,                 ///< the input is not valid UTF-8
+  UNINITIALIZED,              ///< unknown error, or uninitialized document
+  EMPTY,                      ///< no structural element found
+  UNESCAPED_CHARS,            ///< found unescaped characters in a string.
+  UNCLOSED_STRING,            ///< missing quote at the end
+  UNSUPPORTED_ARCHITECTURE,   ///< unsupported architecture
+  INCORRECT_TYPE,             ///< JSON element has a different type than user expected
+  NUMBER_OUT_OF_RANGE,        ///< JSON number does not fit in 64 bits
+  INDEX_OUT_OF_BOUNDS,        ///< JSON array index too large
+  NO_SUCH_FIELD,              ///< JSON field not found in object
+  IO_ERROR,                   ///< Error reading a file
+  INVALID_JSON_POINTER,       ///< Invalid JSON pointer syntax
+  INVALID_URI_FRAGMENT,       ///< Invalid URI fragment
+  UNEXPECTED_ERROR,           ///< indicative of a bug in simdjson
+  PARSER_IN_USE,              ///< parser is already in use.
+  OUT_OF_ORDER_ITERATION,     ///< tried to iterate an array or object out of order (checked when SIMDJSON_DEVELOPMENT_CHECKS=1)
+  INSUFFICIENT_PADDING,       ///< The JSON doesn't have enough padding for simdjson to safely parse it.
+  INCOMPLETE_ARRAY_OR_OBJECT, ///< The document ends early. Fatal/unrecoverable error.
+  SCALAR_DOCUMENT_AS_VALUE,   ///< A scalar document is treated as a value.
+  OUT_OF_BOUNDS,              ///< Attempted to access location outside of document.
+  TRAILING_CONTENT,           ///< Unexpected trailing content in the JSON input
+  OUT_OF_CAPACITY,            ///< The capacity was exceeded, we cannot allocate enough memory.
+  NUM_ERROR_CODES             ///< Placeholder for end of error code list.
+};
+
+/**
+ * Some errors are fatal and invalidate the document. This function returns true if the
+ * error is fatal. It returns true for TAPE_ERROR and INCOMPLETE_ARRAY_OR_OBJECT.
+ * Once a fatal error is encountered, the on-demand document is no longer valid and
+ * processing should stop.
+ */
+ inline bool is_fatal(error_code error) noexcept;
+
+/**
+ * It is the convention throughout the code that  the macro SIMDJSON_DEVELOPMENT_CHECKS determines whether
+ * we check for OUT_OF_ORDER_ITERATION. The logic behind it is that these errors only occurs when the code
+ * that was written while breaking some simdjson::ondemand requirement. They should not occur in released
+ * code after these issues were fixed.
+ */
+
+/**
+ * Get the error message for the given error code.
+ *
+ *   dom::parser parser;
+ *   dom::element doc;
+ *   auto error = parser.parse("foo",3).get(doc);
+ *   if (error) { printf("Error: %s\n", error_message(error)); }
+ *
+ * @return The error message.
+ */
+inline const char *error_message(error_code error) noexcept;
+
+/**
+ * Write the error message to the output stream
+ */
+inline std::ostream& operator<<(std::ostream& out, error_code error) noexcept;
+
+/**
+ * Exception thrown when an exception-supporting simdjson method is called
+ */
+struct simdjson_error : public std::exception {
+  /**
+   * Create an exception from a simdjson error code.
+   * @param error The error code
+   */
+  simdjson_error(error_code error) noexcept : _error{error} { }
+  /** The error message */
+  const char *what() const noexcept override { return error_message(error()); }
+  /** The error code */
+  error_code error() const noexcept { return _error; }
+private:
+  /** The error code that was used */
+  error_code _error;
+};
+
+namespace internal {
+
+/**
+ * The result of a simdjson operation that could fail.
+ *
+ * IMPORTANT: For the ondemand API, we use implementation_simdjson_result_base<T> as a base class
+ * to avoid some compilation issue. Thus, if you modify this class, please ensure that the ondemand
+ * implementation_simdjson_result_base<T> is also modified.
+ *
+ * Gives the option of reading error codes, or throwing an exception by casting to the desired result.
+ *
+ * This is a base class for implementations that want to add functions to the result type for
+ * chaining.
+ *
+ * Override like:
+ *
+ *   struct simdjson_result<T> : public internal::simdjson_result_base<T> {
+ *     simdjson_result() noexcept : internal::simdjson_result_base<T>() {}
+ *     simdjson_result(error_code error) noexcept : internal::simdjson_result_base<T>(error) {}
+ *     simdjson_result(T &&value) noexcept : internal::simdjson_result_base<T>(std::forward(value)) {}
+ *     simdjson_result(T &&value, error_code error) noexcept : internal::simdjson_result_base<T>(value, error) {}
+ *     // Your extra methods here
+ *   }
+ *
+ * Then any method returning simdjson_result<T> will be chainable with your methods.
+ */
+template<typename T>
+struct simdjson_result_base : protected std::pair<T, error_code> {
+
+  /**
+   * Create a new empty result with error = UNINITIALIZED.
+   */
+  simdjson_inline simdjson_result_base() noexcept;
+
+  /**
+   * Create a new error result.
+   */
+  simdjson_inline simdjson_result_base(error_code error) noexcept;
+
+  /**
+   * Create a new successful result.
+   */
+  simdjson_inline simdjson_result_base(T &&value) noexcept;
+
+  /**
+   * Create a new result with both things (use if you don't want to branch when creating the result).
+   */
+  simdjson_inline simdjson_result_base(T &&value, error_code error) noexcept;
+
+  /**
+   * Move the value and the error to the provided variables.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   * @param error The variable to assign the error to. Set to SUCCESS if there is no error.
+   */
+  simdjson_inline void tie(T &value, error_code &error) && noexcept;
+
+  /**
+   * Move the value to the provided variable.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   */
+  simdjson_inline error_code get(T &value) && noexcept;
+
+  /**
+   * The error.
+   */
+  simdjson_inline error_code error() const noexcept;
+
+  /**
+   * Whether there is a value.
+   */
+  simdjson_inline bool has_value() const noexcept;
+#if SIMDJSON_EXCEPTIONS
+
+  /**
+   * Dereference operator to access the contained value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T& operator*() &  noexcept(false);
+  simdjson_inline T&& operator*() &&  noexcept(false);
+  /**
+   * Arrow operator to access members of the contained value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T* operator->() noexcept(false);
+  simdjson_inline const T* operator->() const noexcept(false);
+
+  /**
+   * Get the result value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T& value() & noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& value() && noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& take_value() && noexcept(false);
+
+  /**
+   * Cast to the value (will throw on error).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline operator T&&() && noexcept(false);
+
+#endif // SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   * We discourage the use of value_unsafe().
+   *
+   * The recommended pattern is:
+   *
+   * T value; // where T is the type
+   * auto error = result.get(value);
+   * if (error) {
+   *   // handle error
+   * }
+   *
+   * Or you may call 'value()' which will raise an exception
+   * in case of error:
+   *
+   * T value = result.value();
+   */
+  simdjson_inline const T& value_unsafe() const& noexcept;
+
+  /**
+   * Take the result value (move it). This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   * We discourage the use of value_unsafe().
+   *
+   * The recommended pattern is:
+   *
+   * T value; // where T is the type
+   * auto error = result.get(value);
+   * if (error) {
+   *   // handle error, return, exit, abort
+   * } else {
+   *   // use value here.
+   * }
+   *
+   * Or you may call 'value()' which will raise an exception
+   * in case of error:
+   *
+   * T value = result.value();
+   */
+  simdjson_inline T&& value_unsafe() && noexcept;
+
+  using value_type = T;
+  using error_type = error_code;
+}; // struct simdjson_result_base
+
+} // namespace internal
+
+/**
+ * The result of a simdjson operation that could fail.
+ *
+ * Gives the option of reading error codes, or throwing an exception by casting to the desired result.
+ */
+template<typename T>
+struct simdjson_result : public internal::simdjson_result_base<T> {
+
+  /**
+   * @private Create a new empty result with error = UNINITIALIZED.
+   */
+  simdjson_inline simdjson_result() noexcept;
+  /**
+   * @private Create a new successful result.
+   */
+  simdjson_inline simdjson_result(T &&value) noexcept;
+  /**
+   * @private Create a new error result.
+   */
+  simdjson_inline simdjson_result(error_code error_code) noexcept;
+  /**
+   * @private Create a new result with both things (use if you don't want to branch when creating the result).
+   */
+  simdjson_inline simdjson_result(T &&value, error_code error) noexcept;
+
+  /**
+   * Move the value and the error to the provided variables.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   * @param error The variable to assign the error to. Set to SUCCESS if there is no error.
+   */
+  simdjson_inline void tie(T &value, error_code &error) && noexcept;
+
+  /**
+   * Move the value to the provided variable.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   */
+  simdjson_warn_unused simdjson_inline error_code get(T &value) && noexcept;
+
+  /**
+   * Copy the value to a provided std::string, only enabled for std::string_view.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   */
+  template <typename U = T>
+  simdjson_warn_unused simdjson_inline error_code get(std::string &value) && noexcept {
+    static_assert(std::is_same<U, std::string_view>::value, "SFINAE");
+    std::string_view v;
+    error_code error = std::forward<simdjson_result<T>>(*this).get(v);
+    if (!error) {
+      value.assign(v.data(), v.size());
+    }
+    return error;
+  }
+
+  /**
+   * The error.
+   */
+  simdjson_inline error_code error() const noexcept;
+
+
+
+#if SIMDJSON_EXCEPTIONS
+  using internal::simdjson_result_base<T>::operator*;
+  using internal::simdjson_result_base<T>::operator->;
+  /**
+   * Get the result value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T& value() & noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& value() && noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& take_value() && noexcept(false);
+
+  /**
+   * Cast to the value (will throw on error).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline operator T&&() && noexcept(false);
+#endif // SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline const T& value_unsafe() const& noexcept;
+
+  /**
+   * Take the result value (move it). This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T&& value_unsafe() && noexcept;
+
+  using value_type = T;
+  using error_type = error_code;
+}; // struct simdjson_result
+
+#if SIMDJSON_EXCEPTIONS
+
+template<typename T>
+inline std::ostream& operator<<(std::ostream& out, simdjson_result<T> value) { return out << value.value(); }
+#endif // SIMDJSON_EXCEPTIONS
+
+#ifndef SIMDJSON_DISABLE_DEPRECATED_API
+/**
+ * @deprecated This is an alias and will be removed, use error_code instead
+ */
+using ErrorValues [[deprecated("This is an alias and will be removed, use error_code instead")]] = error_code;
+
+/**
+ * @deprecated Error codes should be stored and returned as `error_code`, use `error_message()` instead.
+ */
+[[deprecated("Error codes should be stored and returned as `error_code`, use `error_message()` instead.")]]
+inline const std::string error_message(int error) noexcept;
+#endif // SIMDJSON_DISABLE_DEPRECATED_API
+} // namespace simdjson
+
+#endif // SIMDJSON_ERROR_H
+/* end file simdjson/error.h */
+/* skipped duplicate #include "simdjson/portability.h" */
+/* including simdjson/concepts.h: #include "simdjson/concepts.h" */
+/* begin file simdjson/concepts.h */
+#ifndef SIMDJSON_CONCEPTS_H
+#define SIMDJSON_CONCEPTS_H
+#if SIMDJSON_SUPPORTS_CONCEPTS
+
+#include <concepts>
+#include <type_traits>
+
+namespace simdjson {
+namespace concepts {
+
+namespace details {
+#define SIMDJSON_IMPL_CONCEPT(name, method)                                    \
+  template <typename T>                                                        \
+  concept supports_##name = !std::is_const_v<T> && requires {                  \
+    typename std::remove_cvref_t<T>::value_type;                               \
+    requires requires(typename std::remove_cvref_t<T>::value_type &&val,       \
+                      T obj) {                                                 \
+      obj.method(std::move(val));                                              \
+      requires !requires { obj = std::move(val); };                            \
+    };                                                                         \
+  };
+
+SIMDJSON_IMPL_CONCEPT(emplace_back, emplace_back)
+SIMDJSON_IMPL_CONCEPT(emplace, emplace)
+SIMDJSON_IMPL_CONCEPT(push_back, push_back)
+SIMDJSON_IMPL_CONCEPT(add, add)
+SIMDJSON_IMPL_CONCEPT(push, push)
+SIMDJSON_IMPL_CONCEPT(append, append)
+SIMDJSON_IMPL_CONCEPT(insert, insert)
+SIMDJSON_IMPL_CONCEPT(op_append, operator+=)
+
+#undef SIMDJSON_IMPL_CONCEPT
+} // namespace details
+
+template <typename T>
+concept is_pair = requires { typename T::first_type; typename T::second_type; } &&
+                  std::same_as<T, std::pair<typename T::first_type, typename T::second_type>>;
+template <typename T>
+concept string_view_like = std::is_convertible_v<T, std::string_view> &&
+                           !std::is_convertible_v<T, const char*>;
+
+template<typename T>
+concept constructible_from_string_view = std::is_constructible_v<T, std::string_view>
+                                        && !std::is_same_v<T, std::string_view>
+                                        && std::is_default_constructible_v<T>;
+
+template<typename M>
+concept string_view_keyed_map = string_view_like<typename M::key_type>
+              && requires(std::remove_cvref_t<M>& m, typename M::key_type sv, typename M::mapped_type v) {
+    { m.emplace(sv, v) } -> std::same_as<std::pair<typename M::iterator, bool>>;
+};
+
+/// Check if T is a container that we can append to, including:
+///   std::vector, std::deque, std::list, std::string, ...
+template <typename T>
+concept appendable_containers =
+    (details::supports_emplace_back<T> || details::supports_emplace<T> ||
+    details::supports_push_back<T> || details::supports_push<T> ||
+    details::supports_add<T> || details::supports_append<T> ||
+    details::supports_insert<T>) && !string_view_keyed_map<T>;
+
+/// Insert into the container however possible
+template <appendable_containers T, typename... Args>
+constexpr decltype(auto) emplace_one(T &vec, Args &&...args) {
+  if constexpr (details::supports_emplace_back<T>) {
+    return vec.emplace_back(std::forward<Args>(args)...);
+  } else if constexpr (details::supports_emplace<T>) {
+    return vec.emplace(std::forward<Args>(args)...);
+  } else if constexpr (details::supports_push_back<T>) {
+    return vec.push_back(std::forward<Args>(args)...);
+  } else if constexpr (details::supports_push<T>) {
+    return vec.push(std::forward<Args>(args)...);
+  } else if constexpr (details::supports_add<T>) {
+    return vec.add(std::forward<Args>(args)...);
+  } else if constexpr (details::supports_append<T>) {
+    return vec.append(std::forward<Args>(args)...);
+  } else if constexpr (details::supports_insert<T>) {
+    return vec.insert(std::forward<Args>(args)...);
+  } else if constexpr (details::supports_op_append<T> && sizeof...(Args) == 1) {
+    return vec.operator+=(std::forward<Args>(args)...);
+  } else {
+    static_assert(!sizeof(T *),
+                  "We don't know how to add things to this container");
+  }
+}
+
+/// This checks if the container will return a reference to the newly added
+/// element after an insert which for example `std::vector::emplace_back` does
+/// since C++17; this will allow some optimizations.
+template <typename T>
+concept returns_reference = appendable_containers<T> && requires {
+  typename std::remove_cvref_t<T>::reference;
+  requires requires(typename std::remove_cvref_t<T>::value_type &&val, T obj) {
+    {
+      emplace_one(obj, std::move(val))
+    } -> std::same_as<typename std::remove_cvref_t<T>::reference>;
+  };
+};
+
+template <typename T>
+concept smart_pointer = requires(std::remove_cvref_t<T> ptr) {
+  // Check if T has a member type named element_type
+  typename std::remove_cvref_t<T>::element_type;
+
+  // Check if T has a get() member function
+  {
+    ptr.get()
+  } -> std::same_as<typename std::remove_cvref_t<T>::element_type *>;
+
+  // Check if T can be dereferenced
+  { *ptr } -> std::same_as<typename std::remove_cvref_t<T>::element_type &>;
+};
+
+template <typename T>
+concept optional_type = requires(std::remove_cvref_t<T> obj) {
+  typename std::remove_cvref_t<T>::value_type;
+  { obj.value() } -> std::same_as<typename std::remove_cvref_t<T>::value_type&>;
+  requires requires(typename std::remove_cvref_t<T>::value_type &&val) {
+    obj.emplace(std::move(val));
+    {
+      obj.value_or(val)
+    } -> std::convertible_to<typename std::remove_cvref_t<T>::value_type>;
+  };
+  { static_cast<bool>(obj) } -> std::same_as<bool>; // convertible to bool
+  { obj.reset() } noexcept -> std::same_as<void>;
+};
+
+
+
+// Types we serialize as JSON strings (not as containers)
+template <typename T>
+concept string_like =
+  std::is_same_v<std::remove_cvref_t<T>, std::string> ||
+  std::is_same_v<std::remove_cvref_t<T>, std::string_view> ||
+  std::is_same_v<std::remove_cvref_t<T>, const char*> ||
+  std::is_same_v<std::remove_cvref_t<T>, char*>;
+
+// Concept that checks if a type is a container but not a string (because
+// strings handling must be handled differently)
+// Now uses iterator-based approach for broader container support
+template <typename T>
+concept container_but_not_string =
+  std::ranges::input_range<T> && !string_like<T> && !concepts::string_view_keyed_map<T>;
+
+
+
+// Concept: Indexable container that is not a string or associative container
+// Accepts: std::vector, std::array, std::deque (have operator[], value_type, not string_like)
+// Rejects: std::string (string_like), std::list (no operator[]), std::map (has key_type)
+template<typename Container>
+concept indexable_container = requires {
+  typename Container::value_type;
+  requires !concepts::string_like<Container>;
+  requires !requires { typename Container::key_type; };  // Reject maps/sets
+  requires requires(Container& c, std::size_t i) {
+    { c[i] } -> std::convertible_to<typename Container::value_type>;
+  };
+};
+
+
+// Variable template to use with std::meta::substitute
+template<typename Container>
+constexpr bool indexable_container_v = indexable_container<Container>;
+
+} // namespace concepts
+
+
+/**
+ * We use tag_invoke as our customization point mechanism.
+ */
+template <typename Tag, typename... Args>
+concept tag_invocable = requires(Tag tag, Args... args) {
+  tag_invoke(std::forward<Tag>(tag), std::forward<Args>(args)...);
+};
+
+template <typename Tag, typename... Args>
+concept nothrow_tag_invocable =
+    tag_invocable<Tag, Args...> && requires(Tag tag, Args... args) {
+      {
+        tag_invoke(std::forward<Tag>(tag), std::forward<Args>(args)...)
+      } noexcept;
+    };
+
+} // namespace simdjson
+#endif // SIMDJSON_SUPPORTS_CONCEPTS
+#endif // SIMDJSON_CONCEPTS_H
+/* end file simdjson/concepts.h */
+/* including simdjson/constevalutil.h: #include "simdjson/constevalutil.h" */
+/* begin file simdjson/constevalutil.h */
+#ifndef SIMDJSON_CONSTEVALUTIL_H
+#define SIMDJSON_CONSTEVALUTIL_H
+
+#include <string>
+#include <string_view>
+#include <array>
+
+namespace simdjson {
+namespace constevalutil {
+#if SIMDJSON_CONSTEVAL
+
+constexpr static std::array<uint8_t, 256> json_quotable_character = {
+  1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+  1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+  0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
+constexpr static std::array<std::string_view, 32> control_chars = {
+    "\\u0000", "\\u0001", "\\u0002", "\\u0003", "\\u0004", "\\u0005", "\\u0006",
+    "\\u0007", "\\b", "\\t",     "\\n",     "\\u000b", "\\f",     "\\r",
+    "\\u000e", "\\u000f", "\\u0010", "\\u0011", "\\u0012", "\\u0013", "\\u0014",
+    "\\u0015", "\\u0016", "\\u0017", "\\u0018", "\\u0019", "\\u001a", "\\u001b",
+    "\\u001c", "\\u001d", "\\u001e", "\\u001f"};
+// unoptimized, meant for compile-time execution
+consteval std::string consteval_to_quoted_escaped(std::string_view input) {
+  std::string out = "\"";
+  for (char c : input) {
+    if (json_quotable_character[uint8_t(c)]) {
+      if (c == '"') {
+        out.append("\\\"");
+      } else if (c == '\\') {
+        out.append("\\\\");
+      } else {
+        std::string_view v = control_chars[uint8_t(c)];
+        out.append(v);
+      }
+    } else {
+      out.push_back(c);
+    }
+  }
+  out.push_back('"');
+  return out;
+}
+#endif  // SIMDJSON_CONSTEVAL
+
+
+#if SIMDJSON_SUPPORTS_CONCEPTS
+template <size_t N>
+struct fixed_string {
+    constexpr fixed_string(const char (&str)[N])  {
+        for (std::size_t i = 0; i < N; ++i) {
+            data[i] = str[i];
+        }
+    }
+    char data[N];
+    constexpr std::string_view view() const { return {data, N - 1}; }
+    constexpr size_t size() const { return N ; }
+    constexpr operator std::string_view() const { return view(); }
+    constexpr char operator[](std::size_t index) const { return data[index]; }
+    constexpr bool operator==(const fixed_string& other) const {
+        if (N != other.size()) {
+            return false;
+        }
+        for (std::size_t i = 0; i < N; ++i) {
+            if (data[i] != other.data[i]) {
+                return false;
+            }
+        }
+        return true;
+    }
+};
+template <std::size_t N>
+fixed_string(const char (&)[N]) -> fixed_string<N>;
+
+template <fixed_string str>
+struct string_constant {
+    static constexpr std::string_view value = str.view();
+};
+#endif // SIMDJSON_SUPPORTS_CONCEPTS
+
+} // namespace constevalutil
+} // namespace simdjson
+#endif // SIMDJSON_CONSTEVALUTIL_H
+/* end file simdjson/constevalutil.h */
+
+/**
+ * @brief The top level simdjson namespace, containing everything the library provides.
+ */
+namespace simdjson {
+
+SIMDJSON_PUSH_DISABLE_UNUSED_WARNINGS
+
+/** The maximum document size supported by simdjson. */
+constexpr size_t SIMDJSON_MAXSIZE_BYTES = 0xFFFFFFFF;
+
+/**
+ * The amount of padding needed in a buffer to parse JSON.
+ *
+ * The input buf should be readable up to buf + SIMDJSON_PADDING
+ * this is a stopgap; there should be a better description of the
+ * main loop and its behavior that abstracts over this
+ * See https://github.com/simdjson/simdjson/issues/174
+ */
+constexpr size_t SIMDJSON_PADDING = 64;
+
+/**
+ * By default, simdjson supports this many nested objects and arrays.
+ *
+ * This is the default for parser::max_depth().
+ */
+constexpr size_t DEFAULT_MAX_DEPTH = 1024;
+
+SIMDJSON_POP_DISABLE_UNUSED_WARNINGS
+
+class implementation;
+struct padded_string;
+class padded_string_view;
+enum class stage1_mode;
+
+namespace internal {
+
+template<typename T>
+class atomic_ptr;
+class dom_parser_implementation;
+class escape_json_string;
+class tape_ref;
+struct value128;
+enum class tape_type;
+
+} // namespace internal
+} // namespace simdjson
+
+#endif // SIMDJSON_BASE_H
+/* end file simdjson/base.h */
+
+#endif // SIMDJSON_SRC_BASE_H
+/* end file base.h */
+
+SIMDJSON_PUSH_DISABLE_UNUSED_WARNINGS
+
+/* including to_chars.cpp: #include <to_chars.cpp> */
+/* begin file to_chars.cpp */
+#ifndef SIMDJSON_SRC_TO_CHARS_CPP
+#define SIMDJSON_SRC_TO_CHARS_CPP
+
+/* skipped duplicate #include <base.h> */
+
+#include <cstring>
+#include <cstdint>
+#include <array>
+#include <cmath>
+
+namespace simdjson {
+namespace internal {
+/*!
+implements the Grisu2 algorithm for binary to decimal floating-point
+conversion.
+Adapted from JSON for Modern C++
+
+This implementation is a slightly modified version of the reference
+implementation which may be obtained from
+http://florian.loitsch.com/publications (bench.tar.gz).
+The code is distributed under the MIT license, Copyright (c) 2009 Florian
+Loitsch. For a detailed description of the algorithm see: [1] Loitsch, "Printing
+Floating-Point Numbers Quickly and Accurately with Integers", Proceedings of the
+ACM SIGPLAN 2010 Conference on Programming Language Design and Implementation,
+PLDI 2010 [2] Burger, Dybvig, "Printing Floating-Point Numbers Quickly and
+Accurately", Proceedings of the ACM SIGPLAN 1996 Conference on Programming
+Language Design and Implementation, PLDI 1996
+*/
+namespace dtoa_impl {
+
+template <typename Target, typename Source>
+Target reinterpret_bits(const Source source) {
+  static_assert(sizeof(Target) == sizeof(Source), "size mismatch");
+
+  Target target;
+  std::memcpy(&target, &source, sizeof(Source));
+  return target;
+}
+
+struct diyfp // f * 2^e
+{
+  static constexpr int kPrecision = 64; // = q
+
+  std::uint64_t f = 0;
+  int e = 0;
+
+  constexpr diyfp(std::uint64_t f_, int e_) noexcept : f(f_), e(e_) {}
+
+  /*!
+  @brief returns x - y
+  @pre x.e == y.e and x.f >= y.f
+  */
+  static diyfp sub(const diyfp &x, const diyfp &y) noexcept {
+
+    return {x.f - y.f, x.e};
+  }
+
+  /*!
+  @brief returns x * y
+  @note The result is rounded. (Only the upper q bits are returned.)
+  */
+  static diyfp mul(const diyfp &x, const diyfp &y) noexcept {
+    static_assert(kPrecision == 64, "internal error");
+
+    // Computes:
+    //  f = round((x.f * y.f) / 2^q)
+    //  e = x.e + y.e + q
+
+    // Emulate the 64-bit * 64-bit multiplication:
+    //
+    // p = u * v
+    //   = (u_lo + 2^32 u_hi) (v_lo + 2^32 v_hi)
+    //   = (u_lo v_lo         ) + 2^32 ((u_lo v_hi         ) + (u_hi v_lo )) +
+    //   2^64 (u_hi v_hi         ) = (p0                ) + 2^32 ((p1 ) + (p2 ))
+    //   + 2^64 (p3                ) = (p0_lo + 2^32 p0_hi) + 2^32 ((p1_lo +
+    //   2^32 p1_hi) + (p2_lo + 2^32 p2_hi)) + 2^64 (p3                ) =
+    //   (p0_lo             ) + 2^32 (p0_hi + p1_lo + p2_lo ) + 2^64 (p1_hi +
+    //   p2_hi + p3) = (p0_lo             ) + 2^32 (Q ) + 2^64 (H ) = (p0_lo ) +
+    //   2^32 (Q_lo + 2^32 Q_hi                           ) + 2^64 (H )
+    //
+    // (Since Q might be larger than 2^32 - 1)
+    //
+    //   = (p0_lo + 2^32 Q_lo) + 2^64 (Q_hi + H)
+    //
+    // (Q_hi + H does not overflow a 64-bit int)
+    //
+    //   = p_lo + 2^64 p_hi
+
+    const std::uint64_t u_lo = x.f & 0xFFFFFFFFu;
+    const std::uint64_t u_hi = x.f >> 32u;
+    const std::uint64_t v_lo = y.f & 0xFFFFFFFFu;
+    const std::uint64_t v_hi = y.f >> 32u;
+
+    const std::uint64_t p0 = u_lo * v_lo;
+    const std::uint64_t p1 = u_lo * v_hi;
+    const std::uint64_t p2 = u_hi * v_lo;
+    const std::uint64_t p3 = u_hi * v_hi;
+
+    const std::uint64_t p0_hi = p0 >> 32u;
+    const std::uint64_t p1_lo = p1 & 0xFFFFFFFFu;
+    const std::uint64_t p1_hi = p1 >> 32u;
+    const std::uint64_t p2_lo = p2 & 0xFFFFFFFFu;
+    const std::uint64_t p2_hi = p2 >> 32u;
+
+    std::uint64_t Q = p0_hi + p1_lo + p2_lo;
+
+    // The full product might now be computed as
+    //
+    // p_hi = p3 + p2_hi + p1_hi + (Q >> 32)
+    // p_lo = p0_lo + (Q << 32)
+    //
+    // But in this particular case here, the full p_lo is not required.
+    // Effectively we only need to add the highest bit in p_lo to p_hi (and
+    // Q_hi + 1 does not overflow).
+
+    Q += std::uint64_t{1} << (64u - 32u - 1u); // round, ties up
+
+    const std::uint64_t h = p3 + p2_hi + p1_hi + (Q >> 32u);
+
+    return {h, x.e + y.e + 64};
+  }
+
+  /*!
+  @brief normalize x such that the significand is >= 2^(q-1)
+  @pre x.f != 0
+  */
+  static diyfp normalize(diyfp x) noexcept {
+
+    while ((x.f >> 63u) == 0) {
+      x.f <<= 1u;
+      x.e--;
+    }
+
+    return x;
+  }
+
+  /*!
+  @brief normalize x such that the result has the exponent E
+  @pre e >= x.e and the upper e - x.e bits of x.f must be zero.
+  */
+  static diyfp normalize_to(const diyfp &x,
+                            const int target_exponent) noexcept {
+    const int delta = x.e - target_exponent;
+
+    return {x.f << delta, target_exponent};
+  }
+};
+
+struct boundaries {
+  diyfp w;
+  diyfp minus;
+  diyfp plus;
+};
+
+/*!
+Compute the (normalized) diyfp representing the input number 'value' and its
+boundaries.
+@pre value must be finite and positive
+*/
+template <typename FloatType> boundaries compute_boundaries(FloatType value) {
+
+  // Convert the IEEE representation into a diyfp.
+  //
+  // If v is denormal:
+  //      value = 0.F * 2^(1 - bias) = (          F) * 2^(1 - bias - (p-1))
+  // If v is normalized:
+  //      value = 1.F * 2^(E - bias) = (2^(p-1) + F) * 2^(E - bias - (p-1))
+
+  static_assert(std::numeric_limits<FloatType>::is_iec559,
+                "internal error: dtoa_short requires an IEEE-754 "
+                "floating-point implementation");
+
+  constexpr int kPrecision =
+      std::numeric_limits<FloatType>::digits; // = p (includes the hidden bit)
+  constexpr int kBias =
+      std::numeric_limits<FloatType>::max_exponent - 1 + (kPrecision - 1);
+  constexpr int kMinExp = 1 - kBias;
+  constexpr std::uint64_t kHiddenBit = std::uint64_t{1}
+                                       << (kPrecision - 1); // = 2^(p-1)
+
+  using bits_type = typename std::conditional<kPrecision == 24, std::uint32_t,
+                                              std::uint64_t>::type;
+
+  const std::uint64_t bits = reinterpret_bits<bits_type>(value);
+  const std::uint64_t E = bits >> (kPrecision - 1);
+  const std::uint64_t F = bits & (kHiddenBit - 1);
+
+  const bool is_denormal = E == 0;
+  const diyfp v = is_denormal
+                      ? diyfp(F, kMinExp)
+                      : diyfp(F + kHiddenBit, static_cast<int>(E) - kBias);
+
+  // Compute the boundaries m- and m+ of the floating-point value
+  // v = f * 2^e.
+  //
+  // Determine v- and v+, the floating-point predecessor and successor if v,
+  // respectively.
+  //
+  //      v- = v - 2^e        if f != 2^(p-1) or e == e_min                (A)
+  //         = v - 2^(e-1)    if f == 2^(p-1) and e > e_min                (B)
+  //
+  //      v+ = v + 2^e
+  //
+  // Let m- = (v- + v) / 2 and m+ = (v + v+) / 2. All real numbers _strictly_
+  // between m- and m+ round to v, regardless of how the input rounding
+  // algorithm breaks ties.
+  //
+  //      ---+-------------+-------------+-------------+-------------+---  (A)
+  //         v-            m-            v             m+            v+
+  //
+  //      -----------------+------+------+-------------+-------------+---  (B)
+  //                       v-     m-     v             m+            v+
+
+  const bool lower_boundary_is_closer = F == 0 && E > 1;
+  const diyfp m_plus = diyfp(2 * v.f + 1, v.e - 1);
+  const diyfp m_minus = lower_boundary_is_closer
+                            ? diyfp(4 * v.f - 1, v.e - 2)  // (B)
+                            : diyfp(2 * v.f - 1, v.e - 1); // (A)
+
+  // Determine the normalized w+ = m+.
+  const diyfp w_plus = diyfp::normalize(m_plus);
+
+  // Determine w- = m- such that e_(w-) = e_(w+).
+  const diyfp w_minus = diyfp::normalize_to(m_minus, w_plus.e);
+
+  return {diyfp::normalize(v), w_minus, w_plus};
+}
+
+// Given normalized diyfp w, Grisu needs to find a (normalized) cached
+// power-of-ten c, such that the exponent of the product c * w = f * 2^e lies
+// within a certain range [alpha, gamma] (Definition 3.2 from [1])
+//
+//      alpha <= e = e_c + e_w + q <= gamma
+//
+// or
+//
+//      f_c * f_w * 2^alpha <= f_c 2^(e_c) * f_w 2^(e_w) * 2^q
+//                          <= f_c * f_w * 2^gamma
+//
+// Since c and w are normalized, i.e. 2^(q-1) <= f < 2^q, this implies
+//
+//      2^(q-1) * 2^(q-1) * 2^alpha <= c * w * 2^q < 2^q * 2^q * 2^gamma
+//
+// or
+//
+//      2^(q - 2 + alpha) <= c * w < 2^(q + gamma)
+//
+// The choice of (alpha,gamma) determines the size of the table and the form of
+// the digit generation procedure. Using (alpha,gamma)=(-60,-32) works out well
+// in practice:
+//
+// The idea is to cut the number c * w = f * 2^e into two parts, which can be
+// processed independently: An integral part p1, and a fractional part p2:
+//
+//      f * 2^e = ( (f div 2^-e) * 2^-e + (f mod 2^-e) ) * 2^e
+//              = (f div 2^-e) + (f mod 2^-e) * 2^e
+//              = p1 + p2 * 2^e
+//
+// The conversion of p1 into decimal form requires a series of divisions and
+// modulos by (a power of) 10. These operations are faster for 32-bit than for
+// 64-bit integers, so p1 should ideally fit into a 32-bit integer. This can be
+// achieved by choosing
+//
+//      -e >= 32   or   e <= -32 := gamma
+//
+// In order to convert the fractional part
+//
+//      p2 * 2^e = p2 / 2^-e = d[-1] / 10^1 + d[-2] / 10^2 + ...
+//
+// into decimal form, the fraction is repeatedly multiplied by 10 and the digits
+// d[-i] are extracted in order:
+//
+//      (10 * p2) div 2^-e = d[-1]
+//      (10 * p2) mod 2^-e = d[-2] / 10^1 + ...
+//
+// The multiplication by 10 must not overflow. It is sufficient to choose
+//
+//      10 * p2 < 16 * p2 = 2^4 * p2 <= 2^64.
+//
+// Since p2 = f mod 2^-e < 2^-e,
+//
+//      -e <= 60   or   e >= -60 := alpha
+
+constexpr int kAlpha = -60;
+constexpr int kGamma = -32;
+
+struct cached_power // c = f * 2^e ~= 10^k
+{
+  std::uint64_t f;
+  int e;
+  int k;
+};
+
+/*!
+For a normalized diyfp w = f * 2^e, this function returns a (normalized) cached
+power-of-ten c = f_c * 2^e_c, such that the exponent of the product w * c
+satisfies (Definition 3.2 from [1])
+     alpha <= e_c + e + q <= gamma.
+*/
+inline cached_power get_cached_power_for_binary_exponent(int e) {
+  // Now
+  //
+  //      alpha <= e_c + e + q <= gamma                                    (1)
+  //      ==> f_c * 2^alpha <= c * 2^e * 2^q
+  //
+  // and since the c's are normalized, 2^(q-1) <= f_c,
+  //
+  //      ==> 2^(q - 1 + alpha) <= c * 2^(e + q)
+  //      ==> 2^(alpha - e - 1) <= c
+  //
+  // If c were an exact power of ten, i.e. c = 10^k, one may determine k as
+  //
+  //      k = ceil( log_10( 2^(alpha - e - 1) ) )
+  //        = ceil( (alpha - e - 1) * log_10(2) )
+  //
+  // From the paper:
+  // "In theory the result of the procedure could be wrong since c is rounded,
+  //  and the computation itself is approximated [...]. In practice, however,
+  //  this simple function is sufficient."
+  //
+  // For IEEE double precision floating-point numbers converted into
+  // normalized diyfp's w = f * 2^e, with q = 64,
+  //
+  //      e >= -1022      (min IEEE exponent)
+  //           -52        (p - 1)
+  //           -52        (p - 1, possibly normalize denormal IEEE numbers)
+  //           -11        (normalize the diyfp)
+  //         = -1137
+  //
+  // and
+  //
+  //      e <= +1023      (max IEEE exponent)
+  //           -52        (p - 1)
+  //           -11        (normalize the diyfp)
+  //         = 960
+  //
+  // This binary exponent range [-1137,960] results in a decimal exponent
+  // range [-307,324]. One does not need to store a cached power for each
+  // k in this range. For each such k it suffices to find a cached power
+  // such that the exponent of the product lies in [alpha,gamma].
+  // This implies that the difference of the decimal exponents of adjacent
+  // table entries must be less than or equal to
+  //
+  //      floor( (gamma - alpha) * log_10(2) ) = 8.
+  //
+  // (A smaller distance gamma-alpha would require a larger table.)
+
+  // NB:
+  // Actually this function returns c, such that -60 <= e_c + e + 64 <= -34.
+
+  constexpr int kCachedPowersMinDecExp = -300;
+  constexpr int kCachedPowersDecStep = 8;
+
+  static constexpr std::array<cached_power, 79> kCachedPowers = {{
+      {0xAB70FE17C79AC6CA, -1060, -300}, {0xFF77B1FCBEBCDC4F, -1034, -292},
+      {0xBE5691EF416BD60C, -1007, -284}, {0x8DD01FAD907FFC3C, -980, -276},
+      {0xD3515C2831559A83, -954, -268},  {0x9D71AC8FADA6C9B5, -927, -260},
+      {0xEA9C227723EE8BCB, -901, -252},  {0xAECC49914078536D, -874, -244},
+      {0x823C12795DB6CE57, -847, -236},  {0xC21094364DFB5637, -821, -228},
+      {0x9096EA6F3848984F, -794, -220},  {0xD77485CB25823AC7, -768, -212},
+      {0xA086CFCD97BF97F4, -741, -204},  {0xEF340A98172AACE5, -715, -196},
+      {0xB23867FB2A35B28E, -688, -188},  {0x84C8D4DFD2C63F3B, -661, -180},
+      {0xC5DD44271AD3CDBA, -635, -172},  {0x936B9FCEBB25C996, -608, -164},
+      {0xDBAC6C247D62A584, -582, -156},  {0xA3AB66580D5FDAF6, -555, -148},
+      {0xF3E2F893DEC3F126, -529, -140},  {0xB5B5ADA8AAFF80B8, -502, -132},
+      {0x87625F056C7C4A8B, -475, -124},  {0xC9BCFF6034C13053, -449, -116},
+      {0x964E858C91BA2655, -422, -108},  {0xDFF9772470297EBD, -396, -100},
+      {0xA6DFBD9FB8E5B88F, -369, -92},   {0xF8A95FCF88747D94, -343, -84},
+      {0xB94470938FA89BCF, -316, -76},   {0x8A08F0F8BF0F156B, -289, -68},
+      {0xCDB02555653131B6, -263, -60},   {0x993FE2C6D07B7FAC, -236, -52},
+      {0xE45C10C42A2B3B06, -210, -44},   {0xAA242499697392D3, -183, -36},
+      {0xFD87B5F28300CA0E, -157, -28},   {0xBCE5086492111AEB, -130, -20},
+      {0x8CBCCC096F5088CC, -103, -12},   {0xD1B71758E219652C, -77, -4},
+      {0x9C40000000000000, -50, 4},      {0xE8D4A51000000000, -24, 12},
+      {0xAD78EBC5AC620000, 3, 20},       {0x813F3978F8940984, 30, 28},
+      {0xC097CE7BC90715B3, 56, 36},      {0x8F7E32CE7BEA5C70, 83, 44},
+      {0xD5D238A4ABE98068, 109, 52},     {0x9F4F2726179A2245, 136, 60},
+      {0xED63A231D4C4FB27, 162, 68},     {0xB0DE65388CC8ADA8, 189, 76},
+      {0x83C7088E1AAB65DB, 216, 84},     {0xC45D1DF942711D9A, 242, 92},
+      {0x924D692CA61BE758, 269, 100},    {0xDA01EE641A708DEA, 295, 108},
+      {0xA26DA3999AEF774A, 322, 116},    {0xF209787BB47D6B85, 348, 124},
+      {0xB454E4A179DD1877, 375, 132},    {0x865B86925B9BC5C2, 402, 140},
+      {0xC83553C5C8965D3D, 428, 148},    {0x952AB45CFA97A0B3, 455, 156},
+      {0xDE469FBD99A05FE3, 481, 164},    {0xA59BC234DB398C25, 508, 172},
+      {0xF6C69A72A3989F5C, 534, 180},    {0xB7DCBF5354E9BECE, 561, 188},
+      {0x88FCF317F22241E2, 588, 196},    {0xCC20CE9BD35C78A5, 614, 204},
+      {0x98165AF37B2153DF, 641, 212},    {0xE2A0B5DC971F303A, 667, 220},
+      {0xA8D9D1535CE3B396, 694, 228},    {0xFB9B7CD9A4A7443C, 720, 236},
+      {0xBB764C4CA7A44410, 747, 244},    {0x8BAB8EEFB6409C1A, 774, 252},
+      {0xD01FEF10A657842C, 800, 260},    {0x9B10A4E5E9913129, 827, 268},
+      {0xE7109BFBA19C0C9D, 853, 276},    {0xAC2820D9623BF429, 880, 284},
+      {0x80444B5E7AA7CF85, 907, 292},    {0xBF21E44003ACDD2D, 933, 300},
+      {0x8E679C2F5E44FF8F, 960, 308},    {0xD433179D9C8CB841, 986, 316},
+      {0x9E19DB92B4E31BA9, 1013, 324},
+  }};
+
+  // This computation gives exactly the same results for k as
+  //      k = ceil((kAlpha - e - 1) * 0.30102999566398114)
+  // for |e| <= 1500, but doesn't require floating-point operations.
+  // NB: log_10(2) ~= 78913 / 2^18
+  const int f = kAlpha - e - 1;
+  const int k = (f * 78913) / (1 << 18) + static_cast<int>(f > 0);
+
+  const int index = (-kCachedPowersMinDecExp + k + (kCachedPowersDecStep - 1)) /
+                    kCachedPowersDecStep;
+
+  const cached_power cached = kCachedPowers[static_cast<std::size_t>(index)];
+
+  return cached;
+}
+
+/*!
+For n != 0, returns k, such that pow10 := 10^(k-1) <= n < 10^k.
+For n == 0, returns 1 and sets pow10 := 1.
+*/
+inline int find_largest_pow10(const std::uint32_t n, std::uint32_t &pow10) {
+  // LCOV_EXCL_START
+  if (n >= 1000000000) {
+    pow10 = 1000000000;
+    return 10;
+  }
+  // LCOV_EXCL_STOP
+  else if (n >= 100000000) {
+    pow10 = 100000000;
+    return 9;
+  } else if (n >= 10000000) {
+    pow10 = 10000000;
+    return 8;
+  } else if (n >= 1000000) {
+    pow10 = 1000000;
+    return 7;
+  } else if (n >= 100000) {
+    pow10 = 100000;
+    return 6;
+  } else if (n >= 10000) {
+    pow10 = 10000;
+    return 5;
+  } else if (n >= 1000) {
+    pow10 = 1000;
+    return 4;
+  } else if (n >= 100) {
+    pow10 = 100;
+    return 3;
+  } else if (n >= 10) {
+    pow10 = 10;
+    return 2;
+  } else {
+    pow10 = 1;
+    return 1;
+  }
+}
+
+inline void grisu2_round(char *buf, int len, std::uint64_t dist,
+                         std::uint64_t delta, std::uint64_t rest,
+                         std::uint64_t ten_k) {
+
+  //               <--------------------------- delta ---->
+  //                                  <---- dist --------->
+  // --------------[------------------+-------------------]--------------
+  //               M-                 w                   M+
+  //
+  //                                  ten_k
+  //                                <------>
+  //                                       <---- rest ---->
+  // --------------[------------------+----+--------------]--------------
+  //                                  w    V
+  //                                       = buf * 10^k
+  //
+  // ten_k represents a unit-in-the-last-place in the decimal representation
+  // stored in buf.
+  // Decrement buf by ten_k while this takes buf closer to w.
+
+  // The tests are written in this order to avoid overflow in unsigned
+  // integer arithmetic.
+
+  while (rest < dist && delta - rest >= ten_k &&
+         (rest + ten_k < dist || dist - rest > rest + ten_k - dist)) {
+    buf[len - 1]--;
+    rest += ten_k;
+  }
+}
+
+/*!
+Generates V = buffer * 10^decimal_exponent, such that M- <= V <= M+.
+M- and M+ must be normalized and share the same exponent -60 <= e <= -32.
+*/
+inline void grisu2_digit_gen(char *buffer, int &length, int &decimal_exponent,
+                             diyfp M_minus, diyfp w, diyfp M_plus) {
+  static_assert(kAlpha >= -60, "internal error");
+  static_assert(kGamma <= -32, "internal error");
+
+  // Generates the digits (and the exponent) of a decimal floating-point
+  // number V = buffer * 10^decimal_exponent in the range [M-, M+]. The diyfp's
+  // w, M- and M+ share the same exponent e, which satisfies alpha <= e <=
+  // gamma.
+  //
+  //               <--------------------------- delta ---->
+  //                                  <---- dist --------->
+  // --------------[------------------+-------------------]--------------
+  //               M-                 w                   M+
+  //
+  // Grisu2 generates the digits of M+ from left to right and stops as soon as
+  // V is in [M-,M+].
+
+  std::uint64_t delta =
+      diyfp::sub(M_plus, M_minus)
+          .f; // (significand of (M+ - M-), implicit exponent is e)
+  std::uint64_t dist =
+      diyfp::sub(M_plus, w)
+          .f; // (significand of (M+ - w ), implicit exponent is e)
+
+  // Split M+ = f * 2^e into two parts p1 and p2 (note: e < 0):
+  //
+  //      M+ = f * 2^e
+  //         = ((f div 2^-e) * 2^-e + (f mod 2^-e)) * 2^e
+  //         = ((p1        ) * 2^-e + (p2        )) * 2^e
+  //         = p1 + p2 * 2^e
+
+  const diyfp one(std::uint64_t{1} << -M_plus.e, M_plus.e);
+
+  auto p1 = static_cast<std::uint32_t>(
+      M_plus.f >>
+      -one.e); // p1 = f div 2^-e (Since -e >= 32, p1 fits into a 32-bit int.)
+  std::uint64_t p2 = M_plus.f & (one.f - 1); // p2 = f mod 2^-e
+
+  // 1)
+  //
+  // Generate the digits of the integral part p1 = d[n-1]...d[1]d[0]
+
+  std::uint32_t pow10;
+  const int k = find_largest_pow10(p1, pow10);
+
+  //      10^(k-1) <= p1 < 10^k, pow10 = 10^(k-1)
+  //
+  //      p1 = (p1 div 10^(k-1)) * 10^(k-1) + (p1 mod 10^(k-1))
+  //         = (d[k-1]         ) * 10^(k-1) + (p1 mod 10^(k-1))
+  //
+  //      M+ = p1                                             + p2 * 2^e
+  //         = d[k-1] * 10^(k-1) + (p1 mod 10^(k-1))          + p2 * 2^e
+  //         = d[k-1] * 10^(k-1) + ((p1 mod 10^(k-1)) * 2^-e + p2) * 2^e
+  //         = d[k-1] * 10^(k-1) + (                         rest) * 2^e
+  //
+  // Now generate the digits d[n] of p1 from left to right (n = k-1,...,0)
+  //
+  //      p1 = d[k-1]...d[n] * 10^n + d[n-1]...d[0]
+  //
+  // but stop as soon as
+  //
+  //      rest * 2^e = (d[n-1]...d[0] * 2^-e + p2) * 2^e <= delta * 2^e
+
+  int n = k;
+  while (n > 0) {
+    // Invariants:
+    //      M+ = buffer * 10^n + (p1 + p2 * 2^e)    (buffer = 0 for n = k)
+    //      pow10 = 10^(n-1) <= p1 < 10^n
+    //
+    const std::uint32_t d = p1 / pow10; // d = p1 div 10^(n-1)
+    const std::uint32_t r = p1 % pow10; // r = p1 mod 10^(n-1)
+    //
+    //      M+ = buffer * 10^n + (d * 10^(n-1) + r) + p2 * 2^e
+    //         = (buffer * 10 + d) * 10^(n-1) + (r + p2 * 2^e)
+    //
+    buffer[length++] = static_cast<char>('0' + d); // buffer := buffer * 10 + d
+    //
+    //      M+ = buffer * 10^(n-1) + (r + p2 * 2^e)
+    //
+    p1 = r;
+    n--;
+    //
+    //      M+ = buffer * 10^n + (p1 + p2 * 2^e)
+    //      pow10 = 10^n
+    //
+
+    // Now check if enough digits have been generated.
+    // Compute
+    //
+    //      p1 + p2 * 2^e = (p1 * 2^-e + p2) * 2^e = rest * 2^e
+    //
+    // Note:
+    // Since rest and delta share the same exponent e, it suffices to
+    // compare the significands.
+    const std::uint64_t rest = (std::uint64_t{p1} << -one.e) + p2;
+    if (rest <= delta) {
+      // V = buffer * 10^n, with M- <= V <= M+.
+
+      decimal_exponent += n;
+
+      // We may now just stop. But instead look if the buffer could be
+      // decremented to bring V closer to w.
+      //
+      // pow10 = 10^n is now 1 ulp in the decimal representation V.
+      // The rounding procedure works with diyfp's with an implicit
+      // exponent of e.
+      //
+      //      10^n = (10^n * 2^-e) * 2^e = ulp * 2^e
+      //
+      const std::uint64_t ten_n = std::uint64_t{pow10} << -one.e;
+      grisu2_round(buffer, length, dist, delta, rest, ten_n);
+
+      return;
+    }
+
+    pow10 /= 10;
+    //
+    //      pow10 = 10^(n-1) <= p1 < 10^n
+    // Invariants restored.
+  }
+
+  // 2)
+  //
+  // The digits of the integral part have been generated:
+  //
+  //      M+ = d[k-1]...d[1]d[0] + p2 * 2^e
+  //         = buffer            + p2 * 2^e
+  //
+  // Now generate the digits of the fractional part p2 * 2^e.
+  //
+  // Note:
+  // No decimal point is generated: the exponent is adjusted instead.
+  //
+  // p2 actually represents the fraction
+  //
+  //      p2 * 2^e
+  //          = p2 / 2^-e
+  //          = d[-1] / 10^1 + d[-2] / 10^2 + ...
+  //
+  // Now generate the digits d[-m] of p1 from left to right (m = 1,2,...)
+  //
+  //      p2 * 2^e = d[-1]d[-2]...d[-m] * 10^-m
+  //                      + 10^-m * (d[-m-1] / 10^1 + d[-m-2] / 10^2 + ...)
+  //
+  // using
+  //
+  //      10^m * p2 = ((10^m * p2) div 2^-e) * 2^-e + ((10^m * p2) mod 2^-e)
+  //                = (                   d) * 2^-e + (                   r)
+  //
+  // or
+  //      10^m * p2 * 2^e = d + r * 2^e
+  //
+  // i.e.
+  //
+  //      M+ = buffer + p2 * 2^e
+  //         = buffer + 10^-m * (d + r * 2^e)
+  //         = (buffer * 10^m + d) * 10^-m + 10^-m * r * 2^e
+  //
+  // and stop as soon as 10^-m * r * 2^e <= delta * 2^e
+
+  int m = 0;
+  for (;;) {
+    // Invariant:
+    //      M+ = buffer * 10^-m + 10^-m * (d[-m-1] / 10 + d[-m-2] / 10^2 + ...)
+    //      * 2^e
+    //         = buffer * 10^-m + 10^-m * (p2                                 )
+    //         * 2^e = buffer * 10^-m + 10^-m * (1/10 * (10 * p2) ) * 2^e =
+    //         buffer * 10^-m + 10^-m * (1/10 * ((10*p2 div 2^-e) * 2^-e +
+    //         (10*p2 mod 2^-e)) * 2^e
+    //
+    p2 *= 10;
+    const std::uint64_t d = p2 >> -one.e;     // d = (10 * p2) div 2^-e
+    const std::uint64_t r = p2 & (one.f - 1); // r = (10 * p2) mod 2^-e
+    //
+    //      M+ = buffer * 10^-m + 10^-m * (1/10 * (d * 2^-e + r) * 2^e
+    //         = buffer * 10^-m + 10^-m * (1/10 * (d + r * 2^e))
+    //         = (buffer * 10 + d) * 10^(-m-1) + 10^(-m-1) * r * 2^e
+    //
+    buffer[length++] = static_cast<char>('0' + d); // buffer := buffer * 10 + d
+    //
+    //      M+ = buffer * 10^(-m-1) + 10^(-m-1) * r * 2^e
+    //
+    p2 = r;
+    m++;
+    //
+    //      M+ = buffer * 10^-m + 10^-m * p2 * 2^e
+    // Invariant restored.
+
+    // Check if enough digits have been generated.
+    //
+    //      10^-m * p2 * 2^e <= delta * 2^e
+    //              p2 * 2^e <= 10^m * delta * 2^e
+    //                    p2 <= 10^m * delta
+    delta *= 10;
+    dist *= 10;
+    if (p2 <= delta) {
+      break;
+    }
+  }
+
+  // V = buffer * 10^-m, with M- <= V <= M+.
+
+  decimal_exponent -= m;
+
+  // 1 ulp in the decimal representation is now 10^-m.
+  // Since delta and dist are now scaled by 10^m, we need to do the
+  // same with ulp in order to keep the units in sync.
+  //
+  //      10^m * 10^-m = 1 = 2^-e * 2^e = ten_m * 2^e
+  //
+  const std::uint64_t ten_m = one.f;
+  grisu2_round(buffer, length, dist, delta, p2, ten_m);
+
+  // By construction this algorithm generates the shortest possible decimal
+  // number (Loitsch, Theorem 6.2) which rounds back to w.
+  // For an input number of precision p, at least
+  //
+  //      N = 1 + ceil(p * log_10(2))
+  //
+  // decimal digits are sufficient to identify all binary floating-point
+  // numbers (Matula, "In-and-Out conversions").
+  // This implies that the algorithm does not produce more than N decimal
+  // digits.
+  //
+  //      N = 17 for p = 53 (IEEE double precision)
+  //      N = 9  for p = 24 (IEEE single precision)
+}
+
+/*!
+v = buf * 10^decimal_exponent
+len is the length of the buffer (number of decimal digits)
+The buffer must be large enough, i.e. >= max_digits10.
+*/
+inline void grisu2(char *buf, int &len, int &decimal_exponent, diyfp m_minus,
+                   diyfp v, diyfp m_plus) {
+
+  //  --------(-----------------------+-----------------------)--------    (A)
+  //          m-                      v                       m+
+  //
+  //  --------------------(-----------+-----------------------)--------    (B)
+  //                      m-          v                       m+
+  //
+  // First scale v (and m- and m+) such that the exponent is in the range
+  // [alpha, gamma].
+
+  const cached_power cached = get_cached_power_for_binary_exponent(m_plus.e);
+
+  const diyfp c_minus_k(cached.f, cached.e); // = c ~= 10^-k
+
+  // The exponent of the products is = v.e + c_minus_k.e + q and is in the range
+  // [alpha,gamma]
+  const diyfp w = diyfp::mul(v, c_minus_k);
+  const diyfp w_minus = diyfp::mul(m_minus, c_minus_k);
+  const diyfp w_plus = diyfp::mul(m_plus, c_minus_k);
+
+  //  ----(---+---)---------------(---+---)---------------(---+---)----
+  //          w-                      w                       w+
+  //          = c*m-                  = c*v                   = c*m+
+  //
+  // diyfp::mul rounds its result and c_minus_k is approximated too. w, w- and
+  // w+ are now off by a small amount.
+  // In fact:
+  //
+  //      w - v * 10^k < 1 ulp
+  //
+  // To account for this inaccuracy, add resp. subtract 1 ulp.
+  //
+  //  --------+---[---------------(---+---)---------------]---+--------
+  //          w-  M-                  w                   M+  w+
+  //
+  // Now any number in [M-, M+] (bounds included) will round to w when input,
+  // regardless of how the input rounding algorithm breaks ties.
+  //
+  // And digit_gen generates the shortest possible such number in [M-, M+].
+  // Note that this does not mean that Grisu2 always generates the shortest
+  // possible number in the interval (m-, m+).
+  const diyfp M_minus(w_minus.f + 1, w_minus.e);
+  const diyfp M_plus(w_plus.f - 1, w_plus.e);
+
+  decimal_exponent = -cached.k; // = -(-k) = k
+
+  grisu2_digit_gen(buf, len, decimal_exponent, M_minus, w, M_plus);
+}
+
+/*!
+v = buf * 10^decimal_exponent
+len is the length of the buffer (number of decimal digits)
+The buffer must be large enough, i.e. >= max_digits10.
+*/
+template <typename FloatType>
+void grisu2(char *buf, int &len, int &decimal_exponent, FloatType value) {
+  static_assert(diyfp::kPrecision >= std::numeric_limits<FloatType>::digits + 3,
+                "internal error: not enough precision");
+
+  // If the neighbors (and boundaries) of 'value' are always computed for
+  // double-precision numbers, all float's can be recovered using strtod (and
+  // strtof). However, the resulting decimal representations are not exactly
+  // "short".
+  //
+  // The documentation for 'std::to_chars'
+  // (https://en.cppreference.com/w/cpp/utility/to_chars) says "value is
+  // converted to a string as if by std::sprintf in the default ("C") locale"
+  // and since sprintf promotes float's to double's, I think this is exactly
+  // what 'std::to_chars' does. On the other hand, the documentation for
+  // 'std::to_chars' requires that "parsing the representation using the
+  // corresponding std::from_chars function recovers value exactly". That
+  // indicates that single precision floating-point numbers should be recovered
+  // using 'std::strtof'.
+  //
+  // NB: If the neighbors are computed for single-precision numbers, there is a
+  // single float
+  //     (7.0385307e-26f) which can't be recovered using strtod. The resulting
+  //     double precision value is off by 1 ulp.
+#if 0
+    const boundaries w = compute_boundaries(static_cast<double>(value));
+#else
+  const boundaries w = compute_boundaries(value);
+#endif
+
+  grisu2(buf, len, decimal_exponent, w.minus, w.w, w.plus);
+}
+
+/*!
+@brief appends a decimal representation of e to buf
+@return a pointer to the element following the exponent.
+@pre -1000 < e < 1000
+*/
+inline char *append_exponent(char *buf, int e) {
+
+  bool isNegative = e < 0;
+  e = isNegative ? -e : e;
+  *buf++ = isNegative  ? '-' : '+';
+
+  auto k = static_cast<std::uint32_t>(e);
+  if (k < 100) {
+    // Always print at least two digits in the exponent.
+    // This is for compatibility with printf("%g").
+    *buf++ = static_cast<char>('0' + k / 10);
+    k %= 10;
+    *buf++ = static_cast<char>('0' + k);
+  } else {
+    *buf++ = static_cast<char>('0' + k / 100);
+    k %= 100;
+    *buf++ = static_cast<char>('0' + k / 10);
+    k %= 10;
+    *buf++ = static_cast<char>('0' + k);
+  }
+
+  return buf;
+}
+
+/*!
+@brief prettify v = buf * 10^decimal_exponent
+If v is in the range [10^min_exp, 10^max_exp) it will be printed in fixed-point
+notation. Otherwise it will be printed in exponential notation.
+@pre min_exp < 0
+@pre max_exp > 0
+*/
+inline char *format_buffer(char *buf, int len, int decimal_exponent,
+                           int min_exp, int max_exp) {
+
+  const int k = len;
+  const int n = len + decimal_exponent;
+
+  // v = buf * 10^(n-k)
+  // k is the length of the buffer (number of decimal digits)
+  // n is the position of the decimal point relative to the start of the buffer.
+
+  if (k <= n && n <= max_exp) {
+    // digits[000]
+    // len <= max_exp + 2
+
+    std::memset(buf + k, '0', static_cast<size_t>(n) - static_cast<size_t>(k));
+    // Make it look like a floating-point number (#362, #378)
+    buf[n + 0] = '.';
+    buf[n + 1] = '0';
+    return buf + (static_cast<size_t>(n)) + 2;
+  }
+
+  if (0 < n && n <= max_exp) {
+    // dig.its
+    // len <= max_digits10 + 1
+    std::memmove(buf + (static_cast<size_t>(n) + 1), buf + n,
+                 static_cast<size_t>(k) - static_cast<size_t>(n));
+    buf[n] = '.';
+    return buf + (static_cast<size_t>(k) + 1U);
+  }
+
+  if (min_exp < n && n <= 0) {
+    // 0.[000]digits
+    // len <= 2 + (-min_exp - 1) + max_digits10
+
+    std::memmove(buf + (2 + static_cast<size_t>(-n)), buf,
+                 static_cast<size_t>(k));
+    buf[0] = '0';
+    buf[1] = '.';
+    std::memset(buf + 2, '0', static_cast<size_t>(-n));
+    return buf + (2U + static_cast<size_t>(-n) + static_cast<size_t>(k));
+  }
+
+  if (k == 1) {
+    // dE+123
+    // len <= 1 + 5
+
+    buf += 1;
+  } else {
+    // d.igitsE+123
+    // len <= max_digits10 + 1 + 5
+
+    std::memmove(buf + 2, buf + 1, static_cast<size_t>(k) - 1);
+    buf[1] = '.';
+    buf += 1 + static_cast<size_t>(k);
+  }
+
+  *buf++ = 'e';
+  return append_exponent(buf, n - 1);
+}
+
+} // namespace dtoa_impl
+
+/*!
+The format of the resulting decimal representation is similar to printf's %g
+format. Returns an iterator pointing past-the-end of the decimal representation.
+@note The input number must be finite, i.e. NaN's and Inf's are not supported.
+@note The buffer must be large enough.
+@note The result is NOT null-terminated.
+*/
+char *to_chars(char *first, const char *last, double value) {
+  static_cast<void>(last); // maybe unused - fix warning
+  bool negative = std::signbit(value);
+  if (negative) {
+    value = -value;
+    *first++ = '-';
+  }
+
+  if (value == 0) // +-0
+  {
+    *first++ = '0';
+    // Make it look like a floating-point number (#362, #378)
+    *first++ = '.';
+    *first++ = '0';
+    return first;
+  }
+  // Compute v = buffer * 10^decimal_exponent.
+  // The decimal digits are stored in the buffer, which needs to be interpreted
+  // as an unsigned decimal integer.
+  // len is the length of the buffer, i.e. the number of decimal digits.
+  int len = 0;
+  int decimal_exponent = 0;
+  dtoa_impl::grisu2(first, len, decimal_exponent, value);
+  // Format the buffer like printf("%.*g", prec, value)
+  constexpr int kMinExp = -4;
+  constexpr int kMaxExp = std::numeric_limits<double>::digits10;
+
+  return dtoa_impl::format_buffer(first, len, decimal_exponent, kMinExp,
+                                  kMaxExp);
+}
+} // namespace internal
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_TO_CHARS_CPP
+/* end file to_chars.cpp */
+/* including from_chars.cpp: #include <from_chars.cpp> */
+/* begin file from_chars.cpp */
+#ifndef SIMDJSON_SRC_FROM_CHARS_CPP
+#define SIMDJSON_SRC_FROM_CHARS_CPP
+
+/* skipped duplicate #include <base.h> */
+
+#include <cstdint>
+#include <cstring>
+#include <limits>
+
+namespace simdjson {
+namespace internal {
+
+/**
+ * The code in the internal::from_chars function is meant to handle the floating-point number parsing
+ * when we have more than 19 digits in the decimal mantissa. This should only be seen
+ * in adversarial scenarios: we do not expect production systems to even produce
+ * such floating-point numbers.
+ *
+ * The parser is based on work by Nigel Tao (at https://github.com/google/wuffs/)
+ * who credits Ken Thompson for the design (via a reference to the Go source
+ * code). See
+ * https://github.com/google/wuffs/blob/aa46859ea40c72516deffa1b146121952d6dfd3b/internal/cgen/base/floatconv-submodule-data.c
+ * https://github.com/google/wuffs/blob/46cd8105f47ca07ae2ba8e6a7818ef9c0df6c152/internal/cgen/base/floatconv-submodule-code.c
+ * It is probably not very fast but it is a fallback that should almost never be
+ * called in real life. Google Wuffs is published under APL 2.0.
+ **/
+
+namespace {
+constexpr uint32_t max_digits = 768;
+constexpr int32_t decimal_point_range = 2047;
+} // namespace
+
+struct adjusted_mantissa {
+  uint64_t mantissa;
+  int power2;
+  adjusted_mantissa() : mantissa(0), power2(0) {}
+};
+
+struct decimal {
+  uint32_t num_digits;
+  int32_t decimal_point;
+  bool negative;
+  bool truncated;
+  uint8_t digits[max_digits];
+};
+
+template <typename T> struct binary_format {
+  static constexpr int mantissa_explicit_bits();
+  static constexpr int minimum_exponent();
+  static constexpr int infinite_power();
+  static constexpr int sign_index();
+};
+
+template <> constexpr int binary_format<double>::mantissa_explicit_bits() {
+  return 52;
+}
+
+template <> constexpr int binary_format<double>::minimum_exponent() {
+  return -1023;
+}
+template <> constexpr int binary_format<double>::infinite_power() {
+  return 0x7FF;
+}
+
+template <> constexpr int binary_format<double>::sign_index() { return 63; }
+
+bool is_integer(char c)  noexcept  { return (c >= '0' && c <= '9'); }
+
+// This should always succeed since it follows a call to parse_number.
+decimal parse_decimal(const char *&p) noexcept {
+  decimal answer;
+  answer.num_digits = 0;
+  answer.decimal_point = 0;
+  answer.truncated = false;
+  answer.negative = (*p == '-');
+  if ((*p == '-') || (*p == '+')) {
+    ++p;
+  }
+
+  while (*p == '0') {
+    ++p;
+  }
+  while (is_integer(*p)) {
+    if (answer.num_digits < max_digits) {
+      answer.digits[answer.num_digits] = uint8_t(*p - '0');
+    }
+    answer.num_digits++;
+    ++p;
+  }
+  if (*p == '.') {
+    ++p;
+    const char *first_after_period = p;
+    // if we have not yet encountered a zero, we have to skip it as well
+    if (answer.num_digits == 0) {
+      // skip zeros
+      while (*p == '0') {
+        ++p;
+      }
+    }
+    while (is_integer(*p)) {
+      if (answer.num_digits < max_digits) {
+        answer.digits[answer.num_digits] = uint8_t(*p - '0');
+      }
+      answer.num_digits++;
+      ++p;
+    }
+    answer.decimal_point = int32_t(first_after_period - p);
+  }
+  if(answer.num_digits > 0) {
+    const char *preverse = p - 1;
+    int32_t trailing_zeros = 0;
+    while ((*preverse == '0') || (*preverse == '.')) {
+      if(*preverse == '0') { trailing_zeros++; };
+      --preverse;
+    }
+    answer.decimal_point += int32_t(answer.num_digits);
+    answer.num_digits -= uint32_t(trailing_zeros);
+  }
+  if(answer.num_digits > max_digits ) {
+    answer.num_digits = max_digits;
+    answer.truncated = true;
+  }
+  if (('e' == *p) || ('E' == *p)) {
+    ++p;
+    bool neg_exp = false;
+    if ('-' == *p) {
+      neg_exp = true;
+      ++p;
+    } else if ('+' == *p) {
+      ++p;
+    }
+    int32_t exp_number = 0; // exponential part
+    while (is_integer(*p)) {
+      uint8_t digit = uint8_t(*p - '0');
+      if (exp_number < 0x10000) {
+        exp_number = 10 * exp_number + digit;
+      }
+      ++p;
+    }
+    answer.decimal_point += (neg_exp ? -exp_number : exp_number);
+  }
+  return answer;
+}
+
+// This should always succeed since it follows a call to parse_number.
+// Will not read at or beyond the "end" pointer.
+decimal parse_decimal(const char *&p, const char * end) noexcept {
+  decimal answer;
+  answer.num_digits = 0;
+  answer.decimal_point = 0;
+  answer.truncated = false;
+  if(p == end) { return answer; } // should never happen
+  answer.negative = (*p == '-');
+  if ((*p == '-') || (*p == '+')) {
+    ++p;
+  }
+
+  while ((p != end) && (*p == '0')) {
+    ++p;
+  }
+  while ((p != end) && is_integer(*p)) {
+    if (answer.num_digits < max_digits) {
+      answer.digits[answer.num_digits] = uint8_t(*p - '0');
+    }
+    answer.num_digits++;
+    ++p;
+  }
+  if ((p != end) && (*p == '.')) {
+    ++p;
+    if(p == end) { return answer; } // should never happen
+    const char *first_after_period = p;
+    // if we have not yet encountered a zero, we have to skip it as well
+    if (answer.num_digits == 0) {
+      // skip zeros
+      while (*p == '0') {
+        ++p;
+      }
+    }
+    while ((p != end) && is_integer(*p)) {
+      if (answer.num_digits < max_digits) {
+        answer.digits[answer.num_digits] = uint8_t(*p - '0');
+      }
+      answer.num_digits++;
+      ++p;
+    }
+    answer.decimal_point = int32_t(first_after_period - p);
+  }
+  if(answer.num_digits > 0) {
+    const char *preverse = p - 1;
+    int32_t trailing_zeros = 0;
+    while ((*preverse == '0') || (*preverse == '.')) {
+      if(*preverse == '0') { trailing_zeros++; };
+      --preverse;
+    }
+    answer.decimal_point += int32_t(answer.num_digits);
+    answer.num_digits -= uint32_t(trailing_zeros);
+  }
+  if(answer.num_digits > max_digits ) {
+    answer.num_digits = max_digits;
+    answer.truncated = true;
+  }
+  if ((p != end) && (('e' == *p) || ('E' == *p))) {
+    ++p;
+    if(p == end) { return answer; } // should never happen
+    bool neg_exp = false;
+    if ('-' == *p) {
+      neg_exp = true;
+      ++p;
+    } else if ('+' == *p) {
+      ++p;
+    }
+    int32_t exp_number = 0; // exponential part
+    while ((p != end) && is_integer(*p)) {
+      uint8_t digit = uint8_t(*p - '0');
+      if (exp_number < 0x10000) {
+        exp_number = 10 * exp_number + digit;
+      }
+      ++p;
+    }
+    answer.decimal_point += (neg_exp ? -exp_number : exp_number);
+  }
+  return answer;
+}
+
+namespace {
+
+// remove all final zeroes
+inline void trim(decimal &h) {
+  while ((h.num_digits > 0) && (h.digits[h.num_digits - 1] == 0)) {
+    h.num_digits--;
+  }
+}
+
+uint32_t number_of_digits_decimal_left_shift(decimal &h, uint32_t shift) {
+  shift &= 63;
+  const static uint16_t number_of_digits_decimal_left_shift_table[65] = {
+      0x0000, 0x0800, 0x0801, 0x0803, 0x1006, 0x1009, 0x100D, 0x1812, 0x1817,
+      0x181D, 0x2024, 0x202B, 0x2033, 0x203C, 0x2846, 0x2850, 0x285B, 0x3067,
+      0x3073, 0x3080, 0x388E, 0x389C, 0x38AB, 0x38BB, 0x40CC, 0x40DD, 0x40EF,
+      0x4902, 0x4915, 0x4929, 0x513E, 0x5153, 0x5169, 0x5180, 0x5998, 0x59B0,
+      0x59C9, 0x61E3, 0x61FD, 0x6218, 0x6A34, 0x6A50, 0x6A6D, 0x6A8B, 0x72AA,
+      0x72C9, 0x72E9, 0x7B0A, 0x7B2B, 0x7B4D, 0x8370, 0x8393, 0x83B7, 0x83DC,
+      0x8C02, 0x8C28, 0x8C4F, 0x9477, 0x949F, 0x94C8, 0x9CF2, 0x051C, 0x051C,
+      0x051C, 0x051C,
+  };
+  uint32_t x_a = number_of_digits_decimal_left_shift_table[shift];
+  uint32_t x_b = number_of_digits_decimal_left_shift_table[shift + 1];
+  uint32_t num_new_digits = x_a >> 11;
+  uint32_t pow5_a = 0x7FF & x_a;
+  uint32_t pow5_b = 0x7FF & x_b;
+  const static uint8_t
+      number_of_digits_decimal_left_shift_table_powers_of_5[0x051C] = {
+          5, 2, 5, 1, 2, 5, 6, 2, 5, 3, 1, 2, 5, 1, 5, 6, 2, 5, 7, 8, 1, 2, 5,
+          3, 9, 0, 6, 2, 5, 1, 9, 5, 3, 1, 2, 5, 9, 7, 6, 5, 6, 2, 5, 4, 8, 8,
+          2, 8, 1, 2, 5, 2, 4, 4, 1, 4, 0, 6, 2, 5, 1, 2, 2, 0, 7, 0, 3, 1, 2,
+          5, 6, 1, 0, 3, 5, 1, 5, 6, 2, 5, 3, 0, 5, 1, 7, 5, 7, 8, 1, 2, 5, 1,
+          5, 2, 5, 8, 7, 8, 9, 0, 6, 2, 5, 7, 6, 2, 9, 3, 9, 4, 5, 3, 1, 2, 5,
+          3, 8, 1, 4, 6, 9, 7, 2, 6, 5, 6, 2, 5, 1, 9, 0, 7, 3, 4, 8, 6, 3, 2,
+          8, 1, 2, 5, 9, 5, 3, 6, 7, 4, 3, 1, 6, 4, 0, 6, 2, 5, 4, 7, 6, 8, 3,
+          7, 1, 5, 8, 2, 0, 3, 1, 2, 5, 2, 3, 8, 4, 1, 8, 5, 7, 9, 1, 0, 1, 5,
+          6, 2, 5, 1, 1, 9, 2, 0, 9, 2, 8, 9, 5, 5, 0, 7, 8, 1, 2, 5, 5, 9, 6,
+          0, 4, 6, 4, 4, 7, 7, 5, 3, 9, 0, 6, 2, 5, 2, 9, 8, 0, 2, 3, 2, 2, 3,
+          8, 7, 6, 9, 5, 3, 1, 2, 5, 1, 4, 9, 0, 1, 1, 6, 1, 1, 9, 3, 8, 4, 7,
+          6, 5, 6, 2, 5, 7, 4, 5, 0, 5, 8, 0, 5, 9, 6, 9, 2, 3, 8, 2, 8, 1, 2,
+          5, 3, 7, 2, 5, 2, 9, 0, 2, 9, 8, 4, 6, 1, 9, 1, 4, 0, 6, 2, 5, 1, 8,
+          6, 2, 6, 4, 5, 1, 4, 9, 2, 3, 0, 9, 5, 7, 0, 3, 1, 2, 5, 9, 3, 1, 3,
+          2, 2, 5, 7, 4, 6, 1, 5, 4, 7, 8, 5, 1, 5, 6, 2, 5, 4, 6, 5, 6, 6, 1,
+          2, 8, 7, 3, 0, 7, 7, 3, 9, 2, 5, 7, 8, 1, 2, 5, 2, 3, 2, 8, 3, 0, 6,
+          4, 3, 6, 5, 3, 8, 6, 9, 6, 2, 8, 9, 0, 6, 2, 5, 1, 1, 6, 4, 1, 5, 3,
+          2, 1, 8, 2, 6, 9, 3, 4, 8, 1, 4, 4, 5, 3, 1, 2, 5, 5, 8, 2, 0, 7, 6,
+          6, 0, 9, 1, 3, 4, 6, 7, 4, 0, 7, 2, 2, 6, 5, 6, 2, 5, 2, 9, 1, 0, 3,
+          8, 3, 0, 4, 5, 6, 7, 3, 3, 7, 0, 3, 6, 1, 3, 2, 8, 1, 2, 5, 1, 4, 5,
+          5, 1, 9, 1, 5, 2, 2, 8, 3, 6, 6, 8, 5, 1, 8, 0, 6, 6, 4, 0, 6, 2, 5,
+          7, 2, 7, 5, 9, 5, 7, 6, 1, 4, 1, 8, 3, 4, 2, 5, 9, 0, 3, 3, 2, 0, 3,
+          1, 2, 5, 3, 6, 3, 7, 9, 7, 8, 8, 0, 7, 0, 9, 1, 7, 1, 2, 9, 5, 1, 6,
+          6, 0, 1, 5, 6, 2, 5, 1, 8, 1, 8, 9, 8, 9, 4, 0, 3, 5, 4, 5, 8, 5, 6,
+          4, 7, 5, 8, 3, 0, 0, 7, 8, 1, 2, 5, 9, 0, 9, 4, 9, 4, 7, 0, 1, 7, 7,
+          2, 9, 2, 8, 2, 3, 7, 9, 1, 5, 0, 3, 9, 0, 6, 2, 5, 4, 5, 4, 7, 4, 7,
+          3, 5, 0, 8, 8, 6, 4, 6, 4, 1, 1, 8, 9, 5, 7, 5, 1, 9, 5, 3, 1, 2, 5,
+          2, 2, 7, 3, 7, 3, 6, 7, 5, 4, 4, 3, 2, 3, 2, 0, 5, 9, 4, 7, 8, 7, 5,
+          9, 7, 6, 5, 6, 2, 5, 1, 1, 3, 6, 8, 6, 8, 3, 7, 7, 2, 1, 6, 1, 6, 0,
+          2, 9, 7, 3, 9, 3, 7, 9, 8, 8, 2, 8, 1, 2, 5, 5, 6, 8, 4, 3, 4, 1, 8,
+          8, 6, 0, 8, 0, 8, 0, 1, 4, 8, 6, 9, 6, 8, 9, 9, 4, 1, 4, 0, 6, 2, 5,
+          2, 8, 4, 2, 1, 7, 0, 9, 4, 3, 0, 4, 0, 4, 0, 0, 7, 4, 3, 4, 8, 4, 4,
+          9, 7, 0, 7, 0, 3, 1, 2, 5, 1, 4, 2, 1, 0, 8, 5, 4, 7, 1, 5, 2, 0, 2,
+          0, 0, 3, 7, 1, 7, 4, 2, 2, 4, 8, 5, 3, 5, 1, 5, 6, 2, 5, 7, 1, 0, 5,
+          4, 2, 7, 3, 5, 7, 6, 0, 1, 0, 0, 1, 8, 5, 8, 7, 1, 1, 2, 4, 2, 6, 7,
+          5, 7, 8, 1, 2, 5, 3, 5, 5, 2, 7, 1, 3, 6, 7, 8, 8, 0, 0, 5, 0, 0, 9,
+          2, 9, 3, 5, 5, 6, 2, 1, 3, 3, 7, 8, 9, 0, 6, 2, 5, 1, 7, 7, 6, 3, 5,
+          6, 8, 3, 9, 4, 0, 0, 2, 5, 0, 4, 6, 4, 6, 7, 7, 8, 1, 0, 6, 6, 8, 9,
+          4, 5, 3, 1, 2, 5, 8, 8, 8, 1, 7, 8, 4, 1, 9, 7, 0, 0, 1, 2, 5, 2, 3,
+          2, 3, 3, 8, 9, 0, 5, 3, 3, 4, 4, 7, 2, 6, 5, 6, 2, 5, 4, 4, 4, 0, 8,
+          9, 2, 0, 9, 8, 5, 0, 0, 6, 2, 6, 1, 6, 1, 6, 9, 4, 5, 2, 6, 6, 7, 2,
+          3, 6, 3, 2, 8, 1, 2, 5, 2, 2, 2, 0, 4, 4, 6, 0, 4, 9, 2, 5, 0, 3, 1,
+          3, 0, 8, 0, 8, 4, 7, 2, 6, 3, 3, 3, 6, 1, 8, 1, 6, 4, 0, 6, 2, 5, 1,
+          1, 1, 0, 2, 2, 3, 0, 2, 4, 6, 2, 5, 1, 5, 6, 5, 4, 0, 4, 2, 3, 6, 3,
+          1, 6, 6, 8, 0, 9, 0, 8, 2, 0, 3, 1, 2, 5, 5, 5, 5, 1, 1, 1, 5, 1, 2,
+          3, 1, 2, 5, 7, 8, 2, 7, 0, 2, 1, 1, 8, 1, 5, 8, 3, 4, 0, 4, 5, 4, 1,
+          0, 1, 5, 6, 2, 5, 2, 7, 7, 5, 5, 5, 7, 5, 6, 1, 5, 6, 2, 8, 9, 1, 3,
+          5, 1, 0, 5, 9, 0, 7, 9, 1, 7, 0, 2, 2, 7, 0, 5, 0, 7, 8, 1, 2, 5, 1,
+          3, 8, 7, 7, 7, 8, 7, 8, 0, 7, 8, 1, 4, 4, 5, 6, 7, 5, 5, 2, 9, 5, 3,
+          9, 5, 8, 5, 1, 1, 3, 5, 2, 5, 3, 9, 0, 6, 2, 5, 6, 9, 3, 8, 8, 9, 3,
+          9, 0, 3, 9, 0, 7, 2, 2, 8, 3, 7, 7, 6, 4, 7, 6, 9, 7, 9, 2, 5, 5, 6,
+          7, 6, 2, 6, 9, 5, 3, 1, 2, 5, 3, 4, 6, 9, 4, 4, 6, 9, 5, 1, 9, 5, 3,
+          6, 1, 4, 1, 8, 8, 8, 2, 3, 8, 4, 8, 9, 6, 2, 7, 8, 3, 8, 1, 3, 4, 7,
+          6, 5, 6, 2, 5, 1, 7, 3, 4, 7, 2, 3, 4, 7, 5, 9, 7, 6, 8, 0, 7, 0, 9,
+          4, 4, 1, 1, 9, 2, 4, 4, 8, 1, 3, 9, 1, 9, 0, 6, 7, 3, 8, 2, 8, 1, 2,
+          5, 8, 6, 7, 3, 6, 1, 7, 3, 7, 9, 8, 8, 4, 0, 3, 5, 4, 7, 2, 0, 5, 9,
+          6, 2, 2, 4, 0, 6, 9, 5, 9, 5, 3, 3, 6, 9, 1, 4, 0, 6, 2, 5,
+      };
+  const uint8_t *pow5 =
+      &number_of_digits_decimal_left_shift_table_powers_of_5[pow5_a];
+  uint32_t i = 0;
+  uint32_t n = pow5_b - pow5_a;
+  for (; i < n; i++) {
+    if (i >= h.num_digits) {
+      return num_new_digits - 1;
+    } else if (h.digits[i] == pow5[i]) {
+      continue;
+    } else if (h.digits[i] < pow5[i]) {
+      return num_new_digits - 1;
+    } else {
+      return num_new_digits;
+    }
+  }
+  return num_new_digits;
+}
+
+} // end of anonymous namespace
+
+uint64_t round(decimal &h) {
+  if ((h.num_digits == 0) || (h.decimal_point < 0)) {
+    return 0;
+  } else if (h.decimal_point > 18) {
+    return UINT64_MAX;
+  }
+  // at this point, we know that h.decimal_point >= 0
+  uint32_t dp = uint32_t(h.decimal_point);
+  uint64_t n = 0;
+  for (uint32_t i = 0; i < dp; i++) {
+    n = (10 * n) + ((i < h.num_digits) ? h.digits[i] : 0);
+  }
+  bool round_up = false;
+  if (dp < h.num_digits) {
+    round_up = h.digits[dp] >= 5; // normally, we round up
+    // but we may need to round to even!
+    if ((h.digits[dp] == 5) && (dp + 1 == h.num_digits)) {
+      round_up = h.truncated || ((dp > 0) && (1 & h.digits[dp - 1]));
+    }
+  }
+  if (round_up) {
+    n++;
+  }
+  return n;
+}
+
+// computes h * 2^-shift
+void decimal_left_shift(decimal &h, uint32_t shift) {
+  if (h.num_digits == 0) {
+    return;
+  }
+  uint32_t num_new_digits = number_of_digits_decimal_left_shift(h, shift);
+  int32_t read_index = int32_t(h.num_digits - 1);
+  uint32_t write_index = h.num_digits - 1 + num_new_digits;
+  uint64_t n = 0;
+
+  while (read_index >= 0) {
+    n += uint64_t(h.digits[read_index]) << shift;
+    uint64_t quotient = n / 10;
+    uint64_t remainder = n - (10 * quotient);
+    if (write_index < max_digits) {
+      h.digits[write_index] = uint8_t(remainder);
+    } else if (remainder > 0) {
+      h.truncated = true;
+    }
+    n = quotient;
+    write_index--;
+    read_index--;
+  }
+  while (n > 0) {
+    uint64_t quotient = n / 10;
+    uint64_t remainder = n - (10 * quotient);
+    if (write_index < max_digits) {
+      h.digits[write_index] = uint8_t(remainder);
+    } else if (remainder > 0) {
+      h.truncated = true;
+    }
+    n = quotient;
+    write_index--;
+  }
+  h.num_digits += num_new_digits;
+  if (h.num_digits > max_digits) {
+    h.num_digits = max_digits;
+  }
+  h.decimal_point += int32_t(num_new_digits);
+  trim(h);
+}
+
+// computes h * 2^shift
+void decimal_right_shift(decimal &h, uint32_t shift) {
+  uint32_t read_index = 0;
+  uint32_t write_index = 0;
+
+  uint64_t n = 0;
+
+  while ((n >> shift) == 0) {
+    if (read_index < h.num_digits) {
+      n = (10 * n) + h.digits[read_index++];
+    } else if (n == 0) {
+      return;
+    } else {
+      while ((n >> shift) == 0) {
+        n = 10 * n;
+        read_index++;
+      }
+      break;
+    }
+  }
+  h.decimal_point -= int32_t(read_index - 1);
+  if (h.decimal_point < -decimal_point_range) { // it is zero
+    h.num_digits = 0;
+    h.decimal_point = 0;
+    h.negative = false;
+    h.truncated = false;
+    return;
+  }
+  uint64_t mask = (uint64_t(1) << shift) - 1;
+  while (read_index < h.num_digits) {
+    uint8_t new_digit = uint8_t(n >> shift);
+    n = (10 * (n & mask)) + h.digits[read_index++];
+    h.digits[write_index++] = new_digit;
+  }
+  while (n > 0) {
+    uint8_t new_digit = uint8_t(n >> shift);
+    n = 10 * (n & mask);
+    if (write_index < max_digits) {
+      h.digits[write_index++] = new_digit;
+    } else if (new_digit > 0) {
+      h.truncated = true;
+    }
+  }
+  h.num_digits = write_index;
+  trim(h);
+}
+
+template <typename binary> adjusted_mantissa compute_float(decimal &d) {
+  adjusted_mantissa answer;
+  if (d.num_digits == 0) {
+    // should be zero
+    answer.power2 = 0;
+    answer.mantissa = 0;
+    return answer;
+  }
+  // At this point, going further, we can assume that d.num_digits > 0.
+  // We want to guard against excessive decimal point values because
+  // they can result in long running times. Indeed, we do
+  // shifts by at most 60 bits. We have that log(10**400)/log(2**60) ~= 22
+  // which is fine, but log(10**299995)/log(2**60) ~= 16609 which is not
+  // fine (runs for a long time).
+  //
+  if(d.decimal_point < -324) {
+    // We have something smaller than 1e-324 which is always zero
+    // in binary64 and binary32.
+    // It should be zero.
+    answer.power2 = 0;
+    answer.mantissa = 0;
+    return answer;
+  } else if(d.decimal_point >= 310) {
+    // We have something at least as large as 0.1e310 which is
+    // always infinite.
+    answer.power2 = binary::infinite_power();
+    answer.mantissa = 0;
+    return answer;
+  }
+
+  static const uint32_t max_shift = 60;
+  static const uint32_t num_powers = 19;
+  static const uint8_t powers[19] = {
+      0,  3,  6,  9,  13, 16, 19, 23, 26, 29, //
+      33, 36, 39, 43, 46, 49, 53, 56, 59,     //
+  };
+  int32_t exp2 = 0;
+  while (d.decimal_point > 0) {
+    uint32_t n = uint32_t(d.decimal_point);
+    uint32_t shift = (n < num_powers) ? powers[n] : max_shift;
+    decimal_right_shift(d, shift);
+    if (d.decimal_point < -decimal_point_range) {
+      // should be zero
+      answer.power2 = 0;
+      answer.mantissa = 0;
+      return answer;
+    }
+    exp2 += int32_t(shift);
+  }
+  // We shift left toward [1/2 ... 1].
+  while (d.decimal_point <= 0) {
+    uint32_t shift;
+    if (d.decimal_point == 0) {
+      if (d.digits[0] >= 5) {
+        break;
+      }
+      shift = (d.digits[0] < 2) ? 2 : 1;
+    } else {
+      uint32_t n = uint32_t(-d.decimal_point);
+      shift = (n < num_powers) ? powers[n] : max_shift;
+    }
+    decimal_left_shift(d, shift);
+    if (d.decimal_point > decimal_point_range) {
+      // we want to get infinity:
+      answer.power2 = 0xFF;
+      answer.mantissa = 0;
+      return answer;
+    }
+    exp2 -= int32_t(shift);
+  }
+  // We are now in the range [1/2 ... 1] but the binary format uses [1 ... 2].
+  exp2--;
+  constexpr int32_t minimum_exponent = binary::minimum_exponent();
+  while ((minimum_exponent + 1) > exp2) {
+    uint32_t n = uint32_t((minimum_exponent + 1) - exp2);
+    if (n > max_shift) {
+      n = max_shift;
+    }
+    decimal_right_shift(d, n);
+    exp2 += int32_t(n);
+  }
+  if ((exp2 - minimum_exponent) >= binary::infinite_power()) {
+    answer.power2 = binary::infinite_power();
+    answer.mantissa = 0;
+    return answer;
+  }
+
+  const int mantissa_size_in_bits = binary::mantissa_explicit_bits() + 1;
+  decimal_left_shift(d, mantissa_size_in_bits);
+
+  uint64_t mantissa = round(d);
+  // It is possible that we have an overflow, in which case we need
+  // to shift back.
+  if (mantissa >= (uint64_t(1) << mantissa_size_in_bits)) {
+    decimal_right_shift(d, 1);
+    exp2 += 1;
+    mantissa = round(d);
+    if ((exp2 - minimum_exponent) >= binary::infinite_power()) {
+      answer.power2 = binary::infinite_power();
+      answer.mantissa = 0;
+      return answer;
+    }
+  }
+  answer.power2 = exp2 - binary::minimum_exponent();
+  if (mantissa < (uint64_t(1) << binary::mantissa_explicit_bits())) {
+    answer.power2--;
+  }
+  answer.mantissa =
+      mantissa & ((uint64_t(1) << binary::mantissa_explicit_bits()) - 1);
+  return answer;
+}
+
+template <typename binary>
+adjusted_mantissa parse_long_mantissa(const char *first) {
+  decimal d = parse_decimal(first);
+  return compute_float<binary>(d);
+}
+
+template <typename binary>
+adjusted_mantissa parse_long_mantissa(const char *first, const char *end) {
+  decimal d = parse_decimal(first, end);
+  return compute_float<binary>(d);
+}
+
+double from_chars(const char *first) noexcept {
+  bool negative = first[0] == '-';
+  if (negative) {
+    first++;
+  }
+  adjusted_mantissa am = parse_long_mantissa<binary_format<double>>(first);
+  uint64_t word = am.mantissa;
+  word |= uint64_t(am.power2)
+          << binary_format<double>::mantissa_explicit_bits();
+  word = negative ? word | (uint64_t(1) << binary_format<double>::sign_index())
+                  : word;
+  double value;
+  std::memcpy(&value, &word, sizeof(double));
+  return value;
+}
+
+
+double from_chars(const char *first, const char *end) noexcept {
+  bool negative = first[0] == '-';
+  if (negative) {
+    first++;
+  }
+  adjusted_mantissa am = parse_long_mantissa<binary_format<double>>(first, end);
+  uint64_t word = am.mantissa;
+  word |= uint64_t(am.power2)
+          << binary_format<double>::mantissa_explicit_bits();
+  word = negative ? word | (uint64_t(1) << binary_format<double>::sign_index())
+                  : word;
+  double value;
+  std::memcpy(&value, &word, sizeof(double));
+  return value;
+}
+
+} // internal
+} // simdjson
+
+#endif // SIMDJSON_SRC_FROM_CHARS_CPP
+/* end file from_chars.cpp */
+/* including internal/error_tables.cpp: #include <internal/error_tables.cpp> */
+/* begin file internal/error_tables.cpp */
+#ifndef SIMDJSON_SRC_ERROR_TABLES_CPP
+#define SIMDJSON_SRC_ERROR_TABLES_CPP
+
+/* including simdjson/internal/jsoncharutils_tables.h: #include <simdjson/internal/jsoncharutils_tables.h> */
+/* begin file simdjson/internal/jsoncharutils_tables.h */
+#ifndef SIMDJSON_INTERNAL_JSONCHARUTILS_TABLES_H
+#define SIMDJSON_INTERNAL_JSONCHARUTILS_TABLES_H
+
+/* skipped duplicate #include "simdjson/base.h" */
+
+#ifdef JSON_TEST_STRINGS
+void found_string(const uint8_t *buf, const uint8_t *parsed_begin,
+                  const uint8_t *parsed_end);
+void found_bad_string(const uint8_t *buf);
+#endif
+
+namespace simdjson {
+namespace internal {
+// structural chars here are
+// they are { 0x7b } 0x7d : 0x3a [ 0x5b ] 0x5d , 0x2c (and NULL)
+// we are also interested in the four whitespace characters
+// space 0x20, linefeed 0x0a, horizontal tab 0x09 and carriage return 0x0d
+
+extern SIMDJSON_DLLIMPORTEXPORT const bool structural_or_whitespace_negated[256];
+extern SIMDJSON_DLLIMPORTEXPORT const bool structural_or_whitespace[256];
+extern SIMDJSON_DLLIMPORTEXPORT const uint32_t digit_to_val32[886];
+
+} // namespace internal
+} // namespace simdjson
+
+#endif // SIMDJSON_INTERNAL_JSONCHARUTILS_TABLES_H
+/* end file simdjson/internal/jsoncharutils_tables.h */
+/* including simdjson/error-inl.h: #include <simdjson/error-inl.h> */
+/* begin file simdjson/error-inl.h */
+#ifndef SIMDJSON_ERROR_INL_H
+#define SIMDJSON_ERROR_INL_H
+
+/* skipped duplicate #include "simdjson/error.h" */
+
+#include <iostream>
+
+namespace simdjson {
+
+inline bool is_fatal(error_code error) noexcept {
+  return error == TAPE_ERROR || error == INCOMPLETE_ARRAY_OR_OBJECT || error == OUT_OF_ORDER_ITERATION || error == DEPTH_ERROR;
+}
+
+namespace internal {
+  // We store the error code so we can validate the error message is associated with the right code
+  struct error_code_info {
+    error_code code;
+    const char* message; // do not use a fancy std::string where a simple C string will do (no alloc, no destructor)
+  };
+  // These MUST match the codes in error_code. We check this constraint in basictests.
+  extern SIMDJSON_DLLIMPORTEXPORT const error_code_info error_codes[];
+} // namespace internal
+
+
+inline const char *error_message(error_code error) noexcept {
+  // If you're using error_code, we're trusting you got it from the enum.
+  return internal::error_codes[int(error)].message;
+}
+
+// deprecated function
+#ifndef SIMDJSON_DISABLE_DEPRECATED_API
+inline const std::string error_message(int error) noexcept {
+  if (error < 0 || error >= error_code::NUM_ERROR_CODES) {
+    return internal::error_codes[UNEXPECTED_ERROR].message;
+  }
+  return internal::error_codes[error].message;
+}
+#endif // SIMDJSON_DISABLE_DEPRECATED_API
+
+inline std::ostream& operator<<(std::ostream& out, error_code error) noexcept {
+  return out << error_message(error);
+}
+
+namespace internal {
+
+//
+// internal::simdjson_result_base<T> inline implementation
+//
+
+template<typename T>
+simdjson_inline void simdjson_result_base<T>::tie(T &value, error_code &error) && noexcept {
+  error = this->second;
+  if (!error) {
+    value = std::forward<simdjson_result_base<T>>(*this).first;
+  }
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code simdjson_result_base<T>::get(T &value) && noexcept {
+  error_code error;
+  std::forward<simdjson_result_base<T>>(*this).tie(value, error);
+  return error;
+}
+
+template<typename T>
+simdjson_inline error_code simdjson_result_base<T>::error() const noexcept {
+  return this->second;
+}
+
+
+template<typename T>
+simdjson_inline bool simdjson_result_base<T>::has_value() const noexcept {
+  return this->error() == SUCCESS;
+}
+
+#if SIMDJSON_EXCEPTIONS
+
+
+template<typename T>
+simdjson_inline T& simdjson_result_base<T>::operator*() &  noexcept(false) {
+  return this->value();
+}
+
+template<typename T>
+simdjson_inline T&& simdjson_result_base<T>::operator*() &&  noexcept(false) {
+  return std::forward<internal::simdjson_result_base<T>>(*this).value();
+}
+
+template<typename T>
+simdjson_inline T* simdjson_result_base<T>::operator->() noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+
+template<typename T>
+simdjson_inline const T* simdjson_result_base<T>::operator->() const noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+template<typename T>
+simdjson_inline T& simdjson_result_base<T>::value() & noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& simdjson_result_base<T>::value() && noexcept(false) {
+  return std::forward<simdjson_result_base<T>>(*this).take_value();
+}
+
+template<typename T>
+simdjson_inline T&& simdjson_result_base<T>::take_value() && noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline simdjson_result_base<T>::operator T&&() && noexcept(false) {
+  return std::forward<simdjson_result_base<T>>(*this).take_value();
+}
+
+#endif // SIMDJSON_EXCEPTIONS
+
+
+template<typename T>
+simdjson_inline const T& simdjson_result_base<T>::value_unsafe() const& noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& simdjson_result_base<T>::value_unsafe() && noexcept {
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline simdjson_result_base<T>::simdjson_result_base(T &&value, error_code error) noexcept
+    : std::pair<T, error_code>(std::forward<T>(value), error) {}
+template<typename T>
+simdjson_inline simdjson_result_base<T>::simdjson_result_base(error_code error) noexcept
+    : simdjson_result_base(T{}, error) {}
+template<typename T>
+simdjson_inline simdjson_result_base<T>::simdjson_result_base(T &&value) noexcept
+    : simdjson_result_base(std::forward<T>(value), SUCCESS) {}
+template<typename T>
+simdjson_inline simdjson_result_base<T>::simdjson_result_base() noexcept
+    : simdjson_result_base(T{}, UNINITIALIZED) {}
+
+} // namespace internal
+
+///
+/// simdjson_result<T> inline implementation
+///
+
+template<typename T>
+simdjson_inline void simdjson_result<T>::tie(T &value, error_code &error) && noexcept {
+  std::forward<internal::simdjson_result_base<T>>(*this).tie(value, error);
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code
+simdjson_result<T>::get(T &value) && noexcept {
+  return std::forward<internal::simdjson_result_base<T>>(*this).get(value);
+}
+
+template<typename T>
+simdjson_inline error_code simdjson_result<T>::error() const noexcept {
+  return internal::simdjson_result_base<T>::error();
+}
+
+#if SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline T& simdjson_result<T>::value() & noexcept(false) {
+  return internal::simdjson_result_base<T>::value();
+}
+
+template<typename T>
+simdjson_inline T&& simdjson_result<T>::value() && noexcept(false) {
+  return std::forward<internal::simdjson_result_base<T>>(*this).value();
+}
+
+template<typename T>
+simdjson_inline T&& simdjson_result<T>::take_value() && noexcept(false) {
+  return std::forward<internal::simdjson_result_base<T>>(*this).take_value();
+}
+
+template<typename T>
+simdjson_inline simdjson_result<T>::operator T&&() && noexcept(false) {
+  return std::forward<internal::simdjson_result_base<T>>(*this).take_value();
+}
+
+#endif // SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline const T& simdjson_result<T>::value_unsafe() const& noexcept {
+  return internal::simdjson_result_base<T>::value_unsafe();
+}
+
+template<typename T>
+simdjson_inline T&& simdjson_result<T>::value_unsafe() && noexcept {
+  return std::forward<internal::simdjson_result_base<T>>(*this).value_unsafe();
+}
+
+template<typename T>
+simdjson_inline simdjson_result<T>::simdjson_result(T &&value, error_code error) noexcept
+    : internal::simdjson_result_base<T>(std::forward<T>(value), error) {}
+template<typename T>
+simdjson_inline simdjson_result<T>::simdjson_result(error_code error) noexcept
+    : internal::simdjson_result_base<T>(error) {}
+template<typename T>
+simdjson_inline simdjson_result<T>::simdjson_result(T &&value) noexcept
+    : internal::simdjson_result_base<T>(std::forward<T>(value)) {}
+template<typename T>
+simdjson_inline simdjson_result<T>::simdjson_result() noexcept
+    : internal::simdjson_result_base<T>() {}
+
+} // namespace simdjson
+
+#endif // SIMDJSON_ERROR_INL_H
+/* end file simdjson/error-inl.h */
+
+namespace simdjson {
+namespace internal {
+
+  SIMDJSON_DLLIMPORTEXPORT const error_code_info error_codes[] {
+    { SUCCESS, "SUCCESS: No error" },
+    { CAPACITY, "CAPACITY: This parser can't support a document that big" },
+    { MEMALLOC, "MEMALLOC: Error allocating memory, we're most likely out of memory" },
+    { TAPE_ERROR, "TAPE_ERROR: The JSON document has an improper structure: missing or superfluous commas, braces, missing keys, etc.  This is a fatal and unrecoverable error." },
+    { DEPTH_ERROR, "DEPTH_ERROR: The JSON document was too deep (too many nested objects and arrays)" },
+    { STRING_ERROR, "STRING_ERROR: Problem while parsing a string" },
+    { T_ATOM_ERROR, "T_ATOM_ERROR: Problem while parsing an atom starting with the letter 't'" },
+    { F_ATOM_ERROR, "F_ATOM_ERROR: Problem while parsing an atom starting with the letter 'f'" },
+    { N_ATOM_ERROR, "N_ATOM_ERROR: Problem while parsing an atom starting with the letter 'n'" },
+    { NUMBER_ERROR, "NUMBER_ERROR: Problem while parsing a number" },
+    { BIGINT_ERROR, "BIGINT_ERROR: Big integer value that cannot be represented using 64 bits" },
+    { UTF8_ERROR, "UTF8_ERROR: The input is not valid UTF-8" },
+    { UNINITIALIZED, "UNINITIALIZED: Uninitialized" },
+    { EMPTY, "EMPTY: no JSON found" },
+    { UNESCAPED_CHARS, "UNESCAPED_CHARS: Within strings, some characters must be escaped, we found unescaped characters" },
+    { UNCLOSED_STRING, "UNCLOSED_STRING: A string is opened, but never closed." },
+    { UNSUPPORTED_ARCHITECTURE, "UNSUPPORTED_ARCHITECTURE: simdjson does not have an implementation supported by this CPU architecture. Please report this error to the core team as it should never happen." },
+    { INCORRECT_TYPE, "INCORRECT_TYPE: The JSON element does not have the requested type." },
+    { NUMBER_OUT_OF_RANGE, "NUMBER_OUT_OF_RANGE: The JSON number is too large or too small to fit within the requested type." },
+    { INDEX_OUT_OF_BOUNDS, "INDEX_OUT_OF_BOUNDS: Attempted to access an element of a JSON array that is beyond its length." },
+    { NO_SUCH_FIELD, "NO_SUCH_FIELD: The JSON field referenced does not exist in this object." },
+    { IO_ERROR, "IO_ERROR: Error reading the file." },
+    { INVALID_JSON_POINTER, "INVALID_JSON_POINTER: Invalid JSON pointer syntax." },
+    { INVALID_URI_FRAGMENT, "INVALID_URI_FRAGMENT: Invalid URI fragment syntax." },
+    { UNEXPECTED_ERROR, "UNEXPECTED_ERROR: Unexpected error, consider reporting this problem as you may have found a bug in simdjson" },
+    { PARSER_IN_USE, "PARSER_IN_USE: Cannot parse a new document while a document is still in use." },
+    { OUT_OF_ORDER_ITERATION, "OUT_OF_ORDER_ITERATION: Objects and arrays can only be iterated when they are first encountered." },
+    { INSUFFICIENT_PADDING, "INSUFFICIENT_PADDING: simdjson requires the input JSON string to have at least SIMDJSON_PADDING extra bytes allocated, beyond the string's length. Consider using the simdjson::padded_string class if needed." },
+    { INCOMPLETE_ARRAY_OR_OBJECT, "INCOMPLETE_ARRAY_OR_OBJECT: JSON document ended early in the middle of an object or array. This is a fatal and unrecoverable error." },
+    { SCALAR_DOCUMENT_AS_VALUE, "SCALAR_DOCUMENT_AS_VALUE: A JSON document made of a scalar (number, Boolean, null or string) is treated as a value. Use get_bool(), get_double(), etc. on the document instead. "},
+    { OUT_OF_BOUNDS, "OUT_OF_BOUNDS: Attempt to access location outside of document."},
+    { TRAILING_CONTENT, "TRAILING_CONTENT: Unexpected trailing content in the JSON input."},
+    { OUT_OF_CAPACITY, "OUT_OF_CAPACITY: The capacity was exceeded, we cannot allocate enough memory."}
+  }; // error_messages[]
+
+} // namespace internal
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_ERROR_TABLES_CPP
+/* end file internal/error_tables.cpp */
+/* including internal/jsoncharutils_tables.cpp: #include <internal/jsoncharutils_tables.cpp> */
+/* begin file internal/jsoncharutils_tables.cpp */
+#ifndef SIMDJSON_SRC_JSONCHARUTILS_TABLES_CPP
+#define SIMDJSON_SRC_JSONCHARUTILS_TABLES_CPP
+
+/* skipped duplicate #include <simdjson/base.h> */
+
+namespace simdjson {
+namespace internal {
+
+// structural chars here are
+// they are { 0x7b } 0x7d : 0x3a [ 0x5b ] 0x5d , 0x2c (and NULL)
+// we are also interested in the four whitespace characters
+// space 0x20, linefeed 0x0a, horizontal tab 0x09 and carriage return 0x0d
+
+SIMDJSON_DLLIMPORTEXPORT const bool structural_or_whitespace_negated[256] = {
+    1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1,
+    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+    0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1,
+
+    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+    1, 1, 1, 1, 1, 0, 1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 0, 1, 1,
+
+    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+
+    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
+    1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1};
+
+SIMDJSON_DLLIMPORTEXPORT const bool structural_or_whitespace[256] = {
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 1, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+    0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
+
+SIMDJSON_DLLIMPORTEXPORT const uint32_t digit_to_val32[886] = {
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0x0,        0x1,        0x2,        0x3,        0x4,        0x5,
+    0x6,        0x7,        0x8,        0x9,        0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xa,
+    0xb,        0xc,        0xd,        0xe,        0xf,        0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xa,        0xb,        0xc,        0xd,        0xe,
+    0xf,        0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0x0,        0x10,       0x20,       0x30,       0x40,       0x50,
+    0x60,       0x70,       0x80,       0x90,       0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xa0,
+    0xb0,       0xc0,       0xd0,       0xe0,       0xf0,       0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xa0,       0xb0,       0xc0,       0xd0,       0xe0,
+    0xf0,       0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0x0,        0x100,      0x200,      0x300,      0x400,      0x500,
+    0x600,      0x700,      0x800,      0x900,      0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xa00,
+    0xb00,      0xc00,      0xd00,      0xe00,      0xf00,      0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xa00,      0xb00,      0xc00,      0xd00,      0xe00,
+    0xf00,      0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0x0,        0x1000,     0x2000,     0x3000,     0x4000,     0x5000,
+    0x6000,     0x7000,     0x8000,     0x9000,     0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xa000,
+    0xb000,     0xc000,     0xd000,     0xe000,     0xf000,     0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xa000,     0xb000,     0xc000,     0xd000,     0xe000,
+    0xf000,     0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF,
+    0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF, 0xFFFFFFFF};
+
+} // namespace internal
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_JSONCHARUTILS_TABLES_CPP
+/* end file internal/jsoncharutils_tables.cpp */
+/* including internal/numberparsing_tables.cpp: #include <internal/numberparsing_tables.cpp> */
+/* begin file internal/numberparsing_tables.cpp */
+#ifndef SIMDJSON_SRC_NUMBERPARSING_TABLES_CPP
+#define SIMDJSON_SRC_NUMBERPARSING_TABLES_CPP
+
+/* skipped duplicate #include <simdjson/base.h> */
+/* including simdjson/internal/numberparsing_tables.h: #include <simdjson/internal/numberparsing_tables.h> */
+/* begin file simdjson/internal/numberparsing_tables.h */
+#ifndef SIMDJSON_INTERNAL_NUMBERPARSING_TABLES_H
+#define SIMDJSON_INTERNAL_NUMBERPARSING_TABLES_H
+
+/* skipped duplicate #include "simdjson/base.h" */
+
+namespace simdjson {
+namespace internal {
+/**
+ * The smallest non-zero float (binary64) is 2^-1074.
+ * We take as input numbers of the form w x 10^q where w < 2^64.
+ * We have that w * 10^-343  <  2^(64-344) 5^-343 < 2^-1076.
+ * However, we have that
+ * (2^64-1) * 10^-342 =  (2^64-1) * 2^-342 * 5^-342 > 2^-1074.
+ * Thus it is possible for a number of the form w * 10^-342 where
+ * w is a 64-bit value to be a non-zero floating-point number.
+ *********
+ * Any number of form w * 10^309 where w>= 1 is going to be
+ * infinite in binary64 so we never need to worry about powers
+ * of 5 greater than 308.
+ */
+constexpr int smallest_power = -342;
+constexpr int largest_power = 308;
+
+/**
+ * Represents a 128-bit value.
+ * low: least significant 64 bits.
+ * high: most significant 64 bits.
+ */
+struct value128 {
+  uint64_t low;
+  uint64_t high;
+};
+
+
+// Precomputed powers of ten from 10^0 to 10^22. These
+// can be represented exactly using the double type.
+extern SIMDJSON_DLLIMPORTEXPORT const double power_of_ten[];
+
+
+/**
+ * When mapping numbers from decimal to binary,
+ * we go from w * 10^q to m * 2^p but we have
+ * 10^q = 5^q * 2^q, so effectively
+ * we are trying to match
+ * w * 2^q * 5^q to m * 2^p. Thus the powers of two
+ * are not a concern since they can be represented
+ * exactly using the binary notation, only the powers of five
+ * affect the binary significand.
+ */
+
+
+// The truncated powers of five from 5^-342 all the way to 5^308
+// The mantissa is truncated to 128 bits, and
+// never rounded up. Uses about 10KB.
+// We use the template trick to allow inclusion in multiple translation units.
+#if SIMDJSON_STATIC_REFLECTION
+template <typename unused = void>  struct powers_template {
+constexpr static uint64_t power_of_five_128[651*2]= {
+        0xeef453d6923bd65a,0x113faa2906a13b3f,
+        0x9558b4661b6565f8,0x4ac7ca59a424c507,
+        0xbaaee17fa23ebf76,0x5d79bcf00d2df649,
+        0xe95a99df8ace6f53,0xf4d82c2c107973dc,
+        0x91d8a02bb6c10594,0x79071b9b8a4be869,
+        0xb64ec836a47146f9,0x9748e2826cdee284,
+        0xe3e27a444d8d98b7,0xfd1b1b2308169b25,
+        0x8e6d8c6ab0787f72,0xfe30f0f5e50e20f7,
+        0xb208ef855c969f4f,0xbdbd2d335e51a935,
+        0xde8b2b66b3bc4723,0xad2c788035e61382,
+        0x8b16fb203055ac76,0x4c3bcb5021afcc31,
+        0xaddcb9e83c6b1793,0xdf4abe242a1bbf3d,
+        0xd953e8624b85dd78,0xd71d6dad34a2af0d,
+        0x87d4713d6f33aa6b,0x8672648c40e5ad68,
+        0xa9c98d8ccb009506,0x680efdaf511f18c2,
+        0xd43bf0effdc0ba48,0x212bd1b2566def2,
+        0x84a57695fe98746d,0x14bb630f7604b57,
+        0xa5ced43b7e3e9188,0x419ea3bd35385e2d,
+        0xcf42894a5dce35ea,0x52064cac828675b9,
+        0x818995ce7aa0e1b2,0x7343efebd1940993,
+        0xa1ebfb4219491a1f,0x1014ebe6c5f90bf8,
+        0xca66fa129f9b60a6,0xd41a26e077774ef6,
+        0xfd00b897478238d0,0x8920b098955522b4,
+        0x9e20735e8cb16382,0x55b46e5f5d5535b0,
+        0xc5a890362fddbc62,0xeb2189f734aa831d,
+        0xf712b443bbd52b7b,0xa5e9ec7501d523e4,
+        0x9a6bb0aa55653b2d,0x47b233c92125366e,
+        0xc1069cd4eabe89f8,0x999ec0bb696e840a,
+        0xf148440a256e2c76,0xc00670ea43ca250d,
+        0x96cd2a865764dbca,0x380406926a5e5728,
+        0xbc807527ed3e12bc,0xc605083704f5ecf2,
+        0xeba09271e88d976b,0xf7864a44c633682e,
+        0x93445b8731587ea3,0x7ab3ee6afbe0211d,
+        0xb8157268fdae9e4c,0x5960ea05bad82964,
+        0xe61acf033d1a45df,0x6fb92487298e33bd,
+        0x8fd0c16206306bab,0xa5d3b6d479f8e056,
+        0xb3c4f1ba87bc8696,0x8f48a4899877186c,
+        0xe0b62e2929aba83c,0x331acdabfe94de87,
+        0x8c71dcd9ba0b4925,0x9ff0c08b7f1d0b14,
+        0xaf8e5410288e1b6f,0x7ecf0ae5ee44dd9,
+        0xdb71e91432b1a24a,0xc9e82cd9f69d6150,
+        0x892731ac9faf056e,0xbe311c083a225cd2,
+        0xab70fe17c79ac6ca,0x6dbd630a48aaf406,
+        0xd64d3d9db981787d,0x92cbbccdad5b108,
+        0x85f0468293f0eb4e,0x25bbf56008c58ea5,
+        0xa76c582338ed2621,0xaf2af2b80af6f24e,
+        0xd1476e2c07286faa,0x1af5af660db4aee1,
+        0x82cca4db847945ca,0x50d98d9fc890ed4d,
+        0xa37fce126597973c,0xe50ff107bab528a0,
+        0xcc5fc196fefd7d0c,0x1e53ed49a96272c8,
+        0xff77b1fcbebcdc4f,0x25e8e89c13bb0f7a,
+        0x9faacf3df73609b1,0x77b191618c54e9ac,
+        0xc795830d75038c1d,0xd59df5b9ef6a2417,
+        0xf97ae3d0d2446f25,0x4b0573286b44ad1d,
+        0x9becce62836ac577,0x4ee367f9430aec32,
+        0xc2e801fb244576d5,0x229c41f793cda73f,
+        0xf3a20279ed56d48a,0x6b43527578c1110f,
+        0x9845418c345644d6,0x830a13896b78aaa9,
+        0xbe5691ef416bd60c,0x23cc986bc656d553,
+        0xedec366b11c6cb8f,0x2cbfbe86b7ec8aa8,
+        0x94b3a202eb1c3f39,0x7bf7d71432f3d6a9,
+        0xb9e08a83a5e34f07,0xdaf5ccd93fb0cc53,
+        0xe858ad248f5c22c9,0xd1b3400f8f9cff68,
+        0x91376c36d99995be,0x23100809b9c21fa1,
+        0xb58547448ffffb2d,0xabd40a0c2832a78a,
+        0xe2e69915b3fff9f9,0x16c90c8f323f516c,
+        0x8dd01fad907ffc3b,0xae3da7d97f6792e3,
+        0xb1442798f49ffb4a,0x99cd11cfdf41779c,
+        0xdd95317f31c7fa1d,0x40405643d711d583,
+        0x8a7d3eef7f1cfc52,0x482835ea666b2572,
+        0xad1c8eab5ee43b66,0xda3243650005eecf,
+        0xd863b256369d4a40,0x90bed43e40076a82,
+        0x873e4f75e2224e68,0x5a7744a6e804a291,
+        0xa90de3535aaae202,0x711515d0a205cb36,
+        0xd3515c2831559a83,0xd5a5b44ca873e03,
+        0x8412d9991ed58091,0xe858790afe9486c2,
+        0xa5178fff668ae0b6,0x626e974dbe39a872,
+        0xce5d73ff402d98e3,0xfb0a3d212dc8128f,
+        0x80fa687f881c7f8e,0x7ce66634bc9d0b99,
+        0xa139029f6a239f72,0x1c1fffc1ebc44e80,
+        0xc987434744ac874e,0xa327ffb266b56220,
+        0xfbe9141915d7a922,0x4bf1ff9f0062baa8,
+        0x9d71ac8fada6c9b5,0x6f773fc3603db4a9,
+        0xc4ce17b399107c22,0xcb550fb4384d21d3,
+        0xf6019da07f549b2b,0x7e2a53a146606a48,
+        0x99c102844f94e0fb,0x2eda7444cbfc426d,
+        0xc0314325637a1939,0xfa911155fefb5308,
+        0xf03d93eebc589f88,0x793555ab7eba27ca,
+        0x96267c7535b763b5,0x4bc1558b2f3458de,
+        0xbbb01b9283253ca2,0x9eb1aaedfb016f16,
+        0xea9c227723ee8bcb,0x465e15a979c1cadc,
+        0x92a1958a7675175f,0xbfacd89ec191ec9,
+        0xb749faed14125d36,0xcef980ec671f667b,
+        0xe51c79a85916f484,0x82b7e12780e7401a,
+        0x8f31cc0937ae58d2,0xd1b2ecb8b0908810,
+        0xb2fe3f0b8599ef07,0x861fa7e6dcb4aa15,
+        0xdfbdcece67006ac9,0x67a791e093e1d49a,
+        0x8bd6a141006042bd,0xe0c8bb2c5c6d24e0,
+        0xaecc49914078536d,0x58fae9f773886e18,
+        0xda7f5bf590966848,0xaf39a475506a899e,
+        0x888f99797a5e012d,0x6d8406c952429603,
+        0xaab37fd7d8f58178,0xc8e5087ba6d33b83,
+        0xd5605fcdcf32e1d6,0xfb1e4a9a90880a64,
+        0x855c3be0a17fcd26,0x5cf2eea09a55067f,
+        0xa6b34ad8c9dfc06f,0xf42faa48c0ea481e,
+        0xd0601d8efc57b08b,0xf13b94daf124da26,
+        0x823c12795db6ce57,0x76c53d08d6b70858,
+        0xa2cb1717b52481ed,0x54768c4b0c64ca6e,
+        0xcb7ddcdda26da268,0xa9942f5dcf7dfd09,
+        0xfe5d54150b090b02,0xd3f93b35435d7c4c,
+        0x9efa548d26e5a6e1,0xc47bc5014a1a6daf,
+        0xc6b8e9b0709f109a,0x359ab6419ca1091b,
+        0xf867241c8cc6d4c0,0xc30163d203c94b62,
+        0x9b407691d7fc44f8,0x79e0de63425dcf1d,
+        0xc21094364dfb5636,0x985915fc12f542e4,
+        0xf294b943e17a2bc4,0x3e6f5b7b17b2939d,
+        0x979cf3ca6cec5b5a,0xa705992ceecf9c42,
+        0xbd8430bd08277231,0x50c6ff782a838353,
+        0xece53cec4a314ebd,0xa4f8bf5635246428,
+        0x940f4613ae5ed136,0x871b7795e136be99,
+        0xb913179899f68584,0x28e2557b59846e3f,
+        0xe757dd7ec07426e5,0x331aeada2fe589cf,
+        0x9096ea6f3848984f,0x3ff0d2c85def7621,
+        0xb4bca50b065abe63,0xfed077a756b53a9,
+        0xe1ebce4dc7f16dfb,0xd3e8495912c62894,
+        0x8d3360f09cf6e4bd,0x64712dd7abbbd95c,
+        0xb080392cc4349dec,0xbd8d794d96aacfb3,
+        0xdca04777f541c567,0xecf0d7a0fc5583a0,
+        0x89e42caaf9491b60,0xf41686c49db57244,
+        0xac5d37d5b79b6239,0x311c2875c522ced5,
+        0xd77485cb25823ac7,0x7d633293366b828b,
+        0x86a8d39ef77164bc,0xae5dff9c02033197,
+        0xa8530886b54dbdeb,0xd9f57f830283fdfc,
+        0xd267caa862a12d66,0xd072df63c324fd7b,
+        0x8380dea93da4bc60,0x4247cb9e59f71e6d,
+        0xa46116538d0deb78,0x52d9be85f074e608,
+        0xcd795be870516656,0x67902e276c921f8b,
+        0x806bd9714632dff6,0xba1cd8a3db53b6,
+        0xa086cfcd97bf97f3,0x80e8a40eccd228a4,
+        0xc8a883c0fdaf7df0,0x6122cd128006b2cd,
+        0xfad2a4b13d1b5d6c,0x796b805720085f81,
+        0x9cc3a6eec6311a63,0xcbe3303674053bb0,
+        0xc3f490aa77bd60fc,0xbedbfc4411068a9c,
+        0xf4f1b4d515acb93b,0xee92fb5515482d44,
+        0x991711052d8bf3c5,0x751bdd152d4d1c4a,
+        0xbf5cd54678eef0b6,0xd262d45a78a0635d,
+        0xef340a98172aace4,0x86fb897116c87c34,
+        0x9580869f0e7aac0e,0xd45d35e6ae3d4da0,
+        0xbae0a846d2195712,0x8974836059cca109,
+        0xe998d258869facd7,0x2bd1a438703fc94b,
+        0x91ff83775423cc06,0x7b6306a34627ddcf,
+        0xb67f6455292cbf08,0x1a3bc84c17b1d542,
+        0xe41f3d6a7377eeca,0x20caba5f1d9e4a93,
+        0x8e938662882af53e,0x547eb47b7282ee9c,
+        0xb23867fb2a35b28d,0xe99e619a4f23aa43,
+        0xdec681f9f4c31f31,0x6405fa00e2ec94d4,
+        0x8b3c113c38f9f37e,0xde83bc408dd3dd04,
+        0xae0b158b4738705e,0x9624ab50b148d445,
+        0xd98ddaee19068c76,0x3badd624dd9b0957,
+        0x87f8a8d4cfa417c9,0xe54ca5d70a80e5d6,
+        0xa9f6d30a038d1dbc,0x5e9fcf4ccd211f4c,
+        0xd47487cc8470652b,0x7647c3200069671f,
+        0x84c8d4dfd2c63f3b,0x29ecd9f40041e073,
+        0xa5fb0a17c777cf09,0xf468107100525890,
+        0xcf79cc9db955c2cc,0x7182148d4066eeb4,
+        0x81ac1fe293d599bf,0xc6f14cd848405530,
+        0xa21727db38cb002f,0xb8ada00e5a506a7c,
+        0xca9cf1d206fdc03b,0xa6d90811f0e4851c,
+        0xfd442e4688bd304a,0x908f4a166d1da663,
+        0x9e4a9cec15763e2e,0x9a598e4e043287fe,
+        0xc5dd44271ad3cdba,0x40eff1e1853f29fd,
+        0xf7549530e188c128,0xd12bee59e68ef47c,
+        0x9a94dd3e8cf578b9,0x82bb74f8301958ce,
+        0xc13a148e3032d6e7,0xe36a52363c1faf01,
+        0xf18899b1bc3f8ca1,0xdc44e6c3cb279ac1,
+        0x96f5600f15a7b7e5,0x29ab103a5ef8c0b9,
+        0xbcb2b812db11a5de,0x7415d448f6b6f0e7,
+        0xebdf661791d60f56,0x111b495b3464ad21,
+        0x936b9fcebb25c995,0xcab10dd900beec34,
+        0xb84687c269ef3bfb,0x3d5d514f40eea742,
+        0xe65829b3046b0afa,0xcb4a5a3112a5112,
+        0x8ff71a0fe2c2e6dc,0x47f0e785eaba72ab,
+        0xb3f4e093db73a093,0x59ed216765690f56,
+        0xe0f218b8d25088b8,0x306869c13ec3532c,
+        0x8c974f7383725573,0x1e414218c73a13fb,
+        0xafbd2350644eeacf,0xe5d1929ef90898fa,
+        0xdbac6c247d62a583,0xdf45f746b74abf39,
+        0x894bc396ce5da772,0x6b8bba8c328eb783,
+        0xab9eb47c81f5114f,0x66ea92f3f326564,
+        0xd686619ba27255a2,0xc80a537b0efefebd,
+        0x8613fd0145877585,0xbd06742ce95f5f36,
+        0xa798fc4196e952e7,0x2c48113823b73704,
+        0xd17f3b51fca3a7a0,0xf75a15862ca504c5,
+        0x82ef85133de648c4,0x9a984d73dbe722fb,
+        0xa3ab66580d5fdaf5,0xc13e60d0d2e0ebba,
+        0xcc963fee10b7d1b3,0x318df905079926a8,
+        0xffbbcfe994e5c61f,0xfdf17746497f7052,
+        0x9fd561f1fd0f9bd3,0xfeb6ea8bedefa633,
+        0xc7caba6e7c5382c8,0xfe64a52ee96b8fc0,
+        0xf9bd690a1b68637b,0x3dfdce7aa3c673b0,
+        0x9c1661a651213e2d,0x6bea10ca65c084e,
+        0xc31bfa0fe5698db8,0x486e494fcff30a62,
+        0xf3e2f893dec3f126,0x5a89dba3c3efccfa,
+        0x986ddb5c6b3a76b7,0xf89629465a75e01c,
+        0xbe89523386091465,0xf6bbb397f1135823,
+        0xee2ba6c0678b597f,0x746aa07ded582e2c,
+        0x94db483840b717ef,0xa8c2a44eb4571cdc,
+        0xba121a4650e4ddeb,0x92f34d62616ce413,
+        0xe896a0d7e51e1566,0x77b020baf9c81d17,
+        0x915e2486ef32cd60,0xace1474dc1d122e,
+        0xb5b5ada8aaff80b8,0xd819992132456ba,
+        0xe3231912d5bf60e6,0x10e1fff697ed6c69,
+        0x8df5efabc5979c8f,0xca8d3ffa1ef463c1,
+        0xb1736b96b6fd83b3,0xbd308ff8a6b17cb2,
+        0xddd0467c64bce4a0,0xac7cb3f6d05ddbde,
+        0x8aa22c0dbef60ee4,0x6bcdf07a423aa96b,
+        0xad4ab7112eb3929d,0x86c16c98d2c953c6,
+        0xd89d64d57a607744,0xe871c7bf077ba8b7,
+        0x87625f056c7c4a8b,0x11471cd764ad4972,
+        0xa93af6c6c79b5d2d,0xd598e40d3dd89bcf,
+        0xd389b47879823479,0x4aff1d108d4ec2c3,
+        0x843610cb4bf160cb,0xcedf722a585139ba,
+        0xa54394fe1eedb8fe,0xc2974eb4ee658828,
+        0xce947a3da6a9273e,0x733d226229feea32,
+        0x811ccc668829b887,0x806357d5a3f525f,
+        0xa163ff802a3426a8,0xca07c2dcb0cf26f7,
+        0xc9bcff6034c13052,0xfc89b393dd02f0b5,
+        0xfc2c3f3841f17c67,0xbbac2078d443ace2,
+        0x9d9ba7832936edc0,0xd54b944b84aa4c0d,
+        0xc5029163f384a931,0xa9e795e65d4df11,
+        0xf64335bcf065d37d,0x4d4617b5ff4a16d5,
+        0x99ea0196163fa42e,0x504bced1bf8e4e45,
+        0xc06481fb9bcf8d39,0xe45ec2862f71e1d6,
+        0xf07da27a82c37088,0x5d767327bb4e5a4c,
+        0x964e858c91ba2655,0x3a6a07f8d510f86f,
+        0xbbe226efb628afea,0x890489f70a55368b,
+        0xeadab0aba3b2dbe5,0x2b45ac74ccea842e,
+        0x92c8ae6b464fc96f,0x3b0b8bc90012929d,
+        0xb77ada0617e3bbcb,0x9ce6ebb40173744,
+        0xe55990879ddcaabd,0xcc420a6a101d0515,
+        0x8f57fa54c2a9eab6,0x9fa946824a12232d,
+        0xb32df8e9f3546564,0x47939822dc96abf9,
+        0xdff9772470297ebd,0x59787e2b93bc56f7,
+        0x8bfbea76c619ef36,0x57eb4edb3c55b65a,
+        0xaefae51477a06b03,0xede622920b6b23f1,
+        0xdab99e59958885c4,0xe95fab368e45eced,
+        0x88b402f7fd75539b,0x11dbcb0218ebb414,
+        0xaae103b5fcd2a881,0xd652bdc29f26a119,
+        0xd59944a37c0752a2,0x4be76d3346f0495f,
+        0x857fcae62d8493a5,0x6f70a4400c562ddb,
+        0xa6dfbd9fb8e5b88e,0xcb4ccd500f6bb952,
+        0xd097ad07a71f26b2,0x7e2000a41346a7a7,
+        0x825ecc24c873782f,0x8ed400668c0c28c8,
+        0xa2f67f2dfa90563b,0x728900802f0f32fa,
+        0xcbb41ef979346bca,0x4f2b40a03ad2ffb9,
+        0xfea126b7d78186bc,0xe2f610c84987bfa8,
+        0x9f24b832e6b0f436,0xdd9ca7d2df4d7c9,
+        0xc6ede63fa05d3143,0x91503d1c79720dbb,
+        0xf8a95fcf88747d94,0x75a44c6397ce912a,
+        0x9b69dbe1b548ce7c,0xc986afbe3ee11aba,
+        0xc24452da229b021b,0xfbe85badce996168,
+        0xf2d56790ab41c2a2,0xfae27299423fb9c3,
+        0x97c560ba6b0919a5,0xdccd879fc967d41a,
+        0xbdb6b8e905cb600f,0x5400e987bbc1c920,
+        0xed246723473e3813,0x290123e9aab23b68,
+        0x9436c0760c86e30b,0xf9a0b6720aaf6521,
+        0xb94470938fa89bce,0xf808e40e8d5b3e69,
+        0xe7958cb87392c2c2,0xb60b1d1230b20e04,
+        0x90bd77f3483bb9b9,0xb1c6f22b5e6f48c2,
+        0xb4ecd5f01a4aa828,0x1e38aeb6360b1af3,
+        0xe2280b6c20dd5232,0x25c6da63c38de1b0,
+        0x8d590723948a535f,0x579c487e5a38ad0e,
+        0xb0af48ec79ace837,0x2d835a9df0c6d851,
+        0xdcdb1b2798182244,0xf8e431456cf88e65,
+        0x8a08f0f8bf0f156b,0x1b8e9ecb641b58ff,
+        0xac8b2d36eed2dac5,0xe272467e3d222f3f,
+        0xd7adf884aa879177,0x5b0ed81dcc6abb0f,
+        0x86ccbb52ea94baea,0x98e947129fc2b4e9,
+        0xa87fea27a539e9a5,0x3f2398d747b36224,
+        0xd29fe4b18e88640e,0x8eec7f0d19a03aad,
+        0x83a3eeeef9153e89,0x1953cf68300424ac,
+        0xa48ceaaab75a8e2b,0x5fa8c3423c052dd7,
+        0xcdb02555653131b6,0x3792f412cb06794d,
+        0x808e17555f3ebf11,0xe2bbd88bbee40bd0,
+        0xa0b19d2ab70e6ed6,0x5b6aceaeae9d0ec4,
+        0xc8de047564d20a8b,0xf245825a5a445275,
+        0xfb158592be068d2e,0xeed6e2f0f0d56712,
+        0x9ced737bb6c4183d,0x55464dd69685606b,
+        0xc428d05aa4751e4c,0xaa97e14c3c26b886,
+        0xf53304714d9265df,0xd53dd99f4b3066a8,
+        0x993fe2c6d07b7fab,0xe546a8038efe4029,
+        0xbf8fdb78849a5f96,0xde98520472bdd033,
+        0xef73d256a5c0f77c,0x963e66858f6d4440,
+        0x95a8637627989aad,0xdde7001379a44aa8,
+        0xbb127c53b17ec159,0x5560c018580d5d52,
+        0xe9d71b689dde71af,0xaab8f01e6e10b4a6,
+        0x9226712162ab070d,0xcab3961304ca70e8,
+        0xb6b00d69bb55c8d1,0x3d607b97c5fd0d22,
+        0xe45c10c42a2b3b05,0x8cb89a7db77c506a,
+        0x8eb98a7a9a5b04e3,0x77f3608e92adb242,
+        0xb267ed1940f1c61c,0x55f038b237591ed3,
+        0xdf01e85f912e37a3,0x6b6c46dec52f6688,
+        0x8b61313bbabce2c6,0x2323ac4b3b3da015,
+        0xae397d8aa96c1b77,0xabec975e0a0d081a,
+        0xd9c7dced53c72255,0x96e7bd358c904a21,
+        0x881cea14545c7575,0x7e50d64177da2e54,
+        0xaa242499697392d2,0xdde50bd1d5d0b9e9,
+        0xd4ad2dbfc3d07787,0x955e4ec64b44e864,
+        0x84ec3c97da624ab4,0xbd5af13bef0b113e,
+        0xa6274bbdd0fadd61,0xecb1ad8aeacdd58e,
+        0xcfb11ead453994ba,0x67de18eda5814af2,
+        0x81ceb32c4b43fcf4,0x80eacf948770ced7,
+        0xa2425ff75e14fc31,0xa1258379a94d028d,
+        0xcad2f7f5359a3b3e,0x96ee45813a04330,
+        0xfd87b5f28300ca0d,0x8bca9d6e188853fc,
+        0x9e74d1b791e07e48,0x775ea264cf55347e,
+        0xc612062576589dda,0x95364afe032a81a0,
+        0xf79687aed3eec551,0x3a83ddbd83f52210,
+        0x9abe14cd44753b52,0xc4926a9672793580,
+        0xc16d9a0095928a27,0x75b7053c0f178400,
+        0xf1c90080baf72cb1,0x5324c68b12dd6800,
+        0x971da05074da7bee,0xd3f6fc16ebca8000,
+        0xbce5086492111aea,0x88f4bb1ca6bd0000,
+        0xec1e4a7db69561a5,0x2b31e9e3d0700000,
+        0x9392ee8e921d5d07,0x3aff322e62600000,
+        0xb877aa3236a4b449,0x9befeb9fad487c3,
+        0xe69594bec44de15b,0x4c2ebe687989a9b4,
+        0x901d7cf73ab0acd9,0xf9d37014bf60a11,
+        0xb424dc35095cd80f,0x538484c19ef38c95,
+        0xe12e13424bb40e13,0x2865a5f206b06fba,
+        0x8cbccc096f5088cb,0xf93f87b7442e45d4,
+        0xafebff0bcb24aafe,0xf78f69a51539d749,
+        0xdbe6fecebdedd5be,0xb573440e5a884d1c,
+        0x89705f4136b4a597,0x31680a88f8953031,
+        0xabcc77118461cefc,0xfdc20d2b36ba7c3e,
+        0xd6bf94d5e57a42bc,0x3d32907604691b4d,
+        0x8637bd05af6c69b5,0xa63f9a49c2c1b110,
+        0xa7c5ac471b478423,0xfcf80dc33721d54,
+        0xd1b71758e219652b,0xd3c36113404ea4a9,
+        0x83126e978d4fdf3b,0x645a1cac083126ea,
+        0xa3d70a3d70a3d70a,0x3d70a3d70a3d70a4,
+        0xcccccccccccccccc,0xcccccccccccccccd,
+        0x8000000000000000,0x0,
+        0xa000000000000000,0x0,
+        0xc800000000000000,0x0,
+        0xfa00000000000000,0x0,
+        0x9c40000000000000,0x0,
+        0xc350000000000000,0x0,
+        0xf424000000000000,0x0,
+        0x9896800000000000,0x0,
+        0xbebc200000000000,0x0,
+        0xee6b280000000000,0x0,
+        0x9502f90000000000,0x0,
+        0xba43b74000000000,0x0,
+        0xe8d4a51000000000,0x0,
+        0x9184e72a00000000,0x0,
+        0xb5e620f480000000,0x0,
+        0xe35fa931a0000000,0x0,
+        0x8e1bc9bf04000000,0x0,
+        0xb1a2bc2ec5000000,0x0,
+        0xde0b6b3a76400000,0x0,
+        0x8ac7230489e80000,0x0,
+        0xad78ebc5ac620000,0x0,
+        0xd8d726b7177a8000,0x0,
+        0x878678326eac9000,0x0,
+        0xa968163f0a57b400,0x0,
+        0xd3c21bcecceda100,0x0,
+        0x84595161401484a0,0x0,
+        0xa56fa5b99019a5c8,0x0,
+        0xcecb8f27f4200f3a,0x0,
+        0x813f3978f8940984,0x4000000000000000,
+        0xa18f07d736b90be5,0x5000000000000000,
+        0xc9f2c9cd04674ede,0xa400000000000000,
+        0xfc6f7c4045812296,0x4d00000000000000,
+        0x9dc5ada82b70b59d,0xf020000000000000,
+        0xc5371912364ce305,0x6c28000000000000,
+        0xf684df56c3e01bc6,0xc732000000000000,
+        0x9a130b963a6c115c,0x3c7f400000000000,
+        0xc097ce7bc90715b3,0x4b9f100000000000,
+        0xf0bdc21abb48db20,0x1e86d40000000000,
+        0x96769950b50d88f4,0x1314448000000000,
+        0xbc143fa4e250eb31,0x17d955a000000000,
+        0xeb194f8e1ae525fd,0x5dcfab0800000000,
+        0x92efd1b8d0cf37be,0x5aa1cae500000000,
+        0xb7abc627050305ad,0xf14a3d9e40000000,
+        0xe596b7b0c643c719,0x6d9ccd05d0000000,
+        0x8f7e32ce7bea5c6f,0xe4820023a2000000,
+        0xb35dbf821ae4f38b,0xdda2802c8a800000,
+        0xe0352f62a19e306e,0xd50b2037ad200000,
+        0x8c213d9da502de45,0x4526f422cc340000,
+        0xaf298d050e4395d6,0x9670b12b7f410000,
+        0xdaf3f04651d47b4c,0x3c0cdd765f114000,
+        0x88d8762bf324cd0f,0xa5880a69fb6ac800,
+        0xab0e93b6efee0053,0x8eea0d047a457a00,
+        0xd5d238a4abe98068,0x72a4904598d6d880,
+        0x85a36366eb71f041,0x47a6da2b7f864750,
+        0xa70c3c40a64e6c51,0x999090b65f67d924,
+        0xd0cf4b50cfe20765,0xfff4b4e3f741cf6d,
+        0x82818f1281ed449f,0xbff8f10e7a8921a4,
+        0xa321f2d7226895c7,0xaff72d52192b6a0d,
+        0xcbea6f8ceb02bb39,0x9bf4f8a69f764490,
+        0xfee50b7025c36a08,0x2f236d04753d5b4,
+        0x9f4f2726179a2245,0x1d762422c946590,
+        0xc722f0ef9d80aad6,0x424d3ad2b7b97ef5,
+        0xf8ebad2b84e0d58b,0xd2e0898765a7deb2,
+        0x9b934c3b330c8577,0x63cc55f49f88eb2f,
+        0xc2781f49ffcfa6d5,0x3cbf6b71c76b25fb,
+        0xf316271c7fc3908a,0x8bef464e3945ef7a,
+        0x97edd871cfda3a56,0x97758bf0e3cbb5ac,
+        0xbde94e8e43d0c8ec,0x3d52eeed1cbea317,
+        0xed63a231d4c4fb27,0x4ca7aaa863ee4bdd,
+        0x945e455f24fb1cf8,0x8fe8caa93e74ef6a,
+        0xb975d6b6ee39e436,0xb3e2fd538e122b44,
+        0xe7d34c64a9c85d44,0x60dbbca87196b616,
+        0x90e40fbeea1d3a4a,0xbc8955e946fe31cd,
+        0xb51d13aea4a488dd,0x6babab6398bdbe41,
+        0xe264589a4dcdab14,0xc696963c7eed2dd1,
+        0x8d7eb76070a08aec,0xfc1e1de5cf543ca2,
+        0xb0de65388cc8ada8,0x3b25a55f43294bcb,
+        0xdd15fe86affad912,0x49ef0eb713f39ebe,
+        0x8a2dbf142dfcc7ab,0x6e3569326c784337,
+        0xacb92ed9397bf996,0x49c2c37f07965404,
+        0xd7e77a8f87daf7fb,0xdc33745ec97be906,
+        0x86f0ac99b4e8dafd,0x69a028bb3ded71a3,
+        0xa8acd7c0222311bc,0xc40832ea0d68ce0c,
+        0xd2d80db02aabd62b,0xf50a3fa490c30190,
+        0x83c7088e1aab65db,0x792667c6da79e0fa,
+        0xa4b8cab1a1563f52,0x577001b891185938,
+        0xcde6fd5e09abcf26,0xed4c0226b55e6f86,
+        0x80b05e5ac60b6178,0x544f8158315b05b4,
+        0xa0dc75f1778e39d6,0x696361ae3db1c721,
+        0xc913936dd571c84c,0x3bc3a19cd1e38e9,
+        0xfb5878494ace3a5f,0x4ab48a04065c723,
+        0x9d174b2dcec0e47b,0x62eb0d64283f9c76,
+        0xc45d1df942711d9a,0x3ba5d0bd324f8394,
+        0xf5746577930d6500,0xca8f44ec7ee36479,
+        0x9968bf6abbe85f20,0x7e998b13cf4e1ecb,
+        0xbfc2ef456ae276e8,0x9e3fedd8c321a67e,
+        0xefb3ab16c59b14a2,0xc5cfe94ef3ea101e,
+        0x95d04aee3b80ece5,0xbba1f1d158724a12,
+        0xbb445da9ca61281f,0x2a8a6e45ae8edc97,
+        0xea1575143cf97226,0xf52d09d71a3293bd,
+        0x924d692ca61be758,0x593c2626705f9c56,
+        0xb6e0c377cfa2e12e,0x6f8b2fb00c77836c,
+        0xe498f455c38b997a,0xb6dfb9c0f956447,
+        0x8edf98b59a373fec,0x4724bd4189bd5eac,
+        0xb2977ee300c50fe7,0x58edec91ec2cb657,
+        0xdf3d5e9bc0f653e1,0x2f2967b66737e3ed,
+        0x8b865b215899f46c,0xbd79e0d20082ee74,
+        0xae67f1e9aec07187,0xecd8590680a3aa11,
+        0xda01ee641a708de9,0xe80e6f4820cc9495,
+        0x884134fe908658b2,0x3109058d147fdcdd,
+        0xaa51823e34a7eede,0xbd4b46f0599fd415,
+        0xd4e5e2cdc1d1ea96,0x6c9e18ac7007c91a,
+        0x850fadc09923329e,0x3e2cf6bc604ddb0,
+        0xa6539930bf6bff45,0x84db8346b786151c,
+        0xcfe87f7cef46ff16,0xe612641865679a63,
+        0x81f14fae158c5f6e,0x4fcb7e8f3f60c07e,
+        0xa26da3999aef7749,0xe3be5e330f38f09d,
+        0xcb090c8001ab551c,0x5cadf5bfd3072cc5,
+        0xfdcb4fa002162a63,0x73d9732fc7c8f7f6,
+        0x9e9f11c4014dda7e,0x2867e7fddcdd9afa,
+        0xc646d63501a1511d,0xb281e1fd541501b8,
+        0xf7d88bc24209a565,0x1f225a7ca91a4226,
+        0x9ae757596946075f,0x3375788de9b06958,
+        0xc1a12d2fc3978937,0x52d6b1641c83ae,
+        0xf209787bb47d6b84,0xc0678c5dbd23a49a,
+        0x9745eb4d50ce6332,0xf840b7ba963646e0,
+        0xbd176620a501fbff,0xb650e5a93bc3d898,
+        0xec5d3fa8ce427aff,0xa3e51f138ab4cebe,
+        0x93ba47c980e98cdf,0xc66f336c36b10137,
+        0xb8a8d9bbe123f017,0xb80b0047445d4184,
+        0xe6d3102ad96cec1d,0xa60dc059157491e5,
+        0x9043ea1ac7e41392,0x87c89837ad68db2f,
+        0xb454e4a179dd1877,0x29babe4598c311fb,
+        0xe16a1dc9d8545e94,0xf4296dd6fef3d67a,
+        0x8ce2529e2734bb1d,0x1899e4a65f58660c,
+        0xb01ae745b101e9e4,0x5ec05dcff72e7f8f,
+        0xdc21a1171d42645d,0x76707543f4fa1f73,
+        0x899504ae72497eba,0x6a06494a791c53a8,
+        0xabfa45da0edbde69,0x487db9d17636892,
+        0xd6f8d7509292d603,0x45a9d2845d3c42b6,
+        0x865b86925b9bc5c2,0xb8a2392ba45a9b2,
+        0xa7f26836f282b732,0x8e6cac7768d7141e,
+        0xd1ef0244af2364ff,0x3207d795430cd926,
+        0x8335616aed761f1f,0x7f44e6bd49e807b8,
+        0xa402b9c5a8d3a6e7,0x5f16206c9c6209a6,
+        0xcd036837130890a1,0x36dba887c37a8c0f,
+        0x802221226be55a64,0xc2494954da2c9789,
+        0xa02aa96b06deb0fd,0xf2db9baa10b7bd6c,
+        0xc83553c5c8965d3d,0x6f92829494e5acc7,
+        0xfa42a8b73abbf48c,0xcb772339ba1f17f9,
+        0x9c69a97284b578d7,0xff2a760414536efb,
+        0xc38413cf25e2d70d,0xfef5138519684aba,
+        0xf46518c2ef5b8cd1,0x7eb258665fc25d69,
+        0x98bf2f79d5993802,0xef2f773ffbd97a61,
+        0xbeeefb584aff8603,0xaafb550ffacfd8fa,
+        0xeeaaba2e5dbf6784,0x95ba2a53f983cf38,
+        0x952ab45cfa97a0b2,0xdd945a747bf26183,
+        0xba756174393d88df,0x94f971119aeef9e4,
+        0xe912b9d1478ceb17,0x7a37cd5601aab85d,
+        0x91abb422ccb812ee,0xac62e055c10ab33a,
+        0xb616a12b7fe617aa,0x577b986b314d6009,
+        0xe39c49765fdf9d94,0xed5a7e85fda0b80b,
+        0x8e41ade9fbebc27d,0x14588f13be847307,
+        0xb1d219647ae6b31c,0x596eb2d8ae258fc8,
+        0xde469fbd99a05fe3,0x6fca5f8ed9aef3bb,
+        0x8aec23d680043bee,0x25de7bb9480d5854,
+        0xada72ccc20054ae9,0xaf561aa79a10ae6a,
+        0xd910f7ff28069da4,0x1b2ba1518094da04,
+        0x87aa9aff79042286,0x90fb44d2f05d0842,
+        0xa99541bf57452b28,0x353a1607ac744a53,
+        0xd3fa922f2d1675f2,0x42889b8997915ce8,
+        0x847c9b5d7c2e09b7,0x69956135febada11,
+        0xa59bc234db398c25,0x43fab9837e699095,
+        0xcf02b2c21207ef2e,0x94f967e45e03f4bb,
+        0x8161afb94b44f57d,0x1d1be0eebac278f5,
+        0xa1ba1ba79e1632dc,0x6462d92a69731732,
+        0xca28a291859bbf93,0x7d7b8f7503cfdcfe,
+        0xfcb2cb35e702af78,0x5cda735244c3d43e,
+        0x9defbf01b061adab,0x3a0888136afa64a7,
+        0xc56baec21c7a1916,0x88aaa1845b8fdd0,
+        0xf6c69a72a3989f5b,0x8aad549e57273d45,
+        0x9a3c2087a63f6399,0x36ac54e2f678864b,
+        0xc0cb28a98fcf3c7f,0x84576a1bb416a7dd,
+        0xf0fdf2d3f3c30b9f,0x656d44a2a11c51d5,
+        0x969eb7c47859e743,0x9f644ae5a4b1b325,
+        0xbc4665b596706114,0x873d5d9f0dde1fee,
+        0xeb57ff22fc0c7959,0xa90cb506d155a7ea,
+        0x9316ff75dd87cbd8,0x9a7f12442d588f2,
+        0xb7dcbf5354e9bece,0xc11ed6d538aeb2f,
+        0xe5d3ef282a242e81,0x8f1668c8a86da5fa,
+        0x8fa475791a569d10,0xf96e017d694487bc,
+        0xb38d92d760ec4455,0x37c981dcc395a9ac,
+        0xe070f78d3927556a,0x85bbe253f47b1417,
+        0x8c469ab843b89562,0x93956d7478ccec8e,
+        0xaf58416654a6babb,0x387ac8d1970027b2,
+        0xdb2e51bfe9d0696a,0x6997b05fcc0319e,
+        0x88fcf317f22241e2,0x441fece3bdf81f03,
+        0xab3c2fddeeaad25a,0xd527e81cad7626c3,
+        0xd60b3bd56a5586f1,0x8a71e223d8d3b074,
+        0x85c7056562757456,0xf6872d5667844e49,
+        0xa738c6bebb12d16c,0xb428f8ac016561db,
+        0xd106f86e69d785c7,0xe13336d701beba52,
+        0x82a45b450226b39c,0xecc0024661173473,
+        0xa34d721642b06084,0x27f002d7f95d0190,
+        0xcc20ce9bd35c78a5,0x31ec038df7b441f4,
+        0xff290242c83396ce,0x7e67047175a15271,
+        0x9f79a169bd203e41,0xf0062c6e984d386,
+        0xc75809c42c684dd1,0x52c07b78a3e60868,
+        0xf92e0c3537826145,0xa7709a56ccdf8a82,
+        0x9bbcc7a142b17ccb,0x88a66076400bb691,
+        0xc2abf989935ddbfe,0x6acff893d00ea435,
+        0xf356f7ebf83552fe,0x583f6b8c4124d43,
+        0x98165af37b2153de,0xc3727a337a8b704a,
+        0xbe1bf1b059e9a8d6,0x744f18c0592e4c5c,
+        0xeda2ee1c7064130c,0x1162def06f79df73,
+        0x9485d4d1c63e8be7,0x8addcb5645ac2ba8,
+        0xb9a74a0637ce2ee1,0x6d953e2bd7173692,
+        0xe8111c87c5c1ba99,0xc8fa8db6ccdd0437,
+        0x910ab1d4db9914a0,0x1d9c9892400a22a2,
+        0xb54d5e4a127f59c8,0x2503beb6d00cab4b,
+        0xe2a0b5dc971f303a,0x2e44ae64840fd61d,
+        0x8da471a9de737e24,0x5ceaecfed289e5d2,
+        0xb10d8e1456105dad,0x7425a83e872c5f47,
+        0xdd50f1996b947518,0xd12f124e28f77719,
+        0x8a5296ffe33cc92f,0x82bd6b70d99aaa6f,
+        0xace73cbfdc0bfb7b,0x636cc64d1001550b,
+        0xd8210befd30efa5a,0x3c47f7e05401aa4e,
+        0x8714a775e3e95c78,0x65acfaec34810a71,
+        0xa8d9d1535ce3b396,0x7f1839a741a14d0d,
+        0xd31045a8341ca07c,0x1ede48111209a050,
+        0x83ea2b892091e44d,0x934aed0aab460432,
+        0xa4e4b66b68b65d60,0xf81da84d5617853f,
+        0xce1de40642e3f4b9,0x36251260ab9d668e,
+        0x80d2ae83e9ce78f3,0xc1d72b7c6b426019,
+        0xa1075a24e4421730,0xb24cf65b8612f81f,
+        0xc94930ae1d529cfc,0xdee033f26797b627,
+        0xfb9b7cd9a4a7443c,0x169840ef017da3b1,
+        0x9d412e0806e88aa5,0x8e1f289560ee864e,
+        0xc491798a08a2ad4e,0xf1a6f2bab92a27e2,
+        0xf5b5d7ec8acb58a2,0xae10af696774b1db,
+        0x9991a6f3d6bf1765,0xacca6da1e0a8ef29,
+        0xbff610b0cc6edd3f,0x17fd090a58d32af3,
+        0xeff394dcff8a948e,0xddfc4b4cef07f5b0,
+        0x95f83d0a1fb69cd9,0x4abdaf101564f98e,
+        0xbb764c4ca7a4440f,0x9d6d1ad41abe37f1,
+        0xea53df5fd18d5513,0x84c86189216dc5ed,
+        0x92746b9be2f8552c,0x32fd3cf5b4e49bb4,
+        0xb7118682dbb66a77,0x3fbc8c33221dc2a1,
+        0xe4d5e82392a40515,0xfabaf3feaa5334a,
+        0x8f05b1163ba6832d,0x29cb4d87f2a7400e,
+        0xb2c71d5bca9023f8,0x743e20e9ef511012,
+        0xdf78e4b2bd342cf6,0x914da9246b255416,
+        0x8bab8eefb6409c1a,0x1ad089b6c2f7548e,
+        0xae9672aba3d0c320,0xa184ac2473b529b1,
+        0xda3c0f568cc4f3e8,0xc9e5d72d90a2741e,
+        0x8865899617fb1871,0x7e2fa67c7a658892,
+        0xaa7eebfb9df9de8d,0xddbb901b98feeab7,
+        0xd51ea6fa85785631,0x552a74227f3ea565,
+        0x8533285c936b35de,0xd53a88958f87275f,
+        0xa67ff273b8460356,0x8a892abaf368f137,
+        0xd01fef10a657842c,0x2d2b7569b0432d85,
+        0x8213f56a67f6b29b,0x9c3b29620e29fc73,
+        0xa298f2c501f45f42,0x8349f3ba91b47b8f,
+        0xcb3f2f7642717713,0x241c70a936219a73,
+        0xfe0efb53d30dd4d7,0xed238cd383aa0110,
+        0x9ec95d1463e8a506,0xf4363804324a40aa,
+        0xc67bb4597ce2ce48,0xb143c6053edcd0d5,
+        0xf81aa16fdc1b81da,0xdd94b7868e94050a,
+        0x9b10a4e5e9913128,0xca7cf2b4191c8326,
+        0xc1d4ce1f63f57d72,0xfd1c2f611f63a3f0,
+        0xf24a01a73cf2dccf,0xbc633b39673c8cec,
+        0x976e41088617ca01,0xd5be0503e085d813,
+        0xbd49d14aa79dbc82,0x4b2d8644d8a74e18,
+        0xec9c459d51852ba2,0xddf8e7d60ed1219e,
+        0x93e1ab8252f33b45,0xcabb90e5c942b503,
+        0xb8da1662e7b00a17,0x3d6a751f3b936243,
+        0xe7109bfba19c0c9d,0xcc512670a783ad4,
+        0x906a617d450187e2,0x27fb2b80668b24c5,
+        0xb484f9dc9641e9da,0xb1f9f660802dedf6,
+        0xe1a63853bbd26451,0x5e7873f8a0396973,
+        0x8d07e33455637eb2,0xdb0b487b6423e1e8,
+        0xb049dc016abc5e5f,0x91ce1a9a3d2cda62,
+        0xdc5c5301c56b75f7,0x7641a140cc7810fb,
+        0x89b9b3e11b6329ba,0xa9e904c87fcb0a9d,
+        0xac2820d9623bf429,0x546345fa9fbdcd44,
+        0xd732290fbacaf133,0xa97c177947ad4095,
+        0x867f59a9d4bed6c0,0x49ed8eabcccc485d,
+        0xa81f301449ee8c70,0x5c68f256bfff5a74,
+        0xd226fc195c6a2f8c,0x73832eec6fff3111,
+        0x83585d8fd9c25db7,0xc831fd53c5ff7eab,
+        0xa42e74f3d032f525,0xba3e7ca8b77f5e55,
+        0xcd3a1230c43fb26f,0x28ce1bd2e55f35eb,
+        0x80444b5e7aa7cf85,0x7980d163cf5b81b3,
+        0xa0555e361951c366,0xd7e105bcc332621f,
+        0xc86ab5c39fa63440,0x8dd9472bf3fefaa7,
+        0xfa856334878fc150,0xb14f98f6f0feb951,
+        0x9c935e00d4b9d8d2,0x6ed1bf9a569f33d3,
+        0xc3b8358109e84f07,0xa862f80ec4700c8,
+        0xf4a642e14c6262c8,0xcd27bb612758c0fa,
+        0x98e7e9cccfbd7dbd,0x8038d51cb897789c,
+        0xbf21e44003acdd2c,0xe0470a63e6bd56c3,
+        0xeeea5d5004981478,0x1858ccfce06cac74,
+        0x95527a5202df0ccb,0xf37801e0c43ebc8,
+        0xbaa718e68396cffd,0xd30560258f54e6ba,
+        0xe950df20247c83fd,0x47c6b82ef32a2069,
+        0x91d28b7416cdd27e,0x4cdc331d57fa5441,
+        0xb6472e511c81471d,0xe0133fe4adf8e952,
+        0xe3d8f9e563a198e5,0x58180fddd97723a6,
+        0x8e679c2f5e44ff8f,0x570f09eaa7ea7648,};
+};
+#endif // SIMDJSON_STATIC_REFLECTION
+
+extern SIMDJSON_DLLIMPORTEXPORT const uint64_t power_of_five_128[651*2];
+} // namespace internal
+} // namespace simdjson
+
+#endif // SIMDJSON_INTERNAL_NUMBERPARSING_TABLES_H
+/* end file simdjson/internal/numberparsing_tables.h */
+
+// Precomputed powers of ten from 10^0 to 10^22. These
+// can be represented exactly using the double type.
+SIMDJSON_DLLIMPORTEXPORT const double simdjson::internal::power_of_ten[] = {
+    1e0,  1e1,  1e2,  1e3,  1e4,  1e5,  1e6,  1e7,  1e8,  1e9,  1e10, 1e11,
+    1e12, 1e13, 1e14, 1e15, 1e16, 1e17, 1e18, 1e19, 1e20, 1e21, 1e22};
+
+/**
+ * When mapping numbers from decimal to binary,
+ * we go from w * 10^q to m * 2^p but we have
+ * 10^q = 5^q * 2^q, so effectively
+ * we are trying to match
+ * w * 2^q * 5^q to m * 2^p. Thus the powers of two
+ * are not a concern since they can be represented
+ * exactly using the binary notation, only the powers of five
+ * affect the binary significand.
+ */
+// The truncated powers of five from 5^-342 all the way to 5^308
+// The mantissa is truncated to 128 bits, and
+// never rounded up. Uses about 10KB.
+SIMDJSON_DLLIMPORTEXPORT const uint64_t simdjson::internal::power_of_five_128[]= {
+        0xeef453d6923bd65a,0x113faa2906a13b3f,
+        0x9558b4661b6565f8,0x4ac7ca59a424c507,
+        0xbaaee17fa23ebf76,0x5d79bcf00d2df649,
+        0xe95a99df8ace6f53,0xf4d82c2c107973dc,
+        0x91d8a02bb6c10594,0x79071b9b8a4be869,
+        0xb64ec836a47146f9,0x9748e2826cdee284,
+        0xe3e27a444d8d98b7,0xfd1b1b2308169b25,
+        0x8e6d8c6ab0787f72,0xfe30f0f5e50e20f7,
+        0xb208ef855c969f4f,0xbdbd2d335e51a935,
+        0xde8b2b66b3bc4723,0xad2c788035e61382,
+        0x8b16fb203055ac76,0x4c3bcb5021afcc31,
+        0xaddcb9e83c6b1793,0xdf4abe242a1bbf3d,
+        0xd953e8624b85dd78,0xd71d6dad34a2af0d,
+        0x87d4713d6f33aa6b,0x8672648c40e5ad68,
+        0xa9c98d8ccb009506,0x680efdaf511f18c2,
+        0xd43bf0effdc0ba48,0x212bd1b2566def2,
+        0x84a57695fe98746d,0x14bb630f7604b57,
+        0xa5ced43b7e3e9188,0x419ea3bd35385e2d,
+        0xcf42894a5dce35ea,0x52064cac828675b9,
+        0x818995ce7aa0e1b2,0x7343efebd1940993,
+        0xa1ebfb4219491a1f,0x1014ebe6c5f90bf8,
+        0xca66fa129f9b60a6,0xd41a26e077774ef6,
+        0xfd00b897478238d0,0x8920b098955522b4,
+        0x9e20735e8cb16382,0x55b46e5f5d5535b0,
+        0xc5a890362fddbc62,0xeb2189f734aa831d,
+        0xf712b443bbd52b7b,0xa5e9ec7501d523e4,
+        0x9a6bb0aa55653b2d,0x47b233c92125366e,
+        0xc1069cd4eabe89f8,0x999ec0bb696e840a,
+        0xf148440a256e2c76,0xc00670ea43ca250d,
+        0x96cd2a865764dbca,0x380406926a5e5728,
+        0xbc807527ed3e12bc,0xc605083704f5ecf2,
+        0xeba09271e88d976b,0xf7864a44c633682e,
+        0x93445b8731587ea3,0x7ab3ee6afbe0211d,
+        0xb8157268fdae9e4c,0x5960ea05bad82964,
+        0xe61acf033d1a45df,0x6fb92487298e33bd,
+        0x8fd0c16206306bab,0xa5d3b6d479f8e056,
+        0xb3c4f1ba87bc8696,0x8f48a4899877186c,
+        0xe0b62e2929aba83c,0x331acdabfe94de87,
+        0x8c71dcd9ba0b4925,0x9ff0c08b7f1d0b14,
+        0xaf8e5410288e1b6f,0x7ecf0ae5ee44dd9,
+        0xdb71e91432b1a24a,0xc9e82cd9f69d6150,
+        0x892731ac9faf056e,0xbe311c083a225cd2,
+        0xab70fe17c79ac6ca,0x6dbd630a48aaf406,
+        0xd64d3d9db981787d,0x92cbbccdad5b108,
+        0x85f0468293f0eb4e,0x25bbf56008c58ea5,
+        0xa76c582338ed2621,0xaf2af2b80af6f24e,
+        0xd1476e2c07286faa,0x1af5af660db4aee1,
+        0x82cca4db847945ca,0x50d98d9fc890ed4d,
+        0xa37fce126597973c,0xe50ff107bab528a0,
+        0xcc5fc196fefd7d0c,0x1e53ed49a96272c8,
+        0xff77b1fcbebcdc4f,0x25e8e89c13bb0f7a,
+        0x9faacf3df73609b1,0x77b191618c54e9ac,
+        0xc795830d75038c1d,0xd59df5b9ef6a2417,
+        0xf97ae3d0d2446f25,0x4b0573286b44ad1d,
+        0x9becce62836ac577,0x4ee367f9430aec32,
+        0xc2e801fb244576d5,0x229c41f793cda73f,
+        0xf3a20279ed56d48a,0x6b43527578c1110f,
+        0x9845418c345644d6,0x830a13896b78aaa9,
+        0xbe5691ef416bd60c,0x23cc986bc656d553,
+        0xedec366b11c6cb8f,0x2cbfbe86b7ec8aa8,
+        0x94b3a202eb1c3f39,0x7bf7d71432f3d6a9,
+        0xb9e08a83a5e34f07,0xdaf5ccd93fb0cc53,
+        0xe858ad248f5c22c9,0xd1b3400f8f9cff68,
+        0x91376c36d99995be,0x23100809b9c21fa1,
+        0xb58547448ffffb2d,0xabd40a0c2832a78a,
+        0xe2e69915b3fff9f9,0x16c90c8f323f516c,
+        0x8dd01fad907ffc3b,0xae3da7d97f6792e3,
+        0xb1442798f49ffb4a,0x99cd11cfdf41779c,
+        0xdd95317f31c7fa1d,0x40405643d711d583,
+        0x8a7d3eef7f1cfc52,0x482835ea666b2572,
+        0xad1c8eab5ee43b66,0xda3243650005eecf,
+        0xd863b256369d4a40,0x90bed43e40076a82,
+        0x873e4f75e2224e68,0x5a7744a6e804a291,
+        0xa90de3535aaae202,0x711515d0a205cb36,
+        0xd3515c2831559a83,0xd5a5b44ca873e03,
+        0x8412d9991ed58091,0xe858790afe9486c2,
+        0xa5178fff668ae0b6,0x626e974dbe39a872,
+        0xce5d73ff402d98e3,0xfb0a3d212dc8128f,
+        0x80fa687f881c7f8e,0x7ce66634bc9d0b99,
+        0xa139029f6a239f72,0x1c1fffc1ebc44e80,
+        0xc987434744ac874e,0xa327ffb266b56220,
+        0xfbe9141915d7a922,0x4bf1ff9f0062baa8,
+        0x9d71ac8fada6c9b5,0x6f773fc3603db4a9,
+        0xc4ce17b399107c22,0xcb550fb4384d21d3,
+        0xf6019da07f549b2b,0x7e2a53a146606a48,
+        0x99c102844f94e0fb,0x2eda7444cbfc426d,
+        0xc0314325637a1939,0xfa911155fefb5308,
+        0xf03d93eebc589f88,0x793555ab7eba27ca,
+        0x96267c7535b763b5,0x4bc1558b2f3458de,
+        0xbbb01b9283253ca2,0x9eb1aaedfb016f16,
+        0xea9c227723ee8bcb,0x465e15a979c1cadc,
+        0x92a1958a7675175f,0xbfacd89ec191ec9,
+        0xb749faed14125d36,0xcef980ec671f667b,
+        0xe51c79a85916f484,0x82b7e12780e7401a,
+        0x8f31cc0937ae58d2,0xd1b2ecb8b0908810,
+        0xb2fe3f0b8599ef07,0x861fa7e6dcb4aa15,
+        0xdfbdcece67006ac9,0x67a791e093e1d49a,
+        0x8bd6a141006042bd,0xe0c8bb2c5c6d24e0,
+        0xaecc49914078536d,0x58fae9f773886e18,
+        0xda7f5bf590966848,0xaf39a475506a899e,
+        0x888f99797a5e012d,0x6d8406c952429603,
+        0xaab37fd7d8f58178,0xc8e5087ba6d33b83,
+        0xd5605fcdcf32e1d6,0xfb1e4a9a90880a64,
+        0x855c3be0a17fcd26,0x5cf2eea09a55067f,
+        0xa6b34ad8c9dfc06f,0xf42faa48c0ea481e,
+        0xd0601d8efc57b08b,0xf13b94daf124da26,
+        0x823c12795db6ce57,0x76c53d08d6b70858,
+        0xa2cb1717b52481ed,0x54768c4b0c64ca6e,
+        0xcb7ddcdda26da268,0xa9942f5dcf7dfd09,
+        0xfe5d54150b090b02,0xd3f93b35435d7c4c,
+        0x9efa548d26e5a6e1,0xc47bc5014a1a6daf,
+        0xc6b8e9b0709f109a,0x359ab6419ca1091b,
+        0xf867241c8cc6d4c0,0xc30163d203c94b62,
+        0x9b407691d7fc44f8,0x79e0de63425dcf1d,
+        0xc21094364dfb5636,0x985915fc12f542e4,
+        0xf294b943e17a2bc4,0x3e6f5b7b17b2939d,
+        0x979cf3ca6cec5b5a,0xa705992ceecf9c42,
+        0xbd8430bd08277231,0x50c6ff782a838353,
+        0xece53cec4a314ebd,0xa4f8bf5635246428,
+        0x940f4613ae5ed136,0x871b7795e136be99,
+        0xb913179899f68584,0x28e2557b59846e3f,
+        0xe757dd7ec07426e5,0x331aeada2fe589cf,
+        0x9096ea6f3848984f,0x3ff0d2c85def7621,
+        0xb4bca50b065abe63,0xfed077a756b53a9,
+        0xe1ebce4dc7f16dfb,0xd3e8495912c62894,
+        0x8d3360f09cf6e4bd,0x64712dd7abbbd95c,
+        0xb080392cc4349dec,0xbd8d794d96aacfb3,
+        0xdca04777f541c567,0xecf0d7a0fc5583a0,
+        0x89e42caaf9491b60,0xf41686c49db57244,
+        0xac5d37d5b79b6239,0x311c2875c522ced5,
+        0xd77485cb25823ac7,0x7d633293366b828b,
+        0x86a8d39ef77164bc,0xae5dff9c02033197,
+        0xa8530886b54dbdeb,0xd9f57f830283fdfc,
+        0xd267caa862a12d66,0xd072df63c324fd7b,
+        0x8380dea93da4bc60,0x4247cb9e59f71e6d,
+        0xa46116538d0deb78,0x52d9be85f074e608,
+        0xcd795be870516656,0x67902e276c921f8b,
+        0x806bd9714632dff6,0xba1cd8a3db53b6,
+        0xa086cfcd97bf97f3,0x80e8a40eccd228a4,
+        0xc8a883c0fdaf7df0,0x6122cd128006b2cd,
+        0xfad2a4b13d1b5d6c,0x796b805720085f81,
+        0x9cc3a6eec6311a63,0xcbe3303674053bb0,
+        0xc3f490aa77bd60fc,0xbedbfc4411068a9c,
+        0xf4f1b4d515acb93b,0xee92fb5515482d44,
+        0x991711052d8bf3c5,0x751bdd152d4d1c4a,
+        0xbf5cd54678eef0b6,0xd262d45a78a0635d,
+        0xef340a98172aace4,0x86fb897116c87c34,
+        0x9580869f0e7aac0e,0xd45d35e6ae3d4da0,
+        0xbae0a846d2195712,0x8974836059cca109,
+        0xe998d258869facd7,0x2bd1a438703fc94b,
+        0x91ff83775423cc06,0x7b6306a34627ddcf,
+        0xb67f6455292cbf08,0x1a3bc84c17b1d542,
+        0xe41f3d6a7377eeca,0x20caba5f1d9e4a93,
+        0x8e938662882af53e,0x547eb47b7282ee9c,
+        0xb23867fb2a35b28d,0xe99e619a4f23aa43,
+        0xdec681f9f4c31f31,0x6405fa00e2ec94d4,
+        0x8b3c113c38f9f37e,0xde83bc408dd3dd04,
+        0xae0b158b4738705e,0x9624ab50b148d445,
+        0xd98ddaee19068c76,0x3badd624dd9b0957,
+        0x87f8a8d4cfa417c9,0xe54ca5d70a80e5d6,
+        0xa9f6d30a038d1dbc,0x5e9fcf4ccd211f4c,
+        0xd47487cc8470652b,0x7647c3200069671f,
+        0x84c8d4dfd2c63f3b,0x29ecd9f40041e073,
+        0xa5fb0a17c777cf09,0xf468107100525890,
+        0xcf79cc9db955c2cc,0x7182148d4066eeb4,
+        0x81ac1fe293d599bf,0xc6f14cd848405530,
+        0xa21727db38cb002f,0xb8ada00e5a506a7c,
+        0xca9cf1d206fdc03b,0xa6d90811f0e4851c,
+        0xfd442e4688bd304a,0x908f4a166d1da663,
+        0x9e4a9cec15763e2e,0x9a598e4e043287fe,
+        0xc5dd44271ad3cdba,0x40eff1e1853f29fd,
+        0xf7549530e188c128,0xd12bee59e68ef47c,
+        0x9a94dd3e8cf578b9,0x82bb74f8301958ce,
+        0xc13a148e3032d6e7,0xe36a52363c1faf01,
+        0xf18899b1bc3f8ca1,0xdc44e6c3cb279ac1,
+        0x96f5600f15a7b7e5,0x29ab103a5ef8c0b9,
+        0xbcb2b812db11a5de,0x7415d448f6b6f0e7,
+        0xebdf661791d60f56,0x111b495b3464ad21,
+        0x936b9fcebb25c995,0xcab10dd900beec34,
+        0xb84687c269ef3bfb,0x3d5d514f40eea742,
+        0xe65829b3046b0afa,0xcb4a5a3112a5112,
+        0x8ff71a0fe2c2e6dc,0x47f0e785eaba72ab,
+        0xb3f4e093db73a093,0x59ed216765690f56,
+        0xe0f218b8d25088b8,0x306869c13ec3532c,
+        0x8c974f7383725573,0x1e414218c73a13fb,
+        0xafbd2350644eeacf,0xe5d1929ef90898fa,
+        0xdbac6c247d62a583,0xdf45f746b74abf39,
+        0x894bc396ce5da772,0x6b8bba8c328eb783,
+        0xab9eb47c81f5114f,0x66ea92f3f326564,
+        0xd686619ba27255a2,0xc80a537b0efefebd,
+        0x8613fd0145877585,0xbd06742ce95f5f36,
+        0xa798fc4196e952e7,0x2c48113823b73704,
+        0xd17f3b51fca3a7a0,0xf75a15862ca504c5,
+        0x82ef85133de648c4,0x9a984d73dbe722fb,
+        0xa3ab66580d5fdaf5,0xc13e60d0d2e0ebba,
+        0xcc963fee10b7d1b3,0x318df905079926a8,
+        0xffbbcfe994e5c61f,0xfdf17746497f7052,
+        0x9fd561f1fd0f9bd3,0xfeb6ea8bedefa633,
+        0xc7caba6e7c5382c8,0xfe64a52ee96b8fc0,
+        0xf9bd690a1b68637b,0x3dfdce7aa3c673b0,
+        0x9c1661a651213e2d,0x6bea10ca65c084e,
+        0xc31bfa0fe5698db8,0x486e494fcff30a62,
+        0xf3e2f893dec3f126,0x5a89dba3c3efccfa,
+        0x986ddb5c6b3a76b7,0xf89629465a75e01c,
+        0xbe89523386091465,0xf6bbb397f1135823,
+        0xee2ba6c0678b597f,0x746aa07ded582e2c,
+        0x94db483840b717ef,0xa8c2a44eb4571cdc,
+        0xba121a4650e4ddeb,0x92f34d62616ce413,
+        0xe896a0d7e51e1566,0x77b020baf9c81d17,
+        0x915e2486ef32cd60,0xace1474dc1d122e,
+        0xb5b5ada8aaff80b8,0xd819992132456ba,
+        0xe3231912d5bf60e6,0x10e1fff697ed6c69,
+        0x8df5efabc5979c8f,0xca8d3ffa1ef463c1,
+        0xb1736b96b6fd83b3,0xbd308ff8a6b17cb2,
+        0xddd0467c64bce4a0,0xac7cb3f6d05ddbde,
+        0x8aa22c0dbef60ee4,0x6bcdf07a423aa96b,
+        0xad4ab7112eb3929d,0x86c16c98d2c953c6,
+        0xd89d64d57a607744,0xe871c7bf077ba8b7,
+        0x87625f056c7c4a8b,0x11471cd764ad4972,
+        0xa93af6c6c79b5d2d,0xd598e40d3dd89bcf,
+        0xd389b47879823479,0x4aff1d108d4ec2c3,
+        0x843610cb4bf160cb,0xcedf722a585139ba,
+        0xa54394fe1eedb8fe,0xc2974eb4ee658828,
+        0xce947a3da6a9273e,0x733d226229feea32,
+        0x811ccc668829b887,0x806357d5a3f525f,
+        0xa163ff802a3426a8,0xca07c2dcb0cf26f7,
+        0xc9bcff6034c13052,0xfc89b393dd02f0b5,
+        0xfc2c3f3841f17c67,0xbbac2078d443ace2,
+        0x9d9ba7832936edc0,0xd54b944b84aa4c0d,
+        0xc5029163f384a931,0xa9e795e65d4df11,
+        0xf64335bcf065d37d,0x4d4617b5ff4a16d5,
+        0x99ea0196163fa42e,0x504bced1bf8e4e45,
+        0xc06481fb9bcf8d39,0xe45ec2862f71e1d6,
+        0xf07da27a82c37088,0x5d767327bb4e5a4c,
+        0x964e858c91ba2655,0x3a6a07f8d510f86f,
+        0xbbe226efb628afea,0x890489f70a55368b,
+        0xeadab0aba3b2dbe5,0x2b45ac74ccea842e,
+        0x92c8ae6b464fc96f,0x3b0b8bc90012929d,
+        0xb77ada0617e3bbcb,0x9ce6ebb40173744,
+        0xe55990879ddcaabd,0xcc420a6a101d0515,
+        0x8f57fa54c2a9eab6,0x9fa946824a12232d,
+        0xb32df8e9f3546564,0x47939822dc96abf9,
+        0xdff9772470297ebd,0x59787e2b93bc56f7,
+        0x8bfbea76c619ef36,0x57eb4edb3c55b65a,
+        0xaefae51477a06b03,0xede622920b6b23f1,
+        0xdab99e59958885c4,0xe95fab368e45eced,
+        0x88b402f7fd75539b,0x11dbcb0218ebb414,
+        0xaae103b5fcd2a881,0xd652bdc29f26a119,
+        0xd59944a37c0752a2,0x4be76d3346f0495f,
+        0x857fcae62d8493a5,0x6f70a4400c562ddb,
+        0xa6dfbd9fb8e5b88e,0xcb4ccd500f6bb952,
+        0xd097ad07a71f26b2,0x7e2000a41346a7a7,
+        0x825ecc24c873782f,0x8ed400668c0c28c8,
+        0xa2f67f2dfa90563b,0x728900802f0f32fa,
+        0xcbb41ef979346bca,0x4f2b40a03ad2ffb9,
+        0xfea126b7d78186bc,0xe2f610c84987bfa8,
+        0x9f24b832e6b0f436,0xdd9ca7d2df4d7c9,
+        0xc6ede63fa05d3143,0x91503d1c79720dbb,
+        0xf8a95fcf88747d94,0x75a44c6397ce912a,
+        0x9b69dbe1b548ce7c,0xc986afbe3ee11aba,
+        0xc24452da229b021b,0xfbe85badce996168,
+        0xf2d56790ab41c2a2,0xfae27299423fb9c3,
+        0x97c560ba6b0919a5,0xdccd879fc967d41a,
+        0xbdb6b8e905cb600f,0x5400e987bbc1c920,
+        0xed246723473e3813,0x290123e9aab23b68,
+        0x9436c0760c86e30b,0xf9a0b6720aaf6521,
+        0xb94470938fa89bce,0xf808e40e8d5b3e69,
+        0xe7958cb87392c2c2,0xb60b1d1230b20e04,
+        0x90bd77f3483bb9b9,0xb1c6f22b5e6f48c2,
+        0xb4ecd5f01a4aa828,0x1e38aeb6360b1af3,
+        0xe2280b6c20dd5232,0x25c6da63c38de1b0,
+        0x8d590723948a535f,0x579c487e5a38ad0e,
+        0xb0af48ec79ace837,0x2d835a9df0c6d851,
+        0xdcdb1b2798182244,0xf8e431456cf88e65,
+        0x8a08f0f8bf0f156b,0x1b8e9ecb641b58ff,
+        0xac8b2d36eed2dac5,0xe272467e3d222f3f,
+        0xd7adf884aa879177,0x5b0ed81dcc6abb0f,
+        0x86ccbb52ea94baea,0x98e947129fc2b4e9,
+        0xa87fea27a539e9a5,0x3f2398d747b36224,
+        0xd29fe4b18e88640e,0x8eec7f0d19a03aad,
+        0x83a3eeeef9153e89,0x1953cf68300424ac,
+        0xa48ceaaab75a8e2b,0x5fa8c3423c052dd7,
+        0xcdb02555653131b6,0x3792f412cb06794d,
+        0x808e17555f3ebf11,0xe2bbd88bbee40bd0,
+        0xa0b19d2ab70e6ed6,0x5b6aceaeae9d0ec4,
+        0xc8de047564d20a8b,0xf245825a5a445275,
+        0xfb158592be068d2e,0xeed6e2f0f0d56712,
+        0x9ced737bb6c4183d,0x55464dd69685606b,
+        0xc428d05aa4751e4c,0xaa97e14c3c26b886,
+        0xf53304714d9265df,0xd53dd99f4b3066a8,
+        0x993fe2c6d07b7fab,0xe546a8038efe4029,
+        0xbf8fdb78849a5f96,0xde98520472bdd033,
+        0xef73d256a5c0f77c,0x963e66858f6d4440,
+        0x95a8637627989aad,0xdde7001379a44aa8,
+        0xbb127c53b17ec159,0x5560c018580d5d52,
+        0xe9d71b689dde71af,0xaab8f01e6e10b4a6,
+        0x9226712162ab070d,0xcab3961304ca70e8,
+        0xb6b00d69bb55c8d1,0x3d607b97c5fd0d22,
+        0xe45c10c42a2b3b05,0x8cb89a7db77c506a,
+        0x8eb98a7a9a5b04e3,0x77f3608e92adb242,
+        0xb267ed1940f1c61c,0x55f038b237591ed3,
+        0xdf01e85f912e37a3,0x6b6c46dec52f6688,
+        0x8b61313bbabce2c6,0x2323ac4b3b3da015,
+        0xae397d8aa96c1b77,0xabec975e0a0d081a,
+        0xd9c7dced53c72255,0x96e7bd358c904a21,
+        0x881cea14545c7575,0x7e50d64177da2e54,
+        0xaa242499697392d2,0xdde50bd1d5d0b9e9,
+        0xd4ad2dbfc3d07787,0x955e4ec64b44e864,
+        0x84ec3c97da624ab4,0xbd5af13bef0b113e,
+        0xa6274bbdd0fadd61,0xecb1ad8aeacdd58e,
+        0xcfb11ead453994ba,0x67de18eda5814af2,
+        0x81ceb32c4b43fcf4,0x80eacf948770ced7,
+        0xa2425ff75e14fc31,0xa1258379a94d028d,
+        0xcad2f7f5359a3b3e,0x96ee45813a04330,
+        0xfd87b5f28300ca0d,0x8bca9d6e188853fc,
+        0x9e74d1b791e07e48,0x775ea264cf55347e,
+        0xc612062576589dda,0x95364afe032a81a0,
+        0xf79687aed3eec551,0x3a83ddbd83f52210,
+        0x9abe14cd44753b52,0xc4926a9672793580,
+        0xc16d9a0095928a27,0x75b7053c0f178400,
+        0xf1c90080baf72cb1,0x5324c68b12dd6800,
+        0x971da05074da7bee,0xd3f6fc16ebca8000,
+        0xbce5086492111aea,0x88f4bb1ca6bd0000,
+        0xec1e4a7db69561a5,0x2b31e9e3d0700000,
+        0x9392ee8e921d5d07,0x3aff322e62600000,
+        0xb877aa3236a4b449,0x9befeb9fad487c3,
+        0xe69594bec44de15b,0x4c2ebe687989a9b4,
+        0x901d7cf73ab0acd9,0xf9d37014bf60a11,
+        0xb424dc35095cd80f,0x538484c19ef38c95,
+        0xe12e13424bb40e13,0x2865a5f206b06fba,
+        0x8cbccc096f5088cb,0xf93f87b7442e45d4,
+        0xafebff0bcb24aafe,0xf78f69a51539d749,
+        0xdbe6fecebdedd5be,0xb573440e5a884d1c,
+        0x89705f4136b4a597,0x31680a88f8953031,
+        0xabcc77118461cefc,0xfdc20d2b36ba7c3e,
+        0xd6bf94d5e57a42bc,0x3d32907604691b4d,
+        0x8637bd05af6c69b5,0xa63f9a49c2c1b110,
+        0xa7c5ac471b478423,0xfcf80dc33721d54,
+        0xd1b71758e219652b,0xd3c36113404ea4a9,
+        0x83126e978d4fdf3b,0x645a1cac083126ea,
+        0xa3d70a3d70a3d70a,0x3d70a3d70a3d70a4,
+        0xcccccccccccccccc,0xcccccccccccccccd,
+        0x8000000000000000,0x0,
+        0xa000000000000000,0x0,
+        0xc800000000000000,0x0,
+        0xfa00000000000000,0x0,
+        0x9c40000000000000,0x0,
+        0xc350000000000000,0x0,
+        0xf424000000000000,0x0,
+        0x9896800000000000,0x0,
+        0xbebc200000000000,0x0,
+        0xee6b280000000000,0x0,
+        0x9502f90000000000,0x0,
+        0xba43b74000000000,0x0,
+        0xe8d4a51000000000,0x0,
+        0x9184e72a00000000,0x0,
+        0xb5e620f480000000,0x0,
+        0xe35fa931a0000000,0x0,
+        0x8e1bc9bf04000000,0x0,
+        0xb1a2bc2ec5000000,0x0,
+        0xde0b6b3a76400000,0x0,
+        0x8ac7230489e80000,0x0,
+        0xad78ebc5ac620000,0x0,
+        0xd8d726b7177a8000,0x0,
+        0x878678326eac9000,0x0,
+        0xa968163f0a57b400,0x0,
+        0xd3c21bcecceda100,0x0,
+        0x84595161401484a0,0x0,
+        0xa56fa5b99019a5c8,0x0,
+        0xcecb8f27f4200f3a,0x0,
+        0x813f3978f8940984,0x4000000000000000,
+        0xa18f07d736b90be5,0x5000000000000000,
+        0xc9f2c9cd04674ede,0xa400000000000000,
+        0xfc6f7c4045812296,0x4d00000000000000,
+        0x9dc5ada82b70b59d,0xf020000000000000,
+        0xc5371912364ce305,0x6c28000000000000,
+        0xf684df56c3e01bc6,0xc732000000000000,
+        0x9a130b963a6c115c,0x3c7f400000000000,
+        0xc097ce7bc90715b3,0x4b9f100000000000,
+        0xf0bdc21abb48db20,0x1e86d40000000000,
+        0x96769950b50d88f4,0x1314448000000000,
+        0xbc143fa4e250eb31,0x17d955a000000000,
+        0xeb194f8e1ae525fd,0x5dcfab0800000000,
+        0x92efd1b8d0cf37be,0x5aa1cae500000000,
+        0xb7abc627050305ad,0xf14a3d9e40000000,
+        0xe596b7b0c643c719,0x6d9ccd05d0000000,
+        0x8f7e32ce7bea5c6f,0xe4820023a2000000,
+        0xb35dbf821ae4f38b,0xdda2802c8a800000,
+        0xe0352f62a19e306e,0xd50b2037ad200000,
+        0x8c213d9da502de45,0x4526f422cc340000,
+        0xaf298d050e4395d6,0x9670b12b7f410000,
+        0xdaf3f04651d47b4c,0x3c0cdd765f114000,
+        0x88d8762bf324cd0f,0xa5880a69fb6ac800,
+        0xab0e93b6efee0053,0x8eea0d047a457a00,
+        0xd5d238a4abe98068,0x72a4904598d6d880,
+        0x85a36366eb71f041,0x47a6da2b7f864750,
+        0xa70c3c40a64e6c51,0x999090b65f67d924,
+        0xd0cf4b50cfe20765,0xfff4b4e3f741cf6d,
+        0x82818f1281ed449f,0xbff8f10e7a8921a4,
+        0xa321f2d7226895c7,0xaff72d52192b6a0d,
+        0xcbea6f8ceb02bb39,0x9bf4f8a69f764490,
+        0xfee50b7025c36a08,0x2f236d04753d5b4,
+        0x9f4f2726179a2245,0x1d762422c946590,
+        0xc722f0ef9d80aad6,0x424d3ad2b7b97ef5,
+        0xf8ebad2b84e0d58b,0xd2e0898765a7deb2,
+        0x9b934c3b330c8577,0x63cc55f49f88eb2f,
+        0xc2781f49ffcfa6d5,0x3cbf6b71c76b25fb,
+        0xf316271c7fc3908a,0x8bef464e3945ef7a,
+        0x97edd871cfda3a56,0x97758bf0e3cbb5ac,
+        0xbde94e8e43d0c8ec,0x3d52eeed1cbea317,
+        0xed63a231d4c4fb27,0x4ca7aaa863ee4bdd,
+        0x945e455f24fb1cf8,0x8fe8caa93e74ef6a,
+        0xb975d6b6ee39e436,0xb3e2fd538e122b44,
+        0xe7d34c64a9c85d44,0x60dbbca87196b616,
+        0x90e40fbeea1d3a4a,0xbc8955e946fe31cd,
+        0xb51d13aea4a488dd,0x6babab6398bdbe41,
+        0xe264589a4dcdab14,0xc696963c7eed2dd1,
+        0x8d7eb76070a08aec,0xfc1e1de5cf543ca2,
+        0xb0de65388cc8ada8,0x3b25a55f43294bcb,
+        0xdd15fe86affad912,0x49ef0eb713f39ebe,
+        0x8a2dbf142dfcc7ab,0x6e3569326c784337,
+        0xacb92ed9397bf996,0x49c2c37f07965404,
+        0xd7e77a8f87daf7fb,0xdc33745ec97be906,
+        0x86f0ac99b4e8dafd,0x69a028bb3ded71a3,
+        0xa8acd7c0222311bc,0xc40832ea0d68ce0c,
+        0xd2d80db02aabd62b,0xf50a3fa490c30190,
+        0x83c7088e1aab65db,0x792667c6da79e0fa,
+        0xa4b8cab1a1563f52,0x577001b891185938,
+        0xcde6fd5e09abcf26,0xed4c0226b55e6f86,
+        0x80b05e5ac60b6178,0x544f8158315b05b4,
+        0xa0dc75f1778e39d6,0x696361ae3db1c721,
+        0xc913936dd571c84c,0x3bc3a19cd1e38e9,
+        0xfb5878494ace3a5f,0x4ab48a04065c723,
+        0x9d174b2dcec0e47b,0x62eb0d64283f9c76,
+        0xc45d1df942711d9a,0x3ba5d0bd324f8394,
+        0xf5746577930d6500,0xca8f44ec7ee36479,
+        0x9968bf6abbe85f20,0x7e998b13cf4e1ecb,
+        0xbfc2ef456ae276e8,0x9e3fedd8c321a67e,
+        0xefb3ab16c59b14a2,0xc5cfe94ef3ea101e,
+        0x95d04aee3b80ece5,0xbba1f1d158724a12,
+        0xbb445da9ca61281f,0x2a8a6e45ae8edc97,
+        0xea1575143cf97226,0xf52d09d71a3293bd,
+        0x924d692ca61be758,0x593c2626705f9c56,
+        0xb6e0c377cfa2e12e,0x6f8b2fb00c77836c,
+        0xe498f455c38b997a,0xb6dfb9c0f956447,
+        0x8edf98b59a373fec,0x4724bd4189bd5eac,
+        0xb2977ee300c50fe7,0x58edec91ec2cb657,
+        0xdf3d5e9bc0f653e1,0x2f2967b66737e3ed,
+        0x8b865b215899f46c,0xbd79e0d20082ee74,
+        0xae67f1e9aec07187,0xecd8590680a3aa11,
+        0xda01ee641a708de9,0xe80e6f4820cc9495,
+        0x884134fe908658b2,0x3109058d147fdcdd,
+        0xaa51823e34a7eede,0xbd4b46f0599fd415,
+        0xd4e5e2cdc1d1ea96,0x6c9e18ac7007c91a,
+        0x850fadc09923329e,0x3e2cf6bc604ddb0,
+        0xa6539930bf6bff45,0x84db8346b786151c,
+        0xcfe87f7cef46ff16,0xe612641865679a63,
+        0x81f14fae158c5f6e,0x4fcb7e8f3f60c07e,
+        0xa26da3999aef7749,0xe3be5e330f38f09d,
+        0xcb090c8001ab551c,0x5cadf5bfd3072cc5,
+        0xfdcb4fa002162a63,0x73d9732fc7c8f7f6,
+        0x9e9f11c4014dda7e,0x2867e7fddcdd9afa,
+        0xc646d63501a1511d,0xb281e1fd541501b8,
+        0xf7d88bc24209a565,0x1f225a7ca91a4226,
+        0x9ae757596946075f,0x3375788de9b06958,
+        0xc1a12d2fc3978937,0x52d6b1641c83ae,
+        0xf209787bb47d6b84,0xc0678c5dbd23a49a,
+        0x9745eb4d50ce6332,0xf840b7ba963646e0,
+        0xbd176620a501fbff,0xb650e5a93bc3d898,
+        0xec5d3fa8ce427aff,0xa3e51f138ab4cebe,
+        0x93ba47c980e98cdf,0xc66f336c36b10137,
+        0xb8a8d9bbe123f017,0xb80b0047445d4184,
+        0xe6d3102ad96cec1d,0xa60dc059157491e5,
+        0x9043ea1ac7e41392,0x87c89837ad68db2f,
+        0xb454e4a179dd1877,0x29babe4598c311fb,
+        0xe16a1dc9d8545e94,0xf4296dd6fef3d67a,
+        0x8ce2529e2734bb1d,0x1899e4a65f58660c,
+        0xb01ae745b101e9e4,0x5ec05dcff72e7f8f,
+        0xdc21a1171d42645d,0x76707543f4fa1f73,
+        0x899504ae72497eba,0x6a06494a791c53a8,
+        0xabfa45da0edbde69,0x487db9d17636892,
+        0xd6f8d7509292d603,0x45a9d2845d3c42b6,
+        0x865b86925b9bc5c2,0xb8a2392ba45a9b2,
+        0xa7f26836f282b732,0x8e6cac7768d7141e,
+        0xd1ef0244af2364ff,0x3207d795430cd926,
+        0x8335616aed761f1f,0x7f44e6bd49e807b8,
+        0xa402b9c5a8d3a6e7,0x5f16206c9c6209a6,
+        0xcd036837130890a1,0x36dba887c37a8c0f,
+        0x802221226be55a64,0xc2494954da2c9789,
+        0xa02aa96b06deb0fd,0xf2db9baa10b7bd6c,
+        0xc83553c5c8965d3d,0x6f92829494e5acc7,
+        0xfa42a8b73abbf48c,0xcb772339ba1f17f9,
+        0x9c69a97284b578d7,0xff2a760414536efb,
+        0xc38413cf25e2d70d,0xfef5138519684aba,
+        0xf46518c2ef5b8cd1,0x7eb258665fc25d69,
+        0x98bf2f79d5993802,0xef2f773ffbd97a61,
+        0xbeeefb584aff8603,0xaafb550ffacfd8fa,
+        0xeeaaba2e5dbf6784,0x95ba2a53f983cf38,
+        0x952ab45cfa97a0b2,0xdd945a747bf26183,
+        0xba756174393d88df,0x94f971119aeef9e4,
+        0xe912b9d1478ceb17,0x7a37cd5601aab85d,
+        0x91abb422ccb812ee,0xac62e055c10ab33a,
+        0xb616a12b7fe617aa,0x577b986b314d6009,
+        0xe39c49765fdf9d94,0xed5a7e85fda0b80b,
+        0x8e41ade9fbebc27d,0x14588f13be847307,
+        0xb1d219647ae6b31c,0x596eb2d8ae258fc8,
+        0xde469fbd99a05fe3,0x6fca5f8ed9aef3bb,
+        0x8aec23d680043bee,0x25de7bb9480d5854,
+        0xada72ccc20054ae9,0xaf561aa79a10ae6a,
+        0xd910f7ff28069da4,0x1b2ba1518094da04,
+        0x87aa9aff79042286,0x90fb44d2f05d0842,
+        0xa99541bf57452b28,0x353a1607ac744a53,
+        0xd3fa922f2d1675f2,0x42889b8997915ce8,
+        0x847c9b5d7c2e09b7,0x69956135febada11,
+        0xa59bc234db398c25,0x43fab9837e699095,
+        0xcf02b2c21207ef2e,0x94f967e45e03f4bb,
+        0x8161afb94b44f57d,0x1d1be0eebac278f5,
+        0xa1ba1ba79e1632dc,0x6462d92a69731732,
+        0xca28a291859bbf93,0x7d7b8f7503cfdcfe,
+        0xfcb2cb35e702af78,0x5cda735244c3d43e,
+        0x9defbf01b061adab,0x3a0888136afa64a7,
+        0xc56baec21c7a1916,0x88aaa1845b8fdd0,
+        0xf6c69a72a3989f5b,0x8aad549e57273d45,
+        0x9a3c2087a63f6399,0x36ac54e2f678864b,
+        0xc0cb28a98fcf3c7f,0x84576a1bb416a7dd,
+        0xf0fdf2d3f3c30b9f,0x656d44a2a11c51d5,
+        0x969eb7c47859e743,0x9f644ae5a4b1b325,
+        0xbc4665b596706114,0x873d5d9f0dde1fee,
+        0xeb57ff22fc0c7959,0xa90cb506d155a7ea,
+        0x9316ff75dd87cbd8,0x9a7f12442d588f2,
+        0xb7dcbf5354e9bece,0xc11ed6d538aeb2f,
+        0xe5d3ef282a242e81,0x8f1668c8a86da5fa,
+        0x8fa475791a569d10,0xf96e017d694487bc,
+        0xb38d92d760ec4455,0x37c981dcc395a9ac,
+        0xe070f78d3927556a,0x85bbe253f47b1417,
+        0x8c469ab843b89562,0x93956d7478ccec8e,
+        0xaf58416654a6babb,0x387ac8d1970027b2,
+        0xdb2e51bfe9d0696a,0x6997b05fcc0319e,
+        0x88fcf317f22241e2,0x441fece3bdf81f03,
+        0xab3c2fddeeaad25a,0xd527e81cad7626c3,
+        0xd60b3bd56a5586f1,0x8a71e223d8d3b074,
+        0x85c7056562757456,0xf6872d5667844e49,
+        0xa738c6bebb12d16c,0xb428f8ac016561db,
+        0xd106f86e69d785c7,0xe13336d701beba52,
+        0x82a45b450226b39c,0xecc0024661173473,
+        0xa34d721642b06084,0x27f002d7f95d0190,
+        0xcc20ce9bd35c78a5,0x31ec038df7b441f4,
+        0xff290242c83396ce,0x7e67047175a15271,
+        0x9f79a169bd203e41,0xf0062c6e984d386,
+        0xc75809c42c684dd1,0x52c07b78a3e60868,
+        0xf92e0c3537826145,0xa7709a56ccdf8a82,
+        0x9bbcc7a142b17ccb,0x88a66076400bb691,
+        0xc2abf989935ddbfe,0x6acff893d00ea435,
+        0xf356f7ebf83552fe,0x583f6b8c4124d43,
+        0x98165af37b2153de,0xc3727a337a8b704a,
+        0xbe1bf1b059e9a8d6,0x744f18c0592e4c5c,
+        0xeda2ee1c7064130c,0x1162def06f79df73,
+        0x9485d4d1c63e8be7,0x8addcb5645ac2ba8,
+        0xb9a74a0637ce2ee1,0x6d953e2bd7173692,
+        0xe8111c87c5c1ba99,0xc8fa8db6ccdd0437,
+        0x910ab1d4db9914a0,0x1d9c9892400a22a2,
+        0xb54d5e4a127f59c8,0x2503beb6d00cab4b,
+        0xe2a0b5dc971f303a,0x2e44ae64840fd61d,
+        0x8da471a9de737e24,0x5ceaecfed289e5d2,
+        0xb10d8e1456105dad,0x7425a83e872c5f47,
+        0xdd50f1996b947518,0xd12f124e28f77719,
+        0x8a5296ffe33cc92f,0x82bd6b70d99aaa6f,
+        0xace73cbfdc0bfb7b,0x636cc64d1001550b,
+        0xd8210befd30efa5a,0x3c47f7e05401aa4e,
+        0x8714a775e3e95c78,0x65acfaec34810a71,
+        0xa8d9d1535ce3b396,0x7f1839a741a14d0d,
+        0xd31045a8341ca07c,0x1ede48111209a050,
+        0x83ea2b892091e44d,0x934aed0aab460432,
+        0xa4e4b66b68b65d60,0xf81da84d5617853f,
+        0xce1de40642e3f4b9,0x36251260ab9d668e,
+        0x80d2ae83e9ce78f3,0xc1d72b7c6b426019,
+        0xa1075a24e4421730,0xb24cf65b8612f81f,
+        0xc94930ae1d529cfc,0xdee033f26797b627,
+        0xfb9b7cd9a4a7443c,0x169840ef017da3b1,
+        0x9d412e0806e88aa5,0x8e1f289560ee864e,
+        0xc491798a08a2ad4e,0xf1a6f2bab92a27e2,
+        0xf5b5d7ec8acb58a2,0xae10af696774b1db,
+        0x9991a6f3d6bf1765,0xacca6da1e0a8ef29,
+        0xbff610b0cc6edd3f,0x17fd090a58d32af3,
+        0xeff394dcff8a948e,0xddfc4b4cef07f5b0,
+        0x95f83d0a1fb69cd9,0x4abdaf101564f98e,
+        0xbb764c4ca7a4440f,0x9d6d1ad41abe37f1,
+        0xea53df5fd18d5513,0x84c86189216dc5ed,
+        0x92746b9be2f8552c,0x32fd3cf5b4e49bb4,
+        0xb7118682dbb66a77,0x3fbc8c33221dc2a1,
+        0xe4d5e82392a40515,0xfabaf3feaa5334a,
+        0x8f05b1163ba6832d,0x29cb4d87f2a7400e,
+        0xb2c71d5bca9023f8,0x743e20e9ef511012,
+        0xdf78e4b2bd342cf6,0x914da9246b255416,
+        0x8bab8eefb6409c1a,0x1ad089b6c2f7548e,
+        0xae9672aba3d0c320,0xa184ac2473b529b1,
+        0xda3c0f568cc4f3e8,0xc9e5d72d90a2741e,
+        0x8865899617fb1871,0x7e2fa67c7a658892,
+        0xaa7eebfb9df9de8d,0xddbb901b98feeab7,
+        0xd51ea6fa85785631,0x552a74227f3ea565,
+        0x8533285c936b35de,0xd53a88958f87275f,
+        0xa67ff273b8460356,0x8a892abaf368f137,
+        0xd01fef10a657842c,0x2d2b7569b0432d85,
+        0x8213f56a67f6b29b,0x9c3b29620e29fc73,
+        0xa298f2c501f45f42,0x8349f3ba91b47b8f,
+        0xcb3f2f7642717713,0x241c70a936219a73,
+        0xfe0efb53d30dd4d7,0xed238cd383aa0110,
+        0x9ec95d1463e8a506,0xf4363804324a40aa,
+        0xc67bb4597ce2ce48,0xb143c6053edcd0d5,
+        0xf81aa16fdc1b81da,0xdd94b7868e94050a,
+        0x9b10a4e5e9913128,0xca7cf2b4191c8326,
+        0xc1d4ce1f63f57d72,0xfd1c2f611f63a3f0,
+        0xf24a01a73cf2dccf,0xbc633b39673c8cec,
+        0x976e41088617ca01,0xd5be0503e085d813,
+        0xbd49d14aa79dbc82,0x4b2d8644d8a74e18,
+        0xec9c459d51852ba2,0xddf8e7d60ed1219e,
+        0x93e1ab8252f33b45,0xcabb90e5c942b503,
+        0xb8da1662e7b00a17,0x3d6a751f3b936243,
+        0xe7109bfba19c0c9d,0xcc512670a783ad4,
+        0x906a617d450187e2,0x27fb2b80668b24c5,
+        0xb484f9dc9641e9da,0xb1f9f660802dedf6,
+        0xe1a63853bbd26451,0x5e7873f8a0396973,
+        0x8d07e33455637eb2,0xdb0b487b6423e1e8,
+        0xb049dc016abc5e5f,0x91ce1a9a3d2cda62,
+        0xdc5c5301c56b75f7,0x7641a140cc7810fb,
+        0x89b9b3e11b6329ba,0xa9e904c87fcb0a9d,
+        0xac2820d9623bf429,0x546345fa9fbdcd44,
+        0xd732290fbacaf133,0xa97c177947ad4095,
+        0x867f59a9d4bed6c0,0x49ed8eabcccc485d,
+        0xa81f301449ee8c70,0x5c68f256bfff5a74,
+        0xd226fc195c6a2f8c,0x73832eec6fff3111,
+        0x83585d8fd9c25db7,0xc831fd53c5ff7eab,
+        0xa42e74f3d032f525,0xba3e7ca8b77f5e55,
+        0xcd3a1230c43fb26f,0x28ce1bd2e55f35eb,
+        0x80444b5e7aa7cf85,0x7980d163cf5b81b3,
+        0xa0555e361951c366,0xd7e105bcc332621f,
+        0xc86ab5c39fa63440,0x8dd9472bf3fefaa7,
+        0xfa856334878fc150,0xb14f98f6f0feb951,
+        0x9c935e00d4b9d8d2,0x6ed1bf9a569f33d3,
+        0xc3b8358109e84f07,0xa862f80ec4700c8,
+        0xf4a642e14c6262c8,0xcd27bb612758c0fa,
+        0x98e7e9cccfbd7dbd,0x8038d51cb897789c,
+        0xbf21e44003acdd2c,0xe0470a63e6bd56c3,
+        0xeeea5d5004981478,0x1858ccfce06cac74,
+        0x95527a5202df0ccb,0xf37801e0c43ebc8,
+        0xbaa718e68396cffd,0xd30560258f54e6ba,
+        0xe950df20247c83fd,0x47c6b82ef32a2069,
+        0x91d28b7416cdd27e,0x4cdc331d57fa5441,
+        0xb6472e511c81471d,0xe0133fe4adf8e952,
+        0xe3d8f9e563a198e5,0x58180fddd97723a6,
+        0x8e679c2f5e44ff8f,0x570f09eaa7ea7648,};
+
+
+#endif // SIMDJSON_SRC_NUMBERPARSING_TABLES_CPP
+/* end file internal/numberparsing_tables.cpp */
+/* including internal/simdprune_tables.cpp: #include <internal/simdprune_tables.cpp> */
+/* begin file internal/simdprune_tables.cpp */
+#ifndef SIMDJSON_SRC_SIMDPRUNE_TABLES_CPP
+#define SIMDJSON_SRC_SIMDPRUNE_TABLES_CPP
+
+/* including simdjson/implementation_detection.h: #include <simdjson/implementation_detection.h> */
+/* begin file simdjson/implementation_detection.h */
+#ifndef SIMDJSON_IMPLEMENTATION_DETECTION_H
+#define SIMDJSON_IMPLEMENTATION_DETECTION_H
+
+/* skipped duplicate #include "simdjson/base.h" */
+
+// 0 is reserved, because undefined SIMDJSON_IMPLEMENTATION equals 0 in preprocessor macros.
+#define SIMDJSON_IMPLEMENTATION_ID_arm64 1
+#define SIMDJSON_IMPLEMENTATION_ID_fallback 2
+#define SIMDJSON_IMPLEMENTATION_ID_haswell 3
+#define SIMDJSON_IMPLEMENTATION_ID_icelake 4
+#define SIMDJSON_IMPLEMENTATION_ID_ppc64 5
+#define SIMDJSON_IMPLEMENTATION_ID_westmere 6
+#define SIMDJSON_IMPLEMENTATION_ID_lsx 7
+#define SIMDJSON_IMPLEMENTATION_ID_lasx 8
+//#define SIMDJSON_IMPLEMENTATION_ID_rvv 9
+#define SIMDJSON_IMPLEMENTATION_ID_rvv_vls 10
+
+#define SIMDJSON_IMPLEMENTATION_ID_FOR(IMPL) SIMDJSON_CAT(SIMDJSON_IMPLEMENTATION_ID_, IMPL)
+#define SIMDJSON_IMPLEMENTATION_ID SIMDJSON_IMPLEMENTATION_ID_FOR(SIMDJSON_IMPLEMENTATION)
+
+#define SIMDJSON_IMPLEMENTATION_IS(IMPL) SIMDJSON_IMPLEMENTATION_ID == SIMDJSON_IMPLEMENTATION_ID_FOR(IMPL)
+
+//
+// First, figure out which implementations can be run. Doing it here makes it so we don't have to worry about the order
+// in which we include them.
+//
+
+#ifndef SIMDJSON_IMPLEMENTATION_ARM64
+#define SIMDJSON_IMPLEMENTATION_ARM64 (SIMDJSON_IS_ARM64)
+#endif
+#if SIMDJSON_IMPLEMENTATION_ARM64 && SIMDJSON_IS_ARM64
+#define SIMDJSON_CAN_ALWAYS_RUN_ARM64 1
+#else
+#define SIMDJSON_CAN_ALWAYS_RUN_ARM64 0
+#endif
+
+// Default Icelake to on if this is x86-64. Even if we're not compiled for it, it could be selected
+// at runtime.
+#ifndef SIMDJSON_IMPLEMENTATION_ICELAKE
+#define SIMDJSON_IMPLEMENTATION_ICELAKE ((SIMDJSON_IS_X86_64) && (SIMDJSON_AVX512_ALLOWED) && (SIMDJSON_COMPILER_SUPPORTS_VBMI2))
+#endif
+
+#ifdef _MSC_VER
+// To see why  (__BMI__) && (__PCLMUL__) && (__LZCNT__) are not part of this next line, see
+// https://github.com/simdjson/simdjson/issues/1247
+#if ((SIMDJSON_IMPLEMENTATION_ICELAKE) && (__AVX2__) && (__AVX512F__) && (__AVX512DQ__) && (__AVX512CD__) && (__AVX512BW__) && (__AVX512VL__) && (__AVX512VBMI2__))
+#define SIMDJSON_CAN_ALWAYS_RUN_ICELAKE 1
+#else
+#define SIMDJSON_CAN_ALWAYS_RUN_ICELAKE 0
+#endif
+
+#else
+
+#if ((SIMDJSON_IMPLEMENTATION_ICELAKE) && (__AVX2__) && (__BMI__) && (__PCLMUL__) && (__LZCNT__) && (__AVX512F__) && (__AVX512DQ__) && (__AVX512CD__) && (__AVX512BW__) && (__AVX512VL__) && (__AVX512VBMI2__))
+#define SIMDJSON_CAN_ALWAYS_RUN_ICELAKE 1
+#else
+#define SIMDJSON_CAN_ALWAYS_RUN_ICELAKE 0
+#endif
+
+#endif
+
+// Default Haswell to on if this is x86-64. Even if we're not compiled for it, it could be selected
+// at runtime.
+#ifndef SIMDJSON_IMPLEMENTATION_HASWELL
+#if SIMDJSON_CAN_ALWAYS_RUN_ICELAKE
+// if icelake is always available, never enable haswell.
+#define SIMDJSON_IMPLEMENTATION_HASWELL 0
+#else
+#define SIMDJSON_IMPLEMENTATION_HASWELL SIMDJSON_IS_X86_64
+#endif
+#endif
+#ifdef _MSC_VER
+// To see why  (__BMI__) && (__PCLMUL__) && (__LZCNT__) are not part of this next line, see
+// https://github.com/simdjson/simdjson/issues/1247
+#if ((SIMDJSON_IMPLEMENTATION_HASWELL) && (SIMDJSON_IS_X86_64) && (__AVX2__))
+#define SIMDJSON_CAN_ALWAYS_RUN_HASWELL 1
+#else
+#define SIMDJSON_CAN_ALWAYS_RUN_HASWELL 0
+#endif
+
+#else
+
+#if ((SIMDJSON_IMPLEMENTATION_HASWELL) && (SIMDJSON_IS_X86_64) && (__AVX2__) && (__BMI__) && (__PCLMUL__) && (__LZCNT__))
+#define SIMDJSON_CAN_ALWAYS_RUN_HASWELL 1
+#else
+#define SIMDJSON_CAN_ALWAYS_RUN_HASWELL 0
+#endif
+
+#endif
+
+// Default Westmere to on if this is x86-64.
+#ifndef SIMDJSON_IMPLEMENTATION_WESTMERE
+#if SIMDJSON_CAN_ALWAYS_RUN_ICELAKE || SIMDJSON_CAN_ALWAYS_RUN_HASWELL
+// if icelake or haswell are always available, never enable westmere.
+#define SIMDJSON_IMPLEMENTATION_WESTMERE 0
+#else
+#define SIMDJSON_IMPLEMENTATION_WESTMERE SIMDJSON_IS_X86_64
+#endif
+#endif
+
+#if (SIMDJSON_IMPLEMENTATION_WESTMERE && SIMDJSON_IS_X86_64 && __SSE4_2__ && __PCLMUL__)
+#define SIMDJSON_CAN_ALWAYS_RUN_WESTMERE 1
+#else
+#define SIMDJSON_CAN_ALWAYS_RUN_WESTMERE 0
+#endif
+
+
+#ifndef SIMDJSON_IMPLEMENTATION_PPC64
+#define SIMDJSON_IMPLEMENTATION_PPC64 (SIMDJSON_IS_PPC64 && SIMDJSON_IS_PPC64_VMX)
+#endif
+#if SIMDJSON_IMPLEMENTATION_PPC64 && SIMDJSON_IS_PPC64 && SIMDJSON_IS_PPC64_VMX
+#define SIMDJSON_CAN_ALWAYS_RUN_PPC64 1
+#else
+#define SIMDJSON_CAN_ALWAYS_RUN_PPC64 0
+#endif
+
+#ifndef SIMDJSON_IMPLEMENTATION_LASX
+#define SIMDJSON_IMPLEMENTATION_LASX (SIMDJSON_IS_LSX)
+#endif
+#define SIMDJSON_CAN_ALWAYS_RUN_LASX (SIMDJSON_IS_LASX)
+
+#ifndef SIMDJSON_IMPLEMENTATION_LSX
+#if SIMDJSON_CAN_ALWAYS_RUN_LASX
+#define SIMDJSON_IMPLEMENTATION_LSX 0
+#else
+#define SIMDJSON_IMPLEMENTATION_LSX (SIMDJSON_IS_LSX)
+#endif
+#endif
+#define SIMDJSON_CAN_ALWAYS_RUN_LSX (SIMDJSON_IMPLEMENTATION_LSX)
+
+#define SIMDJSON_CAN_ALWAYS_RUN_RVV_VLS SIMDJSON_IS_RVV_VLS
+#ifndef SIMDJSON_IMPLEMENTATION_RVV_VLS
+#define SIMDJSON_IMPLEMENTATION_RVV_VLS SIMDJSON_CAN_ALWAYS_RUN_RVV_VLS
+#endif
+
+// Default Fallback to on unless a builtin implementation has already been selected.
+#ifndef SIMDJSON_IMPLEMENTATION_FALLBACK
+#if SIMDJSON_CAN_ALWAYS_RUN_ARM64 || SIMDJSON_CAN_ALWAYS_RUN_ICELAKE || SIMDJSON_CAN_ALWAYS_RUN_HASWELL || SIMDJSON_CAN_ALWAYS_RUN_WESTMERE || SIMDJSON_CAN_ALWAYS_RUN_PPC64 || SIMDJSON_CAN_ALWAYS_RUN_LSX || SIMDJSON_CAN_ALWAYS_RUN_LASX || SIMDJSON_CAN_ALWAYS_RUN_RVV_VLS
+// if anything at all except fallback can always run, then disable fallback.
+#define SIMDJSON_IMPLEMENTATION_FALLBACK 0
+#else
+#define SIMDJSON_IMPLEMENTATION_FALLBACK 1
+#endif
+#endif
+#define SIMDJSON_CAN_ALWAYS_RUN_FALLBACK SIMDJSON_IMPLEMENTATION_FALLBACK
+
+// Determine the best builtin implementation
+#ifndef SIMDJSON_BUILTIN_IMPLEMENTATION
+
+#if SIMDJSON_CAN_ALWAYS_RUN_ICELAKE
+#define SIMDJSON_BUILTIN_IMPLEMENTATION icelake
+#elif SIMDJSON_CAN_ALWAYS_RUN_HASWELL
+#define SIMDJSON_BUILTIN_IMPLEMENTATION haswell
+#elif SIMDJSON_CAN_ALWAYS_RUN_WESTMERE
+#define SIMDJSON_BUILTIN_IMPLEMENTATION westmere
+#elif SIMDJSON_CAN_ALWAYS_RUN_ARM64
+#define SIMDJSON_BUILTIN_IMPLEMENTATION arm64
+#elif SIMDJSON_CAN_ALWAYS_RUN_PPC64
+#define SIMDJSON_BUILTIN_IMPLEMENTATION ppc64
+#elif SIMDJSON_CAN_ALWAYS_RUN_LSX
+#define SIMDJSON_BUILTIN_IMPLEMENTATION lsx
+#elif SIMDJSON_CAN_ALWAYS_RUN_LASX
+#define SIMDJSON_BUILTIN_IMPLEMENTATION lasx
+#elif SIMDJSON_CAN_ALWAYS_RUN_RVV_VLS
+#define SIMDJSON_BUILTIN_IMPLEMENTATION rvv_vls
+#elif SIMDJSON_CAN_ALWAYS_RUN_FALLBACK
+#define SIMDJSON_BUILTIN_IMPLEMENTATION fallback
+#else
+#error "All possible implementations (including fallback) have been disabled! simdjson will not run."
+#endif
+
+#endif // SIMDJSON_BUILTIN_IMPLEMENTATION
+
+#define SIMDJSON_BUILTIN_IMPLEMENTATION_ID SIMDJSON_IMPLEMENTATION_ID_FOR(SIMDJSON_BUILTIN_IMPLEMENTATION)
+#define SIMDJSON_BUILTIN_IMPLEMENTATION_IS(IMPL) SIMDJSON_BUILTIN_IMPLEMENTATION_ID == SIMDJSON_IMPLEMENTATION_ID_FOR(IMPL)
+
+#endif // SIMDJSON_IMPLEMENTATION_DETECTION_H
+/* end file simdjson/implementation_detection.h */
+
+#if SIMDJSON_IMPLEMENTATION_ARM64 || SIMDJSON_IMPLEMENTATION_ICELAKE || SIMDJSON_IMPLEMENTATION_HASWELL || SIMDJSON_IMPLEMENTATION_WESTMERE || SIMDJSON_IMPLEMENTATION_PPC64 || SIMDJSON_IMPLEMENTATION_LSX || SIMDJSON_IMPLEMENTATION_LASX
+
+#include <cstdint>
+
+namespace simdjson { // table modified and copied from
+namespace internal { // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetTable
+SIMDJSON_DLLIMPORTEXPORT  const unsigned char BitsSetTable256mul2[256] = {
+    0,  2,  2,  4,  2,  4,  4,  6,  2,  4,  4,  6,  4,  6,  6,  8,  2,  4,  4,
+    6,  4,  6,  6,  8,  4,  6,  6,  8,  6,  8,  8,  10, 2,  4,  4,  6,  4,  6,
+    6,  8,  4,  6,  6,  8,  6,  8,  8,  10, 4,  6,  6,  8,  6,  8,  8,  10, 6,
+    8,  8,  10, 8,  10, 10, 12, 2,  4,  4,  6,  4,  6,  6,  8,  4,  6,  6,  8,
+    6,  8,  8,  10, 4,  6,  6,  8,  6,  8,  8,  10, 6,  8,  8,  10, 8,  10, 10,
+    12, 4,  6,  6,  8,  6,  8,  8,  10, 6,  8,  8,  10, 8,  10, 10, 12, 6,  8,
+    8,  10, 8,  10, 10, 12, 8,  10, 10, 12, 10, 12, 12, 14, 2,  4,  4,  6,  4,
+    6,  6,  8,  4,  6,  6,  8,  6,  8,  8,  10, 4,  6,  6,  8,  6,  8,  8,  10,
+    6,  8,  8,  10, 8,  10, 10, 12, 4,  6,  6,  8,  6,  8,  8,  10, 6,  8,  8,
+    10, 8,  10, 10, 12, 6,  8,  8,  10, 8,  10, 10, 12, 8,  10, 10, 12, 10, 12,
+    12, 14, 4,  6,  6,  8,  6,  8,  8,  10, 6,  8,  8,  10, 8,  10, 10, 12, 6,
+    8,  8,  10, 8,  10, 10, 12, 8,  10, 10, 12, 10, 12, 12, 14, 6,  8,  8,  10,
+    8,  10, 10, 12, 8,  10, 10, 12, 10, 12, 12, 14, 8,  10, 10, 12, 10, 12, 12,
+    14, 10, 12, 12, 14, 12, 14, 14, 16};
+
+SIMDJSON_DLLIMPORTEXPORT  const uint8_t pshufb_combine_table[272] = {
+    0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, 0x08, 0x09, 0x0a, 0x0b,
+    0x0c, 0x0d, 0x0e, 0x0f, 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x08,
+    0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, 0xff, 0x00, 0x01, 0x02, 0x03,
+    0x04, 0x05, 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, 0xff, 0xff,
+    0x00, 0x01, 0x02, 0x03, 0x04, 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e,
+    0x0f, 0xff, 0xff, 0xff, 0x00, 0x01, 0x02, 0x03, 0x08, 0x09, 0x0a, 0x0b,
+    0x0c, 0x0d, 0x0e, 0x0f, 0xff, 0xff, 0xff, 0xff, 0x00, 0x01, 0x02, 0x08,
+    0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, 0xff, 0xff, 0xff, 0xff, 0xff,
+    0x00, 0x01, 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, 0xff, 0xff,
+    0xff, 0xff, 0xff, 0xff, 0x00, 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e,
+    0x0f, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0x08, 0x09, 0x0a, 0x0b,
+    0x0c, 0x0d, 0x0e, 0x0f, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff, 0xff,
+};
+
+// 256 * 8 bytes = 2kB, easily fits in cache.
+SIMDJSON_DLLIMPORTEXPORT  const uint64_t thintable_epi8[256] = {
+    0x0706050403020100, 0x0007060504030201, 0x0007060504030200,
+    0x0000070605040302, 0x0007060504030100, 0x0000070605040301,
+    0x0000070605040300, 0x0000000706050403, 0x0007060504020100,
+    0x0000070605040201, 0x0000070605040200, 0x0000000706050402,
+    0x0000070605040100, 0x0000000706050401, 0x0000000706050400,
+    0x0000000007060504, 0x0007060503020100, 0x0000070605030201,
+    0x0000070605030200, 0x0000000706050302, 0x0000070605030100,
+    0x0000000706050301, 0x0000000706050300, 0x0000000007060503,
+    0x0000070605020100, 0x0000000706050201, 0x0000000706050200,
+    0x0000000007060502, 0x0000000706050100, 0x0000000007060501,
+    0x0000000007060500, 0x0000000000070605, 0x0007060403020100,
+    0x0000070604030201, 0x0000070604030200, 0x0000000706040302,
+    0x0000070604030100, 0x0000000706040301, 0x0000000706040300,
+    0x0000000007060403, 0x0000070604020100, 0x0000000706040201,
+    0x0000000706040200, 0x0000000007060402, 0x0000000706040100,
+    0x0000000007060401, 0x0000000007060400, 0x0000000000070604,
+    0x0000070603020100, 0x0000000706030201, 0x0000000706030200,
+    0x0000000007060302, 0x0000000706030100, 0x0000000007060301,
+    0x0000000007060300, 0x0000000000070603, 0x0000000706020100,
+    0x0000000007060201, 0x0000000007060200, 0x0000000000070602,
+    0x0000000007060100, 0x0000000000070601, 0x0000000000070600,
+    0x0000000000000706, 0x0007050403020100, 0x0000070504030201,
+    0x0000070504030200, 0x0000000705040302, 0x0000070504030100,
+    0x0000000705040301, 0x0000000705040300, 0x0000000007050403,
+    0x0000070504020100, 0x0000000705040201, 0x0000000705040200,
+    0x0000000007050402, 0x0000000705040100, 0x0000000007050401,
+    0x0000000007050400, 0x0000000000070504, 0x0000070503020100,
+    0x0000000705030201, 0x0000000705030200, 0x0000000007050302,
+    0x0000000705030100, 0x0000000007050301, 0x0000000007050300,
+    0x0000000000070503, 0x0000000705020100, 0x0000000007050201,
+    0x0000000007050200, 0x0000000000070502, 0x0000000007050100,
+    0x0000000000070501, 0x0000000000070500, 0x0000000000000705,
+    0x0000070403020100, 0x0000000704030201, 0x0000000704030200,
+    0x0000000007040302, 0x0000000704030100, 0x0000000007040301,
+    0x0000000007040300, 0x0000000000070403, 0x0000000704020100,
+    0x0000000007040201, 0x0000000007040200, 0x0000000000070402,
+    0x0000000007040100, 0x0000000000070401, 0x0000000000070400,
+    0x0000000000000704, 0x0000000703020100, 0x0000000007030201,
+    0x0000000007030200, 0x0000000000070302, 0x0000000007030100,
+    0x0000000000070301, 0x0000000000070300, 0x0000000000000703,
+    0x0000000007020100, 0x0000000000070201, 0x0000000000070200,
+    0x0000000000000702, 0x0000000000070100, 0x0000000000000701,
+    0x0000000000000700, 0x0000000000000007, 0x0006050403020100,
+    0x0000060504030201, 0x0000060504030200, 0x0000000605040302,
+    0x0000060504030100, 0x0000000605040301, 0x0000000605040300,
+    0x0000000006050403, 0x0000060504020100, 0x0000000605040201,
+    0x0000000605040200, 0x0000000006050402, 0x0000000605040100,
+    0x0000000006050401, 0x0000000006050400, 0x0000000000060504,
+    0x0000060503020100, 0x0000000605030201, 0x0000000605030200,
+    0x0000000006050302, 0x0000000605030100, 0x0000000006050301,
+    0x0000000006050300, 0x0000000000060503, 0x0000000605020100,
+    0x0000000006050201, 0x0000000006050200, 0x0000000000060502,
+    0x0000000006050100, 0x0000000000060501, 0x0000000000060500,
+    0x0000000000000605, 0x0000060403020100, 0x0000000604030201,
+    0x0000000604030200, 0x0000000006040302, 0x0000000604030100,
+    0x0000000006040301, 0x0000000006040300, 0x0000000000060403,
+    0x0000000604020100, 0x0000000006040201, 0x0000000006040200,
+    0x0000000000060402, 0x0000000006040100, 0x0000000000060401,
+    0x0000000000060400, 0x0000000000000604, 0x0000000603020100,
+    0x0000000006030201, 0x0000000006030200, 0x0000000000060302,
+    0x0000000006030100, 0x0000000000060301, 0x0000000000060300,
+    0x0000000000000603, 0x0000000006020100, 0x0000000000060201,
+    0x0000000000060200, 0x0000000000000602, 0x0000000000060100,
+    0x0000000000000601, 0x0000000000000600, 0x0000000000000006,
+    0x0000050403020100, 0x0000000504030201, 0x0000000504030200,
+    0x0000000005040302, 0x0000000504030100, 0x0000000005040301,
+    0x0000000005040300, 0x0000000000050403, 0x0000000504020100,
+    0x0000000005040201, 0x0000000005040200, 0x0000000000050402,
+    0x0000000005040100, 0x0000000000050401, 0x0000000000050400,
+    0x0000000000000504, 0x0000000503020100, 0x0000000005030201,
+    0x0000000005030200, 0x0000000000050302, 0x0000000005030100,
+    0x0000000000050301, 0x0000000000050300, 0x0000000000000503,
+    0x0000000005020100, 0x0000000000050201, 0x0000000000050200,
+    0x0000000000000502, 0x0000000000050100, 0x0000000000000501,
+    0x0000000000000500, 0x0000000000000005, 0x0000000403020100,
+    0x0000000004030201, 0x0000000004030200, 0x0000000000040302,
+    0x0000000004030100, 0x0000000000040301, 0x0000000000040300,
+    0x0000000000000403, 0x0000000004020100, 0x0000000000040201,
+    0x0000000000040200, 0x0000000000000402, 0x0000000000040100,
+    0x0000000000000401, 0x0000000000000400, 0x0000000000000004,
+    0x0000000003020100, 0x0000000000030201, 0x0000000000030200,
+    0x0000000000000302, 0x0000000000030100, 0x0000000000000301,
+    0x0000000000000300, 0x0000000000000003, 0x0000000000020100,
+    0x0000000000000201, 0x0000000000000200, 0x0000000000000002,
+    0x0000000000000100, 0x0000000000000001, 0x0000000000000000,
+    0x0000000000000000,
+}; //static uint64_t thintable_epi8[256]
+
+} // namespace internal
+} // namespace simdjson
+
+#endif //  SIMDJSON_IMPLEMENTATION_ARM64 || SIMDJSON_IMPLEMENTATION_ICELAKE || SIMDJSON_IMPLEMENTATION_HASWELL || SIMDJSON_IMPLEMENTATION_WESTMERE || SIMDJSON_IMPLEMENTATION_PPC64 || SIMDJSON_IMPLEMENTATION_LSX || SIMDJSON_IMPLEMENTATION_LASX
+
+#endif // SIMDJSON_SRC_SIMDPRUNE_TABLES_CPP
+/* end file internal/simdprune_tables.cpp */
+
+/* including simdjson/generic/dependencies.h: #include <simdjson/generic/dependencies.h> */
+/* begin file simdjson/generic/dependencies.h */
+#ifdef SIMDJSON_CONDITIONAL_INCLUDE
+#error simdjson/generic/dependencies.h must be included before defining SIMDJSON_CONDITIONAL_INCLUDE!
+#endif
+
+#ifndef SIMDJSON_GENERIC_DEPENDENCIES_H
+#define SIMDJSON_GENERIC_DEPENDENCIES_H
+
+// Internal headers needed for generics.
+// All includes referencing simdjson headers *not* under simdjson/generic must be here!
+// Otherwise, amalgamation will fail.
+/* skipped duplicate #include "simdjson/base.h" */
+/* including simdjson/implementation.h: #include "simdjson/implementation.h" */
+/* begin file simdjson/implementation.h */
+#ifndef SIMDJSON_IMPLEMENTATION_H
+#define SIMDJSON_IMPLEMENTATION_H
+
+/* including simdjson/internal/atomic_ptr.h: #include "simdjson/internal/atomic_ptr.h" */
+/* begin file simdjson/internal/atomic_ptr.h */
+#ifndef SIMDJSON_INTERNAL_ATOMIC_PTR_H
+#define SIMDJSON_INTERNAL_ATOMIC_PTR_H
+
+/* skipped duplicate #include "simdjson/base.h" */
+#include <atomic>
+
+namespace simdjson {
+namespace internal {
+
+template<typename T>
+class atomic_ptr {
+public:
+  atomic_ptr(T *_ptr) : ptr{_ptr} {}
+
+  operator const T*() const { return ptr.load(); }
+  const T& operator*() const { return *ptr; }
+  const T* operator->() const { return ptr.load(); }
+
+  operator T*() { return ptr.load(); }
+  T& operator*() { return *ptr; }
+  T* operator->() { return ptr.load(); }
+  atomic_ptr& operator=(T *_ptr) { ptr = _ptr; return *this; }
+
+private:
+  std::atomic<T*> ptr;
+};
+
+} // namespace internal
+} // namespace simdjson
+
+#endif // SIMDJSON_INTERNAL_ATOMIC_PTR_H
+/* end file simdjson/internal/atomic_ptr.h */
+/* including simdjson/internal/dom_parser_implementation.h: #include "simdjson/internal/dom_parser_implementation.h" */
+/* begin file simdjson/internal/dom_parser_implementation.h */
+#ifndef SIMDJSON_INTERNAL_DOM_PARSER_IMPLEMENTATION_H
+#define SIMDJSON_INTERNAL_DOM_PARSER_IMPLEMENTATION_H
+
+/* skipped duplicate #include "simdjson/base.h" */
+/* skipped duplicate #include "simdjson/error.h" */
+#include <memory>
+
+namespace simdjson {
+
+namespace dom {
+class document;
+} // namespace dom
+
+/**
+* This enum is used with the dom_parser_implementation::stage1 function.
+* 1) The regular mode expects a fully formed JSON document.
+* 2) The streaming_partial mode expects a possibly truncated
+* input within a stream on JSON documents.
+* 3) The stream_final mode allows us to truncate final
+* unterminated strings. It is useful in conjunction with streaming_partial.
+*/
+enum class stage1_mode { regular, streaming_partial, streaming_final};
+
+/**
+ * Returns true if mode == streaming_partial or mode == streaming_final
+ */
+inline bool is_streaming(stage1_mode mode) {
+  // performance note: it is probably faster to check that mode is different
+  // from regular than checking that it is either streaming_partial or streaming_final.
+  return (mode != stage1_mode::regular);
+  // return (mode == stage1_mode::streaming_partial || mode == stage1_mode::streaming_final);
+}
+
+
+namespace internal {
+
+
+/**
+ * An implementation of simdjson's DOM parser for a particular CPU architecture.
+ *
+ * This class is expected to be accessed only by pointer, and never move in memory (though the
+ * pointer can move).
+ */
+class dom_parser_implementation {
+public:
+
+  /**
+   * @private For internal implementation use
+   *
+   * Run a full JSON parse on a single document (stage1 + stage2).
+   *
+   * Guaranteed only to be called when capacity > document length.
+   *
+   * Overridden by each implementation.
+   *
+   * @param buf The json document to parse. *MUST* be allocated up to len + SIMDJSON_PADDING bytes.
+   * @param len The length of the json document.
+   * @return The error code, or SUCCESS if there was no error.
+   */
+  simdjson_warn_unused virtual error_code parse(const uint8_t *buf, size_t len, dom::document &doc) noexcept = 0;
+
+  /**
+   * @private For internal implementation use
+   *
+   * Stage 1 of the document parser.
+   *
+   * Guaranteed only to be called when capacity > document length.
+   *
+   * Overridden by each implementation.
+   *
+   * @param buf The json document to parse.
+   * @param len The length of the json document.
+   * @param streaming Whether this is being called by parser::parse_many.
+   * @return The error code, or SUCCESS if there was no error.
+   */
+  simdjson_warn_unused virtual error_code stage1(const uint8_t *buf, size_t len, stage1_mode streaming) noexcept = 0;
+
+  /**
+   * @private For internal implementation use
+   *
+   * Stage 2 of the document parser.
+   *
+   * Called after stage1().
+   *
+   * Overridden by each implementation.
+   *
+   * @param doc The document to output to.
+   * @return The error code, or SUCCESS if there was no error.
+   */
+  simdjson_warn_unused virtual error_code stage2(dom::document &doc) noexcept = 0;
+
+  /**
+   * @private For internal implementation use
+   *
+   * Stage 2 of the document parser for parser::parse_many.
+   *
+   * Guaranteed only to be called after stage1().
+   * Overridden by each implementation.
+   *
+   * @param doc The document to output to.
+   * @return The error code, SUCCESS if there was no error, or EMPTY if all documents have been parsed.
+   */
+  simdjson_warn_unused virtual error_code stage2_next(dom::document &doc) noexcept = 0;
+
+  /**
+   * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+   * must be an unescaped quote terminating the string. It returns the final output
+   * position as pointer. In case of error (e.g., the string has bad escaped codes),
+   * then null_ptr is returned. It is assumed that the output buffer is large
+   * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+   * SIMDJSON_PADDING bytes.
+   *
+   * Overridden by each implementation.
+   *
+   * @param str pointer to the beginning of a valid UTF-8 JSON string, must end with an unescaped quote.
+   * @param dst pointer to a destination buffer, it must point a region in memory of sufficient size.
+   * @param allow_replacement whether we allow a replacement character when the UTF-8 contains unmatched surrogate pairs.
+   * @return end of the of the written region (exclusive) or nullptr in case of error.
+   */
+  simdjson_warn_unused virtual uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) const noexcept = 0;
+
+  /**
+   * Unescape a NON-valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+   * must be an unescaped quote terminating the string. It returns the final output
+   * position as pointer. In case of error (e.g., the string has bad escaped codes),
+   * then null_ptr is returned. It is assumed that the output buffer is large
+   * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+   * SIMDJSON_PADDING bytes.
+   *
+   * Overridden by each implementation.
+   *
+   * @param str pointer to the beginning of a possibly invalid UTF-8 JSON string, must end with an unescaped quote.
+   * @param dst pointer to a destination buffer, it must point a region in memory of sufficient size.
+   * @return end of the of the written region (exclusive) or nullptr in case of error.
+   */
+  simdjson_warn_unused virtual uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept = 0;
+
+  /**
+   * Change the capacity of this parser.
+   *
+   * The capacity can never exceed SIMDJSON_MAXSIZE_BYTES (e.g., 4 GB)
+   * and an CAPACITY error is returned if it is attempted.
+   *
+   * Generally used for reallocation.
+   *
+   * @param capacity The new capacity.
+   * @param max_depth The new max_depth.
+   * @return The error code, or SUCCESS if there was no error.
+   */
+  virtual error_code set_capacity(size_t capacity) noexcept = 0;
+
+  /**
+   * Change the max depth of this parser.
+   *
+   * Generally used for reallocation.
+   *
+   * @param capacity The new capacity.
+   * @param max_depth The new max_depth.
+   * @return The error code, or SUCCESS if there was no error.
+   */
+  virtual error_code set_max_depth(size_t max_depth) noexcept = 0;
+
+  /**
+   * Deallocate this parser.
+   */
+  virtual ~dom_parser_implementation() = default;
+
+  /** Number of structural indices passed from stage 1 to stage 2 */
+  uint32_t n_structural_indexes{0};
+  /** Structural indices passed from stage 1 to stage 2 */
+  std::unique_ptr<uint32_t[]> structural_indexes{};
+  /** Next structural index to parse */
+  uint32_t next_structural_index{0};
+
+  /**
+   * The largest document this parser can support without reallocating.
+   *
+   * @return Current capacity, in bytes.
+   */
+  simdjson_pure simdjson_inline size_t capacity() const noexcept;
+
+  /**
+   * The maximum level of nested object and arrays supported by this parser.
+   *
+   * @return Maximum depth, in bytes.
+   */
+  simdjson_pure simdjson_inline size_t max_depth() const noexcept;
+
+  /**
+   * Ensure this parser has enough memory to process JSON documents up to `capacity` bytes in length
+   * and `max_depth` depth.
+   *
+   * @param capacity The new capacity.
+   * @param max_depth The new max_depth. Defaults to DEFAULT_MAX_DEPTH.
+   * @return The error, if there is one.
+   */
+  simdjson_warn_unused inline error_code allocate(size_t capacity, size_t max_depth) noexcept;
+
+
+protected:
+  /**
+   * The maximum document length this parser supports.
+   *
+   * Buffers are large enough to handle any document up to this length.
+   */
+  size_t _capacity{0};
+
+  /**
+   * The maximum depth (number of nested objects and arrays) supported by this parser.
+   *
+   * Defaults to DEFAULT_MAX_DEPTH.
+   */
+  size_t _max_depth{0};
+
+public:
+  /** Whether to store big integers as strings instead of returning BIGINT_ERROR */
+  bool _number_as_string{false};
+
+protected:
+
+  // Declaring these so that subclasses can use them to implement their constructors.
+  simdjson_inline dom_parser_implementation() noexcept;
+  simdjson_inline dom_parser_implementation(dom_parser_implementation &&other) noexcept;
+  simdjson_inline dom_parser_implementation &operator=(dom_parser_implementation &&other) noexcept;
+
+  simdjson_inline dom_parser_implementation(const dom_parser_implementation &) noexcept = delete;
+  simdjson_inline dom_parser_implementation &operator=(const dom_parser_implementation &other) noexcept = delete;
+}; // class dom_parser_implementation
+
+simdjson_inline dom_parser_implementation::dom_parser_implementation() noexcept = default;
+simdjson_inline dom_parser_implementation::dom_parser_implementation(dom_parser_implementation &&other) noexcept = default;
+simdjson_inline dom_parser_implementation &dom_parser_implementation::operator=(dom_parser_implementation &&other) noexcept = default;
+
+simdjson_pure simdjson_inline size_t dom_parser_implementation::capacity() const noexcept {
+  return _capacity;
+}
+
+simdjson_pure simdjson_inline size_t dom_parser_implementation::max_depth() const noexcept {
+  return _max_depth;
+}
+
+simdjson_warn_unused
+inline error_code dom_parser_implementation::allocate(size_t capacity, size_t max_depth) noexcept {
+  if (this->max_depth() != max_depth) {
+    error_code err = set_max_depth(max_depth);
+    if (err) { return err; }
+  }
+  if (_capacity != capacity) {
+    error_code err = set_capacity(capacity);
+    if (err) { return err; }
+  }
+  return SUCCESS;
+}
+
+} // namespace internal
+} // namespace simdjson
+
+#endif // SIMDJSON_INTERNAL_DOM_PARSER_IMPLEMENTATION_H
+/* end file simdjson/internal/dom_parser_implementation.h */
+
+#include <memory>
+
+namespace simdjson {
+
+/**
+ * Validate the UTF-8 string.
+ *
+ * @param buf the string to validate.
+ * @param len the length of the string in bytes.
+ * @return true if the string is valid UTF-8.
+ */
+simdjson_warn_unused bool validate_utf8(const char * buf, size_t len) noexcept;
+/**
+ * Validate the UTF-8 string.
+ *
+ * @param sv the string_view to validate.
+ * @return true if the string is valid UTF-8.
+ */
+simdjson_inline simdjson_warn_unused bool validate_utf8(const std::string_view sv) noexcept {
+  return validate_utf8(sv.data(), sv.size());
+}
+
+/**
+ * Validate the UTF-8 string.
+ *
+ * @param p the string to validate.
+ * @return true if the string is valid UTF-8.
+ */
+simdjson_inline simdjson_warn_unused bool validate_utf8(const std::string& s) noexcept {
+  return validate_utf8(s.data(), s.size());
+}
+
+/**
+ * An implementation of simdjson for a particular CPU architecture.
+ *
+ * Also used to maintain the currently active implementation. The active implementation is
+ * automatically initialized on first use to the most advanced implementation supported by the host.
+ */
+class implementation {
+public:
+
+  /**
+   * The name of this implementation.
+   *
+   *     const implementation *impl = simdjson::get_active_implementation();
+   *     cout << "simdjson is optimized for " << impl->name() << "(" << impl->description() << ")" << endl;
+   *
+   * @return the name of the implementation, e.g. "haswell", "westmere", "arm64".
+   */
+  virtual std::string name() const { return std::string(_name); }
+
+  /**
+   * The description of this implementation.
+   *
+   *     const implementation *impl = simdjson::get_active_implementation();
+   *     cout << "simdjson is optimized for " << impl->name() << "(" << impl->description() << ")" << endl;
+   *
+   * @return the description of the implementation, e.g. "Intel/AMD AVX2", "Intel/AMD SSE4.2", "ARM NEON".
+   */
+  virtual std::string description() const { return std::string(_description); }
+
+  /**
+   * The instruction sets this implementation is compiled against
+   * and the current CPU match. This function may poll the current CPU/system
+   * and should therefore not be called too often if performance is a concern.
+   *
+   * @return true if the implementation can be safely used on the current system (determined at runtime).
+   */
+  bool supported_by_runtime_system() const;
+
+  /**
+   * @private For internal implementation use
+   *
+   * The instruction sets this implementation is compiled against.
+   *
+   * @return a mask of all required `internal::instruction_set::` values.
+   */
+  virtual uint32_t required_instruction_sets() const { return _required_instruction_sets; }
+
+  /**
+   * @private For internal implementation use
+   *
+   *     const implementation *impl = simdjson::get_active_implementation();
+   *     cout << "simdjson is optimized for " << impl->name() << "(" << impl->description() << ")" << endl;
+   *
+   * @param capacity The largest document that will be passed to the parser.
+   * @param max_depth The maximum JSON object/array nesting this parser is expected to handle.
+   * @param dst The place to put the resulting parser implementation.
+   * @return the error code, or SUCCESS if there was no error.
+   */
+  virtual error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_depth,
+    std::unique_ptr<internal::dom_parser_implementation> &dst
+  ) const noexcept = 0;
+
+  /**
+   * @private For internal implementation use
+   *
+   * Minify the input string assuming that it represents a JSON string, does not parse or validate.
+   *
+   * Overridden by each implementation.
+   *
+   * @param buf the json document to minify.
+   * @param len the length of the json document.
+   * @param dst the buffer to write the minified document to. *MUST* be allocated up to len + SIMDJSON_PADDING bytes.
+   * @param dst_len the number of bytes written. Output only.
+   * @return the error code, or SUCCESS if there was no error.
+   */
+  simdjson_warn_unused virtual error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept = 0;
+
+
+  /**
+   * Validate the UTF-8 string.
+   *
+   * Overridden by each implementation.
+   *
+   * @param buf the string to validate.
+   * @param len the length of the string in bytes.
+   * @return true if and only if the string is valid UTF-8.
+   */
+  simdjson_warn_unused virtual bool validate_utf8(const char *buf, size_t len) const noexcept = 0;
+
+protected:
+  /** @private Construct an implementation with the given name and description. For subclasses. */
+  simdjson_inline implementation(
+    std::string_view name,
+    std::string_view description,
+    uint32_t required_instruction_sets
+  ) :
+    _name(name),
+    _description(description),
+    _required_instruction_sets(required_instruction_sets)
+  {
+  }
+protected:
+  ~implementation() = default;
+
+private:
+  /**
+   * The name of this implementation.
+   */
+  std::string_view _name;
+
+  /**
+   * The description of this implementation.
+   */
+  std::string_view _description;
+
+  /**
+   * Instruction sets required for this implementation.
+   */
+  const uint32_t _required_instruction_sets;
+};
+
+/** @private */
+namespace internal {
+
+/**
+ * The list of available implementations compiled into simdjson.
+ */
+class available_implementation_list {
+public:
+  /** Get the list of available implementations compiled into simdjson */
+  simdjson_inline available_implementation_list() {}
+  /** Number of implementations */
+  size_t size() const noexcept;
+  /** STL const begin() iterator */
+  const implementation * const *begin() const noexcept;
+  /** STL const end() iterator */
+  const implementation * const *end() const noexcept;
+
+  /**
+   * Get the implementation with the given name.
+   *
+   * Case sensitive.
+   *
+   *     const implementation *impl = simdjson::get_available_implementations()["westmere"];
+   *     if (!impl) { exit(1); }
+   *     if (!imp->supported_by_runtime_system()) { exit(1); }
+   *     simdjson::get_active_implementation() = impl;
+   *
+   * @param name the implementation to find, e.g. "westmere", "haswell", "arm64"
+   * @return the implementation, or nullptr if the parse failed.
+   */
+  const implementation * operator[](const std::string_view &name) const noexcept {
+    for (const implementation * impl : *this) {
+      if (impl->name() == name) { return impl; }
+    }
+    return nullptr;
+  }
+
+  /**
+   * Detect the most advanced implementation supported by the current host.
+   *
+   * This is used to initialize the implementation on startup.
+   *
+   *     const implementation *impl = simdjson::available_implementation::detect_best_supported();
+   *     simdjson::get_active_implementation() = impl;
+   *
+   * @return the most advanced supported implementation for the current host, or an
+   *         implementation that returns UNSUPPORTED_ARCHITECTURE if there is no supported
+   *         implementation. Will never return nullptr.
+   */
+  const implementation *detect_best_supported() const noexcept;
+};
+
+} // namespace internal
+
+/**
+ * The list of available implementations compiled into simdjson.
+ */
+extern SIMDJSON_DLLIMPORTEXPORT const internal::available_implementation_list& get_available_implementations();
+
+/**
+  * The active implementation.
+  *
+  * Automatically initialized on first use to the most advanced implementation supported by this hardware.
+  */
+extern SIMDJSON_DLLIMPORTEXPORT internal::atomic_ptr<const implementation>& get_active_implementation();
+
+} // namespace simdjson
+
+#endif // SIMDJSON_IMPLEMENTATION_H
+/* end file simdjson/implementation.h */
+/* skipped duplicate #include "simdjson/implementation_detection.h" */
+/* including simdjson/internal/instruction_set.h: #include "simdjson/internal/instruction_set.h" */
+/* begin file simdjson/internal/instruction_set.h */
+/* From
+https://github.com/endorno/pytorch/blob/master/torch/lib/TH/generic/simd/simd.h
+Highly modified.
+
+Copyright (c) 2016-     Facebook, Inc            (Adam Paszke)
+Copyright (c) 2014-     Facebook, Inc            (Soumith Chintala)
+Copyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert)
+Copyright (c) 2012-2014 Deepmind Technologies    (Koray Kavukcuoglu)
+Copyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu)
+Copyright (c) 2011-2013 NYU                      (Clement Farabet)
+Copyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou,
+Iain Melvin, Jason Weston) Copyright (c) 2006      Idiap Research Institute
+(Samy Bengio) Copyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert,
+Samy Bengio, Johnny Mariethoz)
+
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+3. Neither the names of Facebook, Deepmind Technologies, NYU, NEC Laboratories
+America and IDIAP Research Institute nor the names of its contributors may be
+   used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef SIMDJSON_INTERNAL_INSTRUCTION_SET_H
+#define SIMDJSON_INTERNAL_INSTRUCTION_SET_H
+
+namespace simdjson {
+namespace internal {
+
+enum instruction_set {
+  DEFAULT = 0x0,
+  NEON = 0x1,
+  AVX2 = 0x4,
+  SSE42 = 0x8,
+  PCLMULQDQ = 0x10,
+  BMI1 = 0x20,
+  BMI2 = 0x40,
+  ALTIVEC = 0x80,
+  AVX512F = 0x100,
+  AVX512DQ = 0x200,
+  AVX512IFMA = 0x400,
+  AVX512PF = 0x800,
+  AVX512ER = 0x1000,
+  AVX512CD = 0x2000,
+  AVX512BW = 0x4000,
+  AVX512VL = 0x8000,
+  AVX512VBMI2 = 0x10000,
+  LSX = 0x20000,
+  LASX = 0x40000,
+  //RVV = 0x80000,
+  RVV_VLS = 0x100000,
+};
+
+} // namespace internal
+} // namespace simdjson
+
+#endif // SIMDJSON_INTERNAL_INSTRUCTION_SET_H
+/* end file simdjson/internal/instruction_set.h */
+/* skipped duplicate #include "simdjson/internal/dom_parser_implementation.h" */
+/* skipped duplicate #include "simdjson/internal/jsoncharutils_tables.h" */
+/* skipped duplicate #include "simdjson/internal/numberparsing_tables.h" */
+/* including simdjson/internal/simdprune_tables.h: #include "simdjson/internal/simdprune_tables.h" */
+/* begin file simdjson/internal/simdprune_tables.h */
+#ifndef SIMDJSON_INTERNAL_SIMDPRUNE_TABLES_H
+#define SIMDJSON_INTERNAL_SIMDPRUNE_TABLES_H
+
+/* skipped duplicate #include "simdjson/base.h" */
+
+#include <cstdint>
+
+namespace simdjson { // table modified and copied from
+namespace internal { // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetTable
+
+extern SIMDJSON_DLLIMPORTEXPORT const unsigned char BitsSetTable256mul2[256];
+
+extern SIMDJSON_DLLIMPORTEXPORT const uint8_t pshufb_combine_table[272];
+
+// 256 * 8 bytes = 2kB, easily fits in cache.
+extern SIMDJSON_DLLIMPORTEXPORT const uint64_t thintable_epi8[256];
+
+} // namespace internal
+} // namespace simdjson
+
+#endif // SIMDJSON_INTERNAL_SIMDPRUNE_TABLES_H
+/* end file simdjson/internal/simdprune_tables.h */
+#endif // SIMDJSON_GENERIC_DEPENDENCIES_H
+/* end file simdjson/generic/dependencies.h */
+/* including generic/dependencies.h: #include <generic/dependencies.h> */
+/* begin file generic/dependencies.h */
+#ifdef SIMDJSON_CONDITIONAL_INCLUDE
+#error generic/dependencies.h must be included before defining SIMDJSON_CONDITIONAL_INCLUDE!
+#endif
+
+#ifndef SIMDJSON_SRC_GENERIC_DEPENDENCIES_H
+#define SIMDJSON_SRC_GENERIC_DEPENDENCIES_H
+
+/* skipped duplicate #include <base.h> */
+
+#endif // SIMDJSON_SRC_GENERIC_DEPENDENCIES_H
+/* end file generic/dependencies.h */
+/* including generic/stage1/dependencies.h: #include <generic/stage1/dependencies.h> */
+/* begin file generic/stage1/dependencies.h */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_DEPENDENCIES_H
+#define SIMDJSON_SRC_GENERIC_STAGE1_DEPENDENCIES_H
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_DEPENDENCIES_H
+/* end file generic/stage1/dependencies.h */
+/* including generic/stage2/dependencies.h: #include <generic/stage2/dependencies.h> */
+/* begin file generic/stage2/dependencies.h */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_DEPENDENCIES_H
+#define SIMDJSON_SRC_GENERIC_STAGE2_DEPENDENCIES_H
+
+/* including simdjson/dom/document.h: #include <simdjson/dom/document.h> */
+/* begin file simdjson/dom/document.h */
+#ifndef SIMDJSON_DOM_DOCUMENT_H
+#define SIMDJSON_DOM_DOCUMENT_H
+
+/* including simdjson/dom/base.h: #include "simdjson/dom/base.h" */
+/* begin file simdjson/dom/base.h */
+#ifndef SIMDJSON_DOM_BASE_H
+#define SIMDJSON_DOM_BASE_H
+
+/* skipped duplicate #include "simdjson/base.h" */
+
+namespace simdjson {
+
+/**
+ * @brief A DOM API on top of the simdjson parser.
+ */
+namespace dom {
+
+/** The default batch size for parser.parse_many() and parser.load_many() */
+static constexpr size_t DEFAULT_BATCH_SIZE = 1000000;
+/**
+ * Some adversary might try to set the batch size to 0 or 1, which might cause problems.
+ * We set a minimum of 32B since anything else is highly likely to be an error. In practice,
+ * most users will want a much larger batch size.
+ *
+ * All non-negative MINIMAL_BATCH_SIZE values should be 'safe' except that, obviously, no JSON
+ * document can ever span 0 or 1 byte and that very large values would create memory allocation issues.
+ */
+static constexpr size_t MINIMAL_BATCH_SIZE = 32;
+
+/**
+ * It is wasteful to allocate memory for tiny documents (e.g., 4 bytes).
+ */
+static constexpr size_t MINIMAL_DOCUMENT_CAPACITY = 32;
+
+class array;
+class document;
+class document_stream;
+class element;
+class key_value_pair;
+class object;
+class parser;
+
+#ifdef SIMDJSON_THREADS_ENABLED
+struct stage1_worker;
+#endif // SIMDJSON_THREADS_ENABLED
+
+} // namespace dom
+
+namespace internal {
+
+template<typename T>
+class string_builder;
+class tape_ref;
+
+} // namespace internal
+
+} // namespace simdjson
+
+#endif // SIMDJSON_DOM_BASE_H
+/* end file simdjson/dom/base.h */
+
+#include <memory>
+
+namespace simdjson {
+namespace dom {
+
+/**
+ * A parsed JSON document.
+ *
+ * This class cannot be copied, only moved, to avoid unintended allocations.
+ */
+class document {
+public:
+  /**
+   * Create a document container with zero capacity.
+   *
+   * The parser will allocate capacity as needed.
+   */
+  document() noexcept = default;
+  ~document() noexcept = default;
+
+  /**
+   * Take another document's buffers.
+   *
+   * @param other The document to take. Its capacity is zeroed and it is invalidated.
+   */
+  document(document &&other) noexcept = default;
+  /** @private */
+  document(const document &) = delete; // Disallow copying
+  /**
+   * Take another document's buffers.
+   *
+   * @param other The document to take. Its capacity is zeroed.
+   */
+  document &operator=(document &&other) noexcept = default;
+  /** @private */
+  document &operator=(const document &) = delete; // Disallow copying
+
+  /**
+   * Get the root element of this document as a JSON array.
+   */
+  element root() const noexcept;
+
+  /**
+   * @private Dump the raw tape for debugging.
+   *
+   * @param os the stream to output to.
+   * @return false if the tape is likely wrong (e.g., you did not parse a valid JSON).
+   */
+  bool dump_raw_tape(std::ostream &os) const noexcept;
+
+  /** @private Structural values. */
+  std::unique_ptr<uint64_t[]> tape{};
+
+  /** @private String values.
+   *
+   * Should be at least byte_capacity.
+   */
+  std::unique_ptr<uint8_t[]> string_buf{};
+  /** @private Allocate memory to support
+   * input JSON documents of up to len bytes.
+   *
+   * When calling this function, you lose
+   * all the data.
+   *
+   * The memory allocation is strict: you
+   * can you use this function to increase
+   * or lower the amount of allocated memory.
+   * Passing zero clears the memory.
+   */
+  error_code allocate(size_t len) noexcept;
+  /** @private Capacity in bytes, in terms
+   * of how many bytes of input JSON we can
+   * support.
+   */
+  size_t capacity() const noexcept;
+
+
+private:
+  size_t allocated_capacity{0};
+  friend class parser;
+}; // class document
+
+} // namespace dom
+} // namespace simdjson
+
+#endif // SIMDJSON_DOM_DOCUMENT_H
+/* end file simdjson/dom/document.h */
+/* including simdjson/internal/tape_type.h: #include <simdjson/internal/tape_type.h> */
+/* begin file simdjson/internal/tape_type.h */
+#ifndef SIMDJSON_INTERNAL_TAPE_TYPE_H
+#define SIMDJSON_INTERNAL_TAPE_TYPE_H
+
+namespace simdjson {
+namespace internal {
+
+/**
+ * The possible types in the tape.
+ */
+enum class tape_type {
+  ROOT = 'r',
+  START_ARRAY = '[',
+  START_OBJECT = '{',
+  END_ARRAY = ']',
+  END_OBJECT = '}',
+  STRING = '"',
+  INT64 = 'l',
+  UINT64 = 'u',
+  DOUBLE = 'd',
+  TRUE_VALUE = 't',
+  FALSE_VALUE = 'f',
+  NULL_VALUE = 'n',
+  BIGINT = 'Z'  // Big integer stored as string in string buffer
+}; // enum class tape_type
+
+} // namespace internal
+} // namespace simdjson
+
+#endif // SIMDJSON_INTERNAL_TAPE_TYPE_H
+/* end file simdjson/internal/tape_type.h */
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_DEPENDENCIES_H
+/* end file generic/stage2/dependencies.h */
+
+/* including implementation.cpp: #include <implementation.cpp> */
+/* begin file implementation.cpp */
+#ifndef SIMDJSON_SRC_IMPLEMENTATION_CPP
+#define SIMDJSON_SRC_IMPLEMENTATION_CPP
+
+/* skipped duplicate #include <base.h> */
+/* skipped duplicate #include <simdjson/generic/dependencies.h> */
+/* skipped duplicate #include <simdjson/implementation.h> */
+/* including internal/isadetection.h: #include <internal/isadetection.h> */
+/* begin file internal/isadetection.h */
+/* From
+https://github.com/endorno/pytorch/blob/master/torch/lib/TH/generic/simd/simd.h
+Highly modified.
+
+Copyright (c) 2016-     Facebook, Inc            (Adam Paszke)
+Copyright (c) 2014-     Facebook, Inc            (Soumith Chintala)
+Copyright (c) 2011-2014 Idiap Research Institute (Ronan Collobert)
+Copyright (c) 2012-2014 Deepmind Technologies    (Koray Kavukcuoglu)
+Copyright (c) 2011-2012 NEC Laboratories America (Koray Kavukcuoglu)
+Copyright (c) 2011-2013 NYU                      (Clement Farabet)
+Copyright (c) 2006-2010 NEC Laboratories America (Ronan Collobert, Leon Bottou,
+Iain Melvin, Jason Weston) Copyright (c) 2006      Idiap Research Institute
+(Samy Bengio) Copyright (c) 2001-2004 Idiap Research Institute (Ronan Collobert,
+Samy Bengio, Johnny Mariethoz)
+
+All rights reserved.
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright
+   notice, this list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright
+   notice, this list of conditions and the following disclaimer in the
+   documentation and/or other materials provided with the distribution.
+
+3. Neither the names of Facebook, Deepmind Technologies, NYU, NEC Laboratories
+America and IDIAP Research Institute nor the names of its contributors may be
+   used to endorse or promote products derived from this software without
+   specific prior written permission.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE
+LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
+CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
+SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
+CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
+ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
+POSSIBILITY OF SUCH DAMAGE.
+*/
+
+#ifndef SIMDJSON_INTERNAL_ISADETECTION_H
+#define SIMDJSON_INTERNAL_ISADETECTION_H
+
+/* skipped duplicate #include "simdjson/internal/instruction_set.h" */
+
+#include <cstdint>
+#include <cstdlib>
+#if defined(_MSC_VER)
+#include <intrin.h>
+#elif defined(HAVE_GCC_GET_CPUID) && defined(USE_GCC_GET_CPUID)
+#include <cpuid.h>
+#endif
+#if defined(__loongarch__) && defined(__linux__)
+  #include <sys/auxv.h>
+#endif
+
+namespace simdjson {
+namespace internal {
+
+#if defined(__PPC64__)
+
+static inline uint32_t detect_supported_architectures() {
+  return instruction_set::ALTIVEC;
+}
+
+#elif defined(__aarch64__) || defined(_M_ARM64) || defined(_M_ARM64EC)
+
+static inline uint32_t detect_supported_architectures() {
+  return instruction_set::NEON;
+}
+
+#elif defined(__x86_64__) || defined(_M_AMD64) // x64
+
+
+namespace {
+// Can be found on Intel ISA Reference for CPUID
+constexpr uint32_t cpuid_avx2_bit = 1 << 5;         ///< @private Bit 5 of EBX for EAX=0x7
+constexpr uint32_t cpuid_bmi1_bit = 1 << 3;         ///< @private bit 3 of EBX for EAX=0x7
+constexpr uint32_t cpuid_bmi2_bit = 1 << 8;         ///< @private bit 8 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512f_bit = 1 << 16;     ///< @private bit 16 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512dq_bit = 1 << 17;    ///< @private bit 17 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512ifma_bit = 1 << 21;  ///< @private bit 21 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512pf_bit = 1 << 26;    ///< @private bit 26 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512er_bit = 1 << 27;    ///< @private bit 27 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512cd_bit = 1 << 28;    ///< @private bit 28 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512bw_bit = 1 << 30;    ///< @private bit 30 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512vl_bit = 1U << 31;    ///< @private bit 31 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512vbmi2_bit = 1 << 6;  ///< @private bit 6 of ECX for EAX=0x7
+constexpr uint64_t cpuid_avx256_saved = uint64_t(1) << 2; ///< @private bit 2 = AVX
+constexpr uint64_t cpuid_avx512_saved = uint64_t(7) << 5; ///< @private bits 5,6,7 = opmask, ZMM_hi256, hi16_ZMM
+constexpr uint32_t cpuid_sse42_bit = 1 << 20;       ///< @private bit 20 of ECX for EAX=0x1
+constexpr uint32_t cpuid_osxsave = (uint32_t(1) << 26) | (uint32_t(1) << 27); ///< @private bits 26+27 of ECX for EAX=0x1
+constexpr uint32_t cpuid_pclmulqdq_bit = 1 << 1;    ///< @private bit  1 of ECX for EAX=0x1
+}
+
+
+
+static inline void cpuid(uint32_t *eax, uint32_t *ebx, uint32_t *ecx,
+                         uint32_t *edx) {
+#if defined(_MSC_VER)
+  int cpu_info[4];
+  __cpuidex(cpu_info, *eax, *ecx);
+  *eax = cpu_info[0];
+  *ebx = cpu_info[1];
+  *ecx = cpu_info[2];
+  *edx = cpu_info[3];
+#elif defined(HAVE_GCC_GET_CPUID) && defined(USE_GCC_GET_CPUID)
+  uint32_t level = *eax;
+  __get_cpuid(level, eax, ebx, ecx, edx);
+#else
+  uint32_t a = *eax, b, c = *ecx, d;
+  asm volatile("cpuid\n\t" : "+a"(a), "=b"(b), "+c"(c), "=d"(d));
+  *eax = a;
+  *ebx = b;
+  *ecx = c;
+  *edx = d;
+#endif
+}
+
+
+static inline uint64_t xgetbv() {
+#if defined(_MSC_VER)
+  return _xgetbv(0);
+#else
+  uint32_t xcr0_lo, xcr0_hi;
+  asm volatile("xgetbv\n\t" : "=a" (xcr0_lo), "=d" (xcr0_hi) : "c" (0));
+  return xcr0_lo | (uint64_t(xcr0_hi) << 32);
+#endif
+}
+
+static inline uint32_t detect_supported_architectures() {
+  uint32_t eax, ebx, ecx, edx;
+  uint32_t host_isa = 0x0;
+
+  // EBX for EAX=0x1
+  eax = 0x1;
+  ecx = 0x0;
+  cpuid(&eax, &ebx, &ecx, &edx);
+
+  if (ecx & cpuid_sse42_bit) {
+    host_isa |= instruction_set::SSE42;
+  } else {
+    return host_isa; // everything after is redundant
+  }
+
+  if (ecx & cpuid_pclmulqdq_bit) {
+    host_isa |= instruction_set::PCLMULQDQ;
+  }
+
+
+  if ((ecx & cpuid_osxsave) != cpuid_osxsave) {
+    return host_isa;
+  }
+
+  // xgetbv for checking if the OS saves registers
+  uint64_t xcr0 = xgetbv();
+
+  if ((xcr0 & cpuid_avx256_saved) == 0) {
+    return host_isa;
+  }
+
+  // ECX for EAX=0x7
+  eax = 0x7;
+  ecx = 0x0;
+  cpuid(&eax, &ebx, &ecx, &edx);
+  if (ebx & cpuid_avx2_bit) {
+    host_isa |= instruction_set::AVX2;
+  }
+  if (ebx & cpuid_bmi1_bit) {
+    host_isa |= instruction_set::BMI1;
+  }
+
+  if (ebx & cpuid_bmi2_bit) {
+    host_isa |= instruction_set::BMI2;
+  }
+
+  if (!((xcr0 & cpuid_avx512_saved) == cpuid_avx512_saved)) {
+     return host_isa;
+  }
+
+  if (ebx & cpuid_avx512f_bit) {
+    host_isa |= instruction_set::AVX512F;
+  }
+
+  if (ebx & cpuid_avx512dq_bit) {
+    host_isa |= instruction_set::AVX512DQ;
+  }
+
+  if (ebx & cpuid_avx512ifma_bit) {
+    host_isa |= instruction_set::AVX512IFMA;
+  }
+
+  if (ebx & cpuid_avx512pf_bit) {
+    host_isa |= instruction_set::AVX512PF;
+  }
+
+  if (ebx & cpuid_avx512er_bit) {
+    host_isa |= instruction_set::AVX512ER;
+  }
+
+  if (ebx & cpuid_avx512cd_bit) {
+    host_isa |= instruction_set::AVX512CD;
+  }
+
+  if (ebx & cpuid_avx512bw_bit) {
+    host_isa |= instruction_set::AVX512BW;
+  }
+
+  if (ebx & cpuid_avx512vl_bit) {
+    host_isa |= instruction_set::AVX512VL;
+  }
+
+  if (ecx & cpuid_avx512vbmi2_bit) {
+    host_isa |= instruction_set::AVX512VBMI2;
+  }
+
+  return host_isa;
+}
+
+#elif defined(__loongarch__)
+
+static inline uint32_t detect_supported_architectures() {
+  uint32_t host_isa = instruction_set::DEFAULT;
+  #if defined(__linux__)
+  uint64_t hwcap = 0;
+  hwcap = getauxval(AT_HWCAP);
+  if (hwcap & HWCAP_LOONGARCH_LSX) {
+    host_isa |= instruction_set::LSX;
+  }
+  if (hwcap & HWCAP_LOONGARCH_LASX) {
+    host_isa |= instruction_set::LASX;
+  }
+  #endif
+  return host_isa;
+}
+
+#elif SIMDJSON_IS_RISCV64
+
+static inline uint32_t detect_supported_architectures() {
+  uint32_t host_isa = instruction_set::DEFAULT;
+#if SIMDJSON_IS_RVV_VLS
+  host_isa |= instruction_set::RVV_VLS;
+#endif
+  return host_isa;
+}
+
+#else // fallback
+
+
+static inline uint32_t detect_supported_architectures() {
+  return instruction_set::DEFAULT;
+}
+
+
+#endif // end SIMD extension detection code
+
+} // namespace internal
+} // namespace simdjson
+
+#endif // SIMDJSON_INTERNAL_ISADETECTION_H
+/* end file internal/isadetection.h */
+
+#include <initializer_list>
+#include <type_traits>
+
+namespace simdjson {
+
+bool implementation::supported_by_runtime_system() const {
+  uint32_t required_instruction_sets = this->required_instruction_sets();
+  uint32_t supported_instruction_sets = internal::detect_supported_architectures();
+  return ((supported_instruction_sets & required_instruction_sets) == required_instruction_sets);
+}
+
+} // namespace simdjson
+
+/* defining SIMDJSON_CONDITIONAL_INCLUDE */
+#define SIMDJSON_CONDITIONAL_INCLUDE
+
+#if SIMDJSON_IMPLEMENTATION_ARM64
+/* including simdjson/arm64/implementation.h: #include <simdjson/arm64/implementation.h> */
+/* begin file simdjson/arm64/implementation.h */
+#ifndef SIMDJSON_ARM64_IMPLEMENTATION_H
+#define SIMDJSON_ARM64_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/instruction_set.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation("arm64", "ARM NEON", internal::instruction_set::NEON) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_ARM64_IMPLEMENTATION_H
+/* end file simdjson/arm64/implementation.h */
+namespace simdjson {
+namespace internal {
+static const arm64::implementation* get_arm64_singleton() {
+  static const arm64::implementation arm64_singleton{};
+  return &arm64_singleton;
+}
+} // namespace internal
+} // namespace simdjson
+#endif // SIMDJSON_IMPLEMENTATION_ARM64
+
+#if SIMDJSON_IMPLEMENTATION_FALLBACK
+/* including simdjson/fallback/implementation.h: #include <simdjson/fallback/implementation.h> */
+/* begin file simdjson/fallback/implementation.h */
+#ifndef SIMDJSON_FALLBACK_IMPLEMENTATION_H
+#define SIMDJSON_FALLBACK_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace fallback {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation(
+      "fallback",
+      "Generic fallback implementation",
+      0
+  ) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<simdjson::internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_FALLBACK_IMPLEMENTATION_H
+/* end file simdjson/fallback/implementation.h */
+namespace simdjson {
+namespace internal {
+static const fallback::implementation* get_fallback_singleton() {
+  static const fallback::implementation fallback_singleton{};
+  return &fallback_singleton;
+}
+} // namespace internal
+} // namespace simdjson
+#endif // SIMDJSON_IMPLEMENTATION_FALLBACK
+
+
+#if SIMDJSON_IMPLEMENTATION_HASWELL
+/* including simdjson/haswell/implementation.h: #include <simdjson/haswell/implementation.h> */
+/* begin file simdjson/haswell/implementation.h */
+#ifndef SIMDJSON_HASWELL_IMPLEMENTATION_H
+#define SIMDJSON_HASWELL_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/instruction_set.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// The constructor may be executed on any host, so we take care not to use SIMDJSON_TARGET_HASWELL
+namespace simdjson {
+namespace haswell {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation(
+      "haswell",
+      "Intel/AMD AVX2",
+      internal::instruction_set::AVX2 | internal::instruction_set::PCLMULQDQ | internal::instruction_set::BMI1 | internal::instruction_set::BMI2
+  ) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_HASWELL_IMPLEMENTATION_H
+/* end file simdjson/haswell/implementation.h */
+namespace simdjson {
+namespace internal {
+static const haswell::implementation* get_haswell_singleton() {
+  static const haswell::implementation haswell_singleton{};
+  return &haswell_singleton;
+}
+} // namespace internal
+} // namespace simdjson
+#endif
+
+#if SIMDJSON_IMPLEMENTATION_ICELAKE
+/* including simdjson/icelake/implementation.h: #include <simdjson/icelake/implementation.h> */
+/* begin file simdjson/icelake/implementation.h */
+#ifndef SIMDJSON_ICELAKE_IMPLEMENTATION_H
+#define SIMDJSON_ICELAKE_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/instruction_set.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// The constructor may be executed on any host, so we take care not to use SIMDJSON_TARGET_ICELAKE
+namespace simdjson {
+namespace icelake {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation(
+      "icelake",
+      "Intel/AMD AVX512",
+      internal::instruction_set::AVX2 | internal::instruction_set::PCLMULQDQ | internal::instruction_set::BMI1 | internal::instruction_set::BMI2 | internal::instruction_set::AVX512F | internal::instruction_set::AVX512DQ | internal::instruction_set::AVX512CD | internal::instruction_set::AVX512BW | internal::instruction_set::AVX512VL | internal::instruction_set::AVX512VBMI2
+  ) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_ICELAKE_IMPLEMENTATION_H
+/* end file simdjson/icelake/implementation.h */
+namespace simdjson {
+namespace internal {
+static const icelake::implementation* get_icelake_singleton() {
+  static const icelake::implementation icelake_singleton{};
+  return &icelake_singleton;
+}
+} // namespace internal
+} // namespace simdjson
+#endif
+
+#if SIMDJSON_IMPLEMENTATION_PPC64
+/* including simdjson/ppc64/implementation.h: #include <simdjson/ppc64/implementation.h> */
+/* begin file simdjson/ppc64/implementation.h */
+#ifndef SIMDJSON_PPC64_IMPLEMENTATION_H
+#define SIMDJSON_PPC64_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/instruction_set.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+
+/**
+ * Implementation for ALTIVEC (PPC64).
+ */
+namespace ppc64 {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation()
+      : simdjson::implementation("ppc64", "PPC64 ALTIVEC",
+                                 internal::instruction_set::ALTIVEC) {}
+
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+      size_t capacity, size_t max_length,
+      std::unique_ptr<internal::dom_parser_implementation> &dst)
+      const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len,
+                                         uint8_t *dst,
+                                         size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf,
+                                          size_t len) const noexcept final;
+};
+
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_PPC64_IMPLEMENTATION_H
+/* end file simdjson/ppc64/implementation.h */
+namespace simdjson {
+namespace internal {
+static const ppc64::implementation* get_ppc64_singleton() {
+  static const ppc64::implementation ppc64_singleton{};
+  return &ppc64_singleton;
+}
+} // namespace internal
+} // namespace simdjson
+#endif // SIMDJSON_IMPLEMENTATION_PPC64
+
+#if SIMDJSON_IMPLEMENTATION_WESTMERE
+/* including simdjson/westmere/implementation.h: #include <simdjson/westmere/implementation.h> */
+/* begin file simdjson/westmere/implementation.h */
+#ifndef SIMDJSON_WESTMERE_IMPLEMENTATION_H
+#define SIMDJSON_WESTMERE_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/instruction_set.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// The constructor may be executed on any host, so we take care not to use SIMDJSON_TARGET_WESTMERE
+namespace simdjson {
+namespace westmere {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation("westmere", "Intel/AMD SSE4.2", internal::instruction_set::SSE42 | internal::instruction_set::PCLMULQDQ) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_IMPLEMENTATION_H
+/* end file simdjson/westmere/implementation.h */
+namespace simdjson {
+namespace internal {
+static const simdjson::westmere::implementation* get_westmere_singleton() {
+  static const simdjson::westmere::implementation westmere_singleton{};
+  return &westmere_singleton;
+}
+} // namespace internal
+} // namespace simdjson
+#endif // SIMDJSON_IMPLEMENTATION_WESTMERE
+
+#if SIMDJSON_IMPLEMENTATION_LASX
+/* including simdjson/lasx/implementation.h: #include <simdjson/lasx/implementation.h> */
+/* begin file simdjson/lasx/implementation.h */
+#ifndef SIMDJSON_LASX_IMPLEMENTATION_H
+#define SIMDJSON_LASX_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/instruction_set.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation("lasx", "LoongArch ASX", internal::instruction_set::LASX) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_LASX_IMPLEMENTATION_H
+/* end file simdjson/lasx/implementation.h */
+namespace simdjson {
+namespace internal {
+static const simdjson::lasx::implementation* get_lasx_singleton() {
+  static const simdjson::lasx::implementation lasx_singleton{};
+  return &lasx_singleton;
+}
+} // namespace internal
+} // namespace simdjson
+#endif // SIMDJSON_IMPLEMENTATION_LASX
+
+#if SIMDJSON_IMPLEMENTATION_LSX
+/* including simdjson/lsx/implementation.h: #include <simdjson/lsx/implementation.h> */
+/* begin file simdjson/lsx/implementation.h */
+#ifndef SIMDJSON_LSX_IMPLEMENTATION_H
+#define SIMDJSON_LSX_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/instruction_set.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation("lsx", "LoongArch SX", internal::instruction_set::LSX) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_LSX_IMPLEMENTATION_H
+/* end file simdjson/lsx/implementation.h */
+namespace simdjson {
+namespace internal {
+static const simdjson::lsx::implementation* get_lsx_singleton() {
+  static const simdjson::lsx::implementation lsx_singleton{};
+  return &lsx_singleton;
+}
+} // namespace internal
+} // namespace simdjson
+#endif // SIMDJSON_IMPLEMENTATION_LSX
+
+#if SIMDJSON_IMPLEMENTATION_RVV_VLS
+/* including simdjson/rvv-vls/implementation.h: #include <simdjson/rvv-vls/implementation.h> */
+/* begin file simdjson/rvv-vls/implementation.h */
+#ifndef SIMDJSON_RVV_VLS_IMPLEMENTATION_H
+#define SIMDJSON_RVV_VLS_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation(
+      "rvv_vls",
+      "RISC-V V extension",
+      0
+  ) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<simdjson::internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_RVV_VLS_IMPLEMENTATION_H
+/* end file simdjson/rvv-vls/implementation.h */
+namespace simdjson {
+namespace internal {
+static const simdjson::rvv_vls::implementation* get_rvv_vls_singleton() {
+  static const simdjson::rvv_vls::implementation rvv_vls_singleton{};
+  return &rvv_vls_singleton;
+}
+} // namespace internal
+} // namespace simdjson
+#endif // SIMDJSON_IMPLEMENTATION_RVV_VLS
+
+/* undefining SIMDJSON_CONDITIONAL_INCLUDE */
+#undef SIMDJSON_CONDITIONAL_INCLUDE
+
+namespace simdjson {
+namespace internal {
+
+// When there is a single implementation, we should not pay a price
+// for dispatching to the best implementation. We should just use the
+// one we have. This is a compile-time check.
+#define SIMDJSON_SINGLE_IMPLEMENTATION (SIMDJSON_IMPLEMENTATION_ICELAKE \
+             + SIMDJSON_IMPLEMENTATION_HASWELL + SIMDJSON_IMPLEMENTATION_WESTMERE \
+             + SIMDJSON_IMPLEMENTATION_ARM64 + SIMDJSON_IMPLEMENTATION_PPC64 \
+             + SIMDJSON_IMPLEMENTATION_LSX + SIMDJSON_IMPLEMENTATION_LASX \
+             + SIMDJSON_IMPLEMENTATION_RVV_VLS + SIMDJSON_IMPLEMENTATION_FALLBACK == 1)
+
+#if SIMDJSON_SINGLE_IMPLEMENTATION
+  simdjson_really_inline static const implementation* get_single_implementation() {
+    return
+#if SIMDJSON_IMPLEMENTATION_ICELAKE
+    get_icelake_singleton();
+#endif
+#if SIMDJSON_IMPLEMENTATION_HASWELL
+    get_haswell_singleton();
+#endif
+#if SIMDJSON_IMPLEMENTATION_WESTMERE
+    get_westmere_singleton();
+#endif
+#if SIMDJSON_IMPLEMENTATION_ARM64
+    get_arm64_singleton();
+#endif
+#if SIMDJSON_IMPLEMENTATION_PPC64
+    get_ppc64_singleton();
+#endif
+#if SIMDJSON_IMPLEMENTATION_LSX
+    get_lsx_singleton();
+#endif
+#if SIMDJSON_IMPLEMENTATION_LASX
+    get_lasx_singleton();
+#endif
+#if SIMDJSON_IMPLEMENTATION_RVV_VLS
+    get_rvv_vls_singleton();
+#endif
+#if SIMDJSON_IMPLEMENTATION_FALLBACK
+    get_fallback_singleton();
+#endif
+}
+#endif
+
+// Static array of known implementations. We're hoping these get baked into the executable
+// without requiring a static initializer.
+
+/**
+ * @private Detects best supported implementation on first use, and sets it
+ */
+class detect_best_supported_implementation_on_first_use final : public implementation {
+public:
+  std::string name() const noexcept final { return set_best()->name(); }
+  std::string description() const noexcept final { return set_best()->description(); }
+  uint32_t required_instruction_sets() const noexcept final { return set_best()->required_instruction_sets(); }
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<internal::dom_parser_implementation>& dst
+  ) const noexcept final {
+    return set_best()->create_dom_parser_implementation(capacity, max_length, dst);
+  }
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final {
+    return set_best()->minify(buf, len, dst, dst_len);
+  }
+  simdjson_warn_unused bool validate_utf8(const char * buf, size_t len) const noexcept final override {
+    return set_best()->validate_utf8(buf, len);
+  }
+  simdjson_inline detect_best_supported_implementation_on_first_use() noexcept : implementation("best_supported_detector", "Detects the best supported implementation and sets it", 0) {}
+private:
+  const implementation *set_best() const noexcept;
+};
+
+static_assert(std::is_trivially_destructible<detect_best_supported_implementation_on_first_use>::value, "detect_best_supported_implementation_on_first_use should be trivially destructible");
+
+static const std::initializer_list<const implementation *>& get_available_implementation_pointers() {
+  static const std::initializer_list<const implementation *> available_implementation_pointers {
+#if SIMDJSON_IMPLEMENTATION_ICELAKE
+    get_icelake_singleton(),
+#endif
+#if SIMDJSON_IMPLEMENTATION_HASWELL
+    get_haswell_singleton(),
+#endif
+#if SIMDJSON_IMPLEMENTATION_WESTMERE
+    get_westmere_singleton(),
+#endif
+#if SIMDJSON_IMPLEMENTATION_ARM64
+    get_arm64_singleton(),
+#endif
+#if SIMDJSON_IMPLEMENTATION_PPC64
+    get_ppc64_singleton(),
+#endif
+#if SIMDJSON_IMPLEMENTATION_LASX
+    get_lasx_singleton(),
+#endif
+#if SIMDJSON_IMPLEMENTATION_LSX
+    get_lsx_singleton(),
+#endif
+#if SIMDJSON_IMPLEMENTATION_RVV_VLS
+    get_rvv_vls_singleton(),
+#endif
+#if SIMDJSON_IMPLEMENTATION_FALLBACK
+    get_fallback_singleton(),
+#endif
+  }; // available_implementation_pointers
+  return available_implementation_pointers;
+}
+
+// So we can return UNSUPPORTED_ARCHITECTURE from the parser when there is no support
+class unsupported_implementation final : public implementation {
+public:
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t,
+    size_t,
+    std::unique_ptr<internal::dom_parser_implementation>&
+  ) const noexcept final {
+    return UNSUPPORTED_ARCHITECTURE;
+  }
+  simdjson_warn_unused error_code minify(const uint8_t *, size_t, uint8_t *, size_t &) const noexcept final override {
+    return UNSUPPORTED_ARCHITECTURE;
+  }
+  simdjson_warn_unused bool validate_utf8(const char *, size_t) const noexcept final override {
+    return false; // Just refuse to validate. Given that we have a fallback implementation
+    // it seems unlikely that unsupported_implementation will ever be used. If it is used,
+    // then it will flag all strings as invalid. The alternative is to return an error_code
+    // from which the user has to figure out whether the string is valid UTF-8... which seems
+    // like a lot of work just to handle the very unlikely case that we have an unsupported
+    // implementation. And, when it does happen (that we have an unsupported implementation),
+    // what are the chances that the programmer has a fallback? Given that *we* provide the
+    // fallback, it implies that the programmer would need a fallback for our fallback.
+  }
+  unsupported_implementation() : implementation("unsupported", "Unsupported CPU (no detected SIMD instructions)", 0) {}
+};
+
+static_assert(std::is_trivially_destructible<unsupported_implementation>::value, "unsupported_singleton should be trivially destructible");
+
+const unsupported_implementation* get_unsupported_singleton() {
+    static const unsupported_implementation unsupported_singleton{};
+    return &unsupported_singleton;
+}
+
+size_t available_implementation_list::size() const noexcept {
+  return internal::get_available_implementation_pointers().size();
+}
+const implementation * const *available_implementation_list::begin() const noexcept {
+  return internal::get_available_implementation_pointers().begin();
+}
+const implementation * const *available_implementation_list::end() const noexcept {
+  return internal::get_available_implementation_pointers().end();
+}
+const implementation *available_implementation_list::detect_best_supported() const noexcept {
+  // They are prelisted in priority order, so we just go down the list
+  uint32_t supported_instruction_sets = internal::detect_supported_architectures();
+  for (const implementation *impl : internal::get_available_implementation_pointers()) {
+    uint32_t required_instruction_sets = impl->required_instruction_sets();
+    if ((supported_instruction_sets & required_instruction_sets) == required_instruction_sets) { return impl; }
+  }
+  return get_unsupported_singleton(); // this should never happen?
+}
+
+const implementation *detect_best_supported_implementation_on_first_use::set_best() const noexcept {
+  SIMDJSON_PUSH_DISABLE_WARNINGS
+  SIMDJSON_DISABLE_DEPRECATED_WARNING // Disable CRT_SECURE warning on MSVC: manually verified this is safe
+  char *force_implementation_name = getenv("SIMDJSON_FORCE_IMPLEMENTATION");
+  SIMDJSON_POP_DISABLE_WARNINGS
+
+  if (force_implementation_name) {
+    auto force_implementation = get_available_implementations()[force_implementation_name];
+    if (force_implementation) {
+      return get_active_implementation() = force_implementation;
+    } else {
+      // Note: abort() and stderr usage within the library is forbidden.
+      return get_active_implementation() = get_unsupported_singleton();
+    }
+  }
+  return get_active_implementation() = get_available_implementations().detect_best_supported();
+}
+
+} // namespace internal
+
+SIMDJSON_DLLIMPORTEXPORT const internal::available_implementation_list& get_available_implementations() {
+  static const internal::available_implementation_list available_implementations{};
+  return available_implementations;
+}
+
+SIMDJSON_DLLIMPORTEXPORT internal::atomic_ptr<const implementation>& get_active_implementation() {
+#if SIMDJSON_SINGLE_IMPLEMENTATION
+  // We immediately select the only implementation we have, skipping the
+  // detect_best_supported_implementation_on_first_use_singleton.
+  static internal::atomic_ptr<const implementation> active_implementation{internal::get_single_implementation()};
+  return active_implementation;
+#else
+  static const internal::detect_best_supported_implementation_on_first_use detect_best_supported_implementation_on_first_use_singleton;
+  static internal::atomic_ptr<const implementation> active_implementation{&detect_best_supported_implementation_on_first_use_singleton};
+  return active_implementation;
+#endif
+}
+
+simdjson_warn_unused error_code minify(const char *buf, size_t len, char *dst, size_t &dst_len) noexcept {
+  return get_active_implementation()->minify(reinterpret_cast<const uint8_t *>(buf), len, reinterpret_cast<uint8_t *>(dst), dst_len);
+}
+simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) noexcept {
+  return get_active_implementation()->validate_utf8(buf, len);
+}
+const implementation * builtin_implementation() {
+  static const implementation * builtin_impl = get_available_implementations()[SIMDJSON_STRINGIFY(SIMDJSON_BUILTIN_IMPLEMENTATION)];
+  assert(builtin_impl);
+  return builtin_impl;
+}
+
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_IMPLEMENTATION_CPP
+/* end file implementation.cpp */
+
+/* defining SIMDJSON_CONDITIONAL_INCLUDE */
+#define SIMDJSON_CONDITIONAL_INCLUDE
+
+#if SIMDJSON_IMPLEMENTATION_ARM64
+/* including arm64.cpp: #include <arm64.cpp> */
+/* begin file arm64.cpp */
+#ifndef SIMDJSON_SRC_ARM64_CPP
+#define SIMDJSON_SRC_ARM64_CPP
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+/* including simdjson/arm64.h: #include <simdjson/arm64.h> */
+/* begin file simdjson/arm64.h */
+#ifndef SIMDJSON_ARM64_H
+#define SIMDJSON_ARM64_H
+
+/* including simdjson/arm64/begin.h: #include "simdjson/arm64/begin.h" */
+/* begin file simdjson/arm64/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "arm64" */
+#define SIMDJSON_IMPLEMENTATION arm64
+/* including simdjson/arm64/base.h: #include "simdjson/arm64/base.h" */
+/* begin file simdjson/arm64/base.h */
+#ifndef SIMDJSON_ARM64_BASE_H
+#define SIMDJSON_ARM64_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+/**
+ * Implementation for NEON (ARMv8).
+ */
+namespace arm64 {
+
+class implementation;
+
+namespace {
+namespace simd {
+template <typename T> struct simd8;
+template <typename T> struct simd8x64;
+} // namespace simd
+} // unnamed namespace
+
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_ARM64_BASE_H
+/* end file simdjson/arm64/base.h */
+/* including simdjson/arm64/intrinsics.h: #include "simdjson/arm64/intrinsics.h" */
+/* begin file simdjson/arm64/intrinsics.h */
+#ifndef SIMDJSON_ARM64_INTRINSICS_H
+#define SIMDJSON_ARM64_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This should be the correct header whether
+// you use visual studio or other compilers.
+#include <arm_neon.h>
+
+static_assert(sizeof(uint8x16_t) <= simdjson::SIMDJSON_PADDING, "insufficient padding for arm64");
+
+#endif //  SIMDJSON_ARM64_INTRINSICS_H
+/* end file simdjson/arm64/intrinsics.h */
+/* including simdjson/arm64/bitmanipulation.h: #include "simdjson/arm64/bitmanipulation.h" */
+/* begin file simdjson/arm64/bitmanipulation.h */
+#ifndef SIMDJSON_ARM64_BITMANIPULATION_H
+#define SIMDJSON_ARM64_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  unsigned long ret;
+  // Search the mask data from least significant bit (LSB)
+  // to the most significant bit (MSB) for a set bit (1).
+  _BitScanForward64(&ret, input_num);
+  return (int)ret;
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO
+  return __builtin_ctzll(input_num);
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return input_num & (input_num-1);
+}
+
+// We sometimes call leading_zeroes on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+// Applies only when SIMDJSON_PREFER_REVERSE_BITS is defined and true.
+// (See below.)
+SIMDJSON_NO_SANITIZE_UNDEFINED
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  unsigned long leading_zero = 0;
+  // Search the mask data from most significant bit (MSB)
+  // to least significant bit (LSB) for a set bit (1).
+  if (_BitScanReverse64(&leading_zero, input_num))
+    return (int)(63 - leading_zero);
+  else
+    return 64;
+#else
+  return __builtin_clzll(input_num);
+#endif// SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int count_ones(uint64_t input_num) {
+   return vaddv_u8(vcnt_u8(vcreate_u8(input_num)));
+}
+
+
+#if defined(__GNUC__) // catches clang and gcc
+/**
+ * ARM has a fast 64-bit "bit reversal function" that is handy. However,
+ * it is not generally available as an intrinsic function under Visual
+ * Studio (though this might be changing). Even under clang/gcc, we
+ * apparently need to invoke inline assembly.
+ */
+/*
+ * We use SIMDJSON_PREFER_REVERSE_BITS as a hint that algorithms that
+ * work well with bit reversal may use it.
+ */
+#define SIMDJSON_PREFER_REVERSE_BITS 1
+
+/* reverse the bits */
+simdjson_inline uint64_t reverse_bits(uint64_t input_num) {
+  uint64_t rev_bits;
+  __asm("rbit %0, %1" : "=r"(rev_bits) : "r"(input_num));
+  return rev_bits;
+}
+
+/**
+ * Flips bit at index 63 - lz. Thus if you have 'leading_zeroes' leading zeroes,
+ * then this will set to zero the leading bit. It is possible for leading_zeroes to be
+ * greating or equal to 63 in which case we trigger undefined behavior, but the output
+ * of such undefined behavior is never used.
+ **/
+SIMDJSON_NO_SANITIZE_UNDEFINED
+simdjson_inline uint64_t zero_leading_bit(uint64_t rev_bits, int leading_zeroes) {
+  return rev_bits ^ (uint64_t(0x8000000000000000) >> leading_zeroes);
+}
+
+#endif
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2, uint64_t *result) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  *result = value1 + value2;
+  return *result < value1;
+#else
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+#endif
+}
+
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_ARM64_BITMANIPULATION_H
+/* end file simdjson/arm64/bitmanipulation.h */
+/* including simdjson/arm64/bitmask.h: #include "simdjson/arm64/bitmask.h" */
+/* begin file simdjson/arm64/bitmask.h */
+#ifndef SIMDJSON_ARM64_BITMASK_H
+#define SIMDJSON_ARM64_BITMASK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(uint64_t bitmask) {
+  /////////////
+  // We could do this with PMULL, but it is apparently slow.
+  //
+  //#ifdef __ARM_FEATURE_CRYPTO // some ARM processors lack this extension
+  //return vmull_p64(-1ULL, bitmask);
+  //#else
+  // Analysis by @sebpop:
+  // When diffing the assembly for src/stage1_find_marks.cpp I see that the eors are all spread out
+  // in between other vector code, so effectively the extra cycles of the sequence do not matter
+  // because the GPR units are idle otherwise and the critical path is on the FP side.
+  // Also the PMULL requires two extra fmovs: GPR->FP (3 cycles in N1, 5 cycles in A72 )
+  // and FP->GPR (2 cycles on N1 and 5 cycles on A72.)
+  ///////////
+  bitmask ^= bitmask << 1;
+  bitmask ^= bitmask << 2;
+  bitmask ^= bitmask << 4;
+  bitmask ^= bitmask << 8;
+  bitmask ^= bitmask << 16;
+  bitmask ^= bitmask << 32;
+  return bitmask;
+}
+
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif
+/* end file simdjson/arm64/bitmask.h */
+/* including simdjson/arm64/numberparsing_defs.h: #include "simdjson/arm64/numberparsing_defs.h" */
+/* begin file simdjson/arm64/numberparsing_defs.h */
+#ifndef SIMDJSON_ARM64_NUMBERPARSING_DEFS_H
+#define SIMDJSON_ARM64_NUMBERPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+#if SIMDJSON_REGULAR_VISUAL_STUDIO && SIMDJSON_IS_ARM64
+// __umulh requires intrin.h
+#include <intrin.h>
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO && SIMDJSON_IS_ARM64
+
+namespace simdjson {
+namespace arm64 {
+namespace numberparsing {
+
+// we don't have SSE, so let us use a scalar function
+// credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  uint64_t val;
+  std::memcpy(&val, chars, sizeof(uint64_t));
+  val = (val & 0x0F0F0F0F0F0F0F0F) * 2561 >> 8;
+  val = (val & 0x00FF00FF00FF00FF) * 6553601 >> 16;
+  return uint32_t((val & 0x0000FFFF0000FFFF) * 42949672960001 >> 32);
+}
+
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+#if SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+#if SIMDJSON_IS_ARM64
+  // ARM64 has native support for 64-bit multiplications, no need to emultate
+  answer.high = __umulh(value1, value2);
+  answer.low = value1 * value2;
+#else
+  answer.low = _umul128(value1, value2, &answer.high); // _umul128 not available on ARM64
+#endif // SIMDJSON_IS_ARM64
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+#endif
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace arm64
+} // namespace simdjson
+
+#ifndef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_IS_BIG_ENDIAN
+#define SIMDJSON_SWAR_NUMBER_PARSING 0
+#else
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+#endif
+#endif
+
+#endif // SIMDJSON_ARM64_NUMBERPARSING_DEFS_H
+/* end file simdjson/arm64/numberparsing_defs.h */
+/* including simdjson/arm64/simd.h: #include "simdjson/arm64/simd.h" */
+/* begin file simdjson/arm64/simd.h */
+#ifndef SIMDJSON_ARM64_SIMD_H
+#define SIMDJSON_ARM64_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace simd {
+
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+namespace {
+// Start of private section with Visual Studio workaround
+
+
+#ifndef simdjson_make_uint8x16_t
+#define simdjson_make_uint8x16_t(x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, \
+                             x13, x14, x15, x16)                                   \
+   ([=]() {                                                                        \
+     uint8_t array[16] = {x1, x2,  x3,  x4,  x5,  x6,  x7,  x8,                    \
+                                 x9, x10, x11, x12, x13, x14, x15, x16};           \
+     return vld1q_u8(array);                                                       \
+   }())
+#endif
+#ifndef simdjson_make_int8x16_t
+#define simdjson_make_int8x16_t(x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, \
+                             x13, x14, x15, x16)                                  \
+   ([=]() {                                                                       \
+     int8_t array[16] = {x1, x2,  x3,  x4,  x5,  x6,  x7,  x8,                    \
+                                 x9, x10, x11, x12, x13, x14, x15, x16};          \
+     return vld1q_s8(array);                                                      \
+   }())
+#endif
+
+#ifndef simdjson_make_uint8x8_t
+#define simdjson_make_uint8x8_t(x1, x2, x3, x4, x5, x6, x7, x8)                \
+   ([=]() {                                                                    \
+     uint8_t array[8] = {x1, x2,  x3,  x4,  x5,  x6,  x7,  x8};                \
+     return vld1_u8(array);                                                    \
+   }())
+#endif
+#ifndef simdjson_make_int8x8_t
+#define simdjson_make_int8x8_t(x1, x2, x3, x4, x5, x6, x7, x8)                 \
+   ([=]() {                                                                    \
+     int8_t array[8] = {x1, x2,  x3,  x4,  x5,  x6,  x7,  x8};                 \
+     return vld1_s8(array);                                                    \
+   }())
+#endif
+#ifndef simdjson_make_uint16x8_t
+#define simdjson_make_uint16x8_t(x1, x2, x3, x4, x5, x6, x7, x8)               \
+   ([=]() {                                                                    \
+     uint16_t array[8] = {x1, x2,  x3,  x4,  x5,  x6,  x7,  x8};               \
+     return vld1q_u16(array);                                                  \
+   }())
+#endif
+#ifndef simdjson_make_int16x8_t
+#define simdjson_make_int16x8_t(x1, x2, x3, x4, x5, x6, x7, x8)                \
+   ([=]() {                                                                    \
+     int16_t array[8] = {x1, x2,  x3,  x4,  x5,  x6,  x7,  x8};                \
+     return vld1q_s16(array);                                                  \
+   }())
+#endif
+
+// End of private section with Visual Studio workaround
+} // namespace
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+
+
+  template<typename T>
+  struct simd8;
+
+  //
+  // Base class of simd8<uint8_t> and simd8<bool>, both of which use uint8x16_t internally.
+  //
+  template<typename T, typename Mask=simd8<bool>>
+  struct base_u8 {
+    uint8x16_t value;
+    static const int SIZE = sizeof(value);
+
+    // Conversion from/to SIMD register
+    simdjson_inline base_u8(const uint8x16_t _value) : value(_value) {}
+    simdjson_inline operator const uint8x16_t&() const { return this->value; }
+    simdjson_inline operator uint8x16_t&() { return this->value; }
+
+    // Bit operations
+    simdjson_inline simd8<T> operator|(const simd8<T> other) const { return vorrq_u8(*this, other); }
+    simdjson_inline simd8<T> operator&(const simd8<T> other) const { return vandq_u8(*this, other); }
+    simdjson_inline simd8<T> operator^(const simd8<T> other) const { return veorq_u8(*this, other); }
+    simdjson_inline simd8<T> bit_andnot(const simd8<T> other) const { return vbicq_u8(*this, other); }
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+    simdjson_inline simd8<T>& operator|=(const simd8<T> other) { auto this_cast = static_cast<simd8<T>*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline simd8<T>& operator&=(const simd8<T> other) { auto this_cast = static_cast<simd8<T>*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline simd8<T>& operator^=(const simd8<T> other) { auto this_cast = static_cast<simd8<T>*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+
+    friend simdjson_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return vceqq_u8(lhs, rhs); }
+
+    template<int N=1>
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+      return vextq_u8(prev_chunk, *this, 16 - N);
+    }
+  };
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool>: base_u8<bool> {
+    typedef uint16_t bitmask_t;
+    typedef uint32_t bitmask2_t;
+
+    static simdjson_inline simd8<bool> splat(bool _value) { return vmovq_n_u8(uint8_t(-(!!_value))); }
+
+    simdjson_inline simd8(const uint8x16_t _value) : base_u8<bool>(_value) {}
+    // False constructor
+    simdjson_inline simd8() : simd8(vdupq_n_u8(0)) {}
+    // Splat constructor
+    simdjson_inline simd8(bool _value) : simd8(splat(_value)) {}
+
+    // We return uint32_t instead of uint16_t because that seems to be more efficient for most
+    // purposes (cutting it down to uint16_t costs performance in some compilers).
+    simdjson_inline uint32_t to_bitmask() const {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+      const uint8x16_t bit_mask =  simdjson_make_uint8x16_t(0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+                                                   0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80);
+#else
+      const uint8x16_t bit_mask =  {0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+                                    0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80};
+#endif
+      auto minput = *this & bit_mask;
+      uint8x16_t tmp = vpaddq_u8(minput, minput);
+      tmp = vpaddq_u8(tmp, tmp);
+      tmp = vpaddq_u8(tmp, tmp);
+      return vgetq_lane_u16(vreinterpretq_u16_u8(tmp), 0);
+    }
+    // Returns 4-bit out of each byte, alternating between the high 4 bits and low
+    // bits result it is 64 bit.
+    simdjson_inline uint64_t to_bitmask64() const {
+      return vget_lane_u64(
+          vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(*this), 4)), 0);
+    }
+    simdjson_inline bool any() const { return vmaxvq_u32(vreinterpretq_u32_u8(*this)) != 0; }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t>: base_u8<uint8_t> {
+    static simdjson_inline uint8x16_t splat(uint8_t _value) { return vmovq_n_u8(_value); }
+    static simdjson_inline uint8x16_t zero() { return vdupq_n_u8(0); }
+    static simdjson_inline uint8x16_t load(const uint8_t* values) { return vld1q_u8(values); }
+
+    simdjson_inline simd8(const uint8x16_t _value) : base_u8<uint8_t>(_value) {}
+    // Zero constructor
+    simdjson_inline simd8() : simd8(zero()) {}
+    // Array constructor
+    simdjson_inline simd8(const uint8_t values[16]) : simd8(load(values)) {}
+    // Splat constructor
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    // Member-by-member initialization
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) : simd8(simdjson_make_uint8x16_t(
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    )) {}
+#else
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) : simd8(uint8x16_t{
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    }) {}
+#endif
+
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Store to array
+    simdjson_inline void store(uint8_t dst[16]) const { return vst1q_u8(dst, *this); }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return vqaddq_u8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return vqsubq_u8(*this, other); }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<uint8_t> operator+(const simd8<uint8_t> other) const { return vaddq_u8(*this, other); }
+    simdjson_inline simd8<uint8_t> operator-(const simd8<uint8_t> other) const { return vsubq_u8(*this, other); }
+    simdjson_inline simd8<uint8_t>& operator+=(const simd8<uint8_t> other) { *this = *this + other; return *this; }
+    simdjson_inline simd8<uint8_t>& operator-=(const simd8<uint8_t> other) { *this = *this - other; return *this; }
+
+    // Order-specific operations
+    simdjson_inline uint8_t max_val() const { return vmaxvq_u8(*this); }
+    simdjson_inline uint8_t min_val() const { return vminvq_u8(*this); }
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return vmaxq_u8(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return vminq_u8(*this, other); }
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return vcleq_u8(*this, other); }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return vcgeq_u8(*this, other); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return vcltq_u8(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return vcgtq_u8(*this, other); }
+    // Same as >, but instead of guaranteeing all 1's == true, false = 0 and true = nonzero. For ARM, returns all 1's.
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return simd8<uint8_t>(*this > other); }
+    // Same as <, but instead of guaranteeing all 1's == true, false = 0 and true = nonzero. For ARM, returns all 1's.
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return simd8<uint8_t>(*this < other); }
+
+    // Bit-specific operations
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return vtstq_u8(*this, bits); }
+    simdjson_inline bool any_bits_set_anywhere() const { return vmaxvq_u32(vreinterpretq_u32_u8(*this)) != 0; }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return (*this & bits).any_bits_set_anywhere(); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return vshrq_n_u8(*this, N); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return vshlq_n_u8(*this, N); }
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return lookup_table.apply_lookup_16_to(*this);
+    }
+
+    // Returns 4-bit out of each byte, alternating between the high 4 bits and low
+    // bits result it is 64 bit.
+    simdjson_inline uint64_t to_bitmask64() const {
+      return vget_lane_u64(
+          vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(*this), 4)), 0);
+    }
+    // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted as a bitset).
+    // Passing a 0 value for mask would be equivalent to writing out every byte to output.
+    // Only the first 16 - count_ones(mask) bytes of the result are significant but 16 bytes
+    // get written.
+    // Design consideration: it seems like a function with the
+    // signature simd8<L> compress(uint16_t mask) would be
+    // sensible, but the AVX ISA makes this kind of approach difficult.
+    template<typename L>
+    simdjson_inline void compress(uint16_t mask, L * output) const {
+      using internal::thintable_epi8;
+      using internal::BitsSetTable256mul2;
+      using internal::pshufb_combine_table;
+      // this particular implementation was inspired by work done by @animetosho
+      // we do it in two steps, first 8 bytes and then second 8 bytes
+      uint8_t mask1 = uint8_t(mask); // least significant 8 bits
+      uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
+      // next line just loads the 64-bit values thintable_epi8[mask1] and
+      // thintable_epi8[mask2] into a 128-bit register, using only
+      // two instructions on most compilers.
+      uint64x2_t shufmask64 = {thintable_epi8[mask1], thintable_epi8[mask2]};
+      uint8x16_t shufmask = vreinterpretq_u8_u64(shufmask64);
+      // we increment by 0x08 the second half of the mask
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+      uint8x16_t inc = simdjson_make_uint8x16_t(0, 0, 0, 0, 0, 0, 0, 0, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08);
+#else
+      uint8x16_t inc = {0, 0, 0, 0, 0, 0, 0, 0, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08};
+#endif
+      shufmask = vaddq_u8(shufmask, inc);
+      // this is the version "nearly pruned"
+      uint8x16_t pruned = vqtbl1q_u8(*this, shufmask);
+      // we still need to put the two halves together.
+      // we compute the popcount of the first half:
+      int pop1 = BitsSetTable256mul2[mask1];
+      // then load the corresponding mask, what it does is to write
+      // only the first pop1 bytes from the first 8 bytes, and then
+      // it fills in with the bytes from the second 8 bytes + some filling
+      // at the end.
+      uint8x16_t compactmask = vld1q_u8(reinterpret_cast<const uint8_t *>(pshufb_combine_table + pop1 * 8));
+      uint8x16_t answer = vqtbl1q_u8(pruned, compactmask);
+      vst1q_u8(reinterpret_cast<uint8_t*>(output), answer);
+    }
+
+    // Copies all bytes corresponding to a 0 in the low half of the mask (interpreted as a
+    // bitset) to output1, then those corresponding to a 0 in the high half to output2.
+    template<typename L>
+    simdjson_inline void compress_halves(uint16_t mask, L *output1, L *output2) const {
+      using internal::thintable_epi8;
+      uint8_t mask1 = uint8_t(mask); // least significant 8 bits
+      uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
+      uint8x8_t compactmask1 = vcreate_u8(thintable_epi8[mask1]);
+      uint8x8_t compactmask2 = vcreate_u8(thintable_epi8[mask2]);
+      // we increment by 0x08 the second half of the mask
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+      uint8x8_t inc = simdjson_make_uint8x8_t(0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08);
+#else
+      uint8x8_t inc = {0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08};
+#endif
+      compactmask2 = vadd_u8(compactmask2, inc);
+      // store each result (with the second store possibly overlapping the first)
+      vst1_u8((uint8_t*)output1, vqtbl1_u8(*this, compactmask1));
+      vst1_u8((uint8_t*)output2, vqtbl1_u8(*this, compactmask2));
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+
+    template<typename T>
+    simdjson_inline simd8<uint8_t> apply_lookup_16_to(const simd8<T> original) {
+      return vqtbl1q_u8(*this, simd8<uint8_t>(original));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> {
+    int8x16_t value;
+
+    static simdjson_inline simd8<int8_t> splat(int8_t _value) { return vmovq_n_s8(_value); }
+    static simdjson_inline simd8<int8_t> zero() { return vdupq_n_s8(0); }
+    static simdjson_inline simd8<int8_t> load(const int8_t values[16]) { return vld1q_s8(values); }
+
+    // Conversion from/to SIMD register
+    simdjson_inline simd8(const int8x16_t _value) : value{_value} {}
+    simdjson_inline operator const int8x16_t&() const { return this->value; }
+    simdjson_inline operator int8x16_t&() { return this->value; }
+
+    // Zero constructor
+    simdjson_inline simd8() : simd8(zero()) {}
+    // Splat constructor
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const int8_t* values) : simd8(load(values)) {}
+    // Member-by-member initialization
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3, int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) : simd8(simdjson_make_int8x16_t(
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    )) {}
+#else
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3, int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) : simd8(int8x16_t{
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    }) {}
+#endif
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Store to array
+    simdjson_inline void store(int8_t dst[16]) const { return vst1q_s8(dst, *this); }
+
+    // Explicit conversion to/from unsigned
+    //
+    // Under Visual Studio/ARM64 uint8x16_t and int8x16_t are apparently the same type.
+    // In theory, we could check this occurrence with std::same_as and std::enabled_if but it is C++14
+    // and relatively ugly and hard to read.
+#ifndef SIMDJSON_REGULAR_VISUAL_STUDIO
+    simdjson_inline explicit simd8(const uint8x16_t other): simd8(vreinterpretq_s8_u8(other)) {}
+#endif
+    simdjson_inline explicit operator simd8<uint8_t>() const { return vreinterpretq_u8_s8(this->value); }
+
+    // Math
+    simdjson_inline simd8<int8_t> operator+(const simd8<int8_t> other) const { return vaddq_s8(*this, other); }
+    simdjson_inline simd8<int8_t> operator-(const simd8<int8_t> other) const { return vsubq_s8(*this, other); }
+    simdjson_inline simd8<int8_t>& operator+=(const simd8<int8_t> other) { *this = *this + other; return *this; }
+    simdjson_inline simd8<int8_t>& operator-=(const simd8<int8_t> other) { *this = *this - other; return *this; }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return vmaxq_s8(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return vminq_s8(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return vcgtq_s8(*this, other); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return vcltq_s8(*this, other); }
+    simdjson_inline simd8<bool> operator==(const simd8<int8_t> other) const { return vceqq_s8(*this, other); }
+
+    template<int N=1>
+    simdjson_inline simd8<int8_t> prev(const simd8<int8_t> prev_chunk) const {
+      return vextq_s8(prev_chunk, *this, 16 - N);
+    }
+
+    // Perform a lookup assuming no value is larger than 16
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return lookup_table.apply_lookup_16_to(*this);
+    }
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+
+    template<typename T>
+    simdjson_inline simd8<int8_t> apply_lookup_16_to(const simd8<T> original) {
+      return vqtbl1q_s8(*this, simd8<uint8_t>(original));
+    }
+  };
+
+  template<typename T>
+  struct simd8x64 {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+    static_assert(NUM_CHUNKS == 4, "ARM kernel should use four registers per 64-byte block.");
+    const simd8<T> chunks[NUM_CHUNKS];
+    template<int idx> simd8<uint8_t> get() const { return idx < NUM_CHUNKS ? chunks[idx] : simd8<T>(); }
+
+    simd8x64(const simd8x64<T>& o) = delete; // no copy allowed
+    simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1, const simd8<T> chunk2, const simd8<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+16), simd8<T>::load(ptr+32), simd8<T>::load(ptr+48)} {}
+
+    simdjson_inline void store(T ptr[64]) const {
+      this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
+      this->chunks[1].store(ptr+sizeof(simd8<T>)*1);
+      this->chunks[2].store(ptr+sizeof(simd8<T>)*2);
+      this->chunks[3].store(ptr+sizeof(simd8<T>)*3);
+    }
+
+    simdjson_inline simd8<T> reduce_or() const {
+      return (this->chunks[0] | this->chunks[1]) | (this->chunks[2] | this->chunks[3]);
+    }
+
+
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
+      uint64_t popcounts = vget_lane_u64(vreinterpret_u64_u8(vcnt_u8(vcreate_u8(~mask))), 0);
+      // compute the prefix sum of the popcounts of each byte
+      uint64_t offsets = popcounts * 0x0101010101010101;
+      this->chunks[0].compress_halves(uint16_t(mask), output, &output[popcounts & 0xFF]);
+      this->chunks[1].compress_halves(uint16_t(mask >> 16), &output[(offsets >> 8) & 0xFF], &output[(offsets >> 16) & 0xFF]);
+      this->chunks[2].compress_halves(uint16_t(mask >> 32), &output[(offsets >> 24) & 0xFF], &output[(offsets >> 32) & 0xFF]);
+      this->chunks[3].compress_halves(uint16_t(mask >> 48), &output[(offsets >> 40) & 0xFF], &output[(offsets >> 48) & 0xFF]);
+      return offsets >> 56;
+    }
+
+    simdjson_inline uint64_t to_bitmask() const {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+      const uint8x16_t bit_mask = simdjson_make_uint8x16_t(
+        0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80
+      );
+#else
+      const uint8x16_t bit_mask = {
+        0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80
+      };
+#endif
+      // Add each of the elements next to each other, successively, to stuff each 8 byte mask into one.
+      uint8x16_t sum0 = vpaddq_u8(this->chunks[0] & bit_mask, this->chunks[1] & bit_mask);
+      uint8x16_t sum1 = vpaddq_u8(this->chunks[2] & bit_mask, this->chunks[3] & bit_mask);
+      sum0 = vpaddq_u8(sum0, sum1);
+      sum0 = vpaddq_u8(sum0, sum0);
+      return vgetq_lane_u64(vreinterpretq_u64_u8(sum0), 0);
+    }
+
+    simdjson_inline uint64_t eq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] == mask,
+        this->chunks[1] == mask,
+        this->chunks[2] == mask,
+        this->chunks[3] == mask
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t lteq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] <= mask,
+        this->chunks[1] <= mask,
+        this->chunks[2] <= mask,
+        this->chunks[3] <= mask
+      ).to_bitmask();
+    }
+  }; // struct simd8x64<T>
+
+} // namespace simd
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_ARM64_SIMD_H
+/* end file simdjson/arm64/simd.h */
+/* including simdjson/arm64/stringparsing_defs.h: #include "simdjson/arm64/stringparsing_defs.h" */
+/* begin file simdjson/arm64/stringparsing_defs.h */
+#ifndef SIMDJSON_ARM64_STRINGPARSING_DEFS_H
+#define SIMDJSON_ARM64_STRINGPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/simd.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 32;
+  simdjson_inline backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return bs_bits != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+
+  uint32_t bs_bits;
+  uint32_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // this can read up to 31 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v0(src);
+  simd8<uint8_t> v1(src + sizeof(v0));
+  v0.store(dst);
+  v1.store(dst + sizeof(v0));
+
+  // Getting a 64-bit bitmask is much cheaper than multiple 16-bit bitmasks on ARM; therefore, we
+  // smash them together into a 64-byte mask and get the bitmask from there.
+  uint64_t bs_and_quote = simd8x64<bool>(v0 == '\\', v1 == '\\', v0 == '"', v1 == '"').to_bitmask();
+  return {
+    uint32_t(bs_and_quote),      // bs_bits
+    uint32_t(bs_and_quote >> 32) // quote_bits
+  };
+}
+
+struct escaping {
+  static constexpr uint32_t BYTES_PROCESSED = 16;
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits != 0; }
+  simdjson_inline int escape_index() { return trailing_zeroes(escape_bits) / 4; }
+
+  uint64_t escape_bits;
+}; // struct escaping
+
+
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "escaping finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  simd8<bool> is_quote = (v == '"');
+  simd8<bool> is_backslash = (v == '\\');
+  simd8<bool> is_control = (v < 32);
+  return {
+    (is_backslash | is_quote | is_control).to_bitmask64()
+  };
+}
+
+
+
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_ARM64_STRINGPARSING_DEFS_H
+/* end file simdjson/arm64/stringparsing_defs.h */
+
+#define SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT 1
+/* end file simdjson/arm64/begin.h */
+/* including simdjson/generic/amalgamated.h for arm64: #include "simdjson/generic/amalgamated.h" */
+/* begin file simdjson/generic/amalgamated.h for arm64 */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_GENERIC_DEPENDENCIES_H)
+#error simdjson/generic/dependencies.h must be included before simdjson/generic/amalgamated.h!
+#endif
+
+/* including simdjson/generic/base.h for arm64: #include "simdjson/generic/base.h" */
+/* begin file simdjson/generic/base.h for arm64 */
+#ifndef SIMDJSON_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): // If we haven't got an implementation yet, we're in the editor, editing a generic file! Just */
+/* amalgamation skipped (editor-only): // use the most advanced one we can so the most possible stuff can be tested. */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation_detection.h" */
+/* amalgamation skipped (editor-only): #if SIMDJSON_IMPLEMENTATION_ICELAKE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_HASWELL */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_WESTMERE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_ARM64 */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_PPC64 */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LASX */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LSX */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_RVV_VLS */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_FALLBACK */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/begin.h" */
+/* amalgamation skipped (editor-only): #else */
+/* amalgamation skipped (editor-only): #error "All possible implementations (including fallback) have been disabled! simdjson will not run." */
+/* amalgamation skipped (editor-only): #endif */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+
+struct open_container;
+class dom_parser_implementation;
+
+/**
+ * The type of a JSON number
+ */
+enum class number_type {
+    floating_point_number=1, /// a binary64 number
+    signed_integer,          /// a signed integer that fits in a 64-bit word using two's complement
+    unsigned_integer,        /// a positive integer larger or equal to 1<<63
+    big_integer              /// a big integer that does not fit in a 64-bit word
+};
+
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_BASE_H
+/* end file simdjson/generic/base.h for arm64 */
+/* including simdjson/generic/jsoncharutils.h for arm64: #include "simdjson/generic/jsoncharutils.h" */
+/* begin file simdjson/generic/jsoncharutils.h for arm64 */
+#ifndef SIMDJSON_GENERIC_JSONCHARUTILS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_JSONCHARUTILS_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/jsoncharutils_tables.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace jsoncharutils {
+
+// return non-zero if not a structural or whitespace char
+// zero otherwise
+simdjson_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace_negated[c];
+}
+
+simdjson_inline uint32_t is_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace[c];
+}
+
+// returns a value with the high 16 bits set if not valid
+// otherwise returns the conversion of the 4 hex digits at src into the bottom
+// 16 bits of the 32-bit return register
+//
+// see
+// https://lemire.me/blog/2019/04/17/parsing-short-hexadecimal-strings-efficiently/
+static inline uint32_t hex_to_u32_nocheck(
+    const uint8_t *src) { // strictly speaking, static inline is a C-ism
+  uint32_t v1 = internal::digit_to_val32[630 + src[0]];
+  uint32_t v2 = internal::digit_to_val32[420 + src[1]];
+  uint32_t v3 = internal::digit_to_val32[210 + src[2]];
+  uint32_t v4 = internal::digit_to_val32[0 + src[3]];
+  return v1 | v2 | v3 | v4;
+}
+
+// given a code point cp, writes to c
+// the utf-8 code, outputting the length in
+// bytes, if the length is zero, the code point
+// is invalid
+//
+// This can possibly be made faster using pdep
+// and clz and table lookups, but JSON documents
+// have few escaped code points, and the following
+// function looks cheap.
+//
+// Note: we assume that surrogates are treated separately
+//
+simdjson_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
+  if (cp <= 0x7F) {
+    c[0] = uint8_t(cp);
+    return 1; // ascii
+  }
+  if (cp <= 0x7FF) {
+    c[0] = uint8_t((cp >> 6) + 192);
+    c[1] = uint8_t((cp & 63) + 128);
+    return 2; // universal plane
+    //  Surrogates are treated elsewhere...
+    //} //else if (0xd800 <= cp && cp <= 0xdfff) {
+    //  return 0; // surrogates // could put assert here
+  } else if (cp <= 0xFFFF) {
+    c[0] = uint8_t((cp >> 12) + 224);
+    c[1] = uint8_t(((cp >> 6) & 63) + 128);
+    c[2] = uint8_t((cp & 63) + 128);
+    return 3;
+  } else if (cp <= 0x10FFFF) { // if you know you have a valid code point, this
+                               // is not needed
+    c[0] = uint8_t((cp >> 18) + 240);
+    c[1] = uint8_t(((cp >> 12) & 63) + 128);
+    c[2] = uint8_t(((cp >> 6) & 63) + 128);
+    c[3] = uint8_t((cp & 63) + 128);
+    return 4;
+  }
+  // will return 0 when the code point was too large.
+  return 0; // bad r
+}
+
+#if SIMDJSON_IS_32BITS // _umul128 for x86, arm
+// this is a slow emulation routine for 32-bit
+//
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+  return x * (uint64_t)y;
+}
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+  uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
+  uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
+  uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
+  uint64_t adbc_carry = !!(adbc < ad);
+  uint64_t lo = bd + (adbc << 32);
+  *hi = __emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) +
+        (adbc_carry << 32) + !!(lo < bd);
+  return lo;
+}
+#endif
+
+} // namespace jsoncharutils
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_JSONCHARUTILS_H
+/* end file simdjson/generic/jsoncharutils.h for arm64 */
+/* including simdjson/generic/atomparsing.h for arm64: #include "simdjson/generic/atomparsing.h" */
+/* begin file simdjson/generic/atomparsing.h for arm64 */
+#ifndef SIMDJSON_GENERIC_ATOMPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_ATOMPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+/// @private
+namespace atomparsing {
+
+// The string_to_uint32 is exclusively used to map literal strings to 32-bit values.
+// We use memcpy instead of a pointer cast to avoid undefined behaviors since we cannot
+// be certain that the character pointer will be properly aligned.
+// You might think that using memcpy makes this function expensive, but you'd be wrong.
+// All decent optimizing compilers (GCC, clang, Visual Studio) will compile string_to_uint32("false");
+// to the compile-time constant 1936482662.
+simdjson_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
+
+
+// Again in str4ncmp we use a memcpy to avoid undefined behavior. The memcpy may appear expensive.
+// Yet all decent optimizing compilers will compile memcpy to a single instruction, just about.
+simdjson_warn_unused
+simdjson_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
+  uint32_t srcval; // we want to avoid unaligned 32-bit loads (undefined in C/C++)
+  static_assert(sizeof(uint32_t) <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be larger than 4 bytes");
+  std::memcpy(&srcval, src, sizeof(uint32_t));
+  return srcval ^ string_to_uint32(atom);
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src) {
+  return (str4ncmp(src, "true") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_true_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "true"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src) {
+  return (str4ncmp(src+1, "alse") | jsoncharutils::is_not_structural_or_whitespace(src[5])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
+  if (len > 5) { return is_valid_false_atom(src); }
+  else if (len == 5) { return !str4ncmp(src+1, "alse"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src) {
+  return (str4ncmp(src, "null") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_null_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "null"); }
+  else { return false; }
+}
+
+} // namespace atomparsing
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_ATOMPARSING_H
+/* end file simdjson/generic/atomparsing.h for arm64 */
+/* including simdjson/generic/dom_parser_implementation.h for arm64: #include "simdjson/generic/dom_parser_implementation.h" */
+/* begin file simdjson/generic/dom_parser_implementation.h for arm64 */
+#ifndef SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/dom_parser_implementation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+
+// expectation: sizeof(open_container) = 64/8.
+struct open_container {
+  uint32_t tape_index; // where, on the tape, does the scope ([,{) begins
+  uint32_t count; // how many elements in the scope
+}; // struct open_container
+
+static_assert(sizeof(open_container) == 64/8, "Open container must be 64 bits");
+
+class dom_parser_implementation final : public internal::dom_parser_implementation {
+public:
+  /** Tape location of each open { or [ */
+  std::unique_ptr<open_container[]> open_containers{};
+  /** Whether each open container is a [ or { */
+  std::unique_ptr<bool[]> is_array{};
+  /** Buffer passed to stage 1 */
+  const uint8_t *buf{};
+  /** Length passed to stage 1 */
+  size_t len{0};
+  /** Document passed to stage 2 */
+  dom::document *doc{};
+
+  inline dom_parser_implementation() noexcept;
+  inline dom_parser_implementation(dom_parser_implementation &&other) noexcept;
+  inline dom_parser_implementation &operator=(dom_parser_implementation &&other) noexcept;
+  dom_parser_implementation(const dom_parser_implementation &) = delete;
+  dom_parser_implementation &operator=(const dom_parser_implementation &) = delete;
+
+  simdjson_warn_unused error_code parse(const uint8_t *buf, size_t len, dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage1(const uint8_t *buf, size_t len, stage1_mode partial) noexcept final;
+  simdjson_warn_unused error_code stage2(dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage2_next(dom::document &doc) noexcept final;
+  simdjson_warn_unused uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) const noexcept final;
+  simdjson_warn_unused uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept final;
+  inline simdjson_warn_unused error_code set_capacity(size_t capacity) noexcept final;
+  inline simdjson_warn_unused error_code set_max_depth(size_t max_depth) noexcept final;
+private:
+  simdjson_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
+
+};
+
+} // namespace arm64
+} // namespace simdjson
+
+namespace simdjson {
+namespace arm64 {
+
+inline dom_parser_implementation::dom_parser_implementation() noexcept = default;
+inline dom_parser_implementation::dom_parser_implementation(dom_parser_implementation &&other) noexcept = default;
+inline dom_parser_implementation &dom_parser_implementation::operator=(dom_parser_implementation &&other) noexcept = default;
+
+// Leaving these here so they can be inlined if so desired
+inline simdjson_warn_unused error_code dom_parser_implementation::set_capacity(size_t capacity) noexcept {
+  if(capacity > SIMDJSON_MAXSIZE_BYTES) { return CAPACITY; }
+  // Stage 1 index output
+  size_t max_structures = SIMDJSON_ROUNDUP_N(capacity, 64) + 2 + 7;
+  structural_indexes.reset( new (std::nothrow) uint32_t[max_structures] );
+  if (!structural_indexes) { _capacity = 0; return MEMALLOC; }
+  structural_indexes[0] = 0;
+  n_structural_indexes = 0;
+
+  _capacity = capacity;
+  return SUCCESS;
+}
+
+inline simdjson_warn_unused error_code dom_parser_implementation::set_max_depth(size_t max_depth) noexcept {
+  // Stage 2 stacks
+  open_containers.reset(new (std::nothrow) open_container[max_depth]);
+  is_array.reset(new (std::nothrow) bool[max_depth]);
+  if (!is_array || !open_containers) { _max_depth = 0; return MEMALLOC; }
+
+  _max_depth = max_depth;
+  return SUCCESS;
+}
+
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file simdjson/generic/dom_parser_implementation.h for arm64 */
+/* including simdjson/generic/implementation_simdjson_result_base.h for arm64: #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base.h for arm64 */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+
+// This is a near copy of include/error.h's implementation_simdjson_result_base, except it doesn't use std::pair
+// so we can avoid inlining errors
+// TODO reconcile these!
+/**
+ * The result of a simdjson operation that could fail.
+ *
+ * Gives the option of reading error codes, or throwing an exception by casting to the desired result.
+ *
+ * This is a base class for implementations that want to add functions to the result type for
+ * chaining.
+ *
+ * Override like:
+ *
+ *   struct simdjson_result<T> : public internal::implementation_simdjson_result_base<T> {
+ *     simdjson_result() noexcept : internal::implementation_simdjson_result_base<T>() {}
+ *     simdjson_result(error_code error) noexcept : internal::implementation_simdjson_result_base<T>(error) {}
+ *     simdjson_result(T &&value) noexcept : internal::implementation_simdjson_result_base<T>(std::forward(value)) {}
+ *     simdjson_result(T &&value, error_code error) noexcept : internal::implementation_simdjson_result_base<T>(value, error) {}
+ *     // Your extra methods here
+ *   }
+ *
+ * Then any method returning simdjson_result<T> will be chainable with your methods.
+ */
+template<typename T>
+struct implementation_simdjson_result_base {
+
+  /**
+   * Create a new empty result with error = UNINITIALIZED.
+   */
+  simdjson_inline implementation_simdjson_result_base() noexcept = default;
+
+  /**
+   * Create a new error result.
+   */
+  simdjson_inline implementation_simdjson_result_base(error_code error) noexcept;
+
+  /**
+   * Create a new successful result.
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value) noexcept;
+
+  /**
+   * Create a new result with both things (use if you don't want to branch when creating the result).
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value, error_code error) noexcept;
+
+  /**
+   * Move the value and the error to the provided variables.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   * @param error The variable to assign the error to. Set to SUCCESS if there is no error.
+   */
+  simdjson_inline void tie(T &value, error_code &error) && noexcept;
+
+  /**
+   * Move the value to the provided variable.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   */
+  simdjson_warn_unused simdjson_inline error_code get(T &value) && noexcept;
+
+  /**
+   * The error.
+   */
+  simdjson_warn_unused simdjson_inline error_code error() const noexcept;
+
+  /**
+   * Whether there is a value.
+   */
+  simdjson_warn_unused simdjson_inline bool has_value() const noexcept;
+
+#if SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T& operator*() &  noexcept(false);
+  simdjson_inline T&& operator*() &&  noexcept(false);
+  /**
+   * Arrow operator to access members of the contained value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T* operator->() noexcept(false);
+  simdjson_inline const T* operator->() const noexcept(false);
+
+  simdjson_inline T& value() & noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& value() && noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& take_value() && noexcept(false);
+
+  /**
+   * Cast to the value (will throw on error).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline operator T&&() && noexcept(false);
+
+
+#endif // SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline const T& value_unsafe() const& noexcept;
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T& value_unsafe() & noexcept;
+  /**
+   * Take the result value (move it). This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T&& value_unsafe() && noexcept;
+
+  using value_type = T;
+  using error_type = error_code;
+
+protected:
+  /** users should never directly access first and second. **/
+  T first{}; /** Users should never directly access 'first'. **/
+  error_code second{UNINITIALIZED}; /** Users should never directly access 'second'. **/
+}; // struct implementation_simdjson_result_base
+
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+/* end file simdjson/generic/implementation_simdjson_result_base.h for arm64 */
+/* including simdjson/generic/numberparsing.h for arm64: #include "simdjson/generic/numberparsing.h" */
+/* begin file simdjson/generic/numberparsing.h for arm64 */
+#ifndef SIMDJSON_GENERIC_NUMBERPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_NUMBERPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <limits>
+#include <ostream>
+#include <cstring>
+
+namespace simdjson {
+namespace arm64 {
+namespace numberparsing {
+
+#ifdef JSON_TEST_NUMBERS
+#define INVALID_NUMBER(SRC) (found_invalid_number((SRC)), NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (found_integer((VALUE), (SRC)), (WRITER).append_s64((VALUE)))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (found_unsigned_integer((VALUE), (SRC)), (WRITER).append_u64((VALUE)))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (found_float((VALUE), (SRC)), (WRITER).append_double((VALUE)))
+#define BIGINT_NUMBER(SRC) (found_invalid_number((SRC)), BIGINT_ERROR)
+#else
+#define INVALID_NUMBER(SRC) (NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (WRITER).append_s64((VALUE))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (WRITER).append_u64((VALUE))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (WRITER).append_double((VALUE))
+#define BIGINT_NUMBER(SRC) (BIGINT_ERROR)
+#endif
+
+namespace {
+
+// Convert a mantissa, an exponent and a sign bit into an ieee64 double.
+// The real_exponent needs to be in [0, 2046] (technically real_exponent = 2047 would be acceptable).
+// The mantissa should be in [0,1<<53). The bit at index (1ULL << 52) while be zeroed.
+simdjson_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
+    double d;
+    mantissa &= ~(1ULL << 52);
+    mantissa |= real_exponent << 52;
+    mantissa |= ((static_cast<uint64_t>(negative)) << 63);
+    std::memcpy(&d, &mantissa, sizeof(d));
+    return d;
+}
+
+// Attempts to compute i * 10^(power) exactly; and if "negative" is
+// true, negate the result.
+// This function will only work in some cases, when it does not work, success is
+// set to false. This should work *most of the time* (like 99% of the time).
+// We assume that power is in the [smallest_power,
+// largest_power] interval: the caller is responsible for this check.
+simdjson_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
+  // we start with a fast path
+  // It was described in
+  // Clinger WD. How to read floating point numbers accurately.
+  // ACM SIGPLAN Notices. 1990
+#ifndef FLT_EVAL_METHOD
+#error "FLT_EVAL_METHOD should be defined, please include cfloat."
+#endif
+#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
+  // We cannot be certain that x/y is rounded to nearest.
+  if (0 <= power && power <= 22 && i <= 9007199254740991)
+#else
+  if (-22 <= power && power <= 22 && i <= 9007199254740991)
+#endif
+  {
+    // convert the integer into a double. This is lossless since
+    // 0 <= i <= 2^53 - 1.
+    d = double(i);
+    //
+    // The general idea is as follows.
+    // If 0 <= s < 2^53 and if 10^0 <= p <= 10^22 then
+    // 1) Both s and p can be represented exactly as 64-bit floating-point
+    // values
+    // (binary64).
+    // 2) Because s and p can be represented exactly as floating-point values,
+    // then s * p
+    // and s / p will produce correctly rounded values.
+    //
+    if (power < 0) {
+      d = d / simdjson::internal::power_of_ten[-power];
+    } else {
+      d = d * simdjson::internal::power_of_ten[power];
+    }
+    if (negative) {
+      d = -d;
+    }
+    return true;
+  }
+  // When 22 < power && power <  22 + 16, we could
+  // hope for another, secondary fast path.  It was
+  // described by David M. Gay in  "Correctly rounded
+  // binary-decimal and decimal-binary conversions." (1990)
+  // If you need to compute i * 10^(22 + x) for x < 16,
+  // first compute i * 10^x, if you know that result is exact
+  // (e.g., when i * 10^x < 2^53),
+  // then you can still proceed and do (i * 10^x) * 10^22.
+  // Is this worth your time?
+  // You need  22 < power *and* power <  22 + 16 *and* (i * 10^(x-22) < 2^53)
+  // for this second fast path to work.
+  // If you you have 22 < power *and* power <  22 + 16, and then you
+  // optimistically compute "i * 10^(x-22)", there is still a chance that you
+  // have wasted your time if i * 10^(x-22) >= 2^53. It makes the use cases of
+  // this optimization maybe less common than we would like. Source:
+  // http://www.exploringbinary.com/fast-path-decimal-to-floating-point-conversion/
+  // also used in RapidJSON: https://rapidjson.org/strtod_8h_source.html
+
+  // The fast path has now failed, so we are failing back on the slower path.
+
+  // In the slow path, we need to adjust i so that it is > 1<<63 which is always
+  // possible, except if i == 0, so we handle i == 0 separately.
+  if(i == 0) {
+    d = negative ? -0.0 : 0.0;
+    return true;
+  }
+
+
+  // The exponent is 1024 + 63 + power
+  //     + floor(log(5**power)/log(2)).
+  // The 1024 comes from the ieee64 standard.
+  // The 63 comes from the fact that we use a 64-bit word.
+  //
+  // Computing floor(log(5**power)/log(2)) could be
+  // slow. Instead we use a fast function.
+  //
+  // For power in (-400,350), we have that
+  // (((152170 + 65536) * power ) >> 16);
+  // is equal to
+  //  floor(log(5**power)/log(2)) + power when power >= 0
+  // and it is equal to
+  //  ceil(log(5**-power)/log(2)) + power when power < 0
+  //
+  // The 65536 is (1<<16) and corresponds to
+  // (65536 * power) >> 16 ---> power
+  //
+  // ((152170 * power ) >> 16) is equal to
+  // floor(log(5**power)/log(2))
+  //
+  // Note that this is not magic: 152170/(1<<16) is
+  // approximately equal to log(5)/log(2).
+  // The 1<<16 value is a power of two; we could use a
+  // larger power of 2 if we wanted to.
+  //
+  int64_t exponent = (((152170 + 65536) * power) >> 16) + 1024 + 63;
+
+
+  // We want the most significant bit of i to be 1. Shift if needed.
+  int lz = leading_zeroes(i);
+  i <<= lz;
+
+
+  // We are going to need to do some 64-bit arithmetic to get a precise product.
+  // We use a table lookup approach.
+  // It is safe because
+  // power >= smallest_power
+  // and power <= largest_power
+  // We recover the mantissa of the power, it has a leading 1. It is always
+  // rounded down.
+  //
+  // We want the most significant 64 bits of the product. We know
+  // this will be non-zero because the most significant bit of i is
+  // 1.
+  const uint32_t index = 2 * uint32_t(power - simdjson::internal::smallest_power);
+  // Optimization: It may be that materializing the index as a variable might confuse some compilers and prevent effective complex-addressing loads. (Done for code clarity.)
+  //
+  // The full_multiplication function computes the 128-bit product of two 64-bit words
+  // with a returned value of type value128 with a "low component" corresponding to the
+  // 64-bit least significant bits of the product and with a "high component" corresponding
+  // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index]);
+#else
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index]);
+#endif
+
+  // Both i and power_of_five_128[index] have their most significant bit set to 1 which
+  // implies that the either the most or the second most significant bit of the product
+  // is 1. We pack values in this manner for efficiency reasons: it maximizes the use
+  // we make of the product. It also makes it easy to reason about the product: there
+  // is 0 or 1 leading zero in the product.
+
+  // Unless the least significant 9 bits of the high (64-bit) part of the full
+  // product are all 1s, then we know that the most significant 55 bits are
+  // exact and no further work is needed. Having 55 bits is necessary because
+  // we need 53 bits for the mantissa but we have to have one rounding bit and
+  // we can waste a bit if the most significant bit of the product is zero.
+  if((firstproduct.high & 0x1FF) == 0x1FF) {
+    // We want to compute i * 5^q, but only care about the top 55 bits at most.
+    // Consider the scenario where q>=0. Then 5^q may not fit in 64-bits. Doing
+    // the full computation is wasteful. So we do what is called a "truncated
+    // multiplication".
+    // We take the most significant 64-bits, and we put them in
+    // power_of_five_128[index]. Usually, that's good enough to approximate i * 5^q
+    // to the desired approximation using one multiplication. Sometimes it does not suffice.
+    // Then we store the next most significant 64 bits in power_of_five_128[index + 1], and
+    // then we get a better approximation to i * 5^q.
+    //
+    // That's for when q>=0. The logic for q<0 is somewhat similar but it is somewhat
+    // more complicated.
+    //
+    // There is an extra layer of complexity in that we need more than 55 bits of
+    // accuracy in the round-to-even scenario.
+    //
+    // The full_multiplication function computes the 128-bit product of two 64-bit words
+    // with a returned value of type value128 with a "low component" corresponding to the
+    // 64-bit least significant bits of the product and with a "high component" corresponding
+    // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index + 1]);
+#else
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index + 1]);
+#endif
+    firstproduct.low += secondproduct.high;
+    if(secondproduct.high > firstproduct.low) { firstproduct.high++; }
+    // As it has been proven by Noble Mushtak and Daniel Lemire in "Fast Number Parsing Without
+    // Fallback" (https://arxiv.org/abs/2212.06644), at this point we are sure that the product
+    // is sufficiently accurate, and more computation is not needed.
+  }
+  uint64_t lower = firstproduct.low;
+  uint64_t upper = firstproduct.high;
+  // The final mantissa should be 53 bits with a leading 1.
+  // We shift it so that it occupies 54 bits with a leading 1.
+  ///////
+  uint64_t upperbit = upper >> 63;
+  uint64_t mantissa = upper >> (upperbit + 9);
+  lz += int(1 ^ upperbit);
+
+  // Here we have mantissa < (1<<54).
+  int64_t real_exponent = exponent - lz;
+  if (simdjson_unlikely(real_exponent <= 0)) { // we have a subnormal?
+    // Here have that real_exponent <= 0 so -real_exponent >= 0
+    if(-real_exponent + 1 >= 64) { // if we have more than 64 bits below the minimum exponent, you have a zero for sure.
+      d = negative ? -0.0 : 0.0;
+      return true;
+    }
+    // next line is safe because -real_exponent + 1 < 0
+    mantissa >>= -real_exponent + 1;
+    // Thankfully, we can't have both "round-to-even" and subnormals because
+    // "round-to-even" only occurs for powers close to 0.
+    mantissa += (mantissa & 1); // round up
+    mantissa >>= 1;
+    // There is a weird scenario where we don't have a subnormal but just.
+    // Suppose we start with 2.2250738585072013e-308, we end up
+    // with 0x3fffffffffffff x 2^-1023-53 which is technically subnormal
+    // whereas 0x40000000000000 x 2^-1023-53  is normal. Now, we need to round
+    // up 0x3fffffffffffff x 2^-1023-53  and once we do, we are no longer
+    // subnormal, but we can only know this after rounding.
+    // So we only declare a subnormal if we are smaller than the threshold.
+    real_exponent = (mantissa < (uint64_t(1) << 52)) ? 0 : 1;
+    d = to_double(mantissa, real_exponent, negative);
+    return true;
+  }
+  // We have to round to even. The "to even" part
+  // is only a problem when we are right in between two floats
+  // which we guard against.
+  // If we have lots of trailing zeros, we may fall right between two
+  // floating-point values.
+  //
+  // The round-to-even cases take the form of a number 2m+1 which is in (2^53,2^54]
+  // times a power of two. That is, it is right between a number with binary significand
+  // m and another number with binary significand m+1; and it must be the case
+  // that it cannot be represented by a float itself.
+  //
+  // We must have that w * 10 ^q == (2m+1) * 2^p for some power of two 2^p.
+  // Recall that 10^q = 5^q * 2^q.
+  // When q >= 0, we must have that (2m+1) is divible by 5^q, so 5^q <= 2^54. We have that
+  //  5^23 <=  2^54 and it is the last power of five to qualify, so q <= 23.
+  // When q<0, we have  w  >=  (2m+1) x 5^{-q}.  We must have that w<2^{64} so
+  // (2m+1) x 5^{-q} < 2^{64}. We have that 2m+1>2^{53}. Hence, we must have
+  // 2^{53} x 5^{-q} < 2^{64}.
+  // Hence we have 5^{-q} < 2^{11}$ or q>= -4.
+  //
+  // We require lower <= 1 and not lower == 0 because we could not prove that
+  // that lower == 0 is implied; but we could prove that lower <= 1 is a necessary and sufficient test.
+  if (simdjson_unlikely((lower <= 1) && (power >= -4) && (power <= 23) && ((mantissa & 3) == 1))) {
+    if((mantissa  << (upperbit + 64 - 53 - 2)) ==  upper) {
+      mantissa &= ~1;             // flip it so that we do not round up
+    }
+  }
+
+  mantissa += mantissa & 1;
+  mantissa >>= 1;
+
+  // Here we have mantissa < (1<<53), unless there was an overflow
+  if (mantissa >= (1ULL << 53)) {
+    //////////
+    // This will happen when parsing values such as 7.2057594037927933e+16
+    ////////
+    mantissa = (1ULL << 52);
+    real_exponent++;
+  }
+  mantissa &= ~(1ULL << 52);
+  // we have to check that real_exponent is in range, otherwise we bail out
+  if (simdjson_unlikely(real_exponent > 2046)) {
+    // We have an infinite value!!! We could actually throw an error here if we could.
+    return false;
+  }
+  d = to_double(mantissa, real_exponent, negative);
+  return true;
+}
+
+// We call a fallback floating-point parser that might be slow. Note
+// it will accept JSON numbers, but the JSON spec. is more restrictive so
+// before you call parse_float_fallback, you need to have validated the input
+// string with the JSON grammar.
+// It will return an error (false) if the parsed number is infinite.
+// The string parsing itself always succeeds. We know that there is at least
+// one digit.
+static bool parse_float_fallback(const uint8_t *ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+static bool parse_float_fallback(const uint8_t *ptr, const uint8_t *end_ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr), reinterpret_cast<const char *>(end_ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+// check quickly whether the next 8 chars are made of digits
+// at a glance, it looks better than Mula's
+// http://0x80.pl/articles/swar-digits-validate.html
+simdjson_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
+  uint64_t val;
+  // this can read up to 7 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(7 <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be bigger than 7");
+  std::memcpy(&val, chars, 8);
+  // a branchy method might be faster:
+  // return (( val & 0xF0F0F0F0F0F0F0F0 ) == 0x3030303030303030)
+  //  && (( (val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0 ) ==
+  //  0x3030303030303030);
+  return (((val & 0xF0F0F0F0F0F0F0F0) |
+           (((val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0) >> 4)) ==
+          0x3333333333333333);
+}
+
+template<typename I>
+SIMDJSON_NO_SANITIZE_UNDEFINED // We deliberately allow overflow here and check later
+simdjson_inline bool parse_digit(const uint8_t c, I &i) {
+  const uint8_t digit = static_cast<uint8_t>(c - '0');
+  if (digit > 9) {
+    return false;
+  }
+  // PERF NOTE: multiplication by 10 is cheaper than arbitrary integer multiplication
+  i = 10 * i + digit; // might overflow, we will handle the overflow later
+  return true;
+}
+
+simdjson_inline bool is_digit(const uint8_t c) {
+  return static_cast<uint8_t>(c - '0') <= 9;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_decimal_after_separator(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
+  // we continue with the fiction that we have an integer. If the
+  // floating point number is representable as x * 10^z for some integer
+  // z that fits in 53 bits, then we will be able to convert back the
+  // the integer into a float in a lossless manner.
+  const uint8_t *const first_after_period = p;
+
+#ifdef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_SWAR_NUMBER_PARSING
+  // this helps if we have lots of decimals!
+  // this turns out to be frequent enough.
+  if (is_made_of_eight_digits_fast(p)) {
+    i = i * 100000000 + parse_eight_digits_unrolled(p);
+    p += 8;
+  }
+#endif // SIMDJSON_SWAR_NUMBER_PARSING
+#endif // #ifdef SIMDJSON_SWAR_NUMBER_PARSING
+  // Unrolling the first digit makes a small difference on some implementations (e.g. westmere)
+  if (parse_digit(*p, i)) { ++p; }
+  while (parse_digit(*p, i)) { p++; }
+  exponent = first_after_period - p;
+  // Decimal without digits (123.) is illegal
+  if (exponent == 0) {
+    return INVALID_NUMBER(src);
+  }
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
+  // Exp Sign: -123.456e[-]78
+  bool neg_exp = ('-' == *p);
+  if (neg_exp || '+' == *p) { p++; } // Skip + as well
+
+  // Exponent: -123.456e-[78]
+  auto start_exp = p;
+  int64_t exp_number = 0;
+  while (parse_digit(*p, exp_number)) { ++p; }
+  // It is possible for parse_digit to overflow.
+  // In particular, it could overflow to INT64_MIN, and we cannot do - INT64_MIN.
+  // Thus we *must* check for possible overflow before we negate exp_number.
+
+  // Performance notes: it may seem like combining the two "simdjson_unlikely checks" below into
+  // a single simdjson_unlikely path would be faster. The reasoning is sound, but the compiler may
+  // not oblige and may, in fact, generate two distinct paths in any case. It might be
+  // possible to do uint64_t(p - start_exp - 1) >= 18 but it could end up trading off
+  // instructions for a simdjson_likely branch, an unconclusive gain.
+
+  // If there were no digits, it's an error.
+  if (simdjson_unlikely(p == start_exp)) {
+    return INVALID_NUMBER(src);
+  }
+  // We have a valid positive exponent in exp_number at this point, except that
+  // it may have overflowed.
+
+  // If there were more than 18 digits, we may have overflowed the integer. We have to do
+  // something!!!!
+  if (simdjson_unlikely(p > start_exp+18)) {
+    // Skip leading zeroes: 1e000000000000000000001 is technically valid and does not overflow
+    while (*start_exp == '0') { start_exp++; }
+    // 19 digits could overflow int64_t and is kind of absurd anyway. We don't
+    // support exponents smaller than -999,999,999,999,999,999 and bigger
+    // than 999,999,999,999,999,999.
+    // We can truncate.
+    // Note that 999999999999999999 is assuredly too large. The maximal ieee64 value before
+    // infinity is ~1.8e308. The smallest subnormal is ~5e-324. So, actually, we could
+    // truncate at 324.
+    // Note that there is no reason to fail per se at this point in time.
+    // E.g., 0e999999999999999999999 is a fine number.
+    if (p > start_exp+18) { exp_number = 999999999999999999; }
+  }
+  // At this point, we know that exp_number is a sane, positive, signed integer.
+  // It is <= 999,999,999,999,999,999. As long as 'exponent' is in
+  // [-8223372036854775808, 8223372036854775808], we won't overflow. Because 'exponent'
+  // is bounded in magnitude by the size of the JSON input, we are fine in this universe.
+  // To sum it up: the next line should never overflow.
+  exponent += (neg_exp ? -exp_number : exp_number);
+  return SUCCESS;
+}
+
+simdjson_inline bool check_if_integer(const uint8_t *const src, size_t max_length) {
+  const uint8_t *const srcend = src + max_length;
+  bool negative = (*src == '-'); // we can always read at least one character after the '-'
+  const uint8_t *p = src + uint8_t(negative);
+  if(p == srcend) { return false; }
+  if(*p == '0') {
+    ++p;
+    if(p == srcend) { return true; }
+    if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+    return true;
+  }
+  while(p != srcend && is_digit(*p)) { ++p; }
+  if(p == srcend) { return true; }
+  if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+  return true;
+}
+
+simdjson_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
+  // It is possible that the integer had an overflow.
+  // We have to handle the case where we have 0.0000somenumber.
+  const uint8_t *start = start_digits;
+  while ((*start == '0') || (*start == '.')) { ++start; }
+  // we over-decrement by one when there is a '.'
+  return digit_count - size_t(start - start_digits);
+}
+
+} // unnamed namespace
+
+/** @private */
+static error_code slow_float_parsing(simdjson_unused const uint8_t * src, double* answer) {
+  if (parse_float_fallback(src, answer)) {
+    return SUCCESS;
+  }
+  return INVALID_NUMBER(src);
+}
+
+/** @private */
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
+  // If we frequently had to deal with long strings of digits,
+  // we could extend our code by using a 128-bit integer instead
+  // of a 64-bit integer. However, this is uncommon in practice.
+  //
+  // 9999999999999999999 < 2**64 so we can accommodate 19 digits.
+  // If we have a decimal separator, then digit_count - 1 is the number of digits, but we
+  // may not have a decimal separator!
+  if (simdjson_unlikely(digit_count > 19 && significant_digits(start_digits, digit_count) > 19)) {
+    // Ok, chances are good that we had an overflow!
+    // this is almost never going to get called!!!
+    // we start anew, going slowly!!!
+    // This will happen in the following examples:
+    // 10000000000000000000000000000000000000000000e+308
+    // 3.1415926535897932384626433832795028841971693993751
+    //
+    // NOTE: We do not pass a reference to the to slow_float_parsing. If we passed our writer
+    // reference to it, it would force it to be stored in memory, preventing the compiler from
+    // picking it apart and putting into registers. i.e. if we pass it as reference,
+    // it gets slow.
+    double d;
+    error_code error = slow_float_parsing(src, &d);
+    writer.append_double(d);
+    return error;
+  }
+  // NOTE: it's weird that the simdjson_unlikely() only wraps half the if, but it seems to get slower any other
+  // way we've tried: https://github.com/simdjson/simdjson/pull/990#discussion_r448497331
+  // To future reader: we'd love if someone found a better way, or at least could explain this result!
+  if (simdjson_unlikely(exponent < simdjson::internal::smallest_power) || (exponent > simdjson::internal::largest_power)) {
+    //
+    // Important: smallest_power is such that it leads to a zero value.
+    // Observe that 18446744073709551615e-343 == 0, i.e. (2**64 - 1) e -343 is zero
+    // so something x 10^-343 goes to zero, but not so with  something x 10^-342.
+    static_assert(simdjson::internal::smallest_power <= -342, "smallest_power is not small enough");
+    //
+    if((exponent < simdjson::internal::smallest_power) || (i == 0)) {
+      // E.g. Parse "-0.0e-999" into the same value as "-0.0". See https://en.wikipedia.org/wiki/Signed_zero
+      WRITE_DOUBLE(negative ? -0.0 : 0.0, src, writer);
+      return SUCCESS;
+    } else { // (exponent > largest_power) and (i != 0)
+      // We have, for sure, an infinite value and simdjson refuses to parse infinite values.
+      return INVALID_NUMBER(src);
+    }
+  }
+  double d;
+  if (!compute_float_64(exponent, i, negative, d)) {
+    // we are almost never going to get here.
+    if (!parse_float_fallback(src, &d)) { return INVALID_NUMBER(src); }
+  }
+  WRITE_DOUBLE(d, src, writer);
+  return SUCCESS;
+}
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer);
+
+// for performance analysis, it is sometimes  useful to skip parsing
+#ifdef SIMDJSON_SKIPNUMBERPARSING
+
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const, W &writer) {
+  writer.append_s64(0);        // always write zero
+  return SUCCESS;              // always succeeds
+}
+
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept { return number_type::signed_integer; }
+#else
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer) {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  if (digit_count == 0 || ('0' == *start_digits && digit_count > 1)) { return INVALID_NUMBER(src); }
+
+  //
+  // Handle floats if there is a . or e (or both)
+  //
+  int64_t exponent = 0;
+  bool is_float = false;
+  if ('.' == *p) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_decimal_after_separator(src, p, i, exponent) );
+    digit_count = int(p - start_digits); // used later to guard against overflows
+  }
+  if (('e' == *p) || ('E' == *p)) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_exponent(src, p, exponent) );
+  }
+  if (is_float) {
+    const bool dirty_end = jsoncharutils::is_not_structural_or_whitespace(*p);
+    SIMDJSON_TRY( write_float(src, negative, i, start_digits, digit_count, exponent, writer) );
+    if (dirty_end) { return INVALID_NUMBER(src); }
+    return SUCCESS;
+  }
+
+  // The longest negative 64-bit number is 19 digits.
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  size_t longest_digit_count = negative ? 19 : 20;
+  if (digit_count > longest_digit_count) { return BIGINT_NUMBER(src); }
+  if (digit_count == longest_digit_count) {
+    if (negative) {
+      // Anything negative above INT64_MAX+1 is invalid
+      if (i > uint64_t(INT64_MAX)+1) { return BIGINT_NUMBER(src);  }
+      WRITE_INTEGER(~i+1, src, writer);
+      if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+      return SUCCESS;
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    }  else if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INVALID_NUMBER(src); }
+  }
+
+  // Write unsigned if it does not fit in a signed integer.
+  if (i > uint64_t(INT64_MAX)) {
+    WRITE_UNSIGNED(i, src, writer);
+  } else {
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+    if(i == 0 && negative) {
+      // We have to write -0.0 instead of 0
+      WRITE_DOUBLE(-0.0, src, writer);
+    } else {
+      WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+    }
+#else
+  WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+#endif
+  }
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+  return SUCCESS;
+}
+
+// Inlineable functions
+namespace {
+
+// This table can be used to characterize the final character of an integer
+// string. For JSON structural character and allowable white space characters,
+// we return SUCCESS. For 'e', '.' and 'E', we return INCORRECT_TYPE. Otherwise
+// we return NUMBER_ERROR.
+// Optimization note: we could easily reduce the size of the table by half (to 128)
+// at the cost of an extra branch.
+// Optimization note: we want the values to use at most 8 bits (not, e.g., 32 bits):
+static_assert(error_code(uint8_t(NUMBER_ERROR))== NUMBER_ERROR, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(SUCCESS))== SUCCESS, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(INCORRECT_TYPE))== INCORRECT_TYPE, "bad NUMBER_ERROR cast");
+
+const uint8_t integer_string_finisher[256] = {
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   SUCCESS,      NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, INCORRECT_TYPE,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, SUCCESS,        NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR};
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if ((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
+  const uint8_t *p = src + 1;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    // Note: we use src[1] and not src[0] because src[0] is the quote character in this
+    // instance.
+    if (src[1] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  //
+  // Check for minus sign
+  //
+  if(src == src_end) { return NUMBER_ERROR; }
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = src;
+  uint64_t i = 0;
+  while (parse_digit(*src, i)) { src++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(src - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*src)) {
+  //  return (*src == '.' || *src == 'e' || *src == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(*src != '"') { return NUMBER_ERROR; }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept {
+  return (*src == '-');
+}
+
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) { return true; }
+  return false;
+}
+
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  size_t digit_count = size_t(p - src);
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) {
+    static const uint8_t * smaller_big_integer = reinterpret_cast<const uint8_t *>("9223372036854775808");
+    // We have an integer.
+    if(simdjson_unlikely(digit_count > 20)) {
+      return number_type::big_integer;
+    }
+    // If the number is negative and valid, it must be a signed integer.
+    if(negative) {
+      if (simdjson_unlikely(digit_count > 19)) return number_type::big_integer;
+      if (simdjson_unlikely(digit_count == 19 && memcmp(src, smaller_big_integer, 19) > 0)) {
+        return number_type::big_integer;
+      }
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+      if(digit_count == 1 && src[0] == '0') {
+        // We have to write -0.0 instead of 0
+        return number_type::floating_point_number;
+      }
+#endif
+      return number_type::signed_integer;
+    }
+    // Let us check if we have a big integer (>=2**64).
+    static const uint8_t * two_to_sixtyfour = reinterpret_cast<const uint8_t *>("18446744073709551616");
+    if((digit_count > 20) || (digit_count == 20 && memcmp(src, two_to_sixtyfour, 20) >= 0)) {
+      return number_type::big_integer;
+    }
+    // The number is positive and smaller than 18446744073709551616 (or 2**64).
+    // We want values larger or equal to 9223372036854775808 to be unsigned
+    // integers, and the other values to be signed integers.
+    if((digit_count == 20) || (digit_count >= 19 && memcmp(src, smaller_big_integer, 19) >= 0)) {
+      return number_type::unsigned_integer;
+    }
+    return number_type::signed_integer;
+  }
+  // Hopefully, we have 'e' or 'E' or '.'.
+  return number_type::floating_point_number;
+}
+
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
+  if(src == src_end) { return NUMBER_ERROR; }
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  if(p == src_end) { return NUMBER_ERROR; }
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely((p != src_end) && (*p == '.'))) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if ((p == src_end) || !parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while ((p != src_end) && parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = start_digits-src > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if ((p != src_end) && (*p == 'e' || *p == 'E')) {
+    p++;
+    if(p == src_end) { return NUMBER_ERROR; }
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while ((p != src_end) && parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if ((p != src_end) && jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), src_end, &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+} // unnamed namespace
+#endif // SIMDJSON_SKIPNUMBERPARSING
+
+} // namespace numberparsing
+
+inline std::ostream& operator<<(std::ostream& out, number_type type) noexcept {
+    switch (type) {
+        case number_type::signed_integer: out << "integer in [-9223372036854775808,9223372036854775808)"; break;
+        case number_type::unsigned_integer: out << "unsigned integer in [9223372036854775808,18446744073709551616)"; break;
+        case number_type::floating_point_number: out << "floating-point number (binary64)"; break;
+        case number_type::big_integer: out << "big integer"; break;
+        default: SIMDJSON_UNREACHABLE();
+    }
+    return out;
+}
+
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_NUMBERPARSING_H
+/* end file simdjson/generic/numberparsing.h for arm64 */
+
+/* including simdjson/generic/implementation_simdjson_result_base-inl.h for arm64: #include "simdjson/generic/implementation_simdjson_result_base-inl.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base-inl.h for arm64 */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+
+//
+// internal::implementation_simdjson_result_base<T> inline implementation
+//
+
+template<typename T>
+simdjson_inline void implementation_simdjson_result_base<T>::tie(T &value, error_code &error) && noexcept {
+  error = this->second;
+  if (!error) {
+    value = std::forward<implementation_simdjson_result_base<T>>(*this).first;
+  }
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::get(T &value) && noexcept {
+  error_code error;
+  std::forward<implementation_simdjson_result_base<T>>(*this).tie(value, error);
+  return error;
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::error() const noexcept {
+  return this->second;
+}
+
+
+template<typename T>
+simdjson_warn_unused simdjson_inline bool implementation_simdjson_result_base<T>::has_value() const noexcept {
+  return this->error() == SUCCESS;
+}
+
+#if SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::operator*() &  noexcept(false) {
+  return this->value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::operator*() &&  noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).value();
+}
+
+template<typename T>
+simdjson_inline T* implementation_simdjson_result_base<T>::operator->() noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+
+template<typename T>
+simdjson_inline const T* implementation_simdjson_result_base<T>::operator->() const noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value() & noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::take_value() && noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::operator T&&() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+#endif // SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline const T& implementation_simdjson_result_base<T>::value_unsafe() const& noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value_unsafe() & noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value_unsafe() && noexcept {
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value, error_code error) noexcept
+    : first{std::forward<T>(value)}, second{error} {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(error_code error) noexcept
+    : implementation_simdjson_result_base(T{}, error) {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value) noexcept
+    : implementation_simdjson_result_base(std::forward<T>(value), SUCCESS) {}
+
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+/* end file simdjson/generic/implementation_simdjson_result_base-inl.h for arm64 */
+/* end file simdjson/generic/amalgamated.h for arm64 */
+/* including simdjson/arm64/end.h: #include "simdjson/arm64/end.h" */
+/* begin file simdjson/arm64/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#undef SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+/* undefining SIMDJSON_IMPLEMENTATION from "arm64" */
+#undef SIMDJSON_IMPLEMENTATION
+/* end file simdjson/arm64/end.h */
+
+#endif // SIMDJSON_ARM64_H
+/* end file simdjson/arm64.h */
+/* including simdjson/arm64/implementation.h: #include <simdjson/arm64/implementation.h> */
+/* begin file simdjson/arm64/implementation.h */
+#ifndef SIMDJSON_ARM64_IMPLEMENTATION_H
+#define SIMDJSON_ARM64_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/instruction_set.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation("arm64", "ARM NEON", internal::instruction_set::NEON) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_ARM64_IMPLEMENTATION_H
+/* end file simdjson/arm64/implementation.h */
+
+/* including simdjson/arm64/begin.h: #include <simdjson/arm64/begin.h> */
+/* begin file simdjson/arm64/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "arm64" */
+#define SIMDJSON_IMPLEMENTATION arm64
+/* including simdjson/arm64/base.h: #include "simdjson/arm64/base.h" */
+/* begin file simdjson/arm64/base.h */
+#ifndef SIMDJSON_ARM64_BASE_H
+#define SIMDJSON_ARM64_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+/**
+ * Implementation for NEON (ARMv8).
+ */
+namespace arm64 {
+
+class implementation;
+
+namespace {
+namespace simd {
+template <typename T> struct simd8;
+template <typename T> struct simd8x64;
+} // namespace simd
+} // unnamed namespace
+
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_ARM64_BASE_H
+/* end file simdjson/arm64/base.h */
+/* including simdjson/arm64/intrinsics.h: #include "simdjson/arm64/intrinsics.h" */
+/* begin file simdjson/arm64/intrinsics.h */
+#ifndef SIMDJSON_ARM64_INTRINSICS_H
+#define SIMDJSON_ARM64_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This should be the correct header whether
+// you use visual studio or other compilers.
+#include <arm_neon.h>
+
+static_assert(sizeof(uint8x16_t) <= simdjson::SIMDJSON_PADDING, "insufficient padding for arm64");
+
+#endif //  SIMDJSON_ARM64_INTRINSICS_H
+/* end file simdjson/arm64/intrinsics.h */
+/* including simdjson/arm64/bitmanipulation.h: #include "simdjson/arm64/bitmanipulation.h" */
+/* begin file simdjson/arm64/bitmanipulation.h */
+#ifndef SIMDJSON_ARM64_BITMANIPULATION_H
+#define SIMDJSON_ARM64_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  unsigned long ret;
+  // Search the mask data from least significant bit (LSB)
+  // to the most significant bit (MSB) for a set bit (1).
+  _BitScanForward64(&ret, input_num);
+  return (int)ret;
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO
+  return __builtin_ctzll(input_num);
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return input_num & (input_num-1);
+}
+
+// We sometimes call leading_zeroes on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+// Applies only when SIMDJSON_PREFER_REVERSE_BITS is defined and true.
+// (See below.)
+SIMDJSON_NO_SANITIZE_UNDEFINED
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  unsigned long leading_zero = 0;
+  // Search the mask data from most significant bit (MSB)
+  // to least significant bit (LSB) for a set bit (1).
+  if (_BitScanReverse64(&leading_zero, input_num))
+    return (int)(63 - leading_zero);
+  else
+    return 64;
+#else
+  return __builtin_clzll(input_num);
+#endif// SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int count_ones(uint64_t input_num) {
+   return vaddv_u8(vcnt_u8(vcreate_u8(input_num)));
+}
+
+
+#if defined(__GNUC__) // catches clang and gcc
+/**
+ * ARM has a fast 64-bit "bit reversal function" that is handy. However,
+ * it is not generally available as an intrinsic function under Visual
+ * Studio (though this might be changing). Even under clang/gcc, we
+ * apparently need to invoke inline assembly.
+ */
+/*
+ * We use SIMDJSON_PREFER_REVERSE_BITS as a hint that algorithms that
+ * work well with bit reversal may use it.
+ */
+#define SIMDJSON_PREFER_REVERSE_BITS 1
+
+/* reverse the bits */
+simdjson_inline uint64_t reverse_bits(uint64_t input_num) {
+  uint64_t rev_bits;
+  __asm("rbit %0, %1" : "=r"(rev_bits) : "r"(input_num));
+  return rev_bits;
+}
+
+/**
+ * Flips bit at index 63 - lz. Thus if you have 'leading_zeroes' leading zeroes,
+ * then this will set to zero the leading bit. It is possible for leading_zeroes to be
+ * greating or equal to 63 in which case we trigger undefined behavior, but the output
+ * of such undefined behavior is never used.
+ **/
+SIMDJSON_NO_SANITIZE_UNDEFINED
+simdjson_inline uint64_t zero_leading_bit(uint64_t rev_bits, int leading_zeroes) {
+  return rev_bits ^ (uint64_t(0x8000000000000000) >> leading_zeroes);
+}
+
+#endif
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2, uint64_t *result) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  *result = value1 + value2;
+  return *result < value1;
+#else
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+#endif
+}
+
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_ARM64_BITMANIPULATION_H
+/* end file simdjson/arm64/bitmanipulation.h */
+/* including simdjson/arm64/bitmask.h: #include "simdjson/arm64/bitmask.h" */
+/* begin file simdjson/arm64/bitmask.h */
+#ifndef SIMDJSON_ARM64_BITMASK_H
+#define SIMDJSON_ARM64_BITMASK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(uint64_t bitmask) {
+  /////////////
+  // We could do this with PMULL, but it is apparently slow.
+  //
+  //#ifdef __ARM_FEATURE_CRYPTO // some ARM processors lack this extension
+  //return vmull_p64(-1ULL, bitmask);
+  //#else
+  // Analysis by @sebpop:
+  // When diffing the assembly for src/stage1_find_marks.cpp I see that the eors are all spread out
+  // in between other vector code, so effectively the extra cycles of the sequence do not matter
+  // because the GPR units are idle otherwise and the critical path is on the FP side.
+  // Also the PMULL requires two extra fmovs: GPR->FP (3 cycles in N1, 5 cycles in A72 )
+  // and FP->GPR (2 cycles on N1 and 5 cycles on A72.)
+  ///////////
+  bitmask ^= bitmask << 1;
+  bitmask ^= bitmask << 2;
+  bitmask ^= bitmask << 4;
+  bitmask ^= bitmask << 8;
+  bitmask ^= bitmask << 16;
+  bitmask ^= bitmask << 32;
+  return bitmask;
+}
+
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif
+/* end file simdjson/arm64/bitmask.h */
+/* including simdjson/arm64/numberparsing_defs.h: #include "simdjson/arm64/numberparsing_defs.h" */
+/* begin file simdjson/arm64/numberparsing_defs.h */
+#ifndef SIMDJSON_ARM64_NUMBERPARSING_DEFS_H
+#define SIMDJSON_ARM64_NUMBERPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+#if SIMDJSON_REGULAR_VISUAL_STUDIO && SIMDJSON_IS_ARM64
+// __umulh requires intrin.h
+#include <intrin.h>
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO && SIMDJSON_IS_ARM64
+
+namespace simdjson {
+namespace arm64 {
+namespace numberparsing {
+
+// we don't have SSE, so let us use a scalar function
+// credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  uint64_t val;
+  std::memcpy(&val, chars, sizeof(uint64_t));
+  val = (val & 0x0F0F0F0F0F0F0F0F) * 2561 >> 8;
+  val = (val & 0x00FF00FF00FF00FF) * 6553601 >> 16;
+  return uint32_t((val & 0x0000FFFF0000FFFF) * 42949672960001 >> 32);
+}
+
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+#if SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+#if SIMDJSON_IS_ARM64
+  // ARM64 has native support for 64-bit multiplications, no need to emultate
+  answer.high = __umulh(value1, value2);
+  answer.low = value1 * value2;
+#else
+  answer.low = _umul128(value1, value2, &answer.high); // _umul128 not available on ARM64
+#endif // SIMDJSON_IS_ARM64
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+#endif
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace arm64
+} // namespace simdjson
+
+#ifndef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_IS_BIG_ENDIAN
+#define SIMDJSON_SWAR_NUMBER_PARSING 0
+#else
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+#endif
+#endif
+
+#endif // SIMDJSON_ARM64_NUMBERPARSING_DEFS_H
+/* end file simdjson/arm64/numberparsing_defs.h */
+/* including simdjson/arm64/simd.h: #include "simdjson/arm64/simd.h" */
+/* begin file simdjson/arm64/simd.h */
+#ifndef SIMDJSON_ARM64_SIMD_H
+#define SIMDJSON_ARM64_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace simd {
+
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+namespace {
+// Start of private section with Visual Studio workaround
+
+
+#ifndef simdjson_make_uint8x16_t
+#define simdjson_make_uint8x16_t(x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, \
+                             x13, x14, x15, x16)                                   \
+   ([=]() {                                                                        \
+     uint8_t array[16] = {x1, x2,  x3,  x4,  x5,  x6,  x7,  x8,                    \
+                                 x9, x10, x11, x12, x13, x14, x15, x16};           \
+     return vld1q_u8(array);                                                       \
+   }())
+#endif
+#ifndef simdjson_make_int8x16_t
+#define simdjson_make_int8x16_t(x1, x2, x3, x4, x5, x6, x7, x8, x9, x10, x11, x12, \
+                             x13, x14, x15, x16)                                  \
+   ([=]() {                                                                       \
+     int8_t array[16] = {x1, x2,  x3,  x4,  x5,  x6,  x7,  x8,                    \
+                                 x9, x10, x11, x12, x13, x14, x15, x16};          \
+     return vld1q_s8(array);                                                      \
+   }())
+#endif
+
+#ifndef simdjson_make_uint8x8_t
+#define simdjson_make_uint8x8_t(x1, x2, x3, x4, x5, x6, x7, x8)                \
+   ([=]() {                                                                    \
+     uint8_t array[8] = {x1, x2,  x3,  x4,  x5,  x6,  x7,  x8};                \
+     return vld1_u8(array);                                                    \
+   }())
+#endif
+#ifndef simdjson_make_int8x8_t
+#define simdjson_make_int8x8_t(x1, x2, x3, x4, x5, x6, x7, x8)                 \
+   ([=]() {                                                                    \
+     int8_t array[8] = {x1, x2,  x3,  x4,  x5,  x6,  x7,  x8};                 \
+     return vld1_s8(array);                                                    \
+   }())
+#endif
+#ifndef simdjson_make_uint16x8_t
+#define simdjson_make_uint16x8_t(x1, x2, x3, x4, x5, x6, x7, x8)               \
+   ([=]() {                                                                    \
+     uint16_t array[8] = {x1, x2,  x3,  x4,  x5,  x6,  x7,  x8};               \
+     return vld1q_u16(array);                                                  \
+   }())
+#endif
+#ifndef simdjson_make_int16x8_t
+#define simdjson_make_int16x8_t(x1, x2, x3, x4, x5, x6, x7, x8)                \
+   ([=]() {                                                                    \
+     int16_t array[8] = {x1, x2,  x3,  x4,  x5,  x6,  x7,  x8};                \
+     return vld1q_s16(array);                                                  \
+   }())
+#endif
+
+// End of private section with Visual Studio workaround
+} // namespace
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+
+
+  template<typename T>
+  struct simd8;
+
+  //
+  // Base class of simd8<uint8_t> and simd8<bool>, both of which use uint8x16_t internally.
+  //
+  template<typename T, typename Mask=simd8<bool>>
+  struct base_u8 {
+    uint8x16_t value;
+    static const int SIZE = sizeof(value);
+
+    // Conversion from/to SIMD register
+    simdjson_inline base_u8(const uint8x16_t _value) : value(_value) {}
+    simdjson_inline operator const uint8x16_t&() const { return this->value; }
+    simdjson_inline operator uint8x16_t&() { return this->value; }
+
+    // Bit operations
+    simdjson_inline simd8<T> operator|(const simd8<T> other) const { return vorrq_u8(*this, other); }
+    simdjson_inline simd8<T> operator&(const simd8<T> other) const { return vandq_u8(*this, other); }
+    simdjson_inline simd8<T> operator^(const simd8<T> other) const { return veorq_u8(*this, other); }
+    simdjson_inline simd8<T> bit_andnot(const simd8<T> other) const { return vbicq_u8(*this, other); }
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+    simdjson_inline simd8<T>& operator|=(const simd8<T> other) { auto this_cast = static_cast<simd8<T>*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline simd8<T>& operator&=(const simd8<T> other) { auto this_cast = static_cast<simd8<T>*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline simd8<T>& operator^=(const simd8<T> other) { auto this_cast = static_cast<simd8<T>*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+
+    friend simdjson_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return vceqq_u8(lhs, rhs); }
+
+    template<int N=1>
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+      return vextq_u8(prev_chunk, *this, 16 - N);
+    }
+  };
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool>: base_u8<bool> {
+    typedef uint16_t bitmask_t;
+    typedef uint32_t bitmask2_t;
+
+    static simdjson_inline simd8<bool> splat(bool _value) { return vmovq_n_u8(uint8_t(-(!!_value))); }
+
+    simdjson_inline simd8(const uint8x16_t _value) : base_u8<bool>(_value) {}
+    // False constructor
+    simdjson_inline simd8() : simd8(vdupq_n_u8(0)) {}
+    // Splat constructor
+    simdjson_inline simd8(bool _value) : simd8(splat(_value)) {}
+
+    // We return uint32_t instead of uint16_t because that seems to be more efficient for most
+    // purposes (cutting it down to uint16_t costs performance in some compilers).
+    simdjson_inline uint32_t to_bitmask() const {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+      const uint8x16_t bit_mask =  simdjson_make_uint8x16_t(0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+                                                   0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80);
+#else
+      const uint8x16_t bit_mask =  {0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+                                    0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80};
+#endif
+      auto minput = *this & bit_mask;
+      uint8x16_t tmp = vpaddq_u8(minput, minput);
+      tmp = vpaddq_u8(tmp, tmp);
+      tmp = vpaddq_u8(tmp, tmp);
+      return vgetq_lane_u16(vreinterpretq_u16_u8(tmp), 0);
+    }
+    // Returns 4-bit out of each byte, alternating between the high 4 bits and low
+    // bits result it is 64 bit.
+    simdjson_inline uint64_t to_bitmask64() const {
+      return vget_lane_u64(
+          vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(*this), 4)), 0);
+    }
+    simdjson_inline bool any() const { return vmaxvq_u32(vreinterpretq_u32_u8(*this)) != 0; }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t>: base_u8<uint8_t> {
+    static simdjson_inline uint8x16_t splat(uint8_t _value) { return vmovq_n_u8(_value); }
+    static simdjson_inline uint8x16_t zero() { return vdupq_n_u8(0); }
+    static simdjson_inline uint8x16_t load(const uint8_t* values) { return vld1q_u8(values); }
+
+    simdjson_inline simd8(const uint8x16_t _value) : base_u8<uint8_t>(_value) {}
+    // Zero constructor
+    simdjson_inline simd8() : simd8(zero()) {}
+    // Array constructor
+    simdjson_inline simd8(const uint8_t values[16]) : simd8(load(values)) {}
+    // Splat constructor
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    // Member-by-member initialization
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) : simd8(simdjson_make_uint8x16_t(
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    )) {}
+#else
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) : simd8(uint8x16_t{
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    }) {}
+#endif
+
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Store to array
+    simdjson_inline void store(uint8_t dst[16]) const { return vst1q_u8(dst, *this); }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return vqaddq_u8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return vqsubq_u8(*this, other); }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<uint8_t> operator+(const simd8<uint8_t> other) const { return vaddq_u8(*this, other); }
+    simdjson_inline simd8<uint8_t> operator-(const simd8<uint8_t> other) const { return vsubq_u8(*this, other); }
+    simdjson_inline simd8<uint8_t>& operator+=(const simd8<uint8_t> other) { *this = *this + other; return *this; }
+    simdjson_inline simd8<uint8_t>& operator-=(const simd8<uint8_t> other) { *this = *this - other; return *this; }
+
+    // Order-specific operations
+    simdjson_inline uint8_t max_val() const { return vmaxvq_u8(*this); }
+    simdjson_inline uint8_t min_val() const { return vminvq_u8(*this); }
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return vmaxq_u8(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return vminq_u8(*this, other); }
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return vcleq_u8(*this, other); }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return vcgeq_u8(*this, other); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return vcltq_u8(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return vcgtq_u8(*this, other); }
+    // Same as >, but instead of guaranteeing all 1's == true, false = 0 and true = nonzero. For ARM, returns all 1's.
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return simd8<uint8_t>(*this > other); }
+    // Same as <, but instead of guaranteeing all 1's == true, false = 0 and true = nonzero. For ARM, returns all 1's.
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return simd8<uint8_t>(*this < other); }
+
+    // Bit-specific operations
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return vtstq_u8(*this, bits); }
+    simdjson_inline bool any_bits_set_anywhere() const { return vmaxvq_u32(vreinterpretq_u32_u8(*this)) != 0; }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return (*this & bits).any_bits_set_anywhere(); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return vshrq_n_u8(*this, N); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return vshlq_n_u8(*this, N); }
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return lookup_table.apply_lookup_16_to(*this);
+    }
+
+    // Returns 4-bit out of each byte, alternating between the high 4 bits and low
+    // bits result it is 64 bit.
+    simdjson_inline uint64_t to_bitmask64() const {
+      return vget_lane_u64(
+          vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(*this), 4)), 0);
+    }
+    // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted as a bitset).
+    // Passing a 0 value for mask would be equivalent to writing out every byte to output.
+    // Only the first 16 - count_ones(mask) bytes of the result are significant but 16 bytes
+    // get written.
+    // Design consideration: it seems like a function with the
+    // signature simd8<L> compress(uint16_t mask) would be
+    // sensible, but the AVX ISA makes this kind of approach difficult.
+    template<typename L>
+    simdjson_inline void compress(uint16_t mask, L * output) const {
+      using internal::thintable_epi8;
+      using internal::BitsSetTable256mul2;
+      using internal::pshufb_combine_table;
+      // this particular implementation was inspired by work done by @animetosho
+      // we do it in two steps, first 8 bytes and then second 8 bytes
+      uint8_t mask1 = uint8_t(mask); // least significant 8 bits
+      uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
+      // next line just loads the 64-bit values thintable_epi8[mask1] and
+      // thintable_epi8[mask2] into a 128-bit register, using only
+      // two instructions on most compilers.
+      uint64x2_t shufmask64 = {thintable_epi8[mask1], thintable_epi8[mask2]};
+      uint8x16_t shufmask = vreinterpretq_u8_u64(shufmask64);
+      // we increment by 0x08 the second half of the mask
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+      uint8x16_t inc = simdjson_make_uint8x16_t(0, 0, 0, 0, 0, 0, 0, 0, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08);
+#else
+      uint8x16_t inc = {0, 0, 0, 0, 0, 0, 0, 0, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08};
+#endif
+      shufmask = vaddq_u8(shufmask, inc);
+      // this is the version "nearly pruned"
+      uint8x16_t pruned = vqtbl1q_u8(*this, shufmask);
+      // we still need to put the two halves together.
+      // we compute the popcount of the first half:
+      int pop1 = BitsSetTable256mul2[mask1];
+      // then load the corresponding mask, what it does is to write
+      // only the first pop1 bytes from the first 8 bytes, and then
+      // it fills in with the bytes from the second 8 bytes + some filling
+      // at the end.
+      uint8x16_t compactmask = vld1q_u8(reinterpret_cast<const uint8_t *>(pshufb_combine_table + pop1 * 8));
+      uint8x16_t answer = vqtbl1q_u8(pruned, compactmask);
+      vst1q_u8(reinterpret_cast<uint8_t*>(output), answer);
+    }
+
+    // Copies all bytes corresponding to a 0 in the low half of the mask (interpreted as a
+    // bitset) to output1, then those corresponding to a 0 in the high half to output2.
+    template<typename L>
+    simdjson_inline void compress_halves(uint16_t mask, L *output1, L *output2) const {
+      using internal::thintable_epi8;
+      uint8_t mask1 = uint8_t(mask); // least significant 8 bits
+      uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
+      uint8x8_t compactmask1 = vcreate_u8(thintable_epi8[mask1]);
+      uint8x8_t compactmask2 = vcreate_u8(thintable_epi8[mask2]);
+      // we increment by 0x08 the second half of the mask
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+      uint8x8_t inc = simdjson_make_uint8x8_t(0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08);
+#else
+      uint8x8_t inc = {0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08};
+#endif
+      compactmask2 = vadd_u8(compactmask2, inc);
+      // store each result (with the second store possibly overlapping the first)
+      vst1_u8((uint8_t*)output1, vqtbl1_u8(*this, compactmask1));
+      vst1_u8((uint8_t*)output2, vqtbl1_u8(*this, compactmask2));
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+
+    template<typename T>
+    simdjson_inline simd8<uint8_t> apply_lookup_16_to(const simd8<T> original) {
+      return vqtbl1q_u8(*this, simd8<uint8_t>(original));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> {
+    int8x16_t value;
+
+    static simdjson_inline simd8<int8_t> splat(int8_t _value) { return vmovq_n_s8(_value); }
+    static simdjson_inline simd8<int8_t> zero() { return vdupq_n_s8(0); }
+    static simdjson_inline simd8<int8_t> load(const int8_t values[16]) { return vld1q_s8(values); }
+
+    // Conversion from/to SIMD register
+    simdjson_inline simd8(const int8x16_t _value) : value{_value} {}
+    simdjson_inline operator const int8x16_t&() const { return this->value; }
+    simdjson_inline operator int8x16_t&() { return this->value; }
+
+    // Zero constructor
+    simdjson_inline simd8() : simd8(zero()) {}
+    // Splat constructor
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const int8_t* values) : simd8(load(values)) {}
+    // Member-by-member initialization
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3, int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) : simd8(simdjson_make_int8x16_t(
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    )) {}
+#else
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3, int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) : simd8(int8x16_t{
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    }) {}
+#endif
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Store to array
+    simdjson_inline void store(int8_t dst[16]) const { return vst1q_s8(dst, *this); }
+
+    // Explicit conversion to/from unsigned
+    //
+    // Under Visual Studio/ARM64 uint8x16_t and int8x16_t are apparently the same type.
+    // In theory, we could check this occurrence with std::same_as and std::enabled_if but it is C++14
+    // and relatively ugly and hard to read.
+#ifndef SIMDJSON_REGULAR_VISUAL_STUDIO
+    simdjson_inline explicit simd8(const uint8x16_t other): simd8(vreinterpretq_s8_u8(other)) {}
+#endif
+    simdjson_inline explicit operator simd8<uint8_t>() const { return vreinterpretq_u8_s8(this->value); }
+
+    // Math
+    simdjson_inline simd8<int8_t> operator+(const simd8<int8_t> other) const { return vaddq_s8(*this, other); }
+    simdjson_inline simd8<int8_t> operator-(const simd8<int8_t> other) const { return vsubq_s8(*this, other); }
+    simdjson_inline simd8<int8_t>& operator+=(const simd8<int8_t> other) { *this = *this + other; return *this; }
+    simdjson_inline simd8<int8_t>& operator-=(const simd8<int8_t> other) { *this = *this - other; return *this; }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return vmaxq_s8(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return vminq_s8(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return vcgtq_s8(*this, other); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return vcltq_s8(*this, other); }
+    simdjson_inline simd8<bool> operator==(const simd8<int8_t> other) const { return vceqq_s8(*this, other); }
+
+    template<int N=1>
+    simdjson_inline simd8<int8_t> prev(const simd8<int8_t> prev_chunk) const {
+      return vextq_s8(prev_chunk, *this, 16 - N);
+    }
+
+    // Perform a lookup assuming no value is larger than 16
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return lookup_table.apply_lookup_16_to(*this);
+    }
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+
+    template<typename T>
+    simdjson_inline simd8<int8_t> apply_lookup_16_to(const simd8<T> original) {
+      return vqtbl1q_s8(*this, simd8<uint8_t>(original));
+    }
+  };
+
+  template<typename T>
+  struct simd8x64 {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+    static_assert(NUM_CHUNKS == 4, "ARM kernel should use four registers per 64-byte block.");
+    const simd8<T> chunks[NUM_CHUNKS];
+    template<int idx> simd8<uint8_t> get() const { return idx < NUM_CHUNKS ? chunks[idx] : simd8<T>(); }
+
+    simd8x64(const simd8x64<T>& o) = delete; // no copy allowed
+    simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1, const simd8<T> chunk2, const simd8<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+16), simd8<T>::load(ptr+32), simd8<T>::load(ptr+48)} {}
+
+    simdjson_inline void store(T ptr[64]) const {
+      this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
+      this->chunks[1].store(ptr+sizeof(simd8<T>)*1);
+      this->chunks[2].store(ptr+sizeof(simd8<T>)*2);
+      this->chunks[3].store(ptr+sizeof(simd8<T>)*3);
+    }
+
+    simdjson_inline simd8<T> reduce_or() const {
+      return (this->chunks[0] | this->chunks[1]) | (this->chunks[2] | this->chunks[3]);
+    }
+
+
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
+      uint64_t popcounts = vget_lane_u64(vreinterpret_u64_u8(vcnt_u8(vcreate_u8(~mask))), 0);
+      // compute the prefix sum of the popcounts of each byte
+      uint64_t offsets = popcounts * 0x0101010101010101;
+      this->chunks[0].compress_halves(uint16_t(mask), output, &output[popcounts & 0xFF]);
+      this->chunks[1].compress_halves(uint16_t(mask >> 16), &output[(offsets >> 8) & 0xFF], &output[(offsets >> 16) & 0xFF]);
+      this->chunks[2].compress_halves(uint16_t(mask >> 32), &output[(offsets >> 24) & 0xFF], &output[(offsets >> 32) & 0xFF]);
+      this->chunks[3].compress_halves(uint16_t(mask >> 48), &output[(offsets >> 40) & 0xFF], &output[(offsets >> 48) & 0xFF]);
+      return offsets >> 56;
+    }
+
+    simdjson_inline uint64_t to_bitmask() const {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+      const uint8x16_t bit_mask = simdjson_make_uint8x16_t(
+        0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80
+      );
+#else
+      const uint8x16_t bit_mask = {
+        0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
+        0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80
+      };
+#endif
+      // Add each of the elements next to each other, successively, to stuff each 8 byte mask into one.
+      uint8x16_t sum0 = vpaddq_u8(this->chunks[0] & bit_mask, this->chunks[1] & bit_mask);
+      uint8x16_t sum1 = vpaddq_u8(this->chunks[2] & bit_mask, this->chunks[3] & bit_mask);
+      sum0 = vpaddq_u8(sum0, sum1);
+      sum0 = vpaddq_u8(sum0, sum0);
+      return vgetq_lane_u64(vreinterpretq_u64_u8(sum0), 0);
+    }
+
+    simdjson_inline uint64_t eq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] == mask,
+        this->chunks[1] == mask,
+        this->chunks[2] == mask,
+        this->chunks[3] == mask
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t lteq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] <= mask,
+        this->chunks[1] <= mask,
+        this->chunks[2] <= mask,
+        this->chunks[3] <= mask
+      ).to_bitmask();
+    }
+  }; // struct simd8x64<T>
+
+} // namespace simd
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_ARM64_SIMD_H
+/* end file simdjson/arm64/simd.h */
+/* including simdjson/arm64/stringparsing_defs.h: #include "simdjson/arm64/stringparsing_defs.h" */
+/* begin file simdjson/arm64/stringparsing_defs.h */
+#ifndef SIMDJSON_ARM64_STRINGPARSING_DEFS_H
+#define SIMDJSON_ARM64_STRINGPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/simd.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 32;
+  simdjson_inline backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return bs_bits != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+
+  uint32_t bs_bits;
+  uint32_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // this can read up to 31 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v0(src);
+  simd8<uint8_t> v1(src + sizeof(v0));
+  v0.store(dst);
+  v1.store(dst + sizeof(v0));
+
+  // Getting a 64-bit bitmask is much cheaper than multiple 16-bit bitmasks on ARM; therefore, we
+  // smash them together into a 64-byte mask and get the bitmask from there.
+  uint64_t bs_and_quote = simd8x64<bool>(v0 == '\\', v1 == '\\', v0 == '"', v1 == '"').to_bitmask();
+  return {
+    uint32_t(bs_and_quote),      // bs_bits
+    uint32_t(bs_and_quote >> 32) // quote_bits
+  };
+}
+
+struct escaping {
+  static constexpr uint32_t BYTES_PROCESSED = 16;
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits != 0; }
+  simdjson_inline int escape_index() { return trailing_zeroes(escape_bits) / 4; }
+
+  uint64_t escape_bits;
+}; // struct escaping
+
+
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "escaping finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  simd8<bool> is_quote = (v == '"');
+  simd8<bool> is_backslash = (v == '\\');
+  simd8<bool> is_control = (v < 32);
+  return {
+    (is_backslash | is_quote | is_control).to_bitmask64()
+  };
+}
+
+
+
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_ARM64_STRINGPARSING_DEFS_H
+/* end file simdjson/arm64/stringparsing_defs.h */
+
+#define SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT 1
+/* end file simdjson/arm64/begin.h */
+/* including generic/amalgamated.h for arm64: #include <generic/amalgamated.h> */
+/* begin file generic/amalgamated.h for arm64 */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_SRC_GENERIC_DEPENDENCIES_H)
+#error generic/dependencies.h must be included before generic/amalgamated.h!
+#endif
+
+/* including generic/base.h for arm64: #include <generic/base.h> */
+/* begin file generic/base.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+
+struct json_character_block;
+
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_BASE_H
+/* end file generic/base.h for arm64 */
+/* including generic/dom_parser_implementation.h for arm64: #include <generic/dom_parser_implementation.h> */
+/* begin file generic/dom_parser_implementation.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// Interface a dom parser implementation must fulfill
+namespace simdjson {
+namespace arm64 {
+namespace {
+
+simdjson_inline simd8<uint8_t> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3);
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input);
+
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file generic/dom_parser_implementation.h for arm64 */
+/* including generic/json_character_block.h for arm64: #include <generic/json_character_block.h> */
+/* begin file generic/json_character_block.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+
+struct json_character_block {
+  static simdjson_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
+
+  simdjson_inline uint64_t whitespace() const noexcept { return _whitespace; }
+  simdjson_inline uint64_t op() const noexcept { return _op; }
+  simdjson_inline uint64_t scalar() const noexcept { return ~(op() | whitespace()); }
+
+  uint64_t _whitespace;
+  uint64_t _op;
+};
+
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H
+/* end file generic/json_character_block.h for arm64 */
+/* end file generic/amalgamated.h for arm64 */
+/* including generic/stage1/amalgamated.h for arm64: #include <generic/stage1/amalgamated.h> */
+/* begin file generic/stage1/amalgamated.h for arm64 */
+// Stuff other things depend on
+/* including generic/stage1/base.h for arm64: #include <generic/stage1/base.h> */
+/* begin file generic/stage1/base.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_BASE_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace stage1 {
+
+class bit_indexer;
+template<size_t STEP_SIZE>
+struct buf_block_reader;
+struct json_block;
+class json_minifier;
+class json_scanner;
+struct json_string_block;
+class json_string_scanner;
+class json_structural_indexer;
+
+} // namespace stage1
+
+namespace utf8_validation {
+struct utf8_checker;
+} // namespace utf8_validation
+
+using utf8_validation::utf8_checker;
+
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_BASE_H
+/* end file generic/stage1/base.h for arm64 */
+/* including generic/stage1/buf_block_reader.h for arm64: #include <generic/stage1/buf_block_reader.h> */
+/* begin file generic/stage1/buf_block_reader.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace stage1 {
+
+// Walks through a buffer in block-sized increments, loading the last part with spaces
+template<size_t STEP_SIZE>
+struct buf_block_reader {
+public:
+  simdjson_inline buf_block_reader(const uint8_t *_buf, size_t _len);
+  simdjson_inline size_t block_index();
+  simdjson_inline bool has_full_block() const;
+  simdjson_inline const uint8_t *full_block() const;
+  /**
+   * Get the last block, padded with spaces.
+   *
+   * There will always be a last block, with at least 1 byte, unless len == 0 (in which case this
+   * function fills the buffer with spaces and returns 0. In particular, if len == STEP_SIZE there
+   * will be 0 full_blocks and 1 remainder block with STEP_SIZE bytes and no spaces for padding.
+   *
+   * @return the number of effective characters in the last block.
+   */
+  simdjson_inline size_t get_remainder(uint8_t *dst) const;
+  simdjson_inline void advance();
+private:
+  const uint8_t *buf;
+  const size_t len;
+  const size_t lenminusstep;
+  size_t idx;
+};
+
+// Routines to print masks and text for debugging bitmask operations
+simdjson_unused static char * format_input_text_64(const uint8_t *text) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    buf[i] = int8_t(text[i]) < ' ' ? '_' : int8_t(text[i]);
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+// Routines to print masks and text for debugging bitmask operations
+simdjson_unused static char * format_input_text(const simd8x64<uint8_t>& in) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  in.store(reinterpret_cast<uint8_t*>(buf));
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] < ' ') { buf[i] = '_'; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+simdjson_unused static char * format_input_text(const simd8x64<uint8_t>& in, uint64_t mask) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  in.store(reinterpret_cast<uint8_t*>(buf));
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] <= ' ') { buf[i] = '_'; }
+    if (!(mask & (size_t(1) << i))) { buf[i] = ' '; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+simdjson_unused static char * format_mask(uint64_t mask) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  for (size_t i=0; i<64; i++) {
+    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
+  }
+  buf[64] = '\0';
+  return buf;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
+
+template<size_t STEP_SIZE>
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
+
+template<size_t STEP_SIZE>
+simdjson_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
+  return idx < lenminusstep;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
+  return &buf[idx];
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
+  if(len == idx) { return 0; } // memcpy(dst, null, 0) will trigger an error with some sanitizers
+  std::memset(dst, 0x20, STEP_SIZE); // std::memset STEP_SIZE because it's more efficient to write out 8 or 16 bytes at once.
+  std::memcpy(dst, buf + idx, len - idx);
+  return len - idx;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline void buf_block_reader<STEP_SIZE>::advance() {
+  idx += STEP_SIZE;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H
+/* end file generic/stage1/buf_block_reader.h for arm64 */
+/* including generic/stage1/json_escape_scanner.h for arm64: #include <generic/stage1/json_escape_scanner.h> */
+/* begin file generic/stage1/json_escape_scanner.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_ESCAPE_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_ESCAPE_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace stage1 {
+
+/**
+ * Scans for escape characters in JSON, taking care with multiple backslashes (\\n vs. \n).
+ */
+struct json_escape_scanner {
+  /** The actual escape characters (the backslashes themselves). */
+  uint64_t next_is_escaped = 0ULL;
+
+  struct escaped_and_escape {
+    /**
+     * Mask of escaped characters.
+     *
+     * ```
+     * \n \\n \\\n \\\\n \
+     * 0100100010100101000
+     *  n  \   \ n  \ \
+     * ```
+     */
+    uint64_t escaped;
+    /**
+     * Mask of escape characters.
+     *
+     * ```
+     * \n \\n \\\n \\\\n \
+     * 1001000101001010001
+     * \  \   \ \  \ \   \
+     * ```
+     */
+    uint64_t escape;
+  };
+
+  /**
+   * Get a mask of both escape and escaped characters (the characters following a backslash).
+   *
+   * @param potential_escape A mask of the character that can escape others (but could be
+   *        escaped itself). e.g. block.eq('\\')
+   */
+  simdjson_really_inline escaped_and_escape next(uint64_t backslash) noexcept {
+
+#if !SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+    if (!backslash) { return {next_escaped_without_backslashes(), 0}; }
+#endif
+
+    // |                                | Mask (shows characters instead of 1's) | Depth | Instructions        |
+    // |--------------------------------|----------------------------------------|-------|---------------------|
+    // | string                         | `\\n_\\\n___\\\n___\\\\___\\\\__\\\`   |       |                     |
+    // |                                | `    even   odd    even   odd   odd`   |       |                     |
+    // | potential_escape               | ` \  \\\    \\\    \\\\   \\\\  \\\`   | 1     | 1 (backslash & ~first_is_escaped)
+    // | escape_and_terminal_code       | ` \n \ \n   \ \n   \ \    \ \   \ \`   | 5     | 5 (next_escape_and_terminal_code())
+    // | escaped                        | `\    \ n    \ n    \ \    \ \   \ ` X | 6     | 7 (escape_and_terminal_code ^ (potential_escape | first_is_escaped))
+    // | escape                         | `    \ \    \ \    \ \    \ \   \ \`   | 6     | 8 (escape_and_terminal_code & backslash)
+    // | first_is_escaped               | `\                                 `   | 7 (*) | 9 (escape >> 63) ()
+    //                                                                               (*) this is not needed until the next iteration
+    uint64_t escape_and_terminal_code = next_escape_and_terminal_code(backslash & ~this->next_is_escaped);
+    uint64_t escaped = escape_and_terminal_code ^ (backslash | this->next_is_escaped);
+    uint64_t escape = escape_and_terminal_code & backslash;
+    this->next_is_escaped = escape >> 63;
+    return {escaped, escape};
+  }
+
+private:
+  static constexpr const uint64_t ODD_BITS = 0xAAAAAAAAAAAAAAAAULL;
+
+  simdjson_really_inline uint64_t next_escaped_without_backslashes() noexcept {
+    uint64_t escaped = this->next_is_escaped;
+    this->next_is_escaped = 0;
+    return escaped;
+  }
+
+  /**
+   * Returns a mask of the next escape characters (masking out escaped backslashes), along with
+   * any non-backslash escape codes.
+   *
+   * \n \\n \\\n \\\\n returns:
+   * \n \   \ \n \ \
+   * 11 100 1011 10100
+   *
+   * You are expected to mask out the first bit yourself if the previous block had a trailing
+   * escape.
+   *
+   * & the result with potential_escape to get just the escape characters.
+   * ^ the result with (potential_escape | first_is_escaped) to get escaped characters.
+   */
+  static simdjson_really_inline uint64_t next_escape_and_terminal_code(uint64_t potential_escape) noexcept {
+    // If we were to just shift and mask out any odd bits, we'd actually get a *half* right answer:
+    // any even-aligned backslash runs would be correct! Odd-aligned backslash runs would be
+    // inverted (\\\ would be 010 instead of 101).
+    //
+    // ```
+    // string:              | ____\\\\_\\\\_____ |
+    // maybe_escaped | ODD  |     \ \   \ \      |
+    //               even-aligned ^^^  ^^^^ odd-aligned
+    // ```
+    //
+    // Taking that into account, our basic strategy is:
+    //
+    // 1. Use subtraction to produce a mask with 1's for even-aligned runs and 0's for
+    //    odd-aligned runs.
+    // 2. XOR all odd bits, which masks out the odd bits in even-aligned runs, and brings IN the
+    //    odd bits in odd-aligned runs.
+    // 3. & with backslash to clean up any stray bits.
+    // runs are set to 0, and then XORing with "odd":
+    //
+    // |                                | Mask (shows characters instead of 1's) | Instructions        |
+    // |--------------------------------|----------------------------------------|---------------------|
+    // | string                         | `\\n_\\\n___\\\n___\\\\___\\\\__\\\`   |
+    // |                                | `    even   odd    even   odd   odd`   |
+    // | maybe_escaped                  | `  n  \\n    \\n    \\\_   \\\_  \\` X | 1 (potential_escape << 1)
+    // | maybe_escaped_and_odd          | ` \n_ \\n _ \\\n_ _ \\\__ _\\\_ \\\`   | 1 (maybe_escaped | odd)
+    // | even_series_codes_and_odd      | `  n_\\\  _    n_ _\\\\ _     _    `   | 1 (maybe_escaped_and_odd - potential_escape)
+    // | escape_and_terminal_code       | ` \n \ \n   \ \n   \ \    \ \   \ \`   | 1 (^ odd)
+    //
+
+    // Escaped characters are characters following an escape.
+    uint64_t maybe_escaped = potential_escape << 1;
+
+    // To distinguish odd from even escape sequences, therefore, we turn on any *starting*
+    // escapes that are on an odd byte. (We actually bring in all odd bits, for speed.)
+    // - Odd runs of backslashes are 0000, and the code at the end ("n" in \n or \\n) is 1.
+    // - Odd runs of backslashes are 1111, and the code at the end ("n" in \n or \\n) is 0.
+    // - All other odd bytes are 1, and even bytes are 0.
+    uint64_t maybe_escaped_and_odd_bits     = maybe_escaped | ODD_BITS;
+    uint64_t even_series_codes_and_odd_bits = maybe_escaped_and_odd_bits - potential_escape;
+
+    // Now we flip all odd bytes back with xor. This:
+    // - Makes odd runs of backslashes go from 0000 to 1010
+    // - Makes even runs of backslashes go from 1111 to 1010
+    // - Sets actually-escaped codes to 1 (the n in \n and \\n: \n = 11, \\n = 100)
+    // - Resets all other bytes to 0
+    return even_series_codes_and_odd_bits ^ ODD_BITS;
+  }
+};
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+/* end file generic/stage1/json_escape_scanner.h for arm64 */
+/* including generic/stage1/json_string_scanner.h for arm64: #include <generic/stage1/json_string_scanner.h> */
+/* begin file generic/stage1/json_string_scanner.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_escape_scanner.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace stage1 {
+
+struct json_string_block {
+  // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
+  simdjson_really_inline json_string_block(uint64_t escaped, uint64_t quote, uint64_t in_string) :
+  _escaped(escaped), _quote(quote), _in_string(in_string) {}
+
+  // Escaped characters (characters following an escape() character)
+  simdjson_really_inline uint64_t escaped() const { return _escaped; }
+  // Real (non-backslashed) quotes
+  simdjson_really_inline uint64_t quote() const { return _quote; }
+  // Only characters inside the string (not including the quotes)
+  simdjson_really_inline uint64_t string_content() const { return _in_string & ~_quote; }
+  // Return a mask of whether the given characters are inside a string (only works on non-quotes)
+  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
+  // Return a mask of whether the given characters are inside a string (only works on non-quotes)
+  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
+  // Tail of string (everything except the start quote)
+  simdjson_really_inline uint64_t string_tail() const { return _in_string ^ _quote; }
+
+  // escaped characters (backslashed--does not include the hex characters after \u)
+  uint64_t _escaped;
+  // real quotes (non-escaped ones)
+  uint64_t _quote;
+  // string characters (includes start quote but not end quote)
+  uint64_t _in_string;
+};
+
+// Scans blocks for string characters, storing the state necessary to do so
+class json_string_scanner {
+public:
+  simdjson_really_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
+  // Returns either UNCLOSED_STRING or SUCCESS
+  simdjson_really_inline error_code finish();
+
+private:
+  // Scans for escape characters
+  json_escape_scanner escape_scanner{};
+  // Whether the last iteration was still inside a string (all 1's = true, all 0's = false).
+  uint64_t prev_in_string = 0ULL;
+};
+
+//
+// Return a mask of all string characters plus end quotes.
+//
+// prev_escaped is overflow saying whether the next character is escaped.
+// prev_in_string is overflow saying whether we're still in a string.
+//
+// Backslash sequences outside of quotes will be detected in stage 2.
+//
+simdjson_really_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
+  const uint64_t backslash = in.eq('\\');
+  const uint64_t escaped = escape_scanner.next(backslash).escaped;
+  const uint64_t quote = in.eq('"') & ~escaped;
+
+  //
+  // prefix_xor flips on bits inside the string (and flips off the end quote).
+  //
+  // Then we xor with prev_in_string: if we were in a string already, its effect is flipped
+  // (characters inside strings are outside, and characters outside strings are inside).
+  //
+  const uint64_t in_string = prefix_xor(quote) ^ prev_in_string;
+
+  //
+  // Check if we're still in a string at the end of the box so the next block will know
+  //
+  prev_in_string = uint64_t(static_cast<int64_t>(in_string) >> 63);
+
+  // Use ^ to turn the beginning quote off, and the end quote on.
+
+  // We are returning a function-local object so either we get a move constructor
+  // or we get copy elision.
+  return json_string_block(escaped, quote, in_string);
+}
+
+simdjson_really_inline error_code json_string_scanner::finish() {
+  if (prev_in_string) {
+    return UNCLOSED_STRING;
+  }
+  return SUCCESS;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+/* end file generic/stage1/json_string_scanner.h for arm64 */
+/* including generic/stage1/utf8_lookup4_algorithm.h for arm64: #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* begin file generic/stage1/utf8_lookup4_algorithm.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace utf8_validation {
+
+using namespace simd;
+
+  simdjson_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+// Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+// Bit 1 = Too Long (ASCII followed by continuation)
+// Bit 2 = Overlong 3-byte
+// Bit 4 = Surrogate
+// Bit 5 = Overlong 2-byte
+// Bit 7 = Two Continuations
+    constexpr const uint8_t TOO_SHORT   = 1<<0; // 11______ 0_______
+                                                // 11______ 11______
+    constexpr const uint8_t TOO_LONG    = 1<<1; // 0_______ 10______
+    constexpr const uint8_t OVERLONG_3  = 1<<2; // 11100000 100_____
+    constexpr const uint8_t SURROGATE   = 1<<4; // 11101101 101_____
+    constexpr const uint8_t OVERLONG_2  = 1<<5; // 1100000_ 10______
+    constexpr const uint8_t TWO_CONTS   = 1<<7; // 10______ 10______
+    constexpr const uint8_t TOO_LARGE   = 1<<3; // 11110100 1001____
+                                                // 11110100 101_____
+                                                // 11110101 1001____
+                                                // 11110101 101_____
+                                                // 1111011_ 1001____
+                                                // 1111011_ 101_____
+                                                // 11111___ 1001____
+                                                // 11111___ 101_____
+    constexpr const uint8_t TOO_LARGE_1000 = 1<<6;
+                                                // 11110101 1000____
+                                                // 1111011_ 1000____
+                                                // 11111___ 1000____
+    constexpr const uint8_t OVERLONG_4  = 1<<6; // 11110000 1000____
+
+    const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      TOO_SHORT,
+      // 1110____ ________ <three byte lead in byte 1>
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4
+    );
+    constexpr const uint8_t CARRY = TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+    const simd8<uint8_t> byte_1_low = (prev1 & 0x0F).lookup_16<uint8_t>(
+      // ____0000 ________
+      CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+      // ____0001 ________
+      CARRY | OVERLONG_2,
+      // ____001_ ________
+      CARRY,
+      CARRY,
+
+      // ____0100 ________
+      CARRY | TOO_LARGE,
+      // ____0101 ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      // ____011_ ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+
+      // ____1___ ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      // ____1101 ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000
+    );
+    const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE  | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE  | TOO_LARGE,
+
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT
+    );
+    return (byte_1_high & byte_1_low & byte_2_high);
+  }
+  simdjson_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
+      const simd8<uint8_t> prev_input, const simd8<uint8_t> sc) {
+    simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+    simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+    simd8<uint8_t> must23 = must_be_2_3_continuation(prev2, prev3);
+    simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+    return must23_80 ^ sc;
+  }
+
+  //
+  // Return nonzero if there are incomplete multibyte characters at the end of the block:
+  // e.g. if there is a 4-byte character, but it's 3 bytes from the end.
+  //
+  simdjson_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+    // If the previous input's last 3 bytes match this, they're too short (they ended at EOF):
+    // ... 1111____ 111_____ 11______
+#if SIMDJSON_IMPLEMENTATION_ICELAKE || (SIMDJSON_IMPLEMENTATION_RVV_VLS && __riscv_v_fixed_vlen >= 512)
+    static const uint8_t max_array[64] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
+    };
+#else
+    static const uint8_t max_array[32] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
+    };
+#endif
+    const simd8<uint8_t> max_value(&max_array[sizeof(max_array)-sizeof(simd8<uint8_t>)]);
+    return input.gt_bits(max_value);
+  }
+
+  struct utf8_checker {
+    // If this is nonzero, there has been a UTF-8 error.
+    simd8<uint8_t> error;
+    // The last input we received
+    simd8<uint8_t> prev_input_block;
+    // Whether the last input we received was incomplete (used for ASCII fast path)
+    simd8<uint8_t> prev_incomplete;
+
+    //
+    // Check whether the current bytes are valid UTF-8.
+    //
+    simdjson_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
+      // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ lead bytes
+      // (2, 3, 4-byte leads become large positive numbers instead of small negative numbers)
+      simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+      simd8<uint8_t> sc = check_special_cases(input, prev1);
+      this->error |= check_multibyte_lengths(input, prev_input, sc);
+    }
+
+    // The only problem that can happen at EOF is that a multibyte character is too short
+    // or a byte value too large in the last bytes: check_special_cases only checks for bytes
+    // too large in the first of two bytes.
+    simdjson_inline void check_eof() {
+      // If the previous block had incomplete UTF-8 characters at the end, an ASCII block can't
+      // possibly finish them.
+      this->error |= this->prev_incomplete;
+    }
+
+    simdjson_inline void check_next_input(const simd8x64<uint8_t>& input) {
+      if(simdjson_likely(is_ascii(input))) {
+        this->error |= this->prev_incomplete;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio, it is not good enough.
+        static_assert((simd8x64<uint8_t>::NUM_CHUNKS == 1)
+                ||(simd8x64<uint8_t>::NUM_CHUNKS == 2)
+                || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+                "We support one, two or four chunks per 64-byte block.");
+        SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 1) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+          this->check_utf8_bytes(input.get<1>(), input.get<0>());
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+          this->check_utf8_bytes(input.get<1>(), input.get<0>());
+          this->check_utf8_bytes(input.get<2>(), input.get<1>());
+          this->check_utf8_bytes(input.get<3>(), input.get<2>());
+        }
+        this->prev_incomplete = is_incomplete(input.get<simd8x64<uint8_t>::NUM_CHUNKS-1>());
+        this->prev_input_block = input.get<simd8x64<uint8_t>::NUM_CHUNKS-1>();
+      }
+    }
+    // do not forget to call check_eof!
+    simdjson_warn_unused simdjson_inline error_code errors() {
+      return this->error.any_bits_set_anywhere() ? error_code::UTF8_ERROR : error_code::SUCCESS;
+    }
+
+  }; // struct utf8_checker
+} // namespace utf8_validation
+
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H
+/* end file generic/stage1/utf8_lookup4_algorithm.h for arm64 */
+/* including generic/stage1/json_scanner.h for arm64: #include <generic/stage1/json_scanner.h> */
+/* begin file generic/stage1/json_scanner.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/json_character_block.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_string_scanner.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace stage1 {
+
+/**
+ * A block of scanned json, with information on operators and scalars.
+ *
+ * We seek to identify pseudo-structural characters. Anything that is inside
+ * a string must be omitted (hence  & ~_string.string_tail()).
+ * Otherwise, pseudo-structural characters come in two forms.
+ * 1. We have the structural characters ([,],{,},:, comma). The
+ *    term 'structural character' is from the JSON RFC.
+ * 2. We have the 'scalar pseudo-structural characters'.
+ *    Scalars are quotes, and any character except structural characters and white space.
+ *
+ * To identify the scalar pseudo-structural characters, we must look at what comes
+ * before them: it must be a space, a quote or a structural characters.
+ * Starting with simdjson v0.3, we identify them by
+ * negation: we identify everything that is followed by a non-quote scalar,
+ * and we negate that. Whatever remains must be a 'scalar pseudo-structural character'.
+ */
+struct json_block {
+public:
+  // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
+  simdjson_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  _string(std::move(string)), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
+  simdjson_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  _string(string), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
+
+  /**
+   * The start of structurals.
+   * In simdjson prior to v0.3, these were called the pseudo-structural characters.
+   **/
+  simdjson_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
+  /** All JSON whitespace (i.e. not in a string) */
+  simdjson_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
+
+  // Helpers
+
+  /** Whether the given characters are inside a string (only works on non-quotes) */
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
+  /** Whether the given characters are outside a string (only works on non-quotes) */
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
+
+  // string and escape characters
+  json_string_block _string;
+  // whitespace, structural characters ('operators'), scalars
+  json_character_block _characters;
+  // whether the previous character was a scalar
+  uint64_t _follows_potential_nonquote_scalar;
+private:
+  // Potential structurals (i.e. disregarding strings)
+
+  /**
+   * structural elements ([,],{,},:, comma) plus scalar starts like 123, true and "abc".
+   * They may reside inside a string.
+   **/
+  simdjson_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
+  /**
+   * The start of non-operator runs, like 123, true and "abc".
+   * It main reside inside a string.
+   **/
+  simdjson_inline uint64_t potential_scalar_start() const noexcept {
+    // The term "scalar" refers to anything except structural characters and white space
+    // (so letters, numbers, quotes).
+    // Whenever it is preceded by something that is not a structural element ({,},[,],:, ") nor a white-space
+    // then we know that it is irrelevant structurally.
+    return _characters.scalar() & ~follows_potential_scalar();
+  }
+  /**
+   * Whether the given character is immediately after a non-operator like 123, true.
+   * The characters following a quote are not included.
+   */
+  simdjson_inline uint64_t follows_potential_scalar() const noexcept {
+    // _follows_potential_nonquote_scalar: is defined as marking any character that follows a character
+    // that is not a structural element ({,},[,],:, comma) nor a quote (") and that is not a
+    // white space.
+    // It is understood that within quoted region, anything at all could be marked (irrelevant).
+    return _follows_potential_nonquote_scalar;
+  }
+};
+
+/**
+ * Scans JSON for important bits: structural characters or 'operators', strings, and scalars.
+ *
+ * The scanner starts by calculating two distinct things:
+ * - string characters (taking \" into account)
+ * - structural characters or 'operators' ([]{},:, comma)
+ *   and scalars (runs of non-operators like 123, true and "abc")
+ *
+ * To minimize data dependency (a key component of the scanner's speed), it finds these in parallel:
+ * in particular, the operator/scalar bit will find plenty of things that are actually part of
+ * strings. When we're done, json_block will fuse the two together by masking out tokens that are
+ * part of a string.
+ */
+class json_scanner {
+public:
+  json_scanner() = default;
+  simdjson_inline json_block next(const simd::simd8x64<uint8_t>& in);
+  // Returns either UNCLOSED_STRING or SUCCESS
+  simdjson_warn_unused simdjson_inline error_code finish();
+
+private:
+  // Whether the last character of the previous iteration is part of a scalar token
+  // (anything except whitespace or a structural character/'operator').
+  uint64_t prev_scalar = 0ULL;
+  json_string_scanner string_scanner{};
+};
+
+
+//
+// Check if the current character immediately follows a matching character.
+//
+// For example, this checks for quotes with backslashes in front of them:
+//
+//     const uint64_t backslashed_quote = in.eq('"') & immediately_follows(in.eq('\'), prev_backslash);
+//
+simdjson_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
+  const uint64_t result = match << 1 | overflow;
+  overflow = match >> 63;
+  return result;
+}
+
+simdjson_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
+  json_string_block strings = string_scanner.next(in);
+  // identifies the white-space and the structural characters
+  json_character_block characters = json_character_block::classify(in);
+  // The term "scalar" refers to anything except structural characters and white space
+  // (so letters, numbers, quotes).
+  // We want follows_scalar to mark anything that follows a non-quote scalar (so letters and numbers).
+  //
+  // A terminal quote should either be followed by a structural character (comma, brace, bracket, colon)
+  // or nothing. However, we still want ' "a string"true ' to mark the 't' of 'true' as a potential
+  // pseudo-structural character just like we would if we had  ' "a string" true '; otherwise we
+  // may need to add an extra check when parsing strings.
+  //
+  // Performance: there are many ways to skin this cat.
+  const uint64_t nonquote_scalar = characters.scalar() & ~strings.quote();
+  uint64_t follows_nonquote_scalar = follows(nonquote_scalar, prev_scalar);
+  // We are returning a function-local object so either we get a move constructor
+  // or we get copy elision.
+  return json_block(
+    strings,// strings is a function-local object so either it moves or the copy is elided.
+    characters,
+    follows_nonquote_scalar
+  );
+}
+
+simdjson_warn_unused simdjson_inline error_code json_scanner::finish() {
+  return string_scanner.finish();
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H
+/* end file generic/stage1/json_scanner.h for arm64 */
+
+// All other declarations
+/* including generic/stage1/find_next_document_index.h for arm64: #include <generic/stage1/find_next_document_index.h> */
+/* begin file generic/stage1/find_next_document_index.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace stage1 {
+
+/**
+  * This algorithm is used to quickly identify the last structural position that
+  * makes up a complete document.
+  *
+  * It does this by going backwards and finding the last *document boundary* (a
+  * place where one value follows another without a comma between them). If the
+  * last document (the characters after the boundary) has an equal number of
+  * start and end brackets, it is considered complete.
+  *
+  * Simply put, we iterate over the structural characters, starting from
+  * the end. We consider that we found the end of a JSON document when the
+  * first element of the pair is NOT one of these characters: '{' '[' ':' ','
+  * and when the second element is NOT one of these characters: '}' ']' ':' ','.
+  *
+  * This simple comparison works most of the time, but it does not cover cases
+  * where the batch's structural indexes contain a perfect amount of documents.
+  * In such a case, we do not have access to the structural index which follows
+  * the last document, therefore, we do not have access to the second element in
+  * the pair, and that means we cannot identify the last document. To fix this
+  * issue, we keep a count of the open and closed curly/square braces we found
+  * while searching for the pair. When we find a pair AND the count of open and
+  * closed curly/square braces is the same, we know that we just passed a
+  * complete document, therefore the last json buffer location is the end of the
+  * batch.
+  */
+simdjson_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
+  // Variant: do not count separately, just figure out depth
+  if(parser.n_structural_indexes == 0) { return 0; }
+  auto arr_cnt = 0;
+  auto obj_cnt = 0;
+  for (auto i = parser.n_structural_indexes - 1; i > 0; i--) {
+    auto idxb = parser.structural_indexes[i];
+    switch (parser.buf[idxb]) {
+    case ':':
+    case ',':
+      continue;
+    case '}':
+      obj_cnt--;
+      continue;
+    case ']':
+      arr_cnt--;
+      continue;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+    }
+    auto idxa = parser.structural_indexes[i - 1];
+    switch (parser.buf[idxa]) {
+    case '{':
+    case '[':
+    case ':':
+    case ',':
+      continue;
+    }
+    // Last document is complete, so the next document will appear after!
+    if (!arr_cnt && !obj_cnt) {
+      return parser.n_structural_indexes;
+    }
+    // Last document is incomplete; mark the document at i + 1 as the next one
+    return i;
+  }
+  // If we made it to the end, we want to finish counting to see if we have a full document.
+  switch (parser.buf[parser.structural_indexes[0]]) {
+    case '}':
+      obj_cnt--;
+      break;
+    case ']':
+      arr_cnt--;
+      break;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+  }
+  if (!arr_cnt && !obj_cnt) {
+    // We have a complete document.
+    return parser.n_structural_indexes;
+  }
+  return 0;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+/* end file generic/stage1/find_next_document_index.h for arm64 */
+/* including generic/stage1/json_minifier.h for arm64: #include <generic/stage1/json_minifier.h> */
+/* begin file generic/stage1/json_minifier.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses in stage1
+// It is intended to be included multiple times and compiled multiple times
+// We assume the file in which it is included already includes
+// "simdjson/stage1.h" (this simplifies amalgation)
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace stage1 {
+
+class json_minifier {
+public:
+  template<size_t STEP_SIZE>
+  static error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept;
+
+private:
+  simdjson_inline json_minifier(uint8_t *_dst)
+  : dst{_dst}
+  {}
+  template<size_t STEP_SIZE>
+  simdjson_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
+  simdjson_warn_unused simdjson_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
+  json_scanner scanner{};
+  uint8_t *dst;
+};
+
+simdjson_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
+  uint64_t mask = block.whitespace();
+  dst += in.compress(mask, dst);
+}
+
+simdjson_warn_unused simdjson_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
+  error_code error = scanner.finish();
+  if (error) { dst_len = 0; return error; }
+  dst_len = dst - dst_start;
+  return SUCCESS;
+}
+
+template<>
+simdjson_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block_buf);
+  simd::simd8x64<uint8_t> in_2(block_buf+64);
+  json_block block_1 = scanner.next(in_1);
+  json_block block_2 = scanner.next(in_2);
+  this->next(in_1, block_1);
+  this->next(in_2, block_2);
+  reader.advance();
+}
+
+template<>
+simdjson_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block_buf);
+  json_block block_1 = scanner.next(in_1);
+  this->next(block_buf, block_1);
+  reader.advance();
+}
+
+template<size_t STEP_SIZE>
+error_code json_minifier::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept {
+  buf_block_reader<STEP_SIZE> reader(buf, len);
+  json_minifier minifier(dst);
+
+  // Index the first n-1 blocks
+  while (reader.has_full_block()) {
+    minifier.step<STEP_SIZE>(reader.full_block(), reader);
+  }
+
+  // Index the last (remainder) block, padded with spaces
+  uint8_t block[STEP_SIZE];
+  size_t remaining_bytes = reader.get_remainder(block);
+  if (remaining_bytes > 0) {
+    // We do not want to write directly to the output stream. Rather, we write
+    // to a local buffer (for safety).
+    uint8_t out_block[STEP_SIZE];
+    uint8_t * const guarded_dst{minifier.dst};
+    minifier.dst = out_block;
+    minifier.step<STEP_SIZE>(block, reader);
+    size_t to_write = minifier.dst - out_block;
+    // In some cases, we could be enticed to consider the padded spaces
+    // as part of the string. This is fine as long as we do not write more
+    // than we consumed.
+    if(to_write > remaining_bytes) { to_write = remaining_bytes; }
+    memcpy(guarded_dst, out_block, to_write);
+    minifier.dst = guarded_dst + to_write;
+  }
+  return minifier.finish(dst, dst_len);
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H
+/* end file generic/stage1/json_minifier.h for arm64 */
+/* including generic/stage1/json_structural_indexer.h for arm64: #include <generic/stage1/json_structural_indexer.h> */
+/* begin file generic/stage1/json_structural_indexer.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_string_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_minifier.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/find_next_document_index.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses in stage1
+// It is intended to be included multiple times and compiled multiple times
+// We assume the file in which it is included already includes
+// "simdjson/stage1.h" (this simplifies amalgation)
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace stage1 {
+
+class bit_indexer {
+public:
+  uint32_t *tail;
+
+  simdjson_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
+
+#if SIMDJSON_PREFER_REVERSE_BITS
+  /**
+    * ARM lacks a fast trailing zero instruction, but it has a fast
+    * bit reversal instruction and a fast leading zero instruction.
+    * Thus it may be profitable to reverse the bits (once) and then
+    * to rely on a sequence of instructions that call the leading
+    * zero instruction.
+    *
+    * Performance notes:
+    * The chosen routine is not optimal in terms of data dependency
+    * since zero_leading_bit might require two instructions. However,
+    * it tends to minimize the total number of instructions which is
+    * beneficial.
+    */
+  simdjson_inline void write_index(uint32_t idx, uint64_t& rev_bits, int i) {
+    int lz = leading_zeroes(rev_bits);
+    this->tail[i] = static_cast<uint32_t>(idx) + lz;
+    rev_bits = zero_leading_bit(rev_bits, lz);
+  }
+#else
+  /**
+    * Under recent x64 systems, we often have both a fast trailing zero
+    * instruction and a fast 'clear-lower-bit' instruction so the following
+    * algorithm can be competitive.
+    */
+
+  simdjson_inline void write_index(uint32_t idx, uint64_t& bits, int i) {
+    this->tail[i] = idx + trailing_zeroes(bits);
+    bits = clear_lowest_bit(bits);
+  }
+#endif // SIMDJSON_PREFER_REVERSE_BITS
+
+  template <int START, int N>
+  simdjson_inline int write_indexes(uint32_t idx, uint64_t& bits) {
+    write_index(idx, bits, START);
+    SIMDJSON_IF_CONSTEXPR (N > 1) {
+      write_indexes<(N-1>0?START+1:START), (N-1>=0?N-1:1)>(idx, bits);
+    }
+    return START+N;
+  }
+
+  template <int START, int END, int STEP>
+  simdjson_inline int write_indexes_stepped(uint32_t idx, uint64_t& bits, int cnt) {
+    write_indexes<START, STEP>(idx, bits);
+    SIMDJSON_IF_CONSTEXPR ((START+STEP)  < END) {
+      if (simdjson_unlikely((START+STEP) < cnt)) {
+        write_indexes_stepped<(START+STEP<END?START+STEP:END), END, STEP>(idx, bits, cnt);
+      }
+    }
+    return ((END-START) % STEP) == 0 ? END : (END-START) - ((END-START) % STEP) + STEP;
+  }
+
+  // flatten out values in 'bits' assuming that they are are to have values of idx
+  // plus their position in the bitvector, and store these indexes at
+  // base_ptr[base] incrementing base as we go
+  // will potentially store extra values beyond end of valid bits, so base_ptr
+  // needs to be large enough to handle this
+  //
+  // If the kernel sets SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER, then it
+  // will provide its own version of the code.
+#ifdef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+  simdjson_inline void write(uint32_t idx, uint64_t bits);
+#else
+  simdjson_inline void write(uint32_t idx, uint64_t bits) {
+    // In some instances, the next branch is expensive because it is mispredicted.
+    // Unfortunately, in other cases,
+    // it helps tremendously.
+    if (bits == 0)
+        return;
+
+    int cnt = static_cast<int>(count_ones(bits));
+
+#if SIMDJSON_PREFER_REVERSE_BITS
+    bits = reverse_bits(bits);
+#endif
+#ifdef SIMDJSON_STRUCTURAL_INDEXER_STEP
+    static constexpr const int STEP = SIMDJSON_STRUCTURAL_INDEXER_STEP;
+#else
+    static constexpr const int STEP = 4;
+#endif
+    static constexpr const int STEP_UNTIL = 24;
+
+    write_indexes_stepped<0, STEP_UNTIL, STEP>(idx, bits, cnt);
+    SIMDJSON_IF_CONSTEXPR (STEP_UNTIL < 64) {
+      if (simdjson_unlikely(STEP_UNTIL < cnt)) {
+        for (int i=STEP_UNTIL; i<cnt; i++) {
+          write_index(idx, bits, i);
+        }
+      }
+    }
+
+    this->tail += cnt;
+  }
+#endif // SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+};
+
+class json_structural_indexer {
+public:
+  /**
+   * Find the important bits of JSON in a 128-byte chunk, and add them to structural_indexes.
+   *
+   * @param partial Setting the partial parameter to true allows the find_structural_bits to
+   *   tolerate unclosed strings. The caller should still ensure that the input is valid UTF-8. If
+   *   you are processing substrings, you may want to call on a function like trimmed_length_safe_utf8.
+   */
+  template<size_t STEP_SIZE>
+  static error_code index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept;
+
+private:
+  simdjson_inline json_structural_indexer(uint32_t *structural_indexes);
+  template<size_t STEP_SIZE>
+  simdjson_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
+  simdjson_warn_unused simdjson_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
+
+  json_scanner scanner{};
+  utf8_checker checker{};
+  bit_indexer indexer;
+  uint64_t prev_structurals = 0;
+  uint64_t unescaped_chars_error = 0;
+};
+
+simdjson_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
+
+// Skip the last character if it is partial
+simdjson_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
+  if (simdjson_unlikely(len < 3)) {
+    switch (len) {
+      case 2:
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
+        return len;
+      case 1:
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        return len;
+      case 0:
+        return len;
+    }
+  }
+  if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+  if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 1 byte left
+  if (buf[len-3] >= 0xf0) { return len-3; } // 4-byte characters with only 3 bytes left
+  return len;
+}
+
+//
+// PERF NOTES:
+// We pipe 2 inputs through these stages:
+// 1. Load JSON into registers. This takes a long time and is highly parallelizable, so we load
+//    2 inputs' worth at once so that by the time step 2 is looking for them input, it's available.
+// 2. Scan the JSON for critical data: strings, scalars and operators. This is the critical path.
+//    The output of step 1 depends entirely on this information. These functions don't quite use
+//    up enough CPU: the second half of the functions is highly serial, only using 1 execution core
+//    at a time. The second input's scans has some dependency on the first ones finishing it, but
+//    they can make a lot of progress before they need that information.
+// 3. Step 1 does not use enough capacity, so we run some extra stuff while we're waiting for that
+//    to finish: utf-8 checks and generating the output from the last iteration.
+//
+// The reason we run 2 inputs at a time, is steps 2 and 3 are *still* not enough to soak up all
+// available capacity with just one input. Running 2 at a time seems to give the CPU a good enough
+// workout.
+//
+template<size_t STEP_SIZE>
+error_code json_structural_indexer::index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept {
+  if (simdjson_unlikely(len > parser.capacity())) { return CAPACITY; }
+  // We guard the rest of the code so that we can assume that len > 0 throughout.
+  if (len == 0) { return EMPTY; }
+  if (is_streaming(partial)) {
+    len = trim_partial_utf8(buf, len);
+    // If you end up with an empty window after trimming
+    // the partial UTF-8 bytes, then chances are good that you
+    // have an UTF-8 formatting error.
+    if(len == 0) { return UTF8_ERROR; }
+  }
+  buf_block_reader<STEP_SIZE> reader(buf, len);
+  json_structural_indexer indexer(parser.structural_indexes.get());
+
+  // Read all but the last block
+  while (reader.has_full_block()) {
+    indexer.step<STEP_SIZE>(reader.full_block(), reader);
+  }
+  // Take care of the last block (will always be there unless file is empty which is
+  // not supposed to happen.)
+  uint8_t block[STEP_SIZE];
+  if (simdjson_unlikely(reader.get_remainder(block) == 0)) { return UNEXPECTED_ERROR; }
+  indexer.step<STEP_SIZE>(block, reader);
+  return indexer.finish(parser, reader.block_index(), len, partial);
+}
+
+template<>
+simdjson_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block);
+  simd::simd8x64<uint8_t> in_2(block+64);
+  json_block block_1 = scanner.next(in_1);
+  json_block block_2 = scanner.next(in_2);
+  this->next(in_1, block_1, reader.block_index());
+  this->next(in_2, block_2, reader.block_index()+64);
+  reader.advance();
+}
+
+template<>
+simdjson_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block);
+  json_block block_1 = scanner.next(in_1);
+  this->next(in_1, block_1, reader.block_index());
+  reader.advance();
+}
+
+simdjson_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
+  uint64_t unescaped = in.lteq(0x1F);
+#if SIMDJSON_UTF8VALIDATION
+  checker.check_next_input(in);
+#endif
+  indexer.write(uint32_t(idx-64), prev_structurals); // Output *last* iteration's structurals to the parser
+  prev_structurals = block.structural_start();
+  unescaped_chars_error |= block.non_quote_inside_string(unescaped);
+}
+
+simdjson_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
+  // Write out the final iteration's structurals
+  indexer.write(uint32_t(idx-64), prev_structurals);
+  error_code error = scanner.finish();
+  // We deliberately break down the next expression so that it is
+  // human readable.
+  const bool should_we_exit = is_streaming(partial) ?
+    ((error != SUCCESS) && (error != UNCLOSED_STRING)) // when partial we tolerate UNCLOSED_STRING
+    : (error != SUCCESS); // if partial is false, we must have SUCCESS
+  const bool have_unclosed_string = (error == UNCLOSED_STRING);
+  if (simdjson_unlikely(should_we_exit)) { return error; }
+
+  if (unescaped_chars_error) {
+    return UNESCAPED_CHARS;
+  }
+  parser.n_structural_indexes = uint32_t(indexer.tail - parser.structural_indexes.get());
+  /***
+   * The On-Demand API requires special padding.
+   *
+   * This is related to https://github.com/simdjson/simdjson/issues/906
+   * Basically, we want to make sure that if the parsing continues beyond the last (valid)
+   * structural character, it quickly stops.
+   * Only three structural characters can be repeated without triggering an error in JSON:  [,] and }.
+   * We repeat the padding character (at 'len'). We don't know what it is, but if the parsing
+   * continues, then it must be [,] or }.
+   * Suppose it is ] or }. We backtrack to the first character, what could it be that would
+   * not trigger an error? It could be ] or } but no, because you can't start a document that way.
+   * It can't be a comma, a colon or any simple value. So the only way we could continue is
+   * if the repeated character is [. But if so, the document must start with [. But if the document
+   * starts with [, it should end with ]. If we enforce that rule, then we would get
+   * ][[ which is invalid.
+   *
+   * This is illustrated with the test array_iterate_unclosed_error() on the following input:
+   * R"({ "a": [,,)"
+   **/
+  parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len); // used later in partial == stage1_mode::streaming_final
+  parser.structural_indexes[parser.n_structural_indexes + 1] = uint32_t(len);
+  parser.structural_indexes[parser.n_structural_indexes + 2] = 0;
+  parser.next_structural_index = 0;
+  // a valid JSON file cannot have zero structural indexes - we should have found something
+  if (simdjson_unlikely(parser.n_structural_indexes == 0u)) {
+    return EMPTY;
+  }
+  if (simdjson_unlikely(parser.structural_indexes[parser.n_structural_indexes - 1] > len)) {
+    return UNEXPECTED_ERROR;
+  }
+  if (partial == stage1_mode::streaming_partial) {
+    // If we have an unclosed string, then the last structural
+    // will be the quote and we want to make sure to omit it.
+    if(have_unclosed_string) {
+      parser.n_structural_indexes--;
+      // a valid JSON file cannot have zero structural indexes - we should have found something
+      if (simdjson_unlikely(parser.n_structural_indexes == 0u)) { return CAPACITY; }
+    }
+    // We truncate the input to the end of the last complete document (or zero).
+    auto new_structural_indexes = find_next_document_index(parser);
+    if (new_structural_indexes == 0 && parser.n_structural_indexes > 0) {
+      if(parser.structural_indexes[0] == 0) {
+        // If the buffer is partial and we started at index 0 but the document is
+        // incomplete, it's too big to parse.
+        return CAPACITY;
+      } else {
+        // It is possible that the document could be parsed, we just had a lot
+        // of white space.
+        parser.n_structural_indexes = 0;
+        return EMPTY;
+      }
+    }
+
+    parser.n_structural_indexes = new_structural_indexes;
+  } else if (partial == stage1_mode::streaming_final) {
+    if(have_unclosed_string) { parser.n_structural_indexes--; }
+    // We truncate the input to the end of the last complete document (or zero).
+    // Because partial == stage1_mode::streaming_final, it means that we may
+    // silently ignore trailing garbage. Though it sounds bad, we do it
+    // deliberately because many people who have streams of JSON documents
+    // will truncate them for processing. E.g., imagine that you are uncompressing
+    // the data from a size file or receiving it in chunks from the network. You
+    // may not know where exactly the last document will be. Meanwhile the
+    // document_stream instances allow people to know the JSON documents they are
+    // parsing (see the iterator.source() method).
+    parser.n_structural_indexes = find_next_document_index(parser);
+    // We store the initial n_structural_indexes so that the client can see
+    // whether we used truncation. If initial_n_structural_indexes == parser.n_structural_indexes,
+    // then this will query parser.structural_indexes[parser.n_structural_indexes] which is len,
+    // otherwise, it will copy some prior index.
+    parser.structural_indexes[parser.n_structural_indexes + 1] = parser.structural_indexes[parser.n_structural_indexes];
+    // This next line is critical, do not change it unless you understand what you are
+    // doing.
+    parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len);
+    if (simdjson_unlikely(parser.n_structural_indexes == 0u)) {
+        // We tolerate an unclosed string at the very end of the stream. Indeed, users
+        // often load their data in bulk without being careful and they want us to ignore
+        // the trailing garbage.
+        return EMPTY;
+    }
+  }
+  checker.check_eof();
+  return checker.errors();
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+// Clear CUSTOM_BIT_INDEXER so other implementations can set it if they need to.
+#undef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H
+/* end file generic/stage1/json_structural_indexer.h for arm64 */
+/* including generic/stage1/utf8_validator.h for arm64: #include <generic/stage1/utf8_validator.h> */
+/* begin file generic/stage1/utf8_validator.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace stage1 {
+
+/**
+ * Validates that the string is actual UTF-8.
+ */
+template<class checker>
+bool generic_validate_utf8(const uint8_t * input, size_t length) {
+    checker c{};
+    buf_block_reader<64> reader(input, length);
+    while (reader.has_full_block()) {
+      simd::simd8x64<uint8_t> in(reader.full_block());
+      c.check_next_input(in);
+      reader.advance();
+    }
+    uint8_t block[64]{};
+    reader.get_remainder(block);
+    simd::simd8x64<uint8_t> in(block);
+    c.check_next_input(in);
+    reader.advance();
+    c.check_eof();
+    return c.errors() == error_code::SUCCESS;
+}
+
+bool generic_validate_utf8(const char * input, size_t length) {
+    return generic_validate_utf8<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H
+/* end file generic/stage1/utf8_validator.h for arm64 */
+/* end file generic/stage1/amalgamated.h for arm64 */
+/* including generic/stage2/amalgamated.h for arm64: #include <generic/stage2/amalgamated.h> */
+/* begin file generic/stage2/amalgamated.h for arm64 */
+// Stuff other things depend on
+/* including generic/stage2/base.h for arm64: #include <generic/stage2/base.h> */
+/* begin file generic/stage2/base.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_BASE_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace stage2 {
+
+class json_iterator;
+class structural_iterator;
+struct tape_builder;
+struct tape_writer;
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_BASE_H
+/* end file generic/stage2/base.h for arm64 */
+/* including generic/stage2/tape_writer.h for arm64: #include <generic/stage2/tape_writer.h> */
+/* begin file generic/stage2/tape_writer.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/internal/tape_type.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace stage2 {
+
+struct tape_writer {
+  /** The next place to write to tape */
+  uint64_t *next_tape_loc;
+
+  /** Write a signed 64-bit value to tape. */
+  simdjson_inline void append_s64(int64_t value) noexcept;
+
+  /** Write an unsigned 64-bit value to tape. */
+  simdjson_inline void append_u64(uint64_t value) noexcept;
+
+  /** Write a double value to tape. */
+  simdjson_inline void append_double(double value) noexcept;
+
+  /** Write a big integer (as string) to tape. src points to first digit, len is byte count. */
+  simdjson_inline void append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept;
+
+  /**
+   * Append a tape entry (an 8-bit type,and 56 bits worth of value).
+   */
+  simdjson_inline void append(uint64_t val, internal::tape_type t) noexcept;
+
+  /**
+   * Skip the current tape entry without writing.
+   *
+   * Used to skip the start of the container, since we'll come back later to fill it in when the
+   * container ends.
+   */
+  simdjson_inline void skip() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a large u64 or i64.
+   */
+  simdjson_inline void skip_large_integer() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a double.
+   */
+  simdjson_inline void skip_double() noexcept;
+
+  /**
+   * Write a value to a known location on tape.
+   *
+   * Used to go back and write out the start of a container after the container ends.
+   */
+  simdjson_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
+
+private:
+  /**
+   * Append both the tape entry, and a supplementary value following it. Used for types that need
+   * all 64 bits, such as double and uint64_t.
+   */
+  template<typename T>
+  simdjson_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
+}; // struct tape_writer
+
+simdjson_inline void tape_writer::append_s64(int64_t value) noexcept {
+  append2(0, value, internal::tape_type::INT64);
+}
+
+simdjson_inline void tape_writer::append_u64(uint64_t value) noexcept {
+  append(0, internal::tape_type::UINT64);
+  *next_tape_loc = value;
+  next_tape_loc++;
+}
+
+/** Write a double value to tape. */
+simdjson_inline void tape_writer::append_double(double value) noexcept {
+  append2(0, value, internal::tape_type::DOUBLE);
+}
+
+simdjson_inline void tape_writer::skip() noexcept {
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::skip_large_integer() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::skip_double() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
+  *next_tape_loc = val | ((uint64_t(char(t))) << 56);
+  next_tape_loc++;
+}
+
+template<typename T>
+simdjson_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
+  append(val, t);
+  static_assert(sizeof(val2) == sizeof(*next_tape_loc), "Type is not 64 bits!");
+  memcpy(next_tape_loc, &val2, sizeof(val2));
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
+  tape_loc = val | ((uint64_t(char(t))) << 56);
+}
+
+simdjson_inline void tape_writer::append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept {
+  // Write to string buffer: [4-byte LE length][digits][null]
+  uint32_t str_len = uint32_t(len);
+  memcpy(string_buf, &str_len, sizeof(uint32_t));
+  memcpy(string_buf + sizeof(uint32_t), src, len);
+  string_buf[sizeof(uint32_t) + len] = 0;
+  // Tape entry: offset into string buffer
+  // The caller must set the offset relative to doc.string_buf base
+  append(0, internal::tape_type::BIGINT); // placeholder offset, caller patches
+  string_buf += sizeof(uint32_t) + len + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+/* end file generic/stage2/tape_writer.h for arm64 */
+/* including generic/stage2/logger.h for arm64: #include <generic/stage2/logger.h> */
+/* begin file generic/stage2/logger.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+
+// This is for an internal-only stage 2 specific logger.
+// Set LOG_ENABLED = true to log what stage 2 is doing!
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace logger {
+
+  static constexpr const char * DASHES = "----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------";
+
+#if SIMDJSON_VERBOSE_LOGGING
+  static constexpr const bool LOG_ENABLED = true;
+#else
+  static constexpr const bool LOG_ENABLED = false;
+#endif
+  static constexpr const int LOG_EVENT_LEN = 20;
+  static constexpr const int LOG_BUFFER_LEN = 30;
+  static constexpr const int LOG_SMALL_BUFFER_LEN = 10;
+  static constexpr const int LOG_INDEX_LEN = 5;
+
+  static int log_depth; // Not threadsafe. Log only.
+
+  // Helper to turn unprintable or newline characters into spaces
+  static simdjson_inline char printable_char(char c) {
+    if (c >= 0x20) {
+      return c;
+    } else {
+      return ' ';
+    }
+  }
+
+  // Print the header and set up log_start
+  static simdjson_inline void log_start() {
+    if (LOG_ENABLED) {
+      log_depth = 0;
+      printf("\n");
+      printf("| %-*s | %-*s | %-*s | %-*s | Detail |\n", LOG_EVENT_LEN, "Event", LOG_BUFFER_LEN, "Buffer", LOG_SMALL_BUFFER_LEN, "Next", 5, "Next#");
+      printf("|%.*s|%.*s|%.*s|%.*s|--------|\n", LOG_EVENT_LEN+2, DASHES, LOG_BUFFER_LEN+2, DASHES, LOG_SMALL_BUFFER_LEN+2, DASHES, 5+2, DASHES);
+    }
+  }
+
+  simdjson_unused static simdjson_inline void log_string(const char *message) {
+    if (LOG_ENABLED) {
+      printf("%s\n", message);
+    }
+  }
+
+  // Logs a single line from the stage 2 DOM parser
+  template<typename S>
+  static simdjson_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
+    if (LOG_ENABLED) {
+      printf("| %*s%s%-*s ", log_depth*2, "", title_prefix, LOG_EVENT_LEN - log_depth*2 - int(strlen(title_prefix)), title);
+      auto current_index = structurals.at_beginning() ? nullptr : structurals.next_structural-1;
+      auto next_index = structurals.next_structural;
+      auto current = current_index ? &structurals.buf[*current_index] : reinterpret_cast<const uint8_t*>("                                                       ");
+      auto next = &structurals.buf[*next_index];
+      {
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_BUFFER_LEN;i++) {
+          printf("%c", printable_char(current[i]));
+        }
+        printf(" ");
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_SMALL_BUFFER_LEN;i++) {
+          printf("%c", printable_char(next[i]));
+        }
+        printf(" ");
+      }
+      if (current_index) {
+        printf("| %*u ", LOG_INDEX_LEN, *current_index);
+      } else {
+        printf("| %-*s ", LOG_INDEX_LEN, "");
+      }
+      // printf("| %*u ", LOG_INDEX_LEN, structurals.next_tape_index());
+      printf("| %-s ", detail);
+      printf("|\n");
+    }
+  }
+
+} // namespace logger
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+/* end file generic/stage2/logger.h for arm64 */
+
+// All other declarations
+/* including generic/stage2/json_iterator.h for arm64: #include <generic/stage2/json_iterator.h> */
+/* begin file generic/stage2/json_iterator.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/logger.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace stage2 {
+
+class json_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+  uint32_t depth{0};
+
+  /**
+   * Walk the JSON document.
+   *
+   * The visitor receives callbacks when values are encountered. All callbacks pass the iterator as
+   * the first parameter; some callbacks have other parameters as well:
+   *
+   * - visit_document_start() - at the beginning.
+   * - visit_document_end() - at the end (if things were successful).
+   *
+   * - visit_array_start() - at the start `[` of a non-empty array.
+   * - visit_array_end() - at the end `]` of a non-empty array.
+   * - visit_empty_array() - when an empty array is encountered.
+   *
+   * - visit_object_end() - at the start `]` of a non-empty object.
+   * - visit_object_start() - at the end `]` of a non-empty object.
+   * - visit_empty_object() - when an empty object is encountered.
+   * - visit_key(const uint8_t *key) - when a key in an object field is encountered. key is
+   *                                   guaranteed to point at the first quote of the string (`"key"`).
+   * - visit_primitive(const uint8_t *value) - when a value is a string, number, boolean or null.
+   * - visit_root_primitive(iter, uint8_t *value) - when the top-level value is a string, number, boolean or null.
+   *
+   * - increment_count(iter) - each time a value is found in an array or object.
+   */
+  template<bool STREAMING, typename V>
+  simdjson_warn_unused simdjson_inline error_code walk_document(V &visitor) noexcept;
+
+  /**
+   * Create an iterator capable of walking a JSON document.
+   *
+   * The document must have already passed through stage 1.
+   */
+  simdjson_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
+
+  /**
+   * Look at the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *peek() const noexcept;
+  /**
+   * Advance to the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *advance() noexcept;
+  /**
+   * Get the remaining length of the document, from the start of the current token.
+   */
+  simdjson_inline size_t remaining_len() const noexcept;
+  /**
+   * Check if we are at the end of the document.
+   *
+   * If this is true, there are no more tokens.
+   */
+  simdjson_inline bool at_eof() const noexcept;
+  /**
+   * Check if we are at the beginning of the document.
+   */
+  simdjson_inline bool at_beginning() const noexcept;
+  simdjson_inline uint8_t last_structural() const noexcept;
+
+  /**
+   * Log that a value has been found.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_value(const char *type) const noexcept;
+  /**
+   * Log the start of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_start_value(const char *type) const noexcept;
+  /**
+   * Log the end of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_end_value(const char *type) const noexcept;
+  /**
+   * Log an error.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_error(const char *error) const noexcept;
+
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
+};
+
+template<bool STREAMING, typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::walk_document(V &visitor) noexcept {
+  logger::log_start();
+
+  //
+  // Start the document
+  //
+  if (at_eof()) { return EMPTY; }
+  log_start_value("document");
+  SIMDJSON_TRY( visitor.visit_document_start(*this) );
+
+  //
+  // Read first value
+  //
+  {
+    auto value = advance();
+
+    // Make sure the outer object or array is closed before continuing; otherwise, there are ways we
+    // could get into memory corruption. See https://github.com/simdjson/simdjson/issues/906
+    if (!STREAMING) {
+      switch (*value) {
+        case '{': if (last_structural() != '}') { log_value("starting brace unmatched"); return TAPE_ERROR; }; break;
+        case '[': if (last_structural() != ']') { log_value("starting bracket unmatched"); return TAPE_ERROR; }; break;
+      }
+    }
+
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_root_primitive(*this, value) ); break;
+    }
+  }
+  goto document_end;
+
+//
+// Object parser states
+//
+object_begin:
+  log_start_value("object");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = false;
+  SIMDJSON_TRY( visitor.visit_object_start(*this) );
+
+  {
+    auto key = advance();
+    if (*key != '"') { log_error("Object does not start with a key"); return TAPE_ERROR; }
+    SIMDJSON_TRY( visitor.increment_count(*this) );
+    SIMDJSON_TRY( visitor.visit_key(*this, key) );
+  }
+
+object_field:
+  if (simdjson_unlikely( *advance() != ':' )) { log_error("Missing colon after key in object"); return TAPE_ERROR; }
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+object_continue:
+  switch (*advance()) {
+    case ',':
+      SIMDJSON_TRY( visitor.increment_count(*this) );
+      {
+        auto key = advance();
+        if (simdjson_unlikely( *key != '"' )) { log_error("Key string missing at beginning of field in object"); return TAPE_ERROR; }
+        SIMDJSON_TRY( visitor.visit_key(*this, key) );
+      }
+      goto object_field;
+    case '}': log_end_value("object"); SIMDJSON_TRY( visitor.visit_object_end(*this) ); goto scope_end;
+    default: log_error("No comma between object fields"); return TAPE_ERROR;
+  }
+
+scope_end:
+  depth--;
+  if (depth == 0) { goto document_end; }
+  if (dom_parser.is_array[depth]) { goto array_continue; }
+  goto object_continue;
+
+//
+// Array parser states
+//
+array_begin:
+  log_start_value("array");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = true;
+  SIMDJSON_TRY( visitor.visit_array_start(*this) );
+  SIMDJSON_TRY( visitor.increment_count(*this) );
+
+array_value:
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+array_continue:
+  switch (*advance()) {
+    case ',': SIMDJSON_TRY( visitor.increment_count(*this) ); goto array_value;
+    case ']': log_end_value("array"); SIMDJSON_TRY( visitor.visit_array_end(*this) ); goto scope_end;
+    default: log_error("Missing comma between array values"); return TAPE_ERROR;
+  }
+
+document_end:
+  log_end_value("document");
+  SIMDJSON_TRY( visitor.visit_document_end(*this) );
+
+  dom_parser.next_structural_index = uint32_t(next_structural - &dom_parser.structural_indexes[0]);
+
+  // If we didn't make it to the end, it's an error
+  if ( !STREAMING && dom_parser.next_structural_index != dom_parser.n_structural_indexes ) {
+    log_error("More than one JSON value at the root of the document, or extra characters at the end of the JSON!");
+    return TAPE_ERROR;
+  }
+
+  return SUCCESS;
+
+} // walk_document()
+
+simdjson_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+  : buf{_dom_parser.buf},
+    next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+    dom_parser{_dom_parser} {
+}
+
+simdjson_inline const uint8_t *json_iterator::peek() const noexcept {
+  return &buf[*(next_structural)];
+}
+simdjson_inline const uint8_t *json_iterator::advance() noexcept {
+  return &buf[*(next_structural++)];
+}
+simdjson_inline size_t json_iterator::remaining_len() const noexcept {
+  return dom_parser.len - *(next_structural-1);
+}
+
+simdjson_inline bool json_iterator::at_eof() const noexcept {
+  return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+}
+simdjson_inline bool json_iterator::at_beginning() const noexcept {
+  return next_structural == dom_parser.structural_indexes.get();
+}
+simdjson_inline uint8_t json_iterator::last_structural() const noexcept {
+  return buf[dom_parser.structural_indexes[dom_parser.n_structural_indexes - 1]];
+}
+
+simdjson_inline void json_iterator::log_value(const char *type) const noexcept {
+  logger::log_line(*this, "", type, "");
+}
+
+simdjson_inline void json_iterator::log_start_value(const char *type) const noexcept {
+  logger::log_line(*this, "+", type, "");
+  if (logger::LOG_ENABLED) { logger::log_depth++; }
+}
+
+simdjson_inline void json_iterator::log_end_value(const char *type) const noexcept {
+  if (logger::LOG_ENABLED) { logger::log_depth--; }
+  logger::log_line(*this, "-", type, "");
+}
+
+simdjson_inline void json_iterator::log_error(const char *error) const noexcept {
+  logger::log_line(*this, "", "ERROR", error);
+}
+
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
+  switch (*value) {
+    case '"': return visitor.visit_root_string(*this, value);
+    case 't': return visitor.visit_root_true_atom(*this, value);
+    case 'f': return visitor.visit_root_false_atom(*this, value);
+    case 'n': return visitor.visit_root_null_atom(*this, value);
+    case '-':
+    case '0': case '1': case '2': case '3': case '4':
+    case '5': case '6': case '7': case '8': case '9':
+      return visitor.visit_root_number(*this, value);
+    default:
+      log_error("Document starts with a non-value character");
+      return TAPE_ERROR;
+  }
+}
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
+  // Use the fact that most scalars are going to be either strings or numbers.
+  if(*value == '"') {
+    return visitor.visit_string(*this, value);
+  } else if (((*value - '0')  < 10) || (*value == '-')) {
+    return visitor.visit_number(*this, value);
+  }
+  // true, false, null are uncommon.
+  switch (*value) {
+    case 't': return visitor.visit_true_atom(*this, value);
+    case 'f': return visitor.visit_false_atom(*this, value);
+    case 'n': return visitor.visit_null_atom(*this, value);
+    default:
+      log_error("Non-value found when value was expected!");
+      return TAPE_ERROR;
+  }
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+/* end file generic/stage2/json_iterator.h for arm64 */
+/* including generic/stage2/stringparsing.h for arm64: #include <generic/stage2/stringparsing.h> */
+/* begin file generic/stage2/stringparsing.h for arm64 */
+#include <cstdint>
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/jsoncharutils.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses
+// It is intended to be included multiple times and compiled multiple times
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+/// @private
+namespace stringparsing {
+
+// begin copypasta
+// These chars yield themselves: " \ /
+// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
+// u not handled in this table as it's complex
+static const uint8_t escape_map[256] = {
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
+    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
+    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+};
+
+// handle a unicode codepoint
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr, bool allow_replacement) {
+  // Use the default Unicode Character 'REPLACEMENT CHARACTER' (U+FFFD)
+  constexpr uint32_t substitution_code_point = 0xfffd;
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) != ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+    } else {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+
+      // We have already checked that the high surrogate is valid and
+      // (code_point - 0xd800) < 1024.
+      //
+      // Check that code_point_2 is in the range 0xdc00..0xdfff
+      // and that code_point_2 was parsed from valid hex.
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if (low_bit >> 10) {
+        if(!allow_replacement) { return false; }
+        code_point = substitution_code_point;
+      } else {
+        code_point =  (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+
+    }
+  } else if (code_point >= 0xdc00 && code_point <= 0xdfff) {
+      // If we encounter a low surrogate (not preceded by a high surrogate)
+      // then we have an error.
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+  }
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+// handle a unicode codepoint using the wobbly convention
+// https://simonsapin.github.io/wtf-8/
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint_wobbly(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr) {
+  // It is not ideal that this function is nearly identical to handle_unicode_codepoint.
+  //
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) == ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if ((low_bit >> 10) ==  0) {
+        code_point =
+          (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+    }
+  }
+
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+/**
+ * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+ * must be an unescaped quote terminating the string. It returns the final output
+ * position as pointer. In case of error (e.g., the string has bad escaped codes),
+ * then null_ptr is returned. It is assumed that the output buffer is large
+ * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+ * SIMDJSON_PADDING bytes.
+ */
+simdjson_warn_unused simdjson_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) {
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint(&src, &dst, allow_replacement)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+simdjson_warn_unused simdjson_inline uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) {
+  // It is not ideal that this function is nearly identical to parse_string.
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint_wobbly(&src, &dst)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+} // namespace stringparsing
+
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+/* end file generic/stage2/stringparsing.h for arm64 */
+/* including generic/stage2/structural_iterator.h for arm64: #include <generic/stage2/structural_iterator.h> */
+/* begin file generic/stage2/structural_iterator.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace stage2 {
+
+class structural_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+
+  // Start a structural
+  simdjson_inline structural_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+    : buf{_dom_parser.buf},
+      next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+      dom_parser{_dom_parser} {
+  }
+  // Get the buffer position of the current structural character
+  simdjson_inline const uint8_t* current() {
+    return &buf[*(next_structural-1)];
+  }
+  // Get the current structural character
+  simdjson_inline char current_char() {
+    return buf[*(next_structural-1)];
+  }
+  // Get the next structural character without advancing
+  simdjson_inline char peek_next_char() {
+    return buf[*next_structural];
+  }
+  simdjson_inline const uint8_t* peek() {
+    return &buf[*next_structural];
+  }
+  simdjson_inline const uint8_t* advance() {
+    return &buf[*(next_structural++)];
+  }
+  simdjson_inline char advance_char() {
+    return buf[*(next_structural++)];
+  }
+  simdjson_inline size_t remaining_len() {
+    return dom_parser.len - *(next_structural-1);
+  }
+
+  simdjson_inline bool at_end() {
+    return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+  }
+  simdjson_inline bool at_beginning() {
+    return next_structural == dom_parser.structural_indexes.get();
+  }
+};
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H
+/* end file generic/stage2/structural_iterator.h for arm64 */
+/* including generic/stage2/tape_builder.h for arm64: #include <generic/stage2/tape_builder.h> */
+/* begin file generic/stage2/tape_builder.h for arm64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/json_iterator.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/stringparsing.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/tape_writer.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/dom/document.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/atomparsing.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/numberparsing.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+namespace stage2 {
+
+struct tape_builder {
+  template<bool STREAMING>
+  simdjson_warn_unused static simdjson_inline error_code parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept;
+
+  /** Called when a non-empty document starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty document ends without error. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_end(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty array starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty array ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_end(json_iterator &iter) noexcept;
+  /** Called when an empty array is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_array(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty object starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_start(json_iterator &iter) noexcept;
+  /**
+   * Called when a key in a field is encountered.
+   *
+   * primitive, visit_object_start, visit_empty_object, visit_array_start, or visit_empty_array
+   * will be called after this with the field value.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
+  /** Called when a non-empty object ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_end(json_iterator &iter) noexcept;
+  /** Called when an empty object is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_object(json_iterator &iter) noexcept;
+
+  /**
+   * Called when a string, number, boolean or null is found.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  /**
+   * Called when a string, number, boolean or null is found at the top level of a document (i.e.
+   * when there is no array or object and the entire document is a single string, number, boolean or
+   * null.
+   *
+   * This is separate from primitive() because simdjson's normal primitive parsing routines assume
+   * there is at least one more token after the value, which is only true in an array or object.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  /** Called each time a new field or element in an array or object is found. */
+  simdjson_warn_unused simdjson_inline error_code increment_count(json_iterator &iter) noexcept;
+
+  /** Next location to write to tape */
+  tape_writer tape;
+private:
+  /** Next write location in the string buf for stage 2 parsing */
+  uint8_t *current_string_buf_loc;
+
+  simdjson_inline tape_builder(dom::document &doc) noexcept;
+
+  simdjson_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
+  simdjson_inline void start_container(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_warn_unused simdjson_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
+  simdjson_inline void on_end_string(uint8_t *dst) noexcept;
+}; // struct tape_builder
+
+template<bool STREAMING>
+simdjson_warn_unused simdjson_inline error_code tape_builder::parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept {
+  dom_parser.doc = &doc;
+  json_iterator iter(dom_parser, STREAMING ? dom_parser.next_structural_index : 0);
+  tape_builder builder(doc);
+  return iter.walk_document<STREAMING>(builder);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_root_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
+  constexpr uint32_t start_tape_index = 0;
+  tape.append(start_tape_index, internal::tape_type::ROOT);
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter), internal::tape_type::ROOT);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
+  return visit_string(iter, key, true);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].count++; // we have a key value pair in the object at parser.dom_parser.depth - 1
+  return SUCCESS;
+}
+
+simdjson_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
+  iter.log_value(key ? "key" : "string");
+  uint8_t *dst = on_start_string(iter);
+  dst = stringparsing::parse_string(value+1, dst, false); // We do not allow replacement when the escape characters are invalid.
+  if (dst == nullptr) {
+    iter.log_error("Invalid escape in string");
+    return STRING_ERROR;
+  }
+  on_end_string(dst);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
+  return visit_string(iter, value);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("number");
+  error_code err = numberparsing::parse_number(value, tape);
+  if (simdjson_unlikely(err == BIGINT_ERROR &&
+      iter.dom_parser._number_as_string)) {
+    // Write big integer to string buffer using the same format as strings.
+    // Scan digits the same way parse_number does (skip optional '-', then digits).
+    const uint8_t *p = value;
+    if (*p == '-') p++;
+    while (numberparsing::is_digit(*p)) p++;
+    size_t len = size_t(p - value);
+    tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::BIGINT);
+    uint8_t *dst = current_string_buf_loc + sizeof(uint32_t);
+    memcpy(dst, value, len);
+    dst += len;
+    on_end_string(dst);
+    return SUCCESS;
+  }
+  return err;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
+  //
+  // We need to make a copy to make sure that the string is space terminated.
+  // This is not about padding the input, which should already padded up
+  // to len + SIMDJSON_PADDING. However, we have no control at this stage
+  // on how the padding was done. What if the input string was padded with nulls?
+  // It is quite common for an input string to have an extra null character (C string).
+  // We do not want to allow 9\0 (where \0 is the null character) inside a JSON
+  // document, but the string "9\0" by itself is fine. So we make a copy and
+  // pad the input with spaces when we know that there is just one input element.
+  // This copy is relatively expensive, but it will almost never be called in
+  // practice unless you are in the strange scenario where you have many JSON
+  // documents made of single atoms.
+  //
+  std::unique_ptr<uint8_t[]>copy(new (std::nothrow) uint8_t[iter.remaining_len() + SIMDJSON_PADDING]);
+  if (copy.get() == nullptr) { return MEMALLOC; }
+  std::memcpy(copy.get(), value, iter.remaining_len());
+  std::memset(copy.get() + iter.remaining_len(), ' ', SIMDJSON_PADDING);
+  error_code error = visit_number(iter, copy.get());
+  return error;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value)) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value, iter.remaining_len())) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value)) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value, iter.remaining_len())) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value)) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value, iter.remaining_len())) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+// private:
+
+simdjson_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
+  return uint32_t(tape.next_tape_loc - iter.dom_parser.doc->tape.get());
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  auto start_index = next_tape_index(iter);
+  tape.append(start_index+2, start);
+  tape.append(start_index, end);
+  return SUCCESS;
+}
+
+simdjson_inline void tape_builder::start_container(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].tape_index = next_tape_index(iter);
+  iter.dom_parser.open_containers[iter.depth].count = 0;
+  tape.skip(); // We don't actually *write* the start element until the end.
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  // Write the ending tape element, pointing at the start location
+  const uint32_t start_tape_index = iter.dom_parser.open_containers[iter.depth].tape_index;
+  tape.append(start_tape_index, end);
+  // Write the start tape element, pointing at the end location (and including count)
+  // count can overflow if it exceeds 24 bits... so we saturate
+  // the convention being that a cnt of 0xffffff or more is undetermined in value (>=  0xffffff).
+  const uint32_t count = iter.dom_parser.open_containers[iter.depth].count;
+  const uint32_t cntsat = count > 0xFFFFFF ? 0xFFFFFF : count;
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter) | (uint64_t(cntsat) << 32), start);
+  return SUCCESS;
+}
+
+simdjson_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
+  // we advance the point, accounting for the fact that we have a NULL termination
+  tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::STRING);
+  return current_string_buf_loc + sizeof(uint32_t);
+}
+
+simdjson_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
+  uint32_t str_length = uint32_t(dst - (current_string_buf_loc + sizeof(uint32_t)));
+  // TODO check for overflow in case someone has a crazy string (>=4GB?)
+  // But only add the overflow check when the document itself exceeds 4GB
+  // Currently unneeded because we refuse to parse docs larger or equal to 4GB.
+  memcpy(current_string_buf_loc, &str_length, sizeof(uint32_t));
+  // NULL termination is still handy if you expect all your strings to
+  // be NULL terminated? It comes at a small cost
+  *dst = 0;
+  current_string_buf_loc = dst + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+/* end file generic/stage2/tape_builder.h for arm64 */
+/* end file generic/stage2/amalgamated.h for arm64 */
+
+//
+// Stage 1
+//
+namespace simdjson {
+namespace arm64 {
+
+simdjson_warn_unused error_code implementation::create_dom_parser_implementation(
+  size_t capacity,
+  size_t max_depth,
+  std::unique_ptr<internal::dom_parser_implementation>& dst
+) const noexcept {
+  dst.reset( new (std::nothrow) dom_parser_implementation() );
+  if (!dst) { return MEMALLOC; }
+  if (auto err = dst->set_capacity(capacity))
+    return err;
+  if (auto err = dst->set_max_depth(max_depth))
+    return err;
+  return SUCCESS;
+}
+
+namespace {
+
+using namespace simd;
+
+simdjson_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
+  // Functional programming causes trouble with Visual Studio.
+  // Keeping this version in comments since it is much nicer:
+  // auto v = in.map<uint8_t>([&](simd8<uint8_t> chunk) {
+  //  auto nib_lo = chunk & 0xf;
+  //  auto nib_hi = chunk.shr<4>();
+  //  auto shuf_lo = nib_lo.lookup_16<uint8_t>(16, 0, 0, 0, 0, 0, 0, 0, 0, 8, 12, 1, 2, 9, 0, 0);
+  //  auto shuf_hi = nib_hi.lookup_16<uint8_t>(8, 0, 18, 4, 0, 1, 0, 1, 0, 0, 0, 3, 2, 1, 0, 0);
+  //  return shuf_lo & shuf_hi;
+  // });
+  const simd8<uint8_t> table1(16, 0, 0, 0, 0, 0, 0, 0, 0, 8, 12, 1, 2, 9, 0, 0);
+  const simd8<uint8_t> table2(8, 0, 18, 4, 0, 1, 0, 1, 0, 0, 0, 3, 2, 1, 0, 0);
+
+  simd8x64<uint8_t> v(
+     (in.chunks[0] & 0xf).lookup_16(table1) & (in.chunks[0].shr<4>()).lookup_16(table2),
+     (in.chunks[1] & 0xf).lookup_16(table1) & (in.chunks[1].shr<4>()).lookup_16(table2),
+     (in.chunks[2] & 0xf).lookup_16(table1) & (in.chunks[2].shr<4>()).lookup_16(table2),
+     (in.chunks[3] & 0xf).lookup_16(table1) & (in.chunks[3].shr<4>()).lookup_16(table2)
+  );
+
+
+  // We compute whitespace and op separately. If the code later only use one or the
+  // other, given the fact that all functions are aggressively inlined, we can
+  // hope that useless computations will be omitted. This is namely case when
+  // minifying (we only need whitespace). *However* if we only need spaces,
+  // it is likely that we will still compute 'v' above with two lookup_16: one
+  // could do it a bit cheaper. This is in contrast with the x64 implementations
+  // where we can, efficiently, do the white space and structural matching
+  // separately. One reason for this difference is that on ARM NEON, the table
+  // lookups either zero or leave unchanged the characters exceeding 0xF whereas
+  // on x64, the equivalent instruction (pshufb) automatically applies a mask,
+  // ignoring the 4 most significant bits. Thus the x64 implementation is
+  // optimized differently. This being said, if you use this code strictly
+  // just for minification (or just to identify the structural characters),
+  // there is a small untaken optimization opportunity here. We deliberately
+  // do not pick it up.
+
+  uint64_t op = simd8x64<bool>(
+        v.chunks[0].any_bits_set(0x7),
+        v.chunks[1].any_bits_set(0x7),
+        v.chunks[2].any_bits_set(0x7),
+        v.chunks[3].any_bits_set(0x7)
+  ).to_bitmask();
+
+  uint64_t whitespace = simd8x64<bool>(
+        v.chunks[0].any_bits_set(0x18),
+        v.chunks[1].any_bits_set(0x18),
+        v.chunks[2].any_bits_set(0x18),
+        v.chunks[3].any_bits_set(0x18)
+  ).to_bitmask();
+
+  return { whitespace, op };
+}
+
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input) {
+    simd8<uint8_t> bits = input.reduce_or();
+    return bits.max_val() < 0x80u;
+}
+
+simdjson_unused simdjson_inline simd8<bool> must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+    simd8<bool> is_second_byte = prev1 >= uint8_t(0xc0u);
+    simd8<bool> is_third_byte  = prev2 >= uint8_t(0xe0u);
+    simd8<bool> is_fourth_byte = prev3 >= uint8_t(0xf0u);
+    // Use ^ instead of | for is_*_byte, because ^ is commutative, and the caller is using ^ as well.
+    // This will work fine because we only have to report errors for cases with 0-1 lead bytes.
+    // Multiple lead bytes implies 2 overlapping multibyte characters, and if that happens, there is
+    // guaranteed to be at least *one* lead byte that is part of only 1 other multibyte character.
+    // The error will be detected there.
+    return is_second_byte ^ is_third_byte ^ is_fourth_byte;
+}
+
+simdjson_inline simd8<uint8_t> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+    simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-0x80); // Only 111_____ will be >= 0x80
+    simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-0x80); // Only 1111____ will be >= 0x80
+    return is_third_byte | is_fourth_byte;
+}
+
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+
+//
+// Stage 2
+//
+
+//
+// Implementation-specific overrides
+//
+namespace simdjson {
+namespace arm64 {
+
+simdjson_warn_unused error_code implementation::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept {
+  return arm64::stage1::json_minifier::minify<64>(buf, len, dst, dst_len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage1(const uint8_t *_buf, size_t _len, stage1_mode streaming) noexcept {
+  this->buf = _buf;
+  this->len = _len;
+  return arm64::stage1::json_structural_indexer::index<64>(buf, len, *this, streaming);
+}
+
+simdjson_warn_unused bool implementation::validate_utf8(const char *buf, size_t len) const noexcept {
+  return arm64::stage1::generic_validate_utf8(buf,len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<false>(*this, _doc);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2_next(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<true>(*this, _doc);
+}
+
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) const noexcept {
+  return arm64::stringparsing::parse_string(src, dst, allow_replacement);
+}
+
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept {
+  return arm64::stringparsing::parse_wobbly_string(src, dst);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::parse(const uint8_t *_buf, size_t _len, dom::document &_doc) noexcept {
+  auto error = stage1(_buf, _len, stage1_mode::regular);
+  if (error) { return error; }
+  return stage2(_doc);
+}
+
+} // namespace arm64
+} // namespace simdjson
+
+/* including simdjson/arm64/end.h: #include <simdjson/arm64/end.h> */
+/* begin file simdjson/arm64/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#undef SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+/* undefining SIMDJSON_IMPLEMENTATION from "arm64" */
+#undef SIMDJSON_IMPLEMENTATION
+/* end file simdjson/arm64/end.h */
+
+#endif // SIMDJSON_SRC_ARM64_CPP
+/* end file arm64.cpp */
+#endif
+#if SIMDJSON_IMPLEMENTATION_HASWELL
+/* including haswell.cpp: #include <haswell.cpp> */
+/* begin file haswell.cpp */
+#ifndef SIMDJSON_SRC_HASWELL_CPP
+#define SIMDJSON_SRC_HASWELL_CPP
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+/* including simdjson/haswell.h: #include <simdjson/haswell.h> */
+/* begin file simdjson/haswell.h */
+#ifndef SIMDJSON_HASWELL_H
+#define SIMDJSON_HASWELL_H
+
+/* including simdjson/haswell/begin.h: #include "simdjson/haswell/begin.h" */
+/* begin file simdjson/haswell/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "haswell" */
+#define SIMDJSON_IMPLEMENTATION haswell
+
+/* including simdjson/haswell/base.h: #include "simdjson/haswell/base.h" */
+/* begin file simdjson/haswell/base.h */
+#ifndef SIMDJSON_HASWELL_BASE_H
+#define SIMDJSON_HASWELL_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// The constructor may be executed on any host, so we take care not to use SIMDJSON_TARGET_HASWELL
+namespace simdjson {
+/**
+ * Implementation for Haswell (Intel AVX2).
+ */
+namespace haswell {
+
+class implementation;
+
+namespace {
+namespace simd {
+template <typename T> struct simd8;
+template <typename T> struct simd8x64;
+} // namespace simd
+} // unnamed namespace
+
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_HASWELL_BASE_H
+/* end file simdjson/haswell/base.h */
+/* including simdjson/haswell/intrinsics.h: #include "simdjson/haswell/intrinsics.h" */
+/* begin file simdjson/haswell/intrinsics.h */
+#ifndef SIMDJSON_HASWELL_INTRINSICS_H
+#define SIMDJSON_HASWELL_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#if SIMDJSON_VISUAL_STUDIO
+// under clang within visual studio, this will include <x86intrin.h>
+#include <intrin.h>  // visual studio or clang
+#else
+#include <x86intrin.h> // elsewhere
+#endif // SIMDJSON_VISUAL_STUDIO
+
+#if SIMDJSON_CLANG_VISUAL_STUDIO
+/**
+ * You are not supposed, normally, to include these
+ * headers directly. Instead you should either include intrin.h
+ * or x86intrin.h. However, when compiling with clang
+ * under Windows (i.e., when _MSC_VER is set), these headers
+ * only get included *if* the corresponding features are detected
+ * from macros:
+ * e.g., if __AVX2__ is set... in turn,  we normally set these
+ * macros by compiling against the corresponding architecture
+ * (e.g., arch:AVX2, -mavx2, etc.) which compiles the whole
+ * software with these advanced instructions. In simdjson, we
+ * want to compile the whole program for a generic target,
+ * and only target our specific kernels. As a workaround,
+ * we directly include the needed headers. These headers would
+ * normally guard against such usage, but we carefully included
+ * <x86intrin.h>  (or <intrin.h>) before, so the headers
+ * are fooled.
+ */
+#include <bmiintrin.h>   // for _blsr_u64
+#include <lzcntintrin.h> // for  __lzcnt64
+#include <immintrin.h>   // for most things (AVX2, AVX512, _popcnt64)
+#include <smmintrin.h>
+#include <tmmintrin.h>
+#include <avxintrin.h>
+#include <avx2intrin.h>
+#include <wmmintrin.h>   // for  _mm_clmulepi64_si128
+// unfortunately, we may not get _blsr_u64, but, thankfully, clang
+// has it as a macro.
+#ifndef _blsr_u64
+// we roll our own
+#define _blsr_u64(n) ((n - 1) & n)
+#endif //  _blsr_u64
+#endif // SIMDJSON_CLANG_VISUAL_STUDIO
+
+static_assert(sizeof(__m256i) <= simdjson::SIMDJSON_PADDING, "insufficient padding for haswell kernel.");
+
+#endif // SIMDJSON_HASWELL_INTRINSICS_H
+/* end file simdjson/haswell/intrinsics.h */
+
+#if !SIMDJSON_CAN_ALWAYS_RUN_HASWELL
+// We enable bmi2 only if LLVM/clang is used, because GCC may not
+// make good use of it. See https://github.com/simdjson/simdjson/pull/2243
+#if defined(__clang__)
+SIMDJSON_TARGET_REGION("avx2,bmi,bmi2,pclmul,lzcnt,popcnt")
+#else
+SIMDJSON_TARGET_REGION("avx2,bmi,pclmul,lzcnt,popcnt")
+#endif
+#endif
+
+/* including simdjson/haswell/bitmanipulation.h: #include "simdjson/haswell/bitmanipulation.h" */
+/* begin file simdjson/haswell/bitmanipulation.h */
+#ifndef SIMDJSON_HASWELL_BITMANIPULATION_H
+#define SIMDJSON_HASWELL_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/bitmask.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  return (int)_tzcnt_u64(input_num);
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO
+  ////////
+  // You might expect the next line to be equivalent to
+  // return (int)_tzcnt_u64(input_num);
+  // but the generated code differs and might be less efficient?
+  ////////
+  return __builtin_ctzll(input_num);
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return _blsr_u64(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+  return int(_lzcnt_u64(input_num));
+}
+
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+simdjson_inline unsigned __int64 count_ones(uint64_t input_num) {
+  // note: we do not support legacy 32-bit Windows in this kernel
+  return __popcnt64(input_num);// Visual Studio wants two underscores
+}
+#else
+simdjson_inline long long int count_ones(uint64_t input_num) {
+  return _popcnt64(input_num);
+}
+#endif
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2,
+                                uint64_t *result) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  return _addcarry_u64(0, value1, value2,
+                       reinterpret_cast<unsigned __int64 *>(result));
+#else
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+#endif
+}
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_HASWELL_BITMANIPULATION_H
+/* end file simdjson/haswell/bitmanipulation.h */
+/* including simdjson/haswell/bitmask.h: #include "simdjson/haswell/bitmask.h" */
+/* begin file simdjson/haswell/bitmask.h */
+#ifndef SIMDJSON_HASWELL_BITMASK_H
+#define SIMDJSON_HASWELL_BITMASK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(const uint64_t bitmask) {
+  // There should be no such thing with a processor supporting avx2
+  // but not clmul.
+  __m128i all_ones = _mm_set1_epi8('\xFF');
+  __m128i result = _mm_clmulepi64_si128(_mm_set_epi64x(0ULL, bitmask), all_ones, 0);
+  return _mm_cvtsi128_si64(result);
+}
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_HASWELL_BITMASK_H
+/* end file simdjson/haswell/bitmask.h */
+/* including simdjson/haswell/numberparsing_defs.h: #include "simdjson/haswell/numberparsing_defs.h" */
+/* begin file simdjson/haswell/numberparsing_defs.h */
+#ifndef SIMDJSON_HASWELL_NUMBERPARSING_DEFS_H
+#define SIMDJSON_HASWELL_NUMBERPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace numberparsing {
+
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  // this actually computes *16* values so we are being wasteful.
+  const __m128i ascii0 = _mm_set1_epi8('0');
+  const __m128i mul_1_10 =
+      _mm_setr_epi8(10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1);
+  const __m128i mul_1_100 = _mm_setr_epi16(100, 1, 100, 1, 100, 1, 100, 1);
+  const __m128i mul_1_10000 =
+      _mm_setr_epi16(10000, 1, 10000, 1, 10000, 1, 10000, 1);
+  const __m128i input = _mm_sub_epi8(
+      _mm_loadu_si128(reinterpret_cast<const __m128i *>(chars)), ascii0);
+  const __m128i t1 = _mm_maddubs_epi16(input, mul_1_10);
+  const __m128i t2 = _mm_madd_epi16(t1, mul_1_100);
+  const __m128i t3 = _mm_packus_epi32(t2, t2);
+  const __m128i t4 = _mm_madd_epi16(t3, mul_1_10000);
+  return _mm_cvtsi128_si32(
+      t4); // only captures the sum of the first 8 digits, drop the rest
+}
+
+/** @private */
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+#if SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+#if SIMDJSON_IS_ARM64
+  // ARM64 has native support for 64-bit multiplications, no need to emultate
+  answer.high = __umulh(value1, value2);
+  answer.low = value1 * value2;
+#else
+  answer.low = _umul128(value1, value2, &answer.high); // _umul128 not available on ARM64
+#endif // SIMDJSON_IS_ARM64
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+#endif
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace haswell
+} // namespace simdjson
+
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+
+#endif // SIMDJSON_HASWELL_NUMBERPARSING_DEFS_H
+/* end file simdjson/haswell/numberparsing_defs.h */
+/* including simdjson/haswell/simd.h: #include "simdjson/haswell/simd.h" */
+/* begin file simdjson/haswell/simd.h */
+#ifndef SIMDJSON_HASWELL_SIMD_H
+#define SIMDJSON_HASWELL_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace simd {
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename Child>
+  struct base {
+    __m256i value;
+
+    // Zero constructor
+    simdjson_inline base() : value{__m256i()} {}
+
+    // Conversion from SIMD register
+    simdjson_inline base(const __m256i _value) : value(_value) {}
+
+    // Conversion to SIMD register
+    simdjson_inline operator const __m256i&() const { return this->value; }
+    simdjson_inline operator __m256i&() { return this->value; }
+
+    // Bit operations
+    simdjson_inline Child operator|(const Child other) const { return _mm256_or_si256(*this, other); }
+    simdjson_inline Child operator&(const Child other) const { return _mm256_and_si256(*this, other); }
+    simdjson_inline Child operator^(const Child other) const { return _mm256_xor_si256(*this, other); }
+    simdjson_inline Child bit_andnot(const Child other) const { return _mm256_andnot_si256(other, *this); }
+    simdjson_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+  };
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename T>
+  struct simd8;
+
+  template<typename T, typename Mask=simd8<bool>>
+  struct base8: base<simd8<T>> {
+    typedef uint32_t bitmask_t;
+    typedef uint64_t bitmask2_t;
+
+    simdjson_inline base8() : base<simd8<T>>() {}
+    simdjson_inline base8(const __m256i _value) : base<simd8<T>>(_value) {}
+
+    friend simdjson_really_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return _mm256_cmpeq_epi8(lhs, rhs); }
+
+    static const int SIZE = sizeof(base<T>::value);
+
+    template<int N=1>
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+      return _mm256_alignr_epi8(*this, _mm256_permute2x128_si256(prev_chunk, *this, 0x21), 16 - N);
+    }
+  };
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool>: base8<bool> {
+    static simdjson_inline simd8<bool> splat(bool _value) { return _mm256_set1_epi8(uint8_t(-(!!_value))); }
+
+    simdjson_inline simd8() : base8() {}
+    simdjson_inline simd8(const __m256i _value) : base8<bool>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+
+    simdjson_inline int to_bitmask() const { return _mm256_movemask_epi8(*this); }
+    simdjson_inline bool any() const { return !_mm256_testz_si256(*this, *this); }
+    simdjson_inline simd8<bool> operator~() const { return *this ^ true; }
+  };
+
+  template<typename T>
+  struct base8_numeric: base8<T> {
+    static simdjson_inline simd8<T> splat(T _value) { return _mm256_set1_epi8(_value); }
+    static simdjson_inline simd8<T> zero() { return _mm256_setzero_si256(); }
+    static simdjson_inline simd8<T> load(const T values[32]) {
+      return _mm256_loadu_si256(reinterpret_cast<const __m256i *>(values));
+    }
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    static simdjson_inline simd8<T> repeat_16(
+      T v0,  T v1,  T v2,  T v3,  T v4,  T v5,  T v6,  T v7,
+      T v8,  T v9,  T v10, T v11, T v12, T v13, T v14, T v15
+    ) {
+      return simd8<T>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    simdjson_inline base8_numeric() : base8<T>() {}
+    simdjson_inline base8_numeric(const __m256i _value) : base8<T>(_value) {}
+
+    // Store to array
+    simdjson_inline void store(T dst[32]) const { return _mm256_storeu_si256(reinterpret_cast<__m256i *>(dst), *this); }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<T> operator+(const simd8<T> other) const { return _mm256_add_epi8(*this, other); }
+    simdjson_inline simd8<T> operator-(const simd8<T> other) const { return _mm256_sub_epi8(*this, other); }
+    simdjson_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
+
+    // Override to distinguish from bool version
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return _mm256_shuffle_epi8(lookup_table, *this);
+    }
+
+    // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted as a bitset).
+    // Passing a 0 value for mask would be equivalent to writing out every byte to output.
+    // Only the first 32 - count_ones(mask) bytes of the result are significant but 32 bytes
+    // get written.
+    // Design consideration: it seems like a function with the
+    // signature simd8<L> compress(uint32_t mask) would be
+    // sensible, but the AVX ISA makes this kind of approach difficult.
+    template<typename L>
+    simdjson_inline void compress(uint32_t mask, L * output) const {
+      using internal::thintable_epi8;
+      using internal::BitsSetTable256mul2;
+      using internal::pshufb_combine_table;
+      // this particular implementation was inspired by work done by @animetosho
+      // we do it in four steps, first 8 bytes and then second 8 bytes...
+      uint8_t mask1 = uint8_t(mask); // least significant 8 bits
+      uint8_t mask2 = uint8_t(mask >> 8); // second least significant 8 bits
+      uint8_t mask3 = uint8_t(mask >> 16); // ...
+      uint8_t mask4 = uint8_t(mask >> 24); // ...
+      // next line just loads the 64-bit values thintable_epi8[mask1] and
+      // thintable_epi8[mask2] into a 128-bit register, using only
+      // two instructions on most compilers.
+      __m256i shufmask =  _mm256_set_epi64x(thintable_epi8[mask4], thintable_epi8[mask3],
+        thintable_epi8[mask2], thintable_epi8[mask1]);
+      // we increment by 0x08 the second half of the mask and so forth
+      shufmask =
+      _mm256_add_epi8(shufmask, _mm256_set_epi32(0x18181818, 0x18181818,
+         0x10101010, 0x10101010, 0x08080808, 0x08080808, 0, 0));
+      // this is the version "nearly pruned"
+      __m256i pruned = _mm256_shuffle_epi8(*this, shufmask);
+      // we still need to put the  pieces back together.
+      // we compute the popcount of the first words:
+      int pop1 = BitsSetTable256mul2[mask1];
+      int pop3 = BitsSetTable256mul2[mask3];
+
+      // then load the corresponding mask
+      // could be done with _mm256_loadu2_m128i but many standard libraries omit this intrinsic.
+      __m256i v256 = _mm256_castsi128_si256(
+        _mm_loadu_si128(reinterpret_cast<const __m128i *>(pshufb_combine_table + pop1 * 8)));
+      __m256i compactmask = _mm256_insertf128_si256(v256,
+         _mm_loadu_si128(reinterpret_cast<const __m128i *>(pshufb_combine_table + pop3 * 8)), 1);
+      __m256i almostthere =  _mm256_shuffle_epi8(pruned, compactmask);
+      // We just need to write out the result.
+      // This is the tricky bit that is hard to do
+      // if we want to return a SIMD register, since there
+      // is no single-instruction approach to recombine
+      // the two 128-bit lanes with an offset.
+      __m128i v128;
+      v128 = _mm256_castsi256_si128(almostthere);
+      _mm_storeu_si128( reinterpret_cast<__m128i *>(output), v128);
+      v128 = _mm256_extractf128_si256(almostthere, 1);
+      _mm_storeu_si128( reinterpret_cast<__m128i *>(output + 16 - count_ones(mask & 0xFFFF)), v128);
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> : base8_numeric<int8_t> {
+    simdjson_inline simd8() : base8_numeric<int8_t>() {}
+    simdjson_inline simd8(const __m256i _value) : base8_numeric<int8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const int8_t values[32]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15,
+      int8_t v16, int8_t v17, int8_t v18, int8_t v19, int8_t v20, int8_t v21, int8_t v22, int8_t v23,
+      int8_t v24, int8_t v25, int8_t v26, int8_t v27, int8_t v28, int8_t v29, int8_t v30, int8_t v31
+    ) : simd8(_mm256_setr_epi8(
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15,
+      v16,v17,v18,v19,v20,v21,v22,v23,
+      v24,v25,v26,v27,v28,v29,v30,v31
+    )) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm256_max_epi8(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm256_min_epi8(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm256_cmpgt_epi8(*this, other); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return _mm256_cmpgt_epi8(other, *this); }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t>: base8_numeric<uint8_t> {
+    simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+    simdjson_inline simd8(const __m256i _value) : base8_numeric<uint8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const uint8_t values[32]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15,
+      uint8_t v16, uint8_t v17, uint8_t v18, uint8_t v19, uint8_t v20, uint8_t v21, uint8_t v22, uint8_t v23,
+      uint8_t v24, uint8_t v25, uint8_t v26, uint8_t v27, uint8_t v28, uint8_t v29, uint8_t v30, uint8_t v31
+    ) : simd8(_mm256_setr_epi8(
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15,
+      v16,v17,v18,v19,v20,v21,v22,v23,
+      v24,v25,v26,v27,v28,v29,v30,v31
+    )) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm256_adds_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm256_subs_epu8(*this, other); }
+
+    // Order-specific operations
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm256_max_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return _mm256_min_epu8(other, *this); }
+    // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
+    // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->lt_bits(other).any_bits_set(); }
+
+    // Bit-specific operations
+    simdjson_inline simd8<bool> bits_not_set() const { return *this == uint8_t(0); }
+    simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
+    simdjson_inline bool is_ascii() const { return _mm256_movemask_epi8(*this) == 0; }
+    simdjson_inline bool bits_not_set_anywhere() const { return _mm256_testz_si256(*this, *this); }
+    simdjson_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
+    simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return _mm256_testz_si256(*this, bits); }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm256_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(_mm256_slli_epi16(*this, N)) & uint8_t(0xFFu << N); }
+    // Get one of the bits and make a bitmask out of it.
+    // e.g. value.get_bit<7>() gets the high bit
+    template<int N>
+    simdjson_inline int get_bit() const { return _mm256_movemask_epi8(_mm256_slli_epi16(*this, 7-N)); }
+  };
+
+  template<typename T>
+  struct simd8x64 {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+    static_assert(NUM_CHUNKS == 2, "Haswell kernel should use two registers per 64-byte block.");
+    const simd8<T> chunks[NUM_CHUNKS];
+    template<int idx> simd8<uint8_t> get() const { return idx < NUM_CHUNKS ? chunks[idx] : simd8<T>(); }
+
+    simd8x64(const simd8x64<T>& o) = delete; // no copy allowed
+    simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1) : chunks{chunk0, chunk1} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+32)} {}
+
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
+      uint32_t mask1 = uint32_t(mask);
+      uint32_t mask2 = uint32_t(mask >> 32);
+      this->chunks[0].compress(mask1, output);
+      this->chunks[1].compress(mask2, output + 32 - count_ones(mask1));
+      return 64 - count_ones(mask);
+    }
+
+    simdjson_inline void store(T ptr[64]) const {
+      this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
+      this->chunks[1].store(ptr+sizeof(simd8<T>)*1);
+    }
+
+    simdjson_inline uint64_t to_bitmask() const {
+      uint64_t r_lo = uint32_t(this->chunks[0].to_bitmask());
+      uint64_t r_hi =                       this->chunks[1].to_bitmask();
+      return r_lo | (r_hi << 32);
+    }
+
+    simdjson_inline simd8<T> reduce_or() const {
+      return this->chunks[0] | this->chunks[1];
+    }
+
+    simdjson_inline simd8x64<T> bit_or(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return simd8x64<T>(
+        this->chunks[0] | mask,
+        this->chunks[1] | mask
+      );
+    }
+
+    simdjson_inline uint64_t eq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] == mask,
+        this->chunks[1] == mask
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+      return  simd8x64<bool>(
+        this->chunks[0] == other.chunks[0],
+        this->chunks[1] == other.chunks[1]
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t lteq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] <= mask,
+        this->chunks[1] <= mask
+      ).to_bitmask();
+    }
+  }; // struct simd8x64<T>
+
+} // namespace simd
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_HASWELL_SIMD_H
+/* end file simdjson/haswell/simd.h */
+/* including simdjson/haswell/stringparsing_defs.h: #include "simdjson/haswell/stringparsing_defs.h" */
+/* begin file simdjson/haswell/stringparsing_defs.h */
+#ifndef SIMDJSON_HASWELL_STRINGPARSING_DEFS_H
+#define SIMDJSON_HASWELL_STRINGPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/simd.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 32;
+  simdjson_inline backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return ((quote_bits - 1) & bs_bits) != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+
+  uint32_t bs_bits;
+  uint32_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // this can read up to 15 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  // store to dest unconditionally - we can overwrite the bits we don't like later
+  v.store(dst);
+  return {
+      static_cast<uint32_t>((v == '\\').to_bitmask()),     // bs_bits
+      static_cast<uint32_t>((v == '"').to_bitmask()), // quote_bits
+  };
+}
+
+
+struct escaping {
+  static constexpr uint32_t BYTES_PROCESSED = 32;
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits != 0; }
+  simdjson_inline int escape_index() { return trailing_zeroes(escape_bits); }
+
+  uint64_t escape_bits;
+}; // struct escaping
+
+
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "escaping finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  simd8<bool> is_quote = (v == '"');
+  simd8<bool> is_backslash = (v == '\\');
+  simd8<bool> is_control = (v < 32);
+  return {
+    uint64_t((is_backslash | is_quote | is_control).to_bitmask())
+  };
+}
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_HASWELL_STRINGPARSING_DEFS_H
+/* end file simdjson/haswell/stringparsing_defs.h */
+/* end file simdjson/haswell/begin.h */
+/* including simdjson/generic/amalgamated.h for haswell: #include "simdjson/generic/amalgamated.h" */
+/* begin file simdjson/generic/amalgamated.h for haswell */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_GENERIC_DEPENDENCIES_H)
+#error simdjson/generic/dependencies.h must be included before simdjson/generic/amalgamated.h!
+#endif
+
+/* including simdjson/generic/base.h for haswell: #include "simdjson/generic/base.h" */
+/* begin file simdjson/generic/base.h for haswell */
+#ifndef SIMDJSON_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): // If we haven't got an implementation yet, we're in the editor, editing a generic file! Just */
+/* amalgamation skipped (editor-only): // use the most advanced one we can so the most possible stuff can be tested. */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation_detection.h" */
+/* amalgamation skipped (editor-only): #if SIMDJSON_IMPLEMENTATION_ICELAKE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_HASWELL */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_WESTMERE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_ARM64 */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_PPC64 */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LASX */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LSX */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_RVV_VLS */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_FALLBACK */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/begin.h" */
+/* amalgamation skipped (editor-only): #else */
+/* amalgamation skipped (editor-only): #error "All possible implementations (including fallback) have been disabled! simdjson will not run." */
+/* amalgamation skipped (editor-only): #endif */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+
+struct open_container;
+class dom_parser_implementation;
+
+/**
+ * The type of a JSON number
+ */
+enum class number_type {
+    floating_point_number=1, /// a binary64 number
+    signed_integer,          /// a signed integer that fits in a 64-bit word using two's complement
+    unsigned_integer,        /// a positive integer larger or equal to 1<<63
+    big_integer              /// a big integer that does not fit in a 64-bit word
+};
+
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_BASE_H
+/* end file simdjson/generic/base.h for haswell */
+/* including simdjson/generic/jsoncharutils.h for haswell: #include "simdjson/generic/jsoncharutils.h" */
+/* begin file simdjson/generic/jsoncharutils.h for haswell */
+#ifndef SIMDJSON_GENERIC_JSONCHARUTILS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_JSONCHARUTILS_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/jsoncharutils_tables.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace jsoncharutils {
+
+// return non-zero if not a structural or whitespace char
+// zero otherwise
+simdjson_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace_negated[c];
+}
+
+simdjson_inline uint32_t is_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace[c];
+}
+
+// returns a value with the high 16 bits set if not valid
+// otherwise returns the conversion of the 4 hex digits at src into the bottom
+// 16 bits of the 32-bit return register
+//
+// see
+// https://lemire.me/blog/2019/04/17/parsing-short-hexadecimal-strings-efficiently/
+static inline uint32_t hex_to_u32_nocheck(
+    const uint8_t *src) { // strictly speaking, static inline is a C-ism
+  uint32_t v1 = internal::digit_to_val32[630 + src[0]];
+  uint32_t v2 = internal::digit_to_val32[420 + src[1]];
+  uint32_t v3 = internal::digit_to_val32[210 + src[2]];
+  uint32_t v4 = internal::digit_to_val32[0 + src[3]];
+  return v1 | v2 | v3 | v4;
+}
+
+// given a code point cp, writes to c
+// the utf-8 code, outputting the length in
+// bytes, if the length is zero, the code point
+// is invalid
+//
+// This can possibly be made faster using pdep
+// and clz and table lookups, but JSON documents
+// have few escaped code points, and the following
+// function looks cheap.
+//
+// Note: we assume that surrogates are treated separately
+//
+simdjson_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
+  if (cp <= 0x7F) {
+    c[0] = uint8_t(cp);
+    return 1; // ascii
+  }
+  if (cp <= 0x7FF) {
+    c[0] = uint8_t((cp >> 6) + 192);
+    c[1] = uint8_t((cp & 63) + 128);
+    return 2; // universal plane
+    //  Surrogates are treated elsewhere...
+    //} //else if (0xd800 <= cp && cp <= 0xdfff) {
+    //  return 0; // surrogates // could put assert here
+  } else if (cp <= 0xFFFF) {
+    c[0] = uint8_t((cp >> 12) + 224);
+    c[1] = uint8_t(((cp >> 6) & 63) + 128);
+    c[2] = uint8_t((cp & 63) + 128);
+    return 3;
+  } else if (cp <= 0x10FFFF) { // if you know you have a valid code point, this
+                               // is not needed
+    c[0] = uint8_t((cp >> 18) + 240);
+    c[1] = uint8_t(((cp >> 12) & 63) + 128);
+    c[2] = uint8_t(((cp >> 6) & 63) + 128);
+    c[3] = uint8_t((cp & 63) + 128);
+    return 4;
+  }
+  // will return 0 when the code point was too large.
+  return 0; // bad r
+}
+
+#if SIMDJSON_IS_32BITS // _umul128 for x86, arm
+// this is a slow emulation routine for 32-bit
+//
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+  return x * (uint64_t)y;
+}
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+  uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
+  uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
+  uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
+  uint64_t adbc_carry = !!(adbc < ad);
+  uint64_t lo = bd + (adbc << 32);
+  *hi = __emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) +
+        (adbc_carry << 32) + !!(lo < bd);
+  return lo;
+}
+#endif
+
+} // namespace jsoncharutils
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_JSONCHARUTILS_H
+/* end file simdjson/generic/jsoncharutils.h for haswell */
+/* including simdjson/generic/atomparsing.h for haswell: #include "simdjson/generic/atomparsing.h" */
+/* begin file simdjson/generic/atomparsing.h for haswell */
+#ifndef SIMDJSON_GENERIC_ATOMPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_ATOMPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace haswell {
+namespace {
+/// @private
+namespace atomparsing {
+
+// The string_to_uint32 is exclusively used to map literal strings to 32-bit values.
+// We use memcpy instead of a pointer cast to avoid undefined behaviors since we cannot
+// be certain that the character pointer will be properly aligned.
+// You might think that using memcpy makes this function expensive, but you'd be wrong.
+// All decent optimizing compilers (GCC, clang, Visual Studio) will compile string_to_uint32("false");
+// to the compile-time constant 1936482662.
+simdjson_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
+
+
+// Again in str4ncmp we use a memcpy to avoid undefined behavior. The memcpy may appear expensive.
+// Yet all decent optimizing compilers will compile memcpy to a single instruction, just about.
+simdjson_warn_unused
+simdjson_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
+  uint32_t srcval; // we want to avoid unaligned 32-bit loads (undefined in C/C++)
+  static_assert(sizeof(uint32_t) <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be larger than 4 bytes");
+  std::memcpy(&srcval, src, sizeof(uint32_t));
+  return srcval ^ string_to_uint32(atom);
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src) {
+  return (str4ncmp(src, "true") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_true_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "true"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src) {
+  return (str4ncmp(src+1, "alse") | jsoncharutils::is_not_structural_or_whitespace(src[5])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
+  if (len > 5) { return is_valid_false_atom(src); }
+  else if (len == 5) { return !str4ncmp(src+1, "alse"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src) {
+  return (str4ncmp(src, "null") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_null_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "null"); }
+  else { return false; }
+}
+
+} // namespace atomparsing
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_ATOMPARSING_H
+/* end file simdjson/generic/atomparsing.h for haswell */
+/* including simdjson/generic/dom_parser_implementation.h for haswell: #include "simdjson/generic/dom_parser_implementation.h" */
+/* begin file simdjson/generic/dom_parser_implementation.h for haswell */
+#ifndef SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/dom_parser_implementation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+
+// expectation: sizeof(open_container) = 64/8.
+struct open_container {
+  uint32_t tape_index; // where, on the tape, does the scope ([,{) begins
+  uint32_t count; // how many elements in the scope
+}; // struct open_container
+
+static_assert(sizeof(open_container) == 64/8, "Open container must be 64 bits");
+
+class dom_parser_implementation final : public internal::dom_parser_implementation {
+public:
+  /** Tape location of each open { or [ */
+  std::unique_ptr<open_container[]> open_containers{};
+  /** Whether each open container is a [ or { */
+  std::unique_ptr<bool[]> is_array{};
+  /** Buffer passed to stage 1 */
+  const uint8_t *buf{};
+  /** Length passed to stage 1 */
+  size_t len{0};
+  /** Document passed to stage 2 */
+  dom::document *doc{};
+
+  inline dom_parser_implementation() noexcept;
+  inline dom_parser_implementation(dom_parser_implementation &&other) noexcept;
+  inline dom_parser_implementation &operator=(dom_parser_implementation &&other) noexcept;
+  dom_parser_implementation(const dom_parser_implementation &) = delete;
+  dom_parser_implementation &operator=(const dom_parser_implementation &) = delete;
+
+  simdjson_warn_unused error_code parse(const uint8_t *buf, size_t len, dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage1(const uint8_t *buf, size_t len, stage1_mode partial) noexcept final;
+  simdjson_warn_unused error_code stage2(dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage2_next(dom::document &doc) noexcept final;
+  simdjson_warn_unused uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) const noexcept final;
+  simdjson_warn_unused uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept final;
+  inline simdjson_warn_unused error_code set_capacity(size_t capacity) noexcept final;
+  inline simdjson_warn_unused error_code set_max_depth(size_t max_depth) noexcept final;
+private:
+  simdjson_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
+
+};
+
+} // namespace haswell
+} // namespace simdjson
+
+namespace simdjson {
+namespace haswell {
+
+inline dom_parser_implementation::dom_parser_implementation() noexcept = default;
+inline dom_parser_implementation::dom_parser_implementation(dom_parser_implementation &&other) noexcept = default;
+inline dom_parser_implementation &dom_parser_implementation::operator=(dom_parser_implementation &&other) noexcept = default;
+
+// Leaving these here so they can be inlined if so desired
+inline simdjson_warn_unused error_code dom_parser_implementation::set_capacity(size_t capacity) noexcept {
+  if(capacity > SIMDJSON_MAXSIZE_BYTES) { return CAPACITY; }
+  // Stage 1 index output
+  size_t max_structures = SIMDJSON_ROUNDUP_N(capacity, 64) + 2 + 7;
+  structural_indexes.reset( new (std::nothrow) uint32_t[max_structures] );
+  if (!structural_indexes) { _capacity = 0; return MEMALLOC; }
+  structural_indexes[0] = 0;
+  n_structural_indexes = 0;
+
+  _capacity = capacity;
+  return SUCCESS;
+}
+
+inline simdjson_warn_unused error_code dom_parser_implementation::set_max_depth(size_t max_depth) noexcept {
+  // Stage 2 stacks
+  open_containers.reset(new (std::nothrow) open_container[max_depth]);
+  is_array.reset(new (std::nothrow) bool[max_depth]);
+  if (!is_array || !open_containers) { _max_depth = 0; return MEMALLOC; }
+
+  _max_depth = max_depth;
+  return SUCCESS;
+}
+
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file simdjson/generic/dom_parser_implementation.h for haswell */
+/* including simdjson/generic/implementation_simdjson_result_base.h for haswell: #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base.h for haswell */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+
+// This is a near copy of include/error.h's implementation_simdjson_result_base, except it doesn't use std::pair
+// so we can avoid inlining errors
+// TODO reconcile these!
+/**
+ * The result of a simdjson operation that could fail.
+ *
+ * Gives the option of reading error codes, or throwing an exception by casting to the desired result.
+ *
+ * This is a base class for implementations that want to add functions to the result type for
+ * chaining.
+ *
+ * Override like:
+ *
+ *   struct simdjson_result<T> : public internal::implementation_simdjson_result_base<T> {
+ *     simdjson_result() noexcept : internal::implementation_simdjson_result_base<T>() {}
+ *     simdjson_result(error_code error) noexcept : internal::implementation_simdjson_result_base<T>(error) {}
+ *     simdjson_result(T &&value) noexcept : internal::implementation_simdjson_result_base<T>(std::forward(value)) {}
+ *     simdjson_result(T &&value, error_code error) noexcept : internal::implementation_simdjson_result_base<T>(value, error) {}
+ *     // Your extra methods here
+ *   }
+ *
+ * Then any method returning simdjson_result<T> will be chainable with your methods.
+ */
+template<typename T>
+struct implementation_simdjson_result_base {
+
+  /**
+   * Create a new empty result with error = UNINITIALIZED.
+   */
+  simdjson_inline implementation_simdjson_result_base() noexcept = default;
+
+  /**
+   * Create a new error result.
+   */
+  simdjson_inline implementation_simdjson_result_base(error_code error) noexcept;
+
+  /**
+   * Create a new successful result.
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value) noexcept;
+
+  /**
+   * Create a new result with both things (use if you don't want to branch when creating the result).
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value, error_code error) noexcept;
+
+  /**
+   * Move the value and the error to the provided variables.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   * @param error The variable to assign the error to. Set to SUCCESS if there is no error.
+   */
+  simdjson_inline void tie(T &value, error_code &error) && noexcept;
+
+  /**
+   * Move the value to the provided variable.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   */
+  simdjson_warn_unused simdjson_inline error_code get(T &value) && noexcept;
+
+  /**
+   * The error.
+   */
+  simdjson_warn_unused simdjson_inline error_code error() const noexcept;
+
+  /**
+   * Whether there is a value.
+   */
+  simdjson_warn_unused simdjson_inline bool has_value() const noexcept;
+
+#if SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T& operator*() &  noexcept(false);
+  simdjson_inline T&& operator*() &&  noexcept(false);
+  /**
+   * Arrow operator to access members of the contained value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T* operator->() noexcept(false);
+  simdjson_inline const T* operator->() const noexcept(false);
+
+  simdjson_inline T& value() & noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& value() && noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& take_value() && noexcept(false);
+
+  /**
+   * Cast to the value (will throw on error).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline operator T&&() && noexcept(false);
+
+
+#endif // SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline const T& value_unsafe() const& noexcept;
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T& value_unsafe() & noexcept;
+  /**
+   * Take the result value (move it). This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T&& value_unsafe() && noexcept;
+
+  using value_type = T;
+  using error_type = error_code;
+
+protected:
+  /** users should never directly access first and second. **/
+  T first{}; /** Users should never directly access 'first'. **/
+  error_code second{UNINITIALIZED}; /** Users should never directly access 'second'. **/
+}; // struct implementation_simdjson_result_base
+
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+/* end file simdjson/generic/implementation_simdjson_result_base.h for haswell */
+/* including simdjson/generic/numberparsing.h for haswell: #include "simdjson/generic/numberparsing.h" */
+/* begin file simdjson/generic/numberparsing.h for haswell */
+#ifndef SIMDJSON_GENERIC_NUMBERPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_NUMBERPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <limits>
+#include <ostream>
+#include <cstring>
+
+namespace simdjson {
+namespace haswell {
+namespace numberparsing {
+
+#ifdef JSON_TEST_NUMBERS
+#define INVALID_NUMBER(SRC) (found_invalid_number((SRC)), NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (found_integer((VALUE), (SRC)), (WRITER).append_s64((VALUE)))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (found_unsigned_integer((VALUE), (SRC)), (WRITER).append_u64((VALUE)))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (found_float((VALUE), (SRC)), (WRITER).append_double((VALUE)))
+#define BIGINT_NUMBER(SRC) (found_invalid_number((SRC)), BIGINT_ERROR)
+#else
+#define INVALID_NUMBER(SRC) (NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (WRITER).append_s64((VALUE))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (WRITER).append_u64((VALUE))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (WRITER).append_double((VALUE))
+#define BIGINT_NUMBER(SRC) (BIGINT_ERROR)
+#endif
+
+namespace {
+
+// Convert a mantissa, an exponent and a sign bit into an ieee64 double.
+// The real_exponent needs to be in [0, 2046] (technically real_exponent = 2047 would be acceptable).
+// The mantissa should be in [0,1<<53). The bit at index (1ULL << 52) while be zeroed.
+simdjson_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
+    double d;
+    mantissa &= ~(1ULL << 52);
+    mantissa |= real_exponent << 52;
+    mantissa |= ((static_cast<uint64_t>(negative)) << 63);
+    std::memcpy(&d, &mantissa, sizeof(d));
+    return d;
+}
+
+// Attempts to compute i * 10^(power) exactly; and if "negative" is
+// true, negate the result.
+// This function will only work in some cases, when it does not work, success is
+// set to false. This should work *most of the time* (like 99% of the time).
+// We assume that power is in the [smallest_power,
+// largest_power] interval: the caller is responsible for this check.
+simdjson_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
+  // we start with a fast path
+  // It was described in
+  // Clinger WD. How to read floating point numbers accurately.
+  // ACM SIGPLAN Notices. 1990
+#ifndef FLT_EVAL_METHOD
+#error "FLT_EVAL_METHOD should be defined, please include cfloat."
+#endif
+#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
+  // We cannot be certain that x/y is rounded to nearest.
+  if (0 <= power && power <= 22 && i <= 9007199254740991)
+#else
+  if (-22 <= power && power <= 22 && i <= 9007199254740991)
+#endif
+  {
+    // convert the integer into a double. This is lossless since
+    // 0 <= i <= 2^53 - 1.
+    d = double(i);
+    //
+    // The general idea is as follows.
+    // If 0 <= s < 2^53 and if 10^0 <= p <= 10^22 then
+    // 1) Both s and p can be represented exactly as 64-bit floating-point
+    // values
+    // (binary64).
+    // 2) Because s and p can be represented exactly as floating-point values,
+    // then s * p
+    // and s / p will produce correctly rounded values.
+    //
+    if (power < 0) {
+      d = d / simdjson::internal::power_of_ten[-power];
+    } else {
+      d = d * simdjson::internal::power_of_ten[power];
+    }
+    if (negative) {
+      d = -d;
+    }
+    return true;
+  }
+  // When 22 < power && power <  22 + 16, we could
+  // hope for another, secondary fast path.  It was
+  // described by David M. Gay in  "Correctly rounded
+  // binary-decimal and decimal-binary conversions." (1990)
+  // If you need to compute i * 10^(22 + x) for x < 16,
+  // first compute i * 10^x, if you know that result is exact
+  // (e.g., when i * 10^x < 2^53),
+  // then you can still proceed and do (i * 10^x) * 10^22.
+  // Is this worth your time?
+  // You need  22 < power *and* power <  22 + 16 *and* (i * 10^(x-22) < 2^53)
+  // for this second fast path to work.
+  // If you you have 22 < power *and* power <  22 + 16, and then you
+  // optimistically compute "i * 10^(x-22)", there is still a chance that you
+  // have wasted your time if i * 10^(x-22) >= 2^53. It makes the use cases of
+  // this optimization maybe less common than we would like. Source:
+  // http://www.exploringbinary.com/fast-path-decimal-to-floating-point-conversion/
+  // also used in RapidJSON: https://rapidjson.org/strtod_8h_source.html
+
+  // The fast path has now failed, so we are failing back on the slower path.
+
+  // In the slow path, we need to adjust i so that it is > 1<<63 which is always
+  // possible, except if i == 0, so we handle i == 0 separately.
+  if(i == 0) {
+    d = negative ? -0.0 : 0.0;
+    return true;
+  }
+
+
+  // The exponent is 1024 + 63 + power
+  //     + floor(log(5**power)/log(2)).
+  // The 1024 comes from the ieee64 standard.
+  // The 63 comes from the fact that we use a 64-bit word.
+  //
+  // Computing floor(log(5**power)/log(2)) could be
+  // slow. Instead we use a fast function.
+  //
+  // For power in (-400,350), we have that
+  // (((152170 + 65536) * power ) >> 16);
+  // is equal to
+  //  floor(log(5**power)/log(2)) + power when power >= 0
+  // and it is equal to
+  //  ceil(log(5**-power)/log(2)) + power when power < 0
+  //
+  // The 65536 is (1<<16) and corresponds to
+  // (65536 * power) >> 16 ---> power
+  //
+  // ((152170 * power ) >> 16) is equal to
+  // floor(log(5**power)/log(2))
+  //
+  // Note that this is not magic: 152170/(1<<16) is
+  // approximately equal to log(5)/log(2).
+  // The 1<<16 value is a power of two; we could use a
+  // larger power of 2 if we wanted to.
+  //
+  int64_t exponent = (((152170 + 65536) * power) >> 16) + 1024 + 63;
+
+
+  // We want the most significant bit of i to be 1. Shift if needed.
+  int lz = leading_zeroes(i);
+  i <<= lz;
+
+
+  // We are going to need to do some 64-bit arithmetic to get a precise product.
+  // We use a table lookup approach.
+  // It is safe because
+  // power >= smallest_power
+  // and power <= largest_power
+  // We recover the mantissa of the power, it has a leading 1. It is always
+  // rounded down.
+  //
+  // We want the most significant 64 bits of the product. We know
+  // this will be non-zero because the most significant bit of i is
+  // 1.
+  const uint32_t index = 2 * uint32_t(power - simdjson::internal::smallest_power);
+  // Optimization: It may be that materializing the index as a variable might confuse some compilers and prevent effective complex-addressing loads. (Done for code clarity.)
+  //
+  // The full_multiplication function computes the 128-bit product of two 64-bit words
+  // with a returned value of type value128 with a "low component" corresponding to the
+  // 64-bit least significant bits of the product and with a "high component" corresponding
+  // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index]);
+#else
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index]);
+#endif
+
+  // Both i and power_of_five_128[index] have their most significant bit set to 1 which
+  // implies that the either the most or the second most significant bit of the product
+  // is 1. We pack values in this manner for efficiency reasons: it maximizes the use
+  // we make of the product. It also makes it easy to reason about the product: there
+  // is 0 or 1 leading zero in the product.
+
+  // Unless the least significant 9 bits of the high (64-bit) part of the full
+  // product are all 1s, then we know that the most significant 55 bits are
+  // exact and no further work is needed. Having 55 bits is necessary because
+  // we need 53 bits for the mantissa but we have to have one rounding bit and
+  // we can waste a bit if the most significant bit of the product is zero.
+  if((firstproduct.high & 0x1FF) == 0x1FF) {
+    // We want to compute i * 5^q, but only care about the top 55 bits at most.
+    // Consider the scenario where q>=0. Then 5^q may not fit in 64-bits. Doing
+    // the full computation is wasteful. So we do what is called a "truncated
+    // multiplication".
+    // We take the most significant 64-bits, and we put them in
+    // power_of_five_128[index]. Usually, that's good enough to approximate i * 5^q
+    // to the desired approximation using one multiplication. Sometimes it does not suffice.
+    // Then we store the next most significant 64 bits in power_of_five_128[index + 1], and
+    // then we get a better approximation to i * 5^q.
+    //
+    // That's for when q>=0. The logic for q<0 is somewhat similar but it is somewhat
+    // more complicated.
+    //
+    // There is an extra layer of complexity in that we need more than 55 bits of
+    // accuracy in the round-to-even scenario.
+    //
+    // The full_multiplication function computes the 128-bit product of two 64-bit words
+    // with a returned value of type value128 with a "low component" corresponding to the
+    // 64-bit least significant bits of the product and with a "high component" corresponding
+    // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index + 1]);
+#else
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index + 1]);
+#endif
+    firstproduct.low += secondproduct.high;
+    if(secondproduct.high > firstproduct.low) { firstproduct.high++; }
+    // As it has been proven by Noble Mushtak and Daniel Lemire in "Fast Number Parsing Without
+    // Fallback" (https://arxiv.org/abs/2212.06644), at this point we are sure that the product
+    // is sufficiently accurate, and more computation is not needed.
+  }
+  uint64_t lower = firstproduct.low;
+  uint64_t upper = firstproduct.high;
+  // The final mantissa should be 53 bits with a leading 1.
+  // We shift it so that it occupies 54 bits with a leading 1.
+  ///////
+  uint64_t upperbit = upper >> 63;
+  uint64_t mantissa = upper >> (upperbit + 9);
+  lz += int(1 ^ upperbit);
+
+  // Here we have mantissa < (1<<54).
+  int64_t real_exponent = exponent - lz;
+  if (simdjson_unlikely(real_exponent <= 0)) { // we have a subnormal?
+    // Here have that real_exponent <= 0 so -real_exponent >= 0
+    if(-real_exponent + 1 >= 64) { // if we have more than 64 bits below the minimum exponent, you have a zero for sure.
+      d = negative ? -0.0 : 0.0;
+      return true;
+    }
+    // next line is safe because -real_exponent + 1 < 0
+    mantissa >>= -real_exponent + 1;
+    // Thankfully, we can't have both "round-to-even" and subnormals because
+    // "round-to-even" only occurs for powers close to 0.
+    mantissa += (mantissa & 1); // round up
+    mantissa >>= 1;
+    // There is a weird scenario where we don't have a subnormal but just.
+    // Suppose we start with 2.2250738585072013e-308, we end up
+    // with 0x3fffffffffffff x 2^-1023-53 which is technically subnormal
+    // whereas 0x40000000000000 x 2^-1023-53  is normal. Now, we need to round
+    // up 0x3fffffffffffff x 2^-1023-53  and once we do, we are no longer
+    // subnormal, but we can only know this after rounding.
+    // So we only declare a subnormal if we are smaller than the threshold.
+    real_exponent = (mantissa < (uint64_t(1) << 52)) ? 0 : 1;
+    d = to_double(mantissa, real_exponent, negative);
+    return true;
+  }
+  // We have to round to even. The "to even" part
+  // is only a problem when we are right in between two floats
+  // which we guard against.
+  // If we have lots of trailing zeros, we may fall right between two
+  // floating-point values.
+  //
+  // The round-to-even cases take the form of a number 2m+1 which is in (2^53,2^54]
+  // times a power of two. That is, it is right between a number with binary significand
+  // m and another number with binary significand m+1; and it must be the case
+  // that it cannot be represented by a float itself.
+  //
+  // We must have that w * 10 ^q == (2m+1) * 2^p for some power of two 2^p.
+  // Recall that 10^q = 5^q * 2^q.
+  // When q >= 0, we must have that (2m+1) is divible by 5^q, so 5^q <= 2^54. We have that
+  //  5^23 <=  2^54 and it is the last power of five to qualify, so q <= 23.
+  // When q<0, we have  w  >=  (2m+1) x 5^{-q}.  We must have that w<2^{64} so
+  // (2m+1) x 5^{-q} < 2^{64}. We have that 2m+1>2^{53}. Hence, we must have
+  // 2^{53} x 5^{-q} < 2^{64}.
+  // Hence we have 5^{-q} < 2^{11}$ or q>= -4.
+  //
+  // We require lower <= 1 and not lower == 0 because we could not prove that
+  // that lower == 0 is implied; but we could prove that lower <= 1 is a necessary and sufficient test.
+  if (simdjson_unlikely((lower <= 1) && (power >= -4) && (power <= 23) && ((mantissa & 3) == 1))) {
+    if((mantissa  << (upperbit + 64 - 53 - 2)) ==  upper) {
+      mantissa &= ~1;             // flip it so that we do not round up
+    }
+  }
+
+  mantissa += mantissa & 1;
+  mantissa >>= 1;
+
+  // Here we have mantissa < (1<<53), unless there was an overflow
+  if (mantissa >= (1ULL << 53)) {
+    //////////
+    // This will happen when parsing values such as 7.2057594037927933e+16
+    ////////
+    mantissa = (1ULL << 52);
+    real_exponent++;
+  }
+  mantissa &= ~(1ULL << 52);
+  // we have to check that real_exponent is in range, otherwise we bail out
+  if (simdjson_unlikely(real_exponent > 2046)) {
+    // We have an infinite value!!! We could actually throw an error here if we could.
+    return false;
+  }
+  d = to_double(mantissa, real_exponent, negative);
+  return true;
+}
+
+// We call a fallback floating-point parser that might be slow. Note
+// it will accept JSON numbers, but the JSON spec. is more restrictive so
+// before you call parse_float_fallback, you need to have validated the input
+// string with the JSON grammar.
+// It will return an error (false) if the parsed number is infinite.
+// The string parsing itself always succeeds. We know that there is at least
+// one digit.
+static bool parse_float_fallback(const uint8_t *ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+static bool parse_float_fallback(const uint8_t *ptr, const uint8_t *end_ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr), reinterpret_cast<const char *>(end_ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+// check quickly whether the next 8 chars are made of digits
+// at a glance, it looks better than Mula's
+// http://0x80.pl/articles/swar-digits-validate.html
+simdjson_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
+  uint64_t val;
+  // this can read up to 7 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(7 <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be bigger than 7");
+  std::memcpy(&val, chars, 8);
+  // a branchy method might be faster:
+  // return (( val & 0xF0F0F0F0F0F0F0F0 ) == 0x3030303030303030)
+  //  && (( (val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0 ) ==
+  //  0x3030303030303030);
+  return (((val & 0xF0F0F0F0F0F0F0F0) |
+           (((val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0) >> 4)) ==
+          0x3333333333333333);
+}
+
+template<typename I>
+SIMDJSON_NO_SANITIZE_UNDEFINED // We deliberately allow overflow here and check later
+simdjson_inline bool parse_digit(const uint8_t c, I &i) {
+  const uint8_t digit = static_cast<uint8_t>(c - '0');
+  if (digit > 9) {
+    return false;
+  }
+  // PERF NOTE: multiplication by 10 is cheaper than arbitrary integer multiplication
+  i = 10 * i + digit; // might overflow, we will handle the overflow later
+  return true;
+}
+
+simdjson_inline bool is_digit(const uint8_t c) {
+  return static_cast<uint8_t>(c - '0') <= 9;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_decimal_after_separator(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
+  // we continue with the fiction that we have an integer. If the
+  // floating point number is representable as x * 10^z for some integer
+  // z that fits in 53 bits, then we will be able to convert back the
+  // the integer into a float in a lossless manner.
+  const uint8_t *const first_after_period = p;
+
+#ifdef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_SWAR_NUMBER_PARSING
+  // this helps if we have lots of decimals!
+  // this turns out to be frequent enough.
+  if (is_made_of_eight_digits_fast(p)) {
+    i = i * 100000000 + parse_eight_digits_unrolled(p);
+    p += 8;
+  }
+#endif // SIMDJSON_SWAR_NUMBER_PARSING
+#endif // #ifdef SIMDJSON_SWAR_NUMBER_PARSING
+  // Unrolling the first digit makes a small difference on some implementations (e.g. westmere)
+  if (parse_digit(*p, i)) { ++p; }
+  while (parse_digit(*p, i)) { p++; }
+  exponent = first_after_period - p;
+  // Decimal without digits (123.) is illegal
+  if (exponent == 0) {
+    return INVALID_NUMBER(src);
+  }
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
+  // Exp Sign: -123.456e[-]78
+  bool neg_exp = ('-' == *p);
+  if (neg_exp || '+' == *p) { p++; } // Skip + as well
+
+  // Exponent: -123.456e-[78]
+  auto start_exp = p;
+  int64_t exp_number = 0;
+  while (parse_digit(*p, exp_number)) { ++p; }
+  // It is possible for parse_digit to overflow.
+  // In particular, it could overflow to INT64_MIN, and we cannot do - INT64_MIN.
+  // Thus we *must* check for possible overflow before we negate exp_number.
+
+  // Performance notes: it may seem like combining the two "simdjson_unlikely checks" below into
+  // a single simdjson_unlikely path would be faster. The reasoning is sound, but the compiler may
+  // not oblige and may, in fact, generate two distinct paths in any case. It might be
+  // possible to do uint64_t(p - start_exp - 1) >= 18 but it could end up trading off
+  // instructions for a simdjson_likely branch, an unconclusive gain.
+
+  // If there were no digits, it's an error.
+  if (simdjson_unlikely(p == start_exp)) {
+    return INVALID_NUMBER(src);
+  }
+  // We have a valid positive exponent in exp_number at this point, except that
+  // it may have overflowed.
+
+  // If there were more than 18 digits, we may have overflowed the integer. We have to do
+  // something!!!!
+  if (simdjson_unlikely(p > start_exp+18)) {
+    // Skip leading zeroes: 1e000000000000000000001 is technically valid and does not overflow
+    while (*start_exp == '0') { start_exp++; }
+    // 19 digits could overflow int64_t and is kind of absurd anyway. We don't
+    // support exponents smaller than -999,999,999,999,999,999 and bigger
+    // than 999,999,999,999,999,999.
+    // We can truncate.
+    // Note that 999999999999999999 is assuredly too large. The maximal ieee64 value before
+    // infinity is ~1.8e308. The smallest subnormal is ~5e-324. So, actually, we could
+    // truncate at 324.
+    // Note that there is no reason to fail per se at this point in time.
+    // E.g., 0e999999999999999999999 is a fine number.
+    if (p > start_exp+18) { exp_number = 999999999999999999; }
+  }
+  // At this point, we know that exp_number is a sane, positive, signed integer.
+  // It is <= 999,999,999,999,999,999. As long as 'exponent' is in
+  // [-8223372036854775808, 8223372036854775808], we won't overflow. Because 'exponent'
+  // is bounded in magnitude by the size of the JSON input, we are fine in this universe.
+  // To sum it up: the next line should never overflow.
+  exponent += (neg_exp ? -exp_number : exp_number);
+  return SUCCESS;
+}
+
+simdjson_inline bool check_if_integer(const uint8_t *const src, size_t max_length) {
+  const uint8_t *const srcend = src + max_length;
+  bool negative = (*src == '-'); // we can always read at least one character after the '-'
+  const uint8_t *p = src + uint8_t(negative);
+  if(p == srcend) { return false; }
+  if(*p == '0') {
+    ++p;
+    if(p == srcend) { return true; }
+    if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+    return true;
+  }
+  while(p != srcend && is_digit(*p)) { ++p; }
+  if(p == srcend) { return true; }
+  if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+  return true;
+}
+
+simdjson_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
+  // It is possible that the integer had an overflow.
+  // We have to handle the case where we have 0.0000somenumber.
+  const uint8_t *start = start_digits;
+  while ((*start == '0') || (*start == '.')) { ++start; }
+  // we over-decrement by one when there is a '.'
+  return digit_count - size_t(start - start_digits);
+}
+
+} // unnamed namespace
+
+/** @private */
+static error_code slow_float_parsing(simdjson_unused const uint8_t * src, double* answer) {
+  if (parse_float_fallback(src, answer)) {
+    return SUCCESS;
+  }
+  return INVALID_NUMBER(src);
+}
+
+/** @private */
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
+  // If we frequently had to deal with long strings of digits,
+  // we could extend our code by using a 128-bit integer instead
+  // of a 64-bit integer. However, this is uncommon in practice.
+  //
+  // 9999999999999999999 < 2**64 so we can accommodate 19 digits.
+  // If we have a decimal separator, then digit_count - 1 is the number of digits, but we
+  // may not have a decimal separator!
+  if (simdjson_unlikely(digit_count > 19 && significant_digits(start_digits, digit_count) > 19)) {
+    // Ok, chances are good that we had an overflow!
+    // this is almost never going to get called!!!
+    // we start anew, going slowly!!!
+    // This will happen in the following examples:
+    // 10000000000000000000000000000000000000000000e+308
+    // 3.1415926535897932384626433832795028841971693993751
+    //
+    // NOTE: We do not pass a reference to the to slow_float_parsing. If we passed our writer
+    // reference to it, it would force it to be stored in memory, preventing the compiler from
+    // picking it apart and putting into registers. i.e. if we pass it as reference,
+    // it gets slow.
+    double d;
+    error_code error = slow_float_parsing(src, &d);
+    writer.append_double(d);
+    return error;
+  }
+  // NOTE: it's weird that the simdjson_unlikely() only wraps half the if, but it seems to get slower any other
+  // way we've tried: https://github.com/simdjson/simdjson/pull/990#discussion_r448497331
+  // To future reader: we'd love if someone found a better way, or at least could explain this result!
+  if (simdjson_unlikely(exponent < simdjson::internal::smallest_power) || (exponent > simdjson::internal::largest_power)) {
+    //
+    // Important: smallest_power is such that it leads to a zero value.
+    // Observe that 18446744073709551615e-343 == 0, i.e. (2**64 - 1) e -343 is zero
+    // so something x 10^-343 goes to zero, but not so with  something x 10^-342.
+    static_assert(simdjson::internal::smallest_power <= -342, "smallest_power is not small enough");
+    //
+    if((exponent < simdjson::internal::smallest_power) || (i == 0)) {
+      // E.g. Parse "-0.0e-999" into the same value as "-0.0". See https://en.wikipedia.org/wiki/Signed_zero
+      WRITE_DOUBLE(negative ? -0.0 : 0.0, src, writer);
+      return SUCCESS;
+    } else { // (exponent > largest_power) and (i != 0)
+      // We have, for sure, an infinite value and simdjson refuses to parse infinite values.
+      return INVALID_NUMBER(src);
+    }
+  }
+  double d;
+  if (!compute_float_64(exponent, i, negative, d)) {
+    // we are almost never going to get here.
+    if (!parse_float_fallback(src, &d)) { return INVALID_NUMBER(src); }
+  }
+  WRITE_DOUBLE(d, src, writer);
+  return SUCCESS;
+}
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer);
+
+// for performance analysis, it is sometimes  useful to skip parsing
+#ifdef SIMDJSON_SKIPNUMBERPARSING
+
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const, W &writer) {
+  writer.append_s64(0);        // always write zero
+  return SUCCESS;              // always succeeds
+}
+
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept { return number_type::signed_integer; }
+#else
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer) {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  if (digit_count == 0 || ('0' == *start_digits && digit_count > 1)) { return INVALID_NUMBER(src); }
+
+  //
+  // Handle floats if there is a . or e (or both)
+  //
+  int64_t exponent = 0;
+  bool is_float = false;
+  if ('.' == *p) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_decimal_after_separator(src, p, i, exponent) );
+    digit_count = int(p - start_digits); // used later to guard against overflows
+  }
+  if (('e' == *p) || ('E' == *p)) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_exponent(src, p, exponent) );
+  }
+  if (is_float) {
+    const bool dirty_end = jsoncharutils::is_not_structural_or_whitespace(*p);
+    SIMDJSON_TRY( write_float(src, negative, i, start_digits, digit_count, exponent, writer) );
+    if (dirty_end) { return INVALID_NUMBER(src); }
+    return SUCCESS;
+  }
+
+  // The longest negative 64-bit number is 19 digits.
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  size_t longest_digit_count = negative ? 19 : 20;
+  if (digit_count > longest_digit_count) { return BIGINT_NUMBER(src); }
+  if (digit_count == longest_digit_count) {
+    if (negative) {
+      // Anything negative above INT64_MAX+1 is invalid
+      if (i > uint64_t(INT64_MAX)+1) { return BIGINT_NUMBER(src);  }
+      WRITE_INTEGER(~i+1, src, writer);
+      if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+      return SUCCESS;
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    }  else if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INVALID_NUMBER(src); }
+  }
+
+  // Write unsigned if it does not fit in a signed integer.
+  if (i > uint64_t(INT64_MAX)) {
+    WRITE_UNSIGNED(i, src, writer);
+  } else {
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+    if(i == 0 && negative) {
+      // We have to write -0.0 instead of 0
+      WRITE_DOUBLE(-0.0, src, writer);
+    } else {
+      WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+    }
+#else
+  WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+#endif
+  }
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+  return SUCCESS;
+}
+
+// Inlineable functions
+namespace {
+
+// This table can be used to characterize the final character of an integer
+// string. For JSON structural character and allowable white space characters,
+// we return SUCCESS. For 'e', '.' and 'E', we return INCORRECT_TYPE. Otherwise
+// we return NUMBER_ERROR.
+// Optimization note: we could easily reduce the size of the table by half (to 128)
+// at the cost of an extra branch.
+// Optimization note: we want the values to use at most 8 bits (not, e.g., 32 bits):
+static_assert(error_code(uint8_t(NUMBER_ERROR))== NUMBER_ERROR, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(SUCCESS))== SUCCESS, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(INCORRECT_TYPE))== INCORRECT_TYPE, "bad NUMBER_ERROR cast");
+
+const uint8_t integer_string_finisher[256] = {
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   SUCCESS,      NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, INCORRECT_TYPE,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, SUCCESS,        NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR};
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if ((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
+  const uint8_t *p = src + 1;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    // Note: we use src[1] and not src[0] because src[0] is the quote character in this
+    // instance.
+    if (src[1] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  //
+  // Check for minus sign
+  //
+  if(src == src_end) { return NUMBER_ERROR; }
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = src;
+  uint64_t i = 0;
+  while (parse_digit(*src, i)) { src++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(src - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*src)) {
+  //  return (*src == '.' || *src == 'e' || *src == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(*src != '"') { return NUMBER_ERROR; }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept {
+  return (*src == '-');
+}
+
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) { return true; }
+  return false;
+}
+
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  size_t digit_count = size_t(p - src);
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) {
+    static const uint8_t * smaller_big_integer = reinterpret_cast<const uint8_t *>("9223372036854775808");
+    // We have an integer.
+    if(simdjson_unlikely(digit_count > 20)) {
+      return number_type::big_integer;
+    }
+    // If the number is negative and valid, it must be a signed integer.
+    if(negative) {
+      if (simdjson_unlikely(digit_count > 19)) return number_type::big_integer;
+      if (simdjson_unlikely(digit_count == 19 && memcmp(src, smaller_big_integer, 19) > 0)) {
+        return number_type::big_integer;
+      }
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+      if(digit_count == 1 && src[0] == '0') {
+        // We have to write -0.0 instead of 0
+        return number_type::floating_point_number;
+      }
+#endif
+      return number_type::signed_integer;
+    }
+    // Let us check if we have a big integer (>=2**64).
+    static const uint8_t * two_to_sixtyfour = reinterpret_cast<const uint8_t *>("18446744073709551616");
+    if((digit_count > 20) || (digit_count == 20 && memcmp(src, two_to_sixtyfour, 20) >= 0)) {
+      return number_type::big_integer;
+    }
+    // The number is positive and smaller than 18446744073709551616 (or 2**64).
+    // We want values larger or equal to 9223372036854775808 to be unsigned
+    // integers, and the other values to be signed integers.
+    if((digit_count == 20) || (digit_count >= 19 && memcmp(src, smaller_big_integer, 19) >= 0)) {
+      return number_type::unsigned_integer;
+    }
+    return number_type::signed_integer;
+  }
+  // Hopefully, we have 'e' or 'E' or '.'.
+  return number_type::floating_point_number;
+}
+
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
+  if(src == src_end) { return NUMBER_ERROR; }
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  if(p == src_end) { return NUMBER_ERROR; }
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely((p != src_end) && (*p == '.'))) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if ((p == src_end) || !parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while ((p != src_end) && parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = start_digits-src > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if ((p != src_end) && (*p == 'e' || *p == 'E')) {
+    p++;
+    if(p == src_end) { return NUMBER_ERROR; }
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while ((p != src_end) && parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if ((p != src_end) && jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), src_end, &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+} // unnamed namespace
+#endif // SIMDJSON_SKIPNUMBERPARSING
+
+} // namespace numberparsing
+
+inline std::ostream& operator<<(std::ostream& out, number_type type) noexcept {
+    switch (type) {
+        case number_type::signed_integer: out << "integer in [-9223372036854775808,9223372036854775808)"; break;
+        case number_type::unsigned_integer: out << "unsigned integer in [9223372036854775808,18446744073709551616)"; break;
+        case number_type::floating_point_number: out << "floating-point number (binary64)"; break;
+        case number_type::big_integer: out << "big integer"; break;
+        default: SIMDJSON_UNREACHABLE();
+    }
+    return out;
+}
+
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_NUMBERPARSING_H
+/* end file simdjson/generic/numberparsing.h for haswell */
+
+/* including simdjson/generic/implementation_simdjson_result_base-inl.h for haswell: #include "simdjson/generic/implementation_simdjson_result_base-inl.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base-inl.h for haswell */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+
+//
+// internal::implementation_simdjson_result_base<T> inline implementation
+//
+
+template<typename T>
+simdjson_inline void implementation_simdjson_result_base<T>::tie(T &value, error_code &error) && noexcept {
+  error = this->second;
+  if (!error) {
+    value = std::forward<implementation_simdjson_result_base<T>>(*this).first;
+  }
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::get(T &value) && noexcept {
+  error_code error;
+  std::forward<implementation_simdjson_result_base<T>>(*this).tie(value, error);
+  return error;
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::error() const noexcept {
+  return this->second;
+}
+
+
+template<typename T>
+simdjson_warn_unused simdjson_inline bool implementation_simdjson_result_base<T>::has_value() const noexcept {
+  return this->error() == SUCCESS;
+}
+
+#if SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::operator*() &  noexcept(false) {
+  return this->value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::operator*() &&  noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).value();
+}
+
+template<typename T>
+simdjson_inline T* implementation_simdjson_result_base<T>::operator->() noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+
+template<typename T>
+simdjson_inline const T* implementation_simdjson_result_base<T>::operator->() const noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value() & noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::take_value() && noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::operator T&&() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+#endif // SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline const T& implementation_simdjson_result_base<T>::value_unsafe() const& noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value_unsafe() & noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value_unsafe() && noexcept {
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value, error_code error) noexcept
+    : first{std::forward<T>(value)}, second{error} {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(error_code error) noexcept
+    : implementation_simdjson_result_base(T{}, error) {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value) noexcept
+    : implementation_simdjson_result_base(std::forward<T>(value), SUCCESS) {}
+
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+/* end file simdjson/generic/implementation_simdjson_result_base-inl.h for haswell */
+/* end file simdjson/generic/amalgamated.h for haswell */
+/* including simdjson/haswell/end.h: #include "simdjson/haswell/end.h" */
+/* begin file simdjson/haswell/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#if !SIMDJSON_CAN_ALWAYS_RUN_HASWELL
+SIMDJSON_UNTARGET_REGION
+#endif
+
+/* undefining SIMDJSON_IMPLEMENTATION from "haswell" */
+#undef SIMDJSON_IMPLEMENTATION
+/* end file simdjson/haswell/end.h */
+
+#endif // SIMDJSON_HASWELL_H
+/* end file simdjson/haswell.h */
+/* including simdjson/haswell/implementation.h: #include <simdjson/haswell/implementation.h> */
+/* begin file simdjson/haswell/implementation.h */
+#ifndef SIMDJSON_HASWELL_IMPLEMENTATION_H
+#define SIMDJSON_HASWELL_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/instruction_set.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// The constructor may be executed on any host, so we take care not to use SIMDJSON_TARGET_HASWELL
+namespace simdjson {
+namespace haswell {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation(
+      "haswell",
+      "Intel/AMD AVX2",
+      internal::instruction_set::AVX2 | internal::instruction_set::PCLMULQDQ | internal::instruction_set::BMI1 | internal::instruction_set::BMI2
+  ) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_HASWELL_IMPLEMENTATION_H
+/* end file simdjson/haswell/implementation.h */
+
+/* including simdjson/haswell/begin.h: #include <simdjson/haswell/begin.h> */
+/* begin file simdjson/haswell/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "haswell" */
+#define SIMDJSON_IMPLEMENTATION haswell
+
+/* including simdjson/haswell/base.h: #include "simdjson/haswell/base.h" */
+/* begin file simdjson/haswell/base.h */
+#ifndef SIMDJSON_HASWELL_BASE_H
+#define SIMDJSON_HASWELL_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// The constructor may be executed on any host, so we take care not to use SIMDJSON_TARGET_HASWELL
+namespace simdjson {
+/**
+ * Implementation for Haswell (Intel AVX2).
+ */
+namespace haswell {
+
+class implementation;
+
+namespace {
+namespace simd {
+template <typename T> struct simd8;
+template <typename T> struct simd8x64;
+} // namespace simd
+} // unnamed namespace
+
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_HASWELL_BASE_H
+/* end file simdjson/haswell/base.h */
+/* including simdjson/haswell/intrinsics.h: #include "simdjson/haswell/intrinsics.h" */
+/* begin file simdjson/haswell/intrinsics.h */
+#ifndef SIMDJSON_HASWELL_INTRINSICS_H
+#define SIMDJSON_HASWELL_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#if SIMDJSON_VISUAL_STUDIO
+// under clang within visual studio, this will include <x86intrin.h>
+#include <intrin.h>  // visual studio or clang
+#else
+#include <x86intrin.h> // elsewhere
+#endif // SIMDJSON_VISUAL_STUDIO
+
+#if SIMDJSON_CLANG_VISUAL_STUDIO
+/**
+ * You are not supposed, normally, to include these
+ * headers directly. Instead you should either include intrin.h
+ * or x86intrin.h. However, when compiling with clang
+ * under Windows (i.e., when _MSC_VER is set), these headers
+ * only get included *if* the corresponding features are detected
+ * from macros:
+ * e.g., if __AVX2__ is set... in turn,  we normally set these
+ * macros by compiling against the corresponding architecture
+ * (e.g., arch:AVX2, -mavx2, etc.) which compiles the whole
+ * software with these advanced instructions. In simdjson, we
+ * want to compile the whole program for a generic target,
+ * and only target our specific kernels. As a workaround,
+ * we directly include the needed headers. These headers would
+ * normally guard against such usage, but we carefully included
+ * <x86intrin.h>  (or <intrin.h>) before, so the headers
+ * are fooled.
+ */
+#include <bmiintrin.h>   // for _blsr_u64
+#include <lzcntintrin.h> // for  __lzcnt64
+#include <immintrin.h>   // for most things (AVX2, AVX512, _popcnt64)
+#include <smmintrin.h>
+#include <tmmintrin.h>
+#include <avxintrin.h>
+#include <avx2intrin.h>
+#include <wmmintrin.h>   // for  _mm_clmulepi64_si128
+// unfortunately, we may not get _blsr_u64, but, thankfully, clang
+// has it as a macro.
+#ifndef _blsr_u64
+// we roll our own
+#define _blsr_u64(n) ((n - 1) & n)
+#endif //  _blsr_u64
+#endif // SIMDJSON_CLANG_VISUAL_STUDIO
+
+static_assert(sizeof(__m256i) <= simdjson::SIMDJSON_PADDING, "insufficient padding for haswell kernel.");
+
+#endif // SIMDJSON_HASWELL_INTRINSICS_H
+/* end file simdjson/haswell/intrinsics.h */
+
+#if !SIMDJSON_CAN_ALWAYS_RUN_HASWELL
+// We enable bmi2 only if LLVM/clang is used, because GCC may not
+// make good use of it. See https://github.com/simdjson/simdjson/pull/2243
+#if defined(__clang__)
+SIMDJSON_TARGET_REGION("avx2,bmi,bmi2,pclmul,lzcnt,popcnt")
+#else
+SIMDJSON_TARGET_REGION("avx2,bmi,pclmul,lzcnt,popcnt")
+#endif
+#endif
+
+/* including simdjson/haswell/bitmanipulation.h: #include "simdjson/haswell/bitmanipulation.h" */
+/* begin file simdjson/haswell/bitmanipulation.h */
+#ifndef SIMDJSON_HASWELL_BITMANIPULATION_H
+#define SIMDJSON_HASWELL_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/bitmask.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  return (int)_tzcnt_u64(input_num);
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO
+  ////////
+  // You might expect the next line to be equivalent to
+  // return (int)_tzcnt_u64(input_num);
+  // but the generated code differs and might be less efficient?
+  ////////
+  return __builtin_ctzll(input_num);
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return _blsr_u64(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+  return int(_lzcnt_u64(input_num));
+}
+
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+simdjson_inline unsigned __int64 count_ones(uint64_t input_num) {
+  // note: we do not support legacy 32-bit Windows in this kernel
+  return __popcnt64(input_num);// Visual Studio wants two underscores
+}
+#else
+simdjson_inline long long int count_ones(uint64_t input_num) {
+  return _popcnt64(input_num);
+}
+#endif
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2,
+                                uint64_t *result) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  return _addcarry_u64(0, value1, value2,
+                       reinterpret_cast<unsigned __int64 *>(result));
+#else
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+#endif
+}
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_HASWELL_BITMANIPULATION_H
+/* end file simdjson/haswell/bitmanipulation.h */
+/* including simdjson/haswell/bitmask.h: #include "simdjson/haswell/bitmask.h" */
+/* begin file simdjson/haswell/bitmask.h */
+#ifndef SIMDJSON_HASWELL_BITMASK_H
+#define SIMDJSON_HASWELL_BITMASK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(const uint64_t bitmask) {
+  // There should be no such thing with a processor supporting avx2
+  // but not clmul.
+  __m128i all_ones = _mm_set1_epi8('\xFF');
+  __m128i result = _mm_clmulepi64_si128(_mm_set_epi64x(0ULL, bitmask), all_ones, 0);
+  return _mm_cvtsi128_si64(result);
+}
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_HASWELL_BITMASK_H
+/* end file simdjson/haswell/bitmask.h */
+/* including simdjson/haswell/numberparsing_defs.h: #include "simdjson/haswell/numberparsing_defs.h" */
+/* begin file simdjson/haswell/numberparsing_defs.h */
+#ifndef SIMDJSON_HASWELL_NUMBERPARSING_DEFS_H
+#define SIMDJSON_HASWELL_NUMBERPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace numberparsing {
+
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  // this actually computes *16* values so we are being wasteful.
+  const __m128i ascii0 = _mm_set1_epi8('0');
+  const __m128i mul_1_10 =
+      _mm_setr_epi8(10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1);
+  const __m128i mul_1_100 = _mm_setr_epi16(100, 1, 100, 1, 100, 1, 100, 1);
+  const __m128i mul_1_10000 =
+      _mm_setr_epi16(10000, 1, 10000, 1, 10000, 1, 10000, 1);
+  const __m128i input = _mm_sub_epi8(
+      _mm_loadu_si128(reinterpret_cast<const __m128i *>(chars)), ascii0);
+  const __m128i t1 = _mm_maddubs_epi16(input, mul_1_10);
+  const __m128i t2 = _mm_madd_epi16(t1, mul_1_100);
+  const __m128i t3 = _mm_packus_epi32(t2, t2);
+  const __m128i t4 = _mm_madd_epi16(t3, mul_1_10000);
+  return _mm_cvtsi128_si32(
+      t4); // only captures the sum of the first 8 digits, drop the rest
+}
+
+/** @private */
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+#if SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+#if SIMDJSON_IS_ARM64
+  // ARM64 has native support for 64-bit multiplications, no need to emultate
+  answer.high = __umulh(value1, value2);
+  answer.low = value1 * value2;
+#else
+  answer.low = _umul128(value1, value2, &answer.high); // _umul128 not available on ARM64
+#endif // SIMDJSON_IS_ARM64
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+#endif
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace haswell
+} // namespace simdjson
+
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+
+#endif // SIMDJSON_HASWELL_NUMBERPARSING_DEFS_H
+/* end file simdjson/haswell/numberparsing_defs.h */
+/* including simdjson/haswell/simd.h: #include "simdjson/haswell/simd.h" */
+/* begin file simdjson/haswell/simd.h */
+#ifndef SIMDJSON_HASWELL_SIMD_H
+#define SIMDJSON_HASWELL_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace simd {
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename Child>
+  struct base {
+    __m256i value;
+
+    // Zero constructor
+    simdjson_inline base() : value{__m256i()} {}
+
+    // Conversion from SIMD register
+    simdjson_inline base(const __m256i _value) : value(_value) {}
+
+    // Conversion to SIMD register
+    simdjson_inline operator const __m256i&() const { return this->value; }
+    simdjson_inline operator __m256i&() { return this->value; }
+
+    // Bit operations
+    simdjson_inline Child operator|(const Child other) const { return _mm256_or_si256(*this, other); }
+    simdjson_inline Child operator&(const Child other) const { return _mm256_and_si256(*this, other); }
+    simdjson_inline Child operator^(const Child other) const { return _mm256_xor_si256(*this, other); }
+    simdjson_inline Child bit_andnot(const Child other) const { return _mm256_andnot_si256(other, *this); }
+    simdjson_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+  };
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename T>
+  struct simd8;
+
+  template<typename T, typename Mask=simd8<bool>>
+  struct base8: base<simd8<T>> {
+    typedef uint32_t bitmask_t;
+    typedef uint64_t bitmask2_t;
+
+    simdjson_inline base8() : base<simd8<T>>() {}
+    simdjson_inline base8(const __m256i _value) : base<simd8<T>>(_value) {}
+
+    friend simdjson_really_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return _mm256_cmpeq_epi8(lhs, rhs); }
+
+    static const int SIZE = sizeof(base<T>::value);
+
+    template<int N=1>
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+      return _mm256_alignr_epi8(*this, _mm256_permute2x128_si256(prev_chunk, *this, 0x21), 16 - N);
+    }
+  };
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool>: base8<bool> {
+    static simdjson_inline simd8<bool> splat(bool _value) { return _mm256_set1_epi8(uint8_t(-(!!_value))); }
+
+    simdjson_inline simd8() : base8() {}
+    simdjson_inline simd8(const __m256i _value) : base8<bool>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+
+    simdjson_inline int to_bitmask() const { return _mm256_movemask_epi8(*this); }
+    simdjson_inline bool any() const { return !_mm256_testz_si256(*this, *this); }
+    simdjson_inline simd8<bool> operator~() const { return *this ^ true; }
+  };
+
+  template<typename T>
+  struct base8_numeric: base8<T> {
+    static simdjson_inline simd8<T> splat(T _value) { return _mm256_set1_epi8(_value); }
+    static simdjson_inline simd8<T> zero() { return _mm256_setzero_si256(); }
+    static simdjson_inline simd8<T> load(const T values[32]) {
+      return _mm256_loadu_si256(reinterpret_cast<const __m256i *>(values));
+    }
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    static simdjson_inline simd8<T> repeat_16(
+      T v0,  T v1,  T v2,  T v3,  T v4,  T v5,  T v6,  T v7,
+      T v8,  T v9,  T v10, T v11, T v12, T v13, T v14, T v15
+    ) {
+      return simd8<T>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    simdjson_inline base8_numeric() : base8<T>() {}
+    simdjson_inline base8_numeric(const __m256i _value) : base8<T>(_value) {}
+
+    // Store to array
+    simdjson_inline void store(T dst[32]) const { return _mm256_storeu_si256(reinterpret_cast<__m256i *>(dst), *this); }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<T> operator+(const simd8<T> other) const { return _mm256_add_epi8(*this, other); }
+    simdjson_inline simd8<T> operator-(const simd8<T> other) const { return _mm256_sub_epi8(*this, other); }
+    simdjson_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
+
+    // Override to distinguish from bool version
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return _mm256_shuffle_epi8(lookup_table, *this);
+    }
+
+    // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted as a bitset).
+    // Passing a 0 value for mask would be equivalent to writing out every byte to output.
+    // Only the first 32 - count_ones(mask) bytes of the result are significant but 32 bytes
+    // get written.
+    // Design consideration: it seems like a function with the
+    // signature simd8<L> compress(uint32_t mask) would be
+    // sensible, but the AVX ISA makes this kind of approach difficult.
+    template<typename L>
+    simdjson_inline void compress(uint32_t mask, L * output) const {
+      using internal::thintable_epi8;
+      using internal::BitsSetTable256mul2;
+      using internal::pshufb_combine_table;
+      // this particular implementation was inspired by work done by @animetosho
+      // we do it in four steps, first 8 bytes and then second 8 bytes...
+      uint8_t mask1 = uint8_t(mask); // least significant 8 bits
+      uint8_t mask2 = uint8_t(mask >> 8); // second least significant 8 bits
+      uint8_t mask3 = uint8_t(mask >> 16); // ...
+      uint8_t mask4 = uint8_t(mask >> 24); // ...
+      // next line just loads the 64-bit values thintable_epi8[mask1] and
+      // thintable_epi8[mask2] into a 128-bit register, using only
+      // two instructions on most compilers.
+      __m256i shufmask =  _mm256_set_epi64x(thintable_epi8[mask4], thintable_epi8[mask3],
+        thintable_epi8[mask2], thintable_epi8[mask1]);
+      // we increment by 0x08 the second half of the mask and so forth
+      shufmask =
+      _mm256_add_epi8(shufmask, _mm256_set_epi32(0x18181818, 0x18181818,
+         0x10101010, 0x10101010, 0x08080808, 0x08080808, 0, 0));
+      // this is the version "nearly pruned"
+      __m256i pruned = _mm256_shuffle_epi8(*this, shufmask);
+      // we still need to put the  pieces back together.
+      // we compute the popcount of the first words:
+      int pop1 = BitsSetTable256mul2[mask1];
+      int pop3 = BitsSetTable256mul2[mask3];
+
+      // then load the corresponding mask
+      // could be done with _mm256_loadu2_m128i but many standard libraries omit this intrinsic.
+      __m256i v256 = _mm256_castsi128_si256(
+        _mm_loadu_si128(reinterpret_cast<const __m128i *>(pshufb_combine_table + pop1 * 8)));
+      __m256i compactmask = _mm256_insertf128_si256(v256,
+         _mm_loadu_si128(reinterpret_cast<const __m128i *>(pshufb_combine_table + pop3 * 8)), 1);
+      __m256i almostthere =  _mm256_shuffle_epi8(pruned, compactmask);
+      // We just need to write out the result.
+      // This is the tricky bit that is hard to do
+      // if we want to return a SIMD register, since there
+      // is no single-instruction approach to recombine
+      // the two 128-bit lanes with an offset.
+      __m128i v128;
+      v128 = _mm256_castsi256_si128(almostthere);
+      _mm_storeu_si128( reinterpret_cast<__m128i *>(output), v128);
+      v128 = _mm256_extractf128_si256(almostthere, 1);
+      _mm_storeu_si128( reinterpret_cast<__m128i *>(output + 16 - count_ones(mask & 0xFFFF)), v128);
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> : base8_numeric<int8_t> {
+    simdjson_inline simd8() : base8_numeric<int8_t>() {}
+    simdjson_inline simd8(const __m256i _value) : base8_numeric<int8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const int8_t values[32]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15,
+      int8_t v16, int8_t v17, int8_t v18, int8_t v19, int8_t v20, int8_t v21, int8_t v22, int8_t v23,
+      int8_t v24, int8_t v25, int8_t v26, int8_t v27, int8_t v28, int8_t v29, int8_t v30, int8_t v31
+    ) : simd8(_mm256_setr_epi8(
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15,
+      v16,v17,v18,v19,v20,v21,v22,v23,
+      v24,v25,v26,v27,v28,v29,v30,v31
+    )) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm256_max_epi8(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm256_min_epi8(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm256_cmpgt_epi8(*this, other); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return _mm256_cmpgt_epi8(other, *this); }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t>: base8_numeric<uint8_t> {
+    simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+    simdjson_inline simd8(const __m256i _value) : base8_numeric<uint8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const uint8_t values[32]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15,
+      uint8_t v16, uint8_t v17, uint8_t v18, uint8_t v19, uint8_t v20, uint8_t v21, uint8_t v22, uint8_t v23,
+      uint8_t v24, uint8_t v25, uint8_t v26, uint8_t v27, uint8_t v28, uint8_t v29, uint8_t v30, uint8_t v31
+    ) : simd8(_mm256_setr_epi8(
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15,
+      v16,v17,v18,v19,v20,v21,v22,v23,
+      v24,v25,v26,v27,v28,v29,v30,v31
+    )) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm256_adds_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm256_subs_epu8(*this, other); }
+
+    // Order-specific operations
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm256_max_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return _mm256_min_epu8(other, *this); }
+    // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
+    // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->lt_bits(other).any_bits_set(); }
+
+    // Bit-specific operations
+    simdjson_inline simd8<bool> bits_not_set() const { return *this == uint8_t(0); }
+    simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
+    simdjson_inline bool is_ascii() const { return _mm256_movemask_epi8(*this) == 0; }
+    simdjson_inline bool bits_not_set_anywhere() const { return _mm256_testz_si256(*this, *this); }
+    simdjson_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
+    simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return _mm256_testz_si256(*this, bits); }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm256_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(_mm256_slli_epi16(*this, N)) & uint8_t(0xFFu << N); }
+    // Get one of the bits and make a bitmask out of it.
+    // e.g. value.get_bit<7>() gets the high bit
+    template<int N>
+    simdjson_inline int get_bit() const { return _mm256_movemask_epi8(_mm256_slli_epi16(*this, 7-N)); }
+  };
+
+  template<typename T>
+  struct simd8x64 {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+    static_assert(NUM_CHUNKS == 2, "Haswell kernel should use two registers per 64-byte block.");
+    const simd8<T> chunks[NUM_CHUNKS];
+    template<int idx> simd8<uint8_t> get() const { return idx < NUM_CHUNKS ? chunks[idx] : simd8<T>(); }
+
+    simd8x64(const simd8x64<T>& o) = delete; // no copy allowed
+    simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1) : chunks{chunk0, chunk1} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+32)} {}
+
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
+      uint32_t mask1 = uint32_t(mask);
+      uint32_t mask2 = uint32_t(mask >> 32);
+      this->chunks[0].compress(mask1, output);
+      this->chunks[1].compress(mask2, output + 32 - count_ones(mask1));
+      return 64 - count_ones(mask);
+    }
+
+    simdjson_inline void store(T ptr[64]) const {
+      this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
+      this->chunks[1].store(ptr+sizeof(simd8<T>)*1);
+    }
+
+    simdjson_inline uint64_t to_bitmask() const {
+      uint64_t r_lo = uint32_t(this->chunks[0].to_bitmask());
+      uint64_t r_hi =                       this->chunks[1].to_bitmask();
+      return r_lo | (r_hi << 32);
+    }
+
+    simdjson_inline simd8<T> reduce_or() const {
+      return this->chunks[0] | this->chunks[1];
+    }
+
+    simdjson_inline simd8x64<T> bit_or(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return simd8x64<T>(
+        this->chunks[0] | mask,
+        this->chunks[1] | mask
+      );
+    }
+
+    simdjson_inline uint64_t eq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] == mask,
+        this->chunks[1] == mask
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+      return  simd8x64<bool>(
+        this->chunks[0] == other.chunks[0],
+        this->chunks[1] == other.chunks[1]
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t lteq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] <= mask,
+        this->chunks[1] <= mask
+      ).to_bitmask();
+    }
+  }; // struct simd8x64<T>
+
+} // namespace simd
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_HASWELL_SIMD_H
+/* end file simdjson/haswell/simd.h */
+/* including simdjson/haswell/stringparsing_defs.h: #include "simdjson/haswell/stringparsing_defs.h" */
+/* begin file simdjson/haswell/stringparsing_defs.h */
+#ifndef SIMDJSON_HASWELL_STRINGPARSING_DEFS_H
+#define SIMDJSON_HASWELL_STRINGPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/simd.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 32;
+  simdjson_inline backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return ((quote_bits - 1) & bs_bits) != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+
+  uint32_t bs_bits;
+  uint32_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // this can read up to 15 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  // store to dest unconditionally - we can overwrite the bits we don't like later
+  v.store(dst);
+  return {
+      static_cast<uint32_t>((v == '\\').to_bitmask()),     // bs_bits
+      static_cast<uint32_t>((v == '"').to_bitmask()), // quote_bits
+  };
+}
+
+
+struct escaping {
+  static constexpr uint32_t BYTES_PROCESSED = 32;
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits != 0; }
+  simdjson_inline int escape_index() { return trailing_zeroes(escape_bits); }
+
+  uint64_t escape_bits;
+}; // struct escaping
+
+
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "escaping finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  simd8<bool> is_quote = (v == '"');
+  simd8<bool> is_backslash = (v == '\\');
+  simd8<bool> is_control = (v < 32);
+  return {
+    uint64_t((is_backslash | is_quote | is_control).to_bitmask())
+  };
+}
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_HASWELL_STRINGPARSING_DEFS_H
+/* end file simdjson/haswell/stringparsing_defs.h */
+/* end file simdjson/haswell/begin.h */
+/* including generic/amalgamated.h for haswell: #include <generic/amalgamated.h> */
+/* begin file generic/amalgamated.h for haswell */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_SRC_GENERIC_DEPENDENCIES_H)
+#error generic/dependencies.h must be included before generic/amalgamated.h!
+#endif
+
+/* including generic/base.h for haswell: #include <generic/base.h> */
+/* begin file generic/base.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+
+struct json_character_block;
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_BASE_H
+/* end file generic/base.h for haswell */
+/* including generic/dom_parser_implementation.h for haswell: #include <generic/dom_parser_implementation.h> */
+/* begin file generic/dom_parser_implementation.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// Interface a dom parser implementation must fulfill
+namespace simdjson {
+namespace haswell {
+namespace {
+
+simdjson_inline simd8<uint8_t> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3);
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input);
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file generic/dom_parser_implementation.h for haswell */
+/* including generic/json_character_block.h for haswell: #include <generic/json_character_block.h> */
+/* begin file generic/json_character_block.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+
+struct json_character_block {
+  static simdjson_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
+
+  simdjson_inline uint64_t whitespace() const noexcept { return _whitespace; }
+  simdjson_inline uint64_t op() const noexcept { return _op; }
+  simdjson_inline uint64_t scalar() const noexcept { return ~(op() | whitespace()); }
+
+  uint64_t _whitespace;
+  uint64_t _op;
+};
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H
+/* end file generic/json_character_block.h for haswell */
+/* end file generic/amalgamated.h for haswell */
+/* including generic/stage1/amalgamated.h for haswell: #include <generic/stage1/amalgamated.h> */
+/* begin file generic/stage1/amalgamated.h for haswell */
+// Stuff other things depend on
+/* including generic/stage1/base.h for haswell: #include <generic/stage1/base.h> */
+/* begin file generic/stage1/base.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_BASE_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace stage1 {
+
+class bit_indexer;
+template<size_t STEP_SIZE>
+struct buf_block_reader;
+struct json_block;
+class json_minifier;
+class json_scanner;
+struct json_string_block;
+class json_string_scanner;
+class json_structural_indexer;
+
+} // namespace stage1
+
+namespace utf8_validation {
+struct utf8_checker;
+} // namespace utf8_validation
+
+using utf8_validation::utf8_checker;
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_BASE_H
+/* end file generic/stage1/base.h for haswell */
+/* including generic/stage1/buf_block_reader.h for haswell: #include <generic/stage1/buf_block_reader.h> */
+/* begin file generic/stage1/buf_block_reader.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace stage1 {
+
+// Walks through a buffer in block-sized increments, loading the last part with spaces
+template<size_t STEP_SIZE>
+struct buf_block_reader {
+public:
+  simdjson_inline buf_block_reader(const uint8_t *_buf, size_t _len);
+  simdjson_inline size_t block_index();
+  simdjson_inline bool has_full_block() const;
+  simdjson_inline const uint8_t *full_block() const;
+  /**
+   * Get the last block, padded with spaces.
+   *
+   * There will always be a last block, with at least 1 byte, unless len == 0 (in which case this
+   * function fills the buffer with spaces and returns 0. In particular, if len == STEP_SIZE there
+   * will be 0 full_blocks and 1 remainder block with STEP_SIZE bytes and no spaces for padding.
+   *
+   * @return the number of effective characters in the last block.
+   */
+  simdjson_inline size_t get_remainder(uint8_t *dst) const;
+  simdjson_inline void advance();
+private:
+  const uint8_t *buf;
+  const size_t len;
+  const size_t lenminusstep;
+  size_t idx;
+};
+
+// Routines to print masks and text for debugging bitmask operations
+simdjson_unused static char * format_input_text_64(const uint8_t *text) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    buf[i] = int8_t(text[i]) < ' ' ? '_' : int8_t(text[i]);
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+// Routines to print masks and text for debugging bitmask operations
+simdjson_unused static char * format_input_text(const simd8x64<uint8_t>& in) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  in.store(reinterpret_cast<uint8_t*>(buf));
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] < ' ') { buf[i] = '_'; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+simdjson_unused static char * format_input_text(const simd8x64<uint8_t>& in, uint64_t mask) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  in.store(reinterpret_cast<uint8_t*>(buf));
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] <= ' ') { buf[i] = '_'; }
+    if (!(mask & (size_t(1) << i))) { buf[i] = ' '; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+simdjson_unused static char * format_mask(uint64_t mask) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  for (size_t i=0; i<64; i++) {
+    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
+  }
+  buf[64] = '\0';
+  return buf;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
+
+template<size_t STEP_SIZE>
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
+
+template<size_t STEP_SIZE>
+simdjson_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
+  return idx < lenminusstep;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
+  return &buf[idx];
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
+  if(len == idx) { return 0; } // memcpy(dst, null, 0) will trigger an error with some sanitizers
+  std::memset(dst, 0x20, STEP_SIZE); // std::memset STEP_SIZE because it's more efficient to write out 8 or 16 bytes at once.
+  std::memcpy(dst, buf + idx, len - idx);
+  return len - idx;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline void buf_block_reader<STEP_SIZE>::advance() {
+  idx += STEP_SIZE;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H
+/* end file generic/stage1/buf_block_reader.h for haswell */
+/* including generic/stage1/json_escape_scanner.h for haswell: #include <generic/stage1/json_escape_scanner.h> */
+/* begin file generic/stage1/json_escape_scanner.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_ESCAPE_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_ESCAPE_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace stage1 {
+
+/**
+ * Scans for escape characters in JSON, taking care with multiple backslashes (\\n vs. \n).
+ */
+struct json_escape_scanner {
+  /** The actual escape characters (the backslashes themselves). */
+  uint64_t next_is_escaped = 0ULL;
+
+  struct escaped_and_escape {
+    /**
+     * Mask of escaped characters.
+     *
+     * ```
+     * \n \\n \\\n \\\\n \
+     * 0100100010100101000
+     *  n  \   \ n  \ \
+     * ```
+     */
+    uint64_t escaped;
+    /**
+     * Mask of escape characters.
+     *
+     * ```
+     * \n \\n \\\n \\\\n \
+     * 1001000101001010001
+     * \  \   \ \  \ \   \
+     * ```
+     */
+    uint64_t escape;
+  };
+
+  /**
+   * Get a mask of both escape and escaped characters (the characters following a backslash).
+   *
+   * @param potential_escape A mask of the character that can escape others (but could be
+   *        escaped itself). e.g. block.eq('\\')
+   */
+  simdjson_really_inline escaped_and_escape next(uint64_t backslash) noexcept {
+
+#if !SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+    if (!backslash) { return {next_escaped_without_backslashes(), 0}; }
+#endif
+
+    // |                                | Mask (shows characters instead of 1's) | Depth | Instructions        |
+    // |--------------------------------|----------------------------------------|-------|---------------------|
+    // | string                         | `\\n_\\\n___\\\n___\\\\___\\\\__\\\`   |       |                     |
+    // |                                | `    even   odd    even   odd   odd`   |       |                     |
+    // | potential_escape               | ` \  \\\    \\\    \\\\   \\\\  \\\`   | 1     | 1 (backslash & ~first_is_escaped)
+    // | escape_and_terminal_code       | ` \n \ \n   \ \n   \ \    \ \   \ \`   | 5     | 5 (next_escape_and_terminal_code())
+    // | escaped                        | `\    \ n    \ n    \ \    \ \   \ ` X | 6     | 7 (escape_and_terminal_code ^ (potential_escape | first_is_escaped))
+    // | escape                         | `    \ \    \ \    \ \    \ \   \ \`   | 6     | 8 (escape_and_terminal_code & backslash)
+    // | first_is_escaped               | `\                                 `   | 7 (*) | 9 (escape >> 63) ()
+    //                                                                               (*) this is not needed until the next iteration
+    uint64_t escape_and_terminal_code = next_escape_and_terminal_code(backslash & ~this->next_is_escaped);
+    uint64_t escaped = escape_and_terminal_code ^ (backslash | this->next_is_escaped);
+    uint64_t escape = escape_and_terminal_code & backslash;
+    this->next_is_escaped = escape >> 63;
+    return {escaped, escape};
+  }
+
+private:
+  static constexpr const uint64_t ODD_BITS = 0xAAAAAAAAAAAAAAAAULL;
+
+  simdjson_really_inline uint64_t next_escaped_without_backslashes() noexcept {
+    uint64_t escaped = this->next_is_escaped;
+    this->next_is_escaped = 0;
+    return escaped;
+  }
+
+  /**
+   * Returns a mask of the next escape characters (masking out escaped backslashes), along with
+   * any non-backslash escape codes.
+   *
+   * \n \\n \\\n \\\\n returns:
+   * \n \   \ \n \ \
+   * 11 100 1011 10100
+   *
+   * You are expected to mask out the first bit yourself if the previous block had a trailing
+   * escape.
+   *
+   * & the result with potential_escape to get just the escape characters.
+   * ^ the result with (potential_escape | first_is_escaped) to get escaped characters.
+   */
+  static simdjson_really_inline uint64_t next_escape_and_terminal_code(uint64_t potential_escape) noexcept {
+    // If we were to just shift and mask out any odd bits, we'd actually get a *half* right answer:
+    // any even-aligned backslash runs would be correct! Odd-aligned backslash runs would be
+    // inverted (\\\ would be 010 instead of 101).
+    //
+    // ```
+    // string:              | ____\\\\_\\\\_____ |
+    // maybe_escaped | ODD  |     \ \   \ \      |
+    //               even-aligned ^^^  ^^^^ odd-aligned
+    // ```
+    //
+    // Taking that into account, our basic strategy is:
+    //
+    // 1. Use subtraction to produce a mask with 1's for even-aligned runs and 0's for
+    //    odd-aligned runs.
+    // 2. XOR all odd bits, which masks out the odd bits in even-aligned runs, and brings IN the
+    //    odd bits in odd-aligned runs.
+    // 3. & with backslash to clean up any stray bits.
+    // runs are set to 0, and then XORing with "odd":
+    //
+    // |                                | Mask (shows characters instead of 1's) | Instructions        |
+    // |--------------------------------|----------------------------------------|---------------------|
+    // | string                         | `\\n_\\\n___\\\n___\\\\___\\\\__\\\`   |
+    // |                                | `    even   odd    even   odd   odd`   |
+    // | maybe_escaped                  | `  n  \\n    \\n    \\\_   \\\_  \\` X | 1 (potential_escape << 1)
+    // | maybe_escaped_and_odd          | ` \n_ \\n _ \\\n_ _ \\\__ _\\\_ \\\`   | 1 (maybe_escaped | odd)
+    // | even_series_codes_and_odd      | `  n_\\\  _    n_ _\\\\ _     _    `   | 1 (maybe_escaped_and_odd - potential_escape)
+    // | escape_and_terminal_code       | ` \n \ \n   \ \n   \ \    \ \   \ \`   | 1 (^ odd)
+    //
+
+    // Escaped characters are characters following an escape.
+    uint64_t maybe_escaped = potential_escape << 1;
+
+    // To distinguish odd from even escape sequences, therefore, we turn on any *starting*
+    // escapes that are on an odd byte. (We actually bring in all odd bits, for speed.)
+    // - Odd runs of backslashes are 0000, and the code at the end ("n" in \n or \\n) is 1.
+    // - Odd runs of backslashes are 1111, and the code at the end ("n" in \n or \\n) is 0.
+    // - All other odd bytes are 1, and even bytes are 0.
+    uint64_t maybe_escaped_and_odd_bits     = maybe_escaped | ODD_BITS;
+    uint64_t even_series_codes_and_odd_bits = maybe_escaped_and_odd_bits - potential_escape;
+
+    // Now we flip all odd bytes back with xor. This:
+    // - Makes odd runs of backslashes go from 0000 to 1010
+    // - Makes even runs of backslashes go from 1111 to 1010
+    // - Sets actually-escaped codes to 1 (the n in \n and \\n: \n = 11, \\n = 100)
+    // - Resets all other bytes to 0
+    return even_series_codes_and_odd_bits ^ ODD_BITS;
+  }
+};
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+/* end file generic/stage1/json_escape_scanner.h for haswell */
+/* including generic/stage1/json_string_scanner.h for haswell: #include <generic/stage1/json_string_scanner.h> */
+/* begin file generic/stage1/json_string_scanner.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_escape_scanner.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace stage1 {
+
+struct json_string_block {
+  // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
+  simdjson_really_inline json_string_block(uint64_t escaped, uint64_t quote, uint64_t in_string) :
+  _escaped(escaped), _quote(quote), _in_string(in_string) {}
+
+  // Escaped characters (characters following an escape() character)
+  simdjson_really_inline uint64_t escaped() const { return _escaped; }
+  // Real (non-backslashed) quotes
+  simdjson_really_inline uint64_t quote() const { return _quote; }
+  // Only characters inside the string (not including the quotes)
+  simdjson_really_inline uint64_t string_content() const { return _in_string & ~_quote; }
+  // Return a mask of whether the given characters are inside a string (only works on non-quotes)
+  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
+  // Return a mask of whether the given characters are inside a string (only works on non-quotes)
+  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
+  // Tail of string (everything except the start quote)
+  simdjson_really_inline uint64_t string_tail() const { return _in_string ^ _quote; }
+
+  // escaped characters (backslashed--does not include the hex characters after \u)
+  uint64_t _escaped;
+  // real quotes (non-escaped ones)
+  uint64_t _quote;
+  // string characters (includes start quote but not end quote)
+  uint64_t _in_string;
+};
+
+// Scans blocks for string characters, storing the state necessary to do so
+class json_string_scanner {
+public:
+  simdjson_really_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
+  // Returns either UNCLOSED_STRING or SUCCESS
+  simdjson_really_inline error_code finish();
+
+private:
+  // Scans for escape characters
+  json_escape_scanner escape_scanner{};
+  // Whether the last iteration was still inside a string (all 1's = true, all 0's = false).
+  uint64_t prev_in_string = 0ULL;
+};
+
+//
+// Return a mask of all string characters plus end quotes.
+//
+// prev_escaped is overflow saying whether the next character is escaped.
+// prev_in_string is overflow saying whether we're still in a string.
+//
+// Backslash sequences outside of quotes will be detected in stage 2.
+//
+simdjson_really_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
+  const uint64_t backslash = in.eq('\\');
+  const uint64_t escaped = escape_scanner.next(backslash).escaped;
+  const uint64_t quote = in.eq('"') & ~escaped;
+
+  //
+  // prefix_xor flips on bits inside the string (and flips off the end quote).
+  //
+  // Then we xor with prev_in_string: if we were in a string already, its effect is flipped
+  // (characters inside strings are outside, and characters outside strings are inside).
+  //
+  const uint64_t in_string = prefix_xor(quote) ^ prev_in_string;
+
+  //
+  // Check if we're still in a string at the end of the box so the next block will know
+  //
+  prev_in_string = uint64_t(static_cast<int64_t>(in_string) >> 63);
+
+  // Use ^ to turn the beginning quote off, and the end quote on.
+
+  // We are returning a function-local object so either we get a move constructor
+  // or we get copy elision.
+  return json_string_block(escaped, quote, in_string);
+}
+
+simdjson_really_inline error_code json_string_scanner::finish() {
+  if (prev_in_string) {
+    return UNCLOSED_STRING;
+  }
+  return SUCCESS;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+/* end file generic/stage1/json_string_scanner.h for haswell */
+/* including generic/stage1/utf8_lookup4_algorithm.h for haswell: #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* begin file generic/stage1/utf8_lookup4_algorithm.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace utf8_validation {
+
+using namespace simd;
+
+  simdjson_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+// Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+// Bit 1 = Too Long (ASCII followed by continuation)
+// Bit 2 = Overlong 3-byte
+// Bit 4 = Surrogate
+// Bit 5 = Overlong 2-byte
+// Bit 7 = Two Continuations
+    constexpr const uint8_t TOO_SHORT   = 1<<0; // 11______ 0_______
+                                                // 11______ 11______
+    constexpr const uint8_t TOO_LONG    = 1<<1; // 0_______ 10______
+    constexpr const uint8_t OVERLONG_3  = 1<<2; // 11100000 100_____
+    constexpr const uint8_t SURROGATE   = 1<<4; // 11101101 101_____
+    constexpr const uint8_t OVERLONG_2  = 1<<5; // 1100000_ 10______
+    constexpr const uint8_t TWO_CONTS   = 1<<7; // 10______ 10______
+    constexpr const uint8_t TOO_LARGE   = 1<<3; // 11110100 1001____
+                                                // 11110100 101_____
+                                                // 11110101 1001____
+                                                // 11110101 101_____
+                                                // 1111011_ 1001____
+                                                // 1111011_ 101_____
+                                                // 11111___ 1001____
+                                                // 11111___ 101_____
+    constexpr const uint8_t TOO_LARGE_1000 = 1<<6;
+                                                // 11110101 1000____
+                                                // 1111011_ 1000____
+                                                // 11111___ 1000____
+    constexpr const uint8_t OVERLONG_4  = 1<<6; // 11110000 1000____
+
+    const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      TOO_SHORT,
+      // 1110____ ________ <three byte lead in byte 1>
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4
+    );
+    constexpr const uint8_t CARRY = TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+    const simd8<uint8_t> byte_1_low = (prev1 & 0x0F).lookup_16<uint8_t>(
+      // ____0000 ________
+      CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+      // ____0001 ________
+      CARRY | OVERLONG_2,
+      // ____001_ ________
+      CARRY,
+      CARRY,
+
+      // ____0100 ________
+      CARRY | TOO_LARGE,
+      // ____0101 ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      // ____011_ ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+
+      // ____1___ ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      // ____1101 ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000
+    );
+    const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE  | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE  | TOO_LARGE,
+
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT
+    );
+    return (byte_1_high & byte_1_low & byte_2_high);
+  }
+  simdjson_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
+      const simd8<uint8_t> prev_input, const simd8<uint8_t> sc) {
+    simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+    simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+    simd8<uint8_t> must23 = must_be_2_3_continuation(prev2, prev3);
+    simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+    return must23_80 ^ sc;
+  }
+
+  //
+  // Return nonzero if there are incomplete multibyte characters at the end of the block:
+  // e.g. if there is a 4-byte character, but it's 3 bytes from the end.
+  //
+  simdjson_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+    // If the previous input's last 3 bytes match this, they're too short (they ended at EOF):
+    // ... 1111____ 111_____ 11______
+#if SIMDJSON_IMPLEMENTATION_ICELAKE || (SIMDJSON_IMPLEMENTATION_RVV_VLS && __riscv_v_fixed_vlen >= 512)
+    static const uint8_t max_array[64] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
+    };
+#else
+    static const uint8_t max_array[32] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
+    };
+#endif
+    const simd8<uint8_t> max_value(&max_array[sizeof(max_array)-sizeof(simd8<uint8_t>)]);
+    return input.gt_bits(max_value);
+  }
+
+  struct utf8_checker {
+    // If this is nonzero, there has been a UTF-8 error.
+    simd8<uint8_t> error;
+    // The last input we received
+    simd8<uint8_t> prev_input_block;
+    // Whether the last input we received was incomplete (used for ASCII fast path)
+    simd8<uint8_t> prev_incomplete;
+
+    //
+    // Check whether the current bytes are valid UTF-8.
+    //
+    simdjson_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
+      // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ lead bytes
+      // (2, 3, 4-byte leads become large positive numbers instead of small negative numbers)
+      simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+      simd8<uint8_t> sc = check_special_cases(input, prev1);
+      this->error |= check_multibyte_lengths(input, prev_input, sc);
+    }
+
+    // The only problem that can happen at EOF is that a multibyte character is too short
+    // or a byte value too large in the last bytes: check_special_cases only checks for bytes
+    // too large in the first of two bytes.
+    simdjson_inline void check_eof() {
+      // If the previous block had incomplete UTF-8 characters at the end, an ASCII block can't
+      // possibly finish them.
+      this->error |= this->prev_incomplete;
+    }
+
+    simdjson_inline void check_next_input(const simd8x64<uint8_t>& input) {
+      if(simdjson_likely(is_ascii(input))) {
+        this->error |= this->prev_incomplete;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio, it is not good enough.
+        static_assert((simd8x64<uint8_t>::NUM_CHUNKS == 1)
+                ||(simd8x64<uint8_t>::NUM_CHUNKS == 2)
+                || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+                "We support one, two or four chunks per 64-byte block.");
+        SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 1) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+          this->check_utf8_bytes(input.get<1>(), input.get<0>());
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+          this->check_utf8_bytes(input.get<1>(), input.get<0>());
+          this->check_utf8_bytes(input.get<2>(), input.get<1>());
+          this->check_utf8_bytes(input.get<3>(), input.get<2>());
+        }
+        this->prev_incomplete = is_incomplete(input.get<simd8x64<uint8_t>::NUM_CHUNKS-1>());
+        this->prev_input_block = input.get<simd8x64<uint8_t>::NUM_CHUNKS-1>();
+      }
+    }
+    // do not forget to call check_eof!
+    simdjson_warn_unused simdjson_inline error_code errors() {
+      return this->error.any_bits_set_anywhere() ? error_code::UTF8_ERROR : error_code::SUCCESS;
+    }
+
+  }; // struct utf8_checker
+} // namespace utf8_validation
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H
+/* end file generic/stage1/utf8_lookup4_algorithm.h for haswell */
+/* including generic/stage1/json_scanner.h for haswell: #include <generic/stage1/json_scanner.h> */
+/* begin file generic/stage1/json_scanner.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/json_character_block.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_string_scanner.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace stage1 {
+
+/**
+ * A block of scanned json, with information on operators and scalars.
+ *
+ * We seek to identify pseudo-structural characters. Anything that is inside
+ * a string must be omitted (hence  & ~_string.string_tail()).
+ * Otherwise, pseudo-structural characters come in two forms.
+ * 1. We have the structural characters ([,],{,},:, comma). The
+ *    term 'structural character' is from the JSON RFC.
+ * 2. We have the 'scalar pseudo-structural characters'.
+ *    Scalars are quotes, and any character except structural characters and white space.
+ *
+ * To identify the scalar pseudo-structural characters, we must look at what comes
+ * before them: it must be a space, a quote or a structural characters.
+ * Starting with simdjson v0.3, we identify them by
+ * negation: we identify everything that is followed by a non-quote scalar,
+ * and we negate that. Whatever remains must be a 'scalar pseudo-structural character'.
+ */
+struct json_block {
+public:
+  // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
+  simdjson_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  _string(std::move(string)), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
+  simdjson_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  _string(string), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
+
+  /**
+   * The start of structurals.
+   * In simdjson prior to v0.3, these were called the pseudo-structural characters.
+   **/
+  simdjson_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
+  /** All JSON whitespace (i.e. not in a string) */
+  simdjson_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
+
+  // Helpers
+
+  /** Whether the given characters are inside a string (only works on non-quotes) */
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
+  /** Whether the given characters are outside a string (only works on non-quotes) */
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
+
+  // string and escape characters
+  json_string_block _string;
+  // whitespace, structural characters ('operators'), scalars
+  json_character_block _characters;
+  // whether the previous character was a scalar
+  uint64_t _follows_potential_nonquote_scalar;
+private:
+  // Potential structurals (i.e. disregarding strings)
+
+  /**
+   * structural elements ([,],{,},:, comma) plus scalar starts like 123, true and "abc".
+   * They may reside inside a string.
+   **/
+  simdjson_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
+  /**
+   * The start of non-operator runs, like 123, true and "abc".
+   * It main reside inside a string.
+   **/
+  simdjson_inline uint64_t potential_scalar_start() const noexcept {
+    // The term "scalar" refers to anything except structural characters and white space
+    // (so letters, numbers, quotes).
+    // Whenever it is preceded by something that is not a structural element ({,},[,],:, ") nor a white-space
+    // then we know that it is irrelevant structurally.
+    return _characters.scalar() & ~follows_potential_scalar();
+  }
+  /**
+   * Whether the given character is immediately after a non-operator like 123, true.
+   * The characters following a quote are not included.
+   */
+  simdjson_inline uint64_t follows_potential_scalar() const noexcept {
+    // _follows_potential_nonquote_scalar: is defined as marking any character that follows a character
+    // that is not a structural element ({,},[,],:, comma) nor a quote (") and that is not a
+    // white space.
+    // It is understood that within quoted region, anything at all could be marked (irrelevant).
+    return _follows_potential_nonquote_scalar;
+  }
+};
+
+/**
+ * Scans JSON for important bits: structural characters or 'operators', strings, and scalars.
+ *
+ * The scanner starts by calculating two distinct things:
+ * - string characters (taking \" into account)
+ * - structural characters or 'operators' ([]{},:, comma)
+ *   and scalars (runs of non-operators like 123, true and "abc")
+ *
+ * To minimize data dependency (a key component of the scanner's speed), it finds these in parallel:
+ * in particular, the operator/scalar bit will find plenty of things that are actually part of
+ * strings. When we're done, json_block will fuse the two together by masking out tokens that are
+ * part of a string.
+ */
+class json_scanner {
+public:
+  json_scanner() = default;
+  simdjson_inline json_block next(const simd::simd8x64<uint8_t>& in);
+  // Returns either UNCLOSED_STRING or SUCCESS
+  simdjson_warn_unused simdjson_inline error_code finish();
+
+private:
+  // Whether the last character of the previous iteration is part of a scalar token
+  // (anything except whitespace or a structural character/'operator').
+  uint64_t prev_scalar = 0ULL;
+  json_string_scanner string_scanner{};
+};
+
+
+//
+// Check if the current character immediately follows a matching character.
+//
+// For example, this checks for quotes with backslashes in front of them:
+//
+//     const uint64_t backslashed_quote = in.eq('"') & immediately_follows(in.eq('\'), prev_backslash);
+//
+simdjson_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
+  const uint64_t result = match << 1 | overflow;
+  overflow = match >> 63;
+  return result;
+}
+
+simdjson_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
+  json_string_block strings = string_scanner.next(in);
+  // identifies the white-space and the structural characters
+  json_character_block characters = json_character_block::classify(in);
+  // The term "scalar" refers to anything except structural characters and white space
+  // (so letters, numbers, quotes).
+  // We want follows_scalar to mark anything that follows a non-quote scalar (so letters and numbers).
+  //
+  // A terminal quote should either be followed by a structural character (comma, brace, bracket, colon)
+  // or nothing. However, we still want ' "a string"true ' to mark the 't' of 'true' as a potential
+  // pseudo-structural character just like we would if we had  ' "a string" true '; otherwise we
+  // may need to add an extra check when parsing strings.
+  //
+  // Performance: there are many ways to skin this cat.
+  const uint64_t nonquote_scalar = characters.scalar() & ~strings.quote();
+  uint64_t follows_nonquote_scalar = follows(nonquote_scalar, prev_scalar);
+  // We are returning a function-local object so either we get a move constructor
+  // or we get copy elision.
+  return json_block(
+    strings,// strings is a function-local object so either it moves or the copy is elided.
+    characters,
+    follows_nonquote_scalar
+  );
+}
+
+simdjson_warn_unused simdjson_inline error_code json_scanner::finish() {
+  return string_scanner.finish();
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H
+/* end file generic/stage1/json_scanner.h for haswell */
+
+// All other declarations
+/* including generic/stage1/find_next_document_index.h for haswell: #include <generic/stage1/find_next_document_index.h> */
+/* begin file generic/stage1/find_next_document_index.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace stage1 {
+
+/**
+  * This algorithm is used to quickly identify the last structural position that
+  * makes up a complete document.
+  *
+  * It does this by going backwards and finding the last *document boundary* (a
+  * place where one value follows another without a comma between them). If the
+  * last document (the characters after the boundary) has an equal number of
+  * start and end brackets, it is considered complete.
+  *
+  * Simply put, we iterate over the structural characters, starting from
+  * the end. We consider that we found the end of a JSON document when the
+  * first element of the pair is NOT one of these characters: '{' '[' ':' ','
+  * and when the second element is NOT one of these characters: '}' ']' ':' ','.
+  *
+  * This simple comparison works most of the time, but it does not cover cases
+  * where the batch's structural indexes contain a perfect amount of documents.
+  * In such a case, we do not have access to the structural index which follows
+  * the last document, therefore, we do not have access to the second element in
+  * the pair, and that means we cannot identify the last document. To fix this
+  * issue, we keep a count of the open and closed curly/square braces we found
+  * while searching for the pair. When we find a pair AND the count of open and
+  * closed curly/square braces is the same, we know that we just passed a
+  * complete document, therefore the last json buffer location is the end of the
+  * batch.
+  */
+simdjson_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
+  // Variant: do not count separately, just figure out depth
+  if(parser.n_structural_indexes == 0) { return 0; }
+  auto arr_cnt = 0;
+  auto obj_cnt = 0;
+  for (auto i = parser.n_structural_indexes - 1; i > 0; i--) {
+    auto idxb = parser.structural_indexes[i];
+    switch (parser.buf[idxb]) {
+    case ':':
+    case ',':
+      continue;
+    case '}':
+      obj_cnt--;
+      continue;
+    case ']':
+      arr_cnt--;
+      continue;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+    }
+    auto idxa = parser.structural_indexes[i - 1];
+    switch (parser.buf[idxa]) {
+    case '{':
+    case '[':
+    case ':':
+    case ',':
+      continue;
+    }
+    // Last document is complete, so the next document will appear after!
+    if (!arr_cnt && !obj_cnt) {
+      return parser.n_structural_indexes;
+    }
+    // Last document is incomplete; mark the document at i + 1 as the next one
+    return i;
+  }
+  // If we made it to the end, we want to finish counting to see if we have a full document.
+  switch (parser.buf[parser.structural_indexes[0]]) {
+    case '}':
+      obj_cnt--;
+      break;
+    case ']':
+      arr_cnt--;
+      break;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+  }
+  if (!arr_cnt && !obj_cnt) {
+    // We have a complete document.
+    return parser.n_structural_indexes;
+  }
+  return 0;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+/* end file generic/stage1/find_next_document_index.h for haswell */
+/* including generic/stage1/json_minifier.h for haswell: #include <generic/stage1/json_minifier.h> */
+/* begin file generic/stage1/json_minifier.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses in stage1
+// It is intended to be included multiple times and compiled multiple times
+// We assume the file in which it is included already includes
+// "simdjson/stage1.h" (this simplifies amalgation)
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace stage1 {
+
+class json_minifier {
+public:
+  template<size_t STEP_SIZE>
+  static error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept;
+
+private:
+  simdjson_inline json_minifier(uint8_t *_dst)
+  : dst{_dst}
+  {}
+  template<size_t STEP_SIZE>
+  simdjson_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
+  simdjson_warn_unused simdjson_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
+  json_scanner scanner{};
+  uint8_t *dst;
+};
+
+simdjson_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
+  uint64_t mask = block.whitespace();
+  dst += in.compress(mask, dst);
+}
+
+simdjson_warn_unused simdjson_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
+  error_code error = scanner.finish();
+  if (error) { dst_len = 0; return error; }
+  dst_len = dst - dst_start;
+  return SUCCESS;
+}
+
+template<>
+simdjson_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block_buf);
+  simd::simd8x64<uint8_t> in_2(block_buf+64);
+  json_block block_1 = scanner.next(in_1);
+  json_block block_2 = scanner.next(in_2);
+  this->next(in_1, block_1);
+  this->next(in_2, block_2);
+  reader.advance();
+}
+
+template<>
+simdjson_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block_buf);
+  json_block block_1 = scanner.next(in_1);
+  this->next(block_buf, block_1);
+  reader.advance();
+}
+
+template<size_t STEP_SIZE>
+error_code json_minifier::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept {
+  buf_block_reader<STEP_SIZE> reader(buf, len);
+  json_minifier minifier(dst);
+
+  // Index the first n-1 blocks
+  while (reader.has_full_block()) {
+    minifier.step<STEP_SIZE>(reader.full_block(), reader);
+  }
+
+  // Index the last (remainder) block, padded with spaces
+  uint8_t block[STEP_SIZE];
+  size_t remaining_bytes = reader.get_remainder(block);
+  if (remaining_bytes > 0) {
+    // We do not want to write directly to the output stream. Rather, we write
+    // to a local buffer (for safety).
+    uint8_t out_block[STEP_SIZE];
+    uint8_t * const guarded_dst{minifier.dst};
+    minifier.dst = out_block;
+    minifier.step<STEP_SIZE>(block, reader);
+    size_t to_write = minifier.dst - out_block;
+    // In some cases, we could be enticed to consider the padded spaces
+    // as part of the string. This is fine as long as we do not write more
+    // than we consumed.
+    if(to_write > remaining_bytes) { to_write = remaining_bytes; }
+    memcpy(guarded_dst, out_block, to_write);
+    minifier.dst = guarded_dst + to_write;
+  }
+  return minifier.finish(dst, dst_len);
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H
+/* end file generic/stage1/json_minifier.h for haswell */
+/* including generic/stage1/json_structural_indexer.h for haswell: #include <generic/stage1/json_structural_indexer.h> */
+/* begin file generic/stage1/json_structural_indexer.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_string_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_minifier.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/find_next_document_index.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses in stage1
+// It is intended to be included multiple times and compiled multiple times
+// We assume the file in which it is included already includes
+// "simdjson/stage1.h" (this simplifies amalgation)
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace stage1 {
+
+class bit_indexer {
+public:
+  uint32_t *tail;
+
+  simdjson_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
+
+#if SIMDJSON_PREFER_REVERSE_BITS
+  /**
+    * ARM lacks a fast trailing zero instruction, but it has a fast
+    * bit reversal instruction and a fast leading zero instruction.
+    * Thus it may be profitable to reverse the bits (once) and then
+    * to rely on a sequence of instructions that call the leading
+    * zero instruction.
+    *
+    * Performance notes:
+    * The chosen routine is not optimal in terms of data dependency
+    * since zero_leading_bit might require two instructions. However,
+    * it tends to minimize the total number of instructions which is
+    * beneficial.
+    */
+  simdjson_inline void write_index(uint32_t idx, uint64_t& rev_bits, int i) {
+    int lz = leading_zeroes(rev_bits);
+    this->tail[i] = static_cast<uint32_t>(idx) + lz;
+    rev_bits = zero_leading_bit(rev_bits, lz);
+  }
+#else
+  /**
+    * Under recent x64 systems, we often have both a fast trailing zero
+    * instruction and a fast 'clear-lower-bit' instruction so the following
+    * algorithm can be competitive.
+    */
+
+  simdjson_inline void write_index(uint32_t idx, uint64_t& bits, int i) {
+    this->tail[i] = idx + trailing_zeroes(bits);
+    bits = clear_lowest_bit(bits);
+  }
+#endif // SIMDJSON_PREFER_REVERSE_BITS
+
+  template <int START, int N>
+  simdjson_inline int write_indexes(uint32_t idx, uint64_t& bits) {
+    write_index(idx, bits, START);
+    SIMDJSON_IF_CONSTEXPR (N > 1) {
+      write_indexes<(N-1>0?START+1:START), (N-1>=0?N-1:1)>(idx, bits);
+    }
+    return START+N;
+  }
+
+  template <int START, int END, int STEP>
+  simdjson_inline int write_indexes_stepped(uint32_t idx, uint64_t& bits, int cnt) {
+    write_indexes<START, STEP>(idx, bits);
+    SIMDJSON_IF_CONSTEXPR ((START+STEP)  < END) {
+      if (simdjson_unlikely((START+STEP) < cnt)) {
+        write_indexes_stepped<(START+STEP<END?START+STEP:END), END, STEP>(idx, bits, cnt);
+      }
+    }
+    return ((END-START) % STEP) == 0 ? END : (END-START) - ((END-START) % STEP) + STEP;
+  }
+
+  // flatten out values in 'bits' assuming that they are are to have values of idx
+  // plus their position in the bitvector, and store these indexes at
+  // base_ptr[base] incrementing base as we go
+  // will potentially store extra values beyond end of valid bits, so base_ptr
+  // needs to be large enough to handle this
+  //
+  // If the kernel sets SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER, then it
+  // will provide its own version of the code.
+#ifdef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+  simdjson_inline void write(uint32_t idx, uint64_t bits);
+#else
+  simdjson_inline void write(uint32_t idx, uint64_t bits) {
+    // In some instances, the next branch is expensive because it is mispredicted.
+    // Unfortunately, in other cases,
+    // it helps tremendously.
+    if (bits == 0)
+        return;
+
+    int cnt = static_cast<int>(count_ones(bits));
+
+#if SIMDJSON_PREFER_REVERSE_BITS
+    bits = reverse_bits(bits);
+#endif
+#ifdef SIMDJSON_STRUCTURAL_INDEXER_STEP
+    static constexpr const int STEP = SIMDJSON_STRUCTURAL_INDEXER_STEP;
+#else
+    static constexpr const int STEP = 4;
+#endif
+    static constexpr const int STEP_UNTIL = 24;
+
+    write_indexes_stepped<0, STEP_UNTIL, STEP>(idx, bits, cnt);
+    SIMDJSON_IF_CONSTEXPR (STEP_UNTIL < 64) {
+      if (simdjson_unlikely(STEP_UNTIL < cnt)) {
+        for (int i=STEP_UNTIL; i<cnt; i++) {
+          write_index(idx, bits, i);
+        }
+      }
+    }
+
+    this->tail += cnt;
+  }
+#endif // SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+};
+
+class json_structural_indexer {
+public:
+  /**
+   * Find the important bits of JSON in a 128-byte chunk, and add them to structural_indexes.
+   *
+   * @param partial Setting the partial parameter to true allows the find_structural_bits to
+   *   tolerate unclosed strings. The caller should still ensure that the input is valid UTF-8. If
+   *   you are processing substrings, you may want to call on a function like trimmed_length_safe_utf8.
+   */
+  template<size_t STEP_SIZE>
+  static error_code index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept;
+
+private:
+  simdjson_inline json_structural_indexer(uint32_t *structural_indexes);
+  template<size_t STEP_SIZE>
+  simdjson_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
+  simdjson_warn_unused simdjson_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
+
+  json_scanner scanner{};
+  utf8_checker checker{};
+  bit_indexer indexer;
+  uint64_t prev_structurals = 0;
+  uint64_t unescaped_chars_error = 0;
+};
+
+simdjson_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
+
+// Skip the last character if it is partial
+simdjson_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
+  if (simdjson_unlikely(len < 3)) {
+    switch (len) {
+      case 2:
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
+        return len;
+      case 1:
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        return len;
+      case 0:
+        return len;
+    }
+  }
+  if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+  if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 1 byte left
+  if (buf[len-3] >= 0xf0) { return len-3; } // 4-byte characters with only 3 bytes left
+  return len;
+}
+
+//
+// PERF NOTES:
+// We pipe 2 inputs through these stages:
+// 1. Load JSON into registers. This takes a long time and is highly parallelizable, so we load
+//    2 inputs' worth at once so that by the time step 2 is looking for them input, it's available.
+// 2. Scan the JSON for critical data: strings, scalars and operators. This is the critical path.
+//    The output of step 1 depends entirely on this information. These functions don't quite use
+//    up enough CPU: the second half of the functions is highly serial, only using 1 execution core
+//    at a time. The second input's scans has some dependency on the first ones finishing it, but
+//    they can make a lot of progress before they need that information.
+// 3. Step 1 does not use enough capacity, so we run some extra stuff while we're waiting for that
+//    to finish: utf-8 checks and generating the output from the last iteration.
+//
+// The reason we run 2 inputs at a time, is steps 2 and 3 are *still* not enough to soak up all
+// available capacity with just one input. Running 2 at a time seems to give the CPU a good enough
+// workout.
+//
+template<size_t STEP_SIZE>
+error_code json_structural_indexer::index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept {
+  if (simdjson_unlikely(len > parser.capacity())) { return CAPACITY; }
+  // We guard the rest of the code so that we can assume that len > 0 throughout.
+  if (len == 0) { return EMPTY; }
+  if (is_streaming(partial)) {
+    len = trim_partial_utf8(buf, len);
+    // If you end up with an empty window after trimming
+    // the partial UTF-8 bytes, then chances are good that you
+    // have an UTF-8 formatting error.
+    if(len == 0) { return UTF8_ERROR; }
+  }
+  buf_block_reader<STEP_SIZE> reader(buf, len);
+  json_structural_indexer indexer(parser.structural_indexes.get());
+
+  // Read all but the last block
+  while (reader.has_full_block()) {
+    indexer.step<STEP_SIZE>(reader.full_block(), reader);
+  }
+  // Take care of the last block (will always be there unless file is empty which is
+  // not supposed to happen.)
+  uint8_t block[STEP_SIZE];
+  if (simdjson_unlikely(reader.get_remainder(block) == 0)) { return UNEXPECTED_ERROR; }
+  indexer.step<STEP_SIZE>(block, reader);
+  return indexer.finish(parser, reader.block_index(), len, partial);
+}
+
+template<>
+simdjson_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block);
+  simd::simd8x64<uint8_t> in_2(block+64);
+  json_block block_1 = scanner.next(in_1);
+  json_block block_2 = scanner.next(in_2);
+  this->next(in_1, block_1, reader.block_index());
+  this->next(in_2, block_2, reader.block_index()+64);
+  reader.advance();
+}
+
+template<>
+simdjson_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block);
+  json_block block_1 = scanner.next(in_1);
+  this->next(in_1, block_1, reader.block_index());
+  reader.advance();
+}
+
+simdjson_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
+  uint64_t unescaped = in.lteq(0x1F);
+#if SIMDJSON_UTF8VALIDATION
+  checker.check_next_input(in);
+#endif
+  indexer.write(uint32_t(idx-64), prev_structurals); // Output *last* iteration's structurals to the parser
+  prev_structurals = block.structural_start();
+  unescaped_chars_error |= block.non_quote_inside_string(unescaped);
+}
+
+simdjson_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
+  // Write out the final iteration's structurals
+  indexer.write(uint32_t(idx-64), prev_structurals);
+  error_code error = scanner.finish();
+  // We deliberately break down the next expression so that it is
+  // human readable.
+  const bool should_we_exit = is_streaming(partial) ?
+    ((error != SUCCESS) && (error != UNCLOSED_STRING)) // when partial we tolerate UNCLOSED_STRING
+    : (error != SUCCESS); // if partial is false, we must have SUCCESS
+  const bool have_unclosed_string = (error == UNCLOSED_STRING);
+  if (simdjson_unlikely(should_we_exit)) { return error; }
+
+  if (unescaped_chars_error) {
+    return UNESCAPED_CHARS;
+  }
+  parser.n_structural_indexes = uint32_t(indexer.tail - parser.structural_indexes.get());
+  /***
+   * The On-Demand API requires special padding.
+   *
+   * This is related to https://github.com/simdjson/simdjson/issues/906
+   * Basically, we want to make sure that if the parsing continues beyond the last (valid)
+   * structural character, it quickly stops.
+   * Only three structural characters can be repeated without triggering an error in JSON:  [,] and }.
+   * We repeat the padding character (at 'len'). We don't know what it is, but if the parsing
+   * continues, then it must be [,] or }.
+   * Suppose it is ] or }. We backtrack to the first character, what could it be that would
+   * not trigger an error? It could be ] or } but no, because you can't start a document that way.
+   * It can't be a comma, a colon or any simple value. So the only way we could continue is
+   * if the repeated character is [. But if so, the document must start with [. But if the document
+   * starts with [, it should end with ]. If we enforce that rule, then we would get
+   * ][[ which is invalid.
+   *
+   * This is illustrated with the test array_iterate_unclosed_error() on the following input:
+   * R"({ "a": [,,)"
+   **/
+  parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len); // used later in partial == stage1_mode::streaming_final
+  parser.structural_indexes[parser.n_structural_indexes + 1] = uint32_t(len);
+  parser.structural_indexes[parser.n_structural_indexes + 2] = 0;
+  parser.next_structural_index = 0;
+  // a valid JSON file cannot have zero structural indexes - we should have found something
+  if (simdjson_unlikely(parser.n_structural_indexes == 0u)) {
+    return EMPTY;
+  }
+  if (simdjson_unlikely(parser.structural_indexes[parser.n_structural_indexes - 1] > len)) {
+    return UNEXPECTED_ERROR;
+  }
+  if (partial == stage1_mode::streaming_partial) {
+    // If we have an unclosed string, then the last structural
+    // will be the quote and we want to make sure to omit it.
+    if(have_unclosed_string) {
+      parser.n_structural_indexes--;
+      // a valid JSON file cannot have zero structural indexes - we should have found something
+      if (simdjson_unlikely(parser.n_structural_indexes == 0u)) { return CAPACITY; }
+    }
+    // We truncate the input to the end of the last complete document (or zero).
+    auto new_structural_indexes = find_next_document_index(parser);
+    if (new_structural_indexes == 0 && parser.n_structural_indexes > 0) {
+      if(parser.structural_indexes[0] == 0) {
+        // If the buffer is partial and we started at index 0 but the document is
+        // incomplete, it's too big to parse.
+        return CAPACITY;
+      } else {
+        // It is possible that the document could be parsed, we just had a lot
+        // of white space.
+        parser.n_structural_indexes = 0;
+        return EMPTY;
+      }
+    }
+
+    parser.n_structural_indexes = new_structural_indexes;
+  } else if (partial == stage1_mode::streaming_final) {
+    if(have_unclosed_string) { parser.n_structural_indexes--; }
+    // We truncate the input to the end of the last complete document (or zero).
+    // Because partial == stage1_mode::streaming_final, it means that we may
+    // silently ignore trailing garbage. Though it sounds bad, we do it
+    // deliberately because many people who have streams of JSON documents
+    // will truncate them for processing. E.g., imagine that you are uncompressing
+    // the data from a size file or receiving it in chunks from the network. You
+    // may not know where exactly the last document will be. Meanwhile the
+    // document_stream instances allow people to know the JSON documents they are
+    // parsing (see the iterator.source() method).
+    parser.n_structural_indexes = find_next_document_index(parser);
+    // We store the initial n_structural_indexes so that the client can see
+    // whether we used truncation. If initial_n_structural_indexes == parser.n_structural_indexes,
+    // then this will query parser.structural_indexes[parser.n_structural_indexes] which is len,
+    // otherwise, it will copy some prior index.
+    parser.structural_indexes[parser.n_structural_indexes + 1] = parser.structural_indexes[parser.n_structural_indexes];
+    // This next line is critical, do not change it unless you understand what you are
+    // doing.
+    parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len);
+    if (simdjson_unlikely(parser.n_structural_indexes == 0u)) {
+        // We tolerate an unclosed string at the very end of the stream. Indeed, users
+        // often load their data in bulk without being careful and they want us to ignore
+        // the trailing garbage.
+        return EMPTY;
+    }
+  }
+  checker.check_eof();
+  return checker.errors();
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+// Clear CUSTOM_BIT_INDEXER so other implementations can set it if they need to.
+#undef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H
+/* end file generic/stage1/json_structural_indexer.h for haswell */
+/* including generic/stage1/utf8_validator.h for haswell: #include <generic/stage1/utf8_validator.h> */
+/* begin file generic/stage1/utf8_validator.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace stage1 {
+
+/**
+ * Validates that the string is actual UTF-8.
+ */
+template<class checker>
+bool generic_validate_utf8(const uint8_t * input, size_t length) {
+    checker c{};
+    buf_block_reader<64> reader(input, length);
+    while (reader.has_full_block()) {
+      simd::simd8x64<uint8_t> in(reader.full_block());
+      c.check_next_input(in);
+      reader.advance();
+    }
+    uint8_t block[64]{};
+    reader.get_remainder(block);
+    simd::simd8x64<uint8_t> in(block);
+    c.check_next_input(in);
+    reader.advance();
+    c.check_eof();
+    return c.errors() == error_code::SUCCESS;
+}
+
+bool generic_validate_utf8(const char * input, size_t length) {
+    return generic_validate_utf8<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H
+/* end file generic/stage1/utf8_validator.h for haswell */
+/* end file generic/stage1/amalgamated.h for haswell */
+/* including generic/stage2/amalgamated.h for haswell: #include <generic/stage2/amalgamated.h> */
+/* begin file generic/stage2/amalgamated.h for haswell */
+// Stuff other things depend on
+/* including generic/stage2/base.h for haswell: #include <generic/stage2/base.h> */
+/* begin file generic/stage2/base.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_BASE_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace stage2 {
+
+class json_iterator;
+class structural_iterator;
+struct tape_builder;
+struct tape_writer;
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_BASE_H
+/* end file generic/stage2/base.h for haswell */
+/* including generic/stage2/tape_writer.h for haswell: #include <generic/stage2/tape_writer.h> */
+/* begin file generic/stage2/tape_writer.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/internal/tape_type.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace stage2 {
+
+struct tape_writer {
+  /** The next place to write to tape */
+  uint64_t *next_tape_loc;
+
+  /** Write a signed 64-bit value to tape. */
+  simdjson_inline void append_s64(int64_t value) noexcept;
+
+  /** Write an unsigned 64-bit value to tape. */
+  simdjson_inline void append_u64(uint64_t value) noexcept;
+
+  /** Write a double value to tape. */
+  simdjson_inline void append_double(double value) noexcept;
+
+  /** Write a big integer (as string) to tape. src points to first digit, len is byte count. */
+  simdjson_inline void append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept;
+
+  /**
+   * Append a tape entry (an 8-bit type,and 56 bits worth of value).
+   */
+  simdjson_inline void append(uint64_t val, internal::tape_type t) noexcept;
+
+  /**
+   * Skip the current tape entry without writing.
+   *
+   * Used to skip the start of the container, since we'll come back later to fill it in when the
+   * container ends.
+   */
+  simdjson_inline void skip() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a large u64 or i64.
+   */
+  simdjson_inline void skip_large_integer() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a double.
+   */
+  simdjson_inline void skip_double() noexcept;
+
+  /**
+   * Write a value to a known location on tape.
+   *
+   * Used to go back and write out the start of a container after the container ends.
+   */
+  simdjson_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
+
+private:
+  /**
+   * Append both the tape entry, and a supplementary value following it. Used for types that need
+   * all 64 bits, such as double and uint64_t.
+   */
+  template<typename T>
+  simdjson_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
+}; // struct tape_writer
+
+simdjson_inline void tape_writer::append_s64(int64_t value) noexcept {
+  append2(0, value, internal::tape_type::INT64);
+}
+
+simdjson_inline void tape_writer::append_u64(uint64_t value) noexcept {
+  append(0, internal::tape_type::UINT64);
+  *next_tape_loc = value;
+  next_tape_loc++;
+}
+
+/** Write a double value to tape. */
+simdjson_inline void tape_writer::append_double(double value) noexcept {
+  append2(0, value, internal::tape_type::DOUBLE);
+}
+
+simdjson_inline void tape_writer::skip() noexcept {
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::skip_large_integer() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::skip_double() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
+  *next_tape_loc = val | ((uint64_t(char(t))) << 56);
+  next_tape_loc++;
+}
+
+template<typename T>
+simdjson_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
+  append(val, t);
+  static_assert(sizeof(val2) == sizeof(*next_tape_loc), "Type is not 64 bits!");
+  memcpy(next_tape_loc, &val2, sizeof(val2));
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
+  tape_loc = val | ((uint64_t(char(t))) << 56);
+}
+
+simdjson_inline void tape_writer::append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept {
+  // Write to string buffer: [4-byte LE length][digits][null]
+  uint32_t str_len = uint32_t(len);
+  memcpy(string_buf, &str_len, sizeof(uint32_t));
+  memcpy(string_buf + sizeof(uint32_t), src, len);
+  string_buf[sizeof(uint32_t) + len] = 0;
+  // Tape entry: offset into string buffer
+  // The caller must set the offset relative to doc.string_buf base
+  append(0, internal::tape_type::BIGINT); // placeholder offset, caller patches
+  string_buf += sizeof(uint32_t) + len + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+/* end file generic/stage2/tape_writer.h for haswell */
+/* including generic/stage2/logger.h for haswell: #include <generic/stage2/logger.h> */
+/* begin file generic/stage2/logger.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+
+// This is for an internal-only stage 2 specific logger.
+// Set LOG_ENABLED = true to log what stage 2 is doing!
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace logger {
+
+  static constexpr const char * DASHES = "----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------";
+
+#if SIMDJSON_VERBOSE_LOGGING
+  static constexpr const bool LOG_ENABLED = true;
+#else
+  static constexpr const bool LOG_ENABLED = false;
+#endif
+  static constexpr const int LOG_EVENT_LEN = 20;
+  static constexpr const int LOG_BUFFER_LEN = 30;
+  static constexpr const int LOG_SMALL_BUFFER_LEN = 10;
+  static constexpr const int LOG_INDEX_LEN = 5;
+
+  static int log_depth; // Not threadsafe. Log only.
+
+  // Helper to turn unprintable or newline characters into spaces
+  static simdjson_inline char printable_char(char c) {
+    if (c >= 0x20) {
+      return c;
+    } else {
+      return ' ';
+    }
+  }
+
+  // Print the header and set up log_start
+  static simdjson_inline void log_start() {
+    if (LOG_ENABLED) {
+      log_depth = 0;
+      printf("\n");
+      printf("| %-*s | %-*s | %-*s | %-*s | Detail |\n", LOG_EVENT_LEN, "Event", LOG_BUFFER_LEN, "Buffer", LOG_SMALL_BUFFER_LEN, "Next", 5, "Next#");
+      printf("|%.*s|%.*s|%.*s|%.*s|--------|\n", LOG_EVENT_LEN+2, DASHES, LOG_BUFFER_LEN+2, DASHES, LOG_SMALL_BUFFER_LEN+2, DASHES, 5+2, DASHES);
+    }
+  }
+
+  simdjson_unused static simdjson_inline void log_string(const char *message) {
+    if (LOG_ENABLED) {
+      printf("%s\n", message);
+    }
+  }
+
+  // Logs a single line from the stage 2 DOM parser
+  template<typename S>
+  static simdjson_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
+    if (LOG_ENABLED) {
+      printf("| %*s%s%-*s ", log_depth*2, "", title_prefix, LOG_EVENT_LEN - log_depth*2 - int(strlen(title_prefix)), title);
+      auto current_index = structurals.at_beginning() ? nullptr : structurals.next_structural-1;
+      auto next_index = structurals.next_structural;
+      auto current = current_index ? &structurals.buf[*current_index] : reinterpret_cast<const uint8_t*>("                                                       ");
+      auto next = &structurals.buf[*next_index];
+      {
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_BUFFER_LEN;i++) {
+          printf("%c", printable_char(current[i]));
+        }
+        printf(" ");
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_SMALL_BUFFER_LEN;i++) {
+          printf("%c", printable_char(next[i]));
+        }
+        printf(" ");
+      }
+      if (current_index) {
+        printf("| %*u ", LOG_INDEX_LEN, *current_index);
+      } else {
+        printf("| %-*s ", LOG_INDEX_LEN, "");
+      }
+      // printf("| %*u ", LOG_INDEX_LEN, structurals.next_tape_index());
+      printf("| %-s ", detail);
+      printf("|\n");
+    }
+  }
+
+} // namespace logger
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+/* end file generic/stage2/logger.h for haswell */
+
+// All other declarations
+/* including generic/stage2/json_iterator.h for haswell: #include <generic/stage2/json_iterator.h> */
+/* begin file generic/stage2/json_iterator.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/logger.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace stage2 {
+
+class json_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+  uint32_t depth{0};
+
+  /**
+   * Walk the JSON document.
+   *
+   * The visitor receives callbacks when values are encountered. All callbacks pass the iterator as
+   * the first parameter; some callbacks have other parameters as well:
+   *
+   * - visit_document_start() - at the beginning.
+   * - visit_document_end() - at the end (if things were successful).
+   *
+   * - visit_array_start() - at the start `[` of a non-empty array.
+   * - visit_array_end() - at the end `]` of a non-empty array.
+   * - visit_empty_array() - when an empty array is encountered.
+   *
+   * - visit_object_end() - at the start `]` of a non-empty object.
+   * - visit_object_start() - at the end `]` of a non-empty object.
+   * - visit_empty_object() - when an empty object is encountered.
+   * - visit_key(const uint8_t *key) - when a key in an object field is encountered. key is
+   *                                   guaranteed to point at the first quote of the string (`"key"`).
+   * - visit_primitive(const uint8_t *value) - when a value is a string, number, boolean or null.
+   * - visit_root_primitive(iter, uint8_t *value) - when the top-level value is a string, number, boolean or null.
+   *
+   * - increment_count(iter) - each time a value is found in an array or object.
+   */
+  template<bool STREAMING, typename V>
+  simdjson_warn_unused simdjson_inline error_code walk_document(V &visitor) noexcept;
+
+  /**
+   * Create an iterator capable of walking a JSON document.
+   *
+   * The document must have already passed through stage 1.
+   */
+  simdjson_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
+
+  /**
+   * Look at the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *peek() const noexcept;
+  /**
+   * Advance to the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *advance() noexcept;
+  /**
+   * Get the remaining length of the document, from the start of the current token.
+   */
+  simdjson_inline size_t remaining_len() const noexcept;
+  /**
+   * Check if we are at the end of the document.
+   *
+   * If this is true, there are no more tokens.
+   */
+  simdjson_inline bool at_eof() const noexcept;
+  /**
+   * Check if we are at the beginning of the document.
+   */
+  simdjson_inline bool at_beginning() const noexcept;
+  simdjson_inline uint8_t last_structural() const noexcept;
+
+  /**
+   * Log that a value has been found.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_value(const char *type) const noexcept;
+  /**
+   * Log the start of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_start_value(const char *type) const noexcept;
+  /**
+   * Log the end of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_end_value(const char *type) const noexcept;
+  /**
+   * Log an error.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_error(const char *error) const noexcept;
+
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
+};
+
+template<bool STREAMING, typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::walk_document(V &visitor) noexcept {
+  logger::log_start();
+
+  //
+  // Start the document
+  //
+  if (at_eof()) { return EMPTY; }
+  log_start_value("document");
+  SIMDJSON_TRY( visitor.visit_document_start(*this) );
+
+  //
+  // Read first value
+  //
+  {
+    auto value = advance();
+
+    // Make sure the outer object or array is closed before continuing; otherwise, there are ways we
+    // could get into memory corruption. See https://github.com/simdjson/simdjson/issues/906
+    if (!STREAMING) {
+      switch (*value) {
+        case '{': if (last_structural() != '}') { log_value("starting brace unmatched"); return TAPE_ERROR; }; break;
+        case '[': if (last_structural() != ']') { log_value("starting bracket unmatched"); return TAPE_ERROR; }; break;
+      }
+    }
+
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_root_primitive(*this, value) ); break;
+    }
+  }
+  goto document_end;
+
+//
+// Object parser states
+//
+object_begin:
+  log_start_value("object");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = false;
+  SIMDJSON_TRY( visitor.visit_object_start(*this) );
+
+  {
+    auto key = advance();
+    if (*key != '"') { log_error("Object does not start with a key"); return TAPE_ERROR; }
+    SIMDJSON_TRY( visitor.increment_count(*this) );
+    SIMDJSON_TRY( visitor.visit_key(*this, key) );
+  }
+
+object_field:
+  if (simdjson_unlikely( *advance() != ':' )) { log_error("Missing colon after key in object"); return TAPE_ERROR; }
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+object_continue:
+  switch (*advance()) {
+    case ',':
+      SIMDJSON_TRY( visitor.increment_count(*this) );
+      {
+        auto key = advance();
+        if (simdjson_unlikely( *key != '"' )) { log_error("Key string missing at beginning of field in object"); return TAPE_ERROR; }
+        SIMDJSON_TRY( visitor.visit_key(*this, key) );
+      }
+      goto object_field;
+    case '}': log_end_value("object"); SIMDJSON_TRY( visitor.visit_object_end(*this) ); goto scope_end;
+    default: log_error("No comma between object fields"); return TAPE_ERROR;
+  }
+
+scope_end:
+  depth--;
+  if (depth == 0) { goto document_end; }
+  if (dom_parser.is_array[depth]) { goto array_continue; }
+  goto object_continue;
+
+//
+// Array parser states
+//
+array_begin:
+  log_start_value("array");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = true;
+  SIMDJSON_TRY( visitor.visit_array_start(*this) );
+  SIMDJSON_TRY( visitor.increment_count(*this) );
+
+array_value:
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+array_continue:
+  switch (*advance()) {
+    case ',': SIMDJSON_TRY( visitor.increment_count(*this) ); goto array_value;
+    case ']': log_end_value("array"); SIMDJSON_TRY( visitor.visit_array_end(*this) ); goto scope_end;
+    default: log_error("Missing comma between array values"); return TAPE_ERROR;
+  }
+
+document_end:
+  log_end_value("document");
+  SIMDJSON_TRY( visitor.visit_document_end(*this) );
+
+  dom_parser.next_structural_index = uint32_t(next_structural - &dom_parser.structural_indexes[0]);
+
+  // If we didn't make it to the end, it's an error
+  if ( !STREAMING && dom_parser.next_structural_index != dom_parser.n_structural_indexes ) {
+    log_error("More than one JSON value at the root of the document, or extra characters at the end of the JSON!");
+    return TAPE_ERROR;
+  }
+
+  return SUCCESS;
+
+} // walk_document()
+
+simdjson_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+  : buf{_dom_parser.buf},
+    next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+    dom_parser{_dom_parser} {
+}
+
+simdjson_inline const uint8_t *json_iterator::peek() const noexcept {
+  return &buf[*(next_structural)];
+}
+simdjson_inline const uint8_t *json_iterator::advance() noexcept {
+  return &buf[*(next_structural++)];
+}
+simdjson_inline size_t json_iterator::remaining_len() const noexcept {
+  return dom_parser.len - *(next_structural-1);
+}
+
+simdjson_inline bool json_iterator::at_eof() const noexcept {
+  return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+}
+simdjson_inline bool json_iterator::at_beginning() const noexcept {
+  return next_structural == dom_parser.structural_indexes.get();
+}
+simdjson_inline uint8_t json_iterator::last_structural() const noexcept {
+  return buf[dom_parser.structural_indexes[dom_parser.n_structural_indexes - 1]];
+}
+
+simdjson_inline void json_iterator::log_value(const char *type) const noexcept {
+  logger::log_line(*this, "", type, "");
+}
+
+simdjson_inline void json_iterator::log_start_value(const char *type) const noexcept {
+  logger::log_line(*this, "+", type, "");
+  if (logger::LOG_ENABLED) { logger::log_depth++; }
+}
+
+simdjson_inline void json_iterator::log_end_value(const char *type) const noexcept {
+  if (logger::LOG_ENABLED) { logger::log_depth--; }
+  logger::log_line(*this, "-", type, "");
+}
+
+simdjson_inline void json_iterator::log_error(const char *error) const noexcept {
+  logger::log_line(*this, "", "ERROR", error);
+}
+
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
+  switch (*value) {
+    case '"': return visitor.visit_root_string(*this, value);
+    case 't': return visitor.visit_root_true_atom(*this, value);
+    case 'f': return visitor.visit_root_false_atom(*this, value);
+    case 'n': return visitor.visit_root_null_atom(*this, value);
+    case '-':
+    case '0': case '1': case '2': case '3': case '4':
+    case '5': case '6': case '7': case '8': case '9':
+      return visitor.visit_root_number(*this, value);
+    default:
+      log_error("Document starts with a non-value character");
+      return TAPE_ERROR;
+  }
+}
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
+  // Use the fact that most scalars are going to be either strings or numbers.
+  if(*value == '"') {
+    return visitor.visit_string(*this, value);
+  } else if (((*value - '0')  < 10) || (*value == '-')) {
+    return visitor.visit_number(*this, value);
+  }
+  // true, false, null are uncommon.
+  switch (*value) {
+    case 't': return visitor.visit_true_atom(*this, value);
+    case 'f': return visitor.visit_false_atom(*this, value);
+    case 'n': return visitor.visit_null_atom(*this, value);
+    default:
+      log_error("Non-value found when value was expected!");
+      return TAPE_ERROR;
+  }
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+/* end file generic/stage2/json_iterator.h for haswell */
+/* including generic/stage2/stringparsing.h for haswell: #include <generic/stage2/stringparsing.h> */
+/* begin file generic/stage2/stringparsing.h for haswell */
+#include <cstdint>
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/jsoncharutils.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses
+// It is intended to be included multiple times and compiled multiple times
+
+namespace simdjson {
+namespace haswell {
+namespace {
+/// @private
+namespace stringparsing {
+
+// begin copypasta
+// These chars yield themselves: " \ /
+// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
+// u not handled in this table as it's complex
+static const uint8_t escape_map[256] = {
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
+    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
+    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+};
+
+// handle a unicode codepoint
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr, bool allow_replacement) {
+  // Use the default Unicode Character 'REPLACEMENT CHARACTER' (U+FFFD)
+  constexpr uint32_t substitution_code_point = 0xfffd;
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) != ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+    } else {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+
+      // We have already checked that the high surrogate is valid and
+      // (code_point - 0xd800) < 1024.
+      //
+      // Check that code_point_2 is in the range 0xdc00..0xdfff
+      // and that code_point_2 was parsed from valid hex.
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if (low_bit >> 10) {
+        if(!allow_replacement) { return false; }
+        code_point = substitution_code_point;
+      } else {
+        code_point =  (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+
+    }
+  } else if (code_point >= 0xdc00 && code_point <= 0xdfff) {
+      // If we encounter a low surrogate (not preceded by a high surrogate)
+      // then we have an error.
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+  }
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+// handle a unicode codepoint using the wobbly convention
+// https://simonsapin.github.io/wtf-8/
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint_wobbly(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr) {
+  // It is not ideal that this function is nearly identical to handle_unicode_codepoint.
+  //
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) == ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if ((low_bit >> 10) ==  0) {
+        code_point =
+          (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+    }
+  }
+
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+/**
+ * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+ * must be an unescaped quote terminating the string. It returns the final output
+ * position as pointer. In case of error (e.g., the string has bad escaped codes),
+ * then null_ptr is returned. It is assumed that the output buffer is large
+ * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+ * SIMDJSON_PADDING bytes.
+ */
+simdjson_warn_unused simdjson_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) {
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint(&src, &dst, allow_replacement)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+simdjson_warn_unused simdjson_inline uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) {
+  // It is not ideal that this function is nearly identical to parse_string.
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint_wobbly(&src, &dst)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+} // namespace stringparsing
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+/* end file generic/stage2/stringparsing.h for haswell */
+/* including generic/stage2/structural_iterator.h for haswell: #include <generic/stage2/structural_iterator.h> */
+/* begin file generic/stage2/structural_iterator.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace stage2 {
+
+class structural_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+
+  // Start a structural
+  simdjson_inline structural_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+    : buf{_dom_parser.buf},
+      next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+      dom_parser{_dom_parser} {
+  }
+  // Get the buffer position of the current structural character
+  simdjson_inline const uint8_t* current() {
+    return &buf[*(next_structural-1)];
+  }
+  // Get the current structural character
+  simdjson_inline char current_char() {
+    return buf[*(next_structural-1)];
+  }
+  // Get the next structural character without advancing
+  simdjson_inline char peek_next_char() {
+    return buf[*next_structural];
+  }
+  simdjson_inline const uint8_t* peek() {
+    return &buf[*next_structural];
+  }
+  simdjson_inline const uint8_t* advance() {
+    return &buf[*(next_structural++)];
+  }
+  simdjson_inline char advance_char() {
+    return buf[*(next_structural++)];
+  }
+  simdjson_inline size_t remaining_len() {
+    return dom_parser.len - *(next_structural-1);
+  }
+
+  simdjson_inline bool at_end() {
+    return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+  }
+  simdjson_inline bool at_beginning() {
+    return next_structural == dom_parser.structural_indexes.get();
+  }
+};
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H
+/* end file generic/stage2/structural_iterator.h for haswell */
+/* including generic/stage2/tape_builder.h for haswell: #include <generic/stage2/tape_builder.h> */
+/* begin file generic/stage2/tape_builder.h for haswell */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/json_iterator.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/stringparsing.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/tape_writer.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/dom/document.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/atomparsing.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/numberparsing.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+
+namespace simdjson {
+namespace haswell {
+namespace {
+namespace stage2 {
+
+struct tape_builder {
+  template<bool STREAMING>
+  simdjson_warn_unused static simdjson_inline error_code parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept;
+
+  /** Called when a non-empty document starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty document ends without error. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_end(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty array starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty array ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_end(json_iterator &iter) noexcept;
+  /** Called when an empty array is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_array(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty object starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_start(json_iterator &iter) noexcept;
+  /**
+   * Called when a key in a field is encountered.
+   *
+   * primitive, visit_object_start, visit_empty_object, visit_array_start, or visit_empty_array
+   * will be called after this with the field value.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
+  /** Called when a non-empty object ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_end(json_iterator &iter) noexcept;
+  /** Called when an empty object is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_object(json_iterator &iter) noexcept;
+
+  /**
+   * Called when a string, number, boolean or null is found.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  /**
+   * Called when a string, number, boolean or null is found at the top level of a document (i.e.
+   * when there is no array or object and the entire document is a single string, number, boolean or
+   * null.
+   *
+   * This is separate from primitive() because simdjson's normal primitive parsing routines assume
+   * there is at least one more token after the value, which is only true in an array or object.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  /** Called each time a new field or element in an array or object is found. */
+  simdjson_warn_unused simdjson_inline error_code increment_count(json_iterator &iter) noexcept;
+
+  /** Next location to write to tape */
+  tape_writer tape;
+private:
+  /** Next write location in the string buf for stage 2 parsing */
+  uint8_t *current_string_buf_loc;
+
+  simdjson_inline tape_builder(dom::document &doc) noexcept;
+
+  simdjson_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
+  simdjson_inline void start_container(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_warn_unused simdjson_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
+  simdjson_inline void on_end_string(uint8_t *dst) noexcept;
+}; // struct tape_builder
+
+template<bool STREAMING>
+simdjson_warn_unused simdjson_inline error_code tape_builder::parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept {
+  dom_parser.doc = &doc;
+  json_iterator iter(dom_parser, STREAMING ? dom_parser.next_structural_index : 0);
+  tape_builder builder(doc);
+  return iter.walk_document<STREAMING>(builder);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_root_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
+  constexpr uint32_t start_tape_index = 0;
+  tape.append(start_tape_index, internal::tape_type::ROOT);
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter), internal::tape_type::ROOT);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
+  return visit_string(iter, key, true);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].count++; // we have a key value pair in the object at parser.dom_parser.depth - 1
+  return SUCCESS;
+}
+
+simdjson_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
+  iter.log_value(key ? "key" : "string");
+  uint8_t *dst = on_start_string(iter);
+  dst = stringparsing::parse_string(value+1, dst, false); // We do not allow replacement when the escape characters are invalid.
+  if (dst == nullptr) {
+    iter.log_error("Invalid escape in string");
+    return STRING_ERROR;
+  }
+  on_end_string(dst);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
+  return visit_string(iter, value);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("number");
+  error_code err = numberparsing::parse_number(value, tape);
+  if (simdjson_unlikely(err == BIGINT_ERROR &&
+      iter.dom_parser._number_as_string)) {
+    // Write big integer to string buffer using the same format as strings.
+    // Scan digits the same way parse_number does (skip optional '-', then digits).
+    const uint8_t *p = value;
+    if (*p == '-') p++;
+    while (numberparsing::is_digit(*p)) p++;
+    size_t len = size_t(p - value);
+    tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::BIGINT);
+    uint8_t *dst = current_string_buf_loc + sizeof(uint32_t);
+    memcpy(dst, value, len);
+    dst += len;
+    on_end_string(dst);
+    return SUCCESS;
+  }
+  return err;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
+  //
+  // We need to make a copy to make sure that the string is space terminated.
+  // This is not about padding the input, which should already padded up
+  // to len + SIMDJSON_PADDING. However, we have no control at this stage
+  // on how the padding was done. What if the input string was padded with nulls?
+  // It is quite common for an input string to have an extra null character (C string).
+  // We do not want to allow 9\0 (where \0 is the null character) inside a JSON
+  // document, but the string "9\0" by itself is fine. So we make a copy and
+  // pad the input with spaces when we know that there is just one input element.
+  // This copy is relatively expensive, but it will almost never be called in
+  // practice unless you are in the strange scenario where you have many JSON
+  // documents made of single atoms.
+  //
+  std::unique_ptr<uint8_t[]>copy(new (std::nothrow) uint8_t[iter.remaining_len() + SIMDJSON_PADDING]);
+  if (copy.get() == nullptr) { return MEMALLOC; }
+  std::memcpy(copy.get(), value, iter.remaining_len());
+  std::memset(copy.get() + iter.remaining_len(), ' ', SIMDJSON_PADDING);
+  error_code error = visit_number(iter, copy.get());
+  return error;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value)) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value, iter.remaining_len())) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value)) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value, iter.remaining_len())) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value)) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value, iter.remaining_len())) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+// private:
+
+simdjson_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
+  return uint32_t(tape.next_tape_loc - iter.dom_parser.doc->tape.get());
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  auto start_index = next_tape_index(iter);
+  tape.append(start_index+2, start);
+  tape.append(start_index, end);
+  return SUCCESS;
+}
+
+simdjson_inline void tape_builder::start_container(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].tape_index = next_tape_index(iter);
+  iter.dom_parser.open_containers[iter.depth].count = 0;
+  tape.skip(); // We don't actually *write* the start element until the end.
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  // Write the ending tape element, pointing at the start location
+  const uint32_t start_tape_index = iter.dom_parser.open_containers[iter.depth].tape_index;
+  tape.append(start_tape_index, end);
+  // Write the start tape element, pointing at the end location (and including count)
+  // count can overflow if it exceeds 24 bits... so we saturate
+  // the convention being that a cnt of 0xffffff or more is undetermined in value (>=  0xffffff).
+  const uint32_t count = iter.dom_parser.open_containers[iter.depth].count;
+  const uint32_t cntsat = count > 0xFFFFFF ? 0xFFFFFF : count;
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter) | (uint64_t(cntsat) << 32), start);
+  return SUCCESS;
+}
+
+simdjson_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
+  // we advance the point, accounting for the fact that we have a NULL termination
+  tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::STRING);
+  return current_string_buf_loc + sizeof(uint32_t);
+}
+
+simdjson_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
+  uint32_t str_length = uint32_t(dst - (current_string_buf_loc + sizeof(uint32_t)));
+  // TODO check for overflow in case someone has a crazy string (>=4GB?)
+  // But only add the overflow check when the document itself exceeds 4GB
+  // Currently unneeded because we refuse to parse docs larger or equal to 4GB.
+  memcpy(current_string_buf_loc, &str_length, sizeof(uint32_t));
+  // NULL termination is still handy if you expect all your strings to
+  // be NULL terminated? It comes at a small cost
+  *dst = 0;
+  current_string_buf_loc = dst + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+/* end file generic/stage2/tape_builder.h for haswell */
+/* end file generic/stage2/amalgamated.h for haswell */
+
+//
+// Stage 1
+//
+
+namespace simdjson {
+namespace haswell {
+
+simdjson_warn_unused error_code implementation::create_dom_parser_implementation(
+  size_t capacity,
+  size_t max_depth,
+  std::unique_ptr<internal::dom_parser_implementation>& dst
+) const noexcept {
+  dst.reset( new (std::nothrow) dom_parser_implementation() );
+  if (!dst) { return MEMALLOC; }
+  if (auto err = dst->set_capacity(capacity))
+    return err;
+  if (auto err = dst->set_max_depth(max_depth))
+    return err;
+  return SUCCESS;
+}
+
+namespace {
+
+using namespace simd;
+
+// This identifies structural characters (comma, colon, braces, brackets),
+// and ASCII white-space ('\r','\n','\t',' ').
+simdjson_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
+  // These lookups rely on the fact that anything < 127 will match the lower 4 bits, which is why
+  // we can't use the generic lookup_16.
+  const auto whitespace_table = simd8<uint8_t>::repeat_16(' ', 100, 100, 100, 17, 100, 113, 2, 100, '\t', '\n', 112, 100, '\r', 100, 100);
+
+  // The 6 operators (:,[]{}) have these values:
+  //
+  // , 2C
+  // : 3A
+  // [ 5B
+  // { 7B
+  // ] 5D
+  // } 7D
+  //
+  // If you use | 0x20 to turn [ and ] into { and }, the lower 4 bits of each character is unique.
+  // We exploit this, using a simd 4-bit lookup to tell us which character match against, and then
+  // match it (against | 0x20).
+  //
+  // To prevent recognizing other characters, everything else gets compared with 0, which cannot
+  // match due to the | 0x20.
+  //
+  // NOTE: Due to the | 0x20, this ALSO treats <FF> and <SUB> (control characters 0C and 1A) like ,
+  // and :. This gets caught in stage 2, which checks the actual character to ensure the right
+  // operators are in the right places.
+  const auto op_table = simd8<uint8_t>::repeat_16(
+    0, 0, 0, 0,
+    0, 0, 0, 0,
+    0, 0, ':', '{', // : = 3A, [ = 5B, { = 7B
+    ',', '}', 0, 0  // , = 2C, ] = 5D, } = 7D
+  );
+
+  // We compute whitespace and op separately. If later code only uses one or the
+  // other, given the fact that all functions are aggressively inlined, we can
+  // hope that useless computations will be omitted. This is namely case when
+  // minifying (we only need whitespace).
+
+  const uint64_t whitespace = in.eq({
+    _mm256_shuffle_epi8(whitespace_table, in.chunks[0]),
+    _mm256_shuffle_epi8(whitespace_table, in.chunks[1])
+  });
+  // Turn [ and ] into { and }
+  const simd8x64<uint8_t> curlified{
+    in.chunks[0] | 0x20,
+    in.chunks[1] | 0x20
+  };
+  const uint64_t op = curlified.eq({
+    _mm256_shuffle_epi8(op_table, in.chunks[0]),
+    _mm256_shuffle_epi8(op_table, in.chunks[1])
+  });
+
+  return { whitespace, op };
+}
+
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input) {
+  return input.reduce_or().is_ascii();
+}
+
+simdjson_unused simdjson_inline simd8<bool> must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_second_byte = prev1.saturating_sub(0xc0u-1); // Only 11______ will be > 0
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-1); // Only 111_____ will be > 0
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-1); // Only 1111____ will be > 0
+  // Caller requires a bool (all 1's). All values resulting from the subtraction will be <= 64, so signed comparison is fine.
+  return simd8<int8_t>(is_second_byte | is_third_byte | is_fourth_byte) > int8_t(0);
+}
+
+simdjson_inline simd8<uint8_t> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-0x80); // Only 111_____ will be >= 0x80
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-0x80); // Only 1111____ will be >= 0x80
+  return is_third_byte | is_fourth_byte;
+}
+
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+
+//
+// Stage 2
+//
+
+//
+// Implementation-specific overrides
+//
+namespace simdjson {
+namespace haswell {
+
+simdjson_warn_unused error_code implementation::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept {
+  return haswell::stage1::json_minifier::minify<128>(buf, len, dst, dst_len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage1(const uint8_t *_buf, size_t _len, stage1_mode streaming) noexcept {
+  this->buf = _buf;
+  this->len = _len;
+  return haswell::stage1::json_structural_indexer::index<128>(_buf, _len, *this, streaming);
+}
+
+simdjson_warn_unused bool implementation::validate_utf8(const char *buf, size_t len) const noexcept {
+  return haswell::stage1::generic_validate_utf8(buf,len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<false>(*this, _doc);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2_next(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<true>(*this, _doc);
+}
+
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_string(const uint8_t *src, uint8_t *dst, bool replacement_char) const noexcept {
+  return haswell::stringparsing::parse_string(src, dst, replacement_char);
+}
+
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept {
+  return haswell::stringparsing::parse_wobbly_string(src, dst);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::parse(const uint8_t *_buf, size_t _len, dom::document &_doc) noexcept {
+  auto error = stage1(_buf, _len, stage1_mode::regular);
+  if (error) { return error; }
+  return stage2(_doc);
+}
+
+} // namespace haswell
+} // namespace simdjson
+
+/* including simdjson/haswell/end.h: #include <simdjson/haswell/end.h> */
+/* begin file simdjson/haswell/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#if !SIMDJSON_CAN_ALWAYS_RUN_HASWELL
+SIMDJSON_UNTARGET_REGION
+#endif
+
+/* undefining SIMDJSON_IMPLEMENTATION from "haswell" */
+#undef SIMDJSON_IMPLEMENTATION
+/* end file simdjson/haswell/end.h */
+
+#endif // SIMDJSON_SRC_HASWELL_CPP
+/* end file haswell.cpp */
+#endif
+#if SIMDJSON_IMPLEMENTATION_ICELAKE
+/* including icelake.cpp: #include <icelake.cpp> */
+/* begin file icelake.cpp */
+#ifndef SIMDJSON_SRC_ICELAKE_CPP
+#define SIMDJSON_SRC_ICELAKE_CPP
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+/* including simdjson/icelake.h: #include <simdjson/icelake.h> */
+/* begin file simdjson/icelake.h */
+#ifndef SIMDJSON_ICELAKE_H
+#define SIMDJSON_ICELAKE_H
+
+/* including simdjson/icelake/begin.h: #include "simdjson/icelake/begin.h" */
+/* begin file simdjson/icelake/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "icelake" */
+#define SIMDJSON_IMPLEMENTATION icelake
+/* including simdjson/icelake/base.h: #include "simdjson/icelake/base.h" */
+/* begin file simdjson/icelake/base.h */
+#ifndef SIMDJSON_ICELAKE_BASE_H
+#define SIMDJSON_ICELAKE_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// The constructor may be executed on any host, so we take care not to use SIMDJSON_TARGET_ICELAKE
+namespace simdjson {
+/**
+ * Implementation for Icelake (Intel AVX512).
+ */
+namespace icelake {
+
+class implementation;
+
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_ICELAKE_BASE_H
+/* end file simdjson/icelake/base.h */
+/* including simdjson/icelake/intrinsics.h: #include "simdjson/icelake/intrinsics.h" */
+/* begin file simdjson/icelake/intrinsics.h */
+#ifndef SIMDJSON_ICELAKE_INTRINSICS_H
+#define SIMDJSON_ICELAKE_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#if SIMDJSON_VISUAL_STUDIO
+// under clang within visual studio, this will include <x86intrin.h>
+#include <intrin.h>  // visual studio or clang
+#else
+#include <x86intrin.h> // elsewhere
+#endif // SIMDJSON_VISUAL_STUDIO
+
+#if SIMDJSON_CLANG_VISUAL_STUDIO
+/**
+ * You are not supposed, normally, to include these
+ * headers directly. Instead you should either include intrin.h
+ * or x86intrin.h. However, when compiling with clang
+ * under Windows (i.e., when _MSC_VER is set), these headers
+ * only get included *if* the corresponding features are detected
+ * from macros:
+ * e.g., if __AVX2__ is set... in turn,  we normally set these
+ * macros by compiling against the corresponding architecture
+ * (e.g., arch:AVX2, -mavx2, etc.) which compiles the whole
+ * software with these advanced instructions. In simdjson, we
+ * want to compile the whole program for a generic target,
+ * and only target our specific kernels. As a workaround,
+ * we directly include the needed headers. These headers would
+ * normally guard against such usage, but we carefully included
+ * <x86intrin.h>  (or <intrin.h>) before, so the headers
+ * are fooled.
+ */
+#include <bmiintrin.h>   // for _blsr_u64
+#include <lzcntintrin.h> // for  __lzcnt64
+#include <immintrin.h>   // for most things (AVX2, AVX512, _popcnt64)
+#include <smmintrin.h>
+#include <tmmintrin.h>
+#include <avxintrin.h>
+#include <avx2intrin.h>
+#include <wmmintrin.h>   // for  _mm_clmulepi64_si128
+// Important: we need the AVX-512 headers:
+#include <avx512fintrin.h>
+#include <avx512dqintrin.h>
+#include <avx512cdintrin.h>
+#include <avx512bwintrin.h>
+#include <avx512vlintrin.h>
+#include <avx512vbmiintrin.h>
+#include <avx512vbmi2intrin.h>
+// unfortunately, we may not get _blsr_u64, but, thankfully, clang
+// has it as a macro.
+#ifndef _blsr_u64
+// we roll our own
+#define _blsr_u64(n) ((n - 1) & n)
+#endif //  _blsr_u64
+#endif // SIMDJSON_CLANG_VISUAL_STUDIO
+
+static_assert(sizeof(__m512i) <= simdjson::SIMDJSON_PADDING, "insufficient padding for icelake");
+
+#endif // SIMDJSON_ICELAKE_INTRINSICS_H
+/* end file simdjson/icelake/intrinsics.h */
+
+#if !SIMDJSON_CAN_ALWAYS_RUN_ICELAKE
+SIMDJSON_TARGET_REGION("avx512f,avx512dq,avx512cd,avx512bw,avx512vbmi,avx512vbmi2,avx512vl,avx2,bmi,pclmul,lzcnt,popcnt")
+#endif
+
+/* including simdjson/icelake/bitmanipulation.h: #include "simdjson/icelake/bitmanipulation.h" */
+/* begin file simdjson/icelake/bitmanipulation.h */
+#ifndef SIMDJSON_ICELAKE_BITMANIPULATION_H
+#define SIMDJSON_ICELAKE_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  return (int)_tzcnt_u64(input_num);
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO
+  ////////
+  // You might expect the next line to be equivalent to
+  // return (int)_tzcnt_u64(input_num);
+  // but the generated code differs and might be less efficient?
+  ////////
+  return __builtin_ctzll(input_num);
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return _blsr_u64(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+  return int(_lzcnt_u64(input_num));
+}
+
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+simdjson_inline unsigned __int64 count_ones(uint64_t input_num) {
+  // note: we do not support legacy 32-bit Windows
+  return __popcnt64(input_num);// Visual Studio wants two underscores
+}
+#else
+simdjson_inline long long int count_ones(uint64_t input_num) {
+  return _popcnt64(input_num);
+}
+#endif
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2,
+                                uint64_t *result) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  return _addcarry_u64(0, value1, value2,
+                       reinterpret_cast<unsigned __int64 *>(result));
+#else
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+#endif
+}
+
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_ICELAKE_BITMANIPULATION_H
+/* end file simdjson/icelake/bitmanipulation.h */
+/* including simdjson/icelake/bitmask.h: #include "simdjson/icelake/bitmask.h" */
+/* begin file simdjson/icelake/bitmask.h */
+#ifndef SIMDJSON_ICELAKE_BITMASK_H
+#define SIMDJSON_ICELAKE_BITMASK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(const uint64_t bitmask) {
+  // There should be no such thing with a processor supporting avx2
+  // but not clmul.
+  __m128i all_ones = _mm_set1_epi8('\xFF');
+  __m128i result = _mm_clmulepi64_si128(_mm_set_epi64x(0ULL, bitmask), all_ones, 0);
+  return _mm_cvtsi128_si64(result);
+}
+
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_ICELAKE_BITMASK_H
+/* end file simdjson/icelake/bitmask.h */
+/* including simdjson/icelake/simd.h: #include "simdjson/icelake/simd.h" */
+/* begin file simdjson/icelake/simd.h */
+#ifndef SIMDJSON_ICELAKE_SIMD_H
+#define SIMDJSON_ICELAKE_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#if defined(__GNUC__) && !defined(__clang__)
+#if __GNUC__ == 8
+#define SIMDJSON_GCC8 1
+#endif //  __GNUC__ == 8
+#endif // defined(__GNUC__) && !defined(__clang__)
+
+#if SIMDJSON_GCC8
+/**
+ * GCC 8 fails to provide _mm512_set_epi8. We roll our own.
+ */
+inline __m512i _mm512_set_epi8(uint8_t a0, uint8_t a1, uint8_t a2, uint8_t a3, uint8_t a4, uint8_t a5, uint8_t a6, uint8_t a7, uint8_t a8, uint8_t a9, uint8_t a10, uint8_t a11, uint8_t a12, uint8_t a13, uint8_t a14, uint8_t a15, uint8_t a16, uint8_t a17, uint8_t a18, uint8_t a19, uint8_t a20, uint8_t a21, uint8_t a22, uint8_t a23, uint8_t a24, uint8_t a25, uint8_t a26, uint8_t a27, uint8_t a28, uint8_t a29, uint8_t a30, uint8_t a31, uint8_t a32, uint8_t a33, uint8_t a34, uint8_t a35, uint8_t a36, uint8_t a37, uint8_t a38, uint8_t a39, uint8_t a40, uint8_t a41, uint8_t a42, uint8_t a43, uint8_t a44, uint8_t a45, uint8_t a46, uint8_t a47, uint8_t a48, uint8_t a49, uint8_t a50, uint8_t a51, uint8_t a52, uint8_t a53, uint8_t a54, uint8_t a55, uint8_t a56, uint8_t a57, uint8_t a58, uint8_t a59, uint8_t a60, uint8_t a61, uint8_t a62, uint8_t a63) {
+  return _mm512_set_epi64(uint64_t(a7) + (uint64_t(a6) << 8) + (uint64_t(a5) << 16) + (uint64_t(a4) << 24) + (uint64_t(a3) << 32) + (uint64_t(a2) << 40) + (uint64_t(a1) << 48) + (uint64_t(a0) << 56),
+                          uint64_t(a15) + (uint64_t(a14) << 8) + (uint64_t(a13) << 16) + (uint64_t(a12) << 24) + (uint64_t(a11) << 32) + (uint64_t(a10) << 40) + (uint64_t(a9) << 48) + (uint64_t(a8) << 56),
+                          uint64_t(a23) + (uint64_t(a22) << 8) + (uint64_t(a21) << 16) + (uint64_t(a20) << 24) + (uint64_t(a19) << 32) + (uint64_t(a18) << 40) + (uint64_t(a17) << 48) + (uint64_t(a16) << 56),
+                          uint64_t(a31) + (uint64_t(a30) << 8) + (uint64_t(a29) << 16) + (uint64_t(a28) << 24) + (uint64_t(a27) << 32) + (uint64_t(a26) << 40) + (uint64_t(a25) << 48) + (uint64_t(a24) << 56),
+                          uint64_t(a39) + (uint64_t(a38) << 8) + (uint64_t(a37) << 16) + (uint64_t(a36) << 24) + (uint64_t(a35) << 32) + (uint64_t(a34) << 40) + (uint64_t(a33) << 48) + (uint64_t(a32) << 56),
+                          uint64_t(a47) + (uint64_t(a46) << 8) + (uint64_t(a45) << 16) + (uint64_t(a44) << 24) + (uint64_t(a43) << 32) + (uint64_t(a42) << 40) + (uint64_t(a41) << 48) + (uint64_t(a40) << 56),
+                          uint64_t(a55) + (uint64_t(a54) << 8) + (uint64_t(a53) << 16) + (uint64_t(a52) << 24) + (uint64_t(a51) << 32) + (uint64_t(a50) << 40) + (uint64_t(a49) << 48) + (uint64_t(a48) << 56),
+                          uint64_t(a63) + (uint64_t(a62) << 8) + (uint64_t(a61) << 16) + (uint64_t(a60) << 24) + (uint64_t(a59) << 32) + (uint64_t(a58) << 40) + (uint64_t(a57) << 48) + (uint64_t(a56) << 56));
+}
+#endif // SIMDJSON_GCC8
+
+
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace simd {
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename Child>
+  struct base {
+    __m512i value;
+
+    // Zero constructor
+    simdjson_inline base() : value{__m512i()} {}
+
+    // Conversion from SIMD register
+    simdjson_inline base(const __m512i _value) : value(_value) {}
+
+    // Conversion to SIMD register
+    simdjson_inline operator const __m512i&() const { return this->value; }
+    simdjson_inline operator __m512i&() { return this->value; }
+
+    // Bit operations
+    simdjson_inline Child operator|(const Child other) const { return _mm512_or_si512(*this, other); }
+    simdjson_inline Child operator&(const Child other) const { return _mm512_and_si512(*this, other); }
+    simdjson_inline Child operator^(const Child other) const { return _mm512_xor_si512(*this, other); }
+    simdjson_inline Child bit_andnot(const Child other) const { return _mm512_andnot_si512(other, *this); }
+    simdjson_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+  };
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename T>
+  struct simd8;
+
+  template<typename T, typename Mask=simd8<bool>>
+  struct base8: base<simd8<T>> {
+    typedef uint32_t bitmask_t;
+    typedef uint64_t bitmask2_t;
+
+    simdjson_inline base8() : base<simd8<T>>() {}
+    simdjson_inline base8(const __m512i _value) : base<simd8<T>>(_value) {}
+
+    friend simdjson_really_inline uint64_t operator==(const simd8<T> lhs, const simd8<T> rhs) {
+      return _mm512_cmpeq_epi8_mask(lhs, rhs);
+    }
+    static const int SIZE = sizeof(base<T>::value);
+
+    template<int N=1>
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+     // workaround for compilers unable to figure out that 16 - N is a constant (GCC 8)
+      constexpr int shift = 16 - N;
+      return _mm512_alignr_epi8(*this, _mm512_permutex2var_epi64(prev_chunk, _mm512_set_epi64(13, 12, 11, 10, 9, 8, 7, 6), *this), shift);
+    }
+  };
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool>: base8<bool> {
+    static simdjson_inline simd8<bool> splat(bool _value) { return _mm512_set1_epi8(uint8_t(-(!!_value))); }
+
+    simdjson_inline simd8() : base8() {}
+    simdjson_inline simd8(const __m512i _value) : base8<bool>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+    simdjson_inline bool any() const { return !!_mm512_test_epi8_mask (*this, *this); }
+    simdjson_inline simd8<bool> operator~() const { return *this ^ true; }
+  };
+
+  template<typename T>
+  struct base8_numeric: base8<T> {
+    static simdjson_inline simd8<T> splat(T _value) { return _mm512_set1_epi8(_value); }
+    static simdjson_inline simd8<T> zero() { return _mm512_setzero_si512(); }
+    static simdjson_inline simd8<T> load(const T values[64]) {
+      return _mm512_loadu_si512(reinterpret_cast<const __m512i *>(values));
+    }
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    static simdjson_inline simd8<T> repeat_16(
+      T v0,  T v1,  T v2,  T v3,  T v4,  T v5,  T v6,  T v7,
+      T v8,  T v9,  T v10, T v11, T v12, T v13, T v14, T v15
+    ) {
+      return simd8<T>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    simdjson_inline base8_numeric() : base8<T>() {}
+    simdjson_inline base8_numeric(const __m512i _value) : base8<T>(_value) {}
+
+    // Store to array
+    simdjson_inline void store(T dst[64]) const { return _mm512_storeu_si512(reinterpret_cast<__m512i *>(dst), *this); }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<T> operator+(const simd8<T> other) const { return _mm512_add_epi8(*this, other); }
+    simdjson_inline simd8<T> operator-(const simd8<T> other) const { return _mm512_sub_epi8(*this, other); }
+    simdjson_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
+
+    // Override to distinguish from bool version
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return _mm512_shuffle_epi8(lookup_table, *this);
+    }
+
+    // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted as a bitset).
+    // Passing a 0 value for mask would be equivalent to writing out every byte to output.
+    // Only the first 64 - count_ones(mask) bytes of the result are significant but 64 bytes
+    // get written.
+    // Design consideration: it seems like a function with the
+    // signature simd8<L> compress(uint32_t mask) would be
+    // sensible, but the AVX ISA makes this kind of approach difficult.
+    template<typename L>
+    simdjson_inline void compress(uint64_t mask, L * output) const {
+      // we deliberately avoid _mm512_mask_compressstoreu_epi8 for portability
+      // (AMD Zen4 has terrible performance with it, it is effectively broken)
+      // _mm512_mask_compressstoreu_epi8 (output,~mask,*this);
+      __m512i compressed = _mm512_maskz_compress_epi8(~mask, *this);
+      _mm512_storeu_si512(output, compressed); // could use a mask
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> : base8_numeric<int8_t> {
+    simdjson_inline simd8() : base8_numeric<int8_t>() {}
+    simdjson_inline simd8(const __m512i _value) : base8_numeric<int8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const int8_t values[64]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15,
+      int8_t v16, int8_t v17, int8_t v18, int8_t v19, int8_t v20, int8_t v21, int8_t v22, int8_t v23,
+      int8_t v24, int8_t v25, int8_t v26, int8_t v27, int8_t v28, int8_t v29, int8_t v30, int8_t v31,
+      int8_t v32, int8_t v33, int8_t v34, int8_t v35, int8_t v36, int8_t v37, int8_t v38, int8_t v39,
+      int8_t v40, int8_t v41, int8_t v42, int8_t v43, int8_t v44, int8_t v45, int8_t v46, int8_t v47,
+      int8_t v48, int8_t v49, int8_t v50, int8_t v51, int8_t v52, int8_t v53, int8_t v54, int8_t v55,
+      int8_t v56, int8_t v57, int8_t v58, int8_t v59, int8_t v60, int8_t v61, int8_t v62, int8_t v63
+    ) : simd8(_mm512_set_epi8(
+      v63, v62, v61, v60, v59, v58, v57, v56,
+      v55, v54, v53, v52, v51, v50, v49, v48,
+      v47, v46, v45, v44, v43, v42, v41, v40,
+      v39, v38, v37, v36, v35, v34, v33, v32,
+      v31, v30, v29, v28, v27, v26, v25, v24,
+      v23, v22, v21, v20, v19, v18, v17, v16,
+      v15, v14, v13, v12, v11, v10,  v9,  v8,
+       v7,  v6,  v5,  v4,  v3,  v2,  v1,  v0
+    )) {}
+
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm512_max_epi8(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm512_min_epi8(*this, other); }
+
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm512_maskz_abs_epi8(_mm512_cmpgt_epi8_mask(*this, other),_mm512_set1_epi8(uint8_t(0x80))); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return _mm512_maskz_abs_epi8(_mm512_cmpgt_epi8_mask(other, *this),_mm512_set1_epi8(uint8_t(0x80))); }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t>: base8_numeric<uint8_t> {
+    simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+    simdjson_inline simd8(const __m512i _value) : base8_numeric<uint8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const uint8_t values[64]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15,
+      uint8_t v16, uint8_t v17, uint8_t v18, uint8_t v19, uint8_t v20, uint8_t v21, uint8_t v22, uint8_t v23,
+      uint8_t v24, uint8_t v25, uint8_t v26, uint8_t v27, uint8_t v28, uint8_t v29, uint8_t v30, uint8_t v31,
+      uint8_t v32, uint8_t v33, uint8_t v34, uint8_t v35, uint8_t v36, uint8_t v37, uint8_t v38, uint8_t v39,
+      uint8_t v40, uint8_t v41, uint8_t v42, uint8_t v43, uint8_t v44, uint8_t v45, uint8_t v46, uint8_t v47,
+      uint8_t v48, uint8_t v49, uint8_t v50, uint8_t v51, uint8_t v52, uint8_t v53, uint8_t v54, uint8_t v55,
+      uint8_t v56, uint8_t v57, uint8_t v58, uint8_t v59, uint8_t v60, uint8_t v61, uint8_t v62, uint8_t v63
+    ) : simd8(_mm512_set_epi8(
+      v63, v62, v61, v60, v59, v58, v57, v56,
+      v55, v54, v53, v52, v51, v50, v49, v48,
+      v47, v46, v45, v44, v43, v42, v41, v40,
+      v39, v38, v37, v36, v35, v34, v33, v32,
+      v31, v30, v29, v28, v27, v26, v25, v24,
+      v23, v22, v21, v20, v19, v18, v17, v16,
+      v15, v14, v13, v12, v11, v10,  v9,  v8,
+       v7,  v6,  v5,  v4,  v3,  v2,  v1,  v0
+    )) {}
+
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm512_adds_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm512_subs_epu8(*this, other); }
+
+    // Order-specific operations
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm512_max_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return _mm512_min_epu8(other, *this); }
+    // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
+    // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
+    simdjson_inline uint64_t operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
+    simdjson_inline uint64_t operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->lt_bits(other).any_bits_set(); }
+
+    // Bit-specific operations
+    simdjson_inline simd8<bool> bits_not_set() const { return _mm512_mask_blend_epi8(*this == uint8_t(0), _mm512_set1_epi8(0), _mm512_set1_epi8(-1)); }
+    simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
+
+    simdjson_inline bool is_ascii() const { return _mm512_movepi8_mask(*this) == 0; }
+    simdjson_inline bool bits_not_set_anywhere() const {
+      return !_mm512_test_epi8_mask(*this, *this);
+    }
+    simdjson_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
+    simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return !_mm512_test_epi8_mask(*this, bits); }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm512_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(_mm512_slli_epi16(*this, N)) & uint8_t(0xFFu << N); }
+    // Get one of the bits and make a bitmask out of it.
+    // e.g. value.get_bit<7>() gets the high bit
+    template<int N>
+    simdjson_inline uint64_t get_bit() const { return _mm512_movepi8_mask(_mm512_slli_epi16(*this, 7-N)); }
+  };
+
+  template<typename T>
+  struct simd8x64 {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+    static_assert(NUM_CHUNKS == 1, "Icelake kernel should use one register per 64-byte block.");
+    const simd8<T> chunks[NUM_CHUNKS];
+    template<int idx> simd8<uint8_t> get() const { return idx < NUM_CHUNKS ? chunks[idx] : simd8<T>(); }
+
+    simd8x64(const simd8x64<T>& o) = delete; // no copy allowed
+    simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1) : chunks{chunk0, chunk1} {}
+    simdjson_inline simd8x64(const simd8<T> chunk0) : chunks{chunk0} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr)} {}
+
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
+      this->chunks[0].compress(mask, output);
+      return 64 - count_ones(mask);
+    }
+
+    simdjson_inline void store(T ptr[64]) const {
+      this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
+    }
+
+    simdjson_inline simd8<T> reduce_or() const {
+      return this->chunks[0];
+    }
+
+    simdjson_inline simd8x64<T> bit_or(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return simd8x64<T>(
+        this->chunks[0] | mask
+      );
+    }
+
+    simdjson_inline uint64_t eq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return this->chunks[0] == mask;
+    }
+
+    simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+      return this->chunks[0] == other.chunks[0];
+    }
+
+    simdjson_inline uint64_t lteq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return this->chunks[0] <= mask;
+    }
+  }; // struct simd8x64<T>
+
+} // namespace simd
+
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_ICELAKE_SIMD_H
+/* end file simdjson/icelake/simd.h */
+/* including simdjson/icelake/stringparsing_defs.h: #include "simdjson/icelake/stringparsing_defs.h" */
+/* begin file simdjson/icelake/stringparsing_defs.h */
+#ifndef SIMDJSON_ICELAKE_STRINGPARSING_DEFS_H
+#define SIMDJSON_ICELAKE_STRINGPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/simd.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 64;
+  simdjson_inline backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return ((quote_bits - 1) & bs_bits) != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+
+  uint64_t bs_bits;
+  uint64_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // this can read up to 15 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  // store to dest unconditionally - we can overwrite the bits we don't like later
+  v.store(dst);
+  return {
+      static_cast<uint64_t>(v == '\\'), // bs_bits
+      static_cast<uint64_t>(v == '"'), // quote_bits
+  };
+}
+
+
+
+struct escaping {
+  static constexpr uint32_t BYTES_PROCESSED = 64;
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits != 0; }
+  simdjson_inline int escape_index() { return trailing_zeroes(uint64_t(escape_bits)); }
+
+  __mmask64 escape_bits;
+}; // struct escaping
+
+
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "escaping finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  __mmask64 is_quote = _mm512_cmpeq_epi8_mask(v, _mm512_set1_epi8('"'));
+  __mmask64 is_backslash = _mm512_cmpeq_epi8_mask(v, _mm512_set1_epi8('\\'));
+  __mmask64 is_control = _mm512_cmplt_epi8_mask(v, _mm512_set1_epi8(32));
+  return {
+    (is_backslash | is_quote | is_control)
+  };
+}
+
+
+
+
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_ICELAKE_STRINGPARSING_DEFS_H
+/* end file simdjson/icelake/stringparsing_defs.h */
+/* including simdjson/icelake/numberparsing_defs.h: #include "simdjson/icelake/numberparsing_defs.h" */
+/* begin file simdjson/icelake/numberparsing_defs.h */
+#ifndef SIMDJSON_ICELAKE_NUMBERPARSING_DEFS_H
+#define SIMDJSON_ICELAKE_NUMBERPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace numberparsing {
+
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  // this actually computes *16* values so we are being wasteful.
+  const __m128i ascii0 = _mm_set1_epi8('0');
+  const __m128i mul_1_10 =
+      _mm_setr_epi8(10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1);
+  const __m128i mul_1_100 = _mm_setr_epi16(100, 1, 100, 1, 100, 1, 100, 1);
+  const __m128i mul_1_10000 =
+      _mm_setr_epi16(10000, 1, 10000, 1, 10000, 1, 10000, 1);
+  const __m128i input = _mm_sub_epi8(
+      _mm_loadu_si128(reinterpret_cast<const __m128i *>(chars)), ascii0);
+  const __m128i t1 = _mm_maddubs_epi16(input, mul_1_10);
+  const __m128i t2 = _mm_madd_epi16(t1, mul_1_100);
+  const __m128i t3 = _mm_packus_epi32(t2, t2);
+  const __m128i t4 = _mm_madd_epi16(t3, mul_1_10000);
+  return _mm_cvtsi128_si32(
+      t4); // only captures the sum of the first 8 digits, drop the rest
+}
+
+/** @private */
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+#if SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+#if SIMDJSON_IS_ARM64
+  // ARM64 has native support for 64-bit multiplications, no need to emultate
+  answer.high = __umulh(value1, value2);
+  answer.low = value1 * value2;
+#else
+  answer.low = _umul128(value1, value2, &answer.high); // _umul128 not available on ARM64
+#endif // SIMDJSON_IS_ARM64
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+#endif
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace icelake
+} // namespace simdjson
+
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+
+#endif // SIMDJSON_ICELAKE_NUMBERPARSING_DEFS_H
+/* end file simdjson/icelake/numberparsing_defs.h */
+/* end file simdjson/icelake/begin.h */
+/* including simdjson/generic/amalgamated.h for icelake: #include "simdjson/generic/amalgamated.h" */
+/* begin file simdjson/generic/amalgamated.h for icelake */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_GENERIC_DEPENDENCIES_H)
+#error simdjson/generic/dependencies.h must be included before simdjson/generic/amalgamated.h!
+#endif
+
+/* including simdjson/generic/base.h for icelake: #include "simdjson/generic/base.h" */
+/* begin file simdjson/generic/base.h for icelake */
+#ifndef SIMDJSON_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): // If we haven't got an implementation yet, we're in the editor, editing a generic file! Just */
+/* amalgamation skipped (editor-only): // use the most advanced one we can so the most possible stuff can be tested. */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation_detection.h" */
+/* amalgamation skipped (editor-only): #if SIMDJSON_IMPLEMENTATION_ICELAKE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_HASWELL */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_WESTMERE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_ARM64 */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_PPC64 */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LASX */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LSX */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_RVV_VLS */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_FALLBACK */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/begin.h" */
+/* amalgamation skipped (editor-only): #else */
+/* amalgamation skipped (editor-only): #error "All possible implementations (including fallback) have been disabled! simdjson will not run." */
+/* amalgamation skipped (editor-only): #endif */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+
+struct open_container;
+class dom_parser_implementation;
+
+/**
+ * The type of a JSON number
+ */
+enum class number_type {
+    floating_point_number=1, /// a binary64 number
+    signed_integer,          /// a signed integer that fits in a 64-bit word using two's complement
+    unsigned_integer,        /// a positive integer larger or equal to 1<<63
+    big_integer              /// a big integer that does not fit in a 64-bit word
+};
+
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_BASE_H
+/* end file simdjson/generic/base.h for icelake */
+/* including simdjson/generic/jsoncharutils.h for icelake: #include "simdjson/generic/jsoncharutils.h" */
+/* begin file simdjson/generic/jsoncharutils.h for icelake */
+#ifndef SIMDJSON_GENERIC_JSONCHARUTILS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_JSONCHARUTILS_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/jsoncharutils_tables.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace jsoncharutils {
+
+// return non-zero if not a structural or whitespace char
+// zero otherwise
+simdjson_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace_negated[c];
+}
+
+simdjson_inline uint32_t is_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace[c];
+}
+
+// returns a value with the high 16 bits set if not valid
+// otherwise returns the conversion of the 4 hex digits at src into the bottom
+// 16 bits of the 32-bit return register
+//
+// see
+// https://lemire.me/blog/2019/04/17/parsing-short-hexadecimal-strings-efficiently/
+static inline uint32_t hex_to_u32_nocheck(
+    const uint8_t *src) { // strictly speaking, static inline is a C-ism
+  uint32_t v1 = internal::digit_to_val32[630 + src[0]];
+  uint32_t v2 = internal::digit_to_val32[420 + src[1]];
+  uint32_t v3 = internal::digit_to_val32[210 + src[2]];
+  uint32_t v4 = internal::digit_to_val32[0 + src[3]];
+  return v1 | v2 | v3 | v4;
+}
+
+// given a code point cp, writes to c
+// the utf-8 code, outputting the length in
+// bytes, if the length is zero, the code point
+// is invalid
+//
+// This can possibly be made faster using pdep
+// and clz and table lookups, but JSON documents
+// have few escaped code points, and the following
+// function looks cheap.
+//
+// Note: we assume that surrogates are treated separately
+//
+simdjson_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
+  if (cp <= 0x7F) {
+    c[0] = uint8_t(cp);
+    return 1; // ascii
+  }
+  if (cp <= 0x7FF) {
+    c[0] = uint8_t((cp >> 6) + 192);
+    c[1] = uint8_t((cp & 63) + 128);
+    return 2; // universal plane
+    //  Surrogates are treated elsewhere...
+    //} //else if (0xd800 <= cp && cp <= 0xdfff) {
+    //  return 0; // surrogates // could put assert here
+  } else if (cp <= 0xFFFF) {
+    c[0] = uint8_t((cp >> 12) + 224);
+    c[1] = uint8_t(((cp >> 6) & 63) + 128);
+    c[2] = uint8_t((cp & 63) + 128);
+    return 3;
+  } else if (cp <= 0x10FFFF) { // if you know you have a valid code point, this
+                               // is not needed
+    c[0] = uint8_t((cp >> 18) + 240);
+    c[1] = uint8_t(((cp >> 12) & 63) + 128);
+    c[2] = uint8_t(((cp >> 6) & 63) + 128);
+    c[3] = uint8_t((cp & 63) + 128);
+    return 4;
+  }
+  // will return 0 when the code point was too large.
+  return 0; // bad r
+}
+
+#if SIMDJSON_IS_32BITS // _umul128 for x86, arm
+// this is a slow emulation routine for 32-bit
+//
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+  return x * (uint64_t)y;
+}
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+  uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
+  uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
+  uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
+  uint64_t adbc_carry = !!(adbc < ad);
+  uint64_t lo = bd + (adbc << 32);
+  *hi = __emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) +
+        (adbc_carry << 32) + !!(lo < bd);
+  return lo;
+}
+#endif
+
+} // namespace jsoncharutils
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_JSONCHARUTILS_H
+/* end file simdjson/generic/jsoncharutils.h for icelake */
+/* including simdjson/generic/atomparsing.h for icelake: #include "simdjson/generic/atomparsing.h" */
+/* begin file simdjson/generic/atomparsing.h for icelake */
+#ifndef SIMDJSON_GENERIC_ATOMPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_ATOMPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace icelake {
+namespace {
+/// @private
+namespace atomparsing {
+
+// The string_to_uint32 is exclusively used to map literal strings to 32-bit values.
+// We use memcpy instead of a pointer cast to avoid undefined behaviors since we cannot
+// be certain that the character pointer will be properly aligned.
+// You might think that using memcpy makes this function expensive, but you'd be wrong.
+// All decent optimizing compilers (GCC, clang, Visual Studio) will compile string_to_uint32("false");
+// to the compile-time constant 1936482662.
+simdjson_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
+
+
+// Again in str4ncmp we use a memcpy to avoid undefined behavior. The memcpy may appear expensive.
+// Yet all decent optimizing compilers will compile memcpy to a single instruction, just about.
+simdjson_warn_unused
+simdjson_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
+  uint32_t srcval; // we want to avoid unaligned 32-bit loads (undefined in C/C++)
+  static_assert(sizeof(uint32_t) <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be larger than 4 bytes");
+  std::memcpy(&srcval, src, sizeof(uint32_t));
+  return srcval ^ string_to_uint32(atom);
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src) {
+  return (str4ncmp(src, "true") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_true_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "true"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src) {
+  return (str4ncmp(src+1, "alse") | jsoncharutils::is_not_structural_or_whitespace(src[5])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
+  if (len > 5) { return is_valid_false_atom(src); }
+  else if (len == 5) { return !str4ncmp(src+1, "alse"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src) {
+  return (str4ncmp(src, "null") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_null_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "null"); }
+  else { return false; }
+}
+
+} // namespace atomparsing
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_ATOMPARSING_H
+/* end file simdjson/generic/atomparsing.h for icelake */
+/* including simdjson/generic/dom_parser_implementation.h for icelake: #include "simdjson/generic/dom_parser_implementation.h" */
+/* begin file simdjson/generic/dom_parser_implementation.h for icelake */
+#ifndef SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/dom_parser_implementation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+
+// expectation: sizeof(open_container) = 64/8.
+struct open_container {
+  uint32_t tape_index; // where, on the tape, does the scope ([,{) begins
+  uint32_t count; // how many elements in the scope
+}; // struct open_container
+
+static_assert(sizeof(open_container) == 64/8, "Open container must be 64 bits");
+
+class dom_parser_implementation final : public internal::dom_parser_implementation {
+public:
+  /** Tape location of each open { or [ */
+  std::unique_ptr<open_container[]> open_containers{};
+  /** Whether each open container is a [ or { */
+  std::unique_ptr<bool[]> is_array{};
+  /** Buffer passed to stage 1 */
+  const uint8_t *buf{};
+  /** Length passed to stage 1 */
+  size_t len{0};
+  /** Document passed to stage 2 */
+  dom::document *doc{};
+
+  inline dom_parser_implementation() noexcept;
+  inline dom_parser_implementation(dom_parser_implementation &&other) noexcept;
+  inline dom_parser_implementation &operator=(dom_parser_implementation &&other) noexcept;
+  dom_parser_implementation(const dom_parser_implementation &) = delete;
+  dom_parser_implementation &operator=(const dom_parser_implementation &) = delete;
+
+  simdjson_warn_unused error_code parse(const uint8_t *buf, size_t len, dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage1(const uint8_t *buf, size_t len, stage1_mode partial) noexcept final;
+  simdjson_warn_unused error_code stage2(dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage2_next(dom::document &doc) noexcept final;
+  simdjson_warn_unused uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) const noexcept final;
+  simdjson_warn_unused uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept final;
+  inline simdjson_warn_unused error_code set_capacity(size_t capacity) noexcept final;
+  inline simdjson_warn_unused error_code set_max_depth(size_t max_depth) noexcept final;
+private:
+  simdjson_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
+
+};
+
+} // namespace icelake
+} // namespace simdjson
+
+namespace simdjson {
+namespace icelake {
+
+inline dom_parser_implementation::dom_parser_implementation() noexcept = default;
+inline dom_parser_implementation::dom_parser_implementation(dom_parser_implementation &&other) noexcept = default;
+inline dom_parser_implementation &dom_parser_implementation::operator=(dom_parser_implementation &&other) noexcept = default;
+
+// Leaving these here so they can be inlined if so desired
+inline simdjson_warn_unused error_code dom_parser_implementation::set_capacity(size_t capacity) noexcept {
+  if(capacity > SIMDJSON_MAXSIZE_BYTES) { return CAPACITY; }
+  // Stage 1 index output
+  size_t max_structures = SIMDJSON_ROUNDUP_N(capacity, 64) + 2 + 7;
+  structural_indexes.reset( new (std::nothrow) uint32_t[max_structures] );
+  if (!structural_indexes) { _capacity = 0; return MEMALLOC; }
+  structural_indexes[0] = 0;
+  n_structural_indexes = 0;
+
+  _capacity = capacity;
+  return SUCCESS;
+}
+
+inline simdjson_warn_unused error_code dom_parser_implementation::set_max_depth(size_t max_depth) noexcept {
+  // Stage 2 stacks
+  open_containers.reset(new (std::nothrow) open_container[max_depth]);
+  is_array.reset(new (std::nothrow) bool[max_depth]);
+  if (!is_array || !open_containers) { _max_depth = 0; return MEMALLOC; }
+
+  _max_depth = max_depth;
+  return SUCCESS;
+}
+
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file simdjson/generic/dom_parser_implementation.h for icelake */
+/* including simdjson/generic/implementation_simdjson_result_base.h for icelake: #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base.h for icelake */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+
+// This is a near copy of include/error.h's implementation_simdjson_result_base, except it doesn't use std::pair
+// so we can avoid inlining errors
+// TODO reconcile these!
+/**
+ * The result of a simdjson operation that could fail.
+ *
+ * Gives the option of reading error codes, or throwing an exception by casting to the desired result.
+ *
+ * This is a base class for implementations that want to add functions to the result type for
+ * chaining.
+ *
+ * Override like:
+ *
+ *   struct simdjson_result<T> : public internal::implementation_simdjson_result_base<T> {
+ *     simdjson_result() noexcept : internal::implementation_simdjson_result_base<T>() {}
+ *     simdjson_result(error_code error) noexcept : internal::implementation_simdjson_result_base<T>(error) {}
+ *     simdjson_result(T &&value) noexcept : internal::implementation_simdjson_result_base<T>(std::forward(value)) {}
+ *     simdjson_result(T &&value, error_code error) noexcept : internal::implementation_simdjson_result_base<T>(value, error) {}
+ *     // Your extra methods here
+ *   }
+ *
+ * Then any method returning simdjson_result<T> will be chainable with your methods.
+ */
+template<typename T>
+struct implementation_simdjson_result_base {
+
+  /**
+   * Create a new empty result with error = UNINITIALIZED.
+   */
+  simdjson_inline implementation_simdjson_result_base() noexcept = default;
+
+  /**
+   * Create a new error result.
+   */
+  simdjson_inline implementation_simdjson_result_base(error_code error) noexcept;
+
+  /**
+   * Create a new successful result.
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value) noexcept;
+
+  /**
+   * Create a new result with both things (use if you don't want to branch when creating the result).
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value, error_code error) noexcept;
+
+  /**
+   * Move the value and the error to the provided variables.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   * @param error The variable to assign the error to. Set to SUCCESS if there is no error.
+   */
+  simdjson_inline void tie(T &value, error_code &error) && noexcept;
+
+  /**
+   * Move the value to the provided variable.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   */
+  simdjson_warn_unused simdjson_inline error_code get(T &value) && noexcept;
+
+  /**
+   * The error.
+   */
+  simdjson_warn_unused simdjson_inline error_code error() const noexcept;
+
+  /**
+   * Whether there is a value.
+   */
+  simdjson_warn_unused simdjson_inline bool has_value() const noexcept;
+
+#if SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T& operator*() &  noexcept(false);
+  simdjson_inline T&& operator*() &&  noexcept(false);
+  /**
+   * Arrow operator to access members of the contained value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T* operator->() noexcept(false);
+  simdjson_inline const T* operator->() const noexcept(false);
+
+  simdjson_inline T& value() & noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& value() && noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& take_value() && noexcept(false);
+
+  /**
+   * Cast to the value (will throw on error).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline operator T&&() && noexcept(false);
+
+
+#endif // SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline const T& value_unsafe() const& noexcept;
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T& value_unsafe() & noexcept;
+  /**
+   * Take the result value (move it). This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T&& value_unsafe() && noexcept;
+
+  using value_type = T;
+  using error_type = error_code;
+
+protected:
+  /** users should never directly access first and second. **/
+  T first{}; /** Users should never directly access 'first'. **/
+  error_code second{UNINITIALIZED}; /** Users should never directly access 'second'. **/
+}; // struct implementation_simdjson_result_base
+
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+/* end file simdjson/generic/implementation_simdjson_result_base.h for icelake */
+/* including simdjson/generic/numberparsing.h for icelake: #include "simdjson/generic/numberparsing.h" */
+/* begin file simdjson/generic/numberparsing.h for icelake */
+#ifndef SIMDJSON_GENERIC_NUMBERPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_NUMBERPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <limits>
+#include <ostream>
+#include <cstring>
+
+namespace simdjson {
+namespace icelake {
+namespace numberparsing {
+
+#ifdef JSON_TEST_NUMBERS
+#define INVALID_NUMBER(SRC) (found_invalid_number((SRC)), NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (found_integer((VALUE), (SRC)), (WRITER).append_s64((VALUE)))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (found_unsigned_integer((VALUE), (SRC)), (WRITER).append_u64((VALUE)))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (found_float((VALUE), (SRC)), (WRITER).append_double((VALUE)))
+#define BIGINT_NUMBER(SRC) (found_invalid_number((SRC)), BIGINT_ERROR)
+#else
+#define INVALID_NUMBER(SRC) (NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (WRITER).append_s64((VALUE))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (WRITER).append_u64((VALUE))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (WRITER).append_double((VALUE))
+#define BIGINT_NUMBER(SRC) (BIGINT_ERROR)
+#endif
+
+namespace {
+
+// Convert a mantissa, an exponent and a sign bit into an ieee64 double.
+// The real_exponent needs to be in [0, 2046] (technically real_exponent = 2047 would be acceptable).
+// The mantissa should be in [0,1<<53). The bit at index (1ULL << 52) while be zeroed.
+simdjson_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
+    double d;
+    mantissa &= ~(1ULL << 52);
+    mantissa |= real_exponent << 52;
+    mantissa |= ((static_cast<uint64_t>(negative)) << 63);
+    std::memcpy(&d, &mantissa, sizeof(d));
+    return d;
+}
+
+// Attempts to compute i * 10^(power) exactly; and if "negative" is
+// true, negate the result.
+// This function will only work in some cases, when it does not work, success is
+// set to false. This should work *most of the time* (like 99% of the time).
+// We assume that power is in the [smallest_power,
+// largest_power] interval: the caller is responsible for this check.
+simdjson_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
+  // we start with a fast path
+  // It was described in
+  // Clinger WD. How to read floating point numbers accurately.
+  // ACM SIGPLAN Notices. 1990
+#ifndef FLT_EVAL_METHOD
+#error "FLT_EVAL_METHOD should be defined, please include cfloat."
+#endif
+#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
+  // We cannot be certain that x/y is rounded to nearest.
+  if (0 <= power && power <= 22 && i <= 9007199254740991)
+#else
+  if (-22 <= power && power <= 22 && i <= 9007199254740991)
+#endif
+  {
+    // convert the integer into a double. This is lossless since
+    // 0 <= i <= 2^53 - 1.
+    d = double(i);
+    //
+    // The general idea is as follows.
+    // If 0 <= s < 2^53 and if 10^0 <= p <= 10^22 then
+    // 1) Both s and p can be represented exactly as 64-bit floating-point
+    // values
+    // (binary64).
+    // 2) Because s and p can be represented exactly as floating-point values,
+    // then s * p
+    // and s / p will produce correctly rounded values.
+    //
+    if (power < 0) {
+      d = d / simdjson::internal::power_of_ten[-power];
+    } else {
+      d = d * simdjson::internal::power_of_ten[power];
+    }
+    if (negative) {
+      d = -d;
+    }
+    return true;
+  }
+  // When 22 < power && power <  22 + 16, we could
+  // hope for another, secondary fast path.  It was
+  // described by David M. Gay in  "Correctly rounded
+  // binary-decimal and decimal-binary conversions." (1990)
+  // If you need to compute i * 10^(22 + x) for x < 16,
+  // first compute i * 10^x, if you know that result is exact
+  // (e.g., when i * 10^x < 2^53),
+  // then you can still proceed and do (i * 10^x) * 10^22.
+  // Is this worth your time?
+  // You need  22 < power *and* power <  22 + 16 *and* (i * 10^(x-22) < 2^53)
+  // for this second fast path to work.
+  // If you you have 22 < power *and* power <  22 + 16, and then you
+  // optimistically compute "i * 10^(x-22)", there is still a chance that you
+  // have wasted your time if i * 10^(x-22) >= 2^53. It makes the use cases of
+  // this optimization maybe less common than we would like. Source:
+  // http://www.exploringbinary.com/fast-path-decimal-to-floating-point-conversion/
+  // also used in RapidJSON: https://rapidjson.org/strtod_8h_source.html
+
+  // The fast path has now failed, so we are failing back on the slower path.
+
+  // In the slow path, we need to adjust i so that it is > 1<<63 which is always
+  // possible, except if i == 0, so we handle i == 0 separately.
+  if(i == 0) {
+    d = negative ? -0.0 : 0.0;
+    return true;
+  }
+
+
+  // The exponent is 1024 + 63 + power
+  //     + floor(log(5**power)/log(2)).
+  // The 1024 comes from the ieee64 standard.
+  // The 63 comes from the fact that we use a 64-bit word.
+  //
+  // Computing floor(log(5**power)/log(2)) could be
+  // slow. Instead we use a fast function.
+  //
+  // For power in (-400,350), we have that
+  // (((152170 + 65536) * power ) >> 16);
+  // is equal to
+  //  floor(log(5**power)/log(2)) + power when power >= 0
+  // and it is equal to
+  //  ceil(log(5**-power)/log(2)) + power when power < 0
+  //
+  // The 65536 is (1<<16) and corresponds to
+  // (65536 * power) >> 16 ---> power
+  //
+  // ((152170 * power ) >> 16) is equal to
+  // floor(log(5**power)/log(2))
+  //
+  // Note that this is not magic: 152170/(1<<16) is
+  // approximately equal to log(5)/log(2).
+  // The 1<<16 value is a power of two; we could use a
+  // larger power of 2 if we wanted to.
+  //
+  int64_t exponent = (((152170 + 65536) * power) >> 16) + 1024 + 63;
+
+
+  // We want the most significant bit of i to be 1. Shift if needed.
+  int lz = leading_zeroes(i);
+  i <<= lz;
+
+
+  // We are going to need to do some 64-bit arithmetic to get a precise product.
+  // We use a table lookup approach.
+  // It is safe because
+  // power >= smallest_power
+  // and power <= largest_power
+  // We recover the mantissa of the power, it has a leading 1. It is always
+  // rounded down.
+  //
+  // We want the most significant 64 bits of the product. We know
+  // this will be non-zero because the most significant bit of i is
+  // 1.
+  const uint32_t index = 2 * uint32_t(power - simdjson::internal::smallest_power);
+  // Optimization: It may be that materializing the index as a variable might confuse some compilers and prevent effective complex-addressing loads. (Done for code clarity.)
+  //
+  // The full_multiplication function computes the 128-bit product of two 64-bit words
+  // with a returned value of type value128 with a "low component" corresponding to the
+  // 64-bit least significant bits of the product and with a "high component" corresponding
+  // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index]);
+#else
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index]);
+#endif
+
+  // Both i and power_of_five_128[index] have their most significant bit set to 1 which
+  // implies that the either the most or the second most significant bit of the product
+  // is 1. We pack values in this manner for efficiency reasons: it maximizes the use
+  // we make of the product. It also makes it easy to reason about the product: there
+  // is 0 or 1 leading zero in the product.
+
+  // Unless the least significant 9 bits of the high (64-bit) part of the full
+  // product are all 1s, then we know that the most significant 55 bits are
+  // exact and no further work is needed. Having 55 bits is necessary because
+  // we need 53 bits for the mantissa but we have to have one rounding bit and
+  // we can waste a bit if the most significant bit of the product is zero.
+  if((firstproduct.high & 0x1FF) == 0x1FF) {
+    // We want to compute i * 5^q, but only care about the top 55 bits at most.
+    // Consider the scenario where q>=0. Then 5^q may not fit in 64-bits. Doing
+    // the full computation is wasteful. So we do what is called a "truncated
+    // multiplication".
+    // We take the most significant 64-bits, and we put them in
+    // power_of_five_128[index]. Usually, that's good enough to approximate i * 5^q
+    // to the desired approximation using one multiplication. Sometimes it does not suffice.
+    // Then we store the next most significant 64 bits in power_of_five_128[index + 1], and
+    // then we get a better approximation to i * 5^q.
+    //
+    // That's for when q>=0. The logic for q<0 is somewhat similar but it is somewhat
+    // more complicated.
+    //
+    // There is an extra layer of complexity in that we need more than 55 bits of
+    // accuracy in the round-to-even scenario.
+    //
+    // The full_multiplication function computes the 128-bit product of two 64-bit words
+    // with a returned value of type value128 with a "low component" corresponding to the
+    // 64-bit least significant bits of the product and with a "high component" corresponding
+    // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index + 1]);
+#else
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index + 1]);
+#endif
+    firstproduct.low += secondproduct.high;
+    if(secondproduct.high > firstproduct.low) { firstproduct.high++; }
+    // As it has been proven by Noble Mushtak and Daniel Lemire in "Fast Number Parsing Without
+    // Fallback" (https://arxiv.org/abs/2212.06644), at this point we are sure that the product
+    // is sufficiently accurate, and more computation is not needed.
+  }
+  uint64_t lower = firstproduct.low;
+  uint64_t upper = firstproduct.high;
+  // The final mantissa should be 53 bits with a leading 1.
+  // We shift it so that it occupies 54 bits with a leading 1.
+  ///////
+  uint64_t upperbit = upper >> 63;
+  uint64_t mantissa = upper >> (upperbit + 9);
+  lz += int(1 ^ upperbit);
+
+  // Here we have mantissa < (1<<54).
+  int64_t real_exponent = exponent - lz;
+  if (simdjson_unlikely(real_exponent <= 0)) { // we have a subnormal?
+    // Here have that real_exponent <= 0 so -real_exponent >= 0
+    if(-real_exponent + 1 >= 64) { // if we have more than 64 bits below the minimum exponent, you have a zero for sure.
+      d = negative ? -0.0 : 0.0;
+      return true;
+    }
+    // next line is safe because -real_exponent + 1 < 0
+    mantissa >>= -real_exponent + 1;
+    // Thankfully, we can't have both "round-to-even" and subnormals because
+    // "round-to-even" only occurs for powers close to 0.
+    mantissa += (mantissa & 1); // round up
+    mantissa >>= 1;
+    // There is a weird scenario where we don't have a subnormal but just.
+    // Suppose we start with 2.2250738585072013e-308, we end up
+    // with 0x3fffffffffffff x 2^-1023-53 which is technically subnormal
+    // whereas 0x40000000000000 x 2^-1023-53  is normal. Now, we need to round
+    // up 0x3fffffffffffff x 2^-1023-53  and once we do, we are no longer
+    // subnormal, but we can only know this after rounding.
+    // So we only declare a subnormal if we are smaller than the threshold.
+    real_exponent = (mantissa < (uint64_t(1) << 52)) ? 0 : 1;
+    d = to_double(mantissa, real_exponent, negative);
+    return true;
+  }
+  // We have to round to even. The "to even" part
+  // is only a problem when we are right in between two floats
+  // which we guard against.
+  // If we have lots of trailing zeros, we may fall right between two
+  // floating-point values.
+  //
+  // The round-to-even cases take the form of a number 2m+1 which is in (2^53,2^54]
+  // times a power of two. That is, it is right between a number with binary significand
+  // m and another number with binary significand m+1; and it must be the case
+  // that it cannot be represented by a float itself.
+  //
+  // We must have that w * 10 ^q == (2m+1) * 2^p for some power of two 2^p.
+  // Recall that 10^q = 5^q * 2^q.
+  // When q >= 0, we must have that (2m+1) is divible by 5^q, so 5^q <= 2^54. We have that
+  //  5^23 <=  2^54 and it is the last power of five to qualify, so q <= 23.
+  // When q<0, we have  w  >=  (2m+1) x 5^{-q}.  We must have that w<2^{64} so
+  // (2m+1) x 5^{-q} < 2^{64}. We have that 2m+1>2^{53}. Hence, we must have
+  // 2^{53} x 5^{-q} < 2^{64}.
+  // Hence we have 5^{-q} < 2^{11}$ or q>= -4.
+  //
+  // We require lower <= 1 and not lower == 0 because we could not prove that
+  // that lower == 0 is implied; but we could prove that lower <= 1 is a necessary and sufficient test.
+  if (simdjson_unlikely((lower <= 1) && (power >= -4) && (power <= 23) && ((mantissa & 3) == 1))) {
+    if((mantissa  << (upperbit + 64 - 53 - 2)) ==  upper) {
+      mantissa &= ~1;             // flip it so that we do not round up
+    }
+  }
+
+  mantissa += mantissa & 1;
+  mantissa >>= 1;
+
+  // Here we have mantissa < (1<<53), unless there was an overflow
+  if (mantissa >= (1ULL << 53)) {
+    //////////
+    // This will happen when parsing values such as 7.2057594037927933e+16
+    ////////
+    mantissa = (1ULL << 52);
+    real_exponent++;
+  }
+  mantissa &= ~(1ULL << 52);
+  // we have to check that real_exponent is in range, otherwise we bail out
+  if (simdjson_unlikely(real_exponent > 2046)) {
+    // We have an infinite value!!! We could actually throw an error here if we could.
+    return false;
+  }
+  d = to_double(mantissa, real_exponent, negative);
+  return true;
+}
+
+// We call a fallback floating-point parser that might be slow. Note
+// it will accept JSON numbers, but the JSON spec. is more restrictive so
+// before you call parse_float_fallback, you need to have validated the input
+// string with the JSON grammar.
+// It will return an error (false) if the parsed number is infinite.
+// The string parsing itself always succeeds. We know that there is at least
+// one digit.
+static bool parse_float_fallback(const uint8_t *ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+static bool parse_float_fallback(const uint8_t *ptr, const uint8_t *end_ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr), reinterpret_cast<const char *>(end_ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+// check quickly whether the next 8 chars are made of digits
+// at a glance, it looks better than Mula's
+// http://0x80.pl/articles/swar-digits-validate.html
+simdjson_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
+  uint64_t val;
+  // this can read up to 7 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(7 <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be bigger than 7");
+  std::memcpy(&val, chars, 8);
+  // a branchy method might be faster:
+  // return (( val & 0xF0F0F0F0F0F0F0F0 ) == 0x3030303030303030)
+  //  && (( (val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0 ) ==
+  //  0x3030303030303030);
+  return (((val & 0xF0F0F0F0F0F0F0F0) |
+           (((val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0) >> 4)) ==
+          0x3333333333333333);
+}
+
+template<typename I>
+SIMDJSON_NO_SANITIZE_UNDEFINED // We deliberately allow overflow here and check later
+simdjson_inline bool parse_digit(const uint8_t c, I &i) {
+  const uint8_t digit = static_cast<uint8_t>(c - '0');
+  if (digit > 9) {
+    return false;
+  }
+  // PERF NOTE: multiplication by 10 is cheaper than arbitrary integer multiplication
+  i = 10 * i + digit; // might overflow, we will handle the overflow later
+  return true;
+}
+
+simdjson_inline bool is_digit(const uint8_t c) {
+  return static_cast<uint8_t>(c - '0') <= 9;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_decimal_after_separator(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
+  // we continue with the fiction that we have an integer. If the
+  // floating point number is representable as x * 10^z for some integer
+  // z that fits in 53 bits, then we will be able to convert back the
+  // the integer into a float in a lossless manner.
+  const uint8_t *const first_after_period = p;
+
+#ifdef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_SWAR_NUMBER_PARSING
+  // this helps if we have lots of decimals!
+  // this turns out to be frequent enough.
+  if (is_made_of_eight_digits_fast(p)) {
+    i = i * 100000000 + parse_eight_digits_unrolled(p);
+    p += 8;
+  }
+#endif // SIMDJSON_SWAR_NUMBER_PARSING
+#endif // #ifdef SIMDJSON_SWAR_NUMBER_PARSING
+  // Unrolling the first digit makes a small difference on some implementations (e.g. westmere)
+  if (parse_digit(*p, i)) { ++p; }
+  while (parse_digit(*p, i)) { p++; }
+  exponent = first_after_period - p;
+  // Decimal without digits (123.) is illegal
+  if (exponent == 0) {
+    return INVALID_NUMBER(src);
+  }
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
+  // Exp Sign: -123.456e[-]78
+  bool neg_exp = ('-' == *p);
+  if (neg_exp || '+' == *p) { p++; } // Skip + as well
+
+  // Exponent: -123.456e-[78]
+  auto start_exp = p;
+  int64_t exp_number = 0;
+  while (parse_digit(*p, exp_number)) { ++p; }
+  // It is possible for parse_digit to overflow.
+  // In particular, it could overflow to INT64_MIN, and we cannot do - INT64_MIN.
+  // Thus we *must* check for possible overflow before we negate exp_number.
+
+  // Performance notes: it may seem like combining the two "simdjson_unlikely checks" below into
+  // a single simdjson_unlikely path would be faster. The reasoning is sound, but the compiler may
+  // not oblige and may, in fact, generate two distinct paths in any case. It might be
+  // possible to do uint64_t(p - start_exp - 1) >= 18 but it could end up trading off
+  // instructions for a simdjson_likely branch, an unconclusive gain.
+
+  // If there were no digits, it's an error.
+  if (simdjson_unlikely(p == start_exp)) {
+    return INVALID_NUMBER(src);
+  }
+  // We have a valid positive exponent in exp_number at this point, except that
+  // it may have overflowed.
+
+  // If there were more than 18 digits, we may have overflowed the integer. We have to do
+  // something!!!!
+  if (simdjson_unlikely(p > start_exp+18)) {
+    // Skip leading zeroes: 1e000000000000000000001 is technically valid and does not overflow
+    while (*start_exp == '0') { start_exp++; }
+    // 19 digits could overflow int64_t and is kind of absurd anyway. We don't
+    // support exponents smaller than -999,999,999,999,999,999 and bigger
+    // than 999,999,999,999,999,999.
+    // We can truncate.
+    // Note that 999999999999999999 is assuredly too large. The maximal ieee64 value before
+    // infinity is ~1.8e308. The smallest subnormal is ~5e-324. So, actually, we could
+    // truncate at 324.
+    // Note that there is no reason to fail per se at this point in time.
+    // E.g., 0e999999999999999999999 is a fine number.
+    if (p > start_exp+18) { exp_number = 999999999999999999; }
+  }
+  // At this point, we know that exp_number is a sane, positive, signed integer.
+  // It is <= 999,999,999,999,999,999. As long as 'exponent' is in
+  // [-8223372036854775808, 8223372036854775808], we won't overflow. Because 'exponent'
+  // is bounded in magnitude by the size of the JSON input, we are fine in this universe.
+  // To sum it up: the next line should never overflow.
+  exponent += (neg_exp ? -exp_number : exp_number);
+  return SUCCESS;
+}
+
+simdjson_inline bool check_if_integer(const uint8_t *const src, size_t max_length) {
+  const uint8_t *const srcend = src + max_length;
+  bool negative = (*src == '-'); // we can always read at least one character after the '-'
+  const uint8_t *p = src + uint8_t(negative);
+  if(p == srcend) { return false; }
+  if(*p == '0') {
+    ++p;
+    if(p == srcend) { return true; }
+    if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+    return true;
+  }
+  while(p != srcend && is_digit(*p)) { ++p; }
+  if(p == srcend) { return true; }
+  if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+  return true;
+}
+
+simdjson_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
+  // It is possible that the integer had an overflow.
+  // We have to handle the case where we have 0.0000somenumber.
+  const uint8_t *start = start_digits;
+  while ((*start == '0') || (*start == '.')) { ++start; }
+  // we over-decrement by one when there is a '.'
+  return digit_count - size_t(start - start_digits);
+}
+
+} // unnamed namespace
+
+/** @private */
+static error_code slow_float_parsing(simdjson_unused const uint8_t * src, double* answer) {
+  if (parse_float_fallback(src, answer)) {
+    return SUCCESS;
+  }
+  return INVALID_NUMBER(src);
+}
+
+/** @private */
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
+  // If we frequently had to deal with long strings of digits,
+  // we could extend our code by using a 128-bit integer instead
+  // of a 64-bit integer. However, this is uncommon in practice.
+  //
+  // 9999999999999999999 < 2**64 so we can accommodate 19 digits.
+  // If we have a decimal separator, then digit_count - 1 is the number of digits, but we
+  // may not have a decimal separator!
+  if (simdjson_unlikely(digit_count > 19 && significant_digits(start_digits, digit_count) > 19)) {
+    // Ok, chances are good that we had an overflow!
+    // this is almost never going to get called!!!
+    // we start anew, going slowly!!!
+    // This will happen in the following examples:
+    // 10000000000000000000000000000000000000000000e+308
+    // 3.1415926535897932384626433832795028841971693993751
+    //
+    // NOTE: We do not pass a reference to the to slow_float_parsing. If we passed our writer
+    // reference to it, it would force it to be stored in memory, preventing the compiler from
+    // picking it apart and putting into registers. i.e. if we pass it as reference,
+    // it gets slow.
+    double d;
+    error_code error = slow_float_parsing(src, &d);
+    writer.append_double(d);
+    return error;
+  }
+  // NOTE: it's weird that the simdjson_unlikely() only wraps half the if, but it seems to get slower any other
+  // way we've tried: https://github.com/simdjson/simdjson/pull/990#discussion_r448497331
+  // To future reader: we'd love if someone found a better way, or at least could explain this result!
+  if (simdjson_unlikely(exponent < simdjson::internal::smallest_power) || (exponent > simdjson::internal::largest_power)) {
+    //
+    // Important: smallest_power is such that it leads to a zero value.
+    // Observe that 18446744073709551615e-343 == 0, i.e. (2**64 - 1) e -343 is zero
+    // so something x 10^-343 goes to zero, but not so with  something x 10^-342.
+    static_assert(simdjson::internal::smallest_power <= -342, "smallest_power is not small enough");
+    //
+    if((exponent < simdjson::internal::smallest_power) || (i == 0)) {
+      // E.g. Parse "-0.0e-999" into the same value as "-0.0". See https://en.wikipedia.org/wiki/Signed_zero
+      WRITE_DOUBLE(negative ? -0.0 : 0.0, src, writer);
+      return SUCCESS;
+    } else { // (exponent > largest_power) and (i != 0)
+      // We have, for sure, an infinite value and simdjson refuses to parse infinite values.
+      return INVALID_NUMBER(src);
+    }
+  }
+  double d;
+  if (!compute_float_64(exponent, i, negative, d)) {
+    // we are almost never going to get here.
+    if (!parse_float_fallback(src, &d)) { return INVALID_NUMBER(src); }
+  }
+  WRITE_DOUBLE(d, src, writer);
+  return SUCCESS;
+}
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer);
+
+// for performance analysis, it is sometimes  useful to skip parsing
+#ifdef SIMDJSON_SKIPNUMBERPARSING
+
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const, W &writer) {
+  writer.append_s64(0);        // always write zero
+  return SUCCESS;              // always succeeds
+}
+
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept { return number_type::signed_integer; }
+#else
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer) {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  if (digit_count == 0 || ('0' == *start_digits && digit_count > 1)) { return INVALID_NUMBER(src); }
+
+  //
+  // Handle floats if there is a . or e (or both)
+  //
+  int64_t exponent = 0;
+  bool is_float = false;
+  if ('.' == *p) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_decimal_after_separator(src, p, i, exponent) );
+    digit_count = int(p - start_digits); // used later to guard against overflows
+  }
+  if (('e' == *p) || ('E' == *p)) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_exponent(src, p, exponent) );
+  }
+  if (is_float) {
+    const bool dirty_end = jsoncharutils::is_not_structural_or_whitespace(*p);
+    SIMDJSON_TRY( write_float(src, negative, i, start_digits, digit_count, exponent, writer) );
+    if (dirty_end) { return INVALID_NUMBER(src); }
+    return SUCCESS;
+  }
+
+  // The longest negative 64-bit number is 19 digits.
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  size_t longest_digit_count = negative ? 19 : 20;
+  if (digit_count > longest_digit_count) { return BIGINT_NUMBER(src); }
+  if (digit_count == longest_digit_count) {
+    if (negative) {
+      // Anything negative above INT64_MAX+1 is invalid
+      if (i > uint64_t(INT64_MAX)+1) { return BIGINT_NUMBER(src);  }
+      WRITE_INTEGER(~i+1, src, writer);
+      if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+      return SUCCESS;
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    }  else if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INVALID_NUMBER(src); }
+  }
+
+  // Write unsigned if it does not fit in a signed integer.
+  if (i > uint64_t(INT64_MAX)) {
+    WRITE_UNSIGNED(i, src, writer);
+  } else {
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+    if(i == 0 && negative) {
+      // We have to write -0.0 instead of 0
+      WRITE_DOUBLE(-0.0, src, writer);
+    } else {
+      WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+    }
+#else
+  WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+#endif
+  }
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+  return SUCCESS;
+}
+
+// Inlineable functions
+namespace {
+
+// This table can be used to characterize the final character of an integer
+// string. For JSON structural character and allowable white space characters,
+// we return SUCCESS. For 'e', '.' and 'E', we return INCORRECT_TYPE. Otherwise
+// we return NUMBER_ERROR.
+// Optimization note: we could easily reduce the size of the table by half (to 128)
+// at the cost of an extra branch.
+// Optimization note: we want the values to use at most 8 bits (not, e.g., 32 bits):
+static_assert(error_code(uint8_t(NUMBER_ERROR))== NUMBER_ERROR, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(SUCCESS))== SUCCESS, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(INCORRECT_TYPE))== INCORRECT_TYPE, "bad NUMBER_ERROR cast");
+
+const uint8_t integer_string_finisher[256] = {
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   SUCCESS,      NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, INCORRECT_TYPE,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, SUCCESS,        NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR};
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if ((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
+  const uint8_t *p = src + 1;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    // Note: we use src[1] and not src[0] because src[0] is the quote character in this
+    // instance.
+    if (src[1] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  //
+  // Check for minus sign
+  //
+  if(src == src_end) { return NUMBER_ERROR; }
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = src;
+  uint64_t i = 0;
+  while (parse_digit(*src, i)) { src++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(src - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*src)) {
+  //  return (*src == '.' || *src == 'e' || *src == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(*src != '"') { return NUMBER_ERROR; }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept {
+  return (*src == '-');
+}
+
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) { return true; }
+  return false;
+}
+
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  size_t digit_count = size_t(p - src);
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) {
+    static const uint8_t * smaller_big_integer = reinterpret_cast<const uint8_t *>("9223372036854775808");
+    // We have an integer.
+    if(simdjson_unlikely(digit_count > 20)) {
+      return number_type::big_integer;
+    }
+    // If the number is negative and valid, it must be a signed integer.
+    if(negative) {
+      if (simdjson_unlikely(digit_count > 19)) return number_type::big_integer;
+      if (simdjson_unlikely(digit_count == 19 && memcmp(src, smaller_big_integer, 19) > 0)) {
+        return number_type::big_integer;
+      }
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+      if(digit_count == 1 && src[0] == '0') {
+        // We have to write -0.0 instead of 0
+        return number_type::floating_point_number;
+      }
+#endif
+      return number_type::signed_integer;
+    }
+    // Let us check if we have a big integer (>=2**64).
+    static const uint8_t * two_to_sixtyfour = reinterpret_cast<const uint8_t *>("18446744073709551616");
+    if((digit_count > 20) || (digit_count == 20 && memcmp(src, two_to_sixtyfour, 20) >= 0)) {
+      return number_type::big_integer;
+    }
+    // The number is positive and smaller than 18446744073709551616 (or 2**64).
+    // We want values larger or equal to 9223372036854775808 to be unsigned
+    // integers, and the other values to be signed integers.
+    if((digit_count == 20) || (digit_count >= 19 && memcmp(src, smaller_big_integer, 19) >= 0)) {
+      return number_type::unsigned_integer;
+    }
+    return number_type::signed_integer;
+  }
+  // Hopefully, we have 'e' or 'E' or '.'.
+  return number_type::floating_point_number;
+}
+
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
+  if(src == src_end) { return NUMBER_ERROR; }
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  if(p == src_end) { return NUMBER_ERROR; }
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely((p != src_end) && (*p == '.'))) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if ((p == src_end) || !parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while ((p != src_end) && parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = start_digits-src > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if ((p != src_end) && (*p == 'e' || *p == 'E')) {
+    p++;
+    if(p == src_end) { return NUMBER_ERROR; }
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while ((p != src_end) && parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if ((p != src_end) && jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), src_end, &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+} // unnamed namespace
+#endif // SIMDJSON_SKIPNUMBERPARSING
+
+} // namespace numberparsing
+
+inline std::ostream& operator<<(std::ostream& out, number_type type) noexcept {
+    switch (type) {
+        case number_type::signed_integer: out << "integer in [-9223372036854775808,9223372036854775808)"; break;
+        case number_type::unsigned_integer: out << "unsigned integer in [9223372036854775808,18446744073709551616)"; break;
+        case number_type::floating_point_number: out << "floating-point number (binary64)"; break;
+        case number_type::big_integer: out << "big integer"; break;
+        default: SIMDJSON_UNREACHABLE();
+    }
+    return out;
+}
+
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_NUMBERPARSING_H
+/* end file simdjson/generic/numberparsing.h for icelake */
+
+/* including simdjson/generic/implementation_simdjson_result_base-inl.h for icelake: #include "simdjson/generic/implementation_simdjson_result_base-inl.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base-inl.h for icelake */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+
+//
+// internal::implementation_simdjson_result_base<T> inline implementation
+//
+
+template<typename T>
+simdjson_inline void implementation_simdjson_result_base<T>::tie(T &value, error_code &error) && noexcept {
+  error = this->second;
+  if (!error) {
+    value = std::forward<implementation_simdjson_result_base<T>>(*this).first;
+  }
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::get(T &value) && noexcept {
+  error_code error;
+  std::forward<implementation_simdjson_result_base<T>>(*this).tie(value, error);
+  return error;
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::error() const noexcept {
+  return this->second;
+}
+
+
+template<typename T>
+simdjson_warn_unused simdjson_inline bool implementation_simdjson_result_base<T>::has_value() const noexcept {
+  return this->error() == SUCCESS;
+}
+
+#if SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::operator*() &  noexcept(false) {
+  return this->value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::operator*() &&  noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).value();
+}
+
+template<typename T>
+simdjson_inline T* implementation_simdjson_result_base<T>::operator->() noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+
+template<typename T>
+simdjson_inline const T* implementation_simdjson_result_base<T>::operator->() const noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value() & noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::take_value() && noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::operator T&&() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+#endif // SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline const T& implementation_simdjson_result_base<T>::value_unsafe() const& noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value_unsafe() & noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value_unsafe() && noexcept {
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value, error_code error) noexcept
+    : first{std::forward<T>(value)}, second{error} {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(error_code error) noexcept
+    : implementation_simdjson_result_base(T{}, error) {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value) noexcept
+    : implementation_simdjson_result_base(std::forward<T>(value), SUCCESS) {}
+
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+/* end file simdjson/generic/implementation_simdjson_result_base-inl.h for icelake */
+/* end file simdjson/generic/amalgamated.h for icelake */
+/* including simdjson/icelake/end.h: #include "simdjson/icelake/end.h" */
+/* begin file simdjson/icelake/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#if !SIMDJSON_CAN_ALWAYS_RUN_ICELAKE
+SIMDJSON_UNTARGET_REGION
+#endif
+
+/* undefining SIMDJSON_IMPLEMENTATION from "icelake" */
+#undef SIMDJSON_IMPLEMENTATION
+/* end file simdjson/icelake/end.h */
+
+#endif // SIMDJSON_ICELAKE_H
+/* end file simdjson/icelake.h */
+/* including simdjson/icelake/implementation.h: #include <simdjson/icelake/implementation.h> */
+/* begin file simdjson/icelake/implementation.h */
+#ifndef SIMDJSON_ICELAKE_IMPLEMENTATION_H
+#define SIMDJSON_ICELAKE_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/instruction_set.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// The constructor may be executed on any host, so we take care not to use SIMDJSON_TARGET_ICELAKE
+namespace simdjson {
+namespace icelake {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation(
+      "icelake",
+      "Intel/AMD AVX512",
+      internal::instruction_set::AVX2 | internal::instruction_set::PCLMULQDQ | internal::instruction_set::BMI1 | internal::instruction_set::BMI2 | internal::instruction_set::AVX512F | internal::instruction_set::AVX512DQ | internal::instruction_set::AVX512CD | internal::instruction_set::AVX512BW | internal::instruction_set::AVX512VL | internal::instruction_set::AVX512VBMI2
+  ) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_ICELAKE_IMPLEMENTATION_H
+/* end file simdjson/icelake/implementation.h */
+
+// defining SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER allows us to provide our own bit_indexer::write
+#define SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+/* including simdjson/icelake/begin.h: #include <simdjson/icelake/begin.h> */
+/* begin file simdjson/icelake/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "icelake" */
+#define SIMDJSON_IMPLEMENTATION icelake
+/* including simdjson/icelake/base.h: #include "simdjson/icelake/base.h" */
+/* begin file simdjson/icelake/base.h */
+#ifndef SIMDJSON_ICELAKE_BASE_H
+#define SIMDJSON_ICELAKE_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// The constructor may be executed on any host, so we take care not to use SIMDJSON_TARGET_ICELAKE
+namespace simdjson {
+/**
+ * Implementation for Icelake (Intel AVX512).
+ */
+namespace icelake {
+
+class implementation;
+
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_ICELAKE_BASE_H
+/* end file simdjson/icelake/base.h */
+/* including simdjson/icelake/intrinsics.h: #include "simdjson/icelake/intrinsics.h" */
+/* begin file simdjson/icelake/intrinsics.h */
+#ifndef SIMDJSON_ICELAKE_INTRINSICS_H
+#define SIMDJSON_ICELAKE_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#if SIMDJSON_VISUAL_STUDIO
+// under clang within visual studio, this will include <x86intrin.h>
+#include <intrin.h>  // visual studio or clang
+#else
+#include <x86intrin.h> // elsewhere
+#endif // SIMDJSON_VISUAL_STUDIO
+
+#if SIMDJSON_CLANG_VISUAL_STUDIO
+/**
+ * You are not supposed, normally, to include these
+ * headers directly. Instead you should either include intrin.h
+ * or x86intrin.h. However, when compiling with clang
+ * under Windows (i.e., when _MSC_VER is set), these headers
+ * only get included *if* the corresponding features are detected
+ * from macros:
+ * e.g., if __AVX2__ is set... in turn,  we normally set these
+ * macros by compiling against the corresponding architecture
+ * (e.g., arch:AVX2, -mavx2, etc.) which compiles the whole
+ * software with these advanced instructions. In simdjson, we
+ * want to compile the whole program for a generic target,
+ * and only target our specific kernels. As a workaround,
+ * we directly include the needed headers. These headers would
+ * normally guard against such usage, but we carefully included
+ * <x86intrin.h>  (or <intrin.h>) before, so the headers
+ * are fooled.
+ */
+#include <bmiintrin.h>   // for _blsr_u64
+#include <lzcntintrin.h> // for  __lzcnt64
+#include <immintrin.h>   // for most things (AVX2, AVX512, _popcnt64)
+#include <smmintrin.h>
+#include <tmmintrin.h>
+#include <avxintrin.h>
+#include <avx2intrin.h>
+#include <wmmintrin.h>   // for  _mm_clmulepi64_si128
+// Important: we need the AVX-512 headers:
+#include <avx512fintrin.h>
+#include <avx512dqintrin.h>
+#include <avx512cdintrin.h>
+#include <avx512bwintrin.h>
+#include <avx512vlintrin.h>
+#include <avx512vbmiintrin.h>
+#include <avx512vbmi2intrin.h>
+// unfortunately, we may not get _blsr_u64, but, thankfully, clang
+// has it as a macro.
+#ifndef _blsr_u64
+// we roll our own
+#define _blsr_u64(n) ((n - 1) & n)
+#endif //  _blsr_u64
+#endif // SIMDJSON_CLANG_VISUAL_STUDIO
+
+static_assert(sizeof(__m512i) <= simdjson::SIMDJSON_PADDING, "insufficient padding for icelake");
+
+#endif // SIMDJSON_ICELAKE_INTRINSICS_H
+/* end file simdjson/icelake/intrinsics.h */
+
+#if !SIMDJSON_CAN_ALWAYS_RUN_ICELAKE
+SIMDJSON_TARGET_REGION("avx512f,avx512dq,avx512cd,avx512bw,avx512vbmi,avx512vbmi2,avx512vl,avx2,bmi,pclmul,lzcnt,popcnt")
+#endif
+
+/* including simdjson/icelake/bitmanipulation.h: #include "simdjson/icelake/bitmanipulation.h" */
+/* begin file simdjson/icelake/bitmanipulation.h */
+#ifndef SIMDJSON_ICELAKE_BITMANIPULATION_H
+#define SIMDJSON_ICELAKE_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  return (int)_tzcnt_u64(input_num);
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO
+  ////////
+  // You might expect the next line to be equivalent to
+  // return (int)_tzcnt_u64(input_num);
+  // but the generated code differs and might be less efficient?
+  ////////
+  return __builtin_ctzll(input_num);
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return _blsr_u64(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+  return int(_lzcnt_u64(input_num));
+}
+
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+simdjson_inline unsigned __int64 count_ones(uint64_t input_num) {
+  // note: we do not support legacy 32-bit Windows
+  return __popcnt64(input_num);// Visual Studio wants two underscores
+}
+#else
+simdjson_inline long long int count_ones(uint64_t input_num) {
+  return _popcnt64(input_num);
+}
+#endif
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2,
+                                uint64_t *result) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  return _addcarry_u64(0, value1, value2,
+                       reinterpret_cast<unsigned __int64 *>(result));
+#else
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+#endif
+}
+
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_ICELAKE_BITMANIPULATION_H
+/* end file simdjson/icelake/bitmanipulation.h */
+/* including simdjson/icelake/bitmask.h: #include "simdjson/icelake/bitmask.h" */
+/* begin file simdjson/icelake/bitmask.h */
+#ifndef SIMDJSON_ICELAKE_BITMASK_H
+#define SIMDJSON_ICELAKE_BITMASK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(const uint64_t bitmask) {
+  // There should be no such thing with a processor supporting avx2
+  // but not clmul.
+  __m128i all_ones = _mm_set1_epi8('\xFF');
+  __m128i result = _mm_clmulepi64_si128(_mm_set_epi64x(0ULL, bitmask), all_ones, 0);
+  return _mm_cvtsi128_si64(result);
+}
+
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_ICELAKE_BITMASK_H
+/* end file simdjson/icelake/bitmask.h */
+/* including simdjson/icelake/simd.h: #include "simdjson/icelake/simd.h" */
+/* begin file simdjson/icelake/simd.h */
+#ifndef SIMDJSON_ICELAKE_SIMD_H
+#define SIMDJSON_ICELAKE_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#if defined(__GNUC__) && !defined(__clang__)
+#if __GNUC__ == 8
+#define SIMDJSON_GCC8 1
+#endif //  __GNUC__ == 8
+#endif // defined(__GNUC__) && !defined(__clang__)
+
+#if SIMDJSON_GCC8
+/**
+ * GCC 8 fails to provide _mm512_set_epi8. We roll our own.
+ */
+inline __m512i _mm512_set_epi8(uint8_t a0, uint8_t a1, uint8_t a2, uint8_t a3, uint8_t a4, uint8_t a5, uint8_t a6, uint8_t a7, uint8_t a8, uint8_t a9, uint8_t a10, uint8_t a11, uint8_t a12, uint8_t a13, uint8_t a14, uint8_t a15, uint8_t a16, uint8_t a17, uint8_t a18, uint8_t a19, uint8_t a20, uint8_t a21, uint8_t a22, uint8_t a23, uint8_t a24, uint8_t a25, uint8_t a26, uint8_t a27, uint8_t a28, uint8_t a29, uint8_t a30, uint8_t a31, uint8_t a32, uint8_t a33, uint8_t a34, uint8_t a35, uint8_t a36, uint8_t a37, uint8_t a38, uint8_t a39, uint8_t a40, uint8_t a41, uint8_t a42, uint8_t a43, uint8_t a44, uint8_t a45, uint8_t a46, uint8_t a47, uint8_t a48, uint8_t a49, uint8_t a50, uint8_t a51, uint8_t a52, uint8_t a53, uint8_t a54, uint8_t a55, uint8_t a56, uint8_t a57, uint8_t a58, uint8_t a59, uint8_t a60, uint8_t a61, uint8_t a62, uint8_t a63) {
+  return _mm512_set_epi64(uint64_t(a7) + (uint64_t(a6) << 8) + (uint64_t(a5) << 16) + (uint64_t(a4) << 24) + (uint64_t(a3) << 32) + (uint64_t(a2) << 40) + (uint64_t(a1) << 48) + (uint64_t(a0) << 56),
+                          uint64_t(a15) + (uint64_t(a14) << 8) + (uint64_t(a13) << 16) + (uint64_t(a12) << 24) + (uint64_t(a11) << 32) + (uint64_t(a10) << 40) + (uint64_t(a9) << 48) + (uint64_t(a8) << 56),
+                          uint64_t(a23) + (uint64_t(a22) << 8) + (uint64_t(a21) << 16) + (uint64_t(a20) << 24) + (uint64_t(a19) << 32) + (uint64_t(a18) << 40) + (uint64_t(a17) << 48) + (uint64_t(a16) << 56),
+                          uint64_t(a31) + (uint64_t(a30) << 8) + (uint64_t(a29) << 16) + (uint64_t(a28) << 24) + (uint64_t(a27) << 32) + (uint64_t(a26) << 40) + (uint64_t(a25) << 48) + (uint64_t(a24) << 56),
+                          uint64_t(a39) + (uint64_t(a38) << 8) + (uint64_t(a37) << 16) + (uint64_t(a36) << 24) + (uint64_t(a35) << 32) + (uint64_t(a34) << 40) + (uint64_t(a33) << 48) + (uint64_t(a32) << 56),
+                          uint64_t(a47) + (uint64_t(a46) << 8) + (uint64_t(a45) << 16) + (uint64_t(a44) << 24) + (uint64_t(a43) << 32) + (uint64_t(a42) << 40) + (uint64_t(a41) << 48) + (uint64_t(a40) << 56),
+                          uint64_t(a55) + (uint64_t(a54) << 8) + (uint64_t(a53) << 16) + (uint64_t(a52) << 24) + (uint64_t(a51) << 32) + (uint64_t(a50) << 40) + (uint64_t(a49) << 48) + (uint64_t(a48) << 56),
+                          uint64_t(a63) + (uint64_t(a62) << 8) + (uint64_t(a61) << 16) + (uint64_t(a60) << 24) + (uint64_t(a59) << 32) + (uint64_t(a58) << 40) + (uint64_t(a57) << 48) + (uint64_t(a56) << 56));
+}
+#endif // SIMDJSON_GCC8
+
+
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace simd {
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename Child>
+  struct base {
+    __m512i value;
+
+    // Zero constructor
+    simdjson_inline base() : value{__m512i()} {}
+
+    // Conversion from SIMD register
+    simdjson_inline base(const __m512i _value) : value(_value) {}
+
+    // Conversion to SIMD register
+    simdjson_inline operator const __m512i&() const { return this->value; }
+    simdjson_inline operator __m512i&() { return this->value; }
+
+    // Bit operations
+    simdjson_inline Child operator|(const Child other) const { return _mm512_or_si512(*this, other); }
+    simdjson_inline Child operator&(const Child other) const { return _mm512_and_si512(*this, other); }
+    simdjson_inline Child operator^(const Child other) const { return _mm512_xor_si512(*this, other); }
+    simdjson_inline Child bit_andnot(const Child other) const { return _mm512_andnot_si512(other, *this); }
+    simdjson_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+  };
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename T>
+  struct simd8;
+
+  template<typename T, typename Mask=simd8<bool>>
+  struct base8: base<simd8<T>> {
+    typedef uint32_t bitmask_t;
+    typedef uint64_t bitmask2_t;
+
+    simdjson_inline base8() : base<simd8<T>>() {}
+    simdjson_inline base8(const __m512i _value) : base<simd8<T>>(_value) {}
+
+    friend simdjson_really_inline uint64_t operator==(const simd8<T> lhs, const simd8<T> rhs) {
+      return _mm512_cmpeq_epi8_mask(lhs, rhs);
+    }
+    static const int SIZE = sizeof(base<T>::value);
+
+    template<int N=1>
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+     // workaround for compilers unable to figure out that 16 - N is a constant (GCC 8)
+      constexpr int shift = 16 - N;
+      return _mm512_alignr_epi8(*this, _mm512_permutex2var_epi64(prev_chunk, _mm512_set_epi64(13, 12, 11, 10, 9, 8, 7, 6), *this), shift);
+    }
+  };
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool>: base8<bool> {
+    static simdjson_inline simd8<bool> splat(bool _value) { return _mm512_set1_epi8(uint8_t(-(!!_value))); }
+
+    simdjson_inline simd8() : base8() {}
+    simdjson_inline simd8(const __m512i _value) : base8<bool>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+    simdjson_inline bool any() const { return !!_mm512_test_epi8_mask (*this, *this); }
+    simdjson_inline simd8<bool> operator~() const { return *this ^ true; }
+  };
+
+  template<typename T>
+  struct base8_numeric: base8<T> {
+    static simdjson_inline simd8<T> splat(T _value) { return _mm512_set1_epi8(_value); }
+    static simdjson_inline simd8<T> zero() { return _mm512_setzero_si512(); }
+    static simdjson_inline simd8<T> load(const T values[64]) {
+      return _mm512_loadu_si512(reinterpret_cast<const __m512i *>(values));
+    }
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    static simdjson_inline simd8<T> repeat_16(
+      T v0,  T v1,  T v2,  T v3,  T v4,  T v5,  T v6,  T v7,
+      T v8,  T v9,  T v10, T v11, T v12, T v13, T v14, T v15
+    ) {
+      return simd8<T>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    simdjson_inline base8_numeric() : base8<T>() {}
+    simdjson_inline base8_numeric(const __m512i _value) : base8<T>(_value) {}
+
+    // Store to array
+    simdjson_inline void store(T dst[64]) const { return _mm512_storeu_si512(reinterpret_cast<__m512i *>(dst), *this); }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<T> operator+(const simd8<T> other) const { return _mm512_add_epi8(*this, other); }
+    simdjson_inline simd8<T> operator-(const simd8<T> other) const { return _mm512_sub_epi8(*this, other); }
+    simdjson_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
+
+    // Override to distinguish from bool version
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return _mm512_shuffle_epi8(lookup_table, *this);
+    }
+
+    // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted as a bitset).
+    // Passing a 0 value for mask would be equivalent to writing out every byte to output.
+    // Only the first 64 - count_ones(mask) bytes of the result are significant but 64 bytes
+    // get written.
+    // Design consideration: it seems like a function with the
+    // signature simd8<L> compress(uint32_t mask) would be
+    // sensible, but the AVX ISA makes this kind of approach difficult.
+    template<typename L>
+    simdjson_inline void compress(uint64_t mask, L * output) const {
+      // we deliberately avoid _mm512_mask_compressstoreu_epi8 for portability
+      // (AMD Zen4 has terrible performance with it, it is effectively broken)
+      // _mm512_mask_compressstoreu_epi8 (output,~mask,*this);
+      __m512i compressed = _mm512_maskz_compress_epi8(~mask, *this);
+      _mm512_storeu_si512(output, compressed); // could use a mask
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> : base8_numeric<int8_t> {
+    simdjson_inline simd8() : base8_numeric<int8_t>() {}
+    simdjson_inline simd8(const __m512i _value) : base8_numeric<int8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const int8_t values[64]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15,
+      int8_t v16, int8_t v17, int8_t v18, int8_t v19, int8_t v20, int8_t v21, int8_t v22, int8_t v23,
+      int8_t v24, int8_t v25, int8_t v26, int8_t v27, int8_t v28, int8_t v29, int8_t v30, int8_t v31,
+      int8_t v32, int8_t v33, int8_t v34, int8_t v35, int8_t v36, int8_t v37, int8_t v38, int8_t v39,
+      int8_t v40, int8_t v41, int8_t v42, int8_t v43, int8_t v44, int8_t v45, int8_t v46, int8_t v47,
+      int8_t v48, int8_t v49, int8_t v50, int8_t v51, int8_t v52, int8_t v53, int8_t v54, int8_t v55,
+      int8_t v56, int8_t v57, int8_t v58, int8_t v59, int8_t v60, int8_t v61, int8_t v62, int8_t v63
+    ) : simd8(_mm512_set_epi8(
+      v63, v62, v61, v60, v59, v58, v57, v56,
+      v55, v54, v53, v52, v51, v50, v49, v48,
+      v47, v46, v45, v44, v43, v42, v41, v40,
+      v39, v38, v37, v36, v35, v34, v33, v32,
+      v31, v30, v29, v28, v27, v26, v25, v24,
+      v23, v22, v21, v20, v19, v18, v17, v16,
+      v15, v14, v13, v12, v11, v10,  v9,  v8,
+       v7,  v6,  v5,  v4,  v3,  v2,  v1,  v0
+    )) {}
+
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm512_max_epi8(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm512_min_epi8(*this, other); }
+
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm512_maskz_abs_epi8(_mm512_cmpgt_epi8_mask(*this, other),_mm512_set1_epi8(uint8_t(0x80))); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return _mm512_maskz_abs_epi8(_mm512_cmpgt_epi8_mask(other, *this),_mm512_set1_epi8(uint8_t(0x80))); }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t>: base8_numeric<uint8_t> {
+    simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+    simdjson_inline simd8(const __m512i _value) : base8_numeric<uint8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const uint8_t values[64]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15,
+      uint8_t v16, uint8_t v17, uint8_t v18, uint8_t v19, uint8_t v20, uint8_t v21, uint8_t v22, uint8_t v23,
+      uint8_t v24, uint8_t v25, uint8_t v26, uint8_t v27, uint8_t v28, uint8_t v29, uint8_t v30, uint8_t v31,
+      uint8_t v32, uint8_t v33, uint8_t v34, uint8_t v35, uint8_t v36, uint8_t v37, uint8_t v38, uint8_t v39,
+      uint8_t v40, uint8_t v41, uint8_t v42, uint8_t v43, uint8_t v44, uint8_t v45, uint8_t v46, uint8_t v47,
+      uint8_t v48, uint8_t v49, uint8_t v50, uint8_t v51, uint8_t v52, uint8_t v53, uint8_t v54, uint8_t v55,
+      uint8_t v56, uint8_t v57, uint8_t v58, uint8_t v59, uint8_t v60, uint8_t v61, uint8_t v62, uint8_t v63
+    ) : simd8(_mm512_set_epi8(
+      v63, v62, v61, v60, v59, v58, v57, v56,
+      v55, v54, v53, v52, v51, v50, v49, v48,
+      v47, v46, v45, v44, v43, v42, v41, v40,
+      v39, v38, v37, v36, v35, v34, v33, v32,
+      v31, v30, v29, v28, v27, v26, v25, v24,
+      v23, v22, v21, v20, v19, v18, v17, v16,
+      v15, v14, v13, v12, v11, v10,  v9,  v8,
+       v7,  v6,  v5,  v4,  v3,  v2,  v1,  v0
+    )) {}
+
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm512_adds_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm512_subs_epu8(*this, other); }
+
+    // Order-specific operations
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm512_max_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return _mm512_min_epu8(other, *this); }
+    // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
+    // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
+    simdjson_inline uint64_t operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
+    simdjson_inline uint64_t operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->lt_bits(other).any_bits_set(); }
+
+    // Bit-specific operations
+    simdjson_inline simd8<bool> bits_not_set() const { return _mm512_mask_blend_epi8(*this == uint8_t(0), _mm512_set1_epi8(0), _mm512_set1_epi8(-1)); }
+    simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
+
+    simdjson_inline bool is_ascii() const { return _mm512_movepi8_mask(*this) == 0; }
+    simdjson_inline bool bits_not_set_anywhere() const {
+      return !_mm512_test_epi8_mask(*this, *this);
+    }
+    simdjson_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
+    simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return !_mm512_test_epi8_mask(*this, bits); }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm512_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(_mm512_slli_epi16(*this, N)) & uint8_t(0xFFu << N); }
+    // Get one of the bits and make a bitmask out of it.
+    // e.g. value.get_bit<7>() gets the high bit
+    template<int N>
+    simdjson_inline uint64_t get_bit() const { return _mm512_movepi8_mask(_mm512_slli_epi16(*this, 7-N)); }
+  };
+
+  template<typename T>
+  struct simd8x64 {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+    static_assert(NUM_CHUNKS == 1, "Icelake kernel should use one register per 64-byte block.");
+    const simd8<T> chunks[NUM_CHUNKS];
+    template<int idx> simd8<uint8_t> get() const { return idx < NUM_CHUNKS ? chunks[idx] : simd8<T>(); }
+
+    simd8x64(const simd8x64<T>& o) = delete; // no copy allowed
+    simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1) : chunks{chunk0, chunk1} {}
+    simdjson_inline simd8x64(const simd8<T> chunk0) : chunks{chunk0} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr)} {}
+
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
+      this->chunks[0].compress(mask, output);
+      return 64 - count_ones(mask);
+    }
+
+    simdjson_inline void store(T ptr[64]) const {
+      this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
+    }
+
+    simdjson_inline simd8<T> reduce_or() const {
+      return this->chunks[0];
+    }
+
+    simdjson_inline simd8x64<T> bit_or(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return simd8x64<T>(
+        this->chunks[0] | mask
+      );
+    }
+
+    simdjson_inline uint64_t eq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return this->chunks[0] == mask;
+    }
+
+    simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+      return this->chunks[0] == other.chunks[0];
+    }
+
+    simdjson_inline uint64_t lteq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return this->chunks[0] <= mask;
+    }
+  }; // struct simd8x64<T>
+
+} // namespace simd
+
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_ICELAKE_SIMD_H
+/* end file simdjson/icelake/simd.h */
+/* including simdjson/icelake/stringparsing_defs.h: #include "simdjson/icelake/stringparsing_defs.h" */
+/* begin file simdjson/icelake/stringparsing_defs.h */
+#ifndef SIMDJSON_ICELAKE_STRINGPARSING_DEFS_H
+#define SIMDJSON_ICELAKE_STRINGPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/simd.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 64;
+  simdjson_inline backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return ((quote_bits - 1) & bs_bits) != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+
+  uint64_t bs_bits;
+  uint64_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // this can read up to 15 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  // store to dest unconditionally - we can overwrite the bits we don't like later
+  v.store(dst);
+  return {
+      static_cast<uint64_t>(v == '\\'), // bs_bits
+      static_cast<uint64_t>(v == '"'), // quote_bits
+  };
+}
+
+
+
+struct escaping {
+  static constexpr uint32_t BYTES_PROCESSED = 64;
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits != 0; }
+  simdjson_inline int escape_index() { return trailing_zeroes(uint64_t(escape_bits)); }
+
+  __mmask64 escape_bits;
+}; // struct escaping
+
+
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "escaping finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  __mmask64 is_quote = _mm512_cmpeq_epi8_mask(v, _mm512_set1_epi8('"'));
+  __mmask64 is_backslash = _mm512_cmpeq_epi8_mask(v, _mm512_set1_epi8('\\'));
+  __mmask64 is_control = _mm512_cmplt_epi8_mask(v, _mm512_set1_epi8(32));
+  return {
+    (is_backslash | is_quote | is_control)
+  };
+}
+
+
+
+
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_ICELAKE_STRINGPARSING_DEFS_H
+/* end file simdjson/icelake/stringparsing_defs.h */
+/* including simdjson/icelake/numberparsing_defs.h: #include "simdjson/icelake/numberparsing_defs.h" */
+/* begin file simdjson/icelake/numberparsing_defs.h */
+#ifndef SIMDJSON_ICELAKE_NUMBERPARSING_DEFS_H
+#define SIMDJSON_ICELAKE_NUMBERPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace numberparsing {
+
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  // this actually computes *16* values so we are being wasteful.
+  const __m128i ascii0 = _mm_set1_epi8('0');
+  const __m128i mul_1_10 =
+      _mm_setr_epi8(10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1);
+  const __m128i mul_1_100 = _mm_setr_epi16(100, 1, 100, 1, 100, 1, 100, 1);
+  const __m128i mul_1_10000 =
+      _mm_setr_epi16(10000, 1, 10000, 1, 10000, 1, 10000, 1);
+  const __m128i input = _mm_sub_epi8(
+      _mm_loadu_si128(reinterpret_cast<const __m128i *>(chars)), ascii0);
+  const __m128i t1 = _mm_maddubs_epi16(input, mul_1_10);
+  const __m128i t2 = _mm_madd_epi16(t1, mul_1_100);
+  const __m128i t3 = _mm_packus_epi32(t2, t2);
+  const __m128i t4 = _mm_madd_epi16(t3, mul_1_10000);
+  return _mm_cvtsi128_si32(
+      t4); // only captures the sum of the first 8 digits, drop the rest
+}
+
+/** @private */
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+#if SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+#if SIMDJSON_IS_ARM64
+  // ARM64 has native support for 64-bit multiplications, no need to emultate
+  answer.high = __umulh(value1, value2);
+  answer.low = value1 * value2;
+#else
+  answer.low = _umul128(value1, value2, &answer.high); // _umul128 not available on ARM64
+#endif // SIMDJSON_IS_ARM64
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+#endif
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace icelake
+} // namespace simdjson
+
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+
+#endif // SIMDJSON_ICELAKE_NUMBERPARSING_DEFS_H
+/* end file simdjson/icelake/numberparsing_defs.h */
+/* end file simdjson/icelake/begin.h */
+/* including generic/amalgamated.h for icelake: #include <generic/amalgamated.h> */
+/* begin file generic/amalgamated.h for icelake */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_SRC_GENERIC_DEPENDENCIES_H)
+#error generic/dependencies.h must be included before generic/amalgamated.h!
+#endif
+
+/* including generic/base.h for icelake: #include <generic/base.h> */
+/* begin file generic/base.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+
+struct json_character_block;
+
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_BASE_H
+/* end file generic/base.h for icelake */
+/* including generic/dom_parser_implementation.h for icelake: #include <generic/dom_parser_implementation.h> */
+/* begin file generic/dom_parser_implementation.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// Interface a dom parser implementation must fulfill
+namespace simdjson {
+namespace icelake {
+namespace {
+
+simdjson_inline simd8<uint8_t> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3);
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input);
+
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file generic/dom_parser_implementation.h for icelake */
+/* including generic/json_character_block.h for icelake: #include <generic/json_character_block.h> */
+/* begin file generic/json_character_block.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+
+struct json_character_block {
+  static simdjson_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
+
+  simdjson_inline uint64_t whitespace() const noexcept { return _whitespace; }
+  simdjson_inline uint64_t op() const noexcept { return _op; }
+  simdjson_inline uint64_t scalar() const noexcept { return ~(op() | whitespace()); }
+
+  uint64_t _whitespace;
+  uint64_t _op;
+};
+
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H
+/* end file generic/json_character_block.h for icelake */
+/* end file generic/amalgamated.h for icelake */
+/* including generic/stage1/amalgamated.h for icelake: #include <generic/stage1/amalgamated.h> */
+/* begin file generic/stage1/amalgamated.h for icelake */
+// Stuff other things depend on
+/* including generic/stage1/base.h for icelake: #include <generic/stage1/base.h> */
+/* begin file generic/stage1/base.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_BASE_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace stage1 {
+
+class bit_indexer;
+template<size_t STEP_SIZE>
+struct buf_block_reader;
+struct json_block;
+class json_minifier;
+class json_scanner;
+struct json_string_block;
+class json_string_scanner;
+class json_structural_indexer;
+
+} // namespace stage1
+
+namespace utf8_validation {
+struct utf8_checker;
+} // namespace utf8_validation
+
+using utf8_validation::utf8_checker;
+
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_BASE_H
+/* end file generic/stage1/base.h for icelake */
+/* including generic/stage1/buf_block_reader.h for icelake: #include <generic/stage1/buf_block_reader.h> */
+/* begin file generic/stage1/buf_block_reader.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace stage1 {
+
+// Walks through a buffer in block-sized increments, loading the last part with spaces
+template<size_t STEP_SIZE>
+struct buf_block_reader {
+public:
+  simdjson_inline buf_block_reader(const uint8_t *_buf, size_t _len);
+  simdjson_inline size_t block_index();
+  simdjson_inline bool has_full_block() const;
+  simdjson_inline const uint8_t *full_block() const;
+  /**
+   * Get the last block, padded with spaces.
+   *
+   * There will always be a last block, with at least 1 byte, unless len == 0 (in which case this
+   * function fills the buffer with spaces and returns 0. In particular, if len == STEP_SIZE there
+   * will be 0 full_blocks and 1 remainder block with STEP_SIZE bytes and no spaces for padding.
+   *
+   * @return the number of effective characters in the last block.
+   */
+  simdjson_inline size_t get_remainder(uint8_t *dst) const;
+  simdjson_inline void advance();
+private:
+  const uint8_t *buf;
+  const size_t len;
+  const size_t lenminusstep;
+  size_t idx;
+};
+
+// Routines to print masks and text for debugging bitmask operations
+simdjson_unused static char * format_input_text_64(const uint8_t *text) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    buf[i] = int8_t(text[i]) < ' ' ? '_' : int8_t(text[i]);
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+// Routines to print masks and text for debugging bitmask operations
+simdjson_unused static char * format_input_text(const simd8x64<uint8_t>& in) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  in.store(reinterpret_cast<uint8_t*>(buf));
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] < ' ') { buf[i] = '_'; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+simdjson_unused static char * format_input_text(const simd8x64<uint8_t>& in, uint64_t mask) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  in.store(reinterpret_cast<uint8_t*>(buf));
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] <= ' ') { buf[i] = '_'; }
+    if (!(mask & (size_t(1) << i))) { buf[i] = ' '; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+simdjson_unused static char * format_mask(uint64_t mask) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  for (size_t i=0; i<64; i++) {
+    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
+  }
+  buf[64] = '\0';
+  return buf;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
+
+template<size_t STEP_SIZE>
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
+
+template<size_t STEP_SIZE>
+simdjson_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
+  return idx < lenminusstep;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
+  return &buf[idx];
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
+  if(len == idx) { return 0; } // memcpy(dst, null, 0) will trigger an error with some sanitizers
+  std::memset(dst, 0x20, STEP_SIZE); // std::memset STEP_SIZE because it's more efficient to write out 8 or 16 bytes at once.
+  std::memcpy(dst, buf + idx, len - idx);
+  return len - idx;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline void buf_block_reader<STEP_SIZE>::advance() {
+  idx += STEP_SIZE;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H
+/* end file generic/stage1/buf_block_reader.h for icelake */
+/* including generic/stage1/json_escape_scanner.h for icelake: #include <generic/stage1/json_escape_scanner.h> */
+/* begin file generic/stage1/json_escape_scanner.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_ESCAPE_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_ESCAPE_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace stage1 {
+
+/**
+ * Scans for escape characters in JSON, taking care with multiple backslashes (\\n vs. \n).
+ */
+struct json_escape_scanner {
+  /** The actual escape characters (the backslashes themselves). */
+  uint64_t next_is_escaped = 0ULL;
+
+  struct escaped_and_escape {
+    /**
+     * Mask of escaped characters.
+     *
+     * ```
+     * \n \\n \\\n \\\\n \
+     * 0100100010100101000
+     *  n  \   \ n  \ \
+     * ```
+     */
+    uint64_t escaped;
+    /**
+     * Mask of escape characters.
+     *
+     * ```
+     * \n \\n \\\n \\\\n \
+     * 1001000101001010001
+     * \  \   \ \  \ \   \
+     * ```
+     */
+    uint64_t escape;
+  };
+
+  /**
+   * Get a mask of both escape and escaped characters (the characters following a backslash).
+   *
+   * @param potential_escape A mask of the character that can escape others (but could be
+   *        escaped itself). e.g. block.eq('\\')
+   */
+  simdjson_really_inline escaped_and_escape next(uint64_t backslash) noexcept {
+
+#if !SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+    if (!backslash) { return {next_escaped_without_backslashes(), 0}; }
+#endif
+
+    // |                                | Mask (shows characters instead of 1's) | Depth | Instructions        |
+    // |--------------------------------|----------------------------------------|-------|---------------------|
+    // | string                         | `\\n_\\\n___\\\n___\\\\___\\\\__\\\`   |       |                     |
+    // |                                | `    even   odd    even   odd   odd`   |       |                     |
+    // | potential_escape               | ` \  \\\    \\\    \\\\   \\\\  \\\`   | 1     | 1 (backslash & ~first_is_escaped)
+    // | escape_and_terminal_code       | ` \n \ \n   \ \n   \ \    \ \   \ \`   | 5     | 5 (next_escape_and_terminal_code())
+    // | escaped                        | `\    \ n    \ n    \ \    \ \   \ ` X | 6     | 7 (escape_and_terminal_code ^ (potential_escape | first_is_escaped))
+    // | escape                         | `    \ \    \ \    \ \    \ \   \ \`   | 6     | 8 (escape_and_terminal_code & backslash)
+    // | first_is_escaped               | `\                                 `   | 7 (*) | 9 (escape >> 63) ()
+    //                                                                               (*) this is not needed until the next iteration
+    uint64_t escape_and_terminal_code = next_escape_and_terminal_code(backslash & ~this->next_is_escaped);
+    uint64_t escaped = escape_and_terminal_code ^ (backslash | this->next_is_escaped);
+    uint64_t escape = escape_and_terminal_code & backslash;
+    this->next_is_escaped = escape >> 63;
+    return {escaped, escape};
+  }
+
+private:
+  static constexpr const uint64_t ODD_BITS = 0xAAAAAAAAAAAAAAAAULL;
+
+  simdjson_really_inline uint64_t next_escaped_without_backslashes() noexcept {
+    uint64_t escaped = this->next_is_escaped;
+    this->next_is_escaped = 0;
+    return escaped;
+  }
+
+  /**
+   * Returns a mask of the next escape characters (masking out escaped backslashes), along with
+   * any non-backslash escape codes.
+   *
+   * \n \\n \\\n \\\\n returns:
+   * \n \   \ \n \ \
+   * 11 100 1011 10100
+   *
+   * You are expected to mask out the first bit yourself if the previous block had a trailing
+   * escape.
+   *
+   * & the result with potential_escape to get just the escape characters.
+   * ^ the result with (potential_escape | first_is_escaped) to get escaped characters.
+   */
+  static simdjson_really_inline uint64_t next_escape_and_terminal_code(uint64_t potential_escape) noexcept {
+    // If we were to just shift and mask out any odd bits, we'd actually get a *half* right answer:
+    // any even-aligned backslash runs would be correct! Odd-aligned backslash runs would be
+    // inverted (\\\ would be 010 instead of 101).
+    //
+    // ```
+    // string:              | ____\\\\_\\\\_____ |
+    // maybe_escaped | ODD  |     \ \   \ \      |
+    //               even-aligned ^^^  ^^^^ odd-aligned
+    // ```
+    //
+    // Taking that into account, our basic strategy is:
+    //
+    // 1. Use subtraction to produce a mask with 1's for even-aligned runs and 0's for
+    //    odd-aligned runs.
+    // 2. XOR all odd bits, which masks out the odd bits in even-aligned runs, and brings IN the
+    //    odd bits in odd-aligned runs.
+    // 3. & with backslash to clean up any stray bits.
+    // runs are set to 0, and then XORing with "odd":
+    //
+    // |                                | Mask (shows characters instead of 1's) | Instructions        |
+    // |--------------------------------|----------------------------------------|---------------------|
+    // | string                         | `\\n_\\\n___\\\n___\\\\___\\\\__\\\`   |
+    // |                                | `    even   odd    even   odd   odd`   |
+    // | maybe_escaped                  | `  n  \\n    \\n    \\\_   \\\_  \\` X | 1 (potential_escape << 1)
+    // | maybe_escaped_and_odd          | ` \n_ \\n _ \\\n_ _ \\\__ _\\\_ \\\`   | 1 (maybe_escaped | odd)
+    // | even_series_codes_and_odd      | `  n_\\\  _    n_ _\\\\ _     _    `   | 1 (maybe_escaped_and_odd - potential_escape)
+    // | escape_and_terminal_code       | ` \n \ \n   \ \n   \ \    \ \   \ \`   | 1 (^ odd)
+    //
+
+    // Escaped characters are characters following an escape.
+    uint64_t maybe_escaped = potential_escape << 1;
+
+    // To distinguish odd from even escape sequences, therefore, we turn on any *starting*
+    // escapes that are on an odd byte. (We actually bring in all odd bits, for speed.)
+    // - Odd runs of backslashes are 0000, and the code at the end ("n" in \n or \\n) is 1.
+    // - Odd runs of backslashes are 1111, and the code at the end ("n" in \n or \\n) is 0.
+    // - All other odd bytes are 1, and even bytes are 0.
+    uint64_t maybe_escaped_and_odd_bits     = maybe_escaped | ODD_BITS;
+    uint64_t even_series_codes_and_odd_bits = maybe_escaped_and_odd_bits - potential_escape;
+
+    // Now we flip all odd bytes back with xor. This:
+    // - Makes odd runs of backslashes go from 0000 to 1010
+    // - Makes even runs of backslashes go from 1111 to 1010
+    // - Sets actually-escaped codes to 1 (the n in \n and \\n: \n = 11, \\n = 100)
+    // - Resets all other bytes to 0
+    return even_series_codes_and_odd_bits ^ ODD_BITS;
+  }
+};
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+/* end file generic/stage1/json_escape_scanner.h for icelake */
+/* including generic/stage1/json_string_scanner.h for icelake: #include <generic/stage1/json_string_scanner.h> */
+/* begin file generic/stage1/json_string_scanner.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_escape_scanner.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace stage1 {
+
+struct json_string_block {
+  // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
+  simdjson_really_inline json_string_block(uint64_t escaped, uint64_t quote, uint64_t in_string) :
+  _escaped(escaped), _quote(quote), _in_string(in_string) {}
+
+  // Escaped characters (characters following an escape() character)
+  simdjson_really_inline uint64_t escaped() const { return _escaped; }
+  // Real (non-backslashed) quotes
+  simdjson_really_inline uint64_t quote() const { return _quote; }
+  // Only characters inside the string (not including the quotes)
+  simdjson_really_inline uint64_t string_content() const { return _in_string & ~_quote; }
+  // Return a mask of whether the given characters are inside a string (only works on non-quotes)
+  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
+  // Return a mask of whether the given characters are inside a string (only works on non-quotes)
+  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
+  // Tail of string (everything except the start quote)
+  simdjson_really_inline uint64_t string_tail() const { return _in_string ^ _quote; }
+
+  // escaped characters (backslashed--does not include the hex characters after \u)
+  uint64_t _escaped;
+  // real quotes (non-escaped ones)
+  uint64_t _quote;
+  // string characters (includes start quote but not end quote)
+  uint64_t _in_string;
+};
+
+// Scans blocks for string characters, storing the state necessary to do so
+class json_string_scanner {
+public:
+  simdjson_really_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
+  // Returns either UNCLOSED_STRING or SUCCESS
+  simdjson_really_inline error_code finish();
+
+private:
+  // Scans for escape characters
+  json_escape_scanner escape_scanner{};
+  // Whether the last iteration was still inside a string (all 1's = true, all 0's = false).
+  uint64_t prev_in_string = 0ULL;
+};
+
+//
+// Return a mask of all string characters plus end quotes.
+//
+// prev_escaped is overflow saying whether the next character is escaped.
+// prev_in_string is overflow saying whether we're still in a string.
+//
+// Backslash sequences outside of quotes will be detected in stage 2.
+//
+simdjson_really_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
+  const uint64_t backslash = in.eq('\\');
+  const uint64_t escaped = escape_scanner.next(backslash).escaped;
+  const uint64_t quote = in.eq('"') & ~escaped;
+
+  //
+  // prefix_xor flips on bits inside the string (and flips off the end quote).
+  //
+  // Then we xor with prev_in_string: if we were in a string already, its effect is flipped
+  // (characters inside strings are outside, and characters outside strings are inside).
+  //
+  const uint64_t in_string = prefix_xor(quote) ^ prev_in_string;
+
+  //
+  // Check if we're still in a string at the end of the box so the next block will know
+  //
+  prev_in_string = uint64_t(static_cast<int64_t>(in_string) >> 63);
+
+  // Use ^ to turn the beginning quote off, and the end quote on.
+
+  // We are returning a function-local object so either we get a move constructor
+  // or we get copy elision.
+  return json_string_block(escaped, quote, in_string);
+}
+
+simdjson_really_inline error_code json_string_scanner::finish() {
+  if (prev_in_string) {
+    return UNCLOSED_STRING;
+  }
+  return SUCCESS;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+/* end file generic/stage1/json_string_scanner.h for icelake */
+/* including generic/stage1/utf8_lookup4_algorithm.h for icelake: #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* begin file generic/stage1/utf8_lookup4_algorithm.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace utf8_validation {
+
+using namespace simd;
+
+  simdjson_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+// Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+// Bit 1 = Too Long (ASCII followed by continuation)
+// Bit 2 = Overlong 3-byte
+// Bit 4 = Surrogate
+// Bit 5 = Overlong 2-byte
+// Bit 7 = Two Continuations
+    constexpr const uint8_t TOO_SHORT   = 1<<0; // 11______ 0_______
+                                                // 11______ 11______
+    constexpr const uint8_t TOO_LONG    = 1<<1; // 0_______ 10______
+    constexpr const uint8_t OVERLONG_3  = 1<<2; // 11100000 100_____
+    constexpr const uint8_t SURROGATE   = 1<<4; // 11101101 101_____
+    constexpr const uint8_t OVERLONG_2  = 1<<5; // 1100000_ 10______
+    constexpr const uint8_t TWO_CONTS   = 1<<7; // 10______ 10______
+    constexpr const uint8_t TOO_LARGE   = 1<<3; // 11110100 1001____
+                                                // 11110100 101_____
+                                                // 11110101 1001____
+                                                // 11110101 101_____
+                                                // 1111011_ 1001____
+                                                // 1111011_ 101_____
+                                                // 11111___ 1001____
+                                                // 11111___ 101_____
+    constexpr const uint8_t TOO_LARGE_1000 = 1<<6;
+                                                // 11110101 1000____
+                                                // 1111011_ 1000____
+                                                // 11111___ 1000____
+    constexpr const uint8_t OVERLONG_4  = 1<<6; // 11110000 1000____
+
+    const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      TOO_SHORT,
+      // 1110____ ________ <three byte lead in byte 1>
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4
+    );
+    constexpr const uint8_t CARRY = TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+    const simd8<uint8_t> byte_1_low = (prev1 & 0x0F).lookup_16<uint8_t>(
+      // ____0000 ________
+      CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+      // ____0001 ________
+      CARRY | OVERLONG_2,
+      // ____001_ ________
+      CARRY,
+      CARRY,
+
+      // ____0100 ________
+      CARRY | TOO_LARGE,
+      // ____0101 ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      // ____011_ ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+
+      // ____1___ ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      // ____1101 ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000
+    );
+    const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE  | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE  | TOO_LARGE,
+
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT
+    );
+    return (byte_1_high & byte_1_low & byte_2_high);
+  }
+  simdjson_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
+      const simd8<uint8_t> prev_input, const simd8<uint8_t> sc) {
+    simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+    simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+    simd8<uint8_t> must23 = must_be_2_3_continuation(prev2, prev3);
+    simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+    return must23_80 ^ sc;
+  }
+
+  //
+  // Return nonzero if there are incomplete multibyte characters at the end of the block:
+  // e.g. if there is a 4-byte character, but it's 3 bytes from the end.
+  //
+  simdjson_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+    // If the previous input's last 3 bytes match this, they're too short (they ended at EOF):
+    // ... 1111____ 111_____ 11______
+#if SIMDJSON_IMPLEMENTATION_ICELAKE || (SIMDJSON_IMPLEMENTATION_RVV_VLS && __riscv_v_fixed_vlen >= 512)
+    static const uint8_t max_array[64] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
+    };
+#else
+    static const uint8_t max_array[32] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
+    };
+#endif
+    const simd8<uint8_t> max_value(&max_array[sizeof(max_array)-sizeof(simd8<uint8_t>)]);
+    return input.gt_bits(max_value);
+  }
+
+  struct utf8_checker {
+    // If this is nonzero, there has been a UTF-8 error.
+    simd8<uint8_t> error;
+    // The last input we received
+    simd8<uint8_t> prev_input_block;
+    // Whether the last input we received was incomplete (used for ASCII fast path)
+    simd8<uint8_t> prev_incomplete;
+
+    //
+    // Check whether the current bytes are valid UTF-8.
+    //
+    simdjson_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
+      // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ lead bytes
+      // (2, 3, 4-byte leads become large positive numbers instead of small negative numbers)
+      simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+      simd8<uint8_t> sc = check_special_cases(input, prev1);
+      this->error |= check_multibyte_lengths(input, prev_input, sc);
+    }
+
+    // The only problem that can happen at EOF is that a multibyte character is too short
+    // or a byte value too large in the last bytes: check_special_cases only checks for bytes
+    // too large in the first of two bytes.
+    simdjson_inline void check_eof() {
+      // If the previous block had incomplete UTF-8 characters at the end, an ASCII block can't
+      // possibly finish them.
+      this->error |= this->prev_incomplete;
+    }
+
+    simdjson_inline void check_next_input(const simd8x64<uint8_t>& input) {
+      if(simdjson_likely(is_ascii(input))) {
+        this->error |= this->prev_incomplete;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio, it is not good enough.
+        static_assert((simd8x64<uint8_t>::NUM_CHUNKS == 1)
+                ||(simd8x64<uint8_t>::NUM_CHUNKS == 2)
+                || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+                "We support one, two or four chunks per 64-byte block.");
+        SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 1) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+          this->check_utf8_bytes(input.get<1>(), input.get<0>());
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+          this->check_utf8_bytes(input.get<1>(), input.get<0>());
+          this->check_utf8_bytes(input.get<2>(), input.get<1>());
+          this->check_utf8_bytes(input.get<3>(), input.get<2>());
+        }
+        this->prev_incomplete = is_incomplete(input.get<simd8x64<uint8_t>::NUM_CHUNKS-1>());
+        this->prev_input_block = input.get<simd8x64<uint8_t>::NUM_CHUNKS-1>();
+      }
+    }
+    // do not forget to call check_eof!
+    simdjson_warn_unused simdjson_inline error_code errors() {
+      return this->error.any_bits_set_anywhere() ? error_code::UTF8_ERROR : error_code::SUCCESS;
+    }
+
+  }; // struct utf8_checker
+} // namespace utf8_validation
+
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H
+/* end file generic/stage1/utf8_lookup4_algorithm.h for icelake */
+/* including generic/stage1/json_scanner.h for icelake: #include <generic/stage1/json_scanner.h> */
+/* begin file generic/stage1/json_scanner.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/json_character_block.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_string_scanner.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace stage1 {
+
+/**
+ * A block of scanned json, with information on operators and scalars.
+ *
+ * We seek to identify pseudo-structural characters. Anything that is inside
+ * a string must be omitted (hence  & ~_string.string_tail()).
+ * Otherwise, pseudo-structural characters come in two forms.
+ * 1. We have the structural characters ([,],{,},:, comma). The
+ *    term 'structural character' is from the JSON RFC.
+ * 2. We have the 'scalar pseudo-structural characters'.
+ *    Scalars are quotes, and any character except structural characters and white space.
+ *
+ * To identify the scalar pseudo-structural characters, we must look at what comes
+ * before them: it must be a space, a quote or a structural characters.
+ * Starting with simdjson v0.3, we identify them by
+ * negation: we identify everything that is followed by a non-quote scalar,
+ * and we negate that. Whatever remains must be a 'scalar pseudo-structural character'.
+ */
+struct json_block {
+public:
+  // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
+  simdjson_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  _string(std::move(string)), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
+  simdjson_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  _string(string), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
+
+  /**
+   * The start of structurals.
+   * In simdjson prior to v0.3, these were called the pseudo-structural characters.
+   **/
+  simdjson_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
+  /** All JSON whitespace (i.e. not in a string) */
+  simdjson_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
+
+  // Helpers
+
+  /** Whether the given characters are inside a string (only works on non-quotes) */
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
+  /** Whether the given characters are outside a string (only works on non-quotes) */
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
+
+  // string and escape characters
+  json_string_block _string;
+  // whitespace, structural characters ('operators'), scalars
+  json_character_block _characters;
+  // whether the previous character was a scalar
+  uint64_t _follows_potential_nonquote_scalar;
+private:
+  // Potential structurals (i.e. disregarding strings)
+
+  /**
+   * structural elements ([,],{,},:, comma) plus scalar starts like 123, true and "abc".
+   * They may reside inside a string.
+   **/
+  simdjson_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
+  /**
+   * The start of non-operator runs, like 123, true and "abc".
+   * It main reside inside a string.
+   **/
+  simdjson_inline uint64_t potential_scalar_start() const noexcept {
+    // The term "scalar" refers to anything except structural characters and white space
+    // (so letters, numbers, quotes).
+    // Whenever it is preceded by something that is not a structural element ({,},[,],:, ") nor a white-space
+    // then we know that it is irrelevant structurally.
+    return _characters.scalar() & ~follows_potential_scalar();
+  }
+  /**
+   * Whether the given character is immediately after a non-operator like 123, true.
+   * The characters following a quote are not included.
+   */
+  simdjson_inline uint64_t follows_potential_scalar() const noexcept {
+    // _follows_potential_nonquote_scalar: is defined as marking any character that follows a character
+    // that is not a structural element ({,},[,],:, comma) nor a quote (") and that is not a
+    // white space.
+    // It is understood that within quoted region, anything at all could be marked (irrelevant).
+    return _follows_potential_nonquote_scalar;
+  }
+};
+
+/**
+ * Scans JSON for important bits: structural characters or 'operators', strings, and scalars.
+ *
+ * The scanner starts by calculating two distinct things:
+ * - string characters (taking \" into account)
+ * - structural characters or 'operators' ([]{},:, comma)
+ *   and scalars (runs of non-operators like 123, true and "abc")
+ *
+ * To minimize data dependency (a key component of the scanner's speed), it finds these in parallel:
+ * in particular, the operator/scalar bit will find plenty of things that are actually part of
+ * strings. When we're done, json_block will fuse the two together by masking out tokens that are
+ * part of a string.
+ */
+class json_scanner {
+public:
+  json_scanner() = default;
+  simdjson_inline json_block next(const simd::simd8x64<uint8_t>& in);
+  // Returns either UNCLOSED_STRING or SUCCESS
+  simdjson_warn_unused simdjson_inline error_code finish();
+
+private:
+  // Whether the last character of the previous iteration is part of a scalar token
+  // (anything except whitespace or a structural character/'operator').
+  uint64_t prev_scalar = 0ULL;
+  json_string_scanner string_scanner{};
+};
+
+
+//
+// Check if the current character immediately follows a matching character.
+//
+// For example, this checks for quotes with backslashes in front of them:
+//
+//     const uint64_t backslashed_quote = in.eq('"') & immediately_follows(in.eq('\'), prev_backslash);
+//
+simdjson_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
+  const uint64_t result = match << 1 | overflow;
+  overflow = match >> 63;
+  return result;
+}
+
+simdjson_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
+  json_string_block strings = string_scanner.next(in);
+  // identifies the white-space and the structural characters
+  json_character_block characters = json_character_block::classify(in);
+  // The term "scalar" refers to anything except structural characters and white space
+  // (so letters, numbers, quotes).
+  // We want follows_scalar to mark anything that follows a non-quote scalar (so letters and numbers).
+  //
+  // A terminal quote should either be followed by a structural character (comma, brace, bracket, colon)
+  // or nothing. However, we still want ' "a string"true ' to mark the 't' of 'true' as a potential
+  // pseudo-structural character just like we would if we had  ' "a string" true '; otherwise we
+  // may need to add an extra check when parsing strings.
+  //
+  // Performance: there are many ways to skin this cat.
+  const uint64_t nonquote_scalar = characters.scalar() & ~strings.quote();
+  uint64_t follows_nonquote_scalar = follows(nonquote_scalar, prev_scalar);
+  // We are returning a function-local object so either we get a move constructor
+  // or we get copy elision.
+  return json_block(
+    strings,// strings is a function-local object so either it moves or the copy is elided.
+    characters,
+    follows_nonquote_scalar
+  );
+}
+
+simdjson_warn_unused simdjson_inline error_code json_scanner::finish() {
+  return string_scanner.finish();
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H
+/* end file generic/stage1/json_scanner.h for icelake */
+
+// All other declarations
+/* including generic/stage1/find_next_document_index.h for icelake: #include <generic/stage1/find_next_document_index.h> */
+/* begin file generic/stage1/find_next_document_index.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace stage1 {
+
+/**
+  * This algorithm is used to quickly identify the last structural position that
+  * makes up a complete document.
+  *
+  * It does this by going backwards and finding the last *document boundary* (a
+  * place where one value follows another without a comma between them). If the
+  * last document (the characters after the boundary) has an equal number of
+  * start and end brackets, it is considered complete.
+  *
+  * Simply put, we iterate over the structural characters, starting from
+  * the end. We consider that we found the end of a JSON document when the
+  * first element of the pair is NOT one of these characters: '{' '[' ':' ','
+  * and when the second element is NOT one of these characters: '}' ']' ':' ','.
+  *
+  * This simple comparison works most of the time, but it does not cover cases
+  * where the batch's structural indexes contain a perfect amount of documents.
+  * In such a case, we do not have access to the structural index which follows
+  * the last document, therefore, we do not have access to the second element in
+  * the pair, and that means we cannot identify the last document. To fix this
+  * issue, we keep a count of the open and closed curly/square braces we found
+  * while searching for the pair. When we find a pair AND the count of open and
+  * closed curly/square braces is the same, we know that we just passed a
+  * complete document, therefore the last json buffer location is the end of the
+  * batch.
+  */
+simdjson_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
+  // Variant: do not count separately, just figure out depth
+  if(parser.n_structural_indexes == 0) { return 0; }
+  auto arr_cnt = 0;
+  auto obj_cnt = 0;
+  for (auto i = parser.n_structural_indexes - 1; i > 0; i--) {
+    auto idxb = parser.structural_indexes[i];
+    switch (parser.buf[idxb]) {
+    case ':':
+    case ',':
+      continue;
+    case '}':
+      obj_cnt--;
+      continue;
+    case ']':
+      arr_cnt--;
+      continue;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+    }
+    auto idxa = parser.structural_indexes[i - 1];
+    switch (parser.buf[idxa]) {
+    case '{':
+    case '[':
+    case ':':
+    case ',':
+      continue;
+    }
+    // Last document is complete, so the next document will appear after!
+    if (!arr_cnt && !obj_cnt) {
+      return parser.n_structural_indexes;
+    }
+    // Last document is incomplete; mark the document at i + 1 as the next one
+    return i;
+  }
+  // If we made it to the end, we want to finish counting to see if we have a full document.
+  switch (parser.buf[parser.structural_indexes[0]]) {
+    case '}':
+      obj_cnt--;
+      break;
+    case ']':
+      arr_cnt--;
+      break;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+  }
+  if (!arr_cnt && !obj_cnt) {
+    // We have a complete document.
+    return parser.n_structural_indexes;
+  }
+  return 0;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+/* end file generic/stage1/find_next_document_index.h for icelake */
+/* including generic/stage1/json_minifier.h for icelake: #include <generic/stage1/json_minifier.h> */
+/* begin file generic/stage1/json_minifier.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses in stage1
+// It is intended to be included multiple times and compiled multiple times
+// We assume the file in which it is included already includes
+// "simdjson/stage1.h" (this simplifies amalgation)
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace stage1 {
+
+class json_minifier {
+public:
+  template<size_t STEP_SIZE>
+  static error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept;
+
+private:
+  simdjson_inline json_minifier(uint8_t *_dst)
+  : dst{_dst}
+  {}
+  template<size_t STEP_SIZE>
+  simdjson_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
+  simdjson_warn_unused simdjson_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
+  json_scanner scanner{};
+  uint8_t *dst;
+};
+
+simdjson_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
+  uint64_t mask = block.whitespace();
+  dst += in.compress(mask, dst);
+}
+
+simdjson_warn_unused simdjson_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
+  error_code error = scanner.finish();
+  if (error) { dst_len = 0; return error; }
+  dst_len = dst - dst_start;
+  return SUCCESS;
+}
+
+template<>
+simdjson_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block_buf);
+  simd::simd8x64<uint8_t> in_2(block_buf+64);
+  json_block block_1 = scanner.next(in_1);
+  json_block block_2 = scanner.next(in_2);
+  this->next(in_1, block_1);
+  this->next(in_2, block_2);
+  reader.advance();
+}
+
+template<>
+simdjson_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block_buf);
+  json_block block_1 = scanner.next(in_1);
+  this->next(block_buf, block_1);
+  reader.advance();
+}
+
+template<size_t STEP_SIZE>
+error_code json_minifier::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept {
+  buf_block_reader<STEP_SIZE> reader(buf, len);
+  json_minifier minifier(dst);
+
+  // Index the first n-1 blocks
+  while (reader.has_full_block()) {
+    minifier.step<STEP_SIZE>(reader.full_block(), reader);
+  }
+
+  // Index the last (remainder) block, padded with spaces
+  uint8_t block[STEP_SIZE];
+  size_t remaining_bytes = reader.get_remainder(block);
+  if (remaining_bytes > 0) {
+    // We do not want to write directly to the output stream. Rather, we write
+    // to a local buffer (for safety).
+    uint8_t out_block[STEP_SIZE];
+    uint8_t * const guarded_dst{minifier.dst};
+    minifier.dst = out_block;
+    minifier.step<STEP_SIZE>(block, reader);
+    size_t to_write = minifier.dst - out_block;
+    // In some cases, we could be enticed to consider the padded spaces
+    // as part of the string. This is fine as long as we do not write more
+    // than we consumed.
+    if(to_write > remaining_bytes) { to_write = remaining_bytes; }
+    memcpy(guarded_dst, out_block, to_write);
+    minifier.dst = guarded_dst + to_write;
+  }
+  return minifier.finish(dst, dst_len);
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H
+/* end file generic/stage1/json_minifier.h for icelake */
+/* including generic/stage1/json_structural_indexer.h for icelake: #include <generic/stage1/json_structural_indexer.h> */
+/* begin file generic/stage1/json_structural_indexer.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_string_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_minifier.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/find_next_document_index.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses in stage1
+// It is intended to be included multiple times and compiled multiple times
+// We assume the file in which it is included already includes
+// "simdjson/stage1.h" (this simplifies amalgation)
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace stage1 {
+
+class bit_indexer {
+public:
+  uint32_t *tail;
+
+  simdjson_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
+
+#if SIMDJSON_PREFER_REVERSE_BITS
+  /**
+    * ARM lacks a fast trailing zero instruction, but it has a fast
+    * bit reversal instruction and a fast leading zero instruction.
+    * Thus it may be profitable to reverse the bits (once) and then
+    * to rely on a sequence of instructions that call the leading
+    * zero instruction.
+    *
+    * Performance notes:
+    * The chosen routine is not optimal in terms of data dependency
+    * since zero_leading_bit might require two instructions. However,
+    * it tends to minimize the total number of instructions which is
+    * beneficial.
+    */
+  simdjson_inline void write_index(uint32_t idx, uint64_t& rev_bits, int i) {
+    int lz = leading_zeroes(rev_bits);
+    this->tail[i] = static_cast<uint32_t>(idx) + lz;
+    rev_bits = zero_leading_bit(rev_bits, lz);
+  }
+#else
+  /**
+    * Under recent x64 systems, we often have both a fast trailing zero
+    * instruction and a fast 'clear-lower-bit' instruction so the following
+    * algorithm can be competitive.
+    */
+
+  simdjson_inline void write_index(uint32_t idx, uint64_t& bits, int i) {
+    this->tail[i] = idx + trailing_zeroes(bits);
+    bits = clear_lowest_bit(bits);
+  }
+#endif // SIMDJSON_PREFER_REVERSE_BITS
+
+  template <int START, int N>
+  simdjson_inline int write_indexes(uint32_t idx, uint64_t& bits) {
+    write_index(idx, bits, START);
+    SIMDJSON_IF_CONSTEXPR (N > 1) {
+      write_indexes<(N-1>0?START+1:START), (N-1>=0?N-1:1)>(idx, bits);
+    }
+    return START+N;
+  }
+
+  template <int START, int END, int STEP>
+  simdjson_inline int write_indexes_stepped(uint32_t idx, uint64_t& bits, int cnt) {
+    write_indexes<START, STEP>(idx, bits);
+    SIMDJSON_IF_CONSTEXPR ((START+STEP)  < END) {
+      if (simdjson_unlikely((START+STEP) < cnt)) {
+        write_indexes_stepped<(START+STEP<END?START+STEP:END), END, STEP>(idx, bits, cnt);
+      }
+    }
+    return ((END-START) % STEP) == 0 ? END : (END-START) - ((END-START) % STEP) + STEP;
+  }
+
+  // flatten out values in 'bits' assuming that they are are to have values of idx
+  // plus their position in the bitvector, and store these indexes at
+  // base_ptr[base] incrementing base as we go
+  // will potentially store extra values beyond end of valid bits, so base_ptr
+  // needs to be large enough to handle this
+  //
+  // If the kernel sets SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER, then it
+  // will provide its own version of the code.
+#ifdef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+  simdjson_inline void write(uint32_t idx, uint64_t bits);
+#else
+  simdjson_inline void write(uint32_t idx, uint64_t bits) {
+    // In some instances, the next branch is expensive because it is mispredicted.
+    // Unfortunately, in other cases,
+    // it helps tremendously.
+    if (bits == 0)
+        return;
+
+    int cnt = static_cast<int>(count_ones(bits));
+
+#if SIMDJSON_PREFER_REVERSE_BITS
+    bits = reverse_bits(bits);
+#endif
+#ifdef SIMDJSON_STRUCTURAL_INDEXER_STEP
+    static constexpr const int STEP = SIMDJSON_STRUCTURAL_INDEXER_STEP;
+#else
+    static constexpr const int STEP = 4;
+#endif
+    static constexpr const int STEP_UNTIL = 24;
+
+    write_indexes_stepped<0, STEP_UNTIL, STEP>(idx, bits, cnt);
+    SIMDJSON_IF_CONSTEXPR (STEP_UNTIL < 64) {
+      if (simdjson_unlikely(STEP_UNTIL < cnt)) {
+        for (int i=STEP_UNTIL; i<cnt; i++) {
+          write_index(idx, bits, i);
+        }
+      }
+    }
+
+    this->tail += cnt;
+  }
+#endif // SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+};
+
+class json_structural_indexer {
+public:
+  /**
+   * Find the important bits of JSON in a 128-byte chunk, and add them to structural_indexes.
+   *
+   * @param partial Setting the partial parameter to true allows the find_structural_bits to
+   *   tolerate unclosed strings. The caller should still ensure that the input is valid UTF-8. If
+   *   you are processing substrings, you may want to call on a function like trimmed_length_safe_utf8.
+   */
+  template<size_t STEP_SIZE>
+  static error_code index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept;
+
+private:
+  simdjson_inline json_structural_indexer(uint32_t *structural_indexes);
+  template<size_t STEP_SIZE>
+  simdjson_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
+  simdjson_warn_unused simdjson_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
+
+  json_scanner scanner{};
+  utf8_checker checker{};
+  bit_indexer indexer;
+  uint64_t prev_structurals = 0;
+  uint64_t unescaped_chars_error = 0;
+};
+
+simdjson_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
+
+// Skip the last character if it is partial
+simdjson_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
+  if (simdjson_unlikely(len < 3)) {
+    switch (len) {
+      case 2:
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
+        return len;
+      case 1:
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        return len;
+      case 0:
+        return len;
+    }
+  }
+  if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+  if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 1 byte left
+  if (buf[len-3] >= 0xf0) { return len-3; } // 4-byte characters with only 3 bytes left
+  return len;
+}
+
+//
+// PERF NOTES:
+// We pipe 2 inputs through these stages:
+// 1. Load JSON into registers. This takes a long time and is highly parallelizable, so we load
+//    2 inputs' worth at once so that by the time step 2 is looking for them input, it's available.
+// 2. Scan the JSON for critical data: strings, scalars and operators. This is the critical path.
+//    The output of step 1 depends entirely on this information. These functions don't quite use
+//    up enough CPU: the second half of the functions is highly serial, only using 1 execution core
+//    at a time. The second input's scans has some dependency on the first ones finishing it, but
+//    they can make a lot of progress before they need that information.
+// 3. Step 1 does not use enough capacity, so we run some extra stuff while we're waiting for that
+//    to finish: utf-8 checks and generating the output from the last iteration.
+//
+// The reason we run 2 inputs at a time, is steps 2 and 3 are *still* not enough to soak up all
+// available capacity with just one input. Running 2 at a time seems to give the CPU a good enough
+// workout.
+//
+template<size_t STEP_SIZE>
+error_code json_structural_indexer::index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept {
+  if (simdjson_unlikely(len > parser.capacity())) { return CAPACITY; }
+  // We guard the rest of the code so that we can assume that len > 0 throughout.
+  if (len == 0) { return EMPTY; }
+  if (is_streaming(partial)) {
+    len = trim_partial_utf8(buf, len);
+    // If you end up with an empty window after trimming
+    // the partial UTF-8 bytes, then chances are good that you
+    // have an UTF-8 formatting error.
+    if(len == 0) { return UTF8_ERROR; }
+  }
+  buf_block_reader<STEP_SIZE> reader(buf, len);
+  json_structural_indexer indexer(parser.structural_indexes.get());
+
+  // Read all but the last block
+  while (reader.has_full_block()) {
+    indexer.step<STEP_SIZE>(reader.full_block(), reader);
+  }
+  // Take care of the last block (will always be there unless file is empty which is
+  // not supposed to happen.)
+  uint8_t block[STEP_SIZE];
+  if (simdjson_unlikely(reader.get_remainder(block) == 0)) { return UNEXPECTED_ERROR; }
+  indexer.step<STEP_SIZE>(block, reader);
+  return indexer.finish(parser, reader.block_index(), len, partial);
+}
+
+template<>
+simdjson_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block);
+  simd::simd8x64<uint8_t> in_2(block+64);
+  json_block block_1 = scanner.next(in_1);
+  json_block block_2 = scanner.next(in_2);
+  this->next(in_1, block_1, reader.block_index());
+  this->next(in_2, block_2, reader.block_index()+64);
+  reader.advance();
+}
+
+template<>
+simdjson_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block);
+  json_block block_1 = scanner.next(in_1);
+  this->next(in_1, block_1, reader.block_index());
+  reader.advance();
+}
+
+simdjson_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
+  uint64_t unescaped = in.lteq(0x1F);
+#if SIMDJSON_UTF8VALIDATION
+  checker.check_next_input(in);
+#endif
+  indexer.write(uint32_t(idx-64), prev_structurals); // Output *last* iteration's structurals to the parser
+  prev_structurals = block.structural_start();
+  unescaped_chars_error |= block.non_quote_inside_string(unescaped);
+}
+
+simdjson_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
+  // Write out the final iteration's structurals
+  indexer.write(uint32_t(idx-64), prev_structurals);
+  error_code error = scanner.finish();
+  // We deliberately break down the next expression so that it is
+  // human readable.
+  const bool should_we_exit = is_streaming(partial) ?
+    ((error != SUCCESS) && (error != UNCLOSED_STRING)) // when partial we tolerate UNCLOSED_STRING
+    : (error != SUCCESS); // if partial is false, we must have SUCCESS
+  const bool have_unclosed_string = (error == UNCLOSED_STRING);
+  if (simdjson_unlikely(should_we_exit)) { return error; }
+
+  if (unescaped_chars_error) {
+    return UNESCAPED_CHARS;
+  }
+  parser.n_structural_indexes = uint32_t(indexer.tail - parser.structural_indexes.get());
+  /***
+   * The On-Demand API requires special padding.
+   *
+   * This is related to https://github.com/simdjson/simdjson/issues/906
+   * Basically, we want to make sure that if the parsing continues beyond the last (valid)
+   * structural character, it quickly stops.
+   * Only three structural characters can be repeated without triggering an error in JSON:  [,] and }.
+   * We repeat the padding character (at 'len'). We don't know what it is, but if the parsing
+   * continues, then it must be [,] or }.
+   * Suppose it is ] or }. We backtrack to the first character, what could it be that would
+   * not trigger an error? It could be ] or } but no, because you can't start a document that way.
+   * It can't be a comma, a colon or any simple value. So the only way we could continue is
+   * if the repeated character is [. But if so, the document must start with [. But if the document
+   * starts with [, it should end with ]. If we enforce that rule, then we would get
+   * ][[ which is invalid.
+   *
+   * This is illustrated with the test array_iterate_unclosed_error() on the following input:
+   * R"({ "a": [,,)"
+   **/
+  parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len); // used later in partial == stage1_mode::streaming_final
+  parser.structural_indexes[parser.n_structural_indexes + 1] = uint32_t(len);
+  parser.structural_indexes[parser.n_structural_indexes + 2] = 0;
+  parser.next_structural_index = 0;
+  // a valid JSON file cannot have zero structural indexes - we should have found something
+  if (simdjson_unlikely(parser.n_structural_indexes == 0u)) {
+    return EMPTY;
+  }
+  if (simdjson_unlikely(parser.structural_indexes[parser.n_structural_indexes - 1] > len)) {
+    return UNEXPECTED_ERROR;
+  }
+  if (partial == stage1_mode::streaming_partial) {
+    // If we have an unclosed string, then the last structural
+    // will be the quote and we want to make sure to omit it.
+    if(have_unclosed_string) {
+      parser.n_structural_indexes--;
+      // a valid JSON file cannot have zero structural indexes - we should have found something
+      if (simdjson_unlikely(parser.n_structural_indexes == 0u)) { return CAPACITY; }
+    }
+    // We truncate the input to the end of the last complete document (or zero).
+    auto new_structural_indexes = find_next_document_index(parser);
+    if (new_structural_indexes == 0 && parser.n_structural_indexes > 0) {
+      if(parser.structural_indexes[0] == 0) {
+        // If the buffer is partial and we started at index 0 but the document is
+        // incomplete, it's too big to parse.
+        return CAPACITY;
+      } else {
+        // It is possible that the document could be parsed, we just had a lot
+        // of white space.
+        parser.n_structural_indexes = 0;
+        return EMPTY;
+      }
+    }
+
+    parser.n_structural_indexes = new_structural_indexes;
+  } else if (partial == stage1_mode::streaming_final) {
+    if(have_unclosed_string) { parser.n_structural_indexes--; }
+    // We truncate the input to the end of the last complete document (or zero).
+    // Because partial == stage1_mode::streaming_final, it means that we may
+    // silently ignore trailing garbage. Though it sounds bad, we do it
+    // deliberately because many people who have streams of JSON documents
+    // will truncate them for processing. E.g., imagine that you are uncompressing
+    // the data from a size file or receiving it in chunks from the network. You
+    // may not know where exactly the last document will be. Meanwhile the
+    // document_stream instances allow people to know the JSON documents they are
+    // parsing (see the iterator.source() method).
+    parser.n_structural_indexes = find_next_document_index(parser);
+    // We store the initial n_structural_indexes so that the client can see
+    // whether we used truncation. If initial_n_structural_indexes == parser.n_structural_indexes,
+    // then this will query parser.structural_indexes[parser.n_structural_indexes] which is len,
+    // otherwise, it will copy some prior index.
+    parser.structural_indexes[parser.n_structural_indexes + 1] = parser.structural_indexes[parser.n_structural_indexes];
+    // This next line is critical, do not change it unless you understand what you are
+    // doing.
+    parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len);
+    if (simdjson_unlikely(parser.n_structural_indexes == 0u)) {
+        // We tolerate an unclosed string at the very end of the stream. Indeed, users
+        // often load their data in bulk without being careful and they want us to ignore
+        // the trailing garbage.
+        return EMPTY;
+    }
+  }
+  checker.check_eof();
+  return checker.errors();
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+// Clear CUSTOM_BIT_INDEXER so other implementations can set it if they need to.
+#undef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H
+/* end file generic/stage1/json_structural_indexer.h for icelake */
+/* including generic/stage1/utf8_validator.h for icelake: #include <generic/stage1/utf8_validator.h> */
+/* begin file generic/stage1/utf8_validator.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace stage1 {
+
+/**
+ * Validates that the string is actual UTF-8.
+ */
+template<class checker>
+bool generic_validate_utf8(const uint8_t * input, size_t length) {
+    checker c{};
+    buf_block_reader<64> reader(input, length);
+    while (reader.has_full_block()) {
+      simd::simd8x64<uint8_t> in(reader.full_block());
+      c.check_next_input(in);
+      reader.advance();
+    }
+    uint8_t block[64]{};
+    reader.get_remainder(block);
+    simd::simd8x64<uint8_t> in(block);
+    c.check_next_input(in);
+    reader.advance();
+    c.check_eof();
+    return c.errors() == error_code::SUCCESS;
+}
+
+bool generic_validate_utf8(const char * input, size_t length) {
+    return generic_validate_utf8<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H
+/* end file generic/stage1/utf8_validator.h for icelake */
+/* end file generic/stage1/amalgamated.h for icelake */
+/* including generic/stage2/amalgamated.h for icelake: #include <generic/stage2/amalgamated.h> */
+/* begin file generic/stage2/amalgamated.h for icelake */
+// Stuff other things depend on
+/* including generic/stage2/base.h for icelake: #include <generic/stage2/base.h> */
+/* begin file generic/stage2/base.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_BASE_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace stage2 {
+
+class json_iterator;
+class structural_iterator;
+struct tape_builder;
+struct tape_writer;
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_BASE_H
+/* end file generic/stage2/base.h for icelake */
+/* including generic/stage2/tape_writer.h for icelake: #include <generic/stage2/tape_writer.h> */
+/* begin file generic/stage2/tape_writer.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/internal/tape_type.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace stage2 {
+
+struct tape_writer {
+  /** The next place to write to tape */
+  uint64_t *next_tape_loc;
+
+  /** Write a signed 64-bit value to tape. */
+  simdjson_inline void append_s64(int64_t value) noexcept;
+
+  /** Write an unsigned 64-bit value to tape. */
+  simdjson_inline void append_u64(uint64_t value) noexcept;
+
+  /** Write a double value to tape. */
+  simdjson_inline void append_double(double value) noexcept;
+
+  /** Write a big integer (as string) to tape. src points to first digit, len is byte count. */
+  simdjson_inline void append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept;
+
+  /**
+   * Append a tape entry (an 8-bit type,and 56 bits worth of value).
+   */
+  simdjson_inline void append(uint64_t val, internal::tape_type t) noexcept;
+
+  /**
+   * Skip the current tape entry without writing.
+   *
+   * Used to skip the start of the container, since we'll come back later to fill it in when the
+   * container ends.
+   */
+  simdjson_inline void skip() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a large u64 or i64.
+   */
+  simdjson_inline void skip_large_integer() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a double.
+   */
+  simdjson_inline void skip_double() noexcept;
+
+  /**
+   * Write a value to a known location on tape.
+   *
+   * Used to go back and write out the start of a container after the container ends.
+   */
+  simdjson_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
+
+private:
+  /**
+   * Append both the tape entry, and a supplementary value following it. Used for types that need
+   * all 64 bits, such as double and uint64_t.
+   */
+  template<typename T>
+  simdjson_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
+}; // struct tape_writer
+
+simdjson_inline void tape_writer::append_s64(int64_t value) noexcept {
+  append2(0, value, internal::tape_type::INT64);
+}
+
+simdjson_inline void tape_writer::append_u64(uint64_t value) noexcept {
+  append(0, internal::tape_type::UINT64);
+  *next_tape_loc = value;
+  next_tape_loc++;
+}
+
+/** Write a double value to tape. */
+simdjson_inline void tape_writer::append_double(double value) noexcept {
+  append2(0, value, internal::tape_type::DOUBLE);
+}
+
+simdjson_inline void tape_writer::skip() noexcept {
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::skip_large_integer() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::skip_double() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
+  *next_tape_loc = val | ((uint64_t(char(t))) << 56);
+  next_tape_loc++;
+}
+
+template<typename T>
+simdjson_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
+  append(val, t);
+  static_assert(sizeof(val2) == sizeof(*next_tape_loc), "Type is not 64 bits!");
+  memcpy(next_tape_loc, &val2, sizeof(val2));
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
+  tape_loc = val | ((uint64_t(char(t))) << 56);
+}
+
+simdjson_inline void tape_writer::append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept {
+  // Write to string buffer: [4-byte LE length][digits][null]
+  uint32_t str_len = uint32_t(len);
+  memcpy(string_buf, &str_len, sizeof(uint32_t));
+  memcpy(string_buf + sizeof(uint32_t), src, len);
+  string_buf[sizeof(uint32_t) + len] = 0;
+  // Tape entry: offset into string buffer
+  // The caller must set the offset relative to doc.string_buf base
+  append(0, internal::tape_type::BIGINT); // placeholder offset, caller patches
+  string_buf += sizeof(uint32_t) + len + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+/* end file generic/stage2/tape_writer.h for icelake */
+/* including generic/stage2/logger.h for icelake: #include <generic/stage2/logger.h> */
+/* begin file generic/stage2/logger.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+
+// This is for an internal-only stage 2 specific logger.
+// Set LOG_ENABLED = true to log what stage 2 is doing!
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace logger {
+
+  static constexpr const char * DASHES = "----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------";
+
+#if SIMDJSON_VERBOSE_LOGGING
+  static constexpr const bool LOG_ENABLED = true;
+#else
+  static constexpr const bool LOG_ENABLED = false;
+#endif
+  static constexpr const int LOG_EVENT_LEN = 20;
+  static constexpr const int LOG_BUFFER_LEN = 30;
+  static constexpr const int LOG_SMALL_BUFFER_LEN = 10;
+  static constexpr const int LOG_INDEX_LEN = 5;
+
+  static int log_depth; // Not threadsafe. Log only.
+
+  // Helper to turn unprintable or newline characters into spaces
+  static simdjson_inline char printable_char(char c) {
+    if (c >= 0x20) {
+      return c;
+    } else {
+      return ' ';
+    }
+  }
+
+  // Print the header and set up log_start
+  static simdjson_inline void log_start() {
+    if (LOG_ENABLED) {
+      log_depth = 0;
+      printf("\n");
+      printf("| %-*s | %-*s | %-*s | %-*s | Detail |\n", LOG_EVENT_LEN, "Event", LOG_BUFFER_LEN, "Buffer", LOG_SMALL_BUFFER_LEN, "Next", 5, "Next#");
+      printf("|%.*s|%.*s|%.*s|%.*s|--------|\n", LOG_EVENT_LEN+2, DASHES, LOG_BUFFER_LEN+2, DASHES, LOG_SMALL_BUFFER_LEN+2, DASHES, 5+2, DASHES);
+    }
+  }
+
+  simdjson_unused static simdjson_inline void log_string(const char *message) {
+    if (LOG_ENABLED) {
+      printf("%s\n", message);
+    }
+  }
+
+  // Logs a single line from the stage 2 DOM parser
+  template<typename S>
+  static simdjson_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
+    if (LOG_ENABLED) {
+      printf("| %*s%s%-*s ", log_depth*2, "", title_prefix, LOG_EVENT_LEN - log_depth*2 - int(strlen(title_prefix)), title);
+      auto current_index = structurals.at_beginning() ? nullptr : structurals.next_structural-1;
+      auto next_index = structurals.next_structural;
+      auto current = current_index ? &structurals.buf[*current_index] : reinterpret_cast<const uint8_t*>("                                                       ");
+      auto next = &structurals.buf[*next_index];
+      {
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_BUFFER_LEN;i++) {
+          printf("%c", printable_char(current[i]));
+        }
+        printf(" ");
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_SMALL_BUFFER_LEN;i++) {
+          printf("%c", printable_char(next[i]));
+        }
+        printf(" ");
+      }
+      if (current_index) {
+        printf("| %*u ", LOG_INDEX_LEN, *current_index);
+      } else {
+        printf("| %-*s ", LOG_INDEX_LEN, "");
+      }
+      // printf("| %*u ", LOG_INDEX_LEN, structurals.next_tape_index());
+      printf("| %-s ", detail);
+      printf("|\n");
+    }
+  }
+
+} // namespace logger
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+/* end file generic/stage2/logger.h for icelake */
+
+// All other declarations
+/* including generic/stage2/json_iterator.h for icelake: #include <generic/stage2/json_iterator.h> */
+/* begin file generic/stage2/json_iterator.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/logger.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace stage2 {
+
+class json_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+  uint32_t depth{0};
+
+  /**
+   * Walk the JSON document.
+   *
+   * The visitor receives callbacks when values are encountered. All callbacks pass the iterator as
+   * the first parameter; some callbacks have other parameters as well:
+   *
+   * - visit_document_start() - at the beginning.
+   * - visit_document_end() - at the end (if things were successful).
+   *
+   * - visit_array_start() - at the start `[` of a non-empty array.
+   * - visit_array_end() - at the end `]` of a non-empty array.
+   * - visit_empty_array() - when an empty array is encountered.
+   *
+   * - visit_object_end() - at the start `]` of a non-empty object.
+   * - visit_object_start() - at the end `]` of a non-empty object.
+   * - visit_empty_object() - when an empty object is encountered.
+   * - visit_key(const uint8_t *key) - when a key in an object field is encountered. key is
+   *                                   guaranteed to point at the first quote of the string (`"key"`).
+   * - visit_primitive(const uint8_t *value) - when a value is a string, number, boolean or null.
+   * - visit_root_primitive(iter, uint8_t *value) - when the top-level value is a string, number, boolean or null.
+   *
+   * - increment_count(iter) - each time a value is found in an array or object.
+   */
+  template<bool STREAMING, typename V>
+  simdjson_warn_unused simdjson_inline error_code walk_document(V &visitor) noexcept;
+
+  /**
+   * Create an iterator capable of walking a JSON document.
+   *
+   * The document must have already passed through stage 1.
+   */
+  simdjson_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
+
+  /**
+   * Look at the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *peek() const noexcept;
+  /**
+   * Advance to the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *advance() noexcept;
+  /**
+   * Get the remaining length of the document, from the start of the current token.
+   */
+  simdjson_inline size_t remaining_len() const noexcept;
+  /**
+   * Check if we are at the end of the document.
+   *
+   * If this is true, there are no more tokens.
+   */
+  simdjson_inline bool at_eof() const noexcept;
+  /**
+   * Check if we are at the beginning of the document.
+   */
+  simdjson_inline bool at_beginning() const noexcept;
+  simdjson_inline uint8_t last_structural() const noexcept;
+
+  /**
+   * Log that a value has been found.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_value(const char *type) const noexcept;
+  /**
+   * Log the start of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_start_value(const char *type) const noexcept;
+  /**
+   * Log the end of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_end_value(const char *type) const noexcept;
+  /**
+   * Log an error.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_error(const char *error) const noexcept;
+
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
+};
+
+template<bool STREAMING, typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::walk_document(V &visitor) noexcept {
+  logger::log_start();
+
+  //
+  // Start the document
+  //
+  if (at_eof()) { return EMPTY; }
+  log_start_value("document");
+  SIMDJSON_TRY( visitor.visit_document_start(*this) );
+
+  //
+  // Read first value
+  //
+  {
+    auto value = advance();
+
+    // Make sure the outer object or array is closed before continuing; otherwise, there are ways we
+    // could get into memory corruption. See https://github.com/simdjson/simdjson/issues/906
+    if (!STREAMING) {
+      switch (*value) {
+        case '{': if (last_structural() != '}') { log_value("starting brace unmatched"); return TAPE_ERROR; }; break;
+        case '[': if (last_structural() != ']') { log_value("starting bracket unmatched"); return TAPE_ERROR; }; break;
+      }
+    }
+
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_root_primitive(*this, value) ); break;
+    }
+  }
+  goto document_end;
+
+//
+// Object parser states
+//
+object_begin:
+  log_start_value("object");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = false;
+  SIMDJSON_TRY( visitor.visit_object_start(*this) );
+
+  {
+    auto key = advance();
+    if (*key != '"') { log_error("Object does not start with a key"); return TAPE_ERROR; }
+    SIMDJSON_TRY( visitor.increment_count(*this) );
+    SIMDJSON_TRY( visitor.visit_key(*this, key) );
+  }
+
+object_field:
+  if (simdjson_unlikely( *advance() != ':' )) { log_error("Missing colon after key in object"); return TAPE_ERROR; }
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+object_continue:
+  switch (*advance()) {
+    case ',':
+      SIMDJSON_TRY( visitor.increment_count(*this) );
+      {
+        auto key = advance();
+        if (simdjson_unlikely( *key != '"' )) { log_error("Key string missing at beginning of field in object"); return TAPE_ERROR; }
+        SIMDJSON_TRY( visitor.visit_key(*this, key) );
+      }
+      goto object_field;
+    case '}': log_end_value("object"); SIMDJSON_TRY( visitor.visit_object_end(*this) ); goto scope_end;
+    default: log_error("No comma between object fields"); return TAPE_ERROR;
+  }
+
+scope_end:
+  depth--;
+  if (depth == 0) { goto document_end; }
+  if (dom_parser.is_array[depth]) { goto array_continue; }
+  goto object_continue;
+
+//
+// Array parser states
+//
+array_begin:
+  log_start_value("array");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = true;
+  SIMDJSON_TRY( visitor.visit_array_start(*this) );
+  SIMDJSON_TRY( visitor.increment_count(*this) );
+
+array_value:
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+array_continue:
+  switch (*advance()) {
+    case ',': SIMDJSON_TRY( visitor.increment_count(*this) ); goto array_value;
+    case ']': log_end_value("array"); SIMDJSON_TRY( visitor.visit_array_end(*this) ); goto scope_end;
+    default: log_error("Missing comma between array values"); return TAPE_ERROR;
+  }
+
+document_end:
+  log_end_value("document");
+  SIMDJSON_TRY( visitor.visit_document_end(*this) );
+
+  dom_parser.next_structural_index = uint32_t(next_structural - &dom_parser.structural_indexes[0]);
+
+  // If we didn't make it to the end, it's an error
+  if ( !STREAMING && dom_parser.next_structural_index != dom_parser.n_structural_indexes ) {
+    log_error("More than one JSON value at the root of the document, or extra characters at the end of the JSON!");
+    return TAPE_ERROR;
+  }
+
+  return SUCCESS;
+
+} // walk_document()
+
+simdjson_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+  : buf{_dom_parser.buf},
+    next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+    dom_parser{_dom_parser} {
+}
+
+simdjson_inline const uint8_t *json_iterator::peek() const noexcept {
+  return &buf[*(next_structural)];
+}
+simdjson_inline const uint8_t *json_iterator::advance() noexcept {
+  return &buf[*(next_structural++)];
+}
+simdjson_inline size_t json_iterator::remaining_len() const noexcept {
+  return dom_parser.len - *(next_structural-1);
+}
+
+simdjson_inline bool json_iterator::at_eof() const noexcept {
+  return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+}
+simdjson_inline bool json_iterator::at_beginning() const noexcept {
+  return next_structural == dom_parser.structural_indexes.get();
+}
+simdjson_inline uint8_t json_iterator::last_structural() const noexcept {
+  return buf[dom_parser.structural_indexes[dom_parser.n_structural_indexes - 1]];
+}
+
+simdjson_inline void json_iterator::log_value(const char *type) const noexcept {
+  logger::log_line(*this, "", type, "");
+}
+
+simdjson_inline void json_iterator::log_start_value(const char *type) const noexcept {
+  logger::log_line(*this, "+", type, "");
+  if (logger::LOG_ENABLED) { logger::log_depth++; }
+}
+
+simdjson_inline void json_iterator::log_end_value(const char *type) const noexcept {
+  if (logger::LOG_ENABLED) { logger::log_depth--; }
+  logger::log_line(*this, "-", type, "");
+}
+
+simdjson_inline void json_iterator::log_error(const char *error) const noexcept {
+  logger::log_line(*this, "", "ERROR", error);
+}
+
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
+  switch (*value) {
+    case '"': return visitor.visit_root_string(*this, value);
+    case 't': return visitor.visit_root_true_atom(*this, value);
+    case 'f': return visitor.visit_root_false_atom(*this, value);
+    case 'n': return visitor.visit_root_null_atom(*this, value);
+    case '-':
+    case '0': case '1': case '2': case '3': case '4':
+    case '5': case '6': case '7': case '8': case '9':
+      return visitor.visit_root_number(*this, value);
+    default:
+      log_error("Document starts with a non-value character");
+      return TAPE_ERROR;
+  }
+}
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
+  // Use the fact that most scalars are going to be either strings or numbers.
+  if(*value == '"') {
+    return visitor.visit_string(*this, value);
+  } else if (((*value - '0')  < 10) || (*value == '-')) {
+    return visitor.visit_number(*this, value);
+  }
+  // true, false, null are uncommon.
+  switch (*value) {
+    case 't': return visitor.visit_true_atom(*this, value);
+    case 'f': return visitor.visit_false_atom(*this, value);
+    case 'n': return visitor.visit_null_atom(*this, value);
+    default:
+      log_error("Non-value found when value was expected!");
+      return TAPE_ERROR;
+  }
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+/* end file generic/stage2/json_iterator.h for icelake */
+/* including generic/stage2/stringparsing.h for icelake: #include <generic/stage2/stringparsing.h> */
+/* begin file generic/stage2/stringparsing.h for icelake */
+#include <cstdint>
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/jsoncharutils.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses
+// It is intended to be included multiple times and compiled multiple times
+
+namespace simdjson {
+namespace icelake {
+namespace {
+/// @private
+namespace stringparsing {
+
+// begin copypasta
+// These chars yield themselves: " \ /
+// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
+// u not handled in this table as it's complex
+static const uint8_t escape_map[256] = {
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
+    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
+    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+};
+
+// handle a unicode codepoint
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr, bool allow_replacement) {
+  // Use the default Unicode Character 'REPLACEMENT CHARACTER' (U+FFFD)
+  constexpr uint32_t substitution_code_point = 0xfffd;
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) != ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+    } else {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+
+      // We have already checked that the high surrogate is valid and
+      // (code_point - 0xd800) < 1024.
+      //
+      // Check that code_point_2 is in the range 0xdc00..0xdfff
+      // and that code_point_2 was parsed from valid hex.
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if (low_bit >> 10) {
+        if(!allow_replacement) { return false; }
+        code_point = substitution_code_point;
+      } else {
+        code_point =  (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+
+    }
+  } else if (code_point >= 0xdc00 && code_point <= 0xdfff) {
+      // If we encounter a low surrogate (not preceded by a high surrogate)
+      // then we have an error.
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+  }
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+// handle a unicode codepoint using the wobbly convention
+// https://simonsapin.github.io/wtf-8/
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint_wobbly(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr) {
+  // It is not ideal that this function is nearly identical to handle_unicode_codepoint.
+  //
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) == ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if ((low_bit >> 10) ==  0) {
+        code_point =
+          (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+    }
+  }
+
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+/**
+ * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+ * must be an unescaped quote terminating the string. It returns the final output
+ * position as pointer. In case of error (e.g., the string has bad escaped codes),
+ * then null_ptr is returned. It is assumed that the output buffer is large
+ * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+ * SIMDJSON_PADDING bytes.
+ */
+simdjson_warn_unused simdjson_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) {
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint(&src, &dst, allow_replacement)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+simdjson_warn_unused simdjson_inline uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) {
+  // It is not ideal that this function is nearly identical to parse_string.
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint_wobbly(&src, &dst)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+} // namespace stringparsing
+
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+/* end file generic/stage2/stringparsing.h for icelake */
+/* including generic/stage2/structural_iterator.h for icelake: #include <generic/stage2/structural_iterator.h> */
+/* begin file generic/stage2/structural_iterator.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace stage2 {
+
+class structural_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+
+  // Start a structural
+  simdjson_inline structural_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+    : buf{_dom_parser.buf},
+      next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+      dom_parser{_dom_parser} {
+  }
+  // Get the buffer position of the current structural character
+  simdjson_inline const uint8_t* current() {
+    return &buf[*(next_structural-1)];
+  }
+  // Get the current structural character
+  simdjson_inline char current_char() {
+    return buf[*(next_structural-1)];
+  }
+  // Get the next structural character without advancing
+  simdjson_inline char peek_next_char() {
+    return buf[*next_structural];
+  }
+  simdjson_inline const uint8_t* peek() {
+    return &buf[*next_structural];
+  }
+  simdjson_inline const uint8_t* advance() {
+    return &buf[*(next_structural++)];
+  }
+  simdjson_inline char advance_char() {
+    return buf[*(next_structural++)];
+  }
+  simdjson_inline size_t remaining_len() {
+    return dom_parser.len - *(next_structural-1);
+  }
+
+  simdjson_inline bool at_end() {
+    return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+  }
+  simdjson_inline bool at_beginning() {
+    return next_structural == dom_parser.structural_indexes.get();
+  }
+};
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H
+/* end file generic/stage2/structural_iterator.h for icelake */
+/* including generic/stage2/tape_builder.h for icelake: #include <generic/stage2/tape_builder.h> */
+/* begin file generic/stage2/tape_builder.h for icelake */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/json_iterator.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/stringparsing.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/tape_writer.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/dom/document.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/atomparsing.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/numberparsing.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+
+namespace simdjson {
+namespace icelake {
+namespace {
+namespace stage2 {
+
+struct tape_builder {
+  template<bool STREAMING>
+  simdjson_warn_unused static simdjson_inline error_code parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept;
+
+  /** Called when a non-empty document starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty document ends without error. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_end(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty array starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty array ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_end(json_iterator &iter) noexcept;
+  /** Called when an empty array is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_array(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty object starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_start(json_iterator &iter) noexcept;
+  /**
+   * Called when a key in a field is encountered.
+   *
+   * primitive, visit_object_start, visit_empty_object, visit_array_start, or visit_empty_array
+   * will be called after this with the field value.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
+  /** Called when a non-empty object ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_end(json_iterator &iter) noexcept;
+  /** Called when an empty object is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_object(json_iterator &iter) noexcept;
+
+  /**
+   * Called when a string, number, boolean or null is found.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  /**
+   * Called when a string, number, boolean or null is found at the top level of a document (i.e.
+   * when there is no array or object and the entire document is a single string, number, boolean or
+   * null.
+   *
+   * This is separate from primitive() because simdjson's normal primitive parsing routines assume
+   * there is at least one more token after the value, which is only true in an array or object.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  /** Called each time a new field or element in an array or object is found. */
+  simdjson_warn_unused simdjson_inline error_code increment_count(json_iterator &iter) noexcept;
+
+  /** Next location to write to tape */
+  tape_writer tape;
+private:
+  /** Next write location in the string buf for stage 2 parsing */
+  uint8_t *current_string_buf_loc;
+
+  simdjson_inline tape_builder(dom::document &doc) noexcept;
+
+  simdjson_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
+  simdjson_inline void start_container(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_warn_unused simdjson_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
+  simdjson_inline void on_end_string(uint8_t *dst) noexcept;
+}; // struct tape_builder
+
+template<bool STREAMING>
+simdjson_warn_unused simdjson_inline error_code tape_builder::parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept {
+  dom_parser.doc = &doc;
+  json_iterator iter(dom_parser, STREAMING ? dom_parser.next_structural_index : 0);
+  tape_builder builder(doc);
+  return iter.walk_document<STREAMING>(builder);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_root_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
+  constexpr uint32_t start_tape_index = 0;
+  tape.append(start_tape_index, internal::tape_type::ROOT);
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter), internal::tape_type::ROOT);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
+  return visit_string(iter, key, true);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].count++; // we have a key value pair in the object at parser.dom_parser.depth - 1
+  return SUCCESS;
+}
+
+simdjson_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
+  iter.log_value(key ? "key" : "string");
+  uint8_t *dst = on_start_string(iter);
+  dst = stringparsing::parse_string(value+1, dst, false); // We do not allow replacement when the escape characters are invalid.
+  if (dst == nullptr) {
+    iter.log_error("Invalid escape in string");
+    return STRING_ERROR;
+  }
+  on_end_string(dst);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
+  return visit_string(iter, value);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("number");
+  error_code err = numberparsing::parse_number(value, tape);
+  if (simdjson_unlikely(err == BIGINT_ERROR &&
+      iter.dom_parser._number_as_string)) {
+    // Write big integer to string buffer using the same format as strings.
+    // Scan digits the same way parse_number does (skip optional '-', then digits).
+    const uint8_t *p = value;
+    if (*p == '-') p++;
+    while (numberparsing::is_digit(*p)) p++;
+    size_t len = size_t(p - value);
+    tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::BIGINT);
+    uint8_t *dst = current_string_buf_loc + sizeof(uint32_t);
+    memcpy(dst, value, len);
+    dst += len;
+    on_end_string(dst);
+    return SUCCESS;
+  }
+  return err;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
+  //
+  // We need to make a copy to make sure that the string is space terminated.
+  // This is not about padding the input, which should already padded up
+  // to len + SIMDJSON_PADDING. However, we have no control at this stage
+  // on how the padding was done. What if the input string was padded with nulls?
+  // It is quite common for an input string to have an extra null character (C string).
+  // We do not want to allow 9\0 (where \0 is the null character) inside a JSON
+  // document, but the string "9\0" by itself is fine. So we make a copy and
+  // pad the input with spaces when we know that there is just one input element.
+  // This copy is relatively expensive, but it will almost never be called in
+  // practice unless you are in the strange scenario where you have many JSON
+  // documents made of single atoms.
+  //
+  std::unique_ptr<uint8_t[]>copy(new (std::nothrow) uint8_t[iter.remaining_len() + SIMDJSON_PADDING]);
+  if (copy.get() == nullptr) { return MEMALLOC; }
+  std::memcpy(copy.get(), value, iter.remaining_len());
+  std::memset(copy.get() + iter.remaining_len(), ' ', SIMDJSON_PADDING);
+  error_code error = visit_number(iter, copy.get());
+  return error;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value)) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value, iter.remaining_len())) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value)) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value, iter.remaining_len())) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value)) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value, iter.remaining_len())) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+// private:
+
+simdjson_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
+  return uint32_t(tape.next_tape_loc - iter.dom_parser.doc->tape.get());
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  auto start_index = next_tape_index(iter);
+  tape.append(start_index+2, start);
+  tape.append(start_index, end);
+  return SUCCESS;
+}
+
+simdjson_inline void tape_builder::start_container(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].tape_index = next_tape_index(iter);
+  iter.dom_parser.open_containers[iter.depth].count = 0;
+  tape.skip(); // We don't actually *write* the start element until the end.
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  // Write the ending tape element, pointing at the start location
+  const uint32_t start_tape_index = iter.dom_parser.open_containers[iter.depth].tape_index;
+  tape.append(start_tape_index, end);
+  // Write the start tape element, pointing at the end location (and including count)
+  // count can overflow if it exceeds 24 bits... so we saturate
+  // the convention being that a cnt of 0xffffff or more is undetermined in value (>=  0xffffff).
+  const uint32_t count = iter.dom_parser.open_containers[iter.depth].count;
+  const uint32_t cntsat = count > 0xFFFFFF ? 0xFFFFFF : count;
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter) | (uint64_t(cntsat) << 32), start);
+  return SUCCESS;
+}
+
+simdjson_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
+  // we advance the point, accounting for the fact that we have a NULL termination
+  tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::STRING);
+  return current_string_buf_loc + sizeof(uint32_t);
+}
+
+simdjson_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
+  uint32_t str_length = uint32_t(dst - (current_string_buf_loc + sizeof(uint32_t)));
+  // TODO check for overflow in case someone has a crazy string (>=4GB?)
+  // But only add the overflow check when the document itself exceeds 4GB
+  // Currently unneeded because we refuse to parse docs larger or equal to 4GB.
+  memcpy(current_string_buf_loc, &str_length, sizeof(uint32_t));
+  // NULL termination is still handy if you expect all your strings to
+  // be NULL terminated? It comes at a small cost
+  *dst = 0;
+  current_string_buf_loc = dst + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+/* end file generic/stage2/tape_builder.h for icelake */
+/* end file generic/stage2/amalgamated.h for icelake */
+
+#undef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+//
+// Stage 1
+//
+
+namespace simdjson {
+namespace icelake {
+
+simdjson_warn_unused error_code implementation::create_dom_parser_implementation(
+  size_t capacity,
+  size_t max_depth,
+  std::unique_ptr<internal::dom_parser_implementation>& dst
+) const noexcept {
+  dst.reset( new (std::nothrow) dom_parser_implementation() );
+  if (!dst) { return MEMALLOC; }
+  if (auto err = dst->set_capacity(capacity))
+    return err;
+  if (auto err = dst->set_max_depth(max_depth))
+    return err;
+  return SUCCESS;
+}
+
+namespace {
+
+using namespace simd;
+
+// This identifies structural characters (comma, colon, braces, brackets),
+// and ASCII white-space ('\r','\n','\t',' ').
+simdjson_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
+  // These lookups rely on the fact that anything < 127 will match the lower 4 bits, which is why
+  // we can't use the generic lookup_16.
+  const auto whitespace_table = simd8<uint8_t>::repeat_16(' ', 100, 100, 100, 17, 100, 113, 2, 100, '\t', '\n', 112, 100, '\r', 100, 100);
+
+  // The 6 operators (:,[]{}) have these values:
+  //
+  // , 2C
+  // : 3A
+  // [ 5B
+  // { 7B
+  // ] 5D
+  // } 7D
+  //
+  // If you use | 0x20 to turn [ and ] into { and }, the lower 4 bits of each character is unique.
+  // We exploit this, using a simd 4-bit lookup to tell us which character match against, and then
+  // match it (against | 0x20).
+  //
+  // To prevent recognizing other characters, everything else gets compared with 0, which cannot
+  // match due to the | 0x20.
+  //
+  // NOTE: Due to the | 0x20, this ALSO treats <FF> and <SUB> (control characters 0C and 1A) like ,
+  // and :. This gets caught in stage 2, which checks the actual character to ensure the right
+  // operators are in the right places.
+  const auto op_table = simd8<uint8_t>::repeat_16(
+    0, 0, 0, 0,
+    0, 0, 0, 0,
+    0, 0, ':', '{', // : = 3A, [ = 5B, { = 7B
+    ',', '}', 0, 0  // , = 2C, ] = 5D, } = 7D
+  );
+
+  // We compute whitespace and op separately. If later code only uses one or the
+  // other, given the fact that all functions are aggressively inlined, we can
+  // hope that useless computations will be omitted. This is namely case when
+  // minifying (we only need whitespace).
+
+  const uint64_t whitespace = in.eq({
+    _mm512_shuffle_epi8(whitespace_table, in.chunks[0])
+  });
+  // Turn [ and ] into { and }
+  const simd8x64<uint8_t> curlified{
+    in.chunks[0] | 0x20
+  };
+  const uint64_t op = curlified.eq({
+    _mm512_shuffle_epi8(op_table, in.chunks[0])
+  });
+
+  return { whitespace, op };
+}
+
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input) {
+  return input.reduce_or().is_ascii();
+}
+
+simdjson_unused simdjson_inline simd8<bool> must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_second_byte = prev1.saturating_sub(0xc0u-1); // Only 11______ will be > 0
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-1); // Only 111_____ will be > 0
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-1); // Only 1111____ will be > 0
+  // Caller requires a bool (all 1's). All values resulting from the subtraction will be <= 64, so signed comparison is fine.
+  return simd8<int8_t>(is_second_byte | is_third_byte | is_fourth_byte) > int8_t(0);
+}
+
+simdjson_inline simd8<uint8_t> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-0x80); // Only 111_____ will be >= 0x80
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-0x80); // Only 1111____ will be >= 0x80
+  return is_third_byte | is_fourth_byte;
+}
+
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+
+/**
+ * We provide a custom version of bit_indexer::write using
+ * naked intrinsics.
+ * TODO: make this code more elegant.
+ */
+// Under GCC 12, the intrinsic _mm512_extracti32x4_epi32 may generate 'maybe uninitialized'.
+// as a workaround, we disable warnings within the following function.
+SIMDJSON_PUSH_DISABLE_ALL_WARNINGS
+namespace simdjson { namespace icelake { namespace { namespace stage1 {
+simdjson_inline void bit_indexer::write(uint32_t idx, uint64_t bits) {
+    // In some instances, the next branch is expensive because it is mispredicted.
+    // Unfortunately, in other cases,
+    // it helps tremendously.
+    if (bits == 0) { return; }
+
+    const __m512i indexes = _mm512_maskz_compress_epi8(bits, _mm512_set_epi32(
+      0x3f3e3d3c, 0x3b3a3938, 0x37363534, 0x33323130,
+      0x2f2e2d2c, 0x2b2a2928, 0x27262524, 0x23222120,
+      0x1f1e1d1c, 0x1b1a1918, 0x17161514, 0x13121110,
+      0x0f0e0d0c, 0x0b0a0908, 0x07060504, 0x03020100
+    ));
+    const __m512i start_index = _mm512_set1_epi32(idx);
+
+    const auto count = count_ones(bits);
+    __m512i t0 = _mm512_cvtepu8_epi32(_mm512_castsi512_si128(indexes));
+    _mm512_storeu_si512(this->tail, _mm512_add_epi32(t0, start_index));
+
+    if(count > 16) {
+      const __m512i t1 = _mm512_cvtepu8_epi32(_mm512_extracti32x4_epi32(indexes, 1));
+      _mm512_storeu_si512(this->tail + 16, _mm512_add_epi32(t1, start_index));
+      if(count > 32) {
+        const __m512i t2 = _mm512_cvtepu8_epi32(_mm512_extracti32x4_epi32(indexes, 2));
+        _mm512_storeu_si512(this->tail + 32, _mm512_add_epi32(t2, start_index));
+        if(count > 48) {
+          const __m512i t3 = _mm512_cvtepu8_epi32(_mm512_extracti32x4_epi32(indexes, 3));
+          _mm512_storeu_si512(this->tail + 48, _mm512_add_epi32(t3, start_index));
+        }
+      }
+    }
+    this->tail += count;
+}
+}}}}
+SIMDJSON_POP_DISABLE_WARNINGS
+
+//
+// Stage 2
+//
+
+//
+// Implementation-specific overrides
+//
+namespace simdjson {
+namespace icelake {
+
+simdjson_warn_unused error_code implementation::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept {
+  return icelake::stage1::json_minifier::minify<128>(buf, len, dst, dst_len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage1(const uint8_t *_buf, size_t _len, stage1_mode streaming) noexcept {
+  this->buf = _buf;
+  this->len = _len;
+  return icelake::stage1::json_structural_indexer::index<128>(_buf, _len, *this, streaming);
+}
+
+simdjson_warn_unused bool implementation::validate_utf8(const char *buf, size_t len) const noexcept {
+  return icelake::stage1::generic_validate_utf8(buf,len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<false>(*this, _doc);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2_next(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<true>(*this, _doc);
+}
+
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_string(const uint8_t *src, uint8_t *dst, bool replacement_char) const noexcept {
+  return icelake::stringparsing::parse_string(src, dst, replacement_char);
+}
+
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept {
+  return icelake::stringparsing::parse_wobbly_string(src, dst);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::parse(const uint8_t *_buf, size_t _len, dom::document &_doc) noexcept {
+  auto error = stage1(_buf, _len, stage1_mode::regular);
+  if (error) { return error; }
+  return stage2(_doc);
+}
+
+} // namespace icelake
+} // namespace simdjson
+
+/* including simdjson/icelake/end.h: #include <simdjson/icelake/end.h> */
+/* begin file simdjson/icelake/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#if !SIMDJSON_CAN_ALWAYS_RUN_ICELAKE
+SIMDJSON_UNTARGET_REGION
+#endif
+
+/* undefining SIMDJSON_IMPLEMENTATION from "icelake" */
+#undef SIMDJSON_IMPLEMENTATION
+/* end file simdjson/icelake/end.h */
+
+#endif // SIMDJSON_SRC_ICELAKE_CPP
+/* end file icelake.cpp */
+#endif
+#if SIMDJSON_IMPLEMENTATION_PPC64
+/* including ppc64.cpp: #include <ppc64.cpp> */
+/* begin file ppc64.cpp */
+#ifndef SIMDJSON_SRC_PPC64_CPP
+#define SIMDJSON_SRC_PPC64_CPP
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+/* including simdjson/ppc64.h: #include <simdjson/ppc64.h> */
+/* begin file simdjson/ppc64.h */
+#ifndef SIMDJSON_PPC64_H
+#define SIMDJSON_PPC64_H
+
+/* including simdjson/ppc64/begin.h: #include "simdjson/ppc64/begin.h" */
+/* begin file simdjson/ppc64/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "ppc64" */
+#define SIMDJSON_IMPLEMENTATION ppc64
+/* including simdjson/ppc64/base.h: #include "simdjson/ppc64/base.h" */
+/* begin file simdjson/ppc64/base.h */
+#ifndef SIMDJSON_PPC64_BASE_H
+#define SIMDJSON_PPC64_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+/**
+ * Implementation for ALTIVEC (PPC64).
+ */
+namespace ppc64 {
+
+class implementation;
+
+namespace {
+namespace simd {
+template <typename T> struct simd8;
+template <typename T> struct simd8x64;
+} // namespace simd
+} // unnamed namespace
+
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_PPC64_BASE_H
+/* end file simdjson/ppc64/base.h */
+/* including simdjson/ppc64/intrinsics.h: #include "simdjson/ppc64/intrinsics.h" */
+/* begin file simdjson/ppc64/intrinsics.h */
+#ifndef SIMDJSON_PPC64_INTRINSICS_H
+#define SIMDJSON_PPC64_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This should be the correct header whether
+// you use visual studio or other compilers.
+#include <altivec.h>
+
+// These are defined by altivec.h in GCC toolchain, it is safe to undef them.
+#ifdef bool
+#undef bool
+#endif
+
+#ifdef vector
+#undef vector
+#endif
+
+static_assert(sizeof(__vector unsigned char) <= simdjson::SIMDJSON_PADDING, "insufficient padding for ppc64");
+
+#endif //  SIMDJSON_PPC64_INTRINSICS_H
+/* end file simdjson/ppc64/intrinsics.h */
+/* including simdjson/ppc64/bitmanipulation.h: #include "simdjson/ppc64/bitmanipulation.h" */
+/* begin file simdjson/ppc64/bitmanipulation.h */
+#ifndef SIMDJSON_PPC64_BITMANIPULATION_H
+#define SIMDJSON_PPC64_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  unsigned long ret;
+  // Search the mask data from least significant bit (LSB)
+  // to the most significant bit (MSB) for a set bit (1).
+  _BitScanForward64(&ret, input_num);
+  return (int)ret;
+#else  // SIMDJSON_REGULAR_VISUAL_STUDIO
+  return __builtin_ctzll(input_num);
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return input_num & (input_num - 1);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  unsigned long leading_zero = 0;
+  // Search the mask data from most significant bit (MSB)
+  // to least significant bit (LSB) for a set bit (1).
+  if (_BitScanReverse64(&leading_zero, input_num))
+    return (int)(63 - leading_zero);
+  else
+    return 64;
+#else
+  return __builtin_clzll(input_num);
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+simdjson_inline int count_ones(uint64_t input_num) {
+  // note: we do not support legacy 32-bit Windows in this kernel
+  return __popcnt64(input_num); // Visual Studio wants two underscores
+}
+#else
+simdjson_inline int count_ones(uint64_t input_num) {
+  return __builtin_popcountll(input_num);
+}
+#endif
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2,
+                                         uint64_t *result) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  *result = value1 + value2;
+  return *result < value1;
+#else
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+#endif
+}
+
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_PPC64_BITMANIPULATION_H
+/* end file simdjson/ppc64/bitmanipulation.h */
+/* including simdjson/ppc64/bitmask.h: #include "simdjson/ppc64/bitmask.h" */
+/* begin file simdjson/ppc64/bitmask.h */
+#ifndef SIMDJSON_PPC64_BITMASK_H
+#define SIMDJSON_PPC64_BITMASK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is
+// encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(uint64_t bitmask) {
+  // You can use the version below, however gcc sometimes miscompiles
+  // vec_pmsum_be, it happens somewhere around between 8 and 9th version.
+  // The performance boost was not noticeable, falling back to a usual
+  // implementation.
+  //   __vector unsigned long long all_ones = {~0ull, ~0ull};
+  //   __vector unsigned long long mask = {bitmask, 0};
+  //   // Clang and GCC return different values for pmsum for ull so cast it to one.
+  //   // Generally it is not specified by ALTIVEC ISA what is returned by
+  //   // vec_pmsum_be.
+  // #if defined(__LITTLE_ENDIAN__)
+  //   return (uint64_t)(((__vector unsigned long long)vec_pmsum_be(all_ones, mask))[0]);
+  // #else
+  //   return (uint64_t)(((__vector unsigned long long)vec_pmsum_be(all_ones, mask))[1]);
+  // #endif
+  bitmask ^= bitmask << 1;
+  bitmask ^= bitmask << 2;
+  bitmask ^= bitmask << 4;
+  bitmask ^= bitmask << 8;
+  bitmask ^= bitmask << 16;
+  bitmask ^= bitmask << 32;
+  return bitmask;
+}
+
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif
+/* end file simdjson/ppc64/bitmask.h */
+/* including simdjson/ppc64/numberparsing_defs.h: #include "simdjson/ppc64/numberparsing_defs.h" */
+/* begin file simdjson/ppc64/numberparsing_defs.h */
+#ifndef SIMDJSON_PPC64_NUMBERPARSING_DEFS_H
+#define SIMDJSON_PPC64_NUMBERPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+#if defined(__linux__)
+#include <byteswap.h>
+#elif defined(__FreeBSD__)
+#include <sys/endian.h>
+#endif
+
+namespace simdjson {
+namespace ppc64 {
+namespace numberparsing {
+
+// we don't have appropriate instructions, so let us use a scalar function
+// credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  uint64_t val;
+  std::memcpy(&val, chars, sizeof(uint64_t));
+#ifdef __BIG_ENDIAN__
+#if defined(__linux__)
+  val = bswap_64(val);
+#elif defined(__FreeBSD__)
+  val = bswap64(val);
+#endif
+#endif
+  val = (val & 0x0F0F0F0F0F0F0F0F) * 2561 >> 8;
+  val = (val & 0x00FF00FF00FF00FF) * 6553601 >> 16;
+  return uint32_t((val & 0x0000FFFF0000FFFF) * 42949672960001 >> 32);
+}
+
+/** @private */
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+#if SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+#if SIMDJSON_IS_ARM64
+  // ARM64 has native support for 64-bit multiplications, no need to emultate
+  answer.high = __umulh(value1, value2);
+  answer.low = value1 * value2;
+#else
+  answer.low = _umul128(value1, value2, &answer.high); // _umul128 not available on ARM64
+#endif // SIMDJSON_IS_ARM64
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+#endif
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace ppc64
+} // namespace simdjson
+
+#ifndef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_IS_BIG_ENDIAN
+#define SIMDJSON_SWAR_NUMBER_PARSING 0
+#else
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+#endif
+#endif
+
+#endif // SIMDJSON_PPC64_NUMBERPARSING_DEFS_H
+/* end file simdjson/ppc64/numberparsing_defs.h */
+/* including simdjson/ppc64/simd.h: #include "simdjson/ppc64/simd.h" */
+/* begin file simdjson/ppc64/simd.h */
+#ifndef SIMDJSON_PPC64_SIMD_H
+#define SIMDJSON_PPC64_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <type_traits>
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace simd {
+
+using __m128i = __vector unsigned char;
+
+template <typename Child> struct base {
+  __m128i value;
+
+  // Zero constructor
+  simdjson_inline base() : value{__m128i()} {}
+
+  // Conversion from SIMD register
+  simdjson_inline base(const __m128i _value) : value(_value) {}
+
+  // Conversion to SIMD register
+  simdjson_inline operator const __m128i &() const {
+    return this->value;
+  }
+  simdjson_inline operator __m128i &() { return this->value; }
+
+  // Bit operations
+  simdjson_inline Child operator|(const Child other) const {
+    return vec_or(this->value, (__m128i)other);
+  }
+  simdjson_inline Child operator&(const Child other) const {
+    return vec_and(this->value, (__m128i)other);
+  }
+  simdjson_inline Child operator^(const Child other) const {
+    return vec_xor(this->value, (__m128i)other);
+  }
+  simdjson_inline Child bit_andnot(const Child other) const {
+    return vec_andc(this->value, (__m128i)other);
+  }
+  simdjson_inline Child &operator|=(const Child other) {
+    auto this_cast = static_cast<Child*>(this);
+    *this_cast = *this_cast | other;
+    return *this_cast;
+  }
+  simdjson_inline Child &operator&=(const Child other) {
+    auto this_cast = static_cast<Child*>(this);
+    *this_cast = *this_cast & other;
+    return *this_cast;
+  }
+  simdjson_inline Child &operator^=(const Child other) {
+    auto this_cast = static_cast<Child*>(this);
+    *this_cast = *this_cast ^ other;
+    return *this_cast;
+  }
+};
+
+template <typename T, typename Mask = simd8<bool>>
+struct base8 : base<simd8<T>> {
+  typedef uint16_t bitmask_t;
+  typedef uint32_t bitmask2_t;
+
+  simdjson_inline base8() : base<simd8<T>>() {}
+  simdjson_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}
+
+  friend simdjson_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) {
+    return (__m128i)vec_cmpeq(lhs.value, (__m128i)rhs);
+  }
+
+  static const int SIZE = sizeof(base<simd8<T>>::value);
+
+  template <int N = 1>
+  simdjson_inline simd8<T> prev(simd8<T> prev_chunk) const {
+    __m128i chunk = this->value;
+#ifdef __LITTLE_ENDIAN__
+    chunk = (__m128i)vec_reve(this->value);
+    prev_chunk = (__m128i)vec_reve((__m128i)prev_chunk);
+#endif
+    chunk = (__m128i)vec_sld((__m128i)prev_chunk, (__m128i)chunk, 16 - N);
+#ifdef __LITTLE_ENDIAN__
+    chunk = (__m128i)vec_reve((__m128i)chunk);
+#endif
+    return chunk;
+  }
+};
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd8<bool> : base8<bool> {
+  static simdjson_inline simd8<bool> splat(bool _value) {
+    return (__m128i)vec_splats((unsigned char)(-(!!_value)));
+  }
+
+  simdjson_inline simd8() : base8<bool>() {}
+  simdjson_inline simd8(const __m128i _value)
+      : base8<bool>(_value) {}
+  // Splat constructor
+  simdjson_inline simd8(bool _value)
+      : base8<bool>(splat(_value)) {}
+
+  simdjson_inline int to_bitmask() const {
+    __vector unsigned long long result;
+    const __m128i perm_mask = {0x78, 0x70, 0x68, 0x60, 0x58, 0x50, 0x48, 0x40,
+                               0x38, 0x30, 0x28, 0x20, 0x18, 0x10, 0x08, 0x00};
+
+    result = ((__vector unsigned long long)vec_vbpermq((__m128i)this->value,
+                                                       (__m128i)perm_mask));
+#ifdef __LITTLE_ENDIAN__
+    return static_cast<int>(result[1]);
+#else
+    return static_cast<int>(result[0]);
+#endif
+  }
+  simdjson_inline bool any() const {
+    return !vec_all_eq(this->value, (__m128i)vec_splats(0));
+  }
+  simdjson_inline simd8<bool> operator~() const {
+    return this->value ^ (__m128i)splat(true);
+  }
+};
+
+template <typename T> struct base8_numeric : base8<T> {
+  static simdjson_inline simd8<T> splat(T value) {
+    (void)value;
+    return (__m128i)vec_splats(value);
+  }
+  static simdjson_inline simd8<T> zero() { return splat(0); }
+  static simdjson_inline simd8<T> load(const T values[16]) {
+    return (__m128i)(vec_vsx_ld(0, reinterpret_cast<const uint8_t *>(values)));
+  }
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  static simdjson_inline simd8<T> repeat_16(T v0, T v1, T v2, T v3, T v4,
+                                                   T v5, T v6, T v7, T v8, T v9,
+                                                   T v10, T v11, T v12, T v13,
+                                                   T v14, T v15) {
+    return simd8<T>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
+                    v14, v15);
+  }
+
+  simdjson_inline base8_numeric() : base8<T>() {}
+  simdjson_inline base8_numeric(const __m128i _value)
+      : base8<T>(_value) {}
+
+  // Store to array
+  simdjson_inline void store(T dst[16]) const {
+    vec_vsx_st(this->value, 0, reinterpret_cast<__m128i *>(dst));
+  }
+
+  // Override to distinguish from bool version
+  simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+  // Addition/subtraction are the same for signed and unsigned
+  simdjson_inline simd8<T> operator+(const simd8<T> other) const {
+    return (__m128i)((__m128i)this->value + (__m128i)other);
+  }
+  simdjson_inline simd8<T> operator-(const simd8<T> other) const {
+    return (__m128i)((__m128i)this->value - (__m128i)other);
+  }
+  simdjson_inline simd8<T> &operator+=(const simd8<T> other) {
+    *this = *this + other;
+    return *static_cast<simd8<T> *>(this);
+  }
+  simdjson_inline simd8<T> &operator-=(const simd8<T> other) {
+    *this = *this - other;
+    return *static_cast<simd8<T> *>(this);
+  }
+
+  // Perform a lookup assuming the value is between 0 and 16 (undefined behavior
+  // for out of range values)
+  template <typename L>
+  simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+    return (__m128i)vec_perm((__m128i)lookup_table, (__m128i)lookup_table, this->value);
+  }
+
+  // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted
+  // as a bitset). Passing a 0 value for mask would be equivalent to writing out
+  // every byte to output. Only the first 16 - count_ones(mask) bytes of the
+  // result are significant but 16 bytes get written. Design consideration: it
+  // seems like a function with the signature simd8<L> compress(uint32_t mask)
+  // would be sensible, but the AVX ISA makes this kind of approach difficult.
+  template <typename L>
+  simdjson_inline void compress(uint16_t mask, L *output) const {
+    using internal::BitsSetTable256mul2;
+    using internal::pshufb_combine_table;
+    using internal::thintable_epi8;
+    // this particular implementation was inspired by work done by @animetosho
+    // we do it in two steps, first 8 bytes and then second 8 bytes
+    uint8_t mask1 = uint8_t(mask);      // least significant 8 bits
+    uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
+    // next line just loads the 64-bit values thintable_epi8[mask1] and
+    // thintable_epi8[mask2] into a 128-bit register, using only
+    // two instructions on most compilers.
+#ifdef __LITTLE_ENDIAN__
+    __m128i shufmask = (__m128i)(__vector unsigned long long){
+        thintable_epi8[mask1], thintable_epi8[mask2]};
+#else
+    __m128i shufmask = (__m128i)(__vector unsigned long long){
+        thintable_epi8[mask2], thintable_epi8[mask1]};
+    shufmask = (__m128i)vec_reve((__m128i)shufmask);
+#endif
+    // we increment by 0x08 the second half of the mask
+    shufmask = ((__m128i)shufmask) +
+               ((__m128i)(__vector int){0, 0, 0x08080808, 0x08080808});
+
+    // this is the version "nearly pruned"
+    __m128i pruned = vec_perm(this->value, this->value, shufmask);
+    // we still need to put the two halves together.
+    // we compute the popcount of the first half:
+    int pop1 = BitsSetTable256mul2[mask1];
+    // then load the corresponding mask, what it does is to write
+    // only the first pop1 bytes from the first 8 bytes, and then
+    // it fills in with the bytes from the second 8 bytes + some filling
+    // at the end.
+    __m128i compactmask =
+        vec_vsx_ld(0, reinterpret_cast<const uint8_t *>(pshufb_combine_table + pop1 * 8));
+    __m128i answer = vec_perm(pruned, (__m128i)vec_splats(0), compactmask);
+    vec_vsx_st(answer, 0, reinterpret_cast<__m128i *>(output));
+  }
+
+  template <typename L>
+  simdjson_inline simd8<L>
+  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
+            L replace5, L replace6, L replace7, L replace8, L replace9,
+            L replace10, L replace11, L replace12, L replace13, L replace14,
+            L replace15) const {
+    return lookup_16(simd8<L>::repeat_16(
+        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
+        replace7, replace8, replace9, replace10, replace11, replace12,
+        replace13, replace14, replace15));
+  }
+};
+
+// Signed bytes
+template <> struct simd8<int8_t> : base8_numeric<int8_t> {
+  simdjson_inline simd8() : base8_numeric<int8_t>() {}
+  simdjson_inline simd8(const __m128i _value)
+      : base8_numeric<int8_t>(_value) {}
+  // Splat constructor
+  simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+  // Array constructor
+  simdjson_inline simd8(const int8_t *values) : simd8(load(values)) {}
+  // Member-by-member initialization
+  simdjson_inline simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3,
+                               int8_t v4, int8_t v5, int8_t v6, int8_t v7,
+                               int8_t v8, int8_t v9, int8_t v10, int8_t v11,
+                               int8_t v12, int8_t v13, int8_t v14, int8_t v15)
+      : simd8((__m128i)(__vector signed char){v0, v1, v2, v3, v4, v5, v6, v7,
+                                              v8, v9, v10, v11, v12, v13, v14,
+                                              v15}) {}
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdjson_inline static simd8<int8_t>
+  repeat_16(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
+            int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
+            int8_t v12, int8_t v13, int8_t v14, int8_t v15) {
+    return simd8<int8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                         v13, v14, v15);
+  }
+
+  // Order-sensitive comparisons
+  simdjson_inline simd8<int8_t>
+  max_val(const simd8<int8_t> other) const {
+    return (__m128i)vec_max((__vector signed char)this->value,
+                            (__vector signed char)(__m128i)other);
+  }
+  simdjson_inline simd8<int8_t>
+  min_val(const simd8<int8_t> other) const {
+    return (__m128i)vec_min((__vector signed char)this->value,
+                            (__vector signed char)(__m128i)other);
+  }
+  simdjson_inline simd8<bool>
+  operator>(const simd8<int8_t> other) const {
+    return (__m128i)vec_cmpgt((__vector signed char)this->value,
+                              (__vector signed char)(__m128i)other);
+  }
+  simdjson_inline simd8<bool>
+  operator<(const simd8<int8_t> other) const {
+    return (__m128i)vec_cmplt((__vector signed char)this->value,
+                              (__vector signed char)(__m128i)other);
+  }
+};
+
+// Unsigned bytes
+template <> struct simd8<uint8_t> : base8_numeric<uint8_t> {
+  simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+  simdjson_inline simd8(const __m128i _value)
+      : base8_numeric<uint8_t>(_value) {}
+  // Splat constructor
+  simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+  // Array constructor
+  simdjson_inline simd8(const uint8_t *values) : simd8(load(values)) {}
+  // Member-by-member initialization
+  simdjson_inline
+  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
+        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
+        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15)
+      : simd8((__m128i){v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                        v13, v14, v15}) {}
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdjson_inline static simd8<uint8_t>
+  repeat_16(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4,
+            uint8_t v5, uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9,
+            uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14,
+            uint8_t v15) {
+    return simd8<uint8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                          v13, v14, v15);
+  }
+
+  // Saturated math
+  simdjson_inline simd8<uint8_t>
+  saturating_add(const simd8<uint8_t> other) const {
+    return (__m128i)vec_adds(this->value, (__m128i)other);
+  }
+  simdjson_inline simd8<uint8_t>
+  saturating_sub(const simd8<uint8_t> other) const {
+    return (__m128i)vec_subs(this->value, (__m128i)other);
+  }
+
+  // Order-specific operations
+  simdjson_inline simd8<uint8_t>
+  max_val(const simd8<uint8_t> other) const {
+    return (__m128i)vec_max(this->value, (__m128i)other);
+  }
+  simdjson_inline simd8<uint8_t>
+  min_val(const simd8<uint8_t> other) const {
+    return (__m128i)vec_min(this->value, (__m128i)other);
+  }
+  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+  simdjson_inline simd8<uint8_t>
+  gt_bits(const simd8<uint8_t> other) const {
+    return this->saturating_sub(other);
+  }
+  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+  simdjson_inline simd8<uint8_t>
+  lt_bits(const simd8<uint8_t> other) const {
+    return other.saturating_sub(*this);
+  }
+  simdjson_inline simd8<bool>
+  operator<=(const simd8<uint8_t> other) const {
+    return other.max_val(*this) == other;
+  }
+  simdjson_inline simd8<bool>
+  operator>=(const simd8<uint8_t> other) const {
+    return other.min_val(*this) == other;
+  }
+  simdjson_inline simd8<bool>
+  operator>(const simd8<uint8_t> other) const {
+    return this->gt_bits(other).any_bits_set();
+  }
+  simdjson_inline simd8<bool>
+  operator<(const simd8<uint8_t> other) const {
+    return this->gt_bits(other).any_bits_set();
+  }
+
+  // Bit-specific operations
+  simdjson_inline simd8<bool> bits_not_set() const {
+    return (__m128i)vec_cmpeq(this->value, (__m128i)vec_splats(uint8_t(0)));
+  }
+  simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const {
+    return (*this & bits).bits_not_set();
+  }
+  simdjson_inline simd8<bool> any_bits_set() const {
+    return ~this->bits_not_set();
+  }
+  simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
+    return ~this->bits_not_set(bits);
+  }
+  simdjson_inline bool bits_not_set_anywhere() const {
+    return vec_all_eq(this->value, (__m128i)vec_splats(0));
+  }
+  simdjson_inline bool any_bits_set_anywhere() const {
+    return !bits_not_set_anywhere();
+  }
+  simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const {
+    return vec_all_eq(vec_and(this->value, (__m128i)bits),
+                      (__m128i)vec_splats(0));
+  }
+  simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const {
+    return !bits_not_set_anywhere(bits);
+  }
+  template <int N> simdjson_inline simd8<uint8_t> shr() const {
+    return simd8<uint8_t>(
+        (__m128i)vec_sr(this->value, (__m128i)vec_splat_u8(N)));
+  }
+  template <int N> simdjson_inline simd8<uint8_t> shl() const {
+    return simd8<uint8_t>(
+        (__m128i)vec_sl(this->value, (__m128i)vec_splat_u8(N)));
+  }
+};
+
+template <typename T> struct simd8x64 {
+  static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+  static_assert(NUM_CHUNKS == 4,
+                "PPC64 kernel should use four registers per 64-byte block.");
+  const simd8<T> chunks[NUM_CHUNKS];
+  template<int idx> simd8<uint8_t> get() const { return idx < NUM_CHUNKS ? chunks[idx] : simd8<T>(); }
+
+  simd8x64(const simd8x64<T> &o) = delete; // no copy allowed
+  simd8x64<T> &
+  operator=(const simd8<T>& other) = delete; // no assignment allowed
+  simd8x64() = delete;                      // no default constructor allowed
+
+  simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1,
+                                  const simd8<T> chunk2, const simd8<T> chunk3)
+      : chunks{chunk0, chunk1, chunk2, chunk3} {}
+  simdjson_inline simd8x64(const T ptr[64])
+      : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr + 16),
+               simd8<T>::load(ptr + 32), simd8<T>::load(ptr + 48)} {}
+
+  simdjson_inline void store(T ptr[64]) const {
+    this->chunks[0].store(ptr + sizeof(simd8<T>) * 0);
+    this->chunks[1].store(ptr + sizeof(simd8<T>) * 1);
+    this->chunks[2].store(ptr + sizeof(simd8<T>) * 2);
+    this->chunks[3].store(ptr + sizeof(simd8<T>) * 3);
+  }
+
+  simdjson_inline simd8<T> reduce_or() const {
+    return (this->chunks[0] | this->chunks[1]) |
+           (this->chunks[2] | this->chunks[3]);
+  }
+
+  simdjson_inline uint64_t compress(uint64_t mask, T *output) const {
+    this->chunks[0].compress(uint16_t(mask), output);
+    this->chunks[1].compress(uint16_t(mask >> 16),
+                             output + 16 - count_ones(mask & 0xFFFF));
+    this->chunks[2].compress(uint16_t(mask >> 32),
+                             output + 32 - count_ones(mask & 0xFFFFFFFF));
+    this->chunks[3].compress(uint16_t(mask >> 48),
+                             output + 48 - count_ones(mask & 0xFFFFFFFFFFFF));
+    return 64 - count_ones(mask);
+  }
+
+  simdjson_inline uint64_t to_bitmask() const {
+    uint64_t r0 = uint32_t(this->chunks[0].to_bitmask());
+    uint64_t r1 = this->chunks[1].to_bitmask();
+    uint64_t r2 = this->chunks[2].to_bitmask();
+    uint64_t r3 = this->chunks[3].to_bitmask();
+    return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
+  }
+
+  simdjson_inline uint64_t eq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] == mask, this->chunks[1] == mask,
+                          this->chunks[2] == mask, this->chunks[3] == mask)
+        .to_bitmask();
+  }
+
+  simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+    return simd8x64<bool>(this->chunks[0] == other.chunks[0],
+                          this->chunks[1] == other.chunks[1],
+                          this->chunks[2] == other.chunks[2],
+                          this->chunks[3] == other.chunks[3])
+        .to_bitmask();
+  }
+
+  simdjson_inline uint64_t lteq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
+                          this->chunks[2] <= mask, this->chunks[3] <= mask)
+        .to_bitmask();
+  }
+}; // struct simd8x64<T>
+
+} // namespace simd
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_PPC64_SIMD_INPUT_H
+/* end file simdjson/ppc64/simd.h */
+/* including simdjson/ppc64/stringparsing_defs.h: #include "simdjson/ppc64/stringparsing_defs.h" */
+/* begin file simdjson/ppc64/stringparsing_defs.h */
+#ifndef SIMDJSON_PPC64_STRINGPARSING_DEFS_H
+#define SIMDJSON_PPC64_STRINGPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/simd.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 32;
+  simdjson_inline backslash_and_quote
+  copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() {
+    return ((bs_bits - 1) & quote_bits) != 0;
+  }
+  simdjson_inline bool has_backslash() { return bs_bits != 0; }
+  simdjson_inline int quote_index() {
+    return trailing_zeroes(quote_bits);
+  }
+  simdjson_inline int backslash_index() {
+    return trailing_zeroes(bs_bits);
+  }
+
+  uint32_t bs_bits;
+  uint32_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote
+backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // this can read up to 31 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1),
+                "backslash and quote finder must process fewer than "
+                "SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v0(src);
+  simd8<uint8_t> v1(src + sizeof(v0));
+  v0.store(dst);
+  v1.store(dst + sizeof(v0));
+
+  // Getting a 64-bit bitmask is much cheaper than multiple 16-bit bitmasks on
+  // PPC; therefore, we smash them together into a 64-byte mask and get the
+  // bitmask from there.
+  uint64_t bs_and_quote =
+      simd8x64<bool>(v0 == '\\', v1 == '\\', v0 == '"', v1 == '"').to_bitmask();
+  return {
+      uint32_t(bs_and_quote),      // bs_bits
+      uint32_t(bs_and_quote >> 32) // quote_bits
+  };
+}
+
+
+struct escaping {
+  static constexpr uint32_t BYTES_PROCESSED = 16;
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits != 0; }
+  simdjson_inline int escape_index() { return trailing_zeroes(escape_bits); }
+
+  uint64_t escape_bits;
+}; // struct escaping
+
+
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "escaping finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  simd8<bool> is_quote = (v == '"');
+  simd8<bool> is_backslash = (v == '\\');
+  simd8<bool> is_control = (v < 32);
+  return {
+    // We store it as a 64-bit bitmask even though we only need 16 bits.
+    uint64_t((is_backslash | is_quote | is_control).to_bitmask())
+  };
+}
+
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_PPC64_STRINGPARSING_DEFS_H
+/* end file simdjson/ppc64/stringparsing_defs.h */
+
+#define SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT 1
+/* end file simdjson/ppc64/begin.h */
+/* including simdjson/generic/amalgamated.h for ppc64: #include "simdjson/generic/amalgamated.h" */
+/* begin file simdjson/generic/amalgamated.h for ppc64 */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_GENERIC_DEPENDENCIES_H)
+#error simdjson/generic/dependencies.h must be included before simdjson/generic/amalgamated.h!
+#endif
+
+/* including simdjson/generic/base.h for ppc64: #include "simdjson/generic/base.h" */
+/* begin file simdjson/generic/base.h for ppc64 */
+#ifndef SIMDJSON_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): // If we haven't got an implementation yet, we're in the editor, editing a generic file! Just */
+/* amalgamation skipped (editor-only): // use the most advanced one we can so the most possible stuff can be tested. */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation_detection.h" */
+/* amalgamation skipped (editor-only): #if SIMDJSON_IMPLEMENTATION_ICELAKE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_HASWELL */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_WESTMERE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_ARM64 */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_PPC64 */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LASX */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LSX */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_RVV_VLS */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_FALLBACK */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/begin.h" */
+/* amalgamation skipped (editor-only): #else */
+/* amalgamation skipped (editor-only): #error "All possible implementations (including fallback) have been disabled! simdjson will not run." */
+/* amalgamation skipped (editor-only): #endif */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+
+struct open_container;
+class dom_parser_implementation;
+
+/**
+ * The type of a JSON number
+ */
+enum class number_type {
+    floating_point_number=1, /// a binary64 number
+    signed_integer,          /// a signed integer that fits in a 64-bit word using two's complement
+    unsigned_integer,        /// a positive integer larger or equal to 1<<63
+    big_integer              /// a big integer that does not fit in a 64-bit word
+};
+
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_BASE_H
+/* end file simdjson/generic/base.h for ppc64 */
+/* including simdjson/generic/jsoncharutils.h for ppc64: #include "simdjson/generic/jsoncharutils.h" */
+/* begin file simdjson/generic/jsoncharutils.h for ppc64 */
+#ifndef SIMDJSON_GENERIC_JSONCHARUTILS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_JSONCHARUTILS_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/jsoncharutils_tables.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace jsoncharutils {
+
+// return non-zero if not a structural or whitespace char
+// zero otherwise
+simdjson_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace_negated[c];
+}
+
+simdjson_inline uint32_t is_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace[c];
+}
+
+// returns a value with the high 16 bits set if not valid
+// otherwise returns the conversion of the 4 hex digits at src into the bottom
+// 16 bits of the 32-bit return register
+//
+// see
+// https://lemire.me/blog/2019/04/17/parsing-short-hexadecimal-strings-efficiently/
+static inline uint32_t hex_to_u32_nocheck(
+    const uint8_t *src) { // strictly speaking, static inline is a C-ism
+  uint32_t v1 = internal::digit_to_val32[630 + src[0]];
+  uint32_t v2 = internal::digit_to_val32[420 + src[1]];
+  uint32_t v3 = internal::digit_to_val32[210 + src[2]];
+  uint32_t v4 = internal::digit_to_val32[0 + src[3]];
+  return v1 | v2 | v3 | v4;
+}
+
+// given a code point cp, writes to c
+// the utf-8 code, outputting the length in
+// bytes, if the length is zero, the code point
+// is invalid
+//
+// This can possibly be made faster using pdep
+// and clz and table lookups, but JSON documents
+// have few escaped code points, and the following
+// function looks cheap.
+//
+// Note: we assume that surrogates are treated separately
+//
+simdjson_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
+  if (cp <= 0x7F) {
+    c[0] = uint8_t(cp);
+    return 1; // ascii
+  }
+  if (cp <= 0x7FF) {
+    c[0] = uint8_t((cp >> 6) + 192);
+    c[1] = uint8_t((cp & 63) + 128);
+    return 2; // universal plane
+    //  Surrogates are treated elsewhere...
+    //} //else if (0xd800 <= cp && cp <= 0xdfff) {
+    //  return 0; // surrogates // could put assert here
+  } else if (cp <= 0xFFFF) {
+    c[0] = uint8_t((cp >> 12) + 224);
+    c[1] = uint8_t(((cp >> 6) & 63) + 128);
+    c[2] = uint8_t((cp & 63) + 128);
+    return 3;
+  } else if (cp <= 0x10FFFF) { // if you know you have a valid code point, this
+                               // is not needed
+    c[0] = uint8_t((cp >> 18) + 240);
+    c[1] = uint8_t(((cp >> 12) & 63) + 128);
+    c[2] = uint8_t(((cp >> 6) & 63) + 128);
+    c[3] = uint8_t((cp & 63) + 128);
+    return 4;
+  }
+  // will return 0 when the code point was too large.
+  return 0; // bad r
+}
+
+#if SIMDJSON_IS_32BITS // _umul128 for x86, arm
+// this is a slow emulation routine for 32-bit
+//
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+  return x * (uint64_t)y;
+}
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+  uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
+  uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
+  uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
+  uint64_t adbc_carry = !!(adbc < ad);
+  uint64_t lo = bd + (adbc << 32);
+  *hi = __emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) +
+        (adbc_carry << 32) + !!(lo < bd);
+  return lo;
+}
+#endif
+
+} // namespace jsoncharutils
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_JSONCHARUTILS_H
+/* end file simdjson/generic/jsoncharutils.h for ppc64 */
+/* including simdjson/generic/atomparsing.h for ppc64: #include "simdjson/generic/atomparsing.h" */
+/* begin file simdjson/generic/atomparsing.h for ppc64 */
+#ifndef SIMDJSON_GENERIC_ATOMPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_ATOMPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+/// @private
+namespace atomparsing {
+
+// The string_to_uint32 is exclusively used to map literal strings to 32-bit values.
+// We use memcpy instead of a pointer cast to avoid undefined behaviors since we cannot
+// be certain that the character pointer will be properly aligned.
+// You might think that using memcpy makes this function expensive, but you'd be wrong.
+// All decent optimizing compilers (GCC, clang, Visual Studio) will compile string_to_uint32("false");
+// to the compile-time constant 1936482662.
+simdjson_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
+
+
+// Again in str4ncmp we use a memcpy to avoid undefined behavior. The memcpy may appear expensive.
+// Yet all decent optimizing compilers will compile memcpy to a single instruction, just about.
+simdjson_warn_unused
+simdjson_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
+  uint32_t srcval; // we want to avoid unaligned 32-bit loads (undefined in C/C++)
+  static_assert(sizeof(uint32_t) <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be larger than 4 bytes");
+  std::memcpy(&srcval, src, sizeof(uint32_t));
+  return srcval ^ string_to_uint32(atom);
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src) {
+  return (str4ncmp(src, "true") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_true_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "true"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src) {
+  return (str4ncmp(src+1, "alse") | jsoncharutils::is_not_structural_or_whitespace(src[5])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
+  if (len > 5) { return is_valid_false_atom(src); }
+  else if (len == 5) { return !str4ncmp(src+1, "alse"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src) {
+  return (str4ncmp(src, "null") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_null_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "null"); }
+  else { return false; }
+}
+
+} // namespace atomparsing
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_ATOMPARSING_H
+/* end file simdjson/generic/atomparsing.h for ppc64 */
+/* including simdjson/generic/dom_parser_implementation.h for ppc64: #include "simdjson/generic/dom_parser_implementation.h" */
+/* begin file simdjson/generic/dom_parser_implementation.h for ppc64 */
+#ifndef SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/dom_parser_implementation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+
+// expectation: sizeof(open_container) = 64/8.
+struct open_container {
+  uint32_t tape_index; // where, on the tape, does the scope ([,{) begins
+  uint32_t count; // how many elements in the scope
+}; // struct open_container
+
+static_assert(sizeof(open_container) == 64/8, "Open container must be 64 bits");
+
+class dom_parser_implementation final : public internal::dom_parser_implementation {
+public:
+  /** Tape location of each open { or [ */
+  std::unique_ptr<open_container[]> open_containers{};
+  /** Whether each open container is a [ or { */
+  std::unique_ptr<bool[]> is_array{};
+  /** Buffer passed to stage 1 */
+  const uint8_t *buf{};
+  /** Length passed to stage 1 */
+  size_t len{0};
+  /** Document passed to stage 2 */
+  dom::document *doc{};
+
+  inline dom_parser_implementation() noexcept;
+  inline dom_parser_implementation(dom_parser_implementation &&other) noexcept;
+  inline dom_parser_implementation &operator=(dom_parser_implementation &&other) noexcept;
+  dom_parser_implementation(const dom_parser_implementation &) = delete;
+  dom_parser_implementation &operator=(const dom_parser_implementation &) = delete;
+
+  simdjson_warn_unused error_code parse(const uint8_t *buf, size_t len, dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage1(const uint8_t *buf, size_t len, stage1_mode partial) noexcept final;
+  simdjson_warn_unused error_code stage2(dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage2_next(dom::document &doc) noexcept final;
+  simdjson_warn_unused uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) const noexcept final;
+  simdjson_warn_unused uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept final;
+  inline simdjson_warn_unused error_code set_capacity(size_t capacity) noexcept final;
+  inline simdjson_warn_unused error_code set_max_depth(size_t max_depth) noexcept final;
+private:
+  simdjson_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
+
+};
+
+} // namespace ppc64
+} // namespace simdjson
+
+namespace simdjson {
+namespace ppc64 {
+
+inline dom_parser_implementation::dom_parser_implementation() noexcept = default;
+inline dom_parser_implementation::dom_parser_implementation(dom_parser_implementation &&other) noexcept = default;
+inline dom_parser_implementation &dom_parser_implementation::operator=(dom_parser_implementation &&other) noexcept = default;
+
+// Leaving these here so they can be inlined if so desired
+inline simdjson_warn_unused error_code dom_parser_implementation::set_capacity(size_t capacity) noexcept {
+  if(capacity > SIMDJSON_MAXSIZE_BYTES) { return CAPACITY; }
+  // Stage 1 index output
+  size_t max_structures = SIMDJSON_ROUNDUP_N(capacity, 64) + 2 + 7;
+  structural_indexes.reset( new (std::nothrow) uint32_t[max_structures] );
+  if (!structural_indexes) { _capacity = 0; return MEMALLOC; }
+  structural_indexes[0] = 0;
+  n_structural_indexes = 0;
+
+  _capacity = capacity;
+  return SUCCESS;
+}
+
+inline simdjson_warn_unused error_code dom_parser_implementation::set_max_depth(size_t max_depth) noexcept {
+  // Stage 2 stacks
+  open_containers.reset(new (std::nothrow) open_container[max_depth]);
+  is_array.reset(new (std::nothrow) bool[max_depth]);
+  if (!is_array || !open_containers) { _max_depth = 0; return MEMALLOC; }
+
+  _max_depth = max_depth;
+  return SUCCESS;
+}
+
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file simdjson/generic/dom_parser_implementation.h for ppc64 */
+/* including simdjson/generic/implementation_simdjson_result_base.h for ppc64: #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base.h for ppc64 */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+
+// This is a near copy of include/error.h's implementation_simdjson_result_base, except it doesn't use std::pair
+// so we can avoid inlining errors
+// TODO reconcile these!
+/**
+ * The result of a simdjson operation that could fail.
+ *
+ * Gives the option of reading error codes, or throwing an exception by casting to the desired result.
+ *
+ * This is a base class for implementations that want to add functions to the result type for
+ * chaining.
+ *
+ * Override like:
+ *
+ *   struct simdjson_result<T> : public internal::implementation_simdjson_result_base<T> {
+ *     simdjson_result() noexcept : internal::implementation_simdjson_result_base<T>() {}
+ *     simdjson_result(error_code error) noexcept : internal::implementation_simdjson_result_base<T>(error) {}
+ *     simdjson_result(T &&value) noexcept : internal::implementation_simdjson_result_base<T>(std::forward(value)) {}
+ *     simdjson_result(T &&value, error_code error) noexcept : internal::implementation_simdjson_result_base<T>(value, error) {}
+ *     // Your extra methods here
+ *   }
+ *
+ * Then any method returning simdjson_result<T> will be chainable with your methods.
+ */
+template<typename T>
+struct implementation_simdjson_result_base {
+
+  /**
+   * Create a new empty result with error = UNINITIALIZED.
+   */
+  simdjson_inline implementation_simdjson_result_base() noexcept = default;
+
+  /**
+   * Create a new error result.
+   */
+  simdjson_inline implementation_simdjson_result_base(error_code error) noexcept;
+
+  /**
+   * Create a new successful result.
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value) noexcept;
+
+  /**
+   * Create a new result with both things (use if you don't want to branch when creating the result).
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value, error_code error) noexcept;
+
+  /**
+   * Move the value and the error to the provided variables.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   * @param error The variable to assign the error to. Set to SUCCESS if there is no error.
+   */
+  simdjson_inline void tie(T &value, error_code &error) && noexcept;
+
+  /**
+   * Move the value to the provided variable.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   */
+  simdjson_warn_unused simdjson_inline error_code get(T &value) && noexcept;
+
+  /**
+   * The error.
+   */
+  simdjson_warn_unused simdjson_inline error_code error() const noexcept;
+
+  /**
+   * Whether there is a value.
+   */
+  simdjson_warn_unused simdjson_inline bool has_value() const noexcept;
+
+#if SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T& operator*() &  noexcept(false);
+  simdjson_inline T&& operator*() &&  noexcept(false);
+  /**
+   * Arrow operator to access members of the contained value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T* operator->() noexcept(false);
+  simdjson_inline const T* operator->() const noexcept(false);
+
+  simdjson_inline T& value() & noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& value() && noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& take_value() && noexcept(false);
+
+  /**
+   * Cast to the value (will throw on error).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline operator T&&() && noexcept(false);
+
+
+#endif // SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline const T& value_unsafe() const& noexcept;
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T& value_unsafe() & noexcept;
+  /**
+   * Take the result value (move it). This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T&& value_unsafe() && noexcept;
+
+  using value_type = T;
+  using error_type = error_code;
+
+protected:
+  /** users should never directly access first and second. **/
+  T first{}; /** Users should never directly access 'first'. **/
+  error_code second{UNINITIALIZED}; /** Users should never directly access 'second'. **/
+}; // struct implementation_simdjson_result_base
+
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+/* end file simdjson/generic/implementation_simdjson_result_base.h for ppc64 */
+/* including simdjson/generic/numberparsing.h for ppc64: #include "simdjson/generic/numberparsing.h" */
+/* begin file simdjson/generic/numberparsing.h for ppc64 */
+#ifndef SIMDJSON_GENERIC_NUMBERPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_NUMBERPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <limits>
+#include <ostream>
+#include <cstring>
+
+namespace simdjson {
+namespace ppc64 {
+namespace numberparsing {
+
+#ifdef JSON_TEST_NUMBERS
+#define INVALID_NUMBER(SRC) (found_invalid_number((SRC)), NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (found_integer((VALUE), (SRC)), (WRITER).append_s64((VALUE)))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (found_unsigned_integer((VALUE), (SRC)), (WRITER).append_u64((VALUE)))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (found_float((VALUE), (SRC)), (WRITER).append_double((VALUE)))
+#define BIGINT_NUMBER(SRC) (found_invalid_number((SRC)), BIGINT_ERROR)
+#else
+#define INVALID_NUMBER(SRC) (NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (WRITER).append_s64((VALUE))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (WRITER).append_u64((VALUE))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (WRITER).append_double((VALUE))
+#define BIGINT_NUMBER(SRC) (BIGINT_ERROR)
+#endif
+
+namespace {
+
+// Convert a mantissa, an exponent and a sign bit into an ieee64 double.
+// The real_exponent needs to be in [0, 2046] (technically real_exponent = 2047 would be acceptable).
+// The mantissa should be in [0,1<<53). The bit at index (1ULL << 52) while be zeroed.
+simdjson_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
+    double d;
+    mantissa &= ~(1ULL << 52);
+    mantissa |= real_exponent << 52;
+    mantissa |= ((static_cast<uint64_t>(negative)) << 63);
+    std::memcpy(&d, &mantissa, sizeof(d));
+    return d;
+}
+
+// Attempts to compute i * 10^(power) exactly; and if "negative" is
+// true, negate the result.
+// This function will only work in some cases, when it does not work, success is
+// set to false. This should work *most of the time* (like 99% of the time).
+// We assume that power is in the [smallest_power,
+// largest_power] interval: the caller is responsible for this check.
+simdjson_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
+  // we start with a fast path
+  // It was described in
+  // Clinger WD. How to read floating point numbers accurately.
+  // ACM SIGPLAN Notices. 1990
+#ifndef FLT_EVAL_METHOD
+#error "FLT_EVAL_METHOD should be defined, please include cfloat."
+#endif
+#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
+  // We cannot be certain that x/y is rounded to nearest.
+  if (0 <= power && power <= 22 && i <= 9007199254740991)
+#else
+  if (-22 <= power && power <= 22 && i <= 9007199254740991)
+#endif
+  {
+    // convert the integer into a double. This is lossless since
+    // 0 <= i <= 2^53 - 1.
+    d = double(i);
+    //
+    // The general idea is as follows.
+    // If 0 <= s < 2^53 and if 10^0 <= p <= 10^22 then
+    // 1) Both s and p can be represented exactly as 64-bit floating-point
+    // values
+    // (binary64).
+    // 2) Because s and p can be represented exactly as floating-point values,
+    // then s * p
+    // and s / p will produce correctly rounded values.
+    //
+    if (power < 0) {
+      d = d / simdjson::internal::power_of_ten[-power];
+    } else {
+      d = d * simdjson::internal::power_of_ten[power];
+    }
+    if (negative) {
+      d = -d;
+    }
+    return true;
+  }
+  // When 22 < power && power <  22 + 16, we could
+  // hope for another, secondary fast path.  It was
+  // described by David M. Gay in  "Correctly rounded
+  // binary-decimal and decimal-binary conversions." (1990)
+  // If you need to compute i * 10^(22 + x) for x < 16,
+  // first compute i * 10^x, if you know that result is exact
+  // (e.g., when i * 10^x < 2^53),
+  // then you can still proceed and do (i * 10^x) * 10^22.
+  // Is this worth your time?
+  // You need  22 < power *and* power <  22 + 16 *and* (i * 10^(x-22) < 2^53)
+  // for this second fast path to work.
+  // If you you have 22 < power *and* power <  22 + 16, and then you
+  // optimistically compute "i * 10^(x-22)", there is still a chance that you
+  // have wasted your time if i * 10^(x-22) >= 2^53. It makes the use cases of
+  // this optimization maybe less common than we would like. Source:
+  // http://www.exploringbinary.com/fast-path-decimal-to-floating-point-conversion/
+  // also used in RapidJSON: https://rapidjson.org/strtod_8h_source.html
+
+  // The fast path has now failed, so we are failing back on the slower path.
+
+  // In the slow path, we need to adjust i so that it is > 1<<63 which is always
+  // possible, except if i == 0, so we handle i == 0 separately.
+  if(i == 0) {
+    d = negative ? -0.0 : 0.0;
+    return true;
+  }
+
+
+  // The exponent is 1024 + 63 + power
+  //     + floor(log(5**power)/log(2)).
+  // The 1024 comes from the ieee64 standard.
+  // The 63 comes from the fact that we use a 64-bit word.
+  //
+  // Computing floor(log(5**power)/log(2)) could be
+  // slow. Instead we use a fast function.
+  //
+  // For power in (-400,350), we have that
+  // (((152170 + 65536) * power ) >> 16);
+  // is equal to
+  //  floor(log(5**power)/log(2)) + power when power >= 0
+  // and it is equal to
+  //  ceil(log(5**-power)/log(2)) + power when power < 0
+  //
+  // The 65536 is (1<<16) and corresponds to
+  // (65536 * power) >> 16 ---> power
+  //
+  // ((152170 * power ) >> 16) is equal to
+  // floor(log(5**power)/log(2))
+  //
+  // Note that this is not magic: 152170/(1<<16) is
+  // approximately equal to log(5)/log(2).
+  // The 1<<16 value is a power of two; we could use a
+  // larger power of 2 if we wanted to.
+  //
+  int64_t exponent = (((152170 + 65536) * power) >> 16) + 1024 + 63;
+
+
+  // We want the most significant bit of i to be 1. Shift if needed.
+  int lz = leading_zeroes(i);
+  i <<= lz;
+
+
+  // We are going to need to do some 64-bit arithmetic to get a precise product.
+  // We use a table lookup approach.
+  // It is safe because
+  // power >= smallest_power
+  // and power <= largest_power
+  // We recover the mantissa of the power, it has a leading 1. It is always
+  // rounded down.
+  //
+  // We want the most significant 64 bits of the product. We know
+  // this will be non-zero because the most significant bit of i is
+  // 1.
+  const uint32_t index = 2 * uint32_t(power - simdjson::internal::smallest_power);
+  // Optimization: It may be that materializing the index as a variable might confuse some compilers and prevent effective complex-addressing loads. (Done for code clarity.)
+  //
+  // The full_multiplication function computes the 128-bit product of two 64-bit words
+  // with a returned value of type value128 with a "low component" corresponding to the
+  // 64-bit least significant bits of the product and with a "high component" corresponding
+  // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index]);
+#else
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index]);
+#endif
+
+  // Both i and power_of_five_128[index] have their most significant bit set to 1 which
+  // implies that the either the most or the second most significant bit of the product
+  // is 1. We pack values in this manner for efficiency reasons: it maximizes the use
+  // we make of the product. It also makes it easy to reason about the product: there
+  // is 0 or 1 leading zero in the product.
+
+  // Unless the least significant 9 bits of the high (64-bit) part of the full
+  // product are all 1s, then we know that the most significant 55 bits are
+  // exact and no further work is needed. Having 55 bits is necessary because
+  // we need 53 bits for the mantissa but we have to have one rounding bit and
+  // we can waste a bit if the most significant bit of the product is zero.
+  if((firstproduct.high & 0x1FF) == 0x1FF) {
+    // We want to compute i * 5^q, but only care about the top 55 bits at most.
+    // Consider the scenario where q>=0. Then 5^q may not fit in 64-bits. Doing
+    // the full computation is wasteful. So we do what is called a "truncated
+    // multiplication".
+    // We take the most significant 64-bits, and we put them in
+    // power_of_five_128[index]. Usually, that's good enough to approximate i * 5^q
+    // to the desired approximation using one multiplication. Sometimes it does not suffice.
+    // Then we store the next most significant 64 bits in power_of_five_128[index + 1], and
+    // then we get a better approximation to i * 5^q.
+    //
+    // That's for when q>=0. The logic for q<0 is somewhat similar but it is somewhat
+    // more complicated.
+    //
+    // There is an extra layer of complexity in that we need more than 55 bits of
+    // accuracy in the round-to-even scenario.
+    //
+    // The full_multiplication function computes the 128-bit product of two 64-bit words
+    // with a returned value of type value128 with a "low component" corresponding to the
+    // 64-bit least significant bits of the product and with a "high component" corresponding
+    // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index + 1]);
+#else
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index + 1]);
+#endif
+    firstproduct.low += secondproduct.high;
+    if(secondproduct.high > firstproduct.low) { firstproduct.high++; }
+    // As it has been proven by Noble Mushtak and Daniel Lemire in "Fast Number Parsing Without
+    // Fallback" (https://arxiv.org/abs/2212.06644), at this point we are sure that the product
+    // is sufficiently accurate, and more computation is not needed.
+  }
+  uint64_t lower = firstproduct.low;
+  uint64_t upper = firstproduct.high;
+  // The final mantissa should be 53 bits with a leading 1.
+  // We shift it so that it occupies 54 bits with a leading 1.
+  ///////
+  uint64_t upperbit = upper >> 63;
+  uint64_t mantissa = upper >> (upperbit + 9);
+  lz += int(1 ^ upperbit);
+
+  // Here we have mantissa < (1<<54).
+  int64_t real_exponent = exponent - lz;
+  if (simdjson_unlikely(real_exponent <= 0)) { // we have a subnormal?
+    // Here have that real_exponent <= 0 so -real_exponent >= 0
+    if(-real_exponent + 1 >= 64) { // if we have more than 64 bits below the minimum exponent, you have a zero for sure.
+      d = negative ? -0.0 : 0.0;
+      return true;
+    }
+    // next line is safe because -real_exponent + 1 < 0
+    mantissa >>= -real_exponent + 1;
+    // Thankfully, we can't have both "round-to-even" and subnormals because
+    // "round-to-even" only occurs for powers close to 0.
+    mantissa += (mantissa & 1); // round up
+    mantissa >>= 1;
+    // There is a weird scenario where we don't have a subnormal but just.
+    // Suppose we start with 2.2250738585072013e-308, we end up
+    // with 0x3fffffffffffff x 2^-1023-53 which is technically subnormal
+    // whereas 0x40000000000000 x 2^-1023-53  is normal. Now, we need to round
+    // up 0x3fffffffffffff x 2^-1023-53  and once we do, we are no longer
+    // subnormal, but we can only know this after rounding.
+    // So we only declare a subnormal if we are smaller than the threshold.
+    real_exponent = (mantissa < (uint64_t(1) << 52)) ? 0 : 1;
+    d = to_double(mantissa, real_exponent, negative);
+    return true;
+  }
+  // We have to round to even. The "to even" part
+  // is only a problem when we are right in between two floats
+  // which we guard against.
+  // If we have lots of trailing zeros, we may fall right between two
+  // floating-point values.
+  //
+  // The round-to-even cases take the form of a number 2m+1 which is in (2^53,2^54]
+  // times a power of two. That is, it is right between a number with binary significand
+  // m and another number with binary significand m+1; and it must be the case
+  // that it cannot be represented by a float itself.
+  //
+  // We must have that w * 10 ^q == (2m+1) * 2^p for some power of two 2^p.
+  // Recall that 10^q = 5^q * 2^q.
+  // When q >= 0, we must have that (2m+1) is divible by 5^q, so 5^q <= 2^54. We have that
+  //  5^23 <=  2^54 and it is the last power of five to qualify, so q <= 23.
+  // When q<0, we have  w  >=  (2m+1) x 5^{-q}.  We must have that w<2^{64} so
+  // (2m+1) x 5^{-q} < 2^{64}. We have that 2m+1>2^{53}. Hence, we must have
+  // 2^{53} x 5^{-q} < 2^{64}.
+  // Hence we have 5^{-q} < 2^{11}$ or q>= -4.
+  //
+  // We require lower <= 1 and not lower == 0 because we could not prove that
+  // that lower == 0 is implied; but we could prove that lower <= 1 is a necessary and sufficient test.
+  if (simdjson_unlikely((lower <= 1) && (power >= -4) && (power <= 23) && ((mantissa & 3) == 1))) {
+    if((mantissa  << (upperbit + 64 - 53 - 2)) ==  upper) {
+      mantissa &= ~1;             // flip it so that we do not round up
+    }
+  }
+
+  mantissa += mantissa & 1;
+  mantissa >>= 1;
+
+  // Here we have mantissa < (1<<53), unless there was an overflow
+  if (mantissa >= (1ULL << 53)) {
+    //////////
+    // This will happen when parsing values such as 7.2057594037927933e+16
+    ////////
+    mantissa = (1ULL << 52);
+    real_exponent++;
+  }
+  mantissa &= ~(1ULL << 52);
+  // we have to check that real_exponent is in range, otherwise we bail out
+  if (simdjson_unlikely(real_exponent > 2046)) {
+    // We have an infinite value!!! We could actually throw an error here if we could.
+    return false;
+  }
+  d = to_double(mantissa, real_exponent, negative);
+  return true;
+}
+
+// We call a fallback floating-point parser that might be slow. Note
+// it will accept JSON numbers, but the JSON spec. is more restrictive so
+// before you call parse_float_fallback, you need to have validated the input
+// string with the JSON grammar.
+// It will return an error (false) if the parsed number is infinite.
+// The string parsing itself always succeeds. We know that there is at least
+// one digit.
+static bool parse_float_fallback(const uint8_t *ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+static bool parse_float_fallback(const uint8_t *ptr, const uint8_t *end_ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr), reinterpret_cast<const char *>(end_ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+// check quickly whether the next 8 chars are made of digits
+// at a glance, it looks better than Mula's
+// http://0x80.pl/articles/swar-digits-validate.html
+simdjson_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
+  uint64_t val;
+  // this can read up to 7 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(7 <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be bigger than 7");
+  std::memcpy(&val, chars, 8);
+  // a branchy method might be faster:
+  // return (( val & 0xF0F0F0F0F0F0F0F0 ) == 0x3030303030303030)
+  //  && (( (val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0 ) ==
+  //  0x3030303030303030);
+  return (((val & 0xF0F0F0F0F0F0F0F0) |
+           (((val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0) >> 4)) ==
+          0x3333333333333333);
+}
+
+template<typename I>
+SIMDJSON_NO_SANITIZE_UNDEFINED // We deliberately allow overflow here and check later
+simdjson_inline bool parse_digit(const uint8_t c, I &i) {
+  const uint8_t digit = static_cast<uint8_t>(c - '0');
+  if (digit > 9) {
+    return false;
+  }
+  // PERF NOTE: multiplication by 10 is cheaper than arbitrary integer multiplication
+  i = 10 * i + digit; // might overflow, we will handle the overflow later
+  return true;
+}
+
+simdjson_inline bool is_digit(const uint8_t c) {
+  return static_cast<uint8_t>(c - '0') <= 9;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_decimal_after_separator(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
+  // we continue with the fiction that we have an integer. If the
+  // floating point number is representable as x * 10^z for some integer
+  // z that fits in 53 bits, then we will be able to convert back the
+  // the integer into a float in a lossless manner.
+  const uint8_t *const first_after_period = p;
+
+#ifdef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_SWAR_NUMBER_PARSING
+  // this helps if we have lots of decimals!
+  // this turns out to be frequent enough.
+  if (is_made_of_eight_digits_fast(p)) {
+    i = i * 100000000 + parse_eight_digits_unrolled(p);
+    p += 8;
+  }
+#endif // SIMDJSON_SWAR_NUMBER_PARSING
+#endif // #ifdef SIMDJSON_SWAR_NUMBER_PARSING
+  // Unrolling the first digit makes a small difference on some implementations (e.g. westmere)
+  if (parse_digit(*p, i)) { ++p; }
+  while (parse_digit(*p, i)) { p++; }
+  exponent = first_after_period - p;
+  // Decimal without digits (123.) is illegal
+  if (exponent == 0) {
+    return INVALID_NUMBER(src);
+  }
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
+  // Exp Sign: -123.456e[-]78
+  bool neg_exp = ('-' == *p);
+  if (neg_exp || '+' == *p) { p++; } // Skip + as well
+
+  // Exponent: -123.456e-[78]
+  auto start_exp = p;
+  int64_t exp_number = 0;
+  while (parse_digit(*p, exp_number)) { ++p; }
+  // It is possible for parse_digit to overflow.
+  // In particular, it could overflow to INT64_MIN, and we cannot do - INT64_MIN.
+  // Thus we *must* check for possible overflow before we negate exp_number.
+
+  // Performance notes: it may seem like combining the two "simdjson_unlikely checks" below into
+  // a single simdjson_unlikely path would be faster. The reasoning is sound, but the compiler may
+  // not oblige and may, in fact, generate two distinct paths in any case. It might be
+  // possible to do uint64_t(p - start_exp - 1) >= 18 but it could end up trading off
+  // instructions for a simdjson_likely branch, an unconclusive gain.
+
+  // If there were no digits, it's an error.
+  if (simdjson_unlikely(p == start_exp)) {
+    return INVALID_NUMBER(src);
+  }
+  // We have a valid positive exponent in exp_number at this point, except that
+  // it may have overflowed.
+
+  // If there were more than 18 digits, we may have overflowed the integer. We have to do
+  // something!!!!
+  if (simdjson_unlikely(p > start_exp+18)) {
+    // Skip leading zeroes: 1e000000000000000000001 is technically valid and does not overflow
+    while (*start_exp == '0') { start_exp++; }
+    // 19 digits could overflow int64_t and is kind of absurd anyway. We don't
+    // support exponents smaller than -999,999,999,999,999,999 and bigger
+    // than 999,999,999,999,999,999.
+    // We can truncate.
+    // Note that 999999999999999999 is assuredly too large. The maximal ieee64 value before
+    // infinity is ~1.8e308. The smallest subnormal is ~5e-324. So, actually, we could
+    // truncate at 324.
+    // Note that there is no reason to fail per se at this point in time.
+    // E.g., 0e999999999999999999999 is a fine number.
+    if (p > start_exp+18) { exp_number = 999999999999999999; }
+  }
+  // At this point, we know that exp_number is a sane, positive, signed integer.
+  // It is <= 999,999,999,999,999,999. As long as 'exponent' is in
+  // [-8223372036854775808, 8223372036854775808], we won't overflow. Because 'exponent'
+  // is bounded in magnitude by the size of the JSON input, we are fine in this universe.
+  // To sum it up: the next line should never overflow.
+  exponent += (neg_exp ? -exp_number : exp_number);
+  return SUCCESS;
+}
+
+simdjson_inline bool check_if_integer(const uint8_t *const src, size_t max_length) {
+  const uint8_t *const srcend = src + max_length;
+  bool negative = (*src == '-'); // we can always read at least one character after the '-'
+  const uint8_t *p = src + uint8_t(negative);
+  if(p == srcend) { return false; }
+  if(*p == '0') {
+    ++p;
+    if(p == srcend) { return true; }
+    if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+    return true;
+  }
+  while(p != srcend && is_digit(*p)) { ++p; }
+  if(p == srcend) { return true; }
+  if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+  return true;
+}
+
+simdjson_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
+  // It is possible that the integer had an overflow.
+  // We have to handle the case where we have 0.0000somenumber.
+  const uint8_t *start = start_digits;
+  while ((*start == '0') || (*start == '.')) { ++start; }
+  // we over-decrement by one when there is a '.'
+  return digit_count - size_t(start - start_digits);
+}
+
+} // unnamed namespace
+
+/** @private */
+static error_code slow_float_parsing(simdjson_unused const uint8_t * src, double* answer) {
+  if (parse_float_fallback(src, answer)) {
+    return SUCCESS;
+  }
+  return INVALID_NUMBER(src);
+}
+
+/** @private */
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
+  // If we frequently had to deal with long strings of digits,
+  // we could extend our code by using a 128-bit integer instead
+  // of a 64-bit integer. However, this is uncommon in practice.
+  //
+  // 9999999999999999999 < 2**64 so we can accommodate 19 digits.
+  // If we have a decimal separator, then digit_count - 1 is the number of digits, but we
+  // may not have a decimal separator!
+  if (simdjson_unlikely(digit_count > 19 && significant_digits(start_digits, digit_count) > 19)) {
+    // Ok, chances are good that we had an overflow!
+    // this is almost never going to get called!!!
+    // we start anew, going slowly!!!
+    // This will happen in the following examples:
+    // 10000000000000000000000000000000000000000000e+308
+    // 3.1415926535897932384626433832795028841971693993751
+    //
+    // NOTE: We do not pass a reference to the to slow_float_parsing. If we passed our writer
+    // reference to it, it would force it to be stored in memory, preventing the compiler from
+    // picking it apart and putting into registers. i.e. if we pass it as reference,
+    // it gets slow.
+    double d;
+    error_code error = slow_float_parsing(src, &d);
+    writer.append_double(d);
+    return error;
+  }
+  // NOTE: it's weird that the simdjson_unlikely() only wraps half the if, but it seems to get slower any other
+  // way we've tried: https://github.com/simdjson/simdjson/pull/990#discussion_r448497331
+  // To future reader: we'd love if someone found a better way, or at least could explain this result!
+  if (simdjson_unlikely(exponent < simdjson::internal::smallest_power) || (exponent > simdjson::internal::largest_power)) {
+    //
+    // Important: smallest_power is such that it leads to a zero value.
+    // Observe that 18446744073709551615e-343 == 0, i.e. (2**64 - 1) e -343 is zero
+    // so something x 10^-343 goes to zero, but not so with  something x 10^-342.
+    static_assert(simdjson::internal::smallest_power <= -342, "smallest_power is not small enough");
+    //
+    if((exponent < simdjson::internal::smallest_power) || (i == 0)) {
+      // E.g. Parse "-0.0e-999" into the same value as "-0.0". See https://en.wikipedia.org/wiki/Signed_zero
+      WRITE_DOUBLE(negative ? -0.0 : 0.0, src, writer);
+      return SUCCESS;
+    } else { // (exponent > largest_power) and (i != 0)
+      // We have, for sure, an infinite value and simdjson refuses to parse infinite values.
+      return INVALID_NUMBER(src);
+    }
+  }
+  double d;
+  if (!compute_float_64(exponent, i, negative, d)) {
+    // we are almost never going to get here.
+    if (!parse_float_fallback(src, &d)) { return INVALID_NUMBER(src); }
+  }
+  WRITE_DOUBLE(d, src, writer);
+  return SUCCESS;
+}
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer);
+
+// for performance analysis, it is sometimes  useful to skip parsing
+#ifdef SIMDJSON_SKIPNUMBERPARSING
+
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const, W &writer) {
+  writer.append_s64(0);        // always write zero
+  return SUCCESS;              // always succeeds
+}
+
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept { return number_type::signed_integer; }
+#else
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer) {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  if (digit_count == 0 || ('0' == *start_digits && digit_count > 1)) { return INVALID_NUMBER(src); }
+
+  //
+  // Handle floats if there is a . or e (or both)
+  //
+  int64_t exponent = 0;
+  bool is_float = false;
+  if ('.' == *p) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_decimal_after_separator(src, p, i, exponent) );
+    digit_count = int(p - start_digits); // used later to guard against overflows
+  }
+  if (('e' == *p) || ('E' == *p)) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_exponent(src, p, exponent) );
+  }
+  if (is_float) {
+    const bool dirty_end = jsoncharutils::is_not_structural_or_whitespace(*p);
+    SIMDJSON_TRY( write_float(src, negative, i, start_digits, digit_count, exponent, writer) );
+    if (dirty_end) { return INVALID_NUMBER(src); }
+    return SUCCESS;
+  }
+
+  // The longest negative 64-bit number is 19 digits.
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  size_t longest_digit_count = negative ? 19 : 20;
+  if (digit_count > longest_digit_count) { return BIGINT_NUMBER(src); }
+  if (digit_count == longest_digit_count) {
+    if (negative) {
+      // Anything negative above INT64_MAX+1 is invalid
+      if (i > uint64_t(INT64_MAX)+1) { return BIGINT_NUMBER(src);  }
+      WRITE_INTEGER(~i+1, src, writer);
+      if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+      return SUCCESS;
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    }  else if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INVALID_NUMBER(src); }
+  }
+
+  // Write unsigned if it does not fit in a signed integer.
+  if (i > uint64_t(INT64_MAX)) {
+    WRITE_UNSIGNED(i, src, writer);
+  } else {
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+    if(i == 0 && negative) {
+      // We have to write -0.0 instead of 0
+      WRITE_DOUBLE(-0.0, src, writer);
+    } else {
+      WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+    }
+#else
+  WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+#endif
+  }
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+  return SUCCESS;
+}
+
+// Inlineable functions
+namespace {
+
+// This table can be used to characterize the final character of an integer
+// string. For JSON structural character and allowable white space characters,
+// we return SUCCESS. For 'e', '.' and 'E', we return INCORRECT_TYPE. Otherwise
+// we return NUMBER_ERROR.
+// Optimization note: we could easily reduce the size of the table by half (to 128)
+// at the cost of an extra branch.
+// Optimization note: we want the values to use at most 8 bits (not, e.g., 32 bits):
+static_assert(error_code(uint8_t(NUMBER_ERROR))== NUMBER_ERROR, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(SUCCESS))== SUCCESS, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(INCORRECT_TYPE))== INCORRECT_TYPE, "bad NUMBER_ERROR cast");
+
+const uint8_t integer_string_finisher[256] = {
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   SUCCESS,      NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, INCORRECT_TYPE,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, SUCCESS,        NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR};
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if ((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
+  const uint8_t *p = src + 1;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    // Note: we use src[1] and not src[0] because src[0] is the quote character in this
+    // instance.
+    if (src[1] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  //
+  // Check for minus sign
+  //
+  if(src == src_end) { return NUMBER_ERROR; }
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = src;
+  uint64_t i = 0;
+  while (parse_digit(*src, i)) { src++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(src - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*src)) {
+  //  return (*src == '.' || *src == 'e' || *src == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(*src != '"') { return NUMBER_ERROR; }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept {
+  return (*src == '-');
+}
+
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) { return true; }
+  return false;
+}
+
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  size_t digit_count = size_t(p - src);
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) {
+    static const uint8_t * smaller_big_integer = reinterpret_cast<const uint8_t *>("9223372036854775808");
+    // We have an integer.
+    if(simdjson_unlikely(digit_count > 20)) {
+      return number_type::big_integer;
+    }
+    // If the number is negative and valid, it must be a signed integer.
+    if(negative) {
+      if (simdjson_unlikely(digit_count > 19)) return number_type::big_integer;
+      if (simdjson_unlikely(digit_count == 19 && memcmp(src, smaller_big_integer, 19) > 0)) {
+        return number_type::big_integer;
+      }
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+      if(digit_count == 1 && src[0] == '0') {
+        // We have to write -0.0 instead of 0
+        return number_type::floating_point_number;
+      }
+#endif
+      return number_type::signed_integer;
+    }
+    // Let us check if we have a big integer (>=2**64).
+    static const uint8_t * two_to_sixtyfour = reinterpret_cast<const uint8_t *>("18446744073709551616");
+    if((digit_count > 20) || (digit_count == 20 && memcmp(src, two_to_sixtyfour, 20) >= 0)) {
+      return number_type::big_integer;
+    }
+    // The number is positive and smaller than 18446744073709551616 (or 2**64).
+    // We want values larger or equal to 9223372036854775808 to be unsigned
+    // integers, and the other values to be signed integers.
+    if((digit_count == 20) || (digit_count >= 19 && memcmp(src, smaller_big_integer, 19) >= 0)) {
+      return number_type::unsigned_integer;
+    }
+    return number_type::signed_integer;
+  }
+  // Hopefully, we have 'e' or 'E' or '.'.
+  return number_type::floating_point_number;
+}
+
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
+  if(src == src_end) { return NUMBER_ERROR; }
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  if(p == src_end) { return NUMBER_ERROR; }
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely((p != src_end) && (*p == '.'))) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if ((p == src_end) || !parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while ((p != src_end) && parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = start_digits-src > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if ((p != src_end) && (*p == 'e' || *p == 'E')) {
+    p++;
+    if(p == src_end) { return NUMBER_ERROR; }
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while ((p != src_end) && parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if ((p != src_end) && jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), src_end, &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+} // unnamed namespace
+#endif // SIMDJSON_SKIPNUMBERPARSING
+
+} // namespace numberparsing
+
+inline std::ostream& operator<<(std::ostream& out, number_type type) noexcept {
+    switch (type) {
+        case number_type::signed_integer: out << "integer in [-9223372036854775808,9223372036854775808)"; break;
+        case number_type::unsigned_integer: out << "unsigned integer in [9223372036854775808,18446744073709551616)"; break;
+        case number_type::floating_point_number: out << "floating-point number (binary64)"; break;
+        case number_type::big_integer: out << "big integer"; break;
+        default: SIMDJSON_UNREACHABLE();
+    }
+    return out;
+}
+
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_NUMBERPARSING_H
+/* end file simdjson/generic/numberparsing.h for ppc64 */
+
+/* including simdjson/generic/implementation_simdjson_result_base-inl.h for ppc64: #include "simdjson/generic/implementation_simdjson_result_base-inl.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base-inl.h for ppc64 */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+
+//
+// internal::implementation_simdjson_result_base<T> inline implementation
+//
+
+template<typename T>
+simdjson_inline void implementation_simdjson_result_base<T>::tie(T &value, error_code &error) && noexcept {
+  error = this->second;
+  if (!error) {
+    value = std::forward<implementation_simdjson_result_base<T>>(*this).first;
+  }
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::get(T &value) && noexcept {
+  error_code error;
+  std::forward<implementation_simdjson_result_base<T>>(*this).tie(value, error);
+  return error;
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::error() const noexcept {
+  return this->second;
+}
+
+
+template<typename T>
+simdjson_warn_unused simdjson_inline bool implementation_simdjson_result_base<T>::has_value() const noexcept {
+  return this->error() == SUCCESS;
+}
+
+#if SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::operator*() &  noexcept(false) {
+  return this->value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::operator*() &&  noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).value();
+}
+
+template<typename T>
+simdjson_inline T* implementation_simdjson_result_base<T>::operator->() noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+
+template<typename T>
+simdjson_inline const T* implementation_simdjson_result_base<T>::operator->() const noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value() & noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::take_value() && noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::operator T&&() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+#endif // SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline const T& implementation_simdjson_result_base<T>::value_unsafe() const& noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value_unsafe() & noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value_unsafe() && noexcept {
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value, error_code error) noexcept
+    : first{std::forward<T>(value)}, second{error} {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(error_code error) noexcept
+    : implementation_simdjson_result_base(T{}, error) {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value) noexcept
+    : implementation_simdjson_result_base(std::forward<T>(value), SUCCESS) {}
+
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+/* end file simdjson/generic/implementation_simdjson_result_base-inl.h for ppc64 */
+/* end file simdjson/generic/amalgamated.h for ppc64 */
+/* including simdjson/ppc64/end.h: #include "simdjson/ppc64/end.h" */
+/* begin file simdjson/ppc64/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#undef SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+/* undefining SIMDJSON_IMPLEMENTATION from "ppc64" */
+#undef SIMDJSON_IMPLEMENTATION
+/* end file simdjson/ppc64/end.h */
+
+#endif // SIMDJSON_PPC64_H
+/* end file simdjson/ppc64.h */
+/* including simdjson/ppc64/implementation.h: #include <simdjson/ppc64/implementation.h> */
+/* begin file simdjson/ppc64/implementation.h */
+#ifndef SIMDJSON_PPC64_IMPLEMENTATION_H
+#define SIMDJSON_PPC64_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/instruction_set.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+
+/**
+ * Implementation for ALTIVEC (PPC64).
+ */
+namespace ppc64 {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation()
+      : simdjson::implementation("ppc64", "PPC64 ALTIVEC",
+                                 internal::instruction_set::ALTIVEC) {}
+
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+      size_t capacity, size_t max_length,
+      std::unique_ptr<internal::dom_parser_implementation> &dst)
+      const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len,
+                                         uint8_t *dst,
+                                         size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf,
+                                          size_t len) const noexcept final;
+};
+
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_PPC64_IMPLEMENTATION_H
+/* end file simdjson/ppc64/implementation.h */
+
+/* including simdjson/ppc64/begin.h: #include <simdjson/ppc64/begin.h> */
+/* begin file simdjson/ppc64/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "ppc64" */
+#define SIMDJSON_IMPLEMENTATION ppc64
+/* including simdjson/ppc64/base.h: #include "simdjson/ppc64/base.h" */
+/* begin file simdjson/ppc64/base.h */
+#ifndef SIMDJSON_PPC64_BASE_H
+#define SIMDJSON_PPC64_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+/**
+ * Implementation for ALTIVEC (PPC64).
+ */
+namespace ppc64 {
+
+class implementation;
+
+namespace {
+namespace simd {
+template <typename T> struct simd8;
+template <typename T> struct simd8x64;
+} // namespace simd
+} // unnamed namespace
+
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_PPC64_BASE_H
+/* end file simdjson/ppc64/base.h */
+/* including simdjson/ppc64/intrinsics.h: #include "simdjson/ppc64/intrinsics.h" */
+/* begin file simdjson/ppc64/intrinsics.h */
+#ifndef SIMDJSON_PPC64_INTRINSICS_H
+#define SIMDJSON_PPC64_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This should be the correct header whether
+// you use visual studio or other compilers.
+#include <altivec.h>
+
+// These are defined by altivec.h in GCC toolchain, it is safe to undef them.
+#ifdef bool
+#undef bool
+#endif
+
+#ifdef vector
+#undef vector
+#endif
+
+static_assert(sizeof(__vector unsigned char) <= simdjson::SIMDJSON_PADDING, "insufficient padding for ppc64");
+
+#endif //  SIMDJSON_PPC64_INTRINSICS_H
+/* end file simdjson/ppc64/intrinsics.h */
+/* including simdjson/ppc64/bitmanipulation.h: #include "simdjson/ppc64/bitmanipulation.h" */
+/* begin file simdjson/ppc64/bitmanipulation.h */
+#ifndef SIMDJSON_PPC64_BITMANIPULATION_H
+#define SIMDJSON_PPC64_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  unsigned long ret;
+  // Search the mask data from least significant bit (LSB)
+  // to the most significant bit (MSB) for a set bit (1).
+  _BitScanForward64(&ret, input_num);
+  return (int)ret;
+#else  // SIMDJSON_REGULAR_VISUAL_STUDIO
+  return __builtin_ctzll(input_num);
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return input_num & (input_num - 1);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  unsigned long leading_zero = 0;
+  // Search the mask data from most significant bit (MSB)
+  // to least significant bit (LSB) for a set bit (1).
+  if (_BitScanReverse64(&leading_zero, input_num))
+    return (int)(63 - leading_zero);
+  else
+    return 64;
+#else
+  return __builtin_clzll(input_num);
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+simdjson_inline int count_ones(uint64_t input_num) {
+  // note: we do not support legacy 32-bit Windows in this kernel
+  return __popcnt64(input_num); // Visual Studio wants two underscores
+}
+#else
+simdjson_inline int count_ones(uint64_t input_num) {
+  return __builtin_popcountll(input_num);
+}
+#endif
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2,
+                                         uint64_t *result) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  *result = value1 + value2;
+  return *result < value1;
+#else
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+#endif
+}
+
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_PPC64_BITMANIPULATION_H
+/* end file simdjson/ppc64/bitmanipulation.h */
+/* including simdjson/ppc64/bitmask.h: #include "simdjson/ppc64/bitmask.h" */
+/* begin file simdjson/ppc64/bitmask.h */
+#ifndef SIMDJSON_PPC64_BITMASK_H
+#define SIMDJSON_PPC64_BITMASK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is
+// encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(uint64_t bitmask) {
+  // You can use the version below, however gcc sometimes miscompiles
+  // vec_pmsum_be, it happens somewhere around between 8 and 9th version.
+  // The performance boost was not noticeable, falling back to a usual
+  // implementation.
+  //   __vector unsigned long long all_ones = {~0ull, ~0ull};
+  //   __vector unsigned long long mask = {bitmask, 0};
+  //   // Clang and GCC return different values for pmsum for ull so cast it to one.
+  //   // Generally it is not specified by ALTIVEC ISA what is returned by
+  //   // vec_pmsum_be.
+  // #if defined(__LITTLE_ENDIAN__)
+  //   return (uint64_t)(((__vector unsigned long long)vec_pmsum_be(all_ones, mask))[0]);
+  // #else
+  //   return (uint64_t)(((__vector unsigned long long)vec_pmsum_be(all_ones, mask))[1]);
+  // #endif
+  bitmask ^= bitmask << 1;
+  bitmask ^= bitmask << 2;
+  bitmask ^= bitmask << 4;
+  bitmask ^= bitmask << 8;
+  bitmask ^= bitmask << 16;
+  bitmask ^= bitmask << 32;
+  return bitmask;
+}
+
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif
+/* end file simdjson/ppc64/bitmask.h */
+/* including simdjson/ppc64/numberparsing_defs.h: #include "simdjson/ppc64/numberparsing_defs.h" */
+/* begin file simdjson/ppc64/numberparsing_defs.h */
+#ifndef SIMDJSON_PPC64_NUMBERPARSING_DEFS_H
+#define SIMDJSON_PPC64_NUMBERPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+#if defined(__linux__)
+#include <byteswap.h>
+#elif defined(__FreeBSD__)
+#include <sys/endian.h>
+#endif
+
+namespace simdjson {
+namespace ppc64 {
+namespace numberparsing {
+
+// we don't have appropriate instructions, so let us use a scalar function
+// credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  uint64_t val;
+  std::memcpy(&val, chars, sizeof(uint64_t));
+#ifdef __BIG_ENDIAN__
+#if defined(__linux__)
+  val = bswap_64(val);
+#elif defined(__FreeBSD__)
+  val = bswap64(val);
+#endif
+#endif
+  val = (val & 0x0F0F0F0F0F0F0F0F) * 2561 >> 8;
+  val = (val & 0x00FF00FF00FF00FF) * 6553601 >> 16;
+  return uint32_t((val & 0x0000FFFF0000FFFF) * 42949672960001 >> 32);
+}
+
+/** @private */
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+#if SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+#if SIMDJSON_IS_ARM64
+  // ARM64 has native support for 64-bit multiplications, no need to emultate
+  answer.high = __umulh(value1, value2);
+  answer.low = value1 * value2;
+#else
+  answer.low = _umul128(value1, value2, &answer.high); // _umul128 not available on ARM64
+#endif // SIMDJSON_IS_ARM64
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+#endif
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace ppc64
+} // namespace simdjson
+
+#ifndef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_IS_BIG_ENDIAN
+#define SIMDJSON_SWAR_NUMBER_PARSING 0
+#else
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+#endif
+#endif
+
+#endif // SIMDJSON_PPC64_NUMBERPARSING_DEFS_H
+/* end file simdjson/ppc64/numberparsing_defs.h */
+/* including simdjson/ppc64/simd.h: #include "simdjson/ppc64/simd.h" */
+/* begin file simdjson/ppc64/simd.h */
+#ifndef SIMDJSON_PPC64_SIMD_H
+#define SIMDJSON_PPC64_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <type_traits>
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace simd {
+
+using __m128i = __vector unsigned char;
+
+template <typename Child> struct base {
+  __m128i value;
+
+  // Zero constructor
+  simdjson_inline base() : value{__m128i()} {}
+
+  // Conversion from SIMD register
+  simdjson_inline base(const __m128i _value) : value(_value) {}
+
+  // Conversion to SIMD register
+  simdjson_inline operator const __m128i &() const {
+    return this->value;
+  }
+  simdjson_inline operator __m128i &() { return this->value; }
+
+  // Bit operations
+  simdjson_inline Child operator|(const Child other) const {
+    return vec_or(this->value, (__m128i)other);
+  }
+  simdjson_inline Child operator&(const Child other) const {
+    return vec_and(this->value, (__m128i)other);
+  }
+  simdjson_inline Child operator^(const Child other) const {
+    return vec_xor(this->value, (__m128i)other);
+  }
+  simdjson_inline Child bit_andnot(const Child other) const {
+    return vec_andc(this->value, (__m128i)other);
+  }
+  simdjson_inline Child &operator|=(const Child other) {
+    auto this_cast = static_cast<Child*>(this);
+    *this_cast = *this_cast | other;
+    return *this_cast;
+  }
+  simdjson_inline Child &operator&=(const Child other) {
+    auto this_cast = static_cast<Child*>(this);
+    *this_cast = *this_cast & other;
+    return *this_cast;
+  }
+  simdjson_inline Child &operator^=(const Child other) {
+    auto this_cast = static_cast<Child*>(this);
+    *this_cast = *this_cast ^ other;
+    return *this_cast;
+  }
+};
+
+template <typename T, typename Mask = simd8<bool>>
+struct base8 : base<simd8<T>> {
+  typedef uint16_t bitmask_t;
+  typedef uint32_t bitmask2_t;
+
+  simdjson_inline base8() : base<simd8<T>>() {}
+  simdjson_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}
+
+  friend simdjson_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) {
+    return (__m128i)vec_cmpeq(lhs.value, (__m128i)rhs);
+  }
+
+  static const int SIZE = sizeof(base<simd8<T>>::value);
+
+  template <int N = 1>
+  simdjson_inline simd8<T> prev(simd8<T> prev_chunk) const {
+    __m128i chunk = this->value;
+#ifdef __LITTLE_ENDIAN__
+    chunk = (__m128i)vec_reve(this->value);
+    prev_chunk = (__m128i)vec_reve((__m128i)prev_chunk);
+#endif
+    chunk = (__m128i)vec_sld((__m128i)prev_chunk, (__m128i)chunk, 16 - N);
+#ifdef __LITTLE_ENDIAN__
+    chunk = (__m128i)vec_reve((__m128i)chunk);
+#endif
+    return chunk;
+  }
+};
+
+// SIMD byte mask type (returned by things like eq and gt)
+template <> struct simd8<bool> : base8<bool> {
+  static simdjson_inline simd8<bool> splat(bool _value) {
+    return (__m128i)vec_splats((unsigned char)(-(!!_value)));
+  }
+
+  simdjson_inline simd8() : base8<bool>() {}
+  simdjson_inline simd8(const __m128i _value)
+      : base8<bool>(_value) {}
+  // Splat constructor
+  simdjson_inline simd8(bool _value)
+      : base8<bool>(splat(_value)) {}
+
+  simdjson_inline int to_bitmask() const {
+    __vector unsigned long long result;
+    const __m128i perm_mask = {0x78, 0x70, 0x68, 0x60, 0x58, 0x50, 0x48, 0x40,
+                               0x38, 0x30, 0x28, 0x20, 0x18, 0x10, 0x08, 0x00};
+
+    result = ((__vector unsigned long long)vec_vbpermq((__m128i)this->value,
+                                                       (__m128i)perm_mask));
+#ifdef __LITTLE_ENDIAN__
+    return static_cast<int>(result[1]);
+#else
+    return static_cast<int>(result[0]);
+#endif
+  }
+  simdjson_inline bool any() const {
+    return !vec_all_eq(this->value, (__m128i)vec_splats(0));
+  }
+  simdjson_inline simd8<bool> operator~() const {
+    return this->value ^ (__m128i)splat(true);
+  }
+};
+
+template <typename T> struct base8_numeric : base8<T> {
+  static simdjson_inline simd8<T> splat(T value) {
+    (void)value;
+    return (__m128i)vec_splats(value);
+  }
+  static simdjson_inline simd8<T> zero() { return splat(0); }
+  static simdjson_inline simd8<T> load(const T values[16]) {
+    return (__m128i)(vec_vsx_ld(0, reinterpret_cast<const uint8_t *>(values)));
+  }
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  static simdjson_inline simd8<T> repeat_16(T v0, T v1, T v2, T v3, T v4,
+                                                   T v5, T v6, T v7, T v8, T v9,
+                                                   T v10, T v11, T v12, T v13,
+                                                   T v14, T v15) {
+    return simd8<T>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13,
+                    v14, v15);
+  }
+
+  simdjson_inline base8_numeric() : base8<T>() {}
+  simdjson_inline base8_numeric(const __m128i _value)
+      : base8<T>(_value) {}
+
+  // Store to array
+  simdjson_inline void store(T dst[16]) const {
+    vec_vsx_st(this->value, 0, reinterpret_cast<__m128i *>(dst));
+  }
+
+  // Override to distinguish from bool version
+  simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+  // Addition/subtraction are the same for signed and unsigned
+  simdjson_inline simd8<T> operator+(const simd8<T> other) const {
+    return (__m128i)((__m128i)this->value + (__m128i)other);
+  }
+  simdjson_inline simd8<T> operator-(const simd8<T> other) const {
+    return (__m128i)((__m128i)this->value - (__m128i)other);
+  }
+  simdjson_inline simd8<T> &operator+=(const simd8<T> other) {
+    *this = *this + other;
+    return *static_cast<simd8<T> *>(this);
+  }
+  simdjson_inline simd8<T> &operator-=(const simd8<T> other) {
+    *this = *this - other;
+    return *static_cast<simd8<T> *>(this);
+  }
+
+  // Perform a lookup assuming the value is between 0 and 16 (undefined behavior
+  // for out of range values)
+  template <typename L>
+  simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+    return (__m128i)vec_perm((__m128i)lookup_table, (__m128i)lookup_table, this->value);
+  }
+
+  // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted
+  // as a bitset). Passing a 0 value for mask would be equivalent to writing out
+  // every byte to output. Only the first 16 - count_ones(mask) bytes of the
+  // result are significant but 16 bytes get written. Design consideration: it
+  // seems like a function with the signature simd8<L> compress(uint32_t mask)
+  // would be sensible, but the AVX ISA makes this kind of approach difficult.
+  template <typename L>
+  simdjson_inline void compress(uint16_t mask, L *output) const {
+    using internal::BitsSetTable256mul2;
+    using internal::pshufb_combine_table;
+    using internal::thintable_epi8;
+    // this particular implementation was inspired by work done by @animetosho
+    // we do it in two steps, first 8 bytes and then second 8 bytes
+    uint8_t mask1 = uint8_t(mask);      // least significant 8 bits
+    uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
+    // next line just loads the 64-bit values thintable_epi8[mask1] and
+    // thintable_epi8[mask2] into a 128-bit register, using only
+    // two instructions on most compilers.
+#ifdef __LITTLE_ENDIAN__
+    __m128i shufmask = (__m128i)(__vector unsigned long long){
+        thintable_epi8[mask1], thintable_epi8[mask2]};
+#else
+    __m128i shufmask = (__m128i)(__vector unsigned long long){
+        thintable_epi8[mask2], thintable_epi8[mask1]};
+    shufmask = (__m128i)vec_reve((__m128i)shufmask);
+#endif
+    // we increment by 0x08 the second half of the mask
+    shufmask = ((__m128i)shufmask) +
+               ((__m128i)(__vector int){0, 0, 0x08080808, 0x08080808});
+
+    // this is the version "nearly pruned"
+    __m128i pruned = vec_perm(this->value, this->value, shufmask);
+    // we still need to put the two halves together.
+    // we compute the popcount of the first half:
+    int pop1 = BitsSetTable256mul2[mask1];
+    // then load the corresponding mask, what it does is to write
+    // only the first pop1 bytes from the first 8 bytes, and then
+    // it fills in with the bytes from the second 8 bytes + some filling
+    // at the end.
+    __m128i compactmask =
+        vec_vsx_ld(0, reinterpret_cast<const uint8_t *>(pshufb_combine_table + pop1 * 8));
+    __m128i answer = vec_perm(pruned, (__m128i)vec_splats(0), compactmask);
+    vec_vsx_st(answer, 0, reinterpret_cast<__m128i *>(output));
+  }
+
+  template <typename L>
+  simdjson_inline simd8<L>
+  lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
+            L replace5, L replace6, L replace7, L replace8, L replace9,
+            L replace10, L replace11, L replace12, L replace13, L replace14,
+            L replace15) const {
+    return lookup_16(simd8<L>::repeat_16(
+        replace0, replace1, replace2, replace3, replace4, replace5, replace6,
+        replace7, replace8, replace9, replace10, replace11, replace12,
+        replace13, replace14, replace15));
+  }
+};
+
+// Signed bytes
+template <> struct simd8<int8_t> : base8_numeric<int8_t> {
+  simdjson_inline simd8() : base8_numeric<int8_t>() {}
+  simdjson_inline simd8(const __m128i _value)
+      : base8_numeric<int8_t>(_value) {}
+  // Splat constructor
+  simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+  // Array constructor
+  simdjson_inline simd8(const int8_t *values) : simd8(load(values)) {}
+  // Member-by-member initialization
+  simdjson_inline simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3,
+                               int8_t v4, int8_t v5, int8_t v6, int8_t v7,
+                               int8_t v8, int8_t v9, int8_t v10, int8_t v11,
+                               int8_t v12, int8_t v13, int8_t v14, int8_t v15)
+      : simd8((__m128i)(__vector signed char){v0, v1, v2, v3, v4, v5, v6, v7,
+                                              v8, v9, v10, v11, v12, v13, v14,
+                                              v15}) {}
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdjson_inline static simd8<int8_t>
+  repeat_16(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
+            int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
+            int8_t v12, int8_t v13, int8_t v14, int8_t v15) {
+    return simd8<int8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                         v13, v14, v15);
+  }
+
+  // Order-sensitive comparisons
+  simdjson_inline simd8<int8_t>
+  max_val(const simd8<int8_t> other) const {
+    return (__m128i)vec_max((__vector signed char)this->value,
+                            (__vector signed char)(__m128i)other);
+  }
+  simdjson_inline simd8<int8_t>
+  min_val(const simd8<int8_t> other) const {
+    return (__m128i)vec_min((__vector signed char)this->value,
+                            (__vector signed char)(__m128i)other);
+  }
+  simdjson_inline simd8<bool>
+  operator>(const simd8<int8_t> other) const {
+    return (__m128i)vec_cmpgt((__vector signed char)this->value,
+                              (__vector signed char)(__m128i)other);
+  }
+  simdjson_inline simd8<bool>
+  operator<(const simd8<int8_t> other) const {
+    return (__m128i)vec_cmplt((__vector signed char)this->value,
+                              (__vector signed char)(__m128i)other);
+  }
+};
+
+// Unsigned bytes
+template <> struct simd8<uint8_t> : base8_numeric<uint8_t> {
+  simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+  simdjson_inline simd8(const __m128i _value)
+      : base8_numeric<uint8_t>(_value) {}
+  // Splat constructor
+  simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+  // Array constructor
+  simdjson_inline simd8(const uint8_t *values) : simd8(load(values)) {}
+  // Member-by-member initialization
+  simdjson_inline
+  simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
+        uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
+        uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15)
+      : simd8((__m128i){v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                        v13, v14, v15}) {}
+  // Repeat 16 values as many times as necessary (usually for lookup tables)
+  simdjson_inline static simd8<uint8_t>
+  repeat_16(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4,
+            uint8_t v5, uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9,
+            uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14,
+            uint8_t v15) {
+    return simd8<uint8_t>(v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
+                          v13, v14, v15);
+  }
+
+  // Saturated math
+  simdjson_inline simd8<uint8_t>
+  saturating_add(const simd8<uint8_t> other) const {
+    return (__m128i)vec_adds(this->value, (__m128i)other);
+  }
+  simdjson_inline simd8<uint8_t>
+  saturating_sub(const simd8<uint8_t> other) const {
+    return (__m128i)vec_subs(this->value, (__m128i)other);
+  }
+
+  // Order-specific operations
+  simdjson_inline simd8<uint8_t>
+  max_val(const simd8<uint8_t> other) const {
+    return (__m128i)vec_max(this->value, (__m128i)other);
+  }
+  simdjson_inline simd8<uint8_t>
+  min_val(const simd8<uint8_t> other) const {
+    return (__m128i)vec_min(this->value, (__m128i)other);
+  }
+  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+  simdjson_inline simd8<uint8_t>
+  gt_bits(const simd8<uint8_t> other) const {
+    return this->saturating_sub(other);
+  }
+  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+  simdjson_inline simd8<uint8_t>
+  lt_bits(const simd8<uint8_t> other) const {
+    return other.saturating_sub(*this);
+  }
+  simdjson_inline simd8<bool>
+  operator<=(const simd8<uint8_t> other) const {
+    return other.max_val(*this) == other;
+  }
+  simdjson_inline simd8<bool>
+  operator>=(const simd8<uint8_t> other) const {
+    return other.min_val(*this) == other;
+  }
+  simdjson_inline simd8<bool>
+  operator>(const simd8<uint8_t> other) const {
+    return this->gt_bits(other).any_bits_set();
+  }
+  simdjson_inline simd8<bool>
+  operator<(const simd8<uint8_t> other) const {
+    return this->gt_bits(other).any_bits_set();
+  }
+
+  // Bit-specific operations
+  simdjson_inline simd8<bool> bits_not_set() const {
+    return (__m128i)vec_cmpeq(this->value, (__m128i)vec_splats(uint8_t(0)));
+  }
+  simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const {
+    return (*this & bits).bits_not_set();
+  }
+  simdjson_inline simd8<bool> any_bits_set() const {
+    return ~this->bits_not_set();
+  }
+  simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
+    return ~this->bits_not_set(bits);
+  }
+  simdjson_inline bool bits_not_set_anywhere() const {
+    return vec_all_eq(this->value, (__m128i)vec_splats(0));
+  }
+  simdjson_inline bool any_bits_set_anywhere() const {
+    return !bits_not_set_anywhere();
+  }
+  simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const {
+    return vec_all_eq(vec_and(this->value, (__m128i)bits),
+                      (__m128i)vec_splats(0));
+  }
+  simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const {
+    return !bits_not_set_anywhere(bits);
+  }
+  template <int N> simdjson_inline simd8<uint8_t> shr() const {
+    return simd8<uint8_t>(
+        (__m128i)vec_sr(this->value, (__m128i)vec_splat_u8(N)));
+  }
+  template <int N> simdjson_inline simd8<uint8_t> shl() const {
+    return simd8<uint8_t>(
+        (__m128i)vec_sl(this->value, (__m128i)vec_splat_u8(N)));
+  }
+};
+
+template <typename T> struct simd8x64 {
+  static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+  static_assert(NUM_CHUNKS == 4,
+                "PPC64 kernel should use four registers per 64-byte block.");
+  const simd8<T> chunks[NUM_CHUNKS];
+  template<int idx> simd8<uint8_t> get() const { return idx < NUM_CHUNKS ? chunks[idx] : simd8<T>(); }
+
+  simd8x64(const simd8x64<T> &o) = delete; // no copy allowed
+  simd8x64<T> &
+  operator=(const simd8<T>& other) = delete; // no assignment allowed
+  simd8x64() = delete;                      // no default constructor allowed
+
+  simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1,
+                                  const simd8<T> chunk2, const simd8<T> chunk3)
+      : chunks{chunk0, chunk1, chunk2, chunk3} {}
+  simdjson_inline simd8x64(const T ptr[64])
+      : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr + 16),
+               simd8<T>::load(ptr + 32), simd8<T>::load(ptr + 48)} {}
+
+  simdjson_inline void store(T ptr[64]) const {
+    this->chunks[0].store(ptr + sizeof(simd8<T>) * 0);
+    this->chunks[1].store(ptr + sizeof(simd8<T>) * 1);
+    this->chunks[2].store(ptr + sizeof(simd8<T>) * 2);
+    this->chunks[3].store(ptr + sizeof(simd8<T>) * 3);
+  }
+
+  simdjson_inline simd8<T> reduce_or() const {
+    return (this->chunks[0] | this->chunks[1]) |
+           (this->chunks[2] | this->chunks[3]);
+  }
+
+  simdjson_inline uint64_t compress(uint64_t mask, T *output) const {
+    this->chunks[0].compress(uint16_t(mask), output);
+    this->chunks[1].compress(uint16_t(mask >> 16),
+                             output + 16 - count_ones(mask & 0xFFFF));
+    this->chunks[2].compress(uint16_t(mask >> 32),
+                             output + 32 - count_ones(mask & 0xFFFFFFFF));
+    this->chunks[3].compress(uint16_t(mask >> 48),
+                             output + 48 - count_ones(mask & 0xFFFFFFFFFFFF));
+    return 64 - count_ones(mask);
+  }
+
+  simdjson_inline uint64_t to_bitmask() const {
+    uint64_t r0 = uint32_t(this->chunks[0].to_bitmask());
+    uint64_t r1 = this->chunks[1].to_bitmask();
+    uint64_t r2 = this->chunks[2].to_bitmask();
+    uint64_t r3 = this->chunks[3].to_bitmask();
+    return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
+  }
+
+  simdjson_inline uint64_t eq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] == mask, this->chunks[1] == mask,
+                          this->chunks[2] == mask, this->chunks[3] == mask)
+        .to_bitmask();
+  }
+
+  simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+    return simd8x64<bool>(this->chunks[0] == other.chunks[0],
+                          this->chunks[1] == other.chunks[1],
+                          this->chunks[2] == other.chunks[2],
+                          this->chunks[3] == other.chunks[3])
+        .to_bitmask();
+  }
+
+  simdjson_inline uint64_t lteq(const T m) const {
+    const simd8<T> mask = simd8<T>::splat(m);
+    return simd8x64<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
+                          this->chunks[2] <= mask, this->chunks[3] <= mask)
+        .to_bitmask();
+  }
+}; // struct simd8x64<T>
+
+} // namespace simd
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_PPC64_SIMD_INPUT_H
+/* end file simdjson/ppc64/simd.h */
+/* including simdjson/ppc64/stringparsing_defs.h: #include "simdjson/ppc64/stringparsing_defs.h" */
+/* begin file simdjson/ppc64/stringparsing_defs.h */
+#ifndef SIMDJSON_PPC64_STRINGPARSING_DEFS_H
+#define SIMDJSON_PPC64_STRINGPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/simd.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 32;
+  simdjson_inline backslash_and_quote
+  copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() {
+    return ((bs_bits - 1) & quote_bits) != 0;
+  }
+  simdjson_inline bool has_backslash() { return bs_bits != 0; }
+  simdjson_inline int quote_index() {
+    return trailing_zeroes(quote_bits);
+  }
+  simdjson_inline int backslash_index() {
+    return trailing_zeroes(bs_bits);
+  }
+
+  uint32_t bs_bits;
+  uint32_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote
+backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // this can read up to 31 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1),
+                "backslash and quote finder must process fewer than "
+                "SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v0(src);
+  simd8<uint8_t> v1(src + sizeof(v0));
+  v0.store(dst);
+  v1.store(dst + sizeof(v0));
+
+  // Getting a 64-bit bitmask is much cheaper than multiple 16-bit bitmasks on
+  // PPC; therefore, we smash them together into a 64-byte mask and get the
+  // bitmask from there.
+  uint64_t bs_and_quote =
+      simd8x64<bool>(v0 == '\\', v1 == '\\', v0 == '"', v1 == '"').to_bitmask();
+  return {
+      uint32_t(bs_and_quote),      // bs_bits
+      uint32_t(bs_and_quote >> 32) // quote_bits
+  };
+}
+
+
+struct escaping {
+  static constexpr uint32_t BYTES_PROCESSED = 16;
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits != 0; }
+  simdjson_inline int escape_index() { return trailing_zeroes(escape_bits); }
+
+  uint64_t escape_bits;
+}; // struct escaping
+
+
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "escaping finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  simd8<bool> is_quote = (v == '"');
+  simd8<bool> is_backslash = (v == '\\');
+  simd8<bool> is_control = (v < 32);
+  return {
+    // We store it as a 64-bit bitmask even though we only need 16 bits.
+    uint64_t((is_backslash | is_quote | is_control).to_bitmask())
+  };
+}
+
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_PPC64_STRINGPARSING_DEFS_H
+/* end file simdjson/ppc64/stringparsing_defs.h */
+
+#define SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT 1
+/* end file simdjson/ppc64/begin.h */
+/* including generic/amalgamated.h for ppc64: #include <generic/amalgamated.h> */
+/* begin file generic/amalgamated.h for ppc64 */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_SRC_GENERIC_DEPENDENCIES_H)
+#error generic/dependencies.h must be included before generic/amalgamated.h!
+#endif
+
+/* including generic/base.h for ppc64: #include <generic/base.h> */
+/* begin file generic/base.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+
+struct json_character_block;
+
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_BASE_H
+/* end file generic/base.h for ppc64 */
+/* including generic/dom_parser_implementation.h for ppc64: #include <generic/dom_parser_implementation.h> */
+/* begin file generic/dom_parser_implementation.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// Interface a dom parser implementation must fulfill
+namespace simdjson {
+namespace ppc64 {
+namespace {
+
+simdjson_inline simd8<uint8_t> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3);
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input);
+
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file generic/dom_parser_implementation.h for ppc64 */
+/* including generic/json_character_block.h for ppc64: #include <generic/json_character_block.h> */
+/* begin file generic/json_character_block.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+
+struct json_character_block {
+  static simdjson_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
+
+  simdjson_inline uint64_t whitespace() const noexcept { return _whitespace; }
+  simdjson_inline uint64_t op() const noexcept { return _op; }
+  simdjson_inline uint64_t scalar() const noexcept { return ~(op() | whitespace()); }
+
+  uint64_t _whitespace;
+  uint64_t _op;
+};
+
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H
+/* end file generic/json_character_block.h for ppc64 */
+/* end file generic/amalgamated.h for ppc64 */
+/* including generic/stage1/amalgamated.h for ppc64: #include <generic/stage1/amalgamated.h> */
+/* begin file generic/stage1/amalgamated.h for ppc64 */
+// Stuff other things depend on
+/* including generic/stage1/base.h for ppc64: #include <generic/stage1/base.h> */
+/* begin file generic/stage1/base.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_BASE_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace stage1 {
+
+class bit_indexer;
+template<size_t STEP_SIZE>
+struct buf_block_reader;
+struct json_block;
+class json_minifier;
+class json_scanner;
+struct json_string_block;
+class json_string_scanner;
+class json_structural_indexer;
+
+} // namespace stage1
+
+namespace utf8_validation {
+struct utf8_checker;
+} // namespace utf8_validation
+
+using utf8_validation::utf8_checker;
+
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_BASE_H
+/* end file generic/stage1/base.h for ppc64 */
+/* including generic/stage1/buf_block_reader.h for ppc64: #include <generic/stage1/buf_block_reader.h> */
+/* begin file generic/stage1/buf_block_reader.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace stage1 {
+
+// Walks through a buffer in block-sized increments, loading the last part with spaces
+template<size_t STEP_SIZE>
+struct buf_block_reader {
+public:
+  simdjson_inline buf_block_reader(const uint8_t *_buf, size_t _len);
+  simdjson_inline size_t block_index();
+  simdjson_inline bool has_full_block() const;
+  simdjson_inline const uint8_t *full_block() const;
+  /**
+   * Get the last block, padded with spaces.
+   *
+   * There will always be a last block, with at least 1 byte, unless len == 0 (in which case this
+   * function fills the buffer with spaces and returns 0. In particular, if len == STEP_SIZE there
+   * will be 0 full_blocks and 1 remainder block with STEP_SIZE bytes and no spaces for padding.
+   *
+   * @return the number of effective characters in the last block.
+   */
+  simdjson_inline size_t get_remainder(uint8_t *dst) const;
+  simdjson_inline void advance();
+private:
+  const uint8_t *buf;
+  const size_t len;
+  const size_t lenminusstep;
+  size_t idx;
+};
+
+// Routines to print masks and text for debugging bitmask operations
+simdjson_unused static char * format_input_text_64(const uint8_t *text) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    buf[i] = int8_t(text[i]) < ' ' ? '_' : int8_t(text[i]);
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+// Routines to print masks and text for debugging bitmask operations
+simdjson_unused static char * format_input_text(const simd8x64<uint8_t>& in) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  in.store(reinterpret_cast<uint8_t*>(buf));
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] < ' ') { buf[i] = '_'; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+simdjson_unused static char * format_input_text(const simd8x64<uint8_t>& in, uint64_t mask) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  in.store(reinterpret_cast<uint8_t*>(buf));
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] <= ' ') { buf[i] = '_'; }
+    if (!(mask & (size_t(1) << i))) { buf[i] = ' '; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+simdjson_unused static char * format_mask(uint64_t mask) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  for (size_t i=0; i<64; i++) {
+    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
+  }
+  buf[64] = '\0';
+  return buf;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
+
+template<size_t STEP_SIZE>
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
+
+template<size_t STEP_SIZE>
+simdjson_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
+  return idx < lenminusstep;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
+  return &buf[idx];
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
+  if(len == idx) { return 0; } // memcpy(dst, null, 0) will trigger an error with some sanitizers
+  std::memset(dst, 0x20, STEP_SIZE); // std::memset STEP_SIZE because it's more efficient to write out 8 or 16 bytes at once.
+  std::memcpy(dst, buf + idx, len - idx);
+  return len - idx;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline void buf_block_reader<STEP_SIZE>::advance() {
+  idx += STEP_SIZE;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H
+/* end file generic/stage1/buf_block_reader.h for ppc64 */
+/* including generic/stage1/json_escape_scanner.h for ppc64: #include <generic/stage1/json_escape_scanner.h> */
+/* begin file generic/stage1/json_escape_scanner.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_ESCAPE_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_ESCAPE_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace stage1 {
+
+/**
+ * Scans for escape characters in JSON, taking care with multiple backslashes (\\n vs. \n).
+ */
+struct json_escape_scanner {
+  /** The actual escape characters (the backslashes themselves). */
+  uint64_t next_is_escaped = 0ULL;
+
+  struct escaped_and_escape {
+    /**
+     * Mask of escaped characters.
+     *
+     * ```
+     * \n \\n \\\n \\\\n \
+     * 0100100010100101000
+     *  n  \   \ n  \ \
+     * ```
+     */
+    uint64_t escaped;
+    /**
+     * Mask of escape characters.
+     *
+     * ```
+     * \n \\n \\\n \\\\n \
+     * 1001000101001010001
+     * \  \   \ \  \ \   \
+     * ```
+     */
+    uint64_t escape;
+  };
+
+  /**
+   * Get a mask of both escape and escaped characters (the characters following a backslash).
+   *
+   * @param potential_escape A mask of the character that can escape others (but could be
+   *        escaped itself). e.g. block.eq('\\')
+   */
+  simdjson_really_inline escaped_and_escape next(uint64_t backslash) noexcept {
+
+#if !SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+    if (!backslash) { return {next_escaped_without_backslashes(), 0}; }
+#endif
+
+    // |                                | Mask (shows characters instead of 1's) | Depth | Instructions        |
+    // |--------------------------------|----------------------------------------|-------|---------------------|
+    // | string                         | `\\n_\\\n___\\\n___\\\\___\\\\__\\\`   |       |                     |
+    // |                                | `    even   odd    even   odd   odd`   |       |                     |
+    // | potential_escape               | ` \  \\\    \\\    \\\\   \\\\  \\\`   | 1     | 1 (backslash & ~first_is_escaped)
+    // | escape_and_terminal_code       | ` \n \ \n   \ \n   \ \    \ \   \ \`   | 5     | 5 (next_escape_and_terminal_code())
+    // | escaped                        | `\    \ n    \ n    \ \    \ \   \ ` X | 6     | 7 (escape_and_terminal_code ^ (potential_escape | first_is_escaped))
+    // | escape                         | `    \ \    \ \    \ \    \ \   \ \`   | 6     | 8 (escape_and_terminal_code & backslash)
+    // | first_is_escaped               | `\                                 `   | 7 (*) | 9 (escape >> 63) ()
+    //                                                                               (*) this is not needed until the next iteration
+    uint64_t escape_and_terminal_code = next_escape_and_terminal_code(backslash & ~this->next_is_escaped);
+    uint64_t escaped = escape_and_terminal_code ^ (backslash | this->next_is_escaped);
+    uint64_t escape = escape_and_terminal_code & backslash;
+    this->next_is_escaped = escape >> 63;
+    return {escaped, escape};
+  }
+
+private:
+  static constexpr const uint64_t ODD_BITS = 0xAAAAAAAAAAAAAAAAULL;
+
+  simdjson_really_inline uint64_t next_escaped_without_backslashes() noexcept {
+    uint64_t escaped = this->next_is_escaped;
+    this->next_is_escaped = 0;
+    return escaped;
+  }
+
+  /**
+   * Returns a mask of the next escape characters (masking out escaped backslashes), along with
+   * any non-backslash escape codes.
+   *
+   * \n \\n \\\n \\\\n returns:
+   * \n \   \ \n \ \
+   * 11 100 1011 10100
+   *
+   * You are expected to mask out the first bit yourself if the previous block had a trailing
+   * escape.
+   *
+   * & the result with potential_escape to get just the escape characters.
+   * ^ the result with (potential_escape | first_is_escaped) to get escaped characters.
+   */
+  static simdjson_really_inline uint64_t next_escape_and_terminal_code(uint64_t potential_escape) noexcept {
+    // If we were to just shift and mask out any odd bits, we'd actually get a *half* right answer:
+    // any even-aligned backslash runs would be correct! Odd-aligned backslash runs would be
+    // inverted (\\\ would be 010 instead of 101).
+    //
+    // ```
+    // string:              | ____\\\\_\\\\_____ |
+    // maybe_escaped | ODD  |     \ \   \ \      |
+    //               even-aligned ^^^  ^^^^ odd-aligned
+    // ```
+    //
+    // Taking that into account, our basic strategy is:
+    //
+    // 1. Use subtraction to produce a mask with 1's for even-aligned runs and 0's for
+    //    odd-aligned runs.
+    // 2. XOR all odd bits, which masks out the odd bits in even-aligned runs, and brings IN the
+    //    odd bits in odd-aligned runs.
+    // 3. & with backslash to clean up any stray bits.
+    // runs are set to 0, and then XORing with "odd":
+    //
+    // |                                | Mask (shows characters instead of 1's) | Instructions        |
+    // |--------------------------------|----------------------------------------|---------------------|
+    // | string                         | `\\n_\\\n___\\\n___\\\\___\\\\__\\\`   |
+    // |                                | `    even   odd    even   odd   odd`   |
+    // | maybe_escaped                  | `  n  \\n    \\n    \\\_   \\\_  \\` X | 1 (potential_escape << 1)
+    // | maybe_escaped_and_odd          | ` \n_ \\n _ \\\n_ _ \\\__ _\\\_ \\\`   | 1 (maybe_escaped | odd)
+    // | even_series_codes_and_odd      | `  n_\\\  _    n_ _\\\\ _     _    `   | 1 (maybe_escaped_and_odd - potential_escape)
+    // | escape_and_terminal_code       | ` \n \ \n   \ \n   \ \    \ \   \ \`   | 1 (^ odd)
+    //
+
+    // Escaped characters are characters following an escape.
+    uint64_t maybe_escaped = potential_escape << 1;
+
+    // To distinguish odd from even escape sequences, therefore, we turn on any *starting*
+    // escapes that are on an odd byte. (We actually bring in all odd bits, for speed.)
+    // - Odd runs of backslashes are 0000, and the code at the end ("n" in \n or \\n) is 1.
+    // - Odd runs of backslashes are 1111, and the code at the end ("n" in \n or \\n) is 0.
+    // - All other odd bytes are 1, and even bytes are 0.
+    uint64_t maybe_escaped_and_odd_bits     = maybe_escaped | ODD_BITS;
+    uint64_t even_series_codes_and_odd_bits = maybe_escaped_and_odd_bits - potential_escape;
+
+    // Now we flip all odd bytes back with xor. This:
+    // - Makes odd runs of backslashes go from 0000 to 1010
+    // - Makes even runs of backslashes go from 1111 to 1010
+    // - Sets actually-escaped codes to 1 (the n in \n and \\n: \n = 11, \\n = 100)
+    // - Resets all other bytes to 0
+    return even_series_codes_and_odd_bits ^ ODD_BITS;
+  }
+};
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+/* end file generic/stage1/json_escape_scanner.h for ppc64 */
+/* including generic/stage1/json_string_scanner.h for ppc64: #include <generic/stage1/json_string_scanner.h> */
+/* begin file generic/stage1/json_string_scanner.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_escape_scanner.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace stage1 {
+
+struct json_string_block {
+  // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
+  simdjson_really_inline json_string_block(uint64_t escaped, uint64_t quote, uint64_t in_string) :
+  _escaped(escaped), _quote(quote), _in_string(in_string) {}
+
+  // Escaped characters (characters following an escape() character)
+  simdjson_really_inline uint64_t escaped() const { return _escaped; }
+  // Real (non-backslashed) quotes
+  simdjson_really_inline uint64_t quote() const { return _quote; }
+  // Only characters inside the string (not including the quotes)
+  simdjson_really_inline uint64_t string_content() const { return _in_string & ~_quote; }
+  // Return a mask of whether the given characters are inside a string (only works on non-quotes)
+  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
+  // Return a mask of whether the given characters are inside a string (only works on non-quotes)
+  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
+  // Tail of string (everything except the start quote)
+  simdjson_really_inline uint64_t string_tail() const { return _in_string ^ _quote; }
+
+  // escaped characters (backslashed--does not include the hex characters after \u)
+  uint64_t _escaped;
+  // real quotes (non-escaped ones)
+  uint64_t _quote;
+  // string characters (includes start quote but not end quote)
+  uint64_t _in_string;
+};
+
+// Scans blocks for string characters, storing the state necessary to do so
+class json_string_scanner {
+public:
+  simdjson_really_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
+  // Returns either UNCLOSED_STRING or SUCCESS
+  simdjson_really_inline error_code finish();
+
+private:
+  // Scans for escape characters
+  json_escape_scanner escape_scanner{};
+  // Whether the last iteration was still inside a string (all 1's = true, all 0's = false).
+  uint64_t prev_in_string = 0ULL;
+};
+
+//
+// Return a mask of all string characters plus end quotes.
+//
+// prev_escaped is overflow saying whether the next character is escaped.
+// prev_in_string is overflow saying whether we're still in a string.
+//
+// Backslash sequences outside of quotes will be detected in stage 2.
+//
+simdjson_really_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
+  const uint64_t backslash = in.eq('\\');
+  const uint64_t escaped = escape_scanner.next(backslash).escaped;
+  const uint64_t quote = in.eq('"') & ~escaped;
+
+  //
+  // prefix_xor flips on bits inside the string (and flips off the end quote).
+  //
+  // Then we xor with prev_in_string: if we were in a string already, its effect is flipped
+  // (characters inside strings are outside, and characters outside strings are inside).
+  //
+  const uint64_t in_string = prefix_xor(quote) ^ prev_in_string;
+
+  //
+  // Check if we're still in a string at the end of the box so the next block will know
+  //
+  prev_in_string = uint64_t(static_cast<int64_t>(in_string) >> 63);
+
+  // Use ^ to turn the beginning quote off, and the end quote on.
+
+  // We are returning a function-local object so either we get a move constructor
+  // or we get copy elision.
+  return json_string_block(escaped, quote, in_string);
+}
+
+simdjson_really_inline error_code json_string_scanner::finish() {
+  if (prev_in_string) {
+    return UNCLOSED_STRING;
+  }
+  return SUCCESS;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+/* end file generic/stage1/json_string_scanner.h for ppc64 */
+/* including generic/stage1/utf8_lookup4_algorithm.h for ppc64: #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* begin file generic/stage1/utf8_lookup4_algorithm.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace utf8_validation {
+
+using namespace simd;
+
+  simdjson_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+// Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+// Bit 1 = Too Long (ASCII followed by continuation)
+// Bit 2 = Overlong 3-byte
+// Bit 4 = Surrogate
+// Bit 5 = Overlong 2-byte
+// Bit 7 = Two Continuations
+    constexpr const uint8_t TOO_SHORT   = 1<<0; // 11______ 0_______
+                                                // 11______ 11______
+    constexpr const uint8_t TOO_LONG    = 1<<1; // 0_______ 10______
+    constexpr const uint8_t OVERLONG_3  = 1<<2; // 11100000 100_____
+    constexpr const uint8_t SURROGATE   = 1<<4; // 11101101 101_____
+    constexpr const uint8_t OVERLONG_2  = 1<<5; // 1100000_ 10______
+    constexpr const uint8_t TWO_CONTS   = 1<<7; // 10______ 10______
+    constexpr const uint8_t TOO_LARGE   = 1<<3; // 11110100 1001____
+                                                // 11110100 101_____
+                                                // 11110101 1001____
+                                                // 11110101 101_____
+                                                // 1111011_ 1001____
+                                                // 1111011_ 101_____
+                                                // 11111___ 1001____
+                                                // 11111___ 101_____
+    constexpr const uint8_t TOO_LARGE_1000 = 1<<6;
+                                                // 11110101 1000____
+                                                // 1111011_ 1000____
+                                                // 11111___ 1000____
+    constexpr const uint8_t OVERLONG_4  = 1<<6; // 11110000 1000____
+
+    const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      TOO_SHORT,
+      // 1110____ ________ <three byte lead in byte 1>
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4
+    );
+    constexpr const uint8_t CARRY = TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+    const simd8<uint8_t> byte_1_low = (prev1 & 0x0F).lookup_16<uint8_t>(
+      // ____0000 ________
+      CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+      // ____0001 ________
+      CARRY | OVERLONG_2,
+      // ____001_ ________
+      CARRY,
+      CARRY,
+
+      // ____0100 ________
+      CARRY | TOO_LARGE,
+      // ____0101 ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      // ____011_ ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+
+      // ____1___ ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      // ____1101 ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000
+    );
+    const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE  | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE  | TOO_LARGE,
+
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT
+    );
+    return (byte_1_high & byte_1_low & byte_2_high);
+  }
+  simdjson_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
+      const simd8<uint8_t> prev_input, const simd8<uint8_t> sc) {
+    simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+    simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+    simd8<uint8_t> must23 = must_be_2_3_continuation(prev2, prev3);
+    simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+    return must23_80 ^ sc;
+  }
+
+  //
+  // Return nonzero if there are incomplete multibyte characters at the end of the block:
+  // e.g. if there is a 4-byte character, but it's 3 bytes from the end.
+  //
+  simdjson_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+    // If the previous input's last 3 bytes match this, they're too short (they ended at EOF):
+    // ... 1111____ 111_____ 11______
+#if SIMDJSON_IMPLEMENTATION_ICELAKE || (SIMDJSON_IMPLEMENTATION_RVV_VLS && __riscv_v_fixed_vlen >= 512)
+    static const uint8_t max_array[64] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
+    };
+#else
+    static const uint8_t max_array[32] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
+    };
+#endif
+    const simd8<uint8_t> max_value(&max_array[sizeof(max_array)-sizeof(simd8<uint8_t>)]);
+    return input.gt_bits(max_value);
+  }
+
+  struct utf8_checker {
+    // If this is nonzero, there has been a UTF-8 error.
+    simd8<uint8_t> error;
+    // The last input we received
+    simd8<uint8_t> prev_input_block;
+    // Whether the last input we received was incomplete (used for ASCII fast path)
+    simd8<uint8_t> prev_incomplete;
+
+    //
+    // Check whether the current bytes are valid UTF-8.
+    //
+    simdjson_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
+      // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ lead bytes
+      // (2, 3, 4-byte leads become large positive numbers instead of small negative numbers)
+      simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+      simd8<uint8_t> sc = check_special_cases(input, prev1);
+      this->error |= check_multibyte_lengths(input, prev_input, sc);
+    }
+
+    // The only problem that can happen at EOF is that a multibyte character is too short
+    // or a byte value too large in the last bytes: check_special_cases only checks for bytes
+    // too large in the first of two bytes.
+    simdjson_inline void check_eof() {
+      // If the previous block had incomplete UTF-8 characters at the end, an ASCII block can't
+      // possibly finish them.
+      this->error |= this->prev_incomplete;
+    }
+
+    simdjson_inline void check_next_input(const simd8x64<uint8_t>& input) {
+      if(simdjson_likely(is_ascii(input))) {
+        this->error |= this->prev_incomplete;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio, it is not good enough.
+        static_assert((simd8x64<uint8_t>::NUM_CHUNKS == 1)
+                ||(simd8x64<uint8_t>::NUM_CHUNKS == 2)
+                || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+                "We support one, two or four chunks per 64-byte block.");
+        SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 1) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+          this->check_utf8_bytes(input.get<1>(), input.get<0>());
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+          this->check_utf8_bytes(input.get<1>(), input.get<0>());
+          this->check_utf8_bytes(input.get<2>(), input.get<1>());
+          this->check_utf8_bytes(input.get<3>(), input.get<2>());
+        }
+        this->prev_incomplete = is_incomplete(input.get<simd8x64<uint8_t>::NUM_CHUNKS-1>());
+        this->prev_input_block = input.get<simd8x64<uint8_t>::NUM_CHUNKS-1>();
+      }
+    }
+    // do not forget to call check_eof!
+    simdjson_warn_unused simdjson_inline error_code errors() {
+      return this->error.any_bits_set_anywhere() ? error_code::UTF8_ERROR : error_code::SUCCESS;
+    }
+
+  }; // struct utf8_checker
+} // namespace utf8_validation
+
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H
+/* end file generic/stage1/utf8_lookup4_algorithm.h for ppc64 */
+/* including generic/stage1/json_scanner.h for ppc64: #include <generic/stage1/json_scanner.h> */
+/* begin file generic/stage1/json_scanner.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/json_character_block.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_string_scanner.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace stage1 {
+
+/**
+ * A block of scanned json, with information on operators and scalars.
+ *
+ * We seek to identify pseudo-structural characters. Anything that is inside
+ * a string must be omitted (hence  & ~_string.string_tail()).
+ * Otherwise, pseudo-structural characters come in two forms.
+ * 1. We have the structural characters ([,],{,},:, comma). The
+ *    term 'structural character' is from the JSON RFC.
+ * 2. We have the 'scalar pseudo-structural characters'.
+ *    Scalars are quotes, and any character except structural characters and white space.
+ *
+ * To identify the scalar pseudo-structural characters, we must look at what comes
+ * before them: it must be a space, a quote or a structural characters.
+ * Starting with simdjson v0.3, we identify them by
+ * negation: we identify everything that is followed by a non-quote scalar,
+ * and we negate that. Whatever remains must be a 'scalar pseudo-structural character'.
+ */
+struct json_block {
+public:
+  // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
+  simdjson_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  _string(std::move(string)), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
+  simdjson_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  _string(string), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
+
+  /**
+   * The start of structurals.
+   * In simdjson prior to v0.3, these were called the pseudo-structural characters.
+   **/
+  simdjson_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
+  /** All JSON whitespace (i.e. not in a string) */
+  simdjson_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
+
+  // Helpers
+
+  /** Whether the given characters are inside a string (only works on non-quotes) */
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
+  /** Whether the given characters are outside a string (only works on non-quotes) */
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
+
+  // string and escape characters
+  json_string_block _string;
+  // whitespace, structural characters ('operators'), scalars
+  json_character_block _characters;
+  // whether the previous character was a scalar
+  uint64_t _follows_potential_nonquote_scalar;
+private:
+  // Potential structurals (i.e. disregarding strings)
+
+  /**
+   * structural elements ([,],{,},:, comma) plus scalar starts like 123, true and "abc".
+   * They may reside inside a string.
+   **/
+  simdjson_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
+  /**
+   * The start of non-operator runs, like 123, true and "abc".
+   * It main reside inside a string.
+   **/
+  simdjson_inline uint64_t potential_scalar_start() const noexcept {
+    // The term "scalar" refers to anything except structural characters and white space
+    // (so letters, numbers, quotes).
+    // Whenever it is preceded by something that is not a structural element ({,},[,],:, ") nor a white-space
+    // then we know that it is irrelevant structurally.
+    return _characters.scalar() & ~follows_potential_scalar();
+  }
+  /**
+   * Whether the given character is immediately after a non-operator like 123, true.
+   * The characters following a quote are not included.
+   */
+  simdjson_inline uint64_t follows_potential_scalar() const noexcept {
+    // _follows_potential_nonquote_scalar: is defined as marking any character that follows a character
+    // that is not a structural element ({,},[,],:, comma) nor a quote (") and that is not a
+    // white space.
+    // It is understood that within quoted region, anything at all could be marked (irrelevant).
+    return _follows_potential_nonquote_scalar;
+  }
+};
+
+/**
+ * Scans JSON for important bits: structural characters or 'operators', strings, and scalars.
+ *
+ * The scanner starts by calculating two distinct things:
+ * - string characters (taking \" into account)
+ * - structural characters or 'operators' ([]{},:, comma)
+ *   and scalars (runs of non-operators like 123, true and "abc")
+ *
+ * To minimize data dependency (a key component of the scanner's speed), it finds these in parallel:
+ * in particular, the operator/scalar bit will find plenty of things that are actually part of
+ * strings. When we're done, json_block will fuse the two together by masking out tokens that are
+ * part of a string.
+ */
+class json_scanner {
+public:
+  json_scanner() = default;
+  simdjson_inline json_block next(const simd::simd8x64<uint8_t>& in);
+  // Returns either UNCLOSED_STRING or SUCCESS
+  simdjson_warn_unused simdjson_inline error_code finish();
+
+private:
+  // Whether the last character of the previous iteration is part of a scalar token
+  // (anything except whitespace or a structural character/'operator').
+  uint64_t prev_scalar = 0ULL;
+  json_string_scanner string_scanner{};
+};
+
+
+//
+// Check if the current character immediately follows a matching character.
+//
+// For example, this checks for quotes with backslashes in front of them:
+//
+//     const uint64_t backslashed_quote = in.eq('"') & immediately_follows(in.eq('\'), prev_backslash);
+//
+simdjson_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
+  const uint64_t result = match << 1 | overflow;
+  overflow = match >> 63;
+  return result;
+}
+
+simdjson_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
+  json_string_block strings = string_scanner.next(in);
+  // identifies the white-space and the structural characters
+  json_character_block characters = json_character_block::classify(in);
+  // The term "scalar" refers to anything except structural characters and white space
+  // (so letters, numbers, quotes).
+  // We want follows_scalar to mark anything that follows a non-quote scalar (so letters and numbers).
+  //
+  // A terminal quote should either be followed by a structural character (comma, brace, bracket, colon)
+  // or nothing. However, we still want ' "a string"true ' to mark the 't' of 'true' as a potential
+  // pseudo-structural character just like we would if we had  ' "a string" true '; otherwise we
+  // may need to add an extra check when parsing strings.
+  //
+  // Performance: there are many ways to skin this cat.
+  const uint64_t nonquote_scalar = characters.scalar() & ~strings.quote();
+  uint64_t follows_nonquote_scalar = follows(nonquote_scalar, prev_scalar);
+  // We are returning a function-local object so either we get a move constructor
+  // or we get copy elision.
+  return json_block(
+    strings,// strings is a function-local object so either it moves or the copy is elided.
+    characters,
+    follows_nonquote_scalar
+  );
+}
+
+simdjson_warn_unused simdjson_inline error_code json_scanner::finish() {
+  return string_scanner.finish();
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H
+/* end file generic/stage1/json_scanner.h for ppc64 */
+
+// All other declarations
+/* including generic/stage1/find_next_document_index.h for ppc64: #include <generic/stage1/find_next_document_index.h> */
+/* begin file generic/stage1/find_next_document_index.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace stage1 {
+
+/**
+  * This algorithm is used to quickly identify the last structural position that
+  * makes up a complete document.
+  *
+  * It does this by going backwards and finding the last *document boundary* (a
+  * place where one value follows another without a comma between them). If the
+  * last document (the characters after the boundary) has an equal number of
+  * start and end brackets, it is considered complete.
+  *
+  * Simply put, we iterate over the structural characters, starting from
+  * the end. We consider that we found the end of a JSON document when the
+  * first element of the pair is NOT one of these characters: '{' '[' ':' ','
+  * and when the second element is NOT one of these characters: '}' ']' ':' ','.
+  *
+  * This simple comparison works most of the time, but it does not cover cases
+  * where the batch's structural indexes contain a perfect amount of documents.
+  * In such a case, we do not have access to the structural index which follows
+  * the last document, therefore, we do not have access to the second element in
+  * the pair, and that means we cannot identify the last document. To fix this
+  * issue, we keep a count of the open and closed curly/square braces we found
+  * while searching for the pair. When we find a pair AND the count of open and
+  * closed curly/square braces is the same, we know that we just passed a
+  * complete document, therefore the last json buffer location is the end of the
+  * batch.
+  */
+simdjson_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
+  // Variant: do not count separately, just figure out depth
+  if(parser.n_structural_indexes == 0) { return 0; }
+  auto arr_cnt = 0;
+  auto obj_cnt = 0;
+  for (auto i = parser.n_structural_indexes - 1; i > 0; i--) {
+    auto idxb = parser.structural_indexes[i];
+    switch (parser.buf[idxb]) {
+    case ':':
+    case ',':
+      continue;
+    case '}':
+      obj_cnt--;
+      continue;
+    case ']':
+      arr_cnt--;
+      continue;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+    }
+    auto idxa = parser.structural_indexes[i - 1];
+    switch (parser.buf[idxa]) {
+    case '{':
+    case '[':
+    case ':':
+    case ',':
+      continue;
+    }
+    // Last document is complete, so the next document will appear after!
+    if (!arr_cnt && !obj_cnt) {
+      return parser.n_structural_indexes;
+    }
+    // Last document is incomplete; mark the document at i + 1 as the next one
+    return i;
+  }
+  // If we made it to the end, we want to finish counting to see if we have a full document.
+  switch (parser.buf[parser.structural_indexes[0]]) {
+    case '}':
+      obj_cnt--;
+      break;
+    case ']':
+      arr_cnt--;
+      break;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+  }
+  if (!arr_cnt && !obj_cnt) {
+    // We have a complete document.
+    return parser.n_structural_indexes;
+  }
+  return 0;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+/* end file generic/stage1/find_next_document_index.h for ppc64 */
+/* including generic/stage1/json_minifier.h for ppc64: #include <generic/stage1/json_minifier.h> */
+/* begin file generic/stage1/json_minifier.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses in stage1
+// It is intended to be included multiple times and compiled multiple times
+// We assume the file in which it is included already includes
+// "simdjson/stage1.h" (this simplifies amalgation)
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace stage1 {
+
+class json_minifier {
+public:
+  template<size_t STEP_SIZE>
+  static error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept;
+
+private:
+  simdjson_inline json_minifier(uint8_t *_dst)
+  : dst{_dst}
+  {}
+  template<size_t STEP_SIZE>
+  simdjson_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
+  simdjson_warn_unused simdjson_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
+  json_scanner scanner{};
+  uint8_t *dst;
+};
+
+simdjson_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
+  uint64_t mask = block.whitespace();
+  dst += in.compress(mask, dst);
+}
+
+simdjson_warn_unused simdjson_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
+  error_code error = scanner.finish();
+  if (error) { dst_len = 0; return error; }
+  dst_len = dst - dst_start;
+  return SUCCESS;
+}
+
+template<>
+simdjson_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block_buf);
+  simd::simd8x64<uint8_t> in_2(block_buf+64);
+  json_block block_1 = scanner.next(in_1);
+  json_block block_2 = scanner.next(in_2);
+  this->next(in_1, block_1);
+  this->next(in_2, block_2);
+  reader.advance();
+}
+
+template<>
+simdjson_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block_buf);
+  json_block block_1 = scanner.next(in_1);
+  this->next(block_buf, block_1);
+  reader.advance();
+}
+
+template<size_t STEP_SIZE>
+error_code json_minifier::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept {
+  buf_block_reader<STEP_SIZE> reader(buf, len);
+  json_minifier minifier(dst);
+
+  // Index the first n-1 blocks
+  while (reader.has_full_block()) {
+    minifier.step<STEP_SIZE>(reader.full_block(), reader);
+  }
+
+  // Index the last (remainder) block, padded with spaces
+  uint8_t block[STEP_SIZE];
+  size_t remaining_bytes = reader.get_remainder(block);
+  if (remaining_bytes > 0) {
+    // We do not want to write directly to the output stream. Rather, we write
+    // to a local buffer (for safety).
+    uint8_t out_block[STEP_SIZE];
+    uint8_t * const guarded_dst{minifier.dst};
+    minifier.dst = out_block;
+    minifier.step<STEP_SIZE>(block, reader);
+    size_t to_write = minifier.dst - out_block;
+    // In some cases, we could be enticed to consider the padded spaces
+    // as part of the string. This is fine as long as we do not write more
+    // than we consumed.
+    if(to_write > remaining_bytes) { to_write = remaining_bytes; }
+    memcpy(guarded_dst, out_block, to_write);
+    minifier.dst = guarded_dst + to_write;
+  }
+  return minifier.finish(dst, dst_len);
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H
+/* end file generic/stage1/json_minifier.h for ppc64 */
+/* including generic/stage1/json_structural_indexer.h for ppc64: #include <generic/stage1/json_structural_indexer.h> */
+/* begin file generic/stage1/json_structural_indexer.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_string_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_minifier.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/find_next_document_index.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses in stage1
+// It is intended to be included multiple times and compiled multiple times
+// We assume the file in which it is included already includes
+// "simdjson/stage1.h" (this simplifies amalgation)
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace stage1 {
+
+class bit_indexer {
+public:
+  uint32_t *tail;
+
+  simdjson_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
+
+#if SIMDJSON_PREFER_REVERSE_BITS
+  /**
+    * ARM lacks a fast trailing zero instruction, but it has a fast
+    * bit reversal instruction and a fast leading zero instruction.
+    * Thus it may be profitable to reverse the bits (once) and then
+    * to rely on a sequence of instructions that call the leading
+    * zero instruction.
+    *
+    * Performance notes:
+    * The chosen routine is not optimal in terms of data dependency
+    * since zero_leading_bit might require two instructions. However,
+    * it tends to minimize the total number of instructions which is
+    * beneficial.
+    */
+  simdjson_inline void write_index(uint32_t idx, uint64_t& rev_bits, int i) {
+    int lz = leading_zeroes(rev_bits);
+    this->tail[i] = static_cast<uint32_t>(idx) + lz;
+    rev_bits = zero_leading_bit(rev_bits, lz);
+  }
+#else
+  /**
+    * Under recent x64 systems, we often have both a fast trailing zero
+    * instruction and a fast 'clear-lower-bit' instruction so the following
+    * algorithm can be competitive.
+    */
+
+  simdjson_inline void write_index(uint32_t idx, uint64_t& bits, int i) {
+    this->tail[i] = idx + trailing_zeroes(bits);
+    bits = clear_lowest_bit(bits);
+  }
+#endif // SIMDJSON_PREFER_REVERSE_BITS
+
+  template <int START, int N>
+  simdjson_inline int write_indexes(uint32_t idx, uint64_t& bits) {
+    write_index(idx, bits, START);
+    SIMDJSON_IF_CONSTEXPR (N > 1) {
+      write_indexes<(N-1>0?START+1:START), (N-1>=0?N-1:1)>(idx, bits);
+    }
+    return START+N;
+  }
+
+  template <int START, int END, int STEP>
+  simdjson_inline int write_indexes_stepped(uint32_t idx, uint64_t& bits, int cnt) {
+    write_indexes<START, STEP>(idx, bits);
+    SIMDJSON_IF_CONSTEXPR ((START+STEP)  < END) {
+      if (simdjson_unlikely((START+STEP) < cnt)) {
+        write_indexes_stepped<(START+STEP<END?START+STEP:END), END, STEP>(idx, bits, cnt);
+      }
+    }
+    return ((END-START) % STEP) == 0 ? END : (END-START) - ((END-START) % STEP) + STEP;
+  }
+
+  // flatten out values in 'bits' assuming that they are are to have values of idx
+  // plus their position in the bitvector, and store these indexes at
+  // base_ptr[base] incrementing base as we go
+  // will potentially store extra values beyond end of valid bits, so base_ptr
+  // needs to be large enough to handle this
+  //
+  // If the kernel sets SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER, then it
+  // will provide its own version of the code.
+#ifdef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+  simdjson_inline void write(uint32_t idx, uint64_t bits);
+#else
+  simdjson_inline void write(uint32_t idx, uint64_t bits) {
+    // In some instances, the next branch is expensive because it is mispredicted.
+    // Unfortunately, in other cases,
+    // it helps tremendously.
+    if (bits == 0)
+        return;
+
+    int cnt = static_cast<int>(count_ones(bits));
+
+#if SIMDJSON_PREFER_REVERSE_BITS
+    bits = reverse_bits(bits);
+#endif
+#ifdef SIMDJSON_STRUCTURAL_INDEXER_STEP
+    static constexpr const int STEP = SIMDJSON_STRUCTURAL_INDEXER_STEP;
+#else
+    static constexpr const int STEP = 4;
+#endif
+    static constexpr const int STEP_UNTIL = 24;
+
+    write_indexes_stepped<0, STEP_UNTIL, STEP>(idx, bits, cnt);
+    SIMDJSON_IF_CONSTEXPR (STEP_UNTIL < 64) {
+      if (simdjson_unlikely(STEP_UNTIL < cnt)) {
+        for (int i=STEP_UNTIL; i<cnt; i++) {
+          write_index(idx, bits, i);
+        }
+      }
+    }
+
+    this->tail += cnt;
+  }
+#endif // SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+};
+
+class json_structural_indexer {
+public:
+  /**
+   * Find the important bits of JSON in a 128-byte chunk, and add them to structural_indexes.
+   *
+   * @param partial Setting the partial parameter to true allows the find_structural_bits to
+   *   tolerate unclosed strings. The caller should still ensure that the input is valid UTF-8. If
+   *   you are processing substrings, you may want to call on a function like trimmed_length_safe_utf8.
+   */
+  template<size_t STEP_SIZE>
+  static error_code index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept;
+
+private:
+  simdjson_inline json_structural_indexer(uint32_t *structural_indexes);
+  template<size_t STEP_SIZE>
+  simdjson_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
+  simdjson_warn_unused simdjson_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
+
+  json_scanner scanner{};
+  utf8_checker checker{};
+  bit_indexer indexer;
+  uint64_t prev_structurals = 0;
+  uint64_t unescaped_chars_error = 0;
+};
+
+simdjson_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
+
+// Skip the last character if it is partial
+simdjson_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
+  if (simdjson_unlikely(len < 3)) {
+    switch (len) {
+      case 2:
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
+        return len;
+      case 1:
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        return len;
+      case 0:
+        return len;
+    }
+  }
+  if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+  if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 1 byte left
+  if (buf[len-3] >= 0xf0) { return len-3; } // 4-byte characters with only 3 bytes left
+  return len;
+}
+
+//
+// PERF NOTES:
+// We pipe 2 inputs through these stages:
+// 1. Load JSON into registers. This takes a long time and is highly parallelizable, so we load
+//    2 inputs' worth at once so that by the time step 2 is looking for them input, it's available.
+// 2. Scan the JSON for critical data: strings, scalars and operators. This is the critical path.
+//    The output of step 1 depends entirely on this information. These functions don't quite use
+//    up enough CPU: the second half of the functions is highly serial, only using 1 execution core
+//    at a time. The second input's scans has some dependency on the first ones finishing it, but
+//    they can make a lot of progress before they need that information.
+// 3. Step 1 does not use enough capacity, so we run some extra stuff while we're waiting for that
+//    to finish: utf-8 checks and generating the output from the last iteration.
+//
+// The reason we run 2 inputs at a time, is steps 2 and 3 are *still* not enough to soak up all
+// available capacity with just one input. Running 2 at a time seems to give the CPU a good enough
+// workout.
+//
+template<size_t STEP_SIZE>
+error_code json_structural_indexer::index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept {
+  if (simdjson_unlikely(len > parser.capacity())) { return CAPACITY; }
+  // We guard the rest of the code so that we can assume that len > 0 throughout.
+  if (len == 0) { return EMPTY; }
+  if (is_streaming(partial)) {
+    len = trim_partial_utf8(buf, len);
+    // If you end up with an empty window after trimming
+    // the partial UTF-8 bytes, then chances are good that you
+    // have an UTF-8 formatting error.
+    if(len == 0) { return UTF8_ERROR; }
+  }
+  buf_block_reader<STEP_SIZE> reader(buf, len);
+  json_structural_indexer indexer(parser.structural_indexes.get());
+
+  // Read all but the last block
+  while (reader.has_full_block()) {
+    indexer.step<STEP_SIZE>(reader.full_block(), reader);
+  }
+  // Take care of the last block (will always be there unless file is empty which is
+  // not supposed to happen.)
+  uint8_t block[STEP_SIZE];
+  if (simdjson_unlikely(reader.get_remainder(block) == 0)) { return UNEXPECTED_ERROR; }
+  indexer.step<STEP_SIZE>(block, reader);
+  return indexer.finish(parser, reader.block_index(), len, partial);
+}
+
+template<>
+simdjson_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block);
+  simd::simd8x64<uint8_t> in_2(block+64);
+  json_block block_1 = scanner.next(in_1);
+  json_block block_2 = scanner.next(in_2);
+  this->next(in_1, block_1, reader.block_index());
+  this->next(in_2, block_2, reader.block_index()+64);
+  reader.advance();
+}
+
+template<>
+simdjson_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block);
+  json_block block_1 = scanner.next(in_1);
+  this->next(in_1, block_1, reader.block_index());
+  reader.advance();
+}
+
+simdjson_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
+  uint64_t unescaped = in.lteq(0x1F);
+#if SIMDJSON_UTF8VALIDATION
+  checker.check_next_input(in);
+#endif
+  indexer.write(uint32_t(idx-64), prev_structurals); // Output *last* iteration's structurals to the parser
+  prev_structurals = block.structural_start();
+  unescaped_chars_error |= block.non_quote_inside_string(unescaped);
+}
+
+simdjson_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
+  // Write out the final iteration's structurals
+  indexer.write(uint32_t(idx-64), prev_structurals);
+  error_code error = scanner.finish();
+  // We deliberately break down the next expression so that it is
+  // human readable.
+  const bool should_we_exit = is_streaming(partial) ?
+    ((error != SUCCESS) && (error != UNCLOSED_STRING)) // when partial we tolerate UNCLOSED_STRING
+    : (error != SUCCESS); // if partial is false, we must have SUCCESS
+  const bool have_unclosed_string = (error == UNCLOSED_STRING);
+  if (simdjson_unlikely(should_we_exit)) { return error; }
+
+  if (unescaped_chars_error) {
+    return UNESCAPED_CHARS;
+  }
+  parser.n_structural_indexes = uint32_t(indexer.tail - parser.structural_indexes.get());
+  /***
+   * The On-Demand API requires special padding.
+   *
+   * This is related to https://github.com/simdjson/simdjson/issues/906
+   * Basically, we want to make sure that if the parsing continues beyond the last (valid)
+   * structural character, it quickly stops.
+   * Only three structural characters can be repeated without triggering an error in JSON:  [,] and }.
+   * We repeat the padding character (at 'len'). We don't know what it is, but if the parsing
+   * continues, then it must be [,] or }.
+   * Suppose it is ] or }. We backtrack to the first character, what could it be that would
+   * not trigger an error? It could be ] or } but no, because you can't start a document that way.
+   * It can't be a comma, a colon or any simple value. So the only way we could continue is
+   * if the repeated character is [. But if so, the document must start with [. But if the document
+   * starts with [, it should end with ]. If we enforce that rule, then we would get
+   * ][[ which is invalid.
+   *
+   * This is illustrated with the test array_iterate_unclosed_error() on the following input:
+   * R"({ "a": [,,)"
+   **/
+  parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len); // used later in partial == stage1_mode::streaming_final
+  parser.structural_indexes[parser.n_structural_indexes + 1] = uint32_t(len);
+  parser.structural_indexes[parser.n_structural_indexes + 2] = 0;
+  parser.next_structural_index = 0;
+  // a valid JSON file cannot have zero structural indexes - we should have found something
+  if (simdjson_unlikely(parser.n_structural_indexes == 0u)) {
+    return EMPTY;
+  }
+  if (simdjson_unlikely(parser.structural_indexes[parser.n_structural_indexes - 1] > len)) {
+    return UNEXPECTED_ERROR;
+  }
+  if (partial == stage1_mode::streaming_partial) {
+    // If we have an unclosed string, then the last structural
+    // will be the quote and we want to make sure to omit it.
+    if(have_unclosed_string) {
+      parser.n_structural_indexes--;
+      // a valid JSON file cannot have zero structural indexes - we should have found something
+      if (simdjson_unlikely(parser.n_structural_indexes == 0u)) { return CAPACITY; }
+    }
+    // We truncate the input to the end of the last complete document (or zero).
+    auto new_structural_indexes = find_next_document_index(parser);
+    if (new_structural_indexes == 0 && parser.n_structural_indexes > 0) {
+      if(parser.structural_indexes[0] == 0) {
+        // If the buffer is partial and we started at index 0 but the document is
+        // incomplete, it's too big to parse.
+        return CAPACITY;
+      } else {
+        // It is possible that the document could be parsed, we just had a lot
+        // of white space.
+        parser.n_structural_indexes = 0;
+        return EMPTY;
+      }
+    }
+
+    parser.n_structural_indexes = new_structural_indexes;
+  } else if (partial == stage1_mode::streaming_final) {
+    if(have_unclosed_string) { parser.n_structural_indexes--; }
+    // We truncate the input to the end of the last complete document (or zero).
+    // Because partial == stage1_mode::streaming_final, it means that we may
+    // silently ignore trailing garbage. Though it sounds bad, we do it
+    // deliberately because many people who have streams of JSON documents
+    // will truncate them for processing. E.g., imagine that you are uncompressing
+    // the data from a size file or receiving it in chunks from the network. You
+    // may not know where exactly the last document will be. Meanwhile the
+    // document_stream instances allow people to know the JSON documents they are
+    // parsing (see the iterator.source() method).
+    parser.n_structural_indexes = find_next_document_index(parser);
+    // We store the initial n_structural_indexes so that the client can see
+    // whether we used truncation. If initial_n_structural_indexes == parser.n_structural_indexes,
+    // then this will query parser.structural_indexes[parser.n_structural_indexes] which is len,
+    // otherwise, it will copy some prior index.
+    parser.structural_indexes[parser.n_structural_indexes + 1] = parser.structural_indexes[parser.n_structural_indexes];
+    // This next line is critical, do not change it unless you understand what you are
+    // doing.
+    parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len);
+    if (simdjson_unlikely(parser.n_structural_indexes == 0u)) {
+        // We tolerate an unclosed string at the very end of the stream. Indeed, users
+        // often load their data in bulk without being careful and they want us to ignore
+        // the trailing garbage.
+        return EMPTY;
+    }
+  }
+  checker.check_eof();
+  return checker.errors();
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+// Clear CUSTOM_BIT_INDEXER so other implementations can set it if they need to.
+#undef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H
+/* end file generic/stage1/json_structural_indexer.h for ppc64 */
+/* including generic/stage1/utf8_validator.h for ppc64: #include <generic/stage1/utf8_validator.h> */
+/* begin file generic/stage1/utf8_validator.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace stage1 {
+
+/**
+ * Validates that the string is actual UTF-8.
+ */
+template<class checker>
+bool generic_validate_utf8(const uint8_t * input, size_t length) {
+    checker c{};
+    buf_block_reader<64> reader(input, length);
+    while (reader.has_full_block()) {
+      simd::simd8x64<uint8_t> in(reader.full_block());
+      c.check_next_input(in);
+      reader.advance();
+    }
+    uint8_t block[64]{};
+    reader.get_remainder(block);
+    simd::simd8x64<uint8_t> in(block);
+    c.check_next_input(in);
+    reader.advance();
+    c.check_eof();
+    return c.errors() == error_code::SUCCESS;
+}
+
+bool generic_validate_utf8(const char * input, size_t length) {
+    return generic_validate_utf8<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H
+/* end file generic/stage1/utf8_validator.h for ppc64 */
+/* end file generic/stage1/amalgamated.h for ppc64 */
+/* including generic/stage2/amalgamated.h for ppc64: #include <generic/stage2/amalgamated.h> */
+/* begin file generic/stage2/amalgamated.h for ppc64 */
+// Stuff other things depend on
+/* including generic/stage2/base.h for ppc64: #include <generic/stage2/base.h> */
+/* begin file generic/stage2/base.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_BASE_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace stage2 {
+
+class json_iterator;
+class structural_iterator;
+struct tape_builder;
+struct tape_writer;
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_BASE_H
+/* end file generic/stage2/base.h for ppc64 */
+/* including generic/stage2/tape_writer.h for ppc64: #include <generic/stage2/tape_writer.h> */
+/* begin file generic/stage2/tape_writer.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/internal/tape_type.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace stage2 {
+
+struct tape_writer {
+  /** The next place to write to tape */
+  uint64_t *next_tape_loc;
+
+  /** Write a signed 64-bit value to tape. */
+  simdjson_inline void append_s64(int64_t value) noexcept;
+
+  /** Write an unsigned 64-bit value to tape. */
+  simdjson_inline void append_u64(uint64_t value) noexcept;
+
+  /** Write a double value to tape. */
+  simdjson_inline void append_double(double value) noexcept;
+
+  /** Write a big integer (as string) to tape. src points to first digit, len is byte count. */
+  simdjson_inline void append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept;
+
+  /**
+   * Append a tape entry (an 8-bit type,and 56 bits worth of value).
+   */
+  simdjson_inline void append(uint64_t val, internal::tape_type t) noexcept;
+
+  /**
+   * Skip the current tape entry without writing.
+   *
+   * Used to skip the start of the container, since we'll come back later to fill it in when the
+   * container ends.
+   */
+  simdjson_inline void skip() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a large u64 or i64.
+   */
+  simdjson_inline void skip_large_integer() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a double.
+   */
+  simdjson_inline void skip_double() noexcept;
+
+  /**
+   * Write a value to a known location on tape.
+   *
+   * Used to go back and write out the start of a container after the container ends.
+   */
+  simdjson_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
+
+private:
+  /**
+   * Append both the tape entry, and a supplementary value following it. Used for types that need
+   * all 64 bits, such as double and uint64_t.
+   */
+  template<typename T>
+  simdjson_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
+}; // struct tape_writer
+
+simdjson_inline void tape_writer::append_s64(int64_t value) noexcept {
+  append2(0, value, internal::tape_type::INT64);
+}
+
+simdjson_inline void tape_writer::append_u64(uint64_t value) noexcept {
+  append(0, internal::tape_type::UINT64);
+  *next_tape_loc = value;
+  next_tape_loc++;
+}
+
+/** Write a double value to tape. */
+simdjson_inline void tape_writer::append_double(double value) noexcept {
+  append2(0, value, internal::tape_type::DOUBLE);
+}
+
+simdjson_inline void tape_writer::skip() noexcept {
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::skip_large_integer() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::skip_double() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
+  *next_tape_loc = val | ((uint64_t(char(t))) << 56);
+  next_tape_loc++;
+}
+
+template<typename T>
+simdjson_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
+  append(val, t);
+  static_assert(sizeof(val2) == sizeof(*next_tape_loc), "Type is not 64 bits!");
+  memcpy(next_tape_loc, &val2, sizeof(val2));
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
+  tape_loc = val | ((uint64_t(char(t))) << 56);
+}
+
+simdjson_inline void tape_writer::append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept {
+  // Write to string buffer: [4-byte LE length][digits][null]
+  uint32_t str_len = uint32_t(len);
+  memcpy(string_buf, &str_len, sizeof(uint32_t));
+  memcpy(string_buf + sizeof(uint32_t), src, len);
+  string_buf[sizeof(uint32_t) + len] = 0;
+  // Tape entry: offset into string buffer
+  // The caller must set the offset relative to doc.string_buf base
+  append(0, internal::tape_type::BIGINT); // placeholder offset, caller patches
+  string_buf += sizeof(uint32_t) + len + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+/* end file generic/stage2/tape_writer.h for ppc64 */
+/* including generic/stage2/logger.h for ppc64: #include <generic/stage2/logger.h> */
+/* begin file generic/stage2/logger.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+
+// This is for an internal-only stage 2 specific logger.
+// Set LOG_ENABLED = true to log what stage 2 is doing!
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace logger {
+
+  static constexpr const char * DASHES = "----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------";
+
+#if SIMDJSON_VERBOSE_LOGGING
+  static constexpr const bool LOG_ENABLED = true;
+#else
+  static constexpr const bool LOG_ENABLED = false;
+#endif
+  static constexpr const int LOG_EVENT_LEN = 20;
+  static constexpr const int LOG_BUFFER_LEN = 30;
+  static constexpr const int LOG_SMALL_BUFFER_LEN = 10;
+  static constexpr const int LOG_INDEX_LEN = 5;
+
+  static int log_depth; // Not threadsafe. Log only.
+
+  // Helper to turn unprintable or newline characters into spaces
+  static simdjson_inline char printable_char(char c) {
+    if (c >= 0x20) {
+      return c;
+    } else {
+      return ' ';
+    }
+  }
+
+  // Print the header and set up log_start
+  static simdjson_inline void log_start() {
+    if (LOG_ENABLED) {
+      log_depth = 0;
+      printf("\n");
+      printf("| %-*s | %-*s | %-*s | %-*s | Detail |\n", LOG_EVENT_LEN, "Event", LOG_BUFFER_LEN, "Buffer", LOG_SMALL_BUFFER_LEN, "Next", 5, "Next#");
+      printf("|%.*s|%.*s|%.*s|%.*s|--------|\n", LOG_EVENT_LEN+2, DASHES, LOG_BUFFER_LEN+2, DASHES, LOG_SMALL_BUFFER_LEN+2, DASHES, 5+2, DASHES);
+    }
+  }
+
+  simdjson_unused static simdjson_inline void log_string(const char *message) {
+    if (LOG_ENABLED) {
+      printf("%s\n", message);
+    }
+  }
+
+  // Logs a single line from the stage 2 DOM parser
+  template<typename S>
+  static simdjson_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
+    if (LOG_ENABLED) {
+      printf("| %*s%s%-*s ", log_depth*2, "", title_prefix, LOG_EVENT_LEN - log_depth*2 - int(strlen(title_prefix)), title);
+      auto current_index = structurals.at_beginning() ? nullptr : structurals.next_structural-1;
+      auto next_index = structurals.next_structural;
+      auto current = current_index ? &structurals.buf[*current_index] : reinterpret_cast<const uint8_t*>("                                                       ");
+      auto next = &structurals.buf[*next_index];
+      {
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_BUFFER_LEN;i++) {
+          printf("%c", printable_char(current[i]));
+        }
+        printf(" ");
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_SMALL_BUFFER_LEN;i++) {
+          printf("%c", printable_char(next[i]));
+        }
+        printf(" ");
+      }
+      if (current_index) {
+        printf("| %*u ", LOG_INDEX_LEN, *current_index);
+      } else {
+        printf("| %-*s ", LOG_INDEX_LEN, "");
+      }
+      // printf("| %*u ", LOG_INDEX_LEN, structurals.next_tape_index());
+      printf("| %-s ", detail);
+      printf("|\n");
+    }
+  }
+
+} // namespace logger
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+/* end file generic/stage2/logger.h for ppc64 */
+
+// All other declarations
+/* including generic/stage2/json_iterator.h for ppc64: #include <generic/stage2/json_iterator.h> */
+/* begin file generic/stage2/json_iterator.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/logger.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace stage2 {
+
+class json_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+  uint32_t depth{0};
+
+  /**
+   * Walk the JSON document.
+   *
+   * The visitor receives callbacks when values are encountered. All callbacks pass the iterator as
+   * the first parameter; some callbacks have other parameters as well:
+   *
+   * - visit_document_start() - at the beginning.
+   * - visit_document_end() - at the end (if things were successful).
+   *
+   * - visit_array_start() - at the start `[` of a non-empty array.
+   * - visit_array_end() - at the end `]` of a non-empty array.
+   * - visit_empty_array() - when an empty array is encountered.
+   *
+   * - visit_object_end() - at the start `]` of a non-empty object.
+   * - visit_object_start() - at the end `]` of a non-empty object.
+   * - visit_empty_object() - when an empty object is encountered.
+   * - visit_key(const uint8_t *key) - when a key in an object field is encountered. key is
+   *                                   guaranteed to point at the first quote of the string (`"key"`).
+   * - visit_primitive(const uint8_t *value) - when a value is a string, number, boolean or null.
+   * - visit_root_primitive(iter, uint8_t *value) - when the top-level value is a string, number, boolean or null.
+   *
+   * - increment_count(iter) - each time a value is found in an array or object.
+   */
+  template<bool STREAMING, typename V>
+  simdjson_warn_unused simdjson_inline error_code walk_document(V &visitor) noexcept;
+
+  /**
+   * Create an iterator capable of walking a JSON document.
+   *
+   * The document must have already passed through stage 1.
+   */
+  simdjson_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
+
+  /**
+   * Look at the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *peek() const noexcept;
+  /**
+   * Advance to the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *advance() noexcept;
+  /**
+   * Get the remaining length of the document, from the start of the current token.
+   */
+  simdjson_inline size_t remaining_len() const noexcept;
+  /**
+   * Check if we are at the end of the document.
+   *
+   * If this is true, there are no more tokens.
+   */
+  simdjson_inline bool at_eof() const noexcept;
+  /**
+   * Check if we are at the beginning of the document.
+   */
+  simdjson_inline bool at_beginning() const noexcept;
+  simdjson_inline uint8_t last_structural() const noexcept;
+
+  /**
+   * Log that a value has been found.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_value(const char *type) const noexcept;
+  /**
+   * Log the start of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_start_value(const char *type) const noexcept;
+  /**
+   * Log the end of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_end_value(const char *type) const noexcept;
+  /**
+   * Log an error.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_error(const char *error) const noexcept;
+
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
+};
+
+template<bool STREAMING, typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::walk_document(V &visitor) noexcept {
+  logger::log_start();
+
+  //
+  // Start the document
+  //
+  if (at_eof()) { return EMPTY; }
+  log_start_value("document");
+  SIMDJSON_TRY( visitor.visit_document_start(*this) );
+
+  //
+  // Read first value
+  //
+  {
+    auto value = advance();
+
+    // Make sure the outer object or array is closed before continuing; otherwise, there are ways we
+    // could get into memory corruption. See https://github.com/simdjson/simdjson/issues/906
+    if (!STREAMING) {
+      switch (*value) {
+        case '{': if (last_structural() != '}') { log_value("starting brace unmatched"); return TAPE_ERROR; }; break;
+        case '[': if (last_structural() != ']') { log_value("starting bracket unmatched"); return TAPE_ERROR; }; break;
+      }
+    }
+
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_root_primitive(*this, value) ); break;
+    }
+  }
+  goto document_end;
+
+//
+// Object parser states
+//
+object_begin:
+  log_start_value("object");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = false;
+  SIMDJSON_TRY( visitor.visit_object_start(*this) );
+
+  {
+    auto key = advance();
+    if (*key != '"') { log_error("Object does not start with a key"); return TAPE_ERROR; }
+    SIMDJSON_TRY( visitor.increment_count(*this) );
+    SIMDJSON_TRY( visitor.visit_key(*this, key) );
+  }
+
+object_field:
+  if (simdjson_unlikely( *advance() != ':' )) { log_error("Missing colon after key in object"); return TAPE_ERROR; }
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+object_continue:
+  switch (*advance()) {
+    case ',':
+      SIMDJSON_TRY( visitor.increment_count(*this) );
+      {
+        auto key = advance();
+        if (simdjson_unlikely( *key != '"' )) { log_error("Key string missing at beginning of field in object"); return TAPE_ERROR; }
+        SIMDJSON_TRY( visitor.visit_key(*this, key) );
+      }
+      goto object_field;
+    case '}': log_end_value("object"); SIMDJSON_TRY( visitor.visit_object_end(*this) ); goto scope_end;
+    default: log_error("No comma between object fields"); return TAPE_ERROR;
+  }
+
+scope_end:
+  depth--;
+  if (depth == 0) { goto document_end; }
+  if (dom_parser.is_array[depth]) { goto array_continue; }
+  goto object_continue;
+
+//
+// Array parser states
+//
+array_begin:
+  log_start_value("array");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = true;
+  SIMDJSON_TRY( visitor.visit_array_start(*this) );
+  SIMDJSON_TRY( visitor.increment_count(*this) );
+
+array_value:
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+array_continue:
+  switch (*advance()) {
+    case ',': SIMDJSON_TRY( visitor.increment_count(*this) ); goto array_value;
+    case ']': log_end_value("array"); SIMDJSON_TRY( visitor.visit_array_end(*this) ); goto scope_end;
+    default: log_error("Missing comma between array values"); return TAPE_ERROR;
+  }
+
+document_end:
+  log_end_value("document");
+  SIMDJSON_TRY( visitor.visit_document_end(*this) );
+
+  dom_parser.next_structural_index = uint32_t(next_structural - &dom_parser.structural_indexes[0]);
+
+  // If we didn't make it to the end, it's an error
+  if ( !STREAMING && dom_parser.next_structural_index != dom_parser.n_structural_indexes ) {
+    log_error("More than one JSON value at the root of the document, or extra characters at the end of the JSON!");
+    return TAPE_ERROR;
+  }
+
+  return SUCCESS;
+
+} // walk_document()
+
+simdjson_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+  : buf{_dom_parser.buf},
+    next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+    dom_parser{_dom_parser} {
+}
+
+simdjson_inline const uint8_t *json_iterator::peek() const noexcept {
+  return &buf[*(next_structural)];
+}
+simdjson_inline const uint8_t *json_iterator::advance() noexcept {
+  return &buf[*(next_structural++)];
+}
+simdjson_inline size_t json_iterator::remaining_len() const noexcept {
+  return dom_parser.len - *(next_structural-1);
+}
+
+simdjson_inline bool json_iterator::at_eof() const noexcept {
+  return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+}
+simdjson_inline bool json_iterator::at_beginning() const noexcept {
+  return next_structural == dom_parser.structural_indexes.get();
+}
+simdjson_inline uint8_t json_iterator::last_structural() const noexcept {
+  return buf[dom_parser.structural_indexes[dom_parser.n_structural_indexes - 1]];
+}
+
+simdjson_inline void json_iterator::log_value(const char *type) const noexcept {
+  logger::log_line(*this, "", type, "");
+}
+
+simdjson_inline void json_iterator::log_start_value(const char *type) const noexcept {
+  logger::log_line(*this, "+", type, "");
+  if (logger::LOG_ENABLED) { logger::log_depth++; }
+}
+
+simdjson_inline void json_iterator::log_end_value(const char *type) const noexcept {
+  if (logger::LOG_ENABLED) { logger::log_depth--; }
+  logger::log_line(*this, "-", type, "");
+}
+
+simdjson_inline void json_iterator::log_error(const char *error) const noexcept {
+  logger::log_line(*this, "", "ERROR", error);
+}
+
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
+  switch (*value) {
+    case '"': return visitor.visit_root_string(*this, value);
+    case 't': return visitor.visit_root_true_atom(*this, value);
+    case 'f': return visitor.visit_root_false_atom(*this, value);
+    case 'n': return visitor.visit_root_null_atom(*this, value);
+    case '-':
+    case '0': case '1': case '2': case '3': case '4':
+    case '5': case '6': case '7': case '8': case '9':
+      return visitor.visit_root_number(*this, value);
+    default:
+      log_error("Document starts with a non-value character");
+      return TAPE_ERROR;
+  }
+}
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
+  // Use the fact that most scalars are going to be either strings or numbers.
+  if(*value == '"') {
+    return visitor.visit_string(*this, value);
+  } else if (((*value - '0')  < 10) || (*value == '-')) {
+    return visitor.visit_number(*this, value);
+  }
+  // true, false, null are uncommon.
+  switch (*value) {
+    case 't': return visitor.visit_true_atom(*this, value);
+    case 'f': return visitor.visit_false_atom(*this, value);
+    case 'n': return visitor.visit_null_atom(*this, value);
+    default:
+      log_error("Non-value found when value was expected!");
+      return TAPE_ERROR;
+  }
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+/* end file generic/stage2/json_iterator.h for ppc64 */
+/* including generic/stage2/stringparsing.h for ppc64: #include <generic/stage2/stringparsing.h> */
+/* begin file generic/stage2/stringparsing.h for ppc64 */
+#include <cstdint>
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/jsoncharutils.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses
+// It is intended to be included multiple times and compiled multiple times
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+/// @private
+namespace stringparsing {
+
+// begin copypasta
+// These chars yield themselves: " \ /
+// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
+// u not handled in this table as it's complex
+static const uint8_t escape_map[256] = {
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
+    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
+    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+};
+
+// handle a unicode codepoint
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr, bool allow_replacement) {
+  // Use the default Unicode Character 'REPLACEMENT CHARACTER' (U+FFFD)
+  constexpr uint32_t substitution_code_point = 0xfffd;
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) != ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+    } else {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+
+      // We have already checked that the high surrogate is valid and
+      // (code_point - 0xd800) < 1024.
+      //
+      // Check that code_point_2 is in the range 0xdc00..0xdfff
+      // and that code_point_2 was parsed from valid hex.
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if (low_bit >> 10) {
+        if(!allow_replacement) { return false; }
+        code_point = substitution_code_point;
+      } else {
+        code_point =  (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+
+    }
+  } else if (code_point >= 0xdc00 && code_point <= 0xdfff) {
+      // If we encounter a low surrogate (not preceded by a high surrogate)
+      // then we have an error.
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+  }
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+// handle a unicode codepoint using the wobbly convention
+// https://simonsapin.github.io/wtf-8/
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint_wobbly(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr) {
+  // It is not ideal that this function is nearly identical to handle_unicode_codepoint.
+  //
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) == ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if ((low_bit >> 10) ==  0) {
+        code_point =
+          (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+    }
+  }
+
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+/**
+ * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+ * must be an unescaped quote terminating the string. It returns the final output
+ * position as pointer. In case of error (e.g., the string has bad escaped codes),
+ * then null_ptr is returned. It is assumed that the output buffer is large
+ * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+ * SIMDJSON_PADDING bytes.
+ */
+simdjson_warn_unused simdjson_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) {
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint(&src, &dst, allow_replacement)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+simdjson_warn_unused simdjson_inline uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) {
+  // It is not ideal that this function is nearly identical to parse_string.
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint_wobbly(&src, &dst)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+} // namespace stringparsing
+
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+/* end file generic/stage2/stringparsing.h for ppc64 */
+/* including generic/stage2/structural_iterator.h for ppc64: #include <generic/stage2/structural_iterator.h> */
+/* begin file generic/stage2/structural_iterator.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace stage2 {
+
+class structural_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+
+  // Start a structural
+  simdjson_inline structural_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+    : buf{_dom_parser.buf},
+      next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+      dom_parser{_dom_parser} {
+  }
+  // Get the buffer position of the current structural character
+  simdjson_inline const uint8_t* current() {
+    return &buf[*(next_structural-1)];
+  }
+  // Get the current structural character
+  simdjson_inline char current_char() {
+    return buf[*(next_structural-1)];
+  }
+  // Get the next structural character without advancing
+  simdjson_inline char peek_next_char() {
+    return buf[*next_structural];
+  }
+  simdjson_inline const uint8_t* peek() {
+    return &buf[*next_structural];
+  }
+  simdjson_inline const uint8_t* advance() {
+    return &buf[*(next_structural++)];
+  }
+  simdjson_inline char advance_char() {
+    return buf[*(next_structural++)];
+  }
+  simdjson_inline size_t remaining_len() {
+    return dom_parser.len - *(next_structural-1);
+  }
+
+  simdjson_inline bool at_end() {
+    return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+  }
+  simdjson_inline bool at_beginning() {
+    return next_structural == dom_parser.structural_indexes.get();
+  }
+};
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H
+/* end file generic/stage2/structural_iterator.h for ppc64 */
+/* including generic/stage2/tape_builder.h for ppc64: #include <generic/stage2/tape_builder.h> */
+/* begin file generic/stage2/tape_builder.h for ppc64 */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/json_iterator.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/stringparsing.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/tape_writer.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/dom/document.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/atomparsing.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/numberparsing.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+
+namespace simdjson {
+namespace ppc64 {
+namespace {
+namespace stage2 {
+
+struct tape_builder {
+  template<bool STREAMING>
+  simdjson_warn_unused static simdjson_inline error_code parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept;
+
+  /** Called when a non-empty document starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty document ends without error. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_end(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty array starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty array ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_end(json_iterator &iter) noexcept;
+  /** Called when an empty array is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_array(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty object starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_start(json_iterator &iter) noexcept;
+  /**
+   * Called when a key in a field is encountered.
+   *
+   * primitive, visit_object_start, visit_empty_object, visit_array_start, or visit_empty_array
+   * will be called after this with the field value.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
+  /** Called when a non-empty object ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_end(json_iterator &iter) noexcept;
+  /** Called when an empty object is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_object(json_iterator &iter) noexcept;
+
+  /**
+   * Called when a string, number, boolean or null is found.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  /**
+   * Called when a string, number, boolean or null is found at the top level of a document (i.e.
+   * when there is no array or object and the entire document is a single string, number, boolean or
+   * null.
+   *
+   * This is separate from primitive() because simdjson's normal primitive parsing routines assume
+   * there is at least one more token after the value, which is only true in an array or object.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  /** Called each time a new field or element in an array or object is found. */
+  simdjson_warn_unused simdjson_inline error_code increment_count(json_iterator &iter) noexcept;
+
+  /** Next location to write to tape */
+  tape_writer tape;
+private:
+  /** Next write location in the string buf for stage 2 parsing */
+  uint8_t *current_string_buf_loc;
+
+  simdjson_inline tape_builder(dom::document &doc) noexcept;
+
+  simdjson_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
+  simdjson_inline void start_container(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_warn_unused simdjson_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
+  simdjson_inline void on_end_string(uint8_t *dst) noexcept;
+}; // struct tape_builder
+
+template<bool STREAMING>
+simdjson_warn_unused simdjson_inline error_code tape_builder::parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept {
+  dom_parser.doc = &doc;
+  json_iterator iter(dom_parser, STREAMING ? dom_parser.next_structural_index : 0);
+  tape_builder builder(doc);
+  return iter.walk_document<STREAMING>(builder);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_root_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
+  constexpr uint32_t start_tape_index = 0;
+  tape.append(start_tape_index, internal::tape_type::ROOT);
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter), internal::tape_type::ROOT);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
+  return visit_string(iter, key, true);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].count++; // we have a key value pair in the object at parser.dom_parser.depth - 1
+  return SUCCESS;
+}
+
+simdjson_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
+  iter.log_value(key ? "key" : "string");
+  uint8_t *dst = on_start_string(iter);
+  dst = stringparsing::parse_string(value+1, dst, false); // We do not allow replacement when the escape characters are invalid.
+  if (dst == nullptr) {
+    iter.log_error("Invalid escape in string");
+    return STRING_ERROR;
+  }
+  on_end_string(dst);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
+  return visit_string(iter, value);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("number");
+  error_code err = numberparsing::parse_number(value, tape);
+  if (simdjson_unlikely(err == BIGINT_ERROR &&
+      iter.dom_parser._number_as_string)) {
+    // Write big integer to string buffer using the same format as strings.
+    // Scan digits the same way parse_number does (skip optional '-', then digits).
+    const uint8_t *p = value;
+    if (*p == '-') p++;
+    while (numberparsing::is_digit(*p)) p++;
+    size_t len = size_t(p - value);
+    tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::BIGINT);
+    uint8_t *dst = current_string_buf_loc + sizeof(uint32_t);
+    memcpy(dst, value, len);
+    dst += len;
+    on_end_string(dst);
+    return SUCCESS;
+  }
+  return err;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
+  //
+  // We need to make a copy to make sure that the string is space terminated.
+  // This is not about padding the input, which should already padded up
+  // to len + SIMDJSON_PADDING. However, we have no control at this stage
+  // on how the padding was done. What if the input string was padded with nulls?
+  // It is quite common for an input string to have an extra null character (C string).
+  // We do not want to allow 9\0 (where \0 is the null character) inside a JSON
+  // document, but the string "9\0" by itself is fine. So we make a copy and
+  // pad the input with spaces when we know that there is just one input element.
+  // This copy is relatively expensive, but it will almost never be called in
+  // practice unless you are in the strange scenario where you have many JSON
+  // documents made of single atoms.
+  //
+  std::unique_ptr<uint8_t[]>copy(new (std::nothrow) uint8_t[iter.remaining_len() + SIMDJSON_PADDING]);
+  if (copy.get() == nullptr) { return MEMALLOC; }
+  std::memcpy(copy.get(), value, iter.remaining_len());
+  std::memset(copy.get() + iter.remaining_len(), ' ', SIMDJSON_PADDING);
+  error_code error = visit_number(iter, copy.get());
+  return error;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value)) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value, iter.remaining_len())) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value)) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value, iter.remaining_len())) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value)) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value, iter.remaining_len())) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+// private:
+
+simdjson_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
+  return uint32_t(tape.next_tape_loc - iter.dom_parser.doc->tape.get());
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  auto start_index = next_tape_index(iter);
+  tape.append(start_index+2, start);
+  tape.append(start_index, end);
+  return SUCCESS;
+}
+
+simdjson_inline void tape_builder::start_container(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].tape_index = next_tape_index(iter);
+  iter.dom_parser.open_containers[iter.depth].count = 0;
+  tape.skip(); // We don't actually *write* the start element until the end.
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  // Write the ending tape element, pointing at the start location
+  const uint32_t start_tape_index = iter.dom_parser.open_containers[iter.depth].tape_index;
+  tape.append(start_tape_index, end);
+  // Write the start tape element, pointing at the end location (and including count)
+  // count can overflow if it exceeds 24 bits... so we saturate
+  // the convention being that a cnt of 0xffffff or more is undetermined in value (>=  0xffffff).
+  const uint32_t count = iter.dom_parser.open_containers[iter.depth].count;
+  const uint32_t cntsat = count > 0xFFFFFF ? 0xFFFFFF : count;
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter) | (uint64_t(cntsat) << 32), start);
+  return SUCCESS;
+}
+
+simdjson_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
+  // we advance the point, accounting for the fact that we have a NULL termination
+  tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::STRING);
+  return current_string_buf_loc + sizeof(uint32_t);
+}
+
+simdjson_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
+  uint32_t str_length = uint32_t(dst - (current_string_buf_loc + sizeof(uint32_t)));
+  // TODO check for overflow in case someone has a crazy string (>=4GB?)
+  // But only add the overflow check when the document itself exceeds 4GB
+  // Currently unneeded because we refuse to parse docs larger or equal to 4GB.
+  memcpy(current_string_buf_loc, &str_length, sizeof(uint32_t));
+  // NULL termination is still handy if you expect all your strings to
+  // be NULL terminated? It comes at a small cost
+  *dst = 0;
+  current_string_buf_loc = dst + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+/* end file generic/stage2/tape_builder.h for ppc64 */
+/* end file generic/stage2/amalgamated.h for ppc64 */
+
+//
+// Stage 1
+//
+namespace simdjson {
+namespace ppc64 {
+
+simdjson_warn_unused error_code implementation::create_dom_parser_implementation(
+  size_t capacity,
+  size_t max_depth,
+  std::unique_ptr<internal::dom_parser_implementation>& dst
+) const noexcept {
+  dst.reset( new (std::nothrow) dom_parser_implementation() );
+  if (!dst) { return MEMALLOC; }
+  if (auto err = dst->set_capacity(capacity))
+    return err;
+  if (auto err = dst->set_max_depth(max_depth))
+    return err;
+  return SUCCESS;
+}
+
+namespace {
+
+using namespace simd;
+
+simdjson_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
+  const simd8<uint8_t> table1(16, 0, 0, 0, 0, 0, 0, 0, 0, 8, 12, 1, 2, 9, 0, 0);
+  const simd8<uint8_t> table2(8, 0, 18, 4, 0, 1, 0, 1, 0, 0, 0, 3, 2, 1, 0, 0);
+
+  simd8x64<uint8_t> v(
+     (in.chunks[0] & 0xf).lookup_16(table1) & (in.chunks[0].shr<4>()).lookup_16(table2),
+     (in.chunks[1] & 0xf).lookup_16(table1) & (in.chunks[1].shr<4>()).lookup_16(table2),
+     (in.chunks[2] & 0xf).lookup_16(table1) & (in.chunks[2].shr<4>()).lookup_16(table2),
+     (in.chunks[3] & 0xf).lookup_16(table1) & (in.chunks[3].shr<4>()).lookup_16(table2)
+  );
+
+  uint64_t op = simd8x64<bool>(
+        v.chunks[0].any_bits_set(0x7),
+        v.chunks[1].any_bits_set(0x7),
+        v.chunks[2].any_bits_set(0x7),
+        v.chunks[3].any_bits_set(0x7)
+  ).to_bitmask();
+
+  uint64_t whitespace = simd8x64<bool>(
+        v.chunks[0].any_bits_set(0x18),
+        v.chunks[1].any_bits_set(0x18),
+        v.chunks[2].any_bits_set(0x18),
+        v.chunks[3].any_bits_set(0x18)
+  ).to_bitmask();
+
+  return { whitespace, op };
+}
+
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input) {
+  // careful: 0x80 is not ascii.
+  return input.reduce_or().saturating_sub(0x7fu).bits_not_set_anywhere();
+}
+
+simdjson_unused simdjson_inline simd8<bool> must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_second_byte = prev1.saturating_sub(0xc0u-1); // Only 11______ will be > 0
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-1); // Only 111_____ will be > 0
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-1); // Only 1111____ will be > 0
+  // Caller requires a bool (all 1's). All values resulting from the subtraction will be <= 64, so signed comparison is fine.
+  return simd8<int8_t>(is_second_byte | is_third_byte | is_fourth_byte) > int8_t(0);
+}
+
+simdjson_inline simd8<uint8_t> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-0x80); // Only 111_____ will be >= 0x80
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-0x80); // Only 1111____ will be >= 0x80
+  return is_third_byte | is_fourth_byte;
+}
+
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+
+//
+// Stage 2
+//
+
+//
+// Implementation-specific overrides
+//
+namespace simdjson {
+namespace ppc64 {
+
+simdjson_warn_unused error_code implementation::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept {
+  return ppc64::stage1::json_minifier::minify<64>(buf, len, dst, dst_len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage1(const uint8_t *_buf, size_t _len, stage1_mode streaming) noexcept {
+  this->buf = _buf;
+  this->len = _len;
+  return ppc64::stage1::json_structural_indexer::index<64>(buf, len, *this, streaming);
+}
+
+simdjson_warn_unused bool implementation::validate_utf8(const char *buf, size_t len) const noexcept {
+  return ppc64::stage1::generic_validate_utf8(buf,len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<false>(*this, _doc);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2_next(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<true>(*this, _doc);
+}
+
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_string(const uint8_t *src, uint8_t *dst, bool replacement_char) const noexcept {
+  return ppc64::stringparsing::parse_string(src, dst, replacement_char);
+}
+
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept {
+  return ppc64::stringparsing::parse_wobbly_string(src, dst);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::parse(const uint8_t *_buf, size_t _len, dom::document &_doc) noexcept {
+  auto error = stage1(_buf, _len, stage1_mode::regular);
+  if (error) { return error; }
+  return stage2(_doc);
+}
+
+} // namespace ppc64
+} // namespace simdjson
+
+/* including simdjson/ppc64/end.h: #include <simdjson/ppc64/end.h> */
+/* begin file simdjson/ppc64/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#undef SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+/* undefining SIMDJSON_IMPLEMENTATION from "ppc64" */
+#undef SIMDJSON_IMPLEMENTATION
+/* end file simdjson/ppc64/end.h */
+
+#endif // SIMDJSON_SRC_PPC64_CPP
+/* end file ppc64.cpp */
+#endif
+#if SIMDJSON_IMPLEMENTATION_WESTMERE
+/* including westmere.cpp: #include <westmere.cpp> */
+/* begin file westmere.cpp */
+#ifndef SIMDJSON_SRC_WESTMERE_CPP
+#define SIMDJSON_SRC_WESTMERE_CPP
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+/* including simdjson/westmere.h: #include <simdjson/westmere.h> */
+/* begin file simdjson/westmere.h */
+#ifndef SIMDJSON_WESTMERE_H
+#define SIMDJSON_WESTMERE_H
+
+/* including simdjson/westmere/begin.h: #include "simdjson/westmere/begin.h" */
+/* begin file simdjson/westmere/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "westmere" */
+#define SIMDJSON_IMPLEMENTATION westmere
+/* including simdjson/westmere/base.h: #include "simdjson/westmere/base.h" */
+/* begin file simdjson/westmere/base.h */
+#ifndef SIMDJSON_WESTMERE_BASE_H
+#define SIMDJSON_WESTMERE_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// The constructor may be executed on any host, so we take care not to use SIMDJSON_TARGET_WESTMERE
+namespace simdjson {
+/**
+ * Implementation for Westmere (Intel SSE4.2).
+ */
+namespace westmere {
+
+class implementation;
+
+namespace {
+namespace simd {
+
+template <typename T> struct simd8;
+template <typename T> struct simd8x64;
+
+} // namespace simd
+} // unnamed namespace
+
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_BASE_H
+/* end file simdjson/westmere/base.h */
+/* including simdjson/westmere/intrinsics.h: #include "simdjson/westmere/intrinsics.h" */
+/* begin file simdjson/westmere/intrinsics.h */
+#ifndef SIMDJSON_WESTMERE_INTRINSICS_H
+#define SIMDJSON_WESTMERE_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#if SIMDJSON_VISUAL_STUDIO
+// under clang within visual studio, this will include <x86intrin.h>
+#include <intrin.h> // visual studio or clang
+#else
+#include <x86intrin.h> // elsewhere
+#endif // SIMDJSON_VISUAL_STUDIO
+
+
+#if SIMDJSON_CLANG_VISUAL_STUDIO
+/**
+ * You are not supposed, normally, to include these
+ * headers directly. Instead you should either include intrin.h
+ * or x86intrin.h. However, when compiling with clang
+ * under Windows (i.e., when _MSC_VER is set), these headers
+ * only get included *if* the corresponding features are detected
+ * from macros:
+ */
+#include <smmintrin.h>  // for _mm_alignr_epi8
+#include <wmmintrin.h>  // for  _mm_clmulepi64_si128
+#endif
+
+static_assert(sizeof(__m128i) <= simdjson::SIMDJSON_PADDING, "insufficient padding for westmere");
+
+#endif // SIMDJSON_WESTMERE_INTRINSICS_H
+/* end file simdjson/westmere/intrinsics.h */
+
+#if !SIMDJSON_CAN_ALWAYS_RUN_WESTMERE
+SIMDJSON_TARGET_REGION("sse4.2,pclmul,popcnt")
+#endif
+
+/* including simdjson/westmere/bitmanipulation.h: #include "simdjson/westmere/bitmanipulation.h" */
+/* begin file simdjson/westmere/bitmanipulation.h */
+#ifndef SIMDJSON_WESTMERE_BITMANIPULATION_H
+#define SIMDJSON_WESTMERE_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  unsigned long ret;
+  // Search the mask data from least significant bit (LSB)
+  // to the most significant bit (MSB) for a set bit (1).
+  _BitScanForward64(&ret, input_num);
+  return (int)ret;
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO
+  return __builtin_ctzll(input_num);
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return input_num & (input_num-1);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  unsigned long leading_zero = 0;
+  // Search the mask data from most significant bit (MSB)
+  // to least significant bit (LSB) for a set bit (1).
+  if (_BitScanReverse64(&leading_zero, input_num))
+    return (int)(63 - leading_zero);
+  else
+    return 64;
+#else
+  return __builtin_clzll(input_num);
+#endif// SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+simdjson_inline unsigned __int64 count_ones(uint64_t input_num) {
+  // note: we do not support legacy 32-bit Windows in this kernel
+  return __popcnt64(input_num);// Visual Studio wants two underscores
+}
+#else
+simdjson_inline long long int count_ones(uint64_t input_num) {
+  return _popcnt64(input_num);
+}
+#endif
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2,
+                                uint64_t *result) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  return _addcarry_u64(0, value1, value2,
+                       reinterpret_cast<unsigned __int64 *>(result));
+#else
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+#endif
+}
+
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_BITMANIPULATION_H
+/* end file simdjson/westmere/bitmanipulation.h */
+/* including simdjson/westmere/bitmask.h: #include "simdjson/westmere/bitmask.h" */
+/* begin file simdjson/westmere/bitmask.h */
+#ifndef SIMDJSON_WESTMERE_BITMASK_H
+#define SIMDJSON_WESTMERE_BITMASK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(const uint64_t bitmask) {
+  // There should be no such thing with a processing supporting avx2
+  // but not clmul.
+  __m128i all_ones = _mm_set1_epi8('\xFF');
+  __m128i result = _mm_clmulepi64_si128(_mm_set_epi64x(0ULL, bitmask), all_ones, 0);
+  return _mm_cvtsi128_si64(result);
+}
+
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_BITMASK_H
+/* end file simdjson/westmere/bitmask.h */
+/* including simdjson/westmere/numberparsing_defs.h: #include "simdjson/westmere/numberparsing_defs.h" */
+/* begin file simdjson/westmere/numberparsing_defs.h */
+#ifndef SIMDJSON_WESTMERE_NUMBERPARSING_DEFS_H
+#define SIMDJSON_WESTMERE_NUMBERPARSING_DEFS_H
+
+/* including simdjson/westmere/base.h: #include "simdjson/westmere/base.h" */
+/* begin file simdjson/westmere/base.h */
+#ifndef SIMDJSON_WESTMERE_BASE_H
+#define SIMDJSON_WESTMERE_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// The constructor may be executed on any host, so we take care not to use SIMDJSON_TARGET_WESTMERE
+namespace simdjson {
+/**
+ * Implementation for Westmere (Intel SSE4.2).
+ */
+namespace westmere {
+
+class implementation;
+
+namespace {
+namespace simd {
+
+template <typename T> struct simd8;
+template <typename T> struct simd8x64;
+
+} // namespace simd
+} // unnamed namespace
+
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_BASE_H
+/* end file simdjson/westmere/base.h */
+/* including simdjson/westmere/intrinsics.h: #include "simdjson/westmere/intrinsics.h" */
+/* begin file simdjson/westmere/intrinsics.h */
+#ifndef SIMDJSON_WESTMERE_INTRINSICS_H
+#define SIMDJSON_WESTMERE_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#if SIMDJSON_VISUAL_STUDIO
+// under clang within visual studio, this will include <x86intrin.h>
+#include <intrin.h> // visual studio or clang
+#else
+#include <x86intrin.h> // elsewhere
+#endif // SIMDJSON_VISUAL_STUDIO
+
+
+#if SIMDJSON_CLANG_VISUAL_STUDIO
+/**
+ * You are not supposed, normally, to include these
+ * headers directly. Instead you should either include intrin.h
+ * or x86intrin.h. However, when compiling with clang
+ * under Windows (i.e., when _MSC_VER is set), these headers
+ * only get included *if* the corresponding features are detected
+ * from macros:
+ */
+#include <smmintrin.h>  // for _mm_alignr_epi8
+#include <wmmintrin.h>  // for  _mm_clmulepi64_si128
+#endif
+
+static_assert(sizeof(__m128i) <= simdjson::SIMDJSON_PADDING, "insufficient padding for westmere");
+
+#endif // SIMDJSON_WESTMERE_INTRINSICS_H
+/* end file simdjson/westmere/intrinsics.h */
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace numberparsing {
+
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  // this actually computes *16* values so we are being wasteful.
+  const __m128i ascii0 = _mm_set1_epi8('0');
+  const __m128i mul_1_10 =
+      _mm_setr_epi8(10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1);
+  const __m128i mul_1_100 = _mm_setr_epi16(100, 1, 100, 1, 100, 1, 100, 1);
+  const __m128i mul_1_10000 =
+      _mm_setr_epi16(10000, 1, 10000, 1, 10000, 1, 10000, 1);
+  const __m128i input = _mm_sub_epi8(
+      _mm_loadu_si128(reinterpret_cast<const __m128i *>(chars)), ascii0);
+  const __m128i t1 = _mm_maddubs_epi16(input, mul_1_10);
+  const __m128i t2 = _mm_madd_epi16(t1, mul_1_100);
+  const __m128i t3 = _mm_packus_epi32(t2, t2);
+  const __m128i t4 = _mm_madd_epi16(t3, mul_1_10000);
+  return _mm_cvtsi128_si32(
+      t4); // only captures the sum of the first 8 digits, drop the rest
+}
+
+/** @private */
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+#if SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+#if SIMDJSON_IS_ARM64
+  // ARM64 has native support for 64-bit multiplications, no need to emultate
+  answer.high = __umulh(value1, value2);
+  answer.low = value1 * value2;
+#else
+  answer.low = _umul128(value1, value2, &answer.high); // _umul128 not available on ARM64
+#endif // SIMDJSON_IS_ARM64
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+#endif
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace westmere
+} // namespace simdjson
+
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+
+#endif //  SIMDJSON_WESTMERE_NUMBERPARSING_DEFS_H
+/* end file simdjson/westmere/numberparsing_defs.h */
+/* including simdjson/westmere/simd.h: #include "simdjson/westmere/simd.h" */
+/* begin file simdjson/westmere/simd.h */
+#ifndef SIMDJSON_WESTMERE_SIMD_H
+#define SIMDJSON_WESTMERE_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace simd {
+
+  template<typename Child>
+  struct base {
+    __m128i value;
+
+    // Zero constructor
+    simdjson_inline base() : value{__m128i()} {}
+
+    // Conversion from SIMD register
+    simdjson_inline base(const __m128i _value) : value(_value) {}
+
+    // Conversion to SIMD register
+    simdjson_inline operator const __m128i&() const { return this->value; }
+    simdjson_inline operator __m128i&() { return this->value; }
+
+    // Bit operations
+    simdjson_inline Child operator|(const Child other) const { return _mm_or_si128(*this, other); }
+    simdjson_inline Child operator&(const Child other) const { return _mm_and_si128(*this, other); }
+    simdjson_inline Child operator^(const Child other) const { return _mm_xor_si128(*this, other); }
+    simdjson_inline Child bit_andnot(const Child other) const { return _mm_andnot_si128(other, *this); }
+    simdjson_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+  };
+
+  template<typename T, typename Mask=simd8<bool>>
+  struct base8: base<simd8<T>> {
+    typedef uint16_t bitmask_t;
+    typedef uint32_t bitmask2_t;
+
+    simdjson_inline base8() : base<simd8<T>>() {}
+    simdjson_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}
+
+    friend simdjson_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return _mm_cmpeq_epi8(lhs, rhs); }
+
+    static const int SIZE = sizeof(base<simd8<T>>::value);
+
+    template<int N=1>
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+      return _mm_alignr_epi8(*this, prev_chunk, 16 - N);
+    }
+  };
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool>: base8<bool> {
+    static simdjson_inline simd8<bool> splat(bool _value) { return _mm_set1_epi8(uint8_t(-(!!_value))); }
+
+    simdjson_inline simd8() : base8() {}
+    simdjson_inline simd8(const __m128i _value) : base8<bool>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+
+    simdjson_inline int to_bitmask() const { return _mm_movemask_epi8(*this); }
+    simdjson_inline bool any() const { return !_mm_testz_si128(*this, *this); }
+    simdjson_inline simd8<bool> operator~() const { return *this ^ true; }
+  };
+
+  template<typename T>
+  struct base8_numeric: base8<T> {
+    static simdjson_inline simd8<T> splat(T _value) { return _mm_set1_epi8(_value); }
+    static simdjson_inline simd8<T> zero() { return _mm_setzero_si128(); }
+    static simdjson_inline simd8<T> load(const T values[16]) {
+      return _mm_loadu_si128(reinterpret_cast<const __m128i *>(values));
+    }
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    static simdjson_inline simd8<T> repeat_16(
+      T v0,  T v1,  T v2,  T v3,  T v4,  T v5,  T v6,  T v7,
+      T v8,  T v9,  T v10, T v11, T v12, T v13, T v14, T v15
+    ) {
+      return simd8<T>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    simdjson_inline base8_numeric() : base8<T>() {}
+    simdjson_inline base8_numeric(const __m128i _value) : base8<T>(_value) {}
+
+    // Store to array
+    simdjson_inline void store(T dst[16]) const { return _mm_storeu_si128(reinterpret_cast<__m128i *>(dst), *this); }
+
+    // Override to distinguish from bool version
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<T> operator+(const simd8<T> other) const { return _mm_add_epi8(*this, other); }
+    simdjson_inline simd8<T> operator-(const simd8<T> other) const { return _mm_sub_epi8(*this, other); }
+    simdjson_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return _mm_shuffle_epi8(lookup_table, *this);
+    }
+
+    // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted as a bitset).
+    // Passing a 0 value for mask would be equivalent to writing out every byte to output.
+    // Only the first 16 - count_ones(mask) bytes of the result are significant but 16 bytes
+    // get written.
+    // Design consideration: it seems like a function with the
+    // signature simd8<L> compress(uint32_t mask) would be
+    // sensible, but the AVX ISA makes this kind of approach difficult.
+    template<typename L>
+    simdjson_inline void compress(uint16_t mask, L * output) const {
+      using internal::thintable_epi8;
+      using internal::BitsSetTable256mul2;
+      using internal::pshufb_combine_table;
+      // this particular implementation was inspired by work done by @animetosho
+      // we do it in two steps, first 8 bytes and then second 8 bytes
+      uint8_t mask1 = uint8_t(mask); // least significant 8 bits
+      uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
+      // next line just loads the 64-bit values thintable_epi8[mask1] and
+      // thintable_epi8[mask2] into a 128-bit register, using only
+      // two instructions on most compilers.
+      __m128i shufmask =  _mm_set_epi64x(thintable_epi8[mask2], thintable_epi8[mask1]);
+      // we increment by 0x08 the second half of the mask
+      shufmask =
+      _mm_add_epi8(shufmask, _mm_set_epi32(0x08080808, 0x08080808, 0, 0));
+      // this is the version "nearly pruned"
+      __m128i pruned = _mm_shuffle_epi8(*this, shufmask);
+      // we still need to put the two halves together.
+      // we compute the popcount of the first half:
+      int pop1 = BitsSetTable256mul2[mask1];
+      // then load the corresponding mask, what it does is to write
+      // only the first pop1 bytes from the first 8 bytes, and then
+      // it fills in with the bytes from the second 8 bytes + some filling
+      // at the end.
+      __m128i compactmask =
+      _mm_loadu_si128(reinterpret_cast<const __m128i *>(pshufb_combine_table + pop1 * 8));
+      __m128i answer = _mm_shuffle_epi8(pruned, compactmask);
+      _mm_storeu_si128(reinterpret_cast<__m128i *>(output), answer);
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> : base8_numeric<int8_t> {
+    simdjson_inline simd8() : base8_numeric<int8_t>() {}
+    simdjson_inline simd8(const __m128i _value) : base8_numeric<int8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const int8_t* values) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) : simd8(_mm_setr_epi8(
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    )) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm_max_epi8(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm_min_epi8(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm_cmpgt_epi8(*this, other); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return _mm_cmpgt_epi8(other, *this); }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t>: base8_numeric<uint8_t> {
+    simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+    simdjson_inline simd8(const __m128i _value) : base8_numeric<uint8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const uint8_t* values) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) : simd8(_mm_setr_epi8(
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    )) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm_adds_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm_subs_epu8(*this, other); }
+
+    // Order-specific operations
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm_max_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return _mm_min_epu8(*this, other); }
+    // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
+    // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+
+    // Bit-specific operations
+    simdjson_inline simd8<bool> bits_not_set() const { return *this == uint8_t(0); }
+    simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
+    simdjson_inline bool is_ascii() const { return _mm_movemask_epi8(*this) == 0; }
+    simdjson_inline bool bits_not_set_anywhere() const { return _mm_testz_si128(*this, *this); }
+    simdjson_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
+    simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return _mm_testz_si128(*this, bits); }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(_mm_slli_epi16(*this, N)) & uint8_t(0xFFu << N); }
+    // Get one of the bits and make a bitmask out of it.
+    // e.g. value.get_bit<7>() gets the high bit
+    template<int N>
+    simdjson_inline int get_bit() const { return _mm_movemask_epi8(_mm_slli_epi16(*this, 7-N)); }
+  };
+
+  template<typename T>
+  struct simd8x64 {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+    static_assert(NUM_CHUNKS == 4, "Westmere kernel should use four registers per 64-byte block.");
+    const simd8<T> chunks[NUM_CHUNKS];
+    template<int idx> simd8<uint8_t> get() const { return idx < NUM_CHUNKS ? chunks[idx] : simd8<T>(); }
+
+    simd8x64(const simd8x64<T>& o) = delete; // no copy allowed
+    simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1, const simd8<T> chunk2, const simd8<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+16), simd8<T>::load(ptr+32), simd8<T>::load(ptr+48)} {}
+
+    simdjson_inline void store(T ptr[64]) const {
+      this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
+      this->chunks[1].store(ptr+sizeof(simd8<T>)*1);
+      this->chunks[2].store(ptr+sizeof(simd8<T>)*2);
+      this->chunks[3].store(ptr+sizeof(simd8<T>)*3);
+    }
+
+    simdjson_inline simd8<T> reduce_or() const {
+      return (this->chunks[0] | this->chunks[1]) | (this->chunks[2] | this->chunks[3]);
+    }
+
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
+      this->chunks[0].compress(uint16_t(mask), output);
+      this->chunks[1].compress(uint16_t(mask >> 16), output + 16 - count_ones(mask & 0xFFFF));
+      this->chunks[2].compress(uint16_t(mask >> 32), output + 32 - count_ones(mask & 0xFFFFFFFF));
+      this->chunks[3].compress(uint16_t(mask >> 48), output + 48 - count_ones(mask & 0xFFFFFFFFFFFF));
+      return 64 - count_ones(mask);
+    }
+
+    simdjson_inline uint64_t to_bitmask() const {
+      uint64_t r0 = uint32_t(this->chunks[0].to_bitmask() );
+      uint64_t r1 =          this->chunks[1].to_bitmask() ;
+      uint64_t r2 =          this->chunks[2].to_bitmask() ;
+      uint64_t r3 =          this->chunks[3].to_bitmask() ;
+      return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
+    }
+
+    simdjson_inline uint64_t eq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] == mask,
+        this->chunks[1] == mask,
+        this->chunks[2] == mask,
+        this->chunks[3] == mask
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+      return  simd8x64<bool>(
+        this->chunks[0] == other.chunks[0],
+        this->chunks[1] == other.chunks[1],
+        this->chunks[2] == other.chunks[2],
+        this->chunks[3] == other.chunks[3]
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t lteq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] <= mask,
+        this->chunks[1] <= mask,
+        this->chunks[2] <= mask,
+        this->chunks[3] <= mask
+      ).to_bitmask();
+    }
+  }; // struct simd8x64<T>
+
+} // namespace simd
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_SIMD_INPUT_H
+/* end file simdjson/westmere/simd.h */
+/* including simdjson/westmere/stringparsing_defs.h: #include "simdjson/westmere/stringparsing_defs.h" */
+/* begin file simdjson/westmere/stringparsing_defs.h */
+#ifndef SIMDJSON_WESTMERE_STRINGPARSING_DEFS_H
+#define SIMDJSON_WESTMERE_STRINGPARSING_DEFS_H
+
+/* including simdjson/westmere/bitmanipulation.h: #include "simdjson/westmere/bitmanipulation.h" */
+/* begin file simdjson/westmere/bitmanipulation.h */
+#ifndef SIMDJSON_WESTMERE_BITMANIPULATION_H
+#define SIMDJSON_WESTMERE_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  unsigned long ret;
+  // Search the mask data from least significant bit (LSB)
+  // to the most significant bit (MSB) for a set bit (1).
+  _BitScanForward64(&ret, input_num);
+  return (int)ret;
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO
+  return __builtin_ctzll(input_num);
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return input_num & (input_num-1);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  unsigned long leading_zero = 0;
+  // Search the mask data from most significant bit (MSB)
+  // to least significant bit (LSB) for a set bit (1).
+  if (_BitScanReverse64(&leading_zero, input_num))
+    return (int)(63 - leading_zero);
+  else
+    return 64;
+#else
+  return __builtin_clzll(input_num);
+#endif// SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+simdjson_inline unsigned __int64 count_ones(uint64_t input_num) {
+  // note: we do not support legacy 32-bit Windows in this kernel
+  return __popcnt64(input_num);// Visual Studio wants two underscores
+}
+#else
+simdjson_inline long long int count_ones(uint64_t input_num) {
+  return _popcnt64(input_num);
+}
+#endif
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2,
+                                uint64_t *result) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  return _addcarry_u64(0, value1, value2,
+                       reinterpret_cast<unsigned __int64 *>(result));
+#else
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+#endif
+}
+
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_BITMANIPULATION_H
+/* end file simdjson/westmere/bitmanipulation.h */
+/* including simdjson/westmere/simd.h: #include "simdjson/westmere/simd.h" */
+/* begin file simdjson/westmere/simd.h */
+#ifndef SIMDJSON_WESTMERE_SIMD_H
+#define SIMDJSON_WESTMERE_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace simd {
+
+  template<typename Child>
+  struct base {
+    __m128i value;
+
+    // Zero constructor
+    simdjson_inline base() : value{__m128i()} {}
+
+    // Conversion from SIMD register
+    simdjson_inline base(const __m128i _value) : value(_value) {}
+
+    // Conversion to SIMD register
+    simdjson_inline operator const __m128i&() const { return this->value; }
+    simdjson_inline operator __m128i&() { return this->value; }
+
+    // Bit operations
+    simdjson_inline Child operator|(const Child other) const { return _mm_or_si128(*this, other); }
+    simdjson_inline Child operator&(const Child other) const { return _mm_and_si128(*this, other); }
+    simdjson_inline Child operator^(const Child other) const { return _mm_xor_si128(*this, other); }
+    simdjson_inline Child bit_andnot(const Child other) const { return _mm_andnot_si128(other, *this); }
+    simdjson_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+  };
+
+  template<typename T, typename Mask=simd8<bool>>
+  struct base8: base<simd8<T>> {
+    typedef uint16_t bitmask_t;
+    typedef uint32_t bitmask2_t;
+
+    simdjson_inline base8() : base<simd8<T>>() {}
+    simdjson_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}
+
+    friend simdjson_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return _mm_cmpeq_epi8(lhs, rhs); }
+
+    static const int SIZE = sizeof(base<simd8<T>>::value);
+
+    template<int N=1>
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+      return _mm_alignr_epi8(*this, prev_chunk, 16 - N);
+    }
+  };
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool>: base8<bool> {
+    static simdjson_inline simd8<bool> splat(bool _value) { return _mm_set1_epi8(uint8_t(-(!!_value))); }
+
+    simdjson_inline simd8() : base8() {}
+    simdjson_inline simd8(const __m128i _value) : base8<bool>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+
+    simdjson_inline int to_bitmask() const { return _mm_movemask_epi8(*this); }
+    simdjson_inline bool any() const { return !_mm_testz_si128(*this, *this); }
+    simdjson_inline simd8<bool> operator~() const { return *this ^ true; }
+  };
+
+  template<typename T>
+  struct base8_numeric: base8<T> {
+    static simdjson_inline simd8<T> splat(T _value) { return _mm_set1_epi8(_value); }
+    static simdjson_inline simd8<T> zero() { return _mm_setzero_si128(); }
+    static simdjson_inline simd8<T> load(const T values[16]) {
+      return _mm_loadu_si128(reinterpret_cast<const __m128i *>(values));
+    }
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    static simdjson_inline simd8<T> repeat_16(
+      T v0,  T v1,  T v2,  T v3,  T v4,  T v5,  T v6,  T v7,
+      T v8,  T v9,  T v10, T v11, T v12, T v13, T v14, T v15
+    ) {
+      return simd8<T>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    simdjson_inline base8_numeric() : base8<T>() {}
+    simdjson_inline base8_numeric(const __m128i _value) : base8<T>(_value) {}
+
+    // Store to array
+    simdjson_inline void store(T dst[16]) const { return _mm_storeu_si128(reinterpret_cast<__m128i *>(dst), *this); }
+
+    // Override to distinguish from bool version
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<T> operator+(const simd8<T> other) const { return _mm_add_epi8(*this, other); }
+    simdjson_inline simd8<T> operator-(const simd8<T> other) const { return _mm_sub_epi8(*this, other); }
+    simdjson_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return _mm_shuffle_epi8(lookup_table, *this);
+    }
+
+    // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted as a bitset).
+    // Passing a 0 value for mask would be equivalent to writing out every byte to output.
+    // Only the first 16 - count_ones(mask) bytes of the result are significant but 16 bytes
+    // get written.
+    // Design consideration: it seems like a function with the
+    // signature simd8<L> compress(uint32_t mask) would be
+    // sensible, but the AVX ISA makes this kind of approach difficult.
+    template<typename L>
+    simdjson_inline void compress(uint16_t mask, L * output) const {
+      using internal::thintable_epi8;
+      using internal::BitsSetTable256mul2;
+      using internal::pshufb_combine_table;
+      // this particular implementation was inspired by work done by @animetosho
+      // we do it in two steps, first 8 bytes and then second 8 bytes
+      uint8_t mask1 = uint8_t(mask); // least significant 8 bits
+      uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
+      // next line just loads the 64-bit values thintable_epi8[mask1] and
+      // thintable_epi8[mask2] into a 128-bit register, using only
+      // two instructions on most compilers.
+      __m128i shufmask =  _mm_set_epi64x(thintable_epi8[mask2], thintable_epi8[mask1]);
+      // we increment by 0x08 the second half of the mask
+      shufmask =
+      _mm_add_epi8(shufmask, _mm_set_epi32(0x08080808, 0x08080808, 0, 0));
+      // this is the version "nearly pruned"
+      __m128i pruned = _mm_shuffle_epi8(*this, shufmask);
+      // we still need to put the two halves together.
+      // we compute the popcount of the first half:
+      int pop1 = BitsSetTable256mul2[mask1];
+      // then load the corresponding mask, what it does is to write
+      // only the first pop1 bytes from the first 8 bytes, and then
+      // it fills in with the bytes from the second 8 bytes + some filling
+      // at the end.
+      __m128i compactmask =
+      _mm_loadu_si128(reinterpret_cast<const __m128i *>(pshufb_combine_table + pop1 * 8));
+      __m128i answer = _mm_shuffle_epi8(pruned, compactmask);
+      _mm_storeu_si128(reinterpret_cast<__m128i *>(output), answer);
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> : base8_numeric<int8_t> {
+    simdjson_inline simd8() : base8_numeric<int8_t>() {}
+    simdjson_inline simd8(const __m128i _value) : base8_numeric<int8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const int8_t* values) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) : simd8(_mm_setr_epi8(
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    )) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm_max_epi8(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm_min_epi8(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm_cmpgt_epi8(*this, other); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return _mm_cmpgt_epi8(other, *this); }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t>: base8_numeric<uint8_t> {
+    simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+    simdjson_inline simd8(const __m128i _value) : base8_numeric<uint8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const uint8_t* values) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) : simd8(_mm_setr_epi8(
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    )) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm_adds_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm_subs_epu8(*this, other); }
+
+    // Order-specific operations
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm_max_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return _mm_min_epu8(*this, other); }
+    // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
+    // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+
+    // Bit-specific operations
+    simdjson_inline simd8<bool> bits_not_set() const { return *this == uint8_t(0); }
+    simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
+    simdjson_inline bool is_ascii() const { return _mm_movemask_epi8(*this) == 0; }
+    simdjson_inline bool bits_not_set_anywhere() const { return _mm_testz_si128(*this, *this); }
+    simdjson_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
+    simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return _mm_testz_si128(*this, bits); }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(_mm_slli_epi16(*this, N)) & uint8_t(0xFFu << N); }
+    // Get one of the bits and make a bitmask out of it.
+    // e.g. value.get_bit<7>() gets the high bit
+    template<int N>
+    simdjson_inline int get_bit() const { return _mm_movemask_epi8(_mm_slli_epi16(*this, 7-N)); }
+  };
+
+  template<typename T>
+  struct simd8x64 {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+    static_assert(NUM_CHUNKS == 4, "Westmere kernel should use four registers per 64-byte block.");
+    const simd8<T> chunks[NUM_CHUNKS];
+    template<int idx> simd8<uint8_t> get() const { return idx < NUM_CHUNKS ? chunks[idx] : simd8<T>(); }
+
+    simd8x64(const simd8x64<T>& o) = delete; // no copy allowed
+    simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1, const simd8<T> chunk2, const simd8<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+16), simd8<T>::load(ptr+32), simd8<T>::load(ptr+48)} {}
+
+    simdjson_inline void store(T ptr[64]) const {
+      this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
+      this->chunks[1].store(ptr+sizeof(simd8<T>)*1);
+      this->chunks[2].store(ptr+sizeof(simd8<T>)*2);
+      this->chunks[3].store(ptr+sizeof(simd8<T>)*3);
+    }
+
+    simdjson_inline simd8<T> reduce_or() const {
+      return (this->chunks[0] | this->chunks[1]) | (this->chunks[2] | this->chunks[3]);
+    }
+
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
+      this->chunks[0].compress(uint16_t(mask), output);
+      this->chunks[1].compress(uint16_t(mask >> 16), output + 16 - count_ones(mask & 0xFFFF));
+      this->chunks[2].compress(uint16_t(mask >> 32), output + 32 - count_ones(mask & 0xFFFFFFFF));
+      this->chunks[3].compress(uint16_t(mask >> 48), output + 48 - count_ones(mask & 0xFFFFFFFFFFFF));
+      return 64 - count_ones(mask);
+    }
+
+    simdjson_inline uint64_t to_bitmask() const {
+      uint64_t r0 = uint32_t(this->chunks[0].to_bitmask() );
+      uint64_t r1 =          this->chunks[1].to_bitmask() ;
+      uint64_t r2 =          this->chunks[2].to_bitmask() ;
+      uint64_t r3 =          this->chunks[3].to_bitmask() ;
+      return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
+    }
+
+    simdjson_inline uint64_t eq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] == mask,
+        this->chunks[1] == mask,
+        this->chunks[2] == mask,
+        this->chunks[3] == mask
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+      return  simd8x64<bool>(
+        this->chunks[0] == other.chunks[0],
+        this->chunks[1] == other.chunks[1],
+        this->chunks[2] == other.chunks[2],
+        this->chunks[3] == other.chunks[3]
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t lteq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] <= mask,
+        this->chunks[1] <= mask,
+        this->chunks[2] <= mask,
+        this->chunks[3] <= mask
+      ).to_bitmask();
+    }
+  }; // struct simd8x64<T>
+
+} // namespace simd
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_SIMD_INPUT_H
+/* end file simdjson/westmere/simd.h */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 32;
+  simdjson_inline backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return bs_bits != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+
+  uint32_t bs_bits;
+  uint32_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // this can read up to 31 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v0(src);
+  simd8<uint8_t> v1(src + 16);
+  v0.store(dst);
+  v1.store(dst + 16);
+  uint64_t bs_and_quote = simd8x64<bool>(v0 == '\\', v1 == '\\', v0 == '"', v1 == '"').to_bitmask();
+  return {
+    uint32_t(bs_and_quote),      // bs_bits
+    uint32_t(bs_and_quote >> 32) // quote_bits
+  };
+}
+
+
+struct escaping {
+  static constexpr uint32_t BYTES_PROCESSED = 16;
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits != 0; }
+  simdjson_inline int escape_index() { return trailing_zeroes(escape_bits); }
+
+  uint64_t escape_bits;
+}; // struct escaping
+
+
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "escaping finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  simd8<bool> is_quote = (v == '"');
+  simd8<bool> is_backslash = (v == '\\');
+  simd8<bool> is_control = (v < 32);
+  return {
+    uint64_t((is_backslash | is_quote | is_control).to_bitmask())
+  };
+}
+
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_STRINGPARSING_DEFS_H
+/* end file simdjson/westmere/stringparsing_defs.h */
+/* end file simdjson/westmere/begin.h */
+/* including simdjson/generic/amalgamated.h for westmere: #include "simdjson/generic/amalgamated.h" */
+/* begin file simdjson/generic/amalgamated.h for westmere */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_GENERIC_DEPENDENCIES_H)
+#error simdjson/generic/dependencies.h must be included before simdjson/generic/amalgamated.h!
+#endif
+
+/* including simdjson/generic/base.h for westmere: #include "simdjson/generic/base.h" */
+/* begin file simdjson/generic/base.h for westmere */
+#ifndef SIMDJSON_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): // If we haven't got an implementation yet, we're in the editor, editing a generic file! Just */
+/* amalgamation skipped (editor-only): // use the most advanced one we can so the most possible stuff can be tested. */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation_detection.h" */
+/* amalgamation skipped (editor-only): #if SIMDJSON_IMPLEMENTATION_ICELAKE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_HASWELL */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_WESTMERE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_ARM64 */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_PPC64 */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LASX */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LSX */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_RVV_VLS */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_FALLBACK */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/begin.h" */
+/* amalgamation skipped (editor-only): #else */
+/* amalgamation skipped (editor-only): #error "All possible implementations (including fallback) have been disabled! simdjson will not run." */
+/* amalgamation skipped (editor-only): #endif */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+
+struct open_container;
+class dom_parser_implementation;
+
+/**
+ * The type of a JSON number
+ */
+enum class number_type {
+    floating_point_number=1, /// a binary64 number
+    signed_integer,          /// a signed integer that fits in a 64-bit word using two's complement
+    unsigned_integer,        /// a positive integer larger or equal to 1<<63
+    big_integer              /// a big integer that does not fit in a 64-bit word
+};
+
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_BASE_H
+/* end file simdjson/generic/base.h for westmere */
+/* including simdjson/generic/jsoncharutils.h for westmere: #include "simdjson/generic/jsoncharutils.h" */
+/* begin file simdjson/generic/jsoncharutils.h for westmere */
+#ifndef SIMDJSON_GENERIC_JSONCHARUTILS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_JSONCHARUTILS_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/jsoncharutils_tables.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace jsoncharutils {
+
+// return non-zero if not a structural or whitespace char
+// zero otherwise
+simdjson_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace_negated[c];
+}
+
+simdjson_inline uint32_t is_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace[c];
+}
+
+// returns a value with the high 16 bits set if not valid
+// otherwise returns the conversion of the 4 hex digits at src into the bottom
+// 16 bits of the 32-bit return register
+//
+// see
+// https://lemire.me/blog/2019/04/17/parsing-short-hexadecimal-strings-efficiently/
+static inline uint32_t hex_to_u32_nocheck(
+    const uint8_t *src) { // strictly speaking, static inline is a C-ism
+  uint32_t v1 = internal::digit_to_val32[630 + src[0]];
+  uint32_t v2 = internal::digit_to_val32[420 + src[1]];
+  uint32_t v3 = internal::digit_to_val32[210 + src[2]];
+  uint32_t v4 = internal::digit_to_val32[0 + src[3]];
+  return v1 | v2 | v3 | v4;
+}
+
+// given a code point cp, writes to c
+// the utf-8 code, outputting the length in
+// bytes, if the length is zero, the code point
+// is invalid
+//
+// This can possibly be made faster using pdep
+// and clz and table lookups, but JSON documents
+// have few escaped code points, and the following
+// function looks cheap.
+//
+// Note: we assume that surrogates are treated separately
+//
+simdjson_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
+  if (cp <= 0x7F) {
+    c[0] = uint8_t(cp);
+    return 1; // ascii
+  }
+  if (cp <= 0x7FF) {
+    c[0] = uint8_t((cp >> 6) + 192);
+    c[1] = uint8_t((cp & 63) + 128);
+    return 2; // universal plane
+    //  Surrogates are treated elsewhere...
+    //} //else if (0xd800 <= cp && cp <= 0xdfff) {
+    //  return 0; // surrogates // could put assert here
+  } else if (cp <= 0xFFFF) {
+    c[0] = uint8_t((cp >> 12) + 224);
+    c[1] = uint8_t(((cp >> 6) & 63) + 128);
+    c[2] = uint8_t((cp & 63) + 128);
+    return 3;
+  } else if (cp <= 0x10FFFF) { // if you know you have a valid code point, this
+                               // is not needed
+    c[0] = uint8_t((cp >> 18) + 240);
+    c[1] = uint8_t(((cp >> 12) & 63) + 128);
+    c[2] = uint8_t(((cp >> 6) & 63) + 128);
+    c[3] = uint8_t((cp & 63) + 128);
+    return 4;
+  }
+  // will return 0 when the code point was too large.
+  return 0; // bad r
+}
+
+#if SIMDJSON_IS_32BITS // _umul128 for x86, arm
+// this is a slow emulation routine for 32-bit
+//
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+  return x * (uint64_t)y;
+}
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+  uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
+  uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
+  uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
+  uint64_t adbc_carry = !!(adbc < ad);
+  uint64_t lo = bd + (adbc << 32);
+  *hi = __emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) +
+        (adbc_carry << 32) + !!(lo < bd);
+  return lo;
+}
+#endif
+
+} // namespace jsoncharutils
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_JSONCHARUTILS_H
+/* end file simdjson/generic/jsoncharutils.h for westmere */
+/* including simdjson/generic/atomparsing.h for westmere: #include "simdjson/generic/atomparsing.h" */
+/* begin file simdjson/generic/atomparsing.h for westmere */
+#ifndef SIMDJSON_GENERIC_ATOMPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_ATOMPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace westmere {
+namespace {
+/// @private
+namespace atomparsing {
+
+// The string_to_uint32 is exclusively used to map literal strings to 32-bit values.
+// We use memcpy instead of a pointer cast to avoid undefined behaviors since we cannot
+// be certain that the character pointer will be properly aligned.
+// You might think that using memcpy makes this function expensive, but you'd be wrong.
+// All decent optimizing compilers (GCC, clang, Visual Studio) will compile string_to_uint32("false");
+// to the compile-time constant 1936482662.
+simdjson_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
+
+
+// Again in str4ncmp we use a memcpy to avoid undefined behavior. The memcpy may appear expensive.
+// Yet all decent optimizing compilers will compile memcpy to a single instruction, just about.
+simdjson_warn_unused
+simdjson_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
+  uint32_t srcval; // we want to avoid unaligned 32-bit loads (undefined in C/C++)
+  static_assert(sizeof(uint32_t) <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be larger than 4 bytes");
+  std::memcpy(&srcval, src, sizeof(uint32_t));
+  return srcval ^ string_to_uint32(atom);
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src) {
+  return (str4ncmp(src, "true") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_true_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "true"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src) {
+  return (str4ncmp(src+1, "alse") | jsoncharutils::is_not_structural_or_whitespace(src[5])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
+  if (len > 5) { return is_valid_false_atom(src); }
+  else if (len == 5) { return !str4ncmp(src+1, "alse"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src) {
+  return (str4ncmp(src, "null") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_null_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "null"); }
+  else { return false; }
+}
+
+} // namespace atomparsing
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_ATOMPARSING_H
+/* end file simdjson/generic/atomparsing.h for westmere */
+/* including simdjson/generic/dom_parser_implementation.h for westmere: #include "simdjson/generic/dom_parser_implementation.h" */
+/* begin file simdjson/generic/dom_parser_implementation.h for westmere */
+#ifndef SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/dom_parser_implementation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+
+// expectation: sizeof(open_container) = 64/8.
+struct open_container {
+  uint32_t tape_index; // where, on the tape, does the scope ([,{) begins
+  uint32_t count; // how many elements in the scope
+}; // struct open_container
+
+static_assert(sizeof(open_container) == 64/8, "Open container must be 64 bits");
+
+class dom_parser_implementation final : public internal::dom_parser_implementation {
+public:
+  /** Tape location of each open { or [ */
+  std::unique_ptr<open_container[]> open_containers{};
+  /** Whether each open container is a [ or { */
+  std::unique_ptr<bool[]> is_array{};
+  /** Buffer passed to stage 1 */
+  const uint8_t *buf{};
+  /** Length passed to stage 1 */
+  size_t len{0};
+  /** Document passed to stage 2 */
+  dom::document *doc{};
+
+  inline dom_parser_implementation() noexcept;
+  inline dom_parser_implementation(dom_parser_implementation &&other) noexcept;
+  inline dom_parser_implementation &operator=(dom_parser_implementation &&other) noexcept;
+  dom_parser_implementation(const dom_parser_implementation &) = delete;
+  dom_parser_implementation &operator=(const dom_parser_implementation &) = delete;
+
+  simdjson_warn_unused error_code parse(const uint8_t *buf, size_t len, dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage1(const uint8_t *buf, size_t len, stage1_mode partial) noexcept final;
+  simdjson_warn_unused error_code stage2(dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage2_next(dom::document &doc) noexcept final;
+  simdjson_warn_unused uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) const noexcept final;
+  simdjson_warn_unused uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept final;
+  inline simdjson_warn_unused error_code set_capacity(size_t capacity) noexcept final;
+  inline simdjson_warn_unused error_code set_max_depth(size_t max_depth) noexcept final;
+private:
+  simdjson_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
+
+};
+
+} // namespace westmere
+} // namespace simdjson
+
+namespace simdjson {
+namespace westmere {
+
+inline dom_parser_implementation::dom_parser_implementation() noexcept = default;
+inline dom_parser_implementation::dom_parser_implementation(dom_parser_implementation &&other) noexcept = default;
+inline dom_parser_implementation &dom_parser_implementation::operator=(dom_parser_implementation &&other) noexcept = default;
+
+// Leaving these here so they can be inlined if so desired
+inline simdjson_warn_unused error_code dom_parser_implementation::set_capacity(size_t capacity) noexcept {
+  if(capacity > SIMDJSON_MAXSIZE_BYTES) { return CAPACITY; }
+  // Stage 1 index output
+  size_t max_structures = SIMDJSON_ROUNDUP_N(capacity, 64) + 2 + 7;
+  structural_indexes.reset( new (std::nothrow) uint32_t[max_structures] );
+  if (!structural_indexes) { _capacity = 0; return MEMALLOC; }
+  structural_indexes[0] = 0;
+  n_structural_indexes = 0;
+
+  _capacity = capacity;
+  return SUCCESS;
+}
+
+inline simdjson_warn_unused error_code dom_parser_implementation::set_max_depth(size_t max_depth) noexcept {
+  // Stage 2 stacks
+  open_containers.reset(new (std::nothrow) open_container[max_depth]);
+  is_array.reset(new (std::nothrow) bool[max_depth]);
+  if (!is_array || !open_containers) { _max_depth = 0; return MEMALLOC; }
+
+  _max_depth = max_depth;
+  return SUCCESS;
+}
+
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file simdjson/generic/dom_parser_implementation.h for westmere */
+/* including simdjson/generic/implementation_simdjson_result_base.h for westmere: #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base.h for westmere */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+
+// This is a near copy of include/error.h's implementation_simdjson_result_base, except it doesn't use std::pair
+// so we can avoid inlining errors
+// TODO reconcile these!
+/**
+ * The result of a simdjson operation that could fail.
+ *
+ * Gives the option of reading error codes, or throwing an exception by casting to the desired result.
+ *
+ * This is a base class for implementations that want to add functions to the result type for
+ * chaining.
+ *
+ * Override like:
+ *
+ *   struct simdjson_result<T> : public internal::implementation_simdjson_result_base<T> {
+ *     simdjson_result() noexcept : internal::implementation_simdjson_result_base<T>() {}
+ *     simdjson_result(error_code error) noexcept : internal::implementation_simdjson_result_base<T>(error) {}
+ *     simdjson_result(T &&value) noexcept : internal::implementation_simdjson_result_base<T>(std::forward(value)) {}
+ *     simdjson_result(T &&value, error_code error) noexcept : internal::implementation_simdjson_result_base<T>(value, error) {}
+ *     // Your extra methods here
+ *   }
+ *
+ * Then any method returning simdjson_result<T> will be chainable with your methods.
+ */
+template<typename T>
+struct implementation_simdjson_result_base {
+
+  /**
+   * Create a new empty result with error = UNINITIALIZED.
+   */
+  simdjson_inline implementation_simdjson_result_base() noexcept = default;
+
+  /**
+   * Create a new error result.
+   */
+  simdjson_inline implementation_simdjson_result_base(error_code error) noexcept;
+
+  /**
+   * Create a new successful result.
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value) noexcept;
+
+  /**
+   * Create a new result with both things (use if you don't want to branch when creating the result).
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value, error_code error) noexcept;
+
+  /**
+   * Move the value and the error to the provided variables.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   * @param error The variable to assign the error to. Set to SUCCESS if there is no error.
+   */
+  simdjson_inline void tie(T &value, error_code &error) && noexcept;
+
+  /**
+   * Move the value to the provided variable.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   */
+  simdjson_warn_unused simdjson_inline error_code get(T &value) && noexcept;
+
+  /**
+   * The error.
+   */
+  simdjson_warn_unused simdjson_inline error_code error() const noexcept;
+
+  /**
+   * Whether there is a value.
+   */
+  simdjson_warn_unused simdjson_inline bool has_value() const noexcept;
+
+#if SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T& operator*() &  noexcept(false);
+  simdjson_inline T&& operator*() &&  noexcept(false);
+  /**
+   * Arrow operator to access members of the contained value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T* operator->() noexcept(false);
+  simdjson_inline const T* operator->() const noexcept(false);
+
+  simdjson_inline T& value() & noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& value() && noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& take_value() && noexcept(false);
+
+  /**
+   * Cast to the value (will throw on error).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline operator T&&() && noexcept(false);
+
+
+#endif // SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline const T& value_unsafe() const& noexcept;
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T& value_unsafe() & noexcept;
+  /**
+   * Take the result value (move it). This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T&& value_unsafe() && noexcept;
+
+  using value_type = T;
+  using error_type = error_code;
+
+protected:
+  /** users should never directly access first and second. **/
+  T first{}; /** Users should never directly access 'first'. **/
+  error_code second{UNINITIALIZED}; /** Users should never directly access 'second'. **/
+}; // struct implementation_simdjson_result_base
+
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+/* end file simdjson/generic/implementation_simdjson_result_base.h for westmere */
+/* including simdjson/generic/numberparsing.h for westmere: #include "simdjson/generic/numberparsing.h" */
+/* begin file simdjson/generic/numberparsing.h for westmere */
+#ifndef SIMDJSON_GENERIC_NUMBERPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_NUMBERPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <limits>
+#include <ostream>
+#include <cstring>
+
+namespace simdjson {
+namespace westmere {
+namespace numberparsing {
+
+#ifdef JSON_TEST_NUMBERS
+#define INVALID_NUMBER(SRC) (found_invalid_number((SRC)), NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (found_integer((VALUE), (SRC)), (WRITER).append_s64((VALUE)))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (found_unsigned_integer((VALUE), (SRC)), (WRITER).append_u64((VALUE)))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (found_float((VALUE), (SRC)), (WRITER).append_double((VALUE)))
+#define BIGINT_NUMBER(SRC) (found_invalid_number((SRC)), BIGINT_ERROR)
+#else
+#define INVALID_NUMBER(SRC) (NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (WRITER).append_s64((VALUE))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (WRITER).append_u64((VALUE))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (WRITER).append_double((VALUE))
+#define BIGINT_NUMBER(SRC) (BIGINT_ERROR)
+#endif
+
+namespace {
+
+// Convert a mantissa, an exponent and a sign bit into an ieee64 double.
+// The real_exponent needs to be in [0, 2046] (technically real_exponent = 2047 would be acceptable).
+// The mantissa should be in [0,1<<53). The bit at index (1ULL << 52) while be zeroed.
+simdjson_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
+    double d;
+    mantissa &= ~(1ULL << 52);
+    mantissa |= real_exponent << 52;
+    mantissa |= ((static_cast<uint64_t>(negative)) << 63);
+    std::memcpy(&d, &mantissa, sizeof(d));
+    return d;
+}
+
+// Attempts to compute i * 10^(power) exactly; and if "negative" is
+// true, negate the result.
+// This function will only work in some cases, when it does not work, success is
+// set to false. This should work *most of the time* (like 99% of the time).
+// We assume that power is in the [smallest_power,
+// largest_power] interval: the caller is responsible for this check.
+simdjson_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
+  // we start with a fast path
+  // It was described in
+  // Clinger WD. How to read floating point numbers accurately.
+  // ACM SIGPLAN Notices. 1990
+#ifndef FLT_EVAL_METHOD
+#error "FLT_EVAL_METHOD should be defined, please include cfloat."
+#endif
+#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
+  // We cannot be certain that x/y is rounded to nearest.
+  if (0 <= power && power <= 22 && i <= 9007199254740991)
+#else
+  if (-22 <= power && power <= 22 && i <= 9007199254740991)
+#endif
+  {
+    // convert the integer into a double. This is lossless since
+    // 0 <= i <= 2^53 - 1.
+    d = double(i);
+    //
+    // The general idea is as follows.
+    // If 0 <= s < 2^53 and if 10^0 <= p <= 10^22 then
+    // 1) Both s and p can be represented exactly as 64-bit floating-point
+    // values
+    // (binary64).
+    // 2) Because s and p can be represented exactly as floating-point values,
+    // then s * p
+    // and s / p will produce correctly rounded values.
+    //
+    if (power < 0) {
+      d = d / simdjson::internal::power_of_ten[-power];
+    } else {
+      d = d * simdjson::internal::power_of_ten[power];
+    }
+    if (negative) {
+      d = -d;
+    }
+    return true;
+  }
+  // When 22 < power && power <  22 + 16, we could
+  // hope for another, secondary fast path.  It was
+  // described by David M. Gay in  "Correctly rounded
+  // binary-decimal and decimal-binary conversions." (1990)
+  // If you need to compute i * 10^(22 + x) for x < 16,
+  // first compute i * 10^x, if you know that result is exact
+  // (e.g., when i * 10^x < 2^53),
+  // then you can still proceed and do (i * 10^x) * 10^22.
+  // Is this worth your time?
+  // You need  22 < power *and* power <  22 + 16 *and* (i * 10^(x-22) < 2^53)
+  // for this second fast path to work.
+  // If you you have 22 < power *and* power <  22 + 16, and then you
+  // optimistically compute "i * 10^(x-22)", there is still a chance that you
+  // have wasted your time if i * 10^(x-22) >= 2^53. It makes the use cases of
+  // this optimization maybe less common than we would like. Source:
+  // http://www.exploringbinary.com/fast-path-decimal-to-floating-point-conversion/
+  // also used in RapidJSON: https://rapidjson.org/strtod_8h_source.html
+
+  // The fast path has now failed, so we are failing back on the slower path.
+
+  // In the slow path, we need to adjust i so that it is > 1<<63 which is always
+  // possible, except if i == 0, so we handle i == 0 separately.
+  if(i == 0) {
+    d = negative ? -0.0 : 0.0;
+    return true;
+  }
+
+
+  // The exponent is 1024 + 63 + power
+  //     + floor(log(5**power)/log(2)).
+  // The 1024 comes from the ieee64 standard.
+  // The 63 comes from the fact that we use a 64-bit word.
+  //
+  // Computing floor(log(5**power)/log(2)) could be
+  // slow. Instead we use a fast function.
+  //
+  // For power in (-400,350), we have that
+  // (((152170 + 65536) * power ) >> 16);
+  // is equal to
+  //  floor(log(5**power)/log(2)) + power when power >= 0
+  // and it is equal to
+  //  ceil(log(5**-power)/log(2)) + power when power < 0
+  //
+  // The 65536 is (1<<16) and corresponds to
+  // (65536 * power) >> 16 ---> power
+  //
+  // ((152170 * power ) >> 16) is equal to
+  // floor(log(5**power)/log(2))
+  //
+  // Note that this is not magic: 152170/(1<<16) is
+  // approximately equal to log(5)/log(2).
+  // The 1<<16 value is a power of two; we could use a
+  // larger power of 2 if we wanted to.
+  //
+  int64_t exponent = (((152170 + 65536) * power) >> 16) + 1024 + 63;
+
+
+  // We want the most significant bit of i to be 1. Shift if needed.
+  int lz = leading_zeroes(i);
+  i <<= lz;
+
+
+  // We are going to need to do some 64-bit arithmetic to get a precise product.
+  // We use a table lookup approach.
+  // It is safe because
+  // power >= smallest_power
+  // and power <= largest_power
+  // We recover the mantissa of the power, it has a leading 1. It is always
+  // rounded down.
+  //
+  // We want the most significant 64 bits of the product. We know
+  // this will be non-zero because the most significant bit of i is
+  // 1.
+  const uint32_t index = 2 * uint32_t(power - simdjson::internal::smallest_power);
+  // Optimization: It may be that materializing the index as a variable might confuse some compilers and prevent effective complex-addressing loads. (Done for code clarity.)
+  //
+  // The full_multiplication function computes the 128-bit product of two 64-bit words
+  // with a returned value of type value128 with a "low component" corresponding to the
+  // 64-bit least significant bits of the product and with a "high component" corresponding
+  // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index]);
+#else
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index]);
+#endif
+
+  // Both i and power_of_five_128[index] have their most significant bit set to 1 which
+  // implies that the either the most or the second most significant bit of the product
+  // is 1. We pack values in this manner for efficiency reasons: it maximizes the use
+  // we make of the product. It also makes it easy to reason about the product: there
+  // is 0 or 1 leading zero in the product.
+
+  // Unless the least significant 9 bits of the high (64-bit) part of the full
+  // product are all 1s, then we know that the most significant 55 bits are
+  // exact and no further work is needed. Having 55 bits is necessary because
+  // we need 53 bits for the mantissa but we have to have one rounding bit and
+  // we can waste a bit if the most significant bit of the product is zero.
+  if((firstproduct.high & 0x1FF) == 0x1FF) {
+    // We want to compute i * 5^q, but only care about the top 55 bits at most.
+    // Consider the scenario where q>=0. Then 5^q may not fit in 64-bits. Doing
+    // the full computation is wasteful. So we do what is called a "truncated
+    // multiplication".
+    // We take the most significant 64-bits, and we put them in
+    // power_of_five_128[index]. Usually, that's good enough to approximate i * 5^q
+    // to the desired approximation using one multiplication. Sometimes it does not suffice.
+    // Then we store the next most significant 64 bits in power_of_five_128[index + 1], and
+    // then we get a better approximation to i * 5^q.
+    //
+    // That's for when q>=0. The logic for q<0 is somewhat similar but it is somewhat
+    // more complicated.
+    //
+    // There is an extra layer of complexity in that we need more than 55 bits of
+    // accuracy in the round-to-even scenario.
+    //
+    // The full_multiplication function computes the 128-bit product of two 64-bit words
+    // with a returned value of type value128 with a "low component" corresponding to the
+    // 64-bit least significant bits of the product and with a "high component" corresponding
+    // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index + 1]);
+#else
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index + 1]);
+#endif
+    firstproduct.low += secondproduct.high;
+    if(secondproduct.high > firstproduct.low) { firstproduct.high++; }
+    // As it has been proven by Noble Mushtak and Daniel Lemire in "Fast Number Parsing Without
+    // Fallback" (https://arxiv.org/abs/2212.06644), at this point we are sure that the product
+    // is sufficiently accurate, and more computation is not needed.
+  }
+  uint64_t lower = firstproduct.low;
+  uint64_t upper = firstproduct.high;
+  // The final mantissa should be 53 bits with a leading 1.
+  // We shift it so that it occupies 54 bits with a leading 1.
+  ///////
+  uint64_t upperbit = upper >> 63;
+  uint64_t mantissa = upper >> (upperbit + 9);
+  lz += int(1 ^ upperbit);
+
+  // Here we have mantissa < (1<<54).
+  int64_t real_exponent = exponent - lz;
+  if (simdjson_unlikely(real_exponent <= 0)) { // we have a subnormal?
+    // Here have that real_exponent <= 0 so -real_exponent >= 0
+    if(-real_exponent + 1 >= 64) { // if we have more than 64 bits below the minimum exponent, you have a zero for sure.
+      d = negative ? -0.0 : 0.0;
+      return true;
+    }
+    // next line is safe because -real_exponent + 1 < 0
+    mantissa >>= -real_exponent + 1;
+    // Thankfully, we can't have both "round-to-even" and subnormals because
+    // "round-to-even" only occurs for powers close to 0.
+    mantissa += (mantissa & 1); // round up
+    mantissa >>= 1;
+    // There is a weird scenario where we don't have a subnormal but just.
+    // Suppose we start with 2.2250738585072013e-308, we end up
+    // with 0x3fffffffffffff x 2^-1023-53 which is technically subnormal
+    // whereas 0x40000000000000 x 2^-1023-53  is normal. Now, we need to round
+    // up 0x3fffffffffffff x 2^-1023-53  and once we do, we are no longer
+    // subnormal, but we can only know this after rounding.
+    // So we only declare a subnormal if we are smaller than the threshold.
+    real_exponent = (mantissa < (uint64_t(1) << 52)) ? 0 : 1;
+    d = to_double(mantissa, real_exponent, negative);
+    return true;
+  }
+  // We have to round to even. The "to even" part
+  // is only a problem when we are right in between two floats
+  // which we guard against.
+  // If we have lots of trailing zeros, we may fall right between two
+  // floating-point values.
+  //
+  // The round-to-even cases take the form of a number 2m+1 which is in (2^53,2^54]
+  // times a power of two. That is, it is right between a number with binary significand
+  // m and another number with binary significand m+1; and it must be the case
+  // that it cannot be represented by a float itself.
+  //
+  // We must have that w * 10 ^q == (2m+1) * 2^p for some power of two 2^p.
+  // Recall that 10^q = 5^q * 2^q.
+  // When q >= 0, we must have that (2m+1) is divible by 5^q, so 5^q <= 2^54. We have that
+  //  5^23 <=  2^54 and it is the last power of five to qualify, so q <= 23.
+  // When q<0, we have  w  >=  (2m+1) x 5^{-q}.  We must have that w<2^{64} so
+  // (2m+1) x 5^{-q} < 2^{64}. We have that 2m+1>2^{53}. Hence, we must have
+  // 2^{53} x 5^{-q} < 2^{64}.
+  // Hence we have 5^{-q} < 2^{11}$ or q>= -4.
+  //
+  // We require lower <= 1 and not lower == 0 because we could not prove that
+  // that lower == 0 is implied; but we could prove that lower <= 1 is a necessary and sufficient test.
+  if (simdjson_unlikely((lower <= 1) && (power >= -4) && (power <= 23) && ((mantissa & 3) == 1))) {
+    if((mantissa  << (upperbit + 64 - 53 - 2)) ==  upper) {
+      mantissa &= ~1;             // flip it so that we do not round up
+    }
+  }
+
+  mantissa += mantissa & 1;
+  mantissa >>= 1;
+
+  // Here we have mantissa < (1<<53), unless there was an overflow
+  if (mantissa >= (1ULL << 53)) {
+    //////////
+    // This will happen when parsing values such as 7.2057594037927933e+16
+    ////////
+    mantissa = (1ULL << 52);
+    real_exponent++;
+  }
+  mantissa &= ~(1ULL << 52);
+  // we have to check that real_exponent is in range, otherwise we bail out
+  if (simdjson_unlikely(real_exponent > 2046)) {
+    // We have an infinite value!!! We could actually throw an error here if we could.
+    return false;
+  }
+  d = to_double(mantissa, real_exponent, negative);
+  return true;
+}
+
+// We call a fallback floating-point parser that might be slow. Note
+// it will accept JSON numbers, but the JSON spec. is more restrictive so
+// before you call parse_float_fallback, you need to have validated the input
+// string with the JSON grammar.
+// It will return an error (false) if the parsed number is infinite.
+// The string parsing itself always succeeds. We know that there is at least
+// one digit.
+static bool parse_float_fallback(const uint8_t *ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+static bool parse_float_fallback(const uint8_t *ptr, const uint8_t *end_ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr), reinterpret_cast<const char *>(end_ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+// check quickly whether the next 8 chars are made of digits
+// at a glance, it looks better than Mula's
+// http://0x80.pl/articles/swar-digits-validate.html
+simdjson_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
+  uint64_t val;
+  // this can read up to 7 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(7 <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be bigger than 7");
+  std::memcpy(&val, chars, 8);
+  // a branchy method might be faster:
+  // return (( val & 0xF0F0F0F0F0F0F0F0 ) == 0x3030303030303030)
+  //  && (( (val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0 ) ==
+  //  0x3030303030303030);
+  return (((val & 0xF0F0F0F0F0F0F0F0) |
+           (((val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0) >> 4)) ==
+          0x3333333333333333);
+}
+
+template<typename I>
+SIMDJSON_NO_SANITIZE_UNDEFINED // We deliberately allow overflow here and check later
+simdjson_inline bool parse_digit(const uint8_t c, I &i) {
+  const uint8_t digit = static_cast<uint8_t>(c - '0');
+  if (digit > 9) {
+    return false;
+  }
+  // PERF NOTE: multiplication by 10 is cheaper than arbitrary integer multiplication
+  i = 10 * i + digit; // might overflow, we will handle the overflow later
+  return true;
+}
+
+simdjson_inline bool is_digit(const uint8_t c) {
+  return static_cast<uint8_t>(c - '0') <= 9;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_decimal_after_separator(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
+  // we continue with the fiction that we have an integer. If the
+  // floating point number is representable as x * 10^z for some integer
+  // z that fits in 53 bits, then we will be able to convert back the
+  // the integer into a float in a lossless manner.
+  const uint8_t *const first_after_period = p;
+
+#ifdef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_SWAR_NUMBER_PARSING
+  // this helps if we have lots of decimals!
+  // this turns out to be frequent enough.
+  if (is_made_of_eight_digits_fast(p)) {
+    i = i * 100000000 + parse_eight_digits_unrolled(p);
+    p += 8;
+  }
+#endif // SIMDJSON_SWAR_NUMBER_PARSING
+#endif // #ifdef SIMDJSON_SWAR_NUMBER_PARSING
+  // Unrolling the first digit makes a small difference on some implementations (e.g. westmere)
+  if (parse_digit(*p, i)) { ++p; }
+  while (parse_digit(*p, i)) { p++; }
+  exponent = first_after_period - p;
+  // Decimal without digits (123.) is illegal
+  if (exponent == 0) {
+    return INVALID_NUMBER(src);
+  }
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
+  // Exp Sign: -123.456e[-]78
+  bool neg_exp = ('-' == *p);
+  if (neg_exp || '+' == *p) { p++; } // Skip + as well
+
+  // Exponent: -123.456e-[78]
+  auto start_exp = p;
+  int64_t exp_number = 0;
+  while (parse_digit(*p, exp_number)) { ++p; }
+  // It is possible for parse_digit to overflow.
+  // In particular, it could overflow to INT64_MIN, and we cannot do - INT64_MIN.
+  // Thus we *must* check for possible overflow before we negate exp_number.
+
+  // Performance notes: it may seem like combining the two "simdjson_unlikely checks" below into
+  // a single simdjson_unlikely path would be faster. The reasoning is sound, but the compiler may
+  // not oblige and may, in fact, generate two distinct paths in any case. It might be
+  // possible to do uint64_t(p - start_exp - 1) >= 18 but it could end up trading off
+  // instructions for a simdjson_likely branch, an unconclusive gain.
+
+  // If there were no digits, it's an error.
+  if (simdjson_unlikely(p == start_exp)) {
+    return INVALID_NUMBER(src);
+  }
+  // We have a valid positive exponent in exp_number at this point, except that
+  // it may have overflowed.
+
+  // If there were more than 18 digits, we may have overflowed the integer. We have to do
+  // something!!!!
+  if (simdjson_unlikely(p > start_exp+18)) {
+    // Skip leading zeroes: 1e000000000000000000001 is technically valid and does not overflow
+    while (*start_exp == '0') { start_exp++; }
+    // 19 digits could overflow int64_t and is kind of absurd anyway. We don't
+    // support exponents smaller than -999,999,999,999,999,999 and bigger
+    // than 999,999,999,999,999,999.
+    // We can truncate.
+    // Note that 999999999999999999 is assuredly too large. The maximal ieee64 value before
+    // infinity is ~1.8e308. The smallest subnormal is ~5e-324. So, actually, we could
+    // truncate at 324.
+    // Note that there is no reason to fail per se at this point in time.
+    // E.g., 0e999999999999999999999 is a fine number.
+    if (p > start_exp+18) { exp_number = 999999999999999999; }
+  }
+  // At this point, we know that exp_number is a sane, positive, signed integer.
+  // It is <= 999,999,999,999,999,999. As long as 'exponent' is in
+  // [-8223372036854775808, 8223372036854775808], we won't overflow. Because 'exponent'
+  // is bounded in magnitude by the size of the JSON input, we are fine in this universe.
+  // To sum it up: the next line should never overflow.
+  exponent += (neg_exp ? -exp_number : exp_number);
+  return SUCCESS;
+}
+
+simdjson_inline bool check_if_integer(const uint8_t *const src, size_t max_length) {
+  const uint8_t *const srcend = src + max_length;
+  bool negative = (*src == '-'); // we can always read at least one character after the '-'
+  const uint8_t *p = src + uint8_t(negative);
+  if(p == srcend) { return false; }
+  if(*p == '0') {
+    ++p;
+    if(p == srcend) { return true; }
+    if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+    return true;
+  }
+  while(p != srcend && is_digit(*p)) { ++p; }
+  if(p == srcend) { return true; }
+  if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+  return true;
+}
+
+simdjson_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
+  // It is possible that the integer had an overflow.
+  // We have to handle the case where we have 0.0000somenumber.
+  const uint8_t *start = start_digits;
+  while ((*start == '0') || (*start == '.')) { ++start; }
+  // we over-decrement by one when there is a '.'
+  return digit_count - size_t(start - start_digits);
+}
+
+} // unnamed namespace
+
+/** @private */
+static error_code slow_float_parsing(simdjson_unused const uint8_t * src, double* answer) {
+  if (parse_float_fallback(src, answer)) {
+    return SUCCESS;
+  }
+  return INVALID_NUMBER(src);
+}
+
+/** @private */
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
+  // If we frequently had to deal with long strings of digits,
+  // we could extend our code by using a 128-bit integer instead
+  // of a 64-bit integer. However, this is uncommon in practice.
+  //
+  // 9999999999999999999 < 2**64 so we can accommodate 19 digits.
+  // If we have a decimal separator, then digit_count - 1 is the number of digits, but we
+  // may not have a decimal separator!
+  if (simdjson_unlikely(digit_count > 19 && significant_digits(start_digits, digit_count) > 19)) {
+    // Ok, chances are good that we had an overflow!
+    // this is almost never going to get called!!!
+    // we start anew, going slowly!!!
+    // This will happen in the following examples:
+    // 10000000000000000000000000000000000000000000e+308
+    // 3.1415926535897932384626433832795028841971693993751
+    //
+    // NOTE: We do not pass a reference to the to slow_float_parsing. If we passed our writer
+    // reference to it, it would force it to be stored in memory, preventing the compiler from
+    // picking it apart and putting into registers. i.e. if we pass it as reference,
+    // it gets slow.
+    double d;
+    error_code error = slow_float_parsing(src, &d);
+    writer.append_double(d);
+    return error;
+  }
+  // NOTE: it's weird that the simdjson_unlikely() only wraps half the if, but it seems to get slower any other
+  // way we've tried: https://github.com/simdjson/simdjson/pull/990#discussion_r448497331
+  // To future reader: we'd love if someone found a better way, or at least could explain this result!
+  if (simdjson_unlikely(exponent < simdjson::internal::smallest_power) || (exponent > simdjson::internal::largest_power)) {
+    //
+    // Important: smallest_power is such that it leads to a zero value.
+    // Observe that 18446744073709551615e-343 == 0, i.e. (2**64 - 1) e -343 is zero
+    // so something x 10^-343 goes to zero, but not so with  something x 10^-342.
+    static_assert(simdjson::internal::smallest_power <= -342, "smallest_power is not small enough");
+    //
+    if((exponent < simdjson::internal::smallest_power) || (i == 0)) {
+      // E.g. Parse "-0.0e-999" into the same value as "-0.0". See https://en.wikipedia.org/wiki/Signed_zero
+      WRITE_DOUBLE(negative ? -0.0 : 0.0, src, writer);
+      return SUCCESS;
+    } else { // (exponent > largest_power) and (i != 0)
+      // We have, for sure, an infinite value and simdjson refuses to parse infinite values.
+      return INVALID_NUMBER(src);
+    }
+  }
+  double d;
+  if (!compute_float_64(exponent, i, negative, d)) {
+    // we are almost never going to get here.
+    if (!parse_float_fallback(src, &d)) { return INVALID_NUMBER(src); }
+  }
+  WRITE_DOUBLE(d, src, writer);
+  return SUCCESS;
+}
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer);
+
+// for performance analysis, it is sometimes  useful to skip parsing
+#ifdef SIMDJSON_SKIPNUMBERPARSING
+
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const, W &writer) {
+  writer.append_s64(0);        // always write zero
+  return SUCCESS;              // always succeeds
+}
+
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept { return number_type::signed_integer; }
+#else
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer) {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  if (digit_count == 0 || ('0' == *start_digits && digit_count > 1)) { return INVALID_NUMBER(src); }
+
+  //
+  // Handle floats if there is a . or e (or both)
+  //
+  int64_t exponent = 0;
+  bool is_float = false;
+  if ('.' == *p) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_decimal_after_separator(src, p, i, exponent) );
+    digit_count = int(p - start_digits); // used later to guard against overflows
+  }
+  if (('e' == *p) || ('E' == *p)) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_exponent(src, p, exponent) );
+  }
+  if (is_float) {
+    const bool dirty_end = jsoncharutils::is_not_structural_or_whitespace(*p);
+    SIMDJSON_TRY( write_float(src, negative, i, start_digits, digit_count, exponent, writer) );
+    if (dirty_end) { return INVALID_NUMBER(src); }
+    return SUCCESS;
+  }
+
+  // The longest negative 64-bit number is 19 digits.
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  size_t longest_digit_count = negative ? 19 : 20;
+  if (digit_count > longest_digit_count) { return BIGINT_NUMBER(src); }
+  if (digit_count == longest_digit_count) {
+    if (negative) {
+      // Anything negative above INT64_MAX+1 is invalid
+      if (i > uint64_t(INT64_MAX)+1) { return BIGINT_NUMBER(src);  }
+      WRITE_INTEGER(~i+1, src, writer);
+      if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+      return SUCCESS;
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    }  else if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INVALID_NUMBER(src); }
+  }
+
+  // Write unsigned if it does not fit in a signed integer.
+  if (i > uint64_t(INT64_MAX)) {
+    WRITE_UNSIGNED(i, src, writer);
+  } else {
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+    if(i == 0 && negative) {
+      // We have to write -0.0 instead of 0
+      WRITE_DOUBLE(-0.0, src, writer);
+    } else {
+      WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+    }
+#else
+  WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+#endif
+  }
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+  return SUCCESS;
+}
+
+// Inlineable functions
+namespace {
+
+// This table can be used to characterize the final character of an integer
+// string. For JSON structural character and allowable white space characters,
+// we return SUCCESS. For 'e', '.' and 'E', we return INCORRECT_TYPE. Otherwise
+// we return NUMBER_ERROR.
+// Optimization note: we could easily reduce the size of the table by half (to 128)
+// at the cost of an extra branch.
+// Optimization note: we want the values to use at most 8 bits (not, e.g., 32 bits):
+static_assert(error_code(uint8_t(NUMBER_ERROR))== NUMBER_ERROR, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(SUCCESS))== SUCCESS, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(INCORRECT_TYPE))== INCORRECT_TYPE, "bad NUMBER_ERROR cast");
+
+const uint8_t integer_string_finisher[256] = {
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   SUCCESS,      NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, INCORRECT_TYPE,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, SUCCESS,        NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR};
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if ((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
+  const uint8_t *p = src + 1;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    // Note: we use src[1] and not src[0] because src[0] is the quote character in this
+    // instance.
+    if (src[1] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  //
+  // Check for minus sign
+  //
+  if(src == src_end) { return NUMBER_ERROR; }
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = src;
+  uint64_t i = 0;
+  while (parse_digit(*src, i)) { src++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(src - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*src)) {
+  //  return (*src == '.' || *src == 'e' || *src == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(*src != '"') { return NUMBER_ERROR; }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept {
+  return (*src == '-');
+}
+
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) { return true; }
+  return false;
+}
+
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  size_t digit_count = size_t(p - src);
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) {
+    static const uint8_t * smaller_big_integer = reinterpret_cast<const uint8_t *>("9223372036854775808");
+    // We have an integer.
+    if(simdjson_unlikely(digit_count > 20)) {
+      return number_type::big_integer;
+    }
+    // If the number is negative and valid, it must be a signed integer.
+    if(negative) {
+      if (simdjson_unlikely(digit_count > 19)) return number_type::big_integer;
+      if (simdjson_unlikely(digit_count == 19 && memcmp(src, smaller_big_integer, 19) > 0)) {
+        return number_type::big_integer;
+      }
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+      if(digit_count == 1 && src[0] == '0') {
+        // We have to write -0.0 instead of 0
+        return number_type::floating_point_number;
+      }
+#endif
+      return number_type::signed_integer;
+    }
+    // Let us check if we have a big integer (>=2**64).
+    static const uint8_t * two_to_sixtyfour = reinterpret_cast<const uint8_t *>("18446744073709551616");
+    if((digit_count > 20) || (digit_count == 20 && memcmp(src, two_to_sixtyfour, 20) >= 0)) {
+      return number_type::big_integer;
+    }
+    // The number is positive and smaller than 18446744073709551616 (or 2**64).
+    // We want values larger or equal to 9223372036854775808 to be unsigned
+    // integers, and the other values to be signed integers.
+    if((digit_count == 20) || (digit_count >= 19 && memcmp(src, smaller_big_integer, 19) >= 0)) {
+      return number_type::unsigned_integer;
+    }
+    return number_type::signed_integer;
+  }
+  // Hopefully, we have 'e' or 'E' or '.'.
+  return number_type::floating_point_number;
+}
+
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
+  if(src == src_end) { return NUMBER_ERROR; }
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  if(p == src_end) { return NUMBER_ERROR; }
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely((p != src_end) && (*p == '.'))) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if ((p == src_end) || !parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while ((p != src_end) && parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = start_digits-src > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if ((p != src_end) && (*p == 'e' || *p == 'E')) {
+    p++;
+    if(p == src_end) { return NUMBER_ERROR; }
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while ((p != src_end) && parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if ((p != src_end) && jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), src_end, &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+} // unnamed namespace
+#endif // SIMDJSON_SKIPNUMBERPARSING
+
+} // namespace numberparsing
+
+inline std::ostream& operator<<(std::ostream& out, number_type type) noexcept {
+    switch (type) {
+        case number_type::signed_integer: out << "integer in [-9223372036854775808,9223372036854775808)"; break;
+        case number_type::unsigned_integer: out << "unsigned integer in [9223372036854775808,18446744073709551616)"; break;
+        case number_type::floating_point_number: out << "floating-point number (binary64)"; break;
+        case number_type::big_integer: out << "big integer"; break;
+        default: SIMDJSON_UNREACHABLE();
+    }
+    return out;
+}
+
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_NUMBERPARSING_H
+/* end file simdjson/generic/numberparsing.h for westmere */
+
+/* including simdjson/generic/implementation_simdjson_result_base-inl.h for westmere: #include "simdjson/generic/implementation_simdjson_result_base-inl.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base-inl.h for westmere */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+
+//
+// internal::implementation_simdjson_result_base<T> inline implementation
+//
+
+template<typename T>
+simdjson_inline void implementation_simdjson_result_base<T>::tie(T &value, error_code &error) && noexcept {
+  error = this->second;
+  if (!error) {
+    value = std::forward<implementation_simdjson_result_base<T>>(*this).first;
+  }
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::get(T &value) && noexcept {
+  error_code error;
+  std::forward<implementation_simdjson_result_base<T>>(*this).tie(value, error);
+  return error;
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::error() const noexcept {
+  return this->second;
+}
+
+
+template<typename T>
+simdjson_warn_unused simdjson_inline bool implementation_simdjson_result_base<T>::has_value() const noexcept {
+  return this->error() == SUCCESS;
+}
+
+#if SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::operator*() &  noexcept(false) {
+  return this->value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::operator*() &&  noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).value();
+}
+
+template<typename T>
+simdjson_inline T* implementation_simdjson_result_base<T>::operator->() noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+
+template<typename T>
+simdjson_inline const T* implementation_simdjson_result_base<T>::operator->() const noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value() & noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::take_value() && noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::operator T&&() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+#endif // SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline const T& implementation_simdjson_result_base<T>::value_unsafe() const& noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value_unsafe() & noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value_unsafe() && noexcept {
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value, error_code error) noexcept
+    : first{std::forward<T>(value)}, second{error} {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(error_code error) noexcept
+    : implementation_simdjson_result_base(T{}, error) {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value) noexcept
+    : implementation_simdjson_result_base(std::forward<T>(value), SUCCESS) {}
+
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+/* end file simdjson/generic/implementation_simdjson_result_base-inl.h for westmere */
+/* end file simdjson/generic/amalgamated.h for westmere */
+/* including simdjson/westmere/end.h: #include "simdjson/westmere/end.h" */
+/* begin file simdjson/westmere/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#if !SIMDJSON_CAN_ALWAYS_RUN_WESTMERE
+SIMDJSON_UNTARGET_REGION
+#endif
+
+/* undefining SIMDJSON_IMPLEMENTATION from "westmere" */
+#undef SIMDJSON_IMPLEMENTATION
+/* end file simdjson/westmere/end.h */
+
+#endif // SIMDJSON_WESTMERE_H
+/* end file simdjson/westmere.h */
+/* including simdjson/westmere/implementation.h: #include <simdjson/westmere/implementation.h> */
+/* begin file simdjson/westmere/implementation.h */
+#ifndef SIMDJSON_WESTMERE_IMPLEMENTATION_H
+#define SIMDJSON_WESTMERE_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/instruction_set.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// The constructor may be executed on any host, so we take care not to use SIMDJSON_TARGET_WESTMERE
+namespace simdjson {
+namespace westmere {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation("westmere", "Intel/AMD SSE4.2", internal::instruction_set::SSE42 | internal::instruction_set::PCLMULQDQ) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_IMPLEMENTATION_H
+/* end file simdjson/westmere/implementation.h */
+
+/* including simdjson/westmere/begin.h: #include <simdjson/westmere/begin.h> */
+/* begin file simdjson/westmere/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "westmere" */
+#define SIMDJSON_IMPLEMENTATION westmere
+/* including simdjson/westmere/base.h: #include "simdjson/westmere/base.h" */
+/* begin file simdjson/westmere/base.h */
+#ifndef SIMDJSON_WESTMERE_BASE_H
+#define SIMDJSON_WESTMERE_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// The constructor may be executed on any host, so we take care not to use SIMDJSON_TARGET_WESTMERE
+namespace simdjson {
+/**
+ * Implementation for Westmere (Intel SSE4.2).
+ */
+namespace westmere {
+
+class implementation;
+
+namespace {
+namespace simd {
+
+template <typename T> struct simd8;
+template <typename T> struct simd8x64;
+
+} // namespace simd
+} // unnamed namespace
+
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_BASE_H
+/* end file simdjson/westmere/base.h */
+/* including simdjson/westmere/intrinsics.h: #include "simdjson/westmere/intrinsics.h" */
+/* begin file simdjson/westmere/intrinsics.h */
+#ifndef SIMDJSON_WESTMERE_INTRINSICS_H
+#define SIMDJSON_WESTMERE_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#if SIMDJSON_VISUAL_STUDIO
+// under clang within visual studio, this will include <x86intrin.h>
+#include <intrin.h> // visual studio or clang
+#else
+#include <x86intrin.h> // elsewhere
+#endif // SIMDJSON_VISUAL_STUDIO
+
+
+#if SIMDJSON_CLANG_VISUAL_STUDIO
+/**
+ * You are not supposed, normally, to include these
+ * headers directly. Instead you should either include intrin.h
+ * or x86intrin.h. However, when compiling with clang
+ * under Windows (i.e., when _MSC_VER is set), these headers
+ * only get included *if* the corresponding features are detected
+ * from macros:
+ */
+#include <smmintrin.h>  // for _mm_alignr_epi8
+#include <wmmintrin.h>  // for  _mm_clmulepi64_si128
+#endif
+
+static_assert(sizeof(__m128i) <= simdjson::SIMDJSON_PADDING, "insufficient padding for westmere");
+
+#endif // SIMDJSON_WESTMERE_INTRINSICS_H
+/* end file simdjson/westmere/intrinsics.h */
+
+#if !SIMDJSON_CAN_ALWAYS_RUN_WESTMERE
+SIMDJSON_TARGET_REGION("sse4.2,pclmul,popcnt")
+#endif
+
+/* including simdjson/westmere/bitmanipulation.h: #include "simdjson/westmere/bitmanipulation.h" */
+/* begin file simdjson/westmere/bitmanipulation.h */
+#ifndef SIMDJSON_WESTMERE_BITMANIPULATION_H
+#define SIMDJSON_WESTMERE_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  unsigned long ret;
+  // Search the mask data from least significant bit (LSB)
+  // to the most significant bit (MSB) for a set bit (1).
+  _BitScanForward64(&ret, input_num);
+  return (int)ret;
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO
+  return __builtin_ctzll(input_num);
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return input_num & (input_num-1);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  unsigned long leading_zero = 0;
+  // Search the mask data from most significant bit (MSB)
+  // to least significant bit (LSB) for a set bit (1).
+  if (_BitScanReverse64(&leading_zero, input_num))
+    return (int)(63 - leading_zero);
+  else
+    return 64;
+#else
+  return __builtin_clzll(input_num);
+#endif// SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+simdjson_inline unsigned __int64 count_ones(uint64_t input_num) {
+  // note: we do not support legacy 32-bit Windows in this kernel
+  return __popcnt64(input_num);// Visual Studio wants two underscores
+}
+#else
+simdjson_inline long long int count_ones(uint64_t input_num) {
+  return _popcnt64(input_num);
+}
+#endif
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2,
+                                uint64_t *result) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  return _addcarry_u64(0, value1, value2,
+                       reinterpret_cast<unsigned __int64 *>(result));
+#else
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+#endif
+}
+
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_BITMANIPULATION_H
+/* end file simdjson/westmere/bitmanipulation.h */
+/* including simdjson/westmere/bitmask.h: #include "simdjson/westmere/bitmask.h" */
+/* begin file simdjson/westmere/bitmask.h */
+#ifndef SIMDJSON_WESTMERE_BITMASK_H
+#define SIMDJSON_WESTMERE_BITMASK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(const uint64_t bitmask) {
+  // There should be no such thing with a processing supporting avx2
+  // but not clmul.
+  __m128i all_ones = _mm_set1_epi8('\xFF');
+  __m128i result = _mm_clmulepi64_si128(_mm_set_epi64x(0ULL, bitmask), all_ones, 0);
+  return _mm_cvtsi128_si64(result);
+}
+
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_BITMASK_H
+/* end file simdjson/westmere/bitmask.h */
+/* including simdjson/westmere/numberparsing_defs.h: #include "simdjson/westmere/numberparsing_defs.h" */
+/* begin file simdjson/westmere/numberparsing_defs.h */
+#ifndef SIMDJSON_WESTMERE_NUMBERPARSING_DEFS_H
+#define SIMDJSON_WESTMERE_NUMBERPARSING_DEFS_H
+
+/* including simdjson/westmere/base.h: #include "simdjson/westmere/base.h" */
+/* begin file simdjson/westmere/base.h */
+#ifndef SIMDJSON_WESTMERE_BASE_H
+#define SIMDJSON_WESTMERE_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// The constructor may be executed on any host, so we take care not to use SIMDJSON_TARGET_WESTMERE
+namespace simdjson {
+/**
+ * Implementation for Westmere (Intel SSE4.2).
+ */
+namespace westmere {
+
+class implementation;
+
+namespace {
+namespace simd {
+
+template <typename T> struct simd8;
+template <typename T> struct simd8x64;
+
+} // namespace simd
+} // unnamed namespace
+
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_BASE_H
+/* end file simdjson/westmere/base.h */
+/* including simdjson/westmere/intrinsics.h: #include "simdjson/westmere/intrinsics.h" */
+/* begin file simdjson/westmere/intrinsics.h */
+#ifndef SIMDJSON_WESTMERE_INTRINSICS_H
+#define SIMDJSON_WESTMERE_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#if SIMDJSON_VISUAL_STUDIO
+// under clang within visual studio, this will include <x86intrin.h>
+#include <intrin.h> // visual studio or clang
+#else
+#include <x86intrin.h> // elsewhere
+#endif // SIMDJSON_VISUAL_STUDIO
+
+
+#if SIMDJSON_CLANG_VISUAL_STUDIO
+/**
+ * You are not supposed, normally, to include these
+ * headers directly. Instead you should either include intrin.h
+ * or x86intrin.h. However, when compiling with clang
+ * under Windows (i.e., when _MSC_VER is set), these headers
+ * only get included *if* the corresponding features are detected
+ * from macros:
+ */
+#include <smmintrin.h>  // for _mm_alignr_epi8
+#include <wmmintrin.h>  // for  _mm_clmulepi64_si128
+#endif
+
+static_assert(sizeof(__m128i) <= simdjson::SIMDJSON_PADDING, "insufficient padding for westmere");
+
+#endif // SIMDJSON_WESTMERE_INTRINSICS_H
+/* end file simdjson/westmere/intrinsics.h */
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace numberparsing {
+
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  // this actually computes *16* values so we are being wasteful.
+  const __m128i ascii0 = _mm_set1_epi8('0');
+  const __m128i mul_1_10 =
+      _mm_setr_epi8(10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1, 10, 1);
+  const __m128i mul_1_100 = _mm_setr_epi16(100, 1, 100, 1, 100, 1, 100, 1);
+  const __m128i mul_1_10000 =
+      _mm_setr_epi16(10000, 1, 10000, 1, 10000, 1, 10000, 1);
+  const __m128i input = _mm_sub_epi8(
+      _mm_loadu_si128(reinterpret_cast<const __m128i *>(chars)), ascii0);
+  const __m128i t1 = _mm_maddubs_epi16(input, mul_1_10);
+  const __m128i t2 = _mm_madd_epi16(t1, mul_1_100);
+  const __m128i t3 = _mm_packus_epi32(t2, t2);
+  const __m128i t4 = _mm_madd_epi16(t3, mul_1_10000);
+  return _mm_cvtsi128_si32(
+      t4); // only captures the sum of the first 8 digits, drop the rest
+}
+
+/** @private */
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+#if SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+#if SIMDJSON_IS_ARM64
+  // ARM64 has native support for 64-bit multiplications, no need to emultate
+  answer.high = __umulh(value1, value2);
+  answer.low = value1 * value2;
+#else
+  answer.low = _umul128(value1, value2, &answer.high); // _umul128 not available on ARM64
+#endif // SIMDJSON_IS_ARM64
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+#endif
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace westmere
+} // namespace simdjson
+
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+
+#endif //  SIMDJSON_WESTMERE_NUMBERPARSING_DEFS_H
+/* end file simdjson/westmere/numberparsing_defs.h */
+/* including simdjson/westmere/simd.h: #include "simdjson/westmere/simd.h" */
+/* begin file simdjson/westmere/simd.h */
+#ifndef SIMDJSON_WESTMERE_SIMD_H
+#define SIMDJSON_WESTMERE_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace simd {
+
+  template<typename Child>
+  struct base {
+    __m128i value;
+
+    // Zero constructor
+    simdjson_inline base() : value{__m128i()} {}
+
+    // Conversion from SIMD register
+    simdjson_inline base(const __m128i _value) : value(_value) {}
+
+    // Conversion to SIMD register
+    simdjson_inline operator const __m128i&() const { return this->value; }
+    simdjson_inline operator __m128i&() { return this->value; }
+
+    // Bit operations
+    simdjson_inline Child operator|(const Child other) const { return _mm_or_si128(*this, other); }
+    simdjson_inline Child operator&(const Child other) const { return _mm_and_si128(*this, other); }
+    simdjson_inline Child operator^(const Child other) const { return _mm_xor_si128(*this, other); }
+    simdjson_inline Child bit_andnot(const Child other) const { return _mm_andnot_si128(other, *this); }
+    simdjson_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+  };
+
+  template<typename T, typename Mask=simd8<bool>>
+  struct base8: base<simd8<T>> {
+    typedef uint16_t bitmask_t;
+    typedef uint32_t bitmask2_t;
+
+    simdjson_inline base8() : base<simd8<T>>() {}
+    simdjson_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}
+
+    friend simdjson_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return _mm_cmpeq_epi8(lhs, rhs); }
+
+    static const int SIZE = sizeof(base<simd8<T>>::value);
+
+    template<int N=1>
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+      return _mm_alignr_epi8(*this, prev_chunk, 16 - N);
+    }
+  };
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool>: base8<bool> {
+    static simdjson_inline simd8<bool> splat(bool _value) { return _mm_set1_epi8(uint8_t(-(!!_value))); }
+
+    simdjson_inline simd8() : base8() {}
+    simdjson_inline simd8(const __m128i _value) : base8<bool>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+
+    simdjson_inline int to_bitmask() const { return _mm_movemask_epi8(*this); }
+    simdjson_inline bool any() const { return !_mm_testz_si128(*this, *this); }
+    simdjson_inline simd8<bool> operator~() const { return *this ^ true; }
+  };
+
+  template<typename T>
+  struct base8_numeric: base8<T> {
+    static simdjson_inline simd8<T> splat(T _value) { return _mm_set1_epi8(_value); }
+    static simdjson_inline simd8<T> zero() { return _mm_setzero_si128(); }
+    static simdjson_inline simd8<T> load(const T values[16]) {
+      return _mm_loadu_si128(reinterpret_cast<const __m128i *>(values));
+    }
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    static simdjson_inline simd8<T> repeat_16(
+      T v0,  T v1,  T v2,  T v3,  T v4,  T v5,  T v6,  T v7,
+      T v8,  T v9,  T v10, T v11, T v12, T v13, T v14, T v15
+    ) {
+      return simd8<T>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    simdjson_inline base8_numeric() : base8<T>() {}
+    simdjson_inline base8_numeric(const __m128i _value) : base8<T>(_value) {}
+
+    // Store to array
+    simdjson_inline void store(T dst[16]) const { return _mm_storeu_si128(reinterpret_cast<__m128i *>(dst), *this); }
+
+    // Override to distinguish from bool version
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<T> operator+(const simd8<T> other) const { return _mm_add_epi8(*this, other); }
+    simdjson_inline simd8<T> operator-(const simd8<T> other) const { return _mm_sub_epi8(*this, other); }
+    simdjson_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return _mm_shuffle_epi8(lookup_table, *this);
+    }
+
+    // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted as a bitset).
+    // Passing a 0 value for mask would be equivalent to writing out every byte to output.
+    // Only the first 16 - count_ones(mask) bytes of the result are significant but 16 bytes
+    // get written.
+    // Design consideration: it seems like a function with the
+    // signature simd8<L> compress(uint32_t mask) would be
+    // sensible, but the AVX ISA makes this kind of approach difficult.
+    template<typename L>
+    simdjson_inline void compress(uint16_t mask, L * output) const {
+      using internal::thintable_epi8;
+      using internal::BitsSetTable256mul2;
+      using internal::pshufb_combine_table;
+      // this particular implementation was inspired by work done by @animetosho
+      // we do it in two steps, first 8 bytes and then second 8 bytes
+      uint8_t mask1 = uint8_t(mask); // least significant 8 bits
+      uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
+      // next line just loads the 64-bit values thintable_epi8[mask1] and
+      // thintable_epi8[mask2] into a 128-bit register, using only
+      // two instructions on most compilers.
+      __m128i shufmask =  _mm_set_epi64x(thintable_epi8[mask2], thintable_epi8[mask1]);
+      // we increment by 0x08 the second half of the mask
+      shufmask =
+      _mm_add_epi8(shufmask, _mm_set_epi32(0x08080808, 0x08080808, 0, 0));
+      // this is the version "nearly pruned"
+      __m128i pruned = _mm_shuffle_epi8(*this, shufmask);
+      // we still need to put the two halves together.
+      // we compute the popcount of the first half:
+      int pop1 = BitsSetTable256mul2[mask1];
+      // then load the corresponding mask, what it does is to write
+      // only the first pop1 bytes from the first 8 bytes, and then
+      // it fills in with the bytes from the second 8 bytes + some filling
+      // at the end.
+      __m128i compactmask =
+      _mm_loadu_si128(reinterpret_cast<const __m128i *>(pshufb_combine_table + pop1 * 8));
+      __m128i answer = _mm_shuffle_epi8(pruned, compactmask);
+      _mm_storeu_si128(reinterpret_cast<__m128i *>(output), answer);
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> : base8_numeric<int8_t> {
+    simdjson_inline simd8() : base8_numeric<int8_t>() {}
+    simdjson_inline simd8(const __m128i _value) : base8_numeric<int8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const int8_t* values) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) : simd8(_mm_setr_epi8(
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    )) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm_max_epi8(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm_min_epi8(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm_cmpgt_epi8(*this, other); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return _mm_cmpgt_epi8(other, *this); }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t>: base8_numeric<uint8_t> {
+    simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+    simdjson_inline simd8(const __m128i _value) : base8_numeric<uint8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const uint8_t* values) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) : simd8(_mm_setr_epi8(
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    )) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm_adds_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm_subs_epu8(*this, other); }
+
+    // Order-specific operations
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm_max_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return _mm_min_epu8(*this, other); }
+    // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
+    // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+
+    // Bit-specific operations
+    simdjson_inline simd8<bool> bits_not_set() const { return *this == uint8_t(0); }
+    simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
+    simdjson_inline bool is_ascii() const { return _mm_movemask_epi8(*this) == 0; }
+    simdjson_inline bool bits_not_set_anywhere() const { return _mm_testz_si128(*this, *this); }
+    simdjson_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
+    simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return _mm_testz_si128(*this, bits); }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(_mm_slli_epi16(*this, N)) & uint8_t(0xFFu << N); }
+    // Get one of the bits and make a bitmask out of it.
+    // e.g. value.get_bit<7>() gets the high bit
+    template<int N>
+    simdjson_inline int get_bit() const { return _mm_movemask_epi8(_mm_slli_epi16(*this, 7-N)); }
+  };
+
+  template<typename T>
+  struct simd8x64 {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+    static_assert(NUM_CHUNKS == 4, "Westmere kernel should use four registers per 64-byte block.");
+    const simd8<T> chunks[NUM_CHUNKS];
+    template<int idx> simd8<uint8_t> get() const { return idx < NUM_CHUNKS ? chunks[idx] : simd8<T>(); }
+
+    simd8x64(const simd8x64<T>& o) = delete; // no copy allowed
+    simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1, const simd8<T> chunk2, const simd8<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+16), simd8<T>::load(ptr+32), simd8<T>::load(ptr+48)} {}
+
+    simdjson_inline void store(T ptr[64]) const {
+      this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
+      this->chunks[1].store(ptr+sizeof(simd8<T>)*1);
+      this->chunks[2].store(ptr+sizeof(simd8<T>)*2);
+      this->chunks[3].store(ptr+sizeof(simd8<T>)*3);
+    }
+
+    simdjson_inline simd8<T> reduce_or() const {
+      return (this->chunks[0] | this->chunks[1]) | (this->chunks[2] | this->chunks[3]);
+    }
+
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
+      this->chunks[0].compress(uint16_t(mask), output);
+      this->chunks[1].compress(uint16_t(mask >> 16), output + 16 - count_ones(mask & 0xFFFF));
+      this->chunks[2].compress(uint16_t(mask >> 32), output + 32 - count_ones(mask & 0xFFFFFFFF));
+      this->chunks[3].compress(uint16_t(mask >> 48), output + 48 - count_ones(mask & 0xFFFFFFFFFFFF));
+      return 64 - count_ones(mask);
+    }
+
+    simdjson_inline uint64_t to_bitmask() const {
+      uint64_t r0 = uint32_t(this->chunks[0].to_bitmask() );
+      uint64_t r1 =          this->chunks[1].to_bitmask() ;
+      uint64_t r2 =          this->chunks[2].to_bitmask() ;
+      uint64_t r3 =          this->chunks[3].to_bitmask() ;
+      return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
+    }
+
+    simdjson_inline uint64_t eq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] == mask,
+        this->chunks[1] == mask,
+        this->chunks[2] == mask,
+        this->chunks[3] == mask
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+      return  simd8x64<bool>(
+        this->chunks[0] == other.chunks[0],
+        this->chunks[1] == other.chunks[1],
+        this->chunks[2] == other.chunks[2],
+        this->chunks[3] == other.chunks[3]
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t lteq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] <= mask,
+        this->chunks[1] <= mask,
+        this->chunks[2] <= mask,
+        this->chunks[3] <= mask
+      ).to_bitmask();
+    }
+  }; // struct simd8x64<T>
+
+} // namespace simd
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_SIMD_INPUT_H
+/* end file simdjson/westmere/simd.h */
+/* including simdjson/westmere/stringparsing_defs.h: #include "simdjson/westmere/stringparsing_defs.h" */
+/* begin file simdjson/westmere/stringparsing_defs.h */
+#ifndef SIMDJSON_WESTMERE_STRINGPARSING_DEFS_H
+#define SIMDJSON_WESTMERE_STRINGPARSING_DEFS_H
+
+/* including simdjson/westmere/bitmanipulation.h: #include "simdjson/westmere/bitmanipulation.h" */
+/* begin file simdjson/westmere/bitmanipulation.h */
+#ifndef SIMDJSON_WESTMERE_BITMANIPULATION_H
+#define SIMDJSON_WESTMERE_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  unsigned long ret;
+  // Search the mask data from least significant bit (LSB)
+  // to the most significant bit (MSB) for a set bit (1).
+  _BitScanForward64(&ret, input_num);
+  return (int)ret;
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO
+  return __builtin_ctzll(input_num);
+#endif // SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return input_num & (input_num-1);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  unsigned long leading_zero = 0;
+  // Search the mask data from most significant bit (MSB)
+  // to least significant bit (LSB) for a set bit (1).
+  if (_BitScanReverse64(&leading_zero, input_num))
+    return (int)(63 - leading_zero);
+  else
+    return 64;
+#else
+  return __builtin_clzll(input_num);
+#endif// SIMDJSON_REGULAR_VISUAL_STUDIO
+}
+
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+simdjson_inline unsigned __int64 count_ones(uint64_t input_num) {
+  // note: we do not support legacy 32-bit Windows in this kernel
+  return __popcnt64(input_num);// Visual Studio wants two underscores
+}
+#else
+simdjson_inline long long int count_ones(uint64_t input_num) {
+  return _popcnt64(input_num);
+}
+#endif
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2,
+                                uint64_t *result) {
+#if SIMDJSON_REGULAR_VISUAL_STUDIO
+  return _addcarry_u64(0, value1, value2,
+                       reinterpret_cast<unsigned __int64 *>(result));
+#else
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+#endif
+}
+
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_BITMANIPULATION_H
+/* end file simdjson/westmere/bitmanipulation.h */
+/* including simdjson/westmere/simd.h: #include "simdjson/westmere/simd.h" */
+/* begin file simdjson/westmere/simd.h */
+#ifndef SIMDJSON_WESTMERE_SIMD_H
+#define SIMDJSON_WESTMERE_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace simd {
+
+  template<typename Child>
+  struct base {
+    __m128i value;
+
+    // Zero constructor
+    simdjson_inline base() : value{__m128i()} {}
+
+    // Conversion from SIMD register
+    simdjson_inline base(const __m128i _value) : value(_value) {}
+
+    // Conversion to SIMD register
+    simdjson_inline operator const __m128i&() const { return this->value; }
+    simdjson_inline operator __m128i&() { return this->value; }
+
+    // Bit operations
+    simdjson_inline Child operator|(const Child other) const { return _mm_or_si128(*this, other); }
+    simdjson_inline Child operator&(const Child other) const { return _mm_and_si128(*this, other); }
+    simdjson_inline Child operator^(const Child other) const { return _mm_xor_si128(*this, other); }
+    simdjson_inline Child bit_andnot(const Child other) const { return _mm_andnot_si128(other, *this); }
+    simdjson_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+  };
+
+  template<typename T, typename Mask=simd8<bool>>
+  struct base8: base<simd8<T>> {
+    typedef uint16_t bitmask_t;
+    typedef uint32_t bitmask2_t;
+
+    simdjson_inline base8() : base<simd8<T>>() {}
+    simdjson_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}
+
+    friend simdjson_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return _mm_cmpeq_epi8(lhs, rhs); }
+
+    static const int SIZE = sizeof(base<simd8<T>>::value);
+
+    template<int N=1>
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+      return _mm_alignr_epi8(*this, prev_chunk, 16 - N);
+    }
+  };
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool>: base8<bool> {
+    static simdjson_inline simd8<bool> splat(bool _value) { return _mm_set1_epi8(uint8_t(-(!!_value))); }
+
+    simdjson_inline simd8() : base8() {}
+    simdjson_inline simd8(const __m128i _value) : base8<bool>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+
+    simdjson_inline int to_bitmask() const { return _mm_movemask_epi8(*this); }
+    simdjson_inline bool any() const { return !_mm_testz_si128(*this, *this); }
+    simdjson_inline simd8<bool> operator~() const { return *this ^ true; }
+  };
+
+  template<typename T>
+  struct base8_numeric: base8<T> {
+    static simdjson_inline simd8<T> splat(T _value) { return _mm_set1_epi8(_value); }
+    static simdjson_inline simd8<T> zero() { return _mm_setzero_si128(); }
+    static simdjson_inline simd8<T> load(const T values[16]) {
+      return _mm_loadu_si128(reinterpret_cast<const __m128i *>(values));
+    }
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    static simdjson_inline simd8<T> repeat_16(
+      T v0,  T v1,  T v2,  T v3,  T v4,  T v5,  T v6,  T v7,
+      T v8,  T v9,  T v10, T v11, T v12, T v13, T v14, T v15
+    ) {
+      return simd8<T>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    simdjson_inline base8_numeric() : base8<T>() {}
+    simdjson_inline base8_numeric(const __m128i _value) : base8<T>(_value) {}
+
+    // Store to array
+    simdjson_inline void store(T dst[16]) const { return _mm_storeu_si128(reinterpret_cast<__m128i *>(dst), *this); }
+
+    // Override to distinguish from bool version
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<T> operator+(const simd8<T> other) const { return _mm_add_epi8(*this, other); }
+    simdjson_inline simd8<T> operator-(const simd8<T> other) const { return _mm_sub_epi8(*this, other); }
+    simdjson_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return _mm_shuffle_epi8(lookup_table, *this);
+    }
+
+    // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted as a bitset).
+    // Passing a 0 value for mask would be equivalent to writing out every byte to output.
+    // Only the first 16 - count_ones(mask) bytes of the result are significant but 16 bytes
+    // get written.
+    // Design consideration: it seems like a function with the
+    // signature simd8<L> compress(uint32_t mask) would be
+    // sensible, but the AVX ISA makes this kind of approach difficult.
+    template<typename L>
+    simdjson_inline void compress(uint16_t mask, L * output) const {
+      using internal::thintable_epi8;
+      using internal::BitsSetTable256mul2;
+      using internal::pshufb_combine_table;
+      // this particular implementation was inspired by work done by @animetosho
+      // we do it in two steps, first 8 bytes and then second 8 bytes
+      uint8_t mask1 = uint8_t(mask); // least significant 8 bits
+      uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
+      // next line just loads the 64-bit values thintable_epi8[mask1] and
+      // thintable_epi8[mask2] into a 128-bit register, using only
+      // two instructions on most compilers.
+      __m128i shufmask =  _mm_set_epi64x(thintable_epi8[mask2], thintable_epi8[mask1]);
+      // we increment by 0x08 the second half of the mask
+      shufmask =
+      _mm_add_epi8(shufmask, _mm_set_epi32(0x08080808, 0x08080808, 0, 0));
+      // this is the version "nearly pruned"
+      __m128i pruned = _mm_shuffle_epi8(*this, shufmask);
+      // we still need to put the two halves together.
+      // we compute the popcount of the first half:
+      int pop1 = BitsSetTable256mul2[mask1];
+      // then load the corresponding mask, what it does is to write
+      // only the first pop1 bytes from the first 8 bytes, and then
+      // it fills in with the bytes from the second 8 bytes + some filling
+      // at the end.
+      __m128i compactmask =
+      _mm_loadu_si128(reinterpret_cast<const __m128i *>(pshufb_combine_table + pop1 * 8));
+      __m128i answer = _mm_shuffle_epi8(pruned, compactmask);
+      _mm_storeu_si128(reinterpret_cast<__m128i *>(output), answer);
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> : base8_numeric<int8_t> {
+    simdjson_inline simd8() : base8_numeric<int8_t>() {}
+    simdjson_inline simd8(const __m128i _value) : base8_numeric<int8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const int8_t* values) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) : simd8(_mm_setr_epi8(
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    )) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm_max_epi8(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm_min_epi8(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm_cmpgt_epi8(*this, other); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return _mm_cmpgt_epi8(other, *this); }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t>: base8_numeric<uint8_t> {
+    simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+    simdjson_inline simd8(const __m128i _value) : base8_numeric<uint8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const uint8_t* values) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) : simd8(_mm_setr_epi8(
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    )) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm_adds_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm_subs_epu8(*this, other); }
+
+    // Order-specific operations
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm_max_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return _mm_min_epu8(*this, other); }
+    // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
+    // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+
+    // Bit-specific operations
+    simdjson_inline simd8<bool> bits_not_set() const { return *this == uint8_t(0); }
+    simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
+    simdjson_inline bool is_ascii() const { return _mm_movemask_epi8(*this) == 0; }
+    simdjson_inline bool bits_not_set_anywhere() const { return _mm_testz_si128(*this, *this); }
+    simdjson_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
+    simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return _mm_testz_si128(*this, bits); }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(_mm_slli_epi16(*this, N)) & uint8_t(0xFFu << N); }
+    // Get one of the bits and make a bitmask out of it.
+    // e.g. value.get_bit<7>() gets the high bit
+    template<int N>
+    simdjson_inline int get_bit() const { return _mm_movemask_epi8(_mm_slli_epi16(*this, 7-N)); }
+  };
+
+  template<typename T>
+  struct simd8x64 {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+    static_assert(NUM_CHUNKS == 4, "Westmere kernel should use four registers per 64-byte block.");
+    const simd8<T> chunks[NUM_CHUNKS];
+    template<int idx> simd8<uint8_t> get() const { return idx < NUM_CHUNKS ? chunks[idx] : simd8<T>(); }
+
+    simd8x64(const simd8x64<T>& o) = delete; // no copy allowed
+    simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1, const simd8<T> chunk2, const simd8<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+16), simd8<T>::load(ptr+32), simd8<T>::load(ptr+48)} {}
+
+    simdjson_inline void store(T ptr[64]) const {
+      this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
+      this->chunks[1].store(ptr+sizeof(simd8<T>)*1);
+      this->chunks[2].store(ptr+sizeof(simd8<T>)*2);
+      this->chunks[3].store(ptr+sizeof(simd8<T>)*3);
+    }
+
+    simdjson_inline simd8<T> reduce_or() const {
+      return (this->chunks[0] | this->chunks[1]) | (this->chunks[2] | this->chunks[3]);
+    }
+
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
+      this->chunks[0].compress(uint16_t(mask), output);
+      this->chunks[1].compress(uint16_t(mask >> 16), output + 16 - count_ones(mask & 0xFFFF));
+      this->chunks[2].compress(uint16_t(mask >> 32), output + 32 - count_ones(mask & 0xFFFFFFFF));
+      this->chunks[3].compress(uint16_t(mask >> 48), output + 48 - count_ones(mask & 0xFFFFFFFFFFFF));
+      return 64 - count_ones(mask);
+    }
+
+    simdjson_inline uint64_t to_bitmask() const {
+      uint64_t r0 = uint32_t(this->chunks[0].to_bitmask() );
+      uint64_t r1 =          this->chunks[1].to_bitmask() ;
+      uint64_t r2 =          this->chunks[2].to_bitmask() ;
+      uint64_t r3 =          this->chunks[3].to_bitmask() ;
+      return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
+    }
+
+    simdjson_inline uint64_t eq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] == mask,
+        this->chunks[1] == mask,
+        this->chunks[2] == mask,
+        this->chunks[3] == mask
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+      return  simd8x64<bool>(
+        this->chunks[0] == other.chunks[0],
+        this->chunks[1] == other.chunks[1],
+        this->chunks[2] == other.chunks[2],
+        this->chunks[3] == other.chunks[3]
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t lteq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] <= mask,
+        this->chunks[1] <= mask,
+        this->chunks[2] <= mask,
+        this->chunks[3] <= mask
+      ).to_bitmask();
+    }
+  }; // struct simd8x64<T>
+
+} // namespace simd
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_SIMD_INPUT_H
+/* end file simdjson/westmere/simd.h */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 32;
+  simdjson_inline backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return bs_bits != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+
+  uint32_t bs_bits;
+  uint32_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // this can read up to 31 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v0(src);
+  simd8<uint8_t> v1(src + 16);
+  v0.store(dst);
+  v1.store(dst + 16);
+  uint64_t bs_and_quote = simd8x64<bool>(v0 == '\\', v1 == '\\', v0 == '"', v1 == '"').to_bitmask();
+  return {
+    uint32_t(bs_and_quote),      // bs_bits
+    uint32_t(bs_and_quote >> 32) // quote_bits
+  };
+}
+
+
+struct escaping {
+  static constexpr uint32_t BYTES_PROCESSED = 16;
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits != 0; }
+  simdjson_inline int escape_index() { return trailing_zeroes(escape_bits); }
+
+  uint64_t escape_bits;
+}; // struct escaping
+
+
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "escaping finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  simd8<bool> is_quote = (v == '"');
+  simd8<bool> is_backslash = (v == '\\');
+  simd8<bool> is_control = (v < 32);
+  return {
+    uint64_t((is_backslash | is_quote | is_control).to_bitmask())
+  };
+}
+
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_WESTMERE_STRINGPARSING_DEFS_H
+/* end file simdjson/westmere/stringparsing_defs.h */
+/* end file simdjson/westmere/begin.h */
+/* including generic/amalgamated.h for westmere: #include <generic/amalgamated.h> */
+/* begin file generic/amalgamated.h for westmere */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_SRC_GENERIC_DEPENDENCIES_H)
+#error generic/dependencies.h must be included before generic/amalgamated.h!
+#endif
+
+/* including generic/base.h for westmere: #include <generic/base.h> */
+/* begin file generic/base.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+
+struct json_character_block;
+
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_BASE_H
+/* end file generic/base.h for westmere */
+/* including generic/dom_parser_implementation.h for westmere: #include <generic/dom_parser_implementation.h> */
+/* begin file generic/dom_parser_implementation.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// Interface a dom parser implementation must fulfill
+namespace simdjson {
+namespace westmere {
+namespace {
+
+simdjson_inline simd8<uint8_t> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3);
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input);
+
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file generic/dom_parser_implementation.h for westmere */
+/* including generic/json_character_block.h for westmere: #include <generic/json_character_block.h> */
+/* begin file generic/json_character_block.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+
+struct json_character_block {
+  static simdjson_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
+
+  simdjson_inline uint64_t whitespace() const noexcept { return _whitespace; }
+  simdjson_inline uint64_t op() const noexcept { return _op; }
+  simdjson_inline uint64_t scalar() const noexcept { return ~(op() | whitespace()); }
+
+  uint64_t _whitespace;
+  uint64_t _op;
+};
+
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H
+/* end file generic/json_character_block.h for westmere */
+/* end file generic/amalgamated.h for westmere */
+/* including generic/stage1/amalgamated.h for westmere: #include <generic/stage1/amalgamated.h> */
+/* begin file generic/stage1/amalgamated.h for westmere */
+// Stuff other things depend on
+/* including generic/stage1/base.h for westmere: #include <generic/stage1/base.h> */
+/* begin file generic/stage1/base.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_BASE_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace stage1 {
+
+class bit_indexer;
+template<size_t STEP_SIZE>
+struct buf_block_reader;
+struct json_block;
+class json_minifier;
+class json_scanner;
+struct json_string_block;
+class json_string_scanner;
+class json_structural_indexer;
+
+} // namespace stage1
+
+namespace utf8_validation {
+struct utf8_checker;
+} // namespace utf8_validation
+
+using utf8_validation::utf8_checker;
+
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_BASE_H
+/* end file generic/stage1/base.h for westmere */
+/* including generic/stage1/buf_block_reader.h for westmere: #include <generic/stage1/buf_block_reader.h> */
+/* begin file generic/stage1/buf_block_reader.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace stage1 {
+
+// Walks through a buffer in block-sized increments, loading the last part with spaces
+template<size_t STEP_SIZE>
+struct buf_block_reader {
+public:
+  simdjson_inline buf_block_reader(const uint8_t *_buf, size_t _len);
+  simdjson_inline size_t block_index();
+  simdjson_inline bool has_full_block() const;
+  simdjson_inline const uint8_t *full_block() const;
+  /**
+   * Get the last block, padded with spaces.
+   *
+   * There will always be a last block, with at least 1 byte, unless len == 0 (in which case this
+   * function fills the buffer with spaces and returns 0. In particular, if len == STEP_SIZE there
+   * will be 0 full_blocks and 1 remainder block with STEP_SIZE bytes and no spaces for padding.
+   *
+   * @return the number of effective characters in the last block.
+   */
+  simdjson_inline size_t get_remainder(uint8_t *dst) const;
+  simdjson_inline void advance();
+private:
+  const uint8_t *buf;
+  const size_t len;
+  const size_t lenminusstep;
+  size_t idx;
+};
+
+// Routines to print masks and text for debugging bitmask operations
+simdjson_unused static char * format_input_text_64(const uint8_t *text) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    buf[i] = int8_t(text[i]) < ' ' ? '_' : int8_t(text[i]);
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+// Routines to print masks and text for debugging bitmask operations
+simdjson_unused static char * format_input_text(const simd8x64<uint8_t>& in) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  in.store(reinterpret_cast<uint8_t*>(buf));
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] < ' ') { buf[i] = '_'; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+simdjson_unused static char * format_input_text(const simd8x64<uint8_t>& in, uint64_t mask) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  in.store(reinterpret_cast<uint8_t*>(buf));
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] <= ' ') { buf[i] = '_'; }
+    if (!(mask & (size_t(1) << i))) { buf[i] = ' '; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+simdjson_unused static char * format_mask(uint64_t mask) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  for (size_t i=0; i<64; i++) {
+    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
+  }
+  buf[64] = '\0';
+  return buf;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
+
+template<size_t STEP_SIZE>
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
+
+template<size_t STEP_SIZE>
+simdjson_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
+  return idx < lenminusstep;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
+  return &buf[idx];
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
+  if(len == idx) { return 0; } // memcpy(dst, null, 0) will trigger an error with some sanitizers
+  std::memset(dst, 0x20, STEP_SIZE); // std::memset STEP_SIZE because it's more efficient to write out 8 or 16 bytes at once.
+  std::memcpy(dst, buf + idx, len - idx);
+  return len - idx;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline void buf_block_reader<STEP_SIZE>::advance() {
+  idx += STEP_SIZE;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H
+/* end file generic/stage1/buf_block_reader.h for westmere */
+/* including generic/stage1/json_escape_scanner.h for westmere: #include <generic/stage1/json_escape_scanner.h> */
+/* begin file generic/stage1/json_escape_scanner.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_ESCAPE_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_ESCAPE_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace stage1 {
+
+/**
+ * Scans for escape characters in JSON, taking care with multiple backslashes (\\n vs. \n).
+ */
+struct json_escape_scanner {
+  /** The actual escape characters (the backslashes themselves). */
+  uint64_t next_is_escaped = 0ULL;
+
+  struct escaped_and_escape {
+    /**
+     * Mask of escaped characters.
+     *
+     * ```
+     * \n \\n \\\n \\\\n \
+     * 0100100010100101000
+     *  n  \   \ n  \ \
+     * ```
+     */
+    uint64_t escaped;
+    /**
+     * Mask of escape characters.
+     *
+     * ```
+     * \n \\n \\\n \\\\n \
+     * 1001000101001010001
+     * \  \   \ \  \ \   \
+     * ```
+     */
+    uint64_t escape;
+  };
+
+  /**
+   * Get a mask of both escape and escaped characters (the characters following a backslash).
+   *
+   * @param potential_escape A mask of the character that can escape others (but could be
+   *        escaped itself). e.g. block.eq('\\')
+   */
+  simdjson_really_inline escaped_and_escape next(uint64_t backslash) noexcept {
+
+#if !SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+    if (!backslash) { return {next_escaped_without_backslashes(), 0}; }
+#endif
+
+    // |                                | Mask (shows characters instead of 1's) | Depth | Instructions        |
+    // |--------------------------------|----------------------------------------|-------|---------------------|
+    // | string                         | `\\n_\\\n___\\\n___\\\\___\\\\__\\\`   |       |                     |
+    // |                                | `    even   odd    even   odd   odd`   |       |                     |
+    // | potential_escape               | ` \  \\\    \\\    \\\\   \\\\  \\\`   | 1     | 1 (backslash & ~first_is_escaped)
+    // | escape_and_terminal_code       | ` \n \ \n   \ \n   \ \    \ \   \ \`   | 5     | 5 (next_escape_and_terminal_code())
+    // | escaped                        | `\    \ n    \ n    \ \    \ \   \ ` X | 6     | 7 (escape_and_terminal_code ^ (potential_escape | first_is_escaped))
+    // | escape                         | `    \ \    \ \    \ \    \ \   \ \`   | 6     | 8 (escape_and_terminal_code & backslash)
+    // | first_is_escaped               | `\                                 `   | 7 (*) | 9 (escape >> 63) ()
+    //                                                                               (*) this is not needed until the next iteration
+    uint64_t escape_and_terminal_code = next_escape_and_terminal_code(backslash & ~this->next_is_escaped);
+    uint64_t escaped = escape_and_terminal_code ^ (backslash | this->next_is_escaped);
+    uint64_t escape = escape_and_terminal_code & backslash;
+    this->next_is_escaped = escape >> 63;
+    return {escaped, escape};
+  }
+
+private:
+  static constexpr const uint64_t ODD_BITS = 0xAAAAAAAAAAAAAAAAULL;
+
+  simdjson_really_inline uint64_t next_escaped_without_backslashes() noexcept {
+    uint64_t escaped = this->next_is_escaped;
+    this->next_is_escaped = 0;
+    return escaped;
+  }
+
+  /**
+   * Returns a mask of the next escape characters (masking out escaped backslashes), along with
+   * any non-backslash escape codes.
+   *
+   * \n \\n \\\n \\\\n returns:
+   * \n \   \ \n \ \
+   * 11 100 1011 10100
+   *
+   * You are expected to mask out the first bit yourself if the previous block had a trailing
+   * escape.
+   *
+   * & the result with potential_escape to get just the escape characters.
+   * ^ the result with (potential_escape | first_is_escaped) to get escaped characters.
+   */
+  static simdjson_really_inline uint64_t next_escape_and_terminal_code(uint64_t potential_escape) noexcept {
+    // If we were to just shift and mask out any odd bits, we'd actually get a *half* right answer:
+    // any even-aligned backslash runs would be correct! Odd-aligned backslash runs would be
+    // inverted (\\\ would be 010 instead of 101).
+    //
+    // ```
+    // string:              | ____\\\\_\\\\_____ |
+    // maybe_escaped | ODD  |     \ \   \ \      |
+    //               even-aligned ^^^  ^^^^ odd-aligned
+    // ```
+    //
+    // Taking that into account, our basic strategy is:
+    //
+    // 1. Use subtraction to produce a mask with 1's for even-aligned runs and 0's for
+    //    odd-aligned runs.
+    // 2. XOR all odd bits, which masks out the odd bits in even-aligned runs, and brings IN the
+    //    odd bits in odd-aligned runs.
+    // 3. & with backslash to clean up any stray bits.
+    // runs are set to 0, and then XORing with "odd":
+    //
+    // |                                | Mask (shows characters instead of 1's) | Instructions        |
+    // |--------------------------------|----------------------------------------|---------------------|
+    // | string                         | `\\n_\\\n___\\\n___\\\\___\\\\__\\\`   |
+    // |                                | `    even   odd    even   odd   odd`   |
+    // | maybe_escaped                  | `  n  \\n    \\n    \\\_   \\\_  \\` X | 1 (potential_escape << 1)
+    // | maybe_escaped_and_odd          | ` \n_ \\n _ \\\n_ _ \\\__ _\\\_ \\\`   | 1 (maybe_escaped | odd)
+    // | even_series_codes_and_odd      | `  n_\\\  _    n_ _\\\\ _     _    `   | 1 (maybe_escaped_and_odd - potential_escape)
+    // | escape_and_terminal_code       | ` \n \ \n   \ \n   \ \    \ \   \ \`   | 1 (^ odd)
+    //
+
+    // Escaped characters are characters following an escape.
+    uint64_t maybe_escaped = potential_escape << 1;
+
+    // To distinguish odd from even escape sequences, therefore, we turn on any *starting*
+    // escapes that are on an odd byte. (We actually bring in all odd bits, for speed.)
+    // - Odd runs of backslashes are 0000, and the code at the end ("n" in \n or \\n) is 1.
+    // - Odd runs of backslashes are 1111, and the code at the end ("n" in \n or \\n) is 0.
+    // - All other odd bytes are 1, and even bytes are 0.
+    uint64_t maybe_escaped_and_odd_bits     = maybe_escaped | ODD_BITS;
+    uint64_t even_series_codes_and_odd_bits = maybe_escaped_and_odd_bits - potential_escape;
+
+    // Now we flip all odd bytes back with xor. This:
+    // - Makes odd runs of backslashes go from 0000 to 1010
+    // - Makes even runs of backslashes go from 1111 to 1010
+    // - Sets actually-escaped codes to 1 (the n in \n and \\n: \n = 11, \\n = 100)
+    // - Resets all other bytes to 0
+    return even_series_codes_and_odd_bits ^ ODD_BITS;
+  }
+};
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+/* end file generic/stage1/json_escape_scanner.h for westmere */
+/* including generic/stage1/json_string_scanner.h for westmere: #include <generic/stage1/json_string_scanner.h> */
+/* begin file generic/stage1/json_string_scanner.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_escape_scanner.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace stage1 {
+
+struct json_string_block {
+  // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
+  simdjson_really_inline json_string_block(uint64_t escaped, uint64_t quote, uint64_t in_string) :
+  _escaped(escaped), _quote(quote), _in_string(in_string) {}
+
+  // Escaped characters (characters following an escape() character)
+  simdjson_really_inline uint64_t escaped() const { return _escaped; }
+  // Real (non-backslashed) quotes
+  simdjson_really_inline uint64_t quote() const { return _quote; }
+  // Only characters inside the string (not including the quotes)
+  simdjson_really_inline uint64_t string_content() const { return _in_string & ~_quote; }
+  // Return a mask of whether the given characters are inside a string (only works on non-quotes)
+  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
+  // Return a mask of whether the given characters are inside a string (only works on non-quotes)
+  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
+  // Tail of string (everything except the start quote)
+  simdjson_really_inline uint64_t string_tail() const { return _in_string ^ _quote; }
+
+  // escaped characters (backslashed--does not include the hex characters after \u)
+  uint64_t _escaped;
+  // real quotes (non-escaped ones)
+  uint64_t _quote;
+  // string characters (includes start quote but not end quote)
+  uint64_t _in_string;
+};
+
+// Scans blocks for string characters, storing the state necessary to do so
+class json_string_scanner {
+public:
+  simdjson_really_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
+  // Returns either UNCLOSED_STRING or SUCCESS
+  simdjson_really_inline error_code finish();
+
+private:
+  // Scans for escape characters
+  json_escape_scanner escape_scanner{};
+  // Whether the last iteration was still inside a string (all 1's = true, all 0's = false).
+  uint64_t prev_in_string = 0ULL;
+};
+
+//
+// Return a mask of all string characters plus end quotes.
+//
+// prev_escaped is overflow saying whether the next character is escaped.
+// prev_in_string is overflow saying whether we're still in a string.
+//
+// Backslash sequences outside of quotes will be detected in stage 2.
+//
+simdjson_really_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
+  const uint64_t backslash = in.eq('\\');
+  const uint64_t escaped = escape_scanner.next(backslash).escaped;
+  const uint64_t quote = in.eq('"') & ~escaped;
+
+  //
+  // prefix_xor flips on bits inside the string (and flips off the end quote).
+  //
+  // Then we xor with prev_in_string: if we were in a string already, its effect is flipped
+  // (characters inside strings are outside, and characters outside strings are inside).
+  //
+  const uint64_t in_string = prefix_xor(quote) ^ prev_in_string;
+
+  //
+  // Check if we're still in a string at the end of the box so the next block will know
+  //
+  prev_in_string = uint64_t(static_cast<int64_t>(in_string) >> 63);
+
+  // Use ^ to turn the beginning quote off, and the end quote on.
+
+  // We are returning a function-local object so either we get a move constructor
+  // or we get copy elision.
+  return json_string_block(escaped, quote, in_string);
+}
+
+simdjson_really_inline error_code json_string_scanner::finish() {
+  if (prev_in_string) {
+    return UNCLOSED_STRING;
+  }
+  return SUCCESS;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+/* end file generic/stage1/json_string_scanner.h for westmere */
+/* including generic/stage1/utf8_lookup4_algorithm.h for westmere: #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* begin file generic/stage1/utf8_lookup4_algorithm.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace utf8_validation {
+
+using namespace simd;
+
+  simdjson_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+// Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+// Bit 1 = Too Long (ASCII followed by continuation)
+// Bit 2 = Overlong 3-byte
+// Bit 4 = Surrogate
+// Bit 5 = Overlong 2-byte
+// Bit 7 = Two Continuations
+    constexpr const uint8_t TOO_SHORT   = 1<<0; // 11______ 0_______
+                                                // 11______ 11______
+    constexpr const uint8_t TOO_LONG    = 1<<1; // 0_______ 10______
+    constexpr const uint8_t OVERLONG_3  = 1<<2; // 11100000 100_____
+    constexpr const uint8_t SURROGATE   = 1<<4; // 11101101 101_____
+    constexpr const uint8_t OVERLONG_2  = 1<<5; // 1100000_ 10______
+    constexpr const uint8_t TWO_CONTS   = 1<<7; // 10______ 10______
+    constexpr const uint8_t TOO_LARGE   = 1<<3; // 11110100 1001____
+                                                // 11110100 101_____
+                                                // 11110101 1001____
+                                                // 11110101 101_____
+                                                // 1111011_ 1001____
+                                                // 1111011_ 101_____
+                                                // 11111___ 1001____
+                                                // 11111___ 101_____
+    constexpr const uint8_t TOO_LARGE_1000 = 1<<6;
+                                                // 11110101 1000____
+                                                // 1111011_ 1000____
+                                                // 11111___ 1000____
+    constexpr const uint8_t OVERLONG_4  = 1<<6; // 11110000 1000____
+
+    const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      TOO_SHORT,
+      // 1110____ ________ <three byte lead in byte 1>
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4
+    );
+    constexpr const uint8_t CARRY = TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+    const simd8<uint8_t> byte_1_low = (prev1 & 0x0F).lookup_16<uint8_t>(
+      // ____0000 ________
+      CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+      // ____0001 ________
+      CARRY | OVERLONG_2,
+      // ____001_ ________
+      CARRY,
+      CARRY,
+
+      // ____0100 ________
+      CARRY | TOO_LARGE,
+      // ____0101 ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      // ____011_ ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+
+      // ____1___ ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      // ____1101 ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000
+    );
+    const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE  | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE  | TOO_LARGE,
+
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT
+    );
+    return (byte_1_high & byte_1_low & byte_2_high);
+  }
+  simdjson_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
+      const simd8<uint8_t> prev_input, const simd8<uint8_t> sc) {
+    simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+    simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+    simd8<uint8_t> must23 = must_be_2_3_continuation(prev2, prev3);
+    simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+    return must23_80 ^ sc;
+  }
+
+  //
+  // Return nonzero if there are incomplete multibyte characters at the end of the block:
+  // e.g. if there is a 4-byte character, but it's 3 bytes from the end.
+  //
+  simdjson_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+    // If the previous input's last 3 bytes match this, they're too short (they ended at EOF):
+    // ... 1111____ 111_____ 11______
+#if SIMDJSON_IMPLEMENTATION_ICELAKE || (SIMDJSON_IMPLEMENTATION_RVV_VLS && __riscv_v_fixed_vlen >= 512)
+    static const uint8_t max_array[64] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
+    };
+#else
+    static const uint8_t max_array[32] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
+    };
+#endif
+    const simd8<uint8_t> max_value(&max_array[sizeof(max_array)-sizeof(simd8<uint8_t>)]);
+    return input.gt_bits(max_value);
+  }
+
+  struct utf8_checker {
+    // If this is nonzero, there has been a UTF-8 error.
+    simd8<uint8_t> error;
+    // The last input we received
+    simd8<uint8_t> prev_input_block;
+    // Whether the last input we received was incomplete (used for ASCII fast path)
+    simd8<uint8_t> prev_incomplete;
+
+    //
+    // Check whether the current bytes are valid UTF-8.
+    //
+    simdjson_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
+      // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ lead bytes
+      // (2, 3, 4-byte leads become large positive numbers instead of small negative numbers)
+      simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+      simd8<uint8_t> sc = check_special_cases(input, prev1);
+      this->error |= check_multibyte_lengths(input, prev_input, sc);
+    }
+
+    // The only problem that can happen at EOF is that a multibyte character is too short
+    // or a byte value too large in the last bytes: check_special_cases only checks for bytes
+    // too large in the first of two bytes.
+    simdjson_inline void check_eof() {
+      // If the previous block had incomplete UTF-8 characters at the end, an ASCII block can't
+      // possibly finish them.
+      this->error |= this->prev_incomplete;
+    }
+
+    simdjson_inline void check_next_input(const simd8x64<uint8_t>& input) {
+      if(simdjson_likely(is_ascii(input))) {
+        this->error |= this->prev_incomplete;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio, it is not good enough.
+        static_assert((simd8x64<uint8_t>::NUM_CHUNKS == 1)
+                ||(simd8x64<uint8_t>::NUM_CHUNKS == 2)
+                || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+                "We support one, two or four chunks per 64-byte block.");
+        SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 1) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+          this->check_utf8_bytes(input.get<1>(), input.get<0>());
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+          this->check_utf8_bytes(input.get<1>(), input.get<0>());
+          this->check_utf8_bytes(input.get<2>(), input.get<1>());
+          this->check_utf8_bytes(input.get<3>(), input.get<2>());
+        }
+        this->prev_incomplete = is_incomplete(input.get<simd8x64<uint8_t>::NUM_CHUNKS-1>());
+        this->prev_input_block = input.get<simd8x64<uint8_t>::NUM_CHUNKS-1>();
+      }
+    }
+    // do not forget to call check_eof!
+    simdjson_warn_unused simdjson_inline error_code errors() {
+      return this->error.any_bits_set_anywhere() ? error_code::UTF8_ERROR : error_code::SUCCESS;
+    }
+
+  }; // struct utf8_checker
+} // namespace utf8_validation
+
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H
+/* end file generic/stage1/utf8_lookup4_algorithm.h for westmere */
+/* including generic/stage1/json_scanner.h for westmere: #include <generic/stage1/json_scanner.h> */
+/* begin file generic/stage1/json_scanner.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/json_character_block.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_string_scanner.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace stage1 {
+
+/**
+ * A block of scanned json, with information on operators and scalars.
+ *
+ * We seek to identify pseudo-structural characters. Anything that is inside
+ * a string must be omitted (hence  & ~_string.string_tail()).
+ * Otherwise, pseudo-structural characters come in two forms.
+ * 1. We have the structural characters ([,],{,},:, comma). The
+ *    term 'structural character' is from the JSON RFC.
+ * 2. We have the 'scalar pseudo-structural characters'.
+ *    Scalars are quotes, and any character except structural characters and white space.
+ *
+ * To identify the scalar pseudo-structural characters, we must look at what comes
+ * before them: it must be a space, a quote or a structural characters.
+ * Starting with simdjson v0.3, we identify them by
+ * negation: we identify everything that is followed by a non-quote scalar,
+ * and we negate that. Whatever remains must be a 'scalar pseudo-structural character'.
+ */
+struct json_block {
+public:
+  // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
+  simdjson_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  _string(std::move(string)), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
+  simdjson_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  _string(string), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
+
+  /**
+   * The start of structurals.
+   * In simdjson prior to v0.3, these were called the pseudo-structural characters.
+   **/
+  simdjson_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
+  /** All JSON whitespace (i.e. not in a string) */
+  simdjson_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
+
+  // Helpers
+
+  /** Whether the given characters are inside a string (only works on non-quotes) */
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
+  /** Whether the given characters are outside a string (only works on non-quotes) */
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
+
+  // string and escape characters
+  json_string_block _string;
+  // whitespace, structural characters ('operators'), scalars
+  json_character_block _characters;
+  // whether the previous character was a scalar
+  uint64_t _follows_potential_nonquote_scalar;
+private:
+  // Potential structurals (i.e. disregarding strings)
+
+  /**
+   * structural elements ([,],{,},:, comma) plus scalar starts like 123, true and "abc".
+   * They may reside inside a string.
+   **/
+  simdjson_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
+  /**
+   * The start of non-operator runs, like 123, true and "abc".
+   * It main reside inside a string.
+   **/
+  simdjson_inline uint64_t potential_scalar_start() const noexcept {
+    // The term "scalar" refers to anything except structural characters and white space
+    // (so letters, numbers, quotes).
+    // Whenever it is preceded by something that is not a structural element ({,},[,],:, ") nor a white-space
+    // then we know that it is irrelevant structurally.
+    return _characters.scalar() & ~follows_potential_scalar();
+  }
+  /**
+   * Whether the given character is immediately after a non-operator like 123, true.
+   * The characters following a quote are not included.
+   */
+  simdjson_inline uint64_t follows_potential_scalar() const noexcept {
+    // _follows_potential_nonquote_scalar: is defined as marking any character that follows a character
+    // that is not a structural element ({,},[,],:, comma) nor a quote (") and that is not a
+    // white space.
+    // It is understood that within quoted region, anything at all could be marked (irrelevant).
+    return _follows_potential_nonquote_scalar;
+  }
+};
+
+/**
+ * Scans JSON for important bits: structural characters or 'operators', strings, and scalars.
+ *
+ * The scanner starts by calculating two distinct things:
+ * - string characters (taking \" into account)
+ * - structural characters or 'operators' ([]{},:, comma)
+ *   and scalars (runs of non-operators like 123, true and "abc")
+ *
+ * To minimize data dependency (a key component of the scanner's speed), it finds these in parallel:
+ * in particular, the operator/scalar bit will find plenty of things that are actually part of
+ * strings. When we're done, json_block will fuse the two together by masking out tokens that are
+ * part of a string.
+ */
+class json_scanner {
+public:
+  json_scanner() = default;
+  simdjson_inline json_block next(const simd::simd8x64<uint8_t>& in);
+  // Returns either UNCLOSED_STRING or SUCCESS
+  simdjson_warn_unused simdjson_inline error_code finish();
+
+private:
+  // Whether the last character of the previous iteration is part of a scalar token
+  // (anything except whitespace or a structural character/'operator').
+  uint64_t prev_scalar = 0ULL;
+  json_string_scanner string_scanner{};
+};
+
+
+//
+// Check if the current character immediately follows a matching character.
+//
+// For example, this checks for quotes with backslashes in front of them:
+//
+//     const uint64_t backslashed_quote = in.eq('"') & immediately_follows(in.eq('\'), prev_backslash);
+//
+simdjson_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
+  const uint64_t result = match << 1 | overflow;
+  overflow = match >> 63;
+  return result;
+}
+
+simdjson_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
+  json_string_block strings = string_scanner.next(in);
+  // identifies the white-space and the structural characters
+  json_character_block characters = json_character_block::classify(in);
+  // The term "scalar" refers to anything except structural characters and white space
+  // (so letters, numbers, quotes).
+  // We want follows_scalar to mark anything that follows a non-quote scalar (so letters and numbers).
+  //
+  // A terminal quote should either be followed by a structural character (comma, brace, bracket, colon)
+  // or nothing. However, we still want ' "a string"true ' to mark the 't' of 'true' as a potential
+  // pseudo-structural character just like we would if we had  ' "a string" true '; otherwise we
+  // may need to add an extra check when parsing strings.
+  //
+  // Performance: there are many ways to skin this cat.
+  const uint64_t nonquote_scalar = characters.scalar() & ~strings.quote();
+  uint64_t follows_nonquote_scalar = follows(nonquote_scalar, prev_scalar);
+  // We are returning a function-local object so either we get a move constructor
+  // or we get copy elision.
+  return json_block(
+    strings,// strings is a function-local object so either it moves or the copy is elided.
+    characters,
+    follows_nonquote_scalar
+  );
+}
+
+simdjson_warn_unused simdjson_inline error_code json_scanner::finish() {
+  return string_scanner.finish();
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H
+/* end file generic/stage1/json_scanner.h for westmere */
+
+// All other declarations
+/* including generic/stage1/find_next_document_index.h for westmere: #include <generic/stage1/find_next_document_index.h> */
+/* begin file generic/stage1/find_next_document_index.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace stage1 {
+
+/**
+  * This algorithm is used to quickly identify the last structural position that
+  * makes up a complete document.
+  *
+  * It does this by going backwards and finding the last *document boundary* (a
+  * place where one value follows another without a comma between them). If the
+  * last document (the characters after the boundary) has an equal number of
+  * start and end brackets, it is considered complete.
+  *
+  * Simply put, we iterate over the structural characters, starting from
+  * the end. We consider that we found the end of a JSON document when the
+  * first element of the pair is NOT one of these characters: '{' '[' ':' ','
+  * and when the second element is NOT one of these characters: '}' ']' ':' ','.
+  *
+  * This simple comparison works most of the time, but it does not cover cases
+  * where the batch's structural indexes contain a perfect amount of documents.
+  * In such a case, we do not have access to the structural index which follows
+  * the last document, therefore, we do not have access to the second element in
+  * the pair, and that means we cannot identify the last document. To fix this
+  * issue, we keep a count of the open and closed curly/square braces we found
+  * while searching for the pair. When we find a pair AND the count of open and
+  * closed curly/square braces is the same, we know that we just passed a
+  * complete document, therefore the last json buffer location is the end of the
+  * batch.
+  */
+simdjson_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
+  // Variant: do not count separately, just figure out depth
+  if(parser.n_structural_indexes == 0) { return 0; }
+  auto arr_cnt = 0;
+  auto obj_cnt = 0;
+  for (auto i = parser.n_structural_indexes - 1; i > 0; i--) {
+    auto idxb = parser.structural_indexes[i];
+    switch (parser.buf[idxb]) {
+    case ':':
+    case ',':
+      continue;
+    case '}':
+      obj_cnt--;
+      continue;
+    case ']':
+      arr_cnt--;
+      continue;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+    }
+    auto idxa = parser.structural_indexes[i - 1];
+    switch (parser.buf[idxa]) {
+    case '{':
+    case '[':
+    case ':':
+    case ',':
+      continue;
+    }
+    // Last document is complete, so the next document will appear after!
+    if (!arr_cnt && !obj_cnt) {
+      return parser.n_structural_indexes;
+    }
+    // Last document is incomplete; mark the document at i + 1 as the next one
+    return i;
+  }
+  // If we made it to the end, we want to finish counting to see if we have a full document.
+  switch (parser.buf[parser.structural_indexes[0]]) {
+    case '}':
+      obj_cnt--;
+      break;
+    case ']':
+      arr_cnt--;
+      break;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+  }
+  if (!arr_cnt && !obj_cnt) {
+    // We have a complete document.
+    return parser.n_structural_indexes;
+  }
+  return 0;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+/* end file generic/stage1/find_next_document_index.h for westmere */
+/* including generic/stage1/json_minifier.h for westmere: #include <generic/stage1/json_minifier.h> */
+/* begin file generic/stage1/json_minifier.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses in stage1
+// It is intended to be included multiple times and compiled multiple times
+// We assume the file in which it is included already includes
+// "simdjson/stage1.h" (this simplifies amalgation)
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace stage1 {
+
+class json_minifier {
+public:
+  template<size_t STEP_SIZE>
+  static error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept;
+
+private:
+  simdjson_inline json_minifier(uint8_t *_dst)
+  : dst{_dst}
+  {}
+  template<size_t STEP_SIZE>
+  simdjson_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
+  simdjson_warn_unused simdjson_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
+  json_scanner scanner{};
+  uint8_t *dst;
+};
+
+simdjson_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
+  uint64_t mask = block.whitespace();
+  dst += in.compress(mask, dst);
+}
+
+simdjson_warn_unused simdjson_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
+  error_code error = scanner.finish();
+  if (error) { dst_len = 0; return error; }
+  dst_len = dst - dst_start;
+  return SUCCESS;
+}
+
+template<>
+simdjson_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block_buf);
+  simd::simd8x64<uint8_t> in_2(block_buf+64);
+  json_block block_1 = scanner.next(in_1);
+  json_block block_2 = scanner.next(in_2);
+  this->next(in_1, block_1);
+  this->next(in_2, block_2);
+  reader.advance();
+}
+
+template<>
+simdjson_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block_buf);
+  json_block block_1 = scanner.next(in_1);
+  this->next(block_buf, block_1);
+  reader.advance();
+}
+
+template<size_t STEP_SIZE>
+error_code json_minifier::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept {
+  buf_block_reader<STEP_SIZE> reader(buf, len);
+  json_minifier minifier(dst);
+
+  // Index the first n-1 blocks
+  while (reader.has_full_block()) {
+    minifier.step<STEP_SIZE>(reader.full_block(), reader);
+  }
+
+  // Index the last (remainder) block, padded with spaces
+  uint8_t block[STEP_SIZE];
+  size_t remaining_bytes = reader.get_remainder(block);
+  if (remaining_bytes > 0) {
+    // We do not want to write directly to the output stream. Rather, we write
+    // to a local buffer (for safety).
+    uint8_t out_block[STEP_SIZE];
+    uint8_t * const guarded_dst{minifier.dst};
+    minifier.dst = out_block;
+    minifier.step<STEP_SIZE>(block, reader);
+    size_t to_write = minifier.dst - out_block;
+    // In some cases, we could be enticed to consider the padded spaces
+    // as part of the string. This is fine as long as we do not write more
+    // than we consumed.
+    if(to_write > remaining_bytes) { to_write = remaining_bytes; }
+    memcpy(guarded_dst, out_block, to_write);
+    minifier.dst = guarded_dst + to_write;
+  }
+  return minifier.finish(dst, dst_len);
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H
+/* end file generic/stage1/json_minifier.h for westmere */
+/* including generic/stage1/json_structural_indexer.h for westmere: #include <generic/stage1/json_structural_indexer.h> */
+/* begin file generic/stage1/json_structural_indexer.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_string_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_minifier.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/find_next_document_index.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses in stage1
+// It is intended to be included multiple times and compiled multiple times
+// We assume the file in which it is included already includes
+// "simdjson/stage1.h" (this simplifies amalgation)
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace stage1 {
+
+class bit_indexer {
+public:
+  uint32_t *tail;
+
+  simdjson_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
+
+#if SIMDJSON_PREFER_REVERSE_BITS
+  /**
+    * ARM lacks a fast trailing zero instruction, but it has a fast
+    * bit reversal instruction and a fast leading zero instruction.
+    * Thus it may be profitable to reverse the bits (once) and then
+    * to rely on a sequence of instructions that call the leading
+    * zero instruction.
+    *
+    * Performance notes:
+    * The chosen routine is not optimal in terms of data dependency
+    * since zero_leading_bit might require two instructions. However,
+    * it tends to minimize the total number of instructions which is
+    * beneficial.
+    */
+  simdjson_inline void write_index(uint32_t idx, uint64_t& rev_bits, int i) {
+    int lz = leading_zeroes(rev_bits);
+    this->tail[i] = static_cast<uint32_t>(idx) + lz;
+    rev_bits = zero_leading_bit(rev_bits, lz);
+  }
+#else
+  /**
+    * Under recent x64 systems, we often have both a fast trailing zero
+    * instruction and a fast 'clear-lower-bit' instruction so the following
+    * algorithm can be competitive.
+    */
+
+  simdjson_inline void write_index(uint32_t idx, uint64_t& bits, int i) {
+    this->tail[i] = idx + trailing_zeroes(bits);
+    bits = clear_lowest_bit(bits);
+  }
+#endif // SIMDJSON_PREFER_REVERSE_BITS
+
+  template <int START, int N>
+  simdjson_inline int write_indexes(uint32_t idx, uint64_t& bits) {
+    write_index(idx, bits, START);
+    SIMDJSON_IF_CONSTEXPR (N > 1) {
+      write_indexes<(N-1>0?START+1:START), (N-1>=0?N-1:1)>(idx, bits);
+    }
+    return START+N;
+  }
+
+  template <int START, int END, int STEP>
+  simdjson_inline int write_indexes_stepped(uint32_t idx, uint64_t& bits, int cnt) {
+    write_indexes<START, STEP>(idx, bits);
+    SIMDJSON_IF_CONSTEXPR ((START+STEP)  < END) {
+      if (simdjson_unlikely((START+STEP) < cnt)) {
+        write_indexes_stepped<(START+STEP<END?START+STEP:END), END, STEP>(idx, bits, cnt);
+      }
+    }
+    return ((END-START) % STEP) == 0 ? END : (END-START) - ((END-START) % STEP) + STEP;
+  }
+
+  // flatten out values in 'bits' assuming that they are are to have values of idx
+  // plus their position in the bitvector, and store these indexes at
+  // base_ptr[base] incrementing base as we go
+  // will potentially store extra values beyond end of valid bits, so base_ptr
+  // needs to be large enough to handle this
+  //
+  // If the kernel sets SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER, then it
+  // will provide its own version of the code.
+#ifdef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+  simdjson_inline void write(uint32_t idx, uint64_t bits);
+#else
+  simdjson_inline void write(uint32_t idx, uint64_t bits) {
+    // In some instances, the next branch is expensive because it is mispredicted.
+    // Unfortunately, in other cases,
+    // it helps tremendously.
+    if (bits == 0)
+        return;
+
+    int cnt = static_cast<int>(count_ones(bits));
+
+#if SIMDJSON_PREFER_REVERSE_BITS
+    bits = reverse_bits(bits);
+#endif
+#ifdef SIMDJSON_STRUCTURAL_INDEXER_STEP
+    static constexpr const int STEP = SIMDJSON_STRUCTURAL_INDEXER_STEP;
+#else
+    static constexpr const int STEP = 4;
+#endif
+    static constexpr const int STEP_UNTIL = 24;
+
+    write_indexes_stepped<0, STEP_UNTIL, STEP>(idx, bits, cnt);
+    SIMDJSON_IF_CONSTEXPR (STEP_UNTIL < 64) {
+      if (simdjson_unlikely(STEP_UNTIL < cnt)) {
+        for (int i=STEP_UNTIL; i<cnt; i++) {
+          write_index(idx, bits, i);
+        }
+      }
+    }
+
+    this->tail += cnt;
+  }
+#endif // SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+};
+
+class json_structural_indexer {
+public:
+  /**
+   * Find the important bits of JSON in a 128-byte chunk, and add them to structural_indexes.
+   *
+   * @param partial Setting the partial parameter to true allows the find_structural_bits to
+   *   tolerate unclosed strings. The caller should still ensure that the input is valid UTF-8. If
+   *   you are processing substrings, you may want to call on a function like trimmed_length_safe_utf8.
+   */
+  template<size_t STEP_SIZE>
+  static error_code index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept;
+
+private:
+  simdjson_inline json_structural_indexer(uint32_t *structural_indexes);
+  template<size_t STEP_SIZE>
+  simdjson_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
+  simdjson_warn_unused simdjson_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
+
+  json_scanner scanner{};
+  utf8_checker checker{};
+  bit_indexer indexer;
+  uint64_t prev_structurals = 0;
+  uint64_t unescaped_chars_error = 0;
+};
+
+simdjson_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
+
+// Skip the last character if it is partial
+simdjson_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
+  if (simdjson_unlikely(len < 3)) {
+    switch (len) {
+      case 2:
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
+        return len;
+      case 1:
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        return len;
+      case 0:
+        return len;
+    }
+  }
+  if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+  if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 1 byte left
+  if (buf[len-3] >= 0xf0) { return len-3; } // 4-byte characters with only 3 bytes left
+  return len;
+}
+
+//
+// PERF NOTES:
+// We pipe 2 inputs through these stages:
+// 1. Load JSON into registers. This takes a long time and is highly parallelizable, so we load
+//    2 inputs' worth at once so that by the time step 2 is looking for them input, it's available.
+// 2. Scan the JSON for critical data: strings, scalars and operators. This is the critical path.
+//    The output of step 1 depends entirely on this information. These functions don't quite use
+//    up enough CPU: the second half of the functions is highly serial, only using 1 execution core
+//    at a time. The second input's scans has some dependency on the first ones finishing it, but
+//    they can make a lot of progress before they need that information.
+// 3. Step 1 does not use enough capacity, so we run some extra stuff while we're waiting for that
+//    to finish: utf-8 checks and generating the output from the last iteration.
+//
+// The reason we run 2 inputs at a time, is steps 2 and 3 are *still* not enough to soak up all
+// available capacity with just one input. Running 2 at a time seems to give the CPU a good enough
+// workout.
+//
+template<size_t STEP_SIZE>
+error_code json_structural_indexer::index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept {
+  if (simdjson_unlikely(len > parser.capacity())) { return CAPACITY; }
+  // We guard the rest of the code so that we can assume that len > 0 throughout.
+  if (len == 0) { return EMPTY; }
+  if (is_streaming(partial)) {
+    len = trim_partial_utf8(buf, len);
+    // If you end up with an empty window after trimming
+    // the partial UTF-8 bytes, then chances are good that you
+    // have an UTF-8 formatting error.
+    if(len == 0) { return UTF8_ERROR; }
+  }
+  buf_block_reader<STEP_SIZE> reader(buf, len);
+  json_structural_indexer indexer(parser.structural_indexes.get());
+
+  // Read all but the last block
+  while (reader.has_full_block()) {
+    indexer.step<STEP_SIZE>(reader.full_block(), reader);
+  }
+  // Take care of the last block (will always be there unless file is empty which is
+  // not supposed to happen.)
+  uint8_t block[STEP_SIZE];
+  if (simdjson_unlikely(reader.get_remainder(block) == 0)) { return UNEXPECTED_ERROR; }
+  indexer.step<STEP_SIZE>(block, reader);
+  return indexer.finish(parser, reader.block_index(), len, partial);
+}
+
+template<>
+simdjson_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block);
+  simd::simd8x64<uint8_t> in_2(block+64);
+  json_block block_1 = scanner.next(in_1);
+  json_block block_2 = scanner.next(in_2);
+  this->next(in_1, block_1, reader.block_index());
+  this->next(in_2, block_2, reader.block_index()+64);
+  reader.advance();
+}
+
+template<>
+simdjson_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block);
+  json_block block_1 = scanner.next(in_1);
+  this->next(in_1, block_1, reader.block_index());
+  reader.advance();
+}
+
+simdjson_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
+  uint64_t unescaped = in.lteq(0x1F);
+#if SIMDJSON_UTF8VALIDATION
+  checker.check_next_input(in);
+#endif
+  indexer.write(uint32_t(idx-64), prev_structurals); // Output *last* iteration's structurals to the parser
+  prev_structurals = block.structural_start();
+  unescaped_chars_error |= block.non_quote_inside_string(unescaped);
+}
+
+simdjson_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
+  // Write out the final iteration's structurals
+  indexer.write(uint32_t(idx-64), prev_structurals);
+  error_code error = scanner.finish();
+  // We deliberately break down the next expression so that it is
+  // human readable.
+  const bool should_we_exit = is_streaming(partial) ?
+    ((error != SUCCESS) && (error != UNCLOSED_STRING)) // when partial we tolerate UNCLOSED_STRING
+    : (error != SUCCESS); // if partial is false, we must have SUCCESS
+  const bool have_unclosed_string = (error == UNCLOSED_STRING);
+  if (simdjson_unlikely(should_we_exit)) { return error; }
+
+  if (unescaped_chars_error) {
+    return UNESCAPED_CHARS;
+  }
+  parser.n_structural_indexes = uint32_t(indexer.tail - parser.structural_indexes.get());
+  /***
+   * The On-Demand API requires special padding.
+   *
+   * This is related to https://github.com/simdjson/simdjson/issues/906
+   * Basically, we want to make sure that if the parsing continues beyond the last (valid)
+   * structural character, it quickly stops.
+   * Only three structural characters can be repeated without triggering an error in JSON:  [,] and }.
+   * We repeat the padding character (at 'len'). We don't know what it is, but if the parsing
+   * continues, then it must be [,] or }.
+   * Suppose it is ] or }. We backtrack to the first character, what could it be that would
+   * not trigger an error? It could be ] or } but no, because you can't start a document that way.
+   * It can't be a comma, a colon or any simple value. So the only way we could continue is
+   * if the repeated character is [. But if so, the document must start with [. But if the document
+   * starts with [, it should end with ]. If we enforce that rule, then we would get
+   * ][[ which is invalid.
+   *
+   * This is illustrated with the test array_iterate_unclosed_error() on the following input:
+   * R"({ "a": [,,)"
+   **/
+  parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len); // used later in partial == stage1_mode::streaming_final
+  parser.structural_indexes[parser.n_structural_indexes + 1] = uint32_t(len);
+  parser.structural_indexes[parser.n_structural_indexes + 2] = 0;
+  parser.next_structural_index = 0;
+  // a valid JSON file cannot have zero structural indexes - we should have found something
+  if (simdjson_unlikely(parser.n_structural_indexes == 0u)) {
+    return EMPTY;
+  }
+  if (simdjson_unlikely(parser.structural_indexes[parser.n_structural_indexes - 1] > len)) {
+    return UNEXPECTED_ERROR;
+  }
+  if (partial == stage1_mode::streaming_partial) {
+    // If we have an unclosed string, then the last structural
+    // will be the quote and we want to make sure to omit it.
+    if(have_unclosed_string) {
+      parser.n_structural_indexes--;
+      // a valid JSON file cannot have zero structural indexes - we should have found something
+      if (simdjson_unlikely(parser.n_structural_indexes == 0u)) { return CAPACITY; }
+    }
+    // We truncate the input to the end of the last complete document (or zero).
+    auto new_structural_indexes = find_next_document_index(parser);
+    if (new_structural_indexes == 0 && parser.n_structural_indexes > 0) {
+      if(parser.structural_indexes[0] == 0) {
+        // If the buffer is partial and we started at index 0 but the document is
+        // incomplete, it's too big to parse.
+        return CAPACITY;
+      } else {
+        // It is possible that the document could be parsed, we just had a lot
+        // of white space.
+        parser.n_structural_indexes = 0;
+        return EMPTY;
+      }
+    }
+
+    parser.n_structural_indexes = new_structural_indexes;
+  } else if (partial == stage1_mode::streaming_final) {
+    if(have_unclosed_string) { parser.n_structural_indexes--; }
+    // We truncate the input to the end of the last complete document (or zero).
+    // Because partial == stage1_mode::streaming_final, it means that we may
+    // silently ignore trailing garbage. Though it sounds bad, we do it
+    // deliberately because many people who have streams of JSON documents
+    // will truncate them for processing. E.g., imagine that you are uncompressing
+    // the data from a size file or receiving it in chunks from the network. You
+    // may not know where exactly the last document will be. Meanwhile the
+    // document_stream instances allow people to know the JSON documents they are
+    // parsing (see the iterator.source() method).
+    parser.n_structural_indexes = find_next_document_index(parser);
+    // We store the initial n_structural_indexes so that the client can see
+    // whether we used truncation. If initial_n_structural_indexes == parser.n_structural_indexes,
+    // then this will query parser.structural_indexes[parser.n_structural_indexes] which is len,
+    // otherwise, it will copy some prior index.
+    parser.structural_indexes[parser.n_structural_indexes + 1] = parser.structural_indexes[parser.n_structural_indexes];
+    // This next line is critical, do not change it unless you understand what you are
+    // doing.
+    parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len);
+    if (simdjson_unlikely(parser.n_structural_indexes == 0u)) {
+        // We tolerate an unclosed string at the very end of the stream. Indeed, users
+        // often load their data in bulk without being careful and they want us to ignore
+        // the trailing garbage.
+        return EMPTY;
+    }
+  }
+  checker.check_eof();
+  return checker.errors();
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+// Clear CUSTOM_BIT_INDEXER so other implementations can set it if they need to.
+#undef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H
+/* end file generic/stage1/json_structural_indexer.h for westmere */
+/* including generic/stage1/utf8_validator.h for westmere: #include <generic/stage1/utf8_validator.h> */
+/* begin file generic/stage1/utf8_validator.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace stage1 {
+
+/**
+ * Validates that the string is actual UTF-8.
+ */
+template<class checker>
+bool generic_validate_utf8(const uint8_t * input, size_t length) {
+    checker c{};
+    buf_block_reader<64> reader(input, length);
+    while (reader.has_full_block()) {
+      simd::simd8x64<uint8_t> in(reader.full_block());
+      c.check_next_input(in);
+      reader.advance();
+    }
+    uint8_t block[64]{};
+    reader.get_remainder(block);
+    simd::simd8x64<uint8_t> in(block);
+    c.check_next_input(in);
+    reader.advance();
+    c.check_eof();
+    return c.errors() == error_code::SUCCESS;
+}
+
+bool generic_validate_utf8(const char * input, size_t length) {
+    return generic_validate_utf8<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H
+/* end file generic/stage1/utf8_validator.h for westmere */
+/* end file generic/stage1/amalgamated.h for westmere */
+/* including generic/stage2/amalgamated.h for westmere: #include <generic/stage2/amalgamated.h> */
+/* begin file generic/stage2/amalgamated.h for westmere */
+// Stuff other things depend on
+/* including generic/stage2/base.h for westmere: #include <generic/stage2/base.h> */
+/* begin file generic/stage2/base.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_BASE_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace stage2 {
+
+class json_iterator;
+class structural_iterator;
+struct tape_builder;
+struct tape_writer;
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_BASE_H
+/* end file generic/stage2/base.h for westmere */
+/* including generic/stage2/tape_writer.h for westmere: #include <generic/stage2/tape_writer.h> */
+/* begin file generic/stage2/tape_writer.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/internal/tape_type.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace stage2 {
+
+struct tape_writer {
+  /** The next place to write to tape */
+  uint64_t *next_tape_loc;
+
+  /** Write a signed 64-bit value to tape. */
+  simdjson_inline void append_s64(int64_t value) noexcept;
+
+  /** Write an unsigned 64-bit value to tape. */
+  simdjson_inline void append_u64(uint64_t value) noexcept;
+
+  /** Write a double value to tape. */
+  simdjson_inline void append_double(double value) noexcept;
+
+  /** Write a big integer (as string) to tape. src points to first digit, len is byte count. */
+  simdjson_inline void append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept;
+
+  /**
+   * Append a tape entry (an 8-bit type,and 56 bits worth of value).
+   */
+  simdjson_inline void append(uint64_t val, internal::tape_type t) noexcept;
+
+  /**
+   * Skip the current tape entry without writing.
+   *
+   * Used to skip the start of the container, since we'll come back later to fill it in when the
+   * container ends.
+   */
+  simdjson_inline void skip() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a large u64 or i64.
+   */
+  simdjson_inline void skip_large_integer() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a double.
+   */
+  simdjson_inline void skip_double() noexcept;
+
+  /**
+   * Write a value to a known location on tape.
+   *
+   * Used to go back and write out the start of a container after the container ends.
+   */
+  simdjson_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
+
+private:
+  /**
+   * Append both the tape entry, and a supplementary value following it. Used for types that need
+   * all 64 bits, such as double and uint64_t.
+   */
+  template<typename T>
+  simdjson_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
+}; // struct tape_writer
+
+simdjson_inline void tape_writer::append_s64(int64_t value) noexcept {
+  append2(0, value, internal::tape_type::INT64);
+}
+
+simdjson_inline void tape_writer::append_u64(uint64_t value) noexcept {
+  append(0, internal::tape_type::UINT64);
+  *next_tape_loc = value;
+  next_tape_loc++;
+}
+
+/** Write a double value to tape. */
+simdjson_inline void tape_writer::append_double(double value) noexcept {
+  append2(0, value, internal::tape_type::DOUBLE);
+}
+
+simdjson_inline void tape_writer::skip() noexcept {
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::skip_large_integer() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::skip_double() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
+  *next_tape_loc = val | ((uint64_t(char(t))) << 56);
+  next_tape_loc++;
+}
+
+template<typename T>
+simdjson_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
+  append(val, t);
+  static_assert(sizeof(val2) == sizeof(*next_tape_loc), "Type is not 64 bits!");
+  memcpy(next_tape_loc, &val2, sizeof(val2));
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
+  tape_loc = val | ((uint64_t(char(t))) << 56);
+}
+
+simdjson_inline void tape_writer::append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept {
+  // Write to string buffer: [4-byte LE length][digits][null]
+  uint32_t str_len = uint32_t(len);
+  memcpy(string_buf, &str_len, sizeof(uint32_t));
+  memcpy(string_buf + sizeof(uint32_t), src, len);
+  string_buf[sizeof(uint32_t) + len] = 0;
+  // Tape entry: offset into string buffer
+  // The caller must set the offset relative to doc.string_buf base
+  append(0, internal::tape_type::BIGINT); // placeholder offset, caller patches
+  string_buf += sizeof(uint32_t) + len + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+/* end file generic/stage2/tape_writer.h for westmere */
+/* including generic/stage2/logger.h for westmere: #include <generic/stage2/logger.h> */
+/* begin file generic/stage2/logger.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+
+// This is for an internal-only stage 2 specific logger.
+// Set LOG_ENABLED = true to log what stage 2 is doing!
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace logger {
+
+  static constexpr const char * DASHES = "----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------";
+
+#if SIMDJSON_VERBOSE_LOGGING
+  static constexpr const bool LOG_ENABLED = true;
+#else
+  static constexpr const bool LOG_ENABLED = false;
+#endif
+  static constexpr const int LOG_EVENT_LEN = 20;
+  static constexpr const int LOG_BUFFER_LEN = 30;
+  static constexpr const int LOG_SMALL_BUFFER_LEN = 10;
+  static constexpr const int LOG_INDEX_LEN = 5;
+
+  static int log_depth; // Not threadsafe. Log only.
+
+  // Helper to turn unprintable or newline characters into spaces
+  static simdjson_inline char printable_char(char c) {
+    if (c >= 0x20) {
+      return c;
+    } else {
+      return ' ';
+    }
+  }
+
+  // Print the header and set up log_start
+  static simdjson_inline void log_start() {
+    if (LOG_ENABLED) {
+      log_depth = 0;
+      printf("\n");
+      printf("| %-*s | %-*s | %-*s | %-*s | Detail |\n", LOG_EVENT_LEN, "Event", LOG_BUFFER_LEN, "Buffer", LOG_SMALL_BUFFER_LEN, "Next", 5, "Next#");
+      printf("|%.*s|%.*s|%.*s|%.*s|--------|\n", LOG_EVENT_LEN+2, DASHES, LOG_BUFFER_LEN+2, DASHES, LOG_SMALL_BUFFER_LEN+2, DASHES, 5+2, DASHES);
+    }
+  }
+
+  simdjson_unused static simdjson_inline void log_string(const char *message) {
+    if (LOG_ENABLED) {
+      printf("%s\n", message);
+    }
+  }
+
+  // Logs a single line from the stage 2 DOM parser
+  template<typename S>
+  static simdjson_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
+    if (LOG_ENABLED) {
+      printf("| %*s%s%-*s ", log_depth*2, "", title_prefix, LOG_EVENT_LEN - log_depth*2 - int(strlen(title_prefix)), title);
+      auto current_index = structurals.at_beginning() ? nullptr : structurals.next_structural-1;
+      auto next_index = structurals.next_structural;
+      auto current = current_index ? &structurals.buf[*current_index] : reinterpret_cast<const uint8_t*>("                                                       ");
+      auto next = &structurals.buf[*next_index];
+      {
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_BUFFER_LEN;i++) {
+          printf("%c", printable_char(current[i]));
+        }
+        printf(" ");
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_SMALL_BUFFER_LEN;i++) {
+          printf("%c", printable_char(next[i]));
+        }
+        printf(" ");
+      }
+      if (current_index) {
+        printf("| %*u ", LOG_INDEX_LEN, *current_index);
+      } else {
+        printf("| %-*s ", LOG_INDEX_LEN, "");
+      }
+      // printf("| %*u ", LOG_INDEX_LEN, structurals.next_tape_index());
+      printf("| %-s ", detail);
+      printf("|\n");
+    }
+  }
+
+} // namespace logger
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+/* end file generic/stage2/logger.h for westmere */
+
+// All other declarations
+/* including generic/stage2/json_iterator.h for westmere: #include <generic/stage2/json_iterator.h> */
+/* begin file generic/stage2/json_iterator.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/logger.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace stage2 {
+
+class json_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+  uint32_t depth{0};
+
+  /**
+   * Walk the JSON document.
+   *
+   * The visitor receives callbacks when values are encountered. All callbacks pass the iterator as
+   * the first parameter; some callbacks have other parameters as well:
+   *
+   * - visit_document_start() - at the beginning.
+   * - visit_document_end() - at the end (if things were successful).
+   *
+   * - visit_array_start() - at the start `[` of a non-empty array.
+   * - visit_array_end() - at the end `]` of a non-empty array.
+   * - visit_empty_array() - when an empty array is encountered.
+   *
+   * - visit_object_end() - at the start `]` of a non-empty object.
+   * - visit_object_start() - at the end `]` of a non-empty object.
+   * - visit_empty_object() - when an empty object is encountered.
+   * - visit_key(const uint8_t *key) - when a key in an object field is encountered. key is
+   *                                   guaranteed to point at the first quote of the string (`"key"`).
+   * - visit_primitive(const uint8_t *value) - when a value is a string, number, boolean or null.
+   * - visit_root_primitive(iter, uint8_t *value) - when the top-level value is a string, number, boolean or null.
+   *
+   * - increment_count(iter) - each time a value is found in an array or object.
+   */
+  template<bool STREAMING, typename V>
+  simdjson_warn_unused simdjson_inline error_code walk_document(V &visitor) noexcept;
+
+  /**
+   * Create an iterator capable of walking a JSON document.
+   *
+   * The document must have already passed through stage 1.
+   */
+  simdjson_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
+
+  /**
+   * Look at the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *peek() const noexcept;
+  /**
+   * Advance to the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *advance() noexcept;
+  /**
+   * Get the remaining length of the document, from the start of the current token.
+   */
+  simdjson_inline size_t remaining_len() const noexcept;
+  /**
+   * Check if we are at the end of the document.
+   *
+   * If this is true, there are no more tokens.
+   */
+  simdjson_inline bool at_eof() const noexcept;
+  /**
+   * Check if we are at the beginning of the document.
+   */
+  simdjson_inline bool at_beginning() const noexcept;
+  simdjson_inline uint8_t last_structural() const noexcept;
+
+  /**
+   * Log that a value has been found.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_value(const char *type) const noexcept;
+  /**
+   * Log the start of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_start_value(const char *type) const noexcept;
+  /**
+   * Log the end of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_end_value(const char *type) const noexcept;
+  /**
+   * Log an error.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_error(const char *error) const noexcept;
+
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
+};
+
+template<bool STREAMING, typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::walk_document(V &visitor) noexcept {
+  logger::log_start();
+
+  //
+  // Start the document
+  //
+  if (at_eof()) { return EMPTY; }
+  log_start_value("document");
+  SIMDJSON_TRY( visitor.visit_document_start(*this) );
+
+  //
+  // Read first value
+  //
+  {
+    auto value = advance();
+
+    // Make sure the outer object or array is closed before continuing; otherwise, there are ways we
+    // could get into memory corruption. See https://github.com/simdjson/simdjson/issues/906
+    if (!STREAMING) {
+      switch (*value) {
+        case '{': if (last_structural() != '}') { log_value("starting brace unmatched"); return TAPE_ERROR; }; break;
+        case '[': if (last_structural() != ']') { log_value("starting bracket unmatched"); return TAPE_ERROR; }; break;
+      }
+    }
+
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_root_primitive(*this, value) ); break;
+    }
+  }
+  goto document_end;
+
+//
+// Object parser states
+//
+object_begin:
+  log_start_value("object");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = false;
+  SIMDJSON_TRY( visitor.visit_object_start(*this) );
+
+  {
+    auto key = advance();
+    if (*key != '"') { log_error("Object does not start with a key"); return TAPE_ERROR; }
+    SIMDJSON_TRY( visitor.increment_count(*this) );
+    SIMDJSON_TRY( visitor.visit_key(*this, key) );
+  }
+
+object_field:
+  if (simdjson_unlikely( *advance() != ':' )) { log_error("Missing colon after key in object"); return TAPE_ERROR; }
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+object_continue:
+  switch (*advance()) {
+    case ',':
+      SIMDJSON_TRY( visitor.increment_count(*this) );
+      {
+        auto key = advance();
+        if (simdjson_unlikely( *key != '"' )) { log_error("Key string missing at beginning of field in object"); return TAPE_ERROR; }
+        SIMDJSON_TRY( visitor.visit_key(*this, key) );
+      }
+      goto object_field;
+    case '}': log_end_value("object"); SIMDJSON_TRY( visitor.visit_object_end(*this) ); goto scope_end;
+    default: log_error("No comma between object fields"); return TAPE_ERROR;
+  }
+
+scope_end:
+  depth--;
+  if (depth == 0) { goto document_end; }
+  if (dom_parser.is_array[depth]) { goto array_continue; }
+  goto object_continue;
+
+//
+// Array parser states
+//
+array_begin:
+  log_start_value("array");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = true;
+  SIMDJSON_TRY( visitor.visit_array_start(*this) );
+  SIMDJSON_TRY( visitor.increment_count(*this) );
+
+array_value:
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+array_continue:
+  switch (*advance()) {
+    case ',': SIMDJSON_TRY( visitor.increment_count(*this) ); goto array_value;
+    case ']': log_end_value("array"); SIMDJSON_TRY( visitor.visit_array_end(*this) ); goto scope_end;
+    default: log_error("Missing comma between array values"); return TAPE_ERROR;
+  }
+
+document_end:
+  log_end_value("document");
+  SIMDJSON_TRY( visitor.visit_document_end(*this) );
+
+  dom_parser.next_structural_index = uint32_t(next_structural - &dom_parser.structural_indexes[0]);
+
+  // If we didn't make it to the end, it's an error
+  if ( !STREAMING && dom_parser.next_structural_index != dom_parser.n_structural_indexes ) {
+    log_error("More than one JSON value at the root of the document, or extra characters at the end of the JSON!");
+    return TAPE_ERROR;
+  }
+
+  return SUCCESS;
+
+} // walk_document()
+
+simdjson_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+  : buf{_dom_parser.buf},
+    next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+    dom_parser{_dom_parser} {
+}
+
+simdjson_inline const uint8_t *json_iterator::peek() const noexcept {
+  return &buf[*(next_structural)];
+}
+simdjson_inline const uint8_t *json_iterator::advance() noexcept {
+  return &buf[*(next_structural++)];
+}
+simdjson_inline size_t json_iterator::remaining_len() const noexcept {
+  return dom_parser.len - *(next_structural-1);
+}
+
+simdjson_inline bool json_iterator::at_eof() const noexcept {
+  return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+}
+simdjson_inline bool json_iterator::at_beginning() const noexcept {
+  return next_structural == dom_parser.structural_indexes.get();
+}
+simdjson_inline uint8_t json_iterator::last_structural() const noexcept {
+  return buf[dom_parser.structural_indexes[dom_parser.n_structural_indexes - 1]];
+}
+
+simdjson_inline void json_iterator::log_value(const char *type) const noexcept {
+  logger::log_line(*this, "", type, "");
+}
+
+simdjson_inline void json_iterator::log_start_value(const char *type) const noexcept {
+  logger::log_line(*this, "+", type, "");
+  if (logger::LOG_ENABLED) { logger::log_depth++; }
+}
+
+simdjson_inline void json_iterator::log_end_value(const char *type) const noexcept {
+  if (logger::LOG_ENABLED) { logger::log_depth--; }
+  logger::log_line(*this, "-", type, "");
+}
+
+simdjson_inline void json_iterator::log_error(const char *error) const noexcept {
+  logger::log_line(*this, "", "ERROR", error);
+}
+
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
+  switch (*value) {
+    case '"': return visitor.visit_root_string(*this, value);
+    case 't': return visitor.visit_root_true_atom(*this, value);
+    case 'f': return visitor.visit_root_false_atom(*this, value);
+    case 'n': return visitor.visit_root_null_atom(*this, value);
+    case '-':
+    case '0': case '1': case '2': case '3': case '4':
+    case '5': case '6': case '7': case '8': case '9':
+      return visitor.visit_root_number(*this, value);
+    default:
+      log_error("Document starts with a non-value character");
+      return TAPE_ERROR;
+  }
+}
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
+  // Use the fact that most scalars are going to be either strings or numbers.
+  if(*value == '"') {
+    return visitor.visit_string(*this, value);
+  } else if (((*value - '0')  < 10) || (*value == '-')) {
+    return visitor.visit_number(*this, value);
+  }
+  // true, false, null are uncommon.
+  switch (*value) {
+    case 't': return visitor.visit_true_atom(*this, value);
+    case 'f': return visitor.visit_false_atom(*this, value);
+    case 'n': return visitor.visit_null_atom(*this, value);
+    default:
+      log_error("Non-value found when value was expected!");
+      return TAPE_ERROR;
+  }
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+/* end file generic/stage2/json_iterator.h for westmere */
+/* including generic/stage2/stringparsing.h for westmere: #include <generic/stage2/stringparsing.h> */
+/* begin file generic/stage2/stringparsing.h for westmere */
+#include <cstdint>
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/jsoncharutils.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses
+// It is intended to be included multiple times and compiled multiple times
+
+namespace simdjson {
+namespace westmere {
+namespace {
+/// @private
+namespace stringparsing {
+
+// begin copypasta
+// These chars yield themselves: " \ /
+// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
+// u not handled in this table as it's complex
+static const uint8_t escape_map[256] = {
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
+    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
+    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+};
+
+// handle a unicode codepoint
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr, bool allow_replacement) {
+  // Use the default Unicode Character 'REPLACEMENT CHARACTER' (U+FFFD)
+  constexpr uint32_t substitution_code_point = 0xfffd;
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) != ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+    } else {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+
+      // We have already checked that the high surrogate is valid and
+      // (code_point - 0xd800) < 1024.
+      //
+      // Check that code_point_2 is in the range 0xdc00..0xdfff
+      // and that code_point_2 was parsed from valid hex.
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if (low_bit >> 10) {
+        if(!allow_replacement) { return false; }
+        code_point = substitution_code_point;
+      } else {
+        code_point =  (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+
+    }
+  } else if (code_point >= 0xdc00 && code_point <= 0xdfff) {
+      // If we encounter a low surrogate (not preceded by a high surrogate)
+      // then we have an error.
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+  }
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+// handle a unicode codepoint using the wobbly convention
+// https://simonsapin.github.io/wtf-8/
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint_wobbly(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr) {
+  // It is not ideal that this function is nearly identical to handle_unicode_codepoint.
+  //
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) == ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if ((low_bit >> 10) ==  0) {
+        code_point =
+          (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+    }
+  }
+
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+/**
+ * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+ * must be an unescaped quote terminating the string. It returns the final output
+ * position as pointer. In case of error (e.g., the string has bad escaped codes),
+ * then null_ptr is returned. It is assumed that the output buffer is large
+ * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+ * SIMDJSON_PADDING bytes.
+ */
+simdjson_warn_unused simdjson_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) {
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint(&src, &dst, allow_replacement)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+simdjson_warn_unused simdjson_inline uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) {
+  // It is not ideal that this function is nearly identical to parse_string.
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint_wobbly(&src, &dst)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+} // namespace stringparsing
+
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+/* end file generic/stage2/stringparsing.h for westmere */
+/* including generic/stage2/structural_iterator.h for westmere: #include <generic/stage2/structural_iterator.h> */
+/* begin file generic/stage2/structural_iterator.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace stage2 {
+
+class structural_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+
+  // Start a structural
+  simdjson_inline structural_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+    : buf{_dom_parser.buf},
+      next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+      dom_parser{_dom_parser} {
+  }
+  // Get the buffer position of the current structural character
+  simdjson_inline const uint8_t* current() {
+    return &buf[*(next_structural-1)];
+  }
+  // Get the current structural character
+  simdjson_inline char current_char() {
+    return buf[*(next_structural-1)];
+  }
+  // Get the next structural character without advancing
+  simdjson_inline char peek_next_char() {
+    return buf[*next_structural];
+  }
+  simdjson_inline const uint8_t* peek() {
+    return &buf[*next_structural];
+  }
+  simdjson_inline const uint8_t* advance() {
+    return &buf[*(next_structural++)];
+  }
+  simdjson_inline char advance_char() {
+    return buf[*(next_structural++)];
+  }
+  simdjson_inline size_t remaining_len() {
+    return dom_parser.len - *(next_structural-1);
+  }
+
+  simdjson_inline bool at_end() {
+    return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+  }
+  simdjson_inline bool at_beginning() {
+    return next_structural == dom_parser.structural_indexes.get();
+  }
+};
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H
+/* end file generic/stage2/structural_iterator.h for westmere */
+/* including generic/stage2/tape_builder.h for westmere: #include <generic/stage2/tape_builder.h> */
+/* begin file generic/stage2/tape_builder.h for westmere */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/json_iterator.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/stringparsing.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/tape_writer.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/dom/document.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/atomparsing.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/numberparsing.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+
+namespace simdjson {
+namespace westmere {
+namespace {
+namespace stage2 {
+
+struct tape_builder {
+  template<bool STREAMING>
+  simdjson_warn_unused static simdjson_inline error_code parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept;
+
+  /** Called when a non-empty document starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty document ends without error. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_end(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty array starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty array ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_end(json_iterator &iter) noexcept;
+  /** Called when an empty array is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_array(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty object starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_start(json_iterator &iter) noexcept;
+  /**
+   * Called when a key in a field is encountered.
+   *
+   * primitive, visit_object_start, visit_empty_object, visit_array_start, or visit_empty_array
+   * will be called after this with the field value.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
+  /** Called when a non-empty object ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_end(json_iterator &iter) noexcept;
+  /** Called when an empty object is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_object(json_iterator &iter) noexcept;
+
+  /**
+   * Called when a string, number, boolean or null is found.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  /**
+   * Called when a string, number, boolean or null is found at the top level of a document (i.e.
+   * when there is no array or object and the entire document is a single string, number, boolean or
+   * null.
+   *
+   * This is separate from primitive() because simdjson's normal primitive parsing routines assume
+   * there is at least one more token after the value, which is only true in an array or object.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  /** Called each time a new field or element in an array or object is found. */
+  simdjson_warn_unused simdjson_inline error_code increment_count(json_iterator &iter) noexcept;
+
+  /** Next location to write to tape */
+  tape_writer tape;
+private:
+  /** Next write location in the string buf for stage 2 parsing */
+  uint8_t *current_string_buf_loc;
+
+  simdjson_inline tape_builder(dom::document &doc) noexcept;
+
+  simdjson_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
+  simdjson_inline void start_container(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_warn_unused simdjson_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
+  simdjson_inline void on_end_string(uint8_t *dst) noexcept;
+}; // struct tape_builder
+
+template<bool STREAMING>
+simdjson_warn_unused simdjson_inline error_code tape_builder::parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept {
+  dom_parser.doc = &doc;
+  json_iterator iter(dom_parser, STREAMING ? dom_parser.next_structural_index : 0);
+  tape_builder builder(doc);
+  return iter.walk_document<STREAMING>(builder);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_root_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
+  constexpr uint32_t start_tape_index = 0;
+  tape.append(start_tape_index, internal::tape_type::ROOT);
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter), internal::tape_type::ROOT);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
+  return visit_string(iter, key, true);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].count++; // we have a key value pair in the object at parser.dom_parser.depth - 1
+  return SUCCESS;
+}
+
+simdjson_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
+  iter.log_value(key ? "key" : "string");
+  uint8_t *dst = on_start_string(iter);
+  dst = stringparsing::parse_string(value+1, dst, false); // We do not allow replacement when the escape characters are invalid.
+  if (dst == nullptr) {
+    iter.log_error("Invalid escape in string");
+    return STRING_ERROR;
+  }
+  on_end_string(dst);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
+  return visit_string(iter, value);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("number");
+  error_code err = numberparsing::parse_number(value, tape);
+  if (simdjson_unlikely(err == BIGINT_ERROR &&
+      iter.dom_parser._number_as_string)) {
+    // Write big integer to string buffer using the same format as strings.
+    // Scan digits the same way parse_number does (skip optional '-', then digits).
+    const uint8_t *p = value;
+    if (*p == '-') p++;
+    while (numberparsing::is_digit(*p)) p++;
+    size_t len = size_t(p - value);
+    tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::BIGINT);
+    uint8_t *dst = current_string_buf_loc + sizeof(uint32_t);
+    memcpy(dst, value, len);
+    dst += len;
+    on_end_string(dst);
+    return SUCCESS;
+  }
+  return err;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
+  //
+  // We need to make a copy to make sure that the string is space terminated.
+  // This is not about padding the input, which should already padded up
+  // to len + SIMDJSON_PADDING. However, we have no control at this stage
+  // on how the padding was done. What if the input string was padded with nulls?
+  // It is quite common for an input string to have an extra null character (C string).
+  // We do not want to allow 9\0 (where \0 is the null character) inside a JSON
+  // document, but the string "9\0" by itself is fine. So we make a copy and
+  // pad the input with spaces when we know that there is just one input element.
+  // This copy is relatively expensive, but it will almost never be called in
+  // practice unless you are in the strange scenario where you have many JSON
+  // documents made of single atoms.
+  //
+  std::unique_ptr<uint8_t[]>copy(new (std::nothrow) uint8_t[iter.remaining_len() + SIMDJSON_PADDING]);
+  if (copy.get() == nullptr) { return MEMALLOC; }
+  std::memcpy(copy.get(), value, iter.remaining_len());
+  std::memset(copy.get() + iter.remaining_len(), ' ', SIMDJSON_PADDING);
+  error_code error = visit_number(iter, copy.get());
+  return error;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value)) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value, iter.remaining_len())) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value)) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value, iter.remaining_len())) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value)) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value, iter.remaining_len())) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+// private:
+
+simdjson_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
+  return uint32_t(tape.next_tape_loc - iter.dom_parser.doc->tape.get());
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  auto start_index = next_tape_index(iter);
+  tape.append(start_index+2, start);
+  tape.append(start_index, end);
+  return SUCCESS;
+}
+
+simdjson_inline void tape_builder::start_container(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].tape_index = next_tape_index(iter);
+  iter.dom_parser.open_containers[iter.depth].count = 0;
+  tape.skip(); // We don't actually *write* the start element until the end.
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  // Write the ending tape element, pointing at the start location
+  const uint32_t start_tape_index = iter.dom_parser.open_containers[iter.depth].tape_index;
+  tape.append(start_tape_index, end);
+  // Write the start tape element, pointing at the end location (and including count)
+  // count can overflow if it exceeds 24 bits... so we saturate
+  // the convention being that a cnt of 0xffffff or more is undetermined in value (>=  0xffffff).
+  const uint32_t count = iter.dom_parser.open_containers[iter.depth].count;
+  const uint32_t cntsat = count > 0xFFFFFF ? 0xFFFFFF : count;
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter) | (uint64_t(cntsat) << 32), start);
+  return SUCCESS;
+}
+
+simdjson_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
+  // we advance the point, accounting for the fact that we have a NULL termination
+  tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::STRING);
+  return current_string_buf_loc + sizeof(uint32_t);
+}
+
+simdjson_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
+  uint32_t str_length = uint32_t(dst - (current_string_buf_loc + sizeof(uint32_t)));
+  // TODO check for overflow in case someone has a crazy string (>=4GB?)
+  // But only add the overflow check when the document itself exceeds 4GB
+  // Currently unneeded because we refuse to parse docs larger or equal to 4GB.
+  memcpy(current_string_buf_loc, &str_length, sizeof(uint32_t));
+  // NULL termination is still handy if you expect all your strings to
+  // be NULL terminated? It comes at a small cost
+  *dst = 0;
+  current_string_buf_loc = dst + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+/* end file generic/stage2/tape_builder.h for westmere */
+/* end file generic/stage2/amalgamated.h for westmere */
+
+//
+// Stage 1
+//
+
+namespace simdjson {
+namespace westmere {
+
+simdjson_warn_unused error_code implementation::create_dom_parser_implementation(
+  size_t capacity,
+  size_t max_depth,
+  std::unique_ptr<internal::dom_parser_implementation>& dst
+) const noexcept {
+  dst.reset( new (std::nothrow) dom_parser_implementation() );
+  if (!dst) { return MEMALLOC; }
+  if (auto err = dst->set_capacity(capacity))
+    return err;
+  if (auto err = dst->set_max_depth(max_depth))
+    return err;
+  return SUCCESS;
+}
+
+namespace {
+
+using namespace simd;
+
+simdjson_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
+  // These lookups rely on the fact that anything < 127 will match the lower 4 bits, which is why
+  // we can't use the generic lookup_16.
+  auto whitespace_table = simd8<uint8_t>::repeat_16(' ', 100, 100, 100, 17, 100, 113, 2, 100, '\t', '\n', 112, 100, '\r', 100, 100);
+
+  // The 6 operators (:,[]{}) have these values:
+  //
+  // , 2C
+  // : 3A
+  // [ 5B
+  // { 7B
+  // ] 5D
+  // } 7D
+  //
+  // If you use | 0x20 to turn [ and ] into { and }, the lower 4 bits of each character is unique.
+  // We exploit this, using a simd 4-bit lookup to tell us which character match against, and then
+  // match it (against | 0x20).
+  //
+  // To prevent recognizing other characters, everything else gets compared with 0, which cannot
+  // match due to the | 0x20.
+  //
+  // NOTE: Due to the | 0x20, this ALSO treats <FF> and <SUB> (control characters 0C and 1A) like ,
+  // and :. This gets caught in stage 2, which checks the actual character to ensure the right
+  // operators are in the right places.
+  const auto op_table = simd8<uint8_t>::repeat_16(
+    0, 0, 0, 0,
+    0, 0, 0, 0,
+    0, 0, ':', '{', // : = 3A, [ = 5B, { = 7B
+    ',', '}', 0, 0  // , = 2C, ] = 5D, } = 7D
+  );
+
+  // We compute whitespace and op separately. If the code later only use one or the
+  // other, given the fact that all functions are aggressively inlined, we can
+  // hope that useless computations will be omitted. This is namely case when
+  // minifying (we only need whitespace).
+
+
+  const uint64_t whitespace = in.eq({
+    _mm_shuffle_epi8(whitespace_table, in.chunks[0]),
+    _mm_shuffle_epi8(whitespace_table, in.chunks[1]),
+    _mm_shuffle_epi8(whitespace_table, in.chunks[2]),
+    _mm_shuffle_epi8(whitespace_table, in.chunks[3])
+  });
+  // Turn [ and ] into { and }
+  const simd8x64<uint8_t> curlified{
+    in.chunks[0] | 0x20,
+    in.chunks[1] | 0x20,
+    in.chunks[2] | 0x20,
+    in.chunks[3] | 0x20
+  };
+  const uint64_t op = curlified.eq({
+    _mm_shuffle_epi8(op_table, in.chunks[0]),
+    _mm_shuffle_epi8(op_table, in.chunks[1]),
+    _mm_shuffle_epi8(op_table, in.chunks[2]),
+    _mm_shuffle_epi8(op_table, in.chunks[3])
+  });
+    return { whitespace, op };
+}
+
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input) {
+  return input.reduce_or().is_ascii();
+}
+
+simdjson_unused simdjson_inline simd8<bool> must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_second_byte = prev1.saturating_sub(0xc0u-1); // Only 11______ will be > 0
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-1); // Only 111_____ will be > 0
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-1); // Only 1111____ will be > 0
+  // Caller requires a bool (all 1's). All values resulting from the subtraction will be <= 64, so signed comparison is fine.
+  return simd8<int8_t>(is_second_byte | is_third_byte | is_fourth_byte) > int8_t(0);
+}
+
+simdjson_inline simd8<uint8_t> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-0x80); // Only 111_____ will be >= 0x80
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-0x80); // Only 1111____ will be >= 0x80
+  return is_third_byte | is_fourth_byte;
+}
+
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+
+//
+// Stage 2
+//
+
+//
+// Implementation-specific overrides
+//
+
+namespace simdjson {
+namespace westmere {
+
+simdjson_warn_unused error_code implementation::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept {
+  return westmere::stage1::json_minifier::minify<64>(buf, len, dst, dst_len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage1(const uint8_t *_buf, size_t _len, stage1_mode streaming) noexcept {
+  this->buf = _buf;
+  this->len = _len;
+  return westmere::stage1::json_structural_indexer::index<64>(_buf, _len, *this, streaming);
+}
+
+simdjson_warn_unused bool implementation::validate_utf8(const char *buf, size_t len) const noexcept {
+  return westmere::stage1::generic_validate_utf8(buf,len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<false>(*this, _doc);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2_next(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<true>(*this, _doc);
+}
+
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_string(const uint8_t *src, uint8_t *dst, bool replacement_char) const noexcept {
+  return westmere::stringparsing::parse_string(src, dst, replacement_char);
+}
+
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept {
+  return westmere::stringparsing::parse_wobbly_string(src, dst);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::parse(const uint8_t *_buf, size_t _len, dom::document &_doc) noexcept {
+  auto error = stage1(_buf, _len, stage1_mode::regular);
+  if (error) { return error; }
+  return stage2(_doc);
+}
+} // namespace westmere
+} // namespace simdjson
+
+/* including simdjson/westmere/end.h: #include <simdjson/westmere/end.h> */
+/* begin file simdjson/westmere/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#if !SIMDJSON_CAN_ALWAYS_RUN_WESTMERE
+SIMDJSON_UNTARGET_REGION
+#endif
+
+/* undefining SIMDJSON_IMPLEMENTATION from "westmere" */
+#undef SIMDJSON_IMPLEMENTATION
+/* end file simdjson/westmere/end.h */
+
+#endif // SIMDJSON_SRC_WESTMERE_CPP
+/* end file westmere.cpp */
+#endif
+#if SIMDJSON_IMPLEMENTATION_LASX
+/* including lasx.cpp: #include <lasx.cpp> */
+/* begin file lasx.cpp */
+#ifndef SIMDJSON_SRC_LASX_CPP
+#define SIMDJSON_SRC_LASX_CPP
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+/* including simdjson/lasx.h: #include <simdjson/lasx.h> */
+/* begin file simdjson/lasx.h */
+#ifndef SIMDJSON_LASX_H
+#define SIMDJSON_LASX_H
+
+/* including simdjson/lasx/begin.h: #include "simdjson/lasx/begin.h" */
+/* begin file simdjson/lasx/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "lasx" */
+#define SIMDJSON_IMPLEMENTATION lasx
+#include <lsxintrin.h> // This is a hack. We should not need to put this include here.
+#if SIMDJSON_CAN_ALWAYS_RUN_LASX
+// nothing needed.
+#else
+SIMDJSON_TARGET_REGION("lasx,lsx")
+#endif
+
+/* including simdjson/lasx/base.h: #include "simdjson/lasx/base.h" */
+/* begin file simdjson/lasx/base.h */
+#ifndef SIMDJSON_LASX_BASE_H
+#define SIMDJSON_LASX_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+/**
+ * Implementation for LASX.
+ */
+namespace lasx {
+
+class implementation;
+
+namespace {
+namespace simd {
+template <typename T> struct simd8;
+template <typename T> struct simd8x64;
+} // namespace simd
+} // unnamed namespace
+
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_LASX_BASE_H
+/* end file simdjson/lasx/base.h */
+/* including simdjson/lasx/intrinsics.h: #include "simdjson/lasx/intrinsics.h" */
+/* begin file simdjson/lasx/intrinsics.h */
+#ifndef SIMDJSON_LASX_INTRINSICS_H
+#define SIMDJSON_LASX_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <lsxintrin.h>
+#include <lasxintrin.h>
+
+static_assert(sizeof(__m256i) <= simdjson::SIMDJSON_PADDING, "insufficient padding for LoongArch ASX");
+
+#endif //  SIMDJSON_LASX_INTRINSICS_H
+/* end file simdjson/lasx/intrinsics.h */
+/* including simdjson/lasx/bitmanipulation.h: #include "simdjson/lasx/bitmanipulation.h" */
+/* begin file simdjson/lasx/bitmanipulation.h */
+#ifndef SIMDJSON_LASX_BITMANIPULATION_H
+#define SIMDJSON_LASX_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/bitmask.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+  return __builtin_ctzll(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return input_num & (input_num-1);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+  return __builtin_clzll(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int count_ones(uint64_t input_num) {
+  return __lasx_xvpickve2gr_w(__lasx_xvpcnt_d(__m256i(v4u64{input_num, 0, 0, 0})), 0);
+}
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2, uint64_t *result) {
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+}
+
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_LASX_BITMANIPULATION_H
+/* end file simdjson/lasx/bitmanipulation.h */
+/* including simdjson/lasx/bitmask.h: #include "simdjson/lasx/bitmask.h" */
+/* begin file simdjson/lasx/bitmask.h */
+#ifndef SIMDJSON_LASX_BITMASK_H
+#define SIMDJSON_LASX_BITMASK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(uint64_t bitmask) {
+  bitmask ^= bitmask << 1;
+  bitmask ^= bitmask << 2;
+  bitmask ^= bitmask << 4;
+  bitmask ^= bitmask << 8;
+  bitmask ^= bitmask << 16;
+  bitmask ^= bitmask << 32;
+  return bitmask;
+}
+
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif
+/* end file simdjson/lasx/bitmask.h */
+/* including simdjson/lasx/numberparsing_defs.h: #include "simdjson/lasx/numberparsing_defs.h" */
+/* begin file simdjson/lasx/numberparsing_defs.h */
+#ifndef SIMDJSON_LASX_NUMBERPARSING_DEFS_H
+#define SIMDJSON_LASX_NUMBERPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace lasx {
+namespace numberparsing {
+
+// we don't have appropriate instructions, so let us use a scalar function
+// credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  uint64_t val;
+  std::memcpy(&val, chars, sizeof(uint64_t));
+  val = (val & 0x0F0F0F0F0F0F0F0F) * 2561 >> 8;
+  val = (val & 0x00FF00FF00FF00FF) * 6553601 >> 16;
+  return uint32_t((val & 0x0000FFFF0000FFFF) * 42949672960001 >> 32);
+}
+
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace lasx
+} // namespace simdjson
+
+#ifndef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_IS_BIG_ENDIAN
+#define SIMDJSON_SWAR_NUMBER_PARSING 0
+#else
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+#endif
+#endif
+
+#endif // SIMDJSON_LASX_NUMBERPARSING_DEFS_H
+/* end file simdjson/lasx/numberparsing_defs.h */
+/* including simdjson/lasx/simd.h: #include "simdjson/lasx/simd.h" */
+/* begin file simdjson/lasx/simd.h */
+#ifndef SIMDJSON_LASX_SIMD_H
+#define SIMDJSON_LASX_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace simd {
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename Child>
+  struct base {
+    __m256i value;
+
+    // Zero constructor
+    simdjson_inline base() : value{__m256i()} {}
+
+    // Conversion from SIMD register
+    simdjson_inline base(const __m256i _value) : value(_value) {}
+
+    // Conversion to SIMD register
+    simdjson_inline operator const __m256i&() const { return this->value; }
+    simdjson_inline operator __m256i&() { return this->value; }
+    simdjson_inline operator const v32i8&() const { return (v32i8&)this->value; }
+    simdjson_inline operator v32i8&() { return (v32i8&)this->value; }
+
+    // Bit operations
+    simdjson_inline Child operator|(const Child other) const { return __lasx_xvor_v(*this, other); }
+    simdjson_inline Child operator&(const Child other) const { return __lasx_xvand_v(*this, other); }
+    simdjson_inline Child operator^(const Child other) const { return __lasx_xvxor_v(*this, other); }
+    simdjson_inline Child bit_andnot(const Child other) const { return __lasx_xvandn_v(other, *this); }
+    simdjson_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+  };
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename T>
+  struct simd8;
+
+  template<typename T, typename Mask=simd8<bool>>
+  struct base8: base<simd8<T>> {
+    simdjson_inline base8() : base<simd8<T>>() {}
+    simdjson_inline base8(const __m256i _value) : base<simd8<T>>(_value) {}
+
+    friend simdjson_really_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return __lasx_xvseq_b(lhs, rhs); }
+
+    static const int SIZE = sizeof(base<simd8<T>>::value);
+
+    template<int N=1>
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+        __m256i hi = __lasx_xvbsll_v(*this, N);
+        __m256i lo = __lasx_xvbsrl_v(*this, 16 - N);
+        __m256i tmp = __lasx_xvbsrl_v(prev_chunk, 16 - N);
+        lo = __lasx_xvpermi_q(lo, tmp, 0x21);
+        return __lasx_xvor_v(hi, lo);
+    }
+  };
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool>: base8<bool> {
+    static simdjson_inline simd8<bool> splat(bool _value) { return __lasx_xvreplgr2vr_b(uint8_t(-(!!_value))); }
+
+    simdjson_inline simd8() : base8() {}
+    simdjson_inline simd8(const __m256i _value) : base8<bool>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+
+    simdjson_inline int to_bitmask() const {
+      __m256i mask = __lasx_xvmskltz_b(*this);
+      return (__lasx_xvpickve2gr_w(mask, 4) << 16) | (__lasx_xvpickve2gr_w(mask, 0));
+    }
+   simdjson_inline bool any() const {
+      __m256i v = __lasx_xvmsknz_b(*this);
+      return (0 == __lasx_xvpickve2gr_w(v, 0)) && (0 == __lasx_xvpickve2gr_w(v, 4));
+    }
+    simdjson_inline simd8<bool> operator~() const { return *this ^ true; }
+  };
+
+  template<typename T>
+  struct base8_numeric: base8<T> {
+    static simdjson_inline simd8<T> splat(T _value) {
+      return __lasx_xvreplgr2vr_b(_value);
+    }
+    static simdjson_inline simd8<T> zero() { return __lasx_xvldi(0); }
+    static simdjson_inline simd8<T> load(const T values[32]) {
+      return __lasx_xvld(reinterpret_cast<const __m256i *>(values), 0);
+    }
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    static simdjson_inline simd8<T> repeat_16(
+      T v0,  T v1,  T v2,  T v3,  T v4,  T v5,  T v6,  T v7,
+      T v8,  T v9,  T v10, T v11, T v12, T v13, T v14, T v15
+    ) {
+      return simd8<T>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    simdjson_inline base8_numeric() : base8<T>() {}
+    simdjson_inline base8_numeric(const __m256i _value) : base8<T>(_value) {}
+
+    // Store to array
+    simdjson_inline void store(T dst[32]) const {
+      return __lasx_xvst(*this, reinterpret_cast<__m256i *>(dst), 0);
+    }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<T> operator+(const simd8<T> other) const { return __lasx_xvadd_b(*this, other); }
+    simdjson_inline simd8<T> operator-(const simd8<T> other) const { return __lasx_xvsub_b(*this, other); }
+    simdjson_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
+
+    // Override to distinguish from bool version
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return __lasx_xvshuf_b(lookup_table, lookup_table, *this);
+    }
+
+    // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted as a bitset).
+    // Passing a 0 value for mask would be equivalent to writing out every byte to output.
+    // Only the first 16 - count_ones(mask) bytes of the result are significant but 16 bytes
+    // get written.
+    template<typename L>
+    simdjson_inline void compress(uint32_t mask, L * output) const {
+      using internal::thintable_epi8;
+      using internal::BitsSetTable256mul2;
+      using internal::pshufb_combine_table;
+      // this particular implementation was inspired by haswell
+      // lasx do it in 4 steps, first 8 bytes and then second 8 bytes...
+      uint8_t mask1 = uint8_t(mask); // least significant 8 bits
+      uint8_t mask2 = uint8_t(mask >> 8); // second significant 8 bits
+      uint8_t mask3 = uint8_t(mask >> 16); // ...
+      uint8_t mask4 = uint8_t(mask >> 24); // ...
+      // next line just loads the 64-bit values thintable_epi8[mask{1,2,3,4}]
+      // into a 256-bit register.
+      __m256i shufmask = {int64_t(thintable_epi8[mask1]), int64_t(thintable_epi8[mask2]) + 0x0808080808080808, int64_t(thintable_epi8[mask3]), int64_t(thintable_epi8[mask4]) + 0x0808080808080808};
+      // this is the version "nearly pruned"
+      __m256i pruned = __lasx_xvshuf_b(*this, *this, shufmask);
+      // we still need to put the  pieces back together.
+      // we compute the popcount of the first words:
+      int pop1 = BitsSetTable256mul2[mask1];
+      int pop2 = BitsSetTable256mul2[mask2];
+      int pop3 = BitsSetTable256mul2[mask3];
+
+      // then load the corresponding mask
+      __m256i masklo = __lasx_xvldx(reinterpret_cast<void*>(reinterpret_cast<unsigned long>(pshufb_combine_table)), pop1 * 8);
+      __m256i maskhi = __lasx_xvldx(reinterpret_cast<void*>(reinterpret_cast<unsigned long>(pshufb_combine_table)), pop3 * 8);
+      __m256i compactmask = __lasx_xvpermi_q(maskhi, masklo, 0x20);
+      __m256i answer = __lasx_xvshuf_b(pruned, pruned, compactmask);
+      __lasx_xvst(answer, reinterpret_cast<uint8_t*>(output), 0);
+      uint64_t value3 = __lasx_xvpickve2gr_du(answer, 2);
+      uint64_t value4 = __lasx_xvpickve2gr_du(answer, 3);
+      uint64_t *pos = reinterpret_cast<uint64_t*>(reinterpret_cast<uint8_t*>(output) + 16 - (pop1 + pop2) / 2);
+      pos[0] = value3;
+      pos[1] = value4;
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> : base8_numeric<int8_t> {
+    simdjson_inline simd8() : base8_numeric<int8_t>() {}
+    simdjson_inline simd8(const __m256i _value) : base8_numeric<int8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const int8_t values[32]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15,
+      int8_t v16, int8_t v17, int8_t v18, int8_t v19, int8_t v20, int8_t v21, int8_t v22, int8_t v23,
+      int8_t v24, int8_t v25, int8_t v26, int8_t v27, int8_t v28, int8_t v29, int8_t v30, int8_t v31
+    ) : simd8({
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15,
+      v16,v17,v18,v19,v20,v21,v22,v23,
+      v24,v25,v26,v27,v28,v29,v30,v31
+      }) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return __lasx_xvmax_b(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return __lasx_xvmin_b(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return __lasx_xvslt_b(other, *this); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return __lasx_xvslt_b(*this, other); }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t>: base8_numeric<uint8_t> {
+    simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+    simdjson_inline simd8(const __m256i _value) : base8_numeric<uint8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const uint8_t values[32]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15,
+      uint8_t v16, uint8_t v17, uint8_t v18, uint8_t v19, uint8_t v20, uint8_t v21, uint8_t v22, uint8_t v23,
+      uint8_t v24, uint8_t v25, uint8_t v26, uint8_t v27, uint8_t v28, uint8_t v29, uint8_t v30, uint8_t v31
+    ) : simd8(__m256i(v32u8{
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15,
+      v16,v17,v18,v19,v20,v21,v22,v23,
+      v24,v25,v26,v27,v28,v29,v30,v31
+    })) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return __lasx_xvsadd_bu(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return __lasx_xvssub_bu(*this, other); }
+
+    // Order-specific operations
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return __lasx_xvmax_bu(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return __lasx_xvmin_bu(other, *this); }
+    // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
+    // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->lt_bits(other).any_bits_set(); }
+
+    // Bit-specific operations
+    simdjson_inline simd8<bool> bits_not_set() const { return *this == uint8_t(0); }
+    simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
+    simdjson_inline bool is_ascii() const {
+      __m256i mask = __lasx_xvmskltz_b(*this);
+      return (0 == __lasx_xvpickve2gr_w(mask, 0)) && (0 == __lasx_xvpickve2gr_w(mask, 4));
+    }
+    simdjson_inline bool bits_not_set_anywhere() const {
+      __m256i v = __lasx_xvmsknz_b(*this);
+      return (0 == __lasx_xvpickve2gr_w(v, 0)) && (0 == __lasx_xvpickve2gr_w(v, 4));
+    }
+    simdjson_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
+    simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const {
+      __m256i v = __lasx_xvmsknz_b(__lasx_xvand_v(*this, bits));
+      return (0 == __lasx_xvpickve2gr_w(v, 0)) && (0 == __lasx_xvpickve2gr_w(v, 4));
+    }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(__lasx_xvsrli_b(*this, N)); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(__lasx_xvslli_b(*this, N)); }
+  };
+
+  template<typename T>
+  struct simd8x64 {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+    static_assert(NUM_CHUNKS == 2, "LASX kernel should use two registers per 64-byte block.");
+    const simd8<T> chunks[NUM_CHUNKS];
+    template<int idx> simd8<uint8_t> get() const { return idx < NUM_CHUNKS ? chunks[idx] : simd8<T>(); }
+
+    simd8x64(const simd8x64<T>& o) = delete; // no copy allowed
+    simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1) : chunks{chunk0, chunk1} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+32)} {}
+
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
+      uint32_t mask1 = uint32_t(mask);
+      uint32_t mask2 = uint32_t(mask >> 32);
+      __m256i zcnt = __lasx_xvpcnt_w(__m256i(v4u64{~mask, 0, 0, 0}));
+      uint64_t zcnt1 = __lasx_xvpickve2gr_wu(zcnt, 0);
+      uint64_t zcnt2 = __lasx_xvpickve2gr_wu(zcnt, 1);
+      // There should be a critical value which processes in scaler is faster.
+      if (zcnt1)
+        this->chunks[0].compress(mask1, output);
+      if (zcnt2)
+        this->chunks[1].compress(mask2, output + zcnt1);
+      return zcnt1 + zcnt2;
+    }
+
+    simdjson_inline void store(T ptr[64]) const {
+      this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
+      this->chunks[1].store(ptr+sizeof(simd8<T>)*1);
+    }
+
+    simdjson_inline uint64_t to_bitmask() const {
+      __m256i mask0 = __lasx_xvmskltz_b(this->chunks[0]);
+      __m256i mask1 = __lasx_xvmskltz_b(this->chunks[1]);
+      __m256i mask_tmp = __lasx_xvpickve_w(mask0, 4);
+      __m256i tmp = __lasx_xvpickve_w(mask1, 4);
+      mask0 = __lasx_xvinsve0_w(mask0, mask1, 1);
+      mask_tmp = __lasx_xvinsve0_w(mask_tmp, tmp, 1);
+      return __lasx_xvpickve2gr_du(__lasx_xvpackev_h(mask_tmp, mask0), 0);
+    }
+
+    simdjson_inline simd8<T> reduce_or() const {
+      return this->chunks[0] | this->chunks[1];
+    }
+
+    simdjson_inline uint64_t eq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] == mask,
+        this->chunks[1] == mask
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+      return  simd8x64<bool>(
+        this->chunks[0] == other.chunks[0],
+        this->chunks[1] == other.chunks[1]
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t lteq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] <= mask,
+        this->chunks[1] <= mask
+      ).to_bitmask();
+    }
+  }; // struct simd8x64<T>
+
+} // namespace simd
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_LASX_SIMD_H
+/* end file simdjson/lasx/simd.h */
+/* including simdjson/lasx/stringparsing_defs.h: #include "simdjson/lasx/stringparsing_defs.h" */
+/* begin file simdjson/lasx/stringparsing_defs.h */
+#ifndef SIMDJSON_LASX_STRINGPARSING_DEFS_H
+#define SIMDJSON_LASX_STRINGPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/simd.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 32;
+  simdjson_inline backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return bs_bits != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+
+  uint32_t bs_bits;
+  uint32_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // this can read up to 31 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  return {
+      static_cast<uint32_t>((v == '\\').to_bitmask()),     // bs_bits
+      static_cast<uint32_t>((v == '"').to_bitmask()), // quote_bits
+  };
+}
+
+
+struct escaping {
+  static constexpr uint32_t BYTES_PROCESSED = 16;
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits != 0; }
+  simdjson_inline int escape_index() { return trailing_zeroes(escape_bits); }
+
+  uint64_t escape_bits;
+}; // struct escaping
+
+
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "escaping finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  simd8<bool> is_quote = (v == '"');
+  simd8<bool> is_backslash = (v == '\\');
+  simd8<bool> is_control = (v < 32);
+  return {
+    static_cast<uint64_t>((is_backslash | is_quote | is_control).to_bitmask())
+  };
+}
+
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_LASX_STRINGPARSING_DEFS_H
+/* end file simdjson/lasx/stringparsing_defs.h */
+
+#define SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT 1
+
+
+/* end file simdjson/lasx/begin.h */
+/* including simdjson/generic/amalgamated.h for lasx: #include "simdjson/generic/amalgamated.h" */
+/* begin file simdjson/generic/amalgamated.h for lasx */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_GENERIC_DEPENDENCIES_H)
+#error simdjson/generic/dependencies.h must be included before simdjson/generic/amalgamated.h!
+#endif
+
+/* including simdjson/generic/base.h for lasx: #include "simdjson/generic/base.h" */
+/* begin file simdjson/generic/base.h for lasx */
+#ifndef SIMDJSON_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): // If we haven't got an implementation yet, we're in the editor, editing a generic file! Just */
+/* amalgamation skipped (editor-only): // use the most advanced one we can so the most possible stuff can be tested. */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation_detection.h" */
+/* amalgamation skipped (editor-only): #if SIMDJSON_IMPLEMENTATION_ICELAKE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_HASWELL */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_WESTMERE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_ARM64 */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_PPC64 */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LASX */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LSX */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_RVV_VLS */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_FALLBACK */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/begin.h" */
+/* amalgamation skipped (editor-only): #else */
+/* amalgamation skipped (editor-only): #error "All possible implementations (including fallback) have been disabled! simdjson will not run." */
+/* amalgamation skipped (editor-only): #endif */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+
+struct open_container;
+class dom_parser_implementation;
+
+/**
+ * The type of a JSON number
+ */
+enum class number_type {
+    floating_point_number=1, /// a binary64 number
+    signed_integer,          /// a signed integer that fits in a 64-bit word using two's complement
+    unsigned_integer,        /// a positive integer larger or equal to 1<<63
+    big_integer              /// a big integer that does not fit in a 64-bit word
+};
+
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_BASE_H
+/* end file simdjson/generic/base.h for lasx */
+/* including simdjson/generic/jsoncharutils.h for lasx: #include "simdjson/generic/jsoncharutils.h" */
+/* begin file simdjson/generic/jsoncharutils.h for lasx */
+#ifndef SIMDJSON_GENERIC_JSONCHARUTILS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_JSONCHARUTILS_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/jsoncharutils_tables.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace jsoncharutils {
+
+// return non-zero if not a structural or whitespace char
+// zero otherwise
+simdjson_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace_negated[c];
+}
+
+simdjson_inline uint32_t is_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace[c];
+}
+
+// returns a value with the high 16 bits set if not valid
+// otherwise returns the conversion of the 4 hex digits at src into the bottom
+// 16 bits of the 32-bit return register
+//
+// see
+// https://lemire.me/blog/2019/04/17/parsing-short-hexadecimal-strings-efficiently/
+static inline uint32_t hex_to_u32_nocheck(
+    const uint8_t *src) { // strictly speaking, static inline is a C-ism
+  uint32_t v1 = internal::digit_to_val32[630 + src[0]];
+  uint32_t v2 = internal::digit_to_val32[420 + src[1]];
+  uint32_t v3 = internal::digit_to_val32[210 + src[2]];
+  uint32_t v4 = internal::digit_to_val32[0 + src[3]];
+  return v1 | v2 | v3 | v4;
+}
+
+// given a code point cp, writes to c
+// the utf-8 code, outputting the length in
+// bytes, if the length is zero, the code point
+// is invalid
+//
+// This can possibly be made faster using pdep
+// and clz and table lookups, but JSON documents
+// have few escaped code points, and the following
+// function looks cheap.
+//
+// Note: we assume that surrogates are treated separately
+//
+simdjson_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
+  if (cp <= 0x7F) {
+    c[0] = uint8_t(cp);
+    return 1; // ascii
+  }
+  if (cp <= 0x7FF) {
+    c[0] = uint8_t((cp >> 6) + 192);
+    c[1] = uint8_t((cp & 63) + 128);
+    return 2; // universal plane
+    //  Surrogates are treated elsewhere...
+    //} //else if (0xd800 <= cp && cp <= 0xdfff) {
+    //  return 0; // surrogates // could put assert here
+  } else if (cp <= 0xFFFF) {
+    c[0] = uint8_t((cp >> 12) + 224);
+    c[1] = uint8_t(((cp >> 6) & 63) + 128);
+    c[2] = uint8_t((cp & 63) + 128);
+    return 3;
+  } else if (cp <= 0x10FFFF) { // if you know you have a valid code point, this
+                               // is not needed
+    c[0] = uint8_t((cp >> 18) + 240);
+    c[1] = uint8_t(((cp >> 12) & 63) + 128);
+    c[2] = uint8_t(((cp >> 6) & 63) + 128);
+    c[3] = uint8_t((cp & 63) + 128);
+    return 4;
+  }
+  // will return 0 when the code point was too large.
+  return 0; // bad r
+}
+
+#if SIMDJSON_IS_32BITS // _umul128 for x86, arm
+// this is a slow emulation routine for 32-bit
+//
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+  return x * (uint64_t)y;
+}
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+  uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
+  uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
+  uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
+  uint64_t adbc_carry = !!(adbc < ad);
+  uint64_t lo = bd + (adbc << 32);
+  *hi = __emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) +
+        (adbc_carry << 32) + !!(lo < bd);
+  return lo;
+}
+#endif
+
+} // namespace jsoncharutils
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_JSONCHARUTILS_H
+/* end file simdjson/generic/jsoncharutils.h for lasx */
+/* including simdjson/generic/atomparsing.h for lasx: #include "simdjson/generic/atomparsing.h" */
+/* begin file simdjson/generic/atomparsing.h for lasx */
+#ifndef SIMDJSON_GENERIC_ATOMPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_ATOMPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace lasx {
+namespace {
+/// @private
+namespace atomparsing {
+
+// The string_to_uint32 is exclusively used to map literal strings to 32-bit values.
+// We use memcpy instead of a pointer cast to avoid undefined behaviors since we cannot
+// be certain that the character pointer will be properly aligned.
+// You might think that using memcpy makes this function expensive, but you'd be wrong.
+// All decent optimizing compilers (GCC, clang, Visual Studio) will compile string_to_uint32("false");
+// to the compile-time constant 1936482662.
+simdjson_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
+
+
+// Again in str4ncmp we use a memcpy to avoid undefined behavior. The memcpy may appear expensive.
+// Yet all decent optimizing compilers will compile memcpy to a single instruction, just about.
+simdjson_warn_unused
+simdjson_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
+  uint32_t srcval; // we want to avoid unaligned 32-bit loads (undefined in C/C++)
+  static_assert(sizeof(uint32_t) <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be larger than 4 bytes");
+  std::memcpy(&srcval, src, sizeof(uint32_t));
+  return srcval ^ string_to_uint32(atom);
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src) {
+  return (str4ncmp(src, "true") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_true_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "true"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src) {
+  return (str4ncmp(src+1, "alse") | jsoncharutils::is_not_structural_or_whitespace(src[5])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
+  if (len > 5) { return is_valid_false_atom(src); }
+  else if (len == 5) { return !str4ncmp(src+1, "alse"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src) {
+  return (str4ncmp(src, "null") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_null_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "null"); }
+  else { return false; }
+}
+
+} // namespace atomparsing
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_ATOMPARSING_H
+/* end file simdjson/generic/atomparsing.h for lasx */
+/* including simdjson/generic/dom_parser_implementation.h for lasx: #include "simdjson/generic/dom_parser_implementation.h" */
+/* begin file simdjson/generic/dom_parser_implementation.h for lasx */
+#ifndef SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/dom_parser_implementation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+
+// expectation: sizeof(open_container) = 64/8.
+struct open_container {
+  uint32_t tape_index; // where, on the tape, does the scope ([,{) begins
+  uint32_t count; // how many elements in the scope
+}; // struct open_container
+
+static_assert(sizeof(open_container) == 64/8, "Open container must be 64 bits");
+
+class dom_parser_implementation final : public internal::dom_parser_implementation {
+public:
+  /** Tape location of each open { or [ */
+  std::unique_ptr<open_container[]> open_containers{};
+  /** Whether each open container is a [ or { */
+  std::unique_ptr<bool[]> is_array{};
+  /** Buffer passed to stage 1 */
+  const uint8_t *buf{};
+  /** Length passed to stage 1 */
+  size_t len{0};
+  /** Document passed to stage 2 */
+  dom::document *doc{};
+
+  inline dom_parser_implementation() noexcept;
+  inline dom_parser_implementation(dom_parser_implementation &&other) noexcept;
+  inline dom_parser_implementation &operator=(dom_parser_implementation &&other) noexcept;
+  dom_parser_implementation(const dom_parser_implementation &) = delete;
+  dom_parser_implementation &operator=(const dom_parser_implementation &) = delete;
+
+  simdjson_warn_unused error_code parse(const uint8_t *buf, size_t len, dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage1(const uint8_t *buf, size_t len, stage1_mode partial) noexcept final;
+  simdjson_warn_unused error_code stage2(dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage2_next(dom::document &doc) noexcept final;
+  simdjson_warn_unused uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) const noexcept final;
+  simdjson_warn_unused uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept final;
+  inline simdjson_warn_unused error_code set_capacity(size_t capacity) noexcept final;
+  inline simdjson_warn_unused error_code set_max_depth(size_t max_depth) noexcept final;
+private:
+  simdjson_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
+
+};
+
+} // namespace lasx
+} // namespace simdjson
+
+namespace simdjson {
+namespace lasx {
+
+inline dom_parser_implementation::dom_parser_implementation() noexcept = default;
+inline dom_parser_implementation::dom_parser_implementation(dom_parser_implementation &&other) noexcept = default;
+inline dom_parser_implementation &dom_parser_implementation::operator=(dom_parser_implementation &&other) noexcept = default;
+
+// Leaving these here so they can be inlined if so desired
+inline simdjson_warn_unused error_code dom_parser_implementation::set_capacity(size_t capacity) noexcept {
+  if(capacity > SIMDJSON_MAXSIZE_BYTES) { return CAPACITY; }
+  // Stage 1 index output
+  size_t max_structures = SIMDJSON_ROUNDUP_N(capacity, 64) + 2 + 7;
+  structural_indexes.reset( new (std::nothrow) uint32_t[max_structures] );
+  if (!structural_indexes) { _capacity = 0; return MEMALLOC; }
+  structural_indexes[0] = 0;
+  n_structural_indexes = 0;
+
+  _capacity = capacity;
+  return SUCCESS;
+}
+
+inline simdjson_warn_unused error_code dom_parser_implementation::set_max_depth(size_t max_depth) noexcept {
+  // Stage 2 stacks
+  open_containers.reset(new (std::nothrow) open_container[max_depth]);
+  is_array.reset(new (std::nothrow) bool[max_depth]);
+  if (!is_array || !open_containers) { _max_depth = 0; return MEMALLOC; }
+
+  _max_depth = max_depth;
+  return SUCCESS;
+}
+
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file simdjson/generic/dom_parser_implementation.h for lasx */
+/* including simdjson/generic/implementation_simdjson_result_base.h for lasx: #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base.h for lasx */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+
+// This is a near copy of include/error.h's implementation_simdjson_result_base, except it doesn't use std::pair
+// so we can avoid inlining errors
+// TODO reconcile these!
+/**
+ * The result of a simdjson operation that could fail.
+ *
+ * Gives the option of reading error codes, or throwing an exception by casting to the desired result.
+ *
+ * This is a base class for implementations that want to add functions to the result type for
+ * chaining.
+ *
+ * Override like:
+ *
+ *   struct simdjson_result<T> : public internal::implementation_simdjson_result_base<T> {
+ *     simdjson_result() noexcept : internal::implementation_simdjson_result_base<T>() {}
+ *     simdjson_result(error_code error) noexcept : internal::implementation_simdjson_result_base<T>(error) {}
+ *     simdjson_result(T &&value) noexcept : internal::implementation_simdjson_result_base<T>(std::forward(value)) {}
+ *     simdjson_result(T &&value, error_code error) noexcept : internal::implementation_simdjson_result_base<T>(value, error) {}
+ *     // Your extra methods here
+ *   }
+ *
+ * Then any method returning simdjson_result<T> will be chainable with your methods.
+ */
+template<typename T>
+struct implementation_simdjson_result_base {
+
+  /**
+   * Create a new empty result with error = UNINITIALIZED.
+   */
+  simdjson_inline implementation_simdjson_result_base() noexcept = default;
+
+  /**
+   * Create a new error result.
+   */
+  simdjson_inline implementation_simdjson_result_base(error_code error) noexcept;
+
+  /**
+   * Create a new successful result.
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value) noexcept;
+
+  /**
+   * Create a new result with both things (use if you don't want to branch when creating the result).
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value, error_code error) noexcept;
+
+  /**
+   * Move the value and the error to the provided variables.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   * @param error The variable to assign the error to. Set to SUCCESS if there is no error.
+   */
+  simdjson_inline void tie(T &value, error_code &error) && noexcept;
+
+  /**
+   * Move the value to the provided variable.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   */
+  simdjson_warn_unused simdjson_inline error_code get(T &value) && noexcept;
+
+  /**
+   * The error.
+   */
+  simdjson_warn_unused simdjson_inline error_code error() const noexcept;
+
+  /**
+   * Whether there is a value.
+   */
+  simdjson_warn_unused simdjson_inline bool has_value() const noexcept;
+
+#if SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T& operator*() &  noexcept(false);
+  simdjson_inline T&& operator*() &&  noexcept(false);
+  /**
+   * Arrow operator to access members of the contained value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T* operator->() noexcept(false);
+  simdjson_inline const T* operator->() const noexcept(false);
+
+  simdjson_inline T& value() & noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& value() && noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& take_value() && noexcept(false);
+
+  /**
+   * Cast to the value (will throw on error).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline operator T&&() && noexcept(false);
+
+
+#endif // SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline const T& value_unsafe() const& noexcept;
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T& value_unsafe() & noexcept;
+  /**
+   * Take the result value (move it). This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T&& value_unsafe() && noexcept;
+
+  using value_type = T;
+  using error_type = error_code;
+
+protected:
+  /** users should never directly access first and second. **/
+  T first{}; /** Users should never directly access 'first'. **/
+  error_code second{UNINITIALIZED}; /** Users should never directly access 'second'. **/
+}; // struct implementation_simdjson_result_base
+
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+/* end file simdjson/generic/implementation_simdjson_result_base.h for lasx */
+/* including simdjson/generic/numberparsing.h for lasx: #include "simdjson/generic/numberparsing.h" */
+/* begin file simdjson/generic/numberparsing.h for lasx */
+#ifndef SIMDJSON_GENERIC_NUMBERPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_NUMBERPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <limits>
+#include <ostream>
+#include <cstring>
+
+namespace simdjson {
+namespace lasx {
+namespace numberparsing {
+
+#ifdef JSON_TEST_NUMBERS
+#define INVALID_NUMBER(SRC) (found_invalid_number((SRC)), NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (found_integer((VALUE), (SRC)), (WRITER).append_s64((VALUE)))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (found_unsigned_integer((VALUE), (SRC)), (WRITER).append_u64((VALUE)))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (found_float((VALUE), (SRC)), (WRITER).append_double((VALUE)))
+#define BIGINT_NUMBER(SRC) (found_invalid_number((SRC)), BIGINT_ERROR)
+#else
+#define INVALID_NUMBER(SRC) (NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (WRITER).append_s64((VALUE))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (WRITER).append_u64((VALUE))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (WRITER).append_double((VALUE))
+#define BIGINT_NUMBER(SRC) (BIGINT_ERROR)
+#endif
+
+namespace {
+
+// Convert a mantissa, an exponent and a sign bit into an ieee64 double.
+// The real_exponent needs to be in [0, 2046] (technically real_exponent = 2047 would be acceptable).
+// The mantissa should be in [0,1<<53). The bit at index (1ULL << 52) while be zeroed.
+simdjson_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
+    double d;
+    mantissa &= ~(1ULL << 52);
+    mantissa |= real_exponent << 52;
+    mantissa |= ((static_cast<uint64_t>(negative)) << 63);
+    std::memcpy(&d, &mantissa, sizeof(d));
+    return d;
+}
+
+// Attempts to compute i * 10^(power) exactly; and if "negative" is
+// true, negate the result.
+// This function will only work in some cases, when it does not work, success is
+// set to false. This should work *most of the time* (like 99% of the time).
+// We assume that power is in the [smallest_power,
+// largest_power] interval: the caller is responsible for this check.
+simdjson_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
+  // we start with a fast path
+  // It was described in
+  // Clinger WD. How to read floating point numbers accurately.
+  // ACM SIGPLAN Notices. 1990
+#ifndef FLT_EVAL_METHOD
+#error "FLT_EVAL_METHOD should be defined, please include cfloat."
+#endif
+#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
+  // We cannot be certain that x/y is rounded to nearest.
+  if (0 <= power && power <= 22 && i <= 9007199254740991)
+#else
+  if (-22 <= power && power <= 22 && i <= 9007199254740991)
+#endif
+  {
+    // convert the integer into a double. This is lossless since
+    // 0 <= i <= 2^53 - 1.
+    d = double(i);
+    //
+    // The general idea is as follows.
+    // If 0 <= s < 2^53 and if 10^0 <= p <= 10^22 then
+    // 1) Both s and p can be represented exactly as 64-bit floating-point
+    // values
+    // (binary64).
+    // 2) Because s and p can be represented exactly as floating-point values,
+    // then s * p
+    // and s / p will produce correctly rounded values.
+    //
+    if (power < 0) {
+      d = d / simdjson::internal::power_of_ten[-power];
+    } else {
+      d = d * simdjson::internal::power_of_ten[power];
+    }
+    if (negative) {
+      d = -d;
+    }
+    return true;
+  }
+  // When 22 < power && power <  22 + 16, we could
+  // hope for another, secondary fast path.  It was
+  // described by David M. Gay in  "Correctly rounded
+  // binary-decimal and decimal-binary conversions." (1990)
+  // If you need to compute i * 10^(22 + x) for x < 16,
+  // first compute i * 10^x, if you know that result is exact
+  // (e.g., when i * 10^x < 2^53),
+  // then you can still proceed and do (i * 10^x) * 10^22.
+  // Is this worth your time?
+  // You need  22 < power *and* power <  22 + 16 *and* (i * 10^(x-22) < 2^53)
+  // for this second fast path to work.
+  // If you you have 22 < power *and* power <  22 + 16, and then you
+  // optimistically compute "i * 10^(x-22)", there is still a chance that you
+  // have wasted your time if i * 10^(x-22) >= 2^53. It makes the use cases of
+  // this optimization maybe less common than we would like. Source:
+  // http://www.exploringbinary.com/fast-path-decimal-to-floating-point-conversion/
+  // also used in RapidJSON: https://rapidjson.org/strtod_8h_source.html
+
+  // The fast path has now failed, so we are failing back on the slower path.
+
+  // In the slow path, we need to adjust i so that it is > 1<<63 which is always
+  // possible, except if i == 0, so we handle i == 0 separately.
+  if(i == 0) {
+    d = negative ? -0.0 : 0.0;
+    return true;
+  }
+
+
+  // The exponent is 1024 + 63 + power
+  //     + floor(log(5**power)/log(2)).
+  // The 1024 comes from the ieee64 standard.
+  // The 63 comes from the fact that we use a 64-bit word.
+  //
+  // Computing floor(log(5**power)/log(2)) could be
+  // slow. Instead we use a fast function.
+  //
+  // For power in (-400,350), we have that
+  // (((152170 + 65536) * power ) >> 16);
+  // is equal to
+  //  floor(log(5**power)/log(2)) + power when power >= 0
+  // and it is equal to
+  //  ceil(log(5**-power)/log(2)) + power when power < 0
+  //
+  // The 65536 is (1<<16) and corresponds to
+  // (65536 * power) >> 16 ---> power
+  //
+  // ((152170 * power ) >> 16) is equal to
+  // floor(log(5**power)/log(2))
+  //
+  // Note that this is not magic: 152170/(1<<16) is
+  // approximately equal to log(5)/log(2).
+  // The 1<<16 value is a power of two; we could use a
+  // larger power of 2 if we wanted to.
+  //
+  int64_t exponent = (((152170 + 65536) * power) >> 16) + 1024 + 63;
+
+
+  // We want the most significant bit of i to be 1. Shift if needed.
+  int lz = leading_zeroes(i);
+  i <<= lz;
+
+
+  // We are going to need to do some 64-bit arithmetic to get a precise product.
+  // We use a table lookup approach.
+  // It is safe because
+  // power >= smallest_power
+  // and power <= largest_power
+  // We recover the mantissa of the power, it has a leading 1. It is always
+  // rounded down.
+  //
+  // We want the most significant 64 bits of the product. We know
+  // this will be non-zero because the most significant bit of i is
+  // 1.
+  const uint32_t index = 2 * uint32_t(power - simdjson::internal::smallest_power);
+  // Optimization: It may be that materializing the index as a variable might confuse some compilers and prevent effective complex-addressing loads. (Done for code clarity.)
+  //
+  // The full_multiplication function computes the 128-bit product of two 64-bit words
+  // with a returned value of type value128 with a "low component" corresponding to the
+  // 64-bit least significant bits of the product and with a "high component" corresponding
+  // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index]);
+#else
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index]);
+#endif
+
+  // Both i and power_of_five_128[index] have their most significant bit set to 1 which
+  // implies that the either the most or the second most significant bit of the product
+  // is 1. We pack values in this manner for efficiency reasons: it maximizes the use
+  // we make of the product. It also makes it easy to reason about the product: there
+  // is 0 or 1 leading zero in the product.
+
+  // Unless the least significant 9 bits of the high (64-bit) part of the full
+  // product are all 1s, then we know that the most significant 55 bits are
+  // exact and no further work is needed. Having 55 bits is necessary because
+  // we need 53 bits for the mantissa but we have to have one rounding bit and
+  // we can waste a bit if the most significant bit of the product is zero.
+  if((firstproduct.high & 0x1FF) == 0x1FF) {
+    // We want to compute i * 5^q, but only care about the top 55 bits at most.
+    // Consider the scenario where q>=0. Then 5^q may not fit in 64-bits. Doing
+    // the full computation is wasteful. So we do what is called a "truncated
+    // multiplication".
+    // We take the most significant 64-bits, and we put them in
+    // power_of_five_128[index]. Usually, that's good enough to approximate i * 5^q
+    // to the desired approximation using one multiplication. Sometimes it does not suffice.
+    // Then we store the next most significant 64 bits in power_of_five_128[index + 1], and
+    // then we get a better approximation to i * 5^q.
+    //
+    // That's for when q>=0. The logic for q<0 is somewhat similar but it is somewhat
+    // more complicated.
+    //
+    // There is an extra layer of complexity in that we need more than 55 bits of
+    // accuracy in the round-to-even scenario.
+    //
+    // The full_multiplication function computes the 128-bit product of two 64-bit words
+    // with a returned value of type value128 with a "low component" corresponding to the
+    // 64-bit least significant bits of the product and with a "high component" corresponding
+    // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index + 1]);
+#else
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index + 1]);
+#endif
+    firstproduct.low += secondproduct.high;
+    if(secondproduct.high > firstproduct.low) { firstproduct.high++; }
+    // As it has been proven by Noble Mushtak and Daniel Lemire in "Fast Number Parsing Without
+    // Fallback" (https://arxiv.org/abs/2212.06644), at this point we are sure that the product
+    // is sufficiently accurate, and more computation is not needed.
+  }
+  uint64_t lower = firstproduct.low;
+  uint64_t upper = firstproduct.high;
+  // The final mantissa should be 53 bits with a leading 1.
+  // We shift it so that it occupies 54 bits with a leading 1.
+  ///////
+  uint64_t upperbit = upper >> 63;
+  uint64_t mantissa = upper >> (upperbit + 9);
+  lz += int(1 ^ upperbit);
+
+  // Here we have mantissa < (1<<54).
+  int64_t real_exponent = exponent - lz;
+  if (simdjson_unlikely(real_exponent <= 0)) { // we have a subnormal?
+    // Here have that real_exponent <= 0 so -real_exponent >= 0
+    if(-real_exponent + 1 >= 64) { // if we have more than 64 bits below the minimum exponent, you have a zero for sure.
+      d = negative ? -0.0 : 0.0;
+      return true;
+    }
+    // next line is safe because -real_exponent + 1 < 0
+    mantissa >>= -real_exponent + 1;
+    // Thankfully, we can't have both "round-to-even" and subnormals because
+    // "round-to-even" only occurs for powers close to 0.
+    mantissa += (mantissa & 1); // round up
+    mantissa >>= 1;
+    // There is a weird scenario where we don't have a subnormal but just.
+    // Suppose we start with 2.2250738585072013e-308, we end up
+    // with 0x3fffffffffffff x 2^-1023-53 which is technically subnormal
+    // whereas 0x40000000000000 x 2^-1023-53  is normal. Now, we need to round
+    // up 0x3fffffffffffff x 2^-1023-53  and once we do, we are no longer
+    // subnormal, but we can only know this after rounding.
+    // So we only declare a subnormal if we are smaller than the threshold.
+    real_exponent = (mantissa < (uint64_t(1) << 52)) ? 0 : 1;
+    d = to_double(mantissa, real_exponent, negative);
+    return true;
+  }
+  // We have to round to even. The "to even" part
+  // is only a problem when we are right in between two floats
+  // which we guard against.
+  // If we have lots of trailing zeros, we may fall right between two
+  // floating-point values.
+  //
+  // The round-to-even cases take the form of a number 2m+1 which is in (2^53,2^54]
+  // times a power of two. That is, it is right between a number with binary significand
+  // m and another number with binary significand m+1; and it must be the case
+  // that it cannot be represented by a float itself.
+  //
+  // We must have that w * 10 ^q == (2m+1) * 2^p for some power of two 2^p.
+  // Recall that 10^q = 5^q * 2^q.
+  // When q >= 0, we must have that (2m+1) is divible by 5^q, so 5^q <= 2^54. We have that
+  //  5^23 <=  2^54 and it is the last power of five to qualify, so q <= 23.
+  // When q<0, we have  w  >=  (2m+1) x 5^{-q}.  We must have that w<2^{64} so
+  // (2m+1) x 5^{-q} < 2^{64}. We have that 2m+1>2^{53}. Hence, we must have
+  // 2^{53} x 5^{-q} < 2^{64}.
+  // Hence we have 5^{-q} < 2^{11}$ or q>= -4.
+  //
+  // We require lower <= 1 and not lower == 0 because we could not prove that
+  // that lower == 0 is implied; but we could prove that lower <= 1 is a necessary and sufficient test.
+  if (simdjson_unlikely((lower <= 1) && (power >= -4) && (power <= 23) && ((mantissa & 3) == 1))) {
+    if((mantissa  << (upperbit + 64 - 53 - 2)) ==  upper) {
+      mantissa &= ~1;             // flip it so that we do not round up
+    }
+  }
+
+  mantissa += mantissa & 1;
+  mantissa >>= 1;
+
+  // Here we have mantissa < (1<<53), unless there was an overflow
+  if (mantissa >= (1ULL << 53)) {
+    //////////
+    // This will happen when parsing values such as 7.2057594037927933e+16
+    ////////
+    mantissa = (1ULL << 52);
+    real_exponent++;
+  }
+  mantissa &= ~(1ULL << 52);
+  // we have to check that real_exponent is in range, otherwise we bail out
+  if (simdjson_unlikely(real_exponent > 2046)) {
+    // We have an infinite value!!! We could actually throw an error here if we could.
+    return false;
+  }
+  d = to_double(mantissa, real_exponent, negative);
+  return true;
+}
+
+// We call a fallback floating-point parser that might be slow. Note
+// it will accept JSON numbers, but the JSON spec. is more restrictive so
+// before you call parse_float_fallback, you need to have validated the input
+// string with the JSON grammar.
+// It will return an error (false) if the parsed number is infinite.
+// The string parsing itself always succeeds. We know that there is at least
+// one digit.
+static bool parse_float_fallback(const uint8_t *ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+static bool parse_float_fallback(const uint8_t *ptr, const uint8_t *end_ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr), reinterpret_cast<const char *>(end_ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+// check quickly whether the next 8 chars are made of digits
+// at a glance, it looks better than Mula's
+// http://0x80.pl/articles/swar-digits-validate.html
+simdjson_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
+  uint64_t val;
+  // this can read up to 7 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(7 <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be bigger than 7");
+  std::memcpy(&val, chars, 8);
+  // a branchy method might be faster:
+  // return (( val & 0xF0F0F0F0F0F0F0F0 ) == 0x3030303030303030)
+  //  && (( (val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0 ) ==
+  //  0x3030303030303030);
+  return (((val & 0xF0F0F0F0F0F0F0F0) |
+           (((val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0) >> 4)) ==
+          0x3333333333333333);
+}
+
+template<typename I>
+SIMDJSON_NO_SANITIZE_UNDEFINED // We deliberately allow overflow here and check later
+simdjson_inline bool parse_digit(const uint8_t c, I &i) {
+  const uint8_t digit = static_cast<uint8_t>(c - '0');
+  if (digit > 9) {
+    return false;
+  }
+  // PERF NOTE: multiplication by 10 is cheaper than arbitrary integer multiplication
+  i = 10 * i + digit; // might overflow, we will handle the overflow later
+  return true;
+}
+
+simdjson_inline bool is_digit(const uint8_t c) {
+  return static_cast<uint8_t>(c - '0') <= 9;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_decimal_after_separator(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
+  // we continue with the fiction that we have an integer. If the
+  // floating point number is representable as x * 10^z for some integer
+  // z that fits in 53 bits, then we will be able to convert back the
+  // the integer into a float in a lossless manner.
+  const uint8_t *const first_after_period = p;
+
+#ifdef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_SWAR_NUMBER_PARSING
+  // this helps if we have lots of decimals!
+  // this turns out to be frequent enough.
+  if (is_made_of_eight_digits_fast(p)) {
+    i = i * 100000000 + parse_eight_digits_unrolled(p);
+    p += 8;
+  }
+#endif // SIMDJSON_SWAR_NUMBER_PARSING
+#endif // #ifdef SIMDJSON_SWAR_NUMBER_PARSING
+  // Unrolling the first digit makes a small difference on some implementations (e.g. westmere)
+  if (parse_digit(*p, i)) { ++p; }
+  while (parse_digit(*p, i)) { p++; }
+  exponent = first_after_period - p;
+  // Decimal without digits (123.) is illegal
+  if (exponent == 0) {
+    return INVALID_NUMBER(src);
+  }
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
+  // Exp Sign: -123.456e[-]78
+  bool neg_exp = ('-' == *p);
+  if (neg_exp || '+' == *p) { p++; } // Skip + as well
+
+  // Exponent: -123.456e-[78]
+  auto start_exp = p;
+  int64_t exp_number = 0;
+  while (parse_digit(*p, exp_number)) { ++p; }
+  // It is possible for parse_digit to overflow.
+  // In particular, it could overflow to INT64_MIN, and we cannot do - INT64_MIN.
+  // Thus we *must* check for possible overflow before we negate exp_number.
+
+  // Performance notes: it may seem like combining the two "simdjson_unlikely checks" below into
+  // a single simdjson_unlikely path would be faster. The reasoning is sound, but the compiler may
+  // not oblige and may, in fact, generate two distinct paths in any case. It might be
+  // possible to do uint64_t(p - start_exp - 1) >= 18 but it could end up trading off
+  // instructions for a simdjson_likely branch, an unconclusive gain.
+
+  // If there were no digits, it's an error.
+  if (simdjson_unlikely(p == start_exp)) {
+    return INVALID_NUMBER(src);
+  }
+  // We have a valid positive exponent in exp_number at this point, except that
+  // it may have overflowed.
+
+  // If there were more than 18 digits, we may have overflowed the integer. We have to do
+  // something!!!!
+  if (simdjson_unlikely(p > start_exp+18)) {
+    // Skip leading zeroes: 1e000000000000000000001 is technically valid and does not overflow
+    while (*start_exp == '0') { start_exp++; }
+    // 19 digits could overflow int64_t and is kind of absurd anyway. We don't
+    // support exponents smaller than -999,999,999,999,999,999 and bigger
+    // than 999,999,999,999,999,999.
+    // We can truncate.
+    // Note that 999999999999999999 is assuredly too large. The maximal ieee64 value before
+    // infinity is ~1.8e308. The smallest subnormal is ~5e-324. So, actually, we could
+    // truncate at 324.
+    // Note that there is no reason to fail per se at this point in time.
+    // E.g., 0e999999999999999999999 is a fine number.
+    if (p > start_exp+18) { exp_number = 999999999999999999; }
+  }
+  // At this point, we know that exp_number is a sane, positive, signed integer.
+  // It is <= 999,999,999,999,999,999. As long as 'exponent' is in
+  // [-8223372036854775808, 8223372036854775808], we won't overflow. Because 'exponent'
+  // is bounded in magnitude by the size of the JSON input, we are fine in this universe.
+  // To sum it up: the next line should never overflow.
+  exponent += (neg_exp ? -exp_number : exp_number);
+  return SUCCESS;
+}
+
+simdjson_inline bool check_if_integer(const uint8_t *const src, size_t max_length) {
+  const uint8_t *const srcend = src + max_length;
+  bool negative = (*src == '-'); // we can always read at least one character after the '-'
+  const uint8_t *p = src + uint8_t(negative);
+  if(p == srcend) { return false; }
+  if(*p == '0') {
+    ++p;
+    if(p == srcend) { return true; }
+    if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+    return true;
+  }
+  while(p != srcend && is_digit(*p)) { ++p; }
+  if(p == srcend) { return true; }
+  if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+  return true;
+}
+
+simdjson_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
+  // It is possible that the integer had an overflow.
+  // We have to handle the case where we have 0.0000somenumber.
+  const uint8_t *start = start_digits;
+  while ((*start == '0') || (*start == '.')) { ++start; }
+  // we over-decrement by one when there is a '.'
+  return digit_count - size_t(start - start_digits);
+}
+
+} // unnamed namespace
+
+/** @private */
+static error_code slow_float_parsing(simdjson_unused const uint8_t * src, double* answer) {
+  if (parse_float_fallback(src, answer)) {
+    return SUCCESS;
+  }
+  return INVALID_NUMBER(src);
+}
+
+/** @private */
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
+  // If we frequently had to deal with long strings of digits,
+  // we could extend our code by using a 128-bit integer instead
+  // of a 64-bit integer. However, this is uncommon in practice.
+  //
+  // 9999999999999999999 < 2**64 so we can accommodate 19 digits.
+  // If we have a decimal separator, then digit_count - 1 is the number of digits, but we
+  // may not have a decimal separator!
+  if (simdjson_unlikely(digit_count > 19 && significant_digits(start_digits, digit_count) > 19)) {
+    // Ok, chances are good that we had an overflow!
+    // this is almost never going to get called!!!
+    // we start anew, going slowly!!!
+    // This will happen in the following examples:
+    // 10000000000000000000000000000000000000000000e+308
+    // 3.1415926535897932384626433832795028841971693993751
+    //
+    // NOTE: We do not pass a reference to the to slow_float_parsing. If we passed our writer
+    // reference to it, it would force it to be stored in memory, preventing the compiler from
+    // picking it apart and putting into registers. i.e. if we pass it as reference,
+    // it gets slow.
+    double d;
+    error_code error = slow_float_parsing(src, &d);
+    writer.append_double(d);
+    return error;
+  }
+  // NOTE: it's weird that the simdjson_unlikely() only wraps half the if, but it seems to get slower any other
+  // way we've tried: https://github.com/simdjson/simdjson/pull/990#discussion_r448497331
+  // To future reader: we'd love if someone found a better way, or at least could explain this result!
+  if (simdjson_unlikely(exponent < simdjson::internal::smallest_power) || (exponent > simdjson::internal::largest_power)) {
+    //
+    // Important: smallest_power is such that it leads to a zero value.
+    // Observe that 18446744073709551615e-343 == 0, i.e. (2**64 - 1) e -343 is zero
+    // so something x 10^-343 goes to zero, but not so with  something x 10^-342.
+    static_assert(simdjson::internal::smallest_power <= -342, "smallest_power is not small enough");
+    //
+    if((exponent < simdjson::internal::smallest_power) || (i == 0)) {
+      // E.g. Parse "-0.0e-999" into the same value as "-0.0". See https://en.wikipedia.org/wiki/Signed_zero
+      WRITE_DOUBLE(negative ? -0.0 : 0.0, src, writer);
+      return SUCCESS;
+    } else { // (exponent > largest_power) and (i != 0)
+      // We have, for sure, an infinite value and simdjson refuses to parse infinite values.
+      return INVALID_NUMBER(src);
+    }
+  }
+  double d;
+  if (!compute_float_64(exponent, i, negative, d)) {
+    // we are almost never going to get here.
+    if (!parse_float_fallback(src, &d)) { return INVALID_NUMBER(src); }
+  }
+  WRITE_DOUBLE(d, src, writer);
+  return SUCCESS;
+}
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer);
+
+// for performance analysis, it is sometimes  useful to skip parsing
+#ifdef SIMDJSON_SKIPNUMBERPARSING
+
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const, W &writer) {
+  writer.append_s64(0);        // always write zero
+  return SUCCESS;              // always succeeds
+}
+
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept { return number_type::signed_integer; }
+#else
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer) {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  if (digit_count == 0 || ('0' == *start_digits && digit_count > 1)) { return INVALID_NUMBER(src); }
+
+  //
+  // Handle floats if there is a . or e (or both)
+  //
+  int64_t exponent = 0;
+  bool is_float = false;
+  if ('.' == *p) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_decimal_after_separator(src, p, i, exponent) );
+    digit_count = int(p - start_digits); // used later to guard against overflows
+  }
+  if (('e' == *p) || ('E' == *p)) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_exponent(src, p, exponent) );
+  }
+  if (is_float) {
+    const bool dirty_end = jsoncharutils::is_not_structural_or_whitespace(*p);
+    SIMDJSON_TRY( write_float(src, negative, i, start_digits, digit_count, exponent, writer) );
+    if (dirty_end) { return INVALID_NUMBER(src); }
+    return SUCCESS;
+  }
+
+  // The longest negative 64-bit number is 19 digits.
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  size_t longest_digit_count = negative ? 19 : 20;
+  if (digit_count > longest_digit_count) { return BIGINT_NUMBER(src); }
+  if (digit_count == longest_digit_count) {
+    if (negative) {
+      // Anything negative above INT64_MAX+1 is invalid
+      if (i > uint64_t(INT64_MAX)+1) { return BIGINT_NUMBER(src);  }
+      WRITE_INTEGER(~i+1, src, writer);
+      if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+      return SUCCESS;
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    }  else if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INVALID_NUMBER(src); }
+  }
+
+  // Write unsigned if it does not fit in a signed integer.
+  if (i > uint64_t(INT64_MAX)) {
+    WRITE_UNSIGNED(i, src, writer);
+  } else {
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+    if(i == 0 && negative) {
+      // We have to write -0.0 instead of 0
+      WRITE_DOUBLE(-0.0, src, writer);
+    } else {
+      WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+    }
+#else
+  WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+#endif
+  }
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+  return SUCCESS;
+}
+
+// Inlineable functions
+namespace {
+
+// This table can be used to characterize the final character of an integer
+// string. For JSON structural character and allowable white space characters,
+// we return SUCCESS. For 'e', '.' and 'E', we return INCORRECT_TYPE. Otherwise
+// we return NUMBER_ERROR.
+// Optimization note: we could easily reduce the size of the table by half (to 128)
+// at the cost of an extra branch.
+// Optimization note: we want the values to use at most 8 bits (not, e.g., 32 bits):
+static_assert(error_code(uint8_t(NUMBER_ERROR))== NUMBER_ERROR, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(SUCCESS))== SUCCESS, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(INCORRECT_TYPE))== INCORRECT_TYPE, "bad NUMBER_ERROR cast");
+
+const uint8_t integer_string_finisher[256] = {
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   SUCCESS,      NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, INCORRECT_TYPE,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, SUCCESS,        NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR};
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if ((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
+  const uint8_t *p = src + 1;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    // Note: we use src[1] and not src[0] because src[0] is the quote character in this
+    // instance.
+    if (src[1] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  //
+  // Check for minus sign
+  //
+  if(src == src_end) { return NUMBER_ERROR; }
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = src;
+  uint64_t i = 0;
+  while (parse_digit(*src, i)) { src++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(src - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*src)) {
+  //  return (*src == '.' || *src == 'e' || *src == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(*src != '"') { return NUMBER_ERROR; }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept {
+  return (*src == '-');
+}
+
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) { return true; }
+  return false;
+}
+
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  size_t digit_count = size_t(p - src);
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) {
+    static const uint8_t * smaller_big_integer = reinterpret_cast<const uint8_t *>("9223372036854775808");
+    // We have an integer.
+    if(simdjson_unlikely(digit_count > 20)) {
+      return number_type::big_integer;
+    }
+    // If the number is negative and valid, it must be a signed integer.
+    if(negative) {
+      if (simdjson_unlikely(digit_count > 19)) return number_type::big_integer;
+      if (simdjson_unlikely(digit_count == 19 && memcmp(src, smaller_big_integer, 19) > 0)) {
+        return number_type::big_integer;
+      }
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+      if(digit_count == 1 && src[0] == '0') {
+        // We have to write -0.0 instead of 0
+        return number_type::floating_point_number;
+      }
+#endif
+      return number_type::signed_integer;
+    }
+    // Let us check if we have a big integer (>=2**64).
+    static const uint8_t * two_to_sixtyfour = reinterpret_cast<const uint8_t *>("18446744073709551616");
+    if((digit_count > 20) || (digit_count == 20 && memcmp(src, two_to_sixtyfour, 20) >= 0)) {
+      return number_type::big_integer;
+    }
+    // The number is positive and smaller than 18446744073709551616 (or 2**64).
+    // We want values larger or equal to 9223372036854775808 to be unsigned
+    // integers, and the other values to be signed integers.
+    if((digit_count == 20) || (digit_count >= 19 && memcmp(src, smaller_big_integer, 19) >= 0)) {
+      return number_type::unsigned_integer;
+    }
+    return number_type::signed_integer;
+  }
+  // Hopefully, we have 'e' or 'E' or '.'.
+  return number_type::floating_point_number;
+}
+
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
+  if(src == src_end) { return NUMBER_ERROR; }
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  if(p == src_end) { return NUMBER_ERROR; }
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely((p != src_end) && (*p == '.'))) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if ((p == src_end) || !parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while ((p != src_end) && parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = start_digits-src > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if ((p != src_end) && (*p == 'e' || *p == 'E')) {
+    p++;
+    if(p == src_end) { return NUMBER_ERROR; }
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while ((p != src_end) && parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if ((p != src_end) && jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), src_end, &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+} // unnamed namespace
+#endif // SIMDJSON_SKIPNUMBERPARSING
+
+} // namespace numberparsing
+
+inline std::ostream& operator<<(std::ostream& out, number_type type) noexcept {
+    switch (type) {
+        case number_type::signed_integer: out << "integer in [-9223372036854775808,9223372036854775808)"; break;
+        case number_type::unsigned_integer: out << "unsigned integer in [9223372036854775808,18446744073709551616)"; break;
+        case number_type::floating_point_number: out << "floating-point number (binary64)"; break;
+        case number_type::big_integer: out << "big integer"; break;
+        default: SIMDJSON_UNREACHABLE();
+    }
+    return out;
+}
+
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_NUMBERPARSING_H
+/* end file simdjson/generic/numberparsing.h for lasx */
+
+/* including simdjson/generic/implementation_simdjson_result_base-inl.h for lasx: #include "simdjson/generic/implementation_simdjson_result_base-inl.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base-inl.h for lasx */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+
+//
+// internal::implementation_simdjson_result_base<T> inline implementation
+//
+
+template<typename T>
+simdjson_inline void implementation_simdjson_result_base<T>::tie(T &value, error_code &error) && noexcept {
+  error = this->second;
+  if (!error) {
+    value = std::forward<implementation_simdjson_result_base<T>>(*this).first;
+  }
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::get(T &value) && noexcept {
+  error_code error;
+  std::forward<implementation_simdjson_result_base<T>>(*this).tie(value, error);
+  return error;
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::error() const noexcept {
+  return this->second;
+}
+
+
+template<typename T>
+simdjson_warn_unused simdjson_inline bool implementation_simdjson_result_base<T>::has_value() const noexcept {
+  return this->error() == SUCCESS;
+}
+
+#if SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::operator*() &  noexcept(false) {
+  return this->value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::operator*() &&  noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).value();
+}
+
+template<typename T>
+simdjson_inline T* implementation_simdjson_result_base<T>::operator->() noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+
+template<typename T>
+simdjson_inline const T* implementation_simdjson_result_base<T>::operator->() const noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value() & noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::take_value() && noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::operator T&&() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+#endif // SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline const T& implementation_simdjson_result_base<T>::value_unsafe() const& noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value_unsafe() & noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value_unsafe() && noexcept {
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value, error_code error) noexcept
+    : first{std::forward<T>(value)}, second{error} {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(error_code error) noexcept
+    : implementation_simdjson_result_base(T{}, error) {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value) noexcept
+    : implementation_simdjson_result_base(std::forward<T>(value), SUCCESS) {}
+
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+/* end file simdjson/generic/implementation_simdjson_result_base-inl.h for lasx */
+/* end file simdjson/generic/amalgamated.h for lasx */
+/* including simdjson/lasx/end.h: #include "simdjson/lasx/end.h" */
+/* begin file simdjson/lasx/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#undef SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+/* undefining SIMDJSON_IMPLEMENTATION from "lasx" */
+#undef SIMDJSON_IMPLEMENTATION
+
+
+#if SIMDJSON_CAN_ALWAYS_RUN_LASX
+// nothing needed.
+#else
+SIMDJSON_UNTARGET_REGION
+#endif
+/* end file simdjson/lasx/end.h */
+
+#endif // SIMDJSON_LASX_H
+/* end file simdjson/lasx.h */
+/* including simdjson/lasx/implementation.h: #include <simdjson/lasx/implementation.h> */
+/* begin file simdjson/lasx/implementation.h */
+#ifndef SIMDJSON_LASX_IMPLEMENTATION_H
+#define SIMDJSON_LASX_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/instruction_set.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation("lasx", "LoongArch ASX", internal::instruction_set::LASX) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_LASX_IMPLEMENTATION_H
+/* end file simdjson/lasx/implementation.h */
+
+/* including simdjson/lasx/begin.h: #include <simdjson/lasx/begin.h> */
+/* begin file simdjson/lasx/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "lasx" */
+#define SIMDJSON_IMPLEMENTATION lasx
+#include <lsxintrin.h> // This is a hack. We should not need to put this include here.
+#if SIMDJSON_CAN_ALWAYS_RUN_LASX
+// nothing needed.
+#else
+SIMDJSON_TARGET_REGION("lasx,lsx")
+#endif
+
+/* including simdjson/lasx/base.h: #include "simdjson/lasx/base.h" */
+/* begin file simdjson/lasx/base.h */
+#ifndef SIMDJSON_LASX_BASE_H
+#define SIMDJSON_LASX_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+/**
+ * Implementation for LASX.
+ */
+namespace lasx {
+
+class implementation;
+
+namespace {
+namespace simd {
+template <typename T> struct simd8;
+template <typename T> struct simd8x64;
+} // namespace simd
+} // unnamed namespace
+
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_LASX_BASE_H
+/* end file simdjson/lasx/base.h */
+/* including simdjson/lasx/intrinsics.h: #include "simdjson/lasx/intrinsics.h" */
+/* begin file simdjson/lasx/intrinsics.h */
+#ifndef SIMDJSON_LASX_INTRINSICS_H
+#define SIMDJSON_LASX_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <lsxintrin.h>
+#include <lasxintrin.h>
+
+static_assert(sizeof(__m256i) <= simdjson::SIMDJSON_PADDING, "insufficient padding for LoongArch ASX");
+
+#endif //  SIMDJSON_LASX_INTRINSICS_H
+/* end file simdjson/lasx/intrinsics.h */
+/* including simdjson/lasx/bitmanipulation.h: #include "simdjson/lasx/bitmanipulation.h" */
+/* begin file simdjson/lasx/bitmanipulation.h */
+#ifndef SIMDJSON_LASX_BITMANIPULATION_H
+#define SIMDJSON_LASX_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/bitmask.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+  return __builtin_ctzll(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return input_num & (input_num-1);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+  return __builtin_clzll(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int count_ones(uint64_t input_num) {
+  return __lasx_xvpickve2gr_w(__lasx_xvpcnt_d(__m256i(v4u64{input_num, 0, 0, 0})), 0);
+}
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2, uint64_t *result) {
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+}
+
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_LASX_BITMANIPULATION_H
+/* end file simdjson/lasx/bitmanipulation.h */
+/* including simdjson/lasx/bitmask.h: #include "simdjson/lasx/bitmask.h" */
+/* begin file simdjson/lasx/bitmask.h */
+#ifndef SIMDJSON_LASX_BITMASK_H
+#define SIMDJSON_LASX_BITMASK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(uint64_t bitmask) {
+  bitmask ^= bitmask << 1;
+  bitmask ^= bitmask << 2;
+  bitmask ^= bitmask << 4;
+  bitmask ^= bitmask << 8;
+  bitmask ^= bitmask << 16;
+  bitmask ^= bitmask << 32;
+  return bitmask;
+}
+
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif
+/* end file simdjson/lasx/bitmask.h */
+/* including simdjson/lasx/numberparsing_defs.h: #include "simdjson/lasx/numberparsing_defs.h" */
+/* begin file simdjson/lasx/numberparsing_defs.h */
+#ifndef SIMDJSON_LASX_NUMBERPARSING_DEFS_H
+#define SIMDJSON_LASX_NUMBERPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace lasx {
+namespace numberparsing {
+
+// we don't have appropriate instructions, so let us use a scalar function
+// credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  uint64_t val;
+  std::memcpy(&val, chars, sizeof(uint64_t));
+  val = (val & 0x0F0F0F0F0F0F0F0F) * 2561 >> 8;
+  val = (val & 0x00FF00FF00FF00FF) * 6553601 >> 16;
+  return uint32_t((val & 0x0000FFFF0000FFFF) * 42949672960001 >> 32);
+}
+
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace lasx
+} // namespace simdjson
+
+#ifndef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_IS_BIG_ENDIAN
+#define SIMDJSON_SWAR_NUMBER_PARSING 0
+#else
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+#endif
+#endif
+
+#endif // SIMDJSON_LASX_NUMBERPARSING_DEFS_H
+/* end file simdjson/lasx/numberparsing_defs.h */
+/* including simdjson/lasx/simd.h: #include "simdjson/lasx/simd.h" */
+/* begin file simdjson/lasx/simd.h */
+#ifndef SIMDJSON_LASX_SIMD_H
+#define SIMDJSON_LASX_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace simd {
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename Child>
+  struct base {
+    __m256i value;
+
+    // Zero constructor
+    simdjson_inline base() : value{__m256i()} {}
+
+    // Conversion from SIMD register
+    simdjson_inline base(const __m256i _value) : value(_value) {}
+
+    // Conversion to SIMD register
+    simdjson_inline operator const __m256i&() const { return this->value; }
+    simdjson_inline operator __m256i&() { return this->value; }
+    simdjson_inline operator const v32i8&() const { return (v32i8&)this->value; }
+    simdjson_inline operator v32i8&() { return (v32i8&)this->value; }
+
+    // Bit operations
+    simdjson_inline Child operator|(const Child other) const { return __lasx_xvor_v(*this, other); }
+    simdjson_inline Child operator&(const Child other) const { return __lasx_xvand_v(*this, other); }
+    simdjson_inline Child operator^(const Child other) const { return __lasx_xvxor_v(*this, other); }
+    simdjson_inline Child bit_andnot(const Child other) const { return __lasx_xvandn_v(other, *this); }
+    simdjson_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+  };
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename T>
+  struct simd8;
+
+  template<typename T, typename Mask=simd8<bool>>
+  struct base8: base<simd8<T>> {
+    simdjson_inline base8() : base<simd8<T>>() {}
+    simdjson_inline base8(const __m256i _value) : base<simd8<T>>(_value) {}
+
+    friend simdjson_really_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return __lasx_xvseq_b(lhs, rhs); }
+
+    static const int SIZE = sizeof(base<simd8<T>>::value);
+
+    template<int N=1>
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+        __m256i hi = __lasx_xvbsll_v(*this, N);
+        __m256i lo = __lasx_xvbsrl_v(*this, 16 - N);
+        __m256i tmp = __lasx_xvbsrl_v(prev_chunk, 16 - N);
+        lo = __lasx_xvpermi_q(lo, tmp, 0x21);
+        return __lasx_xvor_v(hi, lo);
+    }
+  };
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool>: base8<bool> {
+    static simdjson_inline simd8<bool> splat(bool _value) { return __lasx_xvreplgr2vr_b(uint8_t(-(!!_value))); }
+
+    simdjson_inline simd8() : base8() {}
+    simdjson_inline simd8(const __m256i _value) : base8<bool>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+
+    simdjson_inline int to_bitmask() const {
+      __m256i mask = __lasx_xvmskltz_b(*this);
+      return (__lasx_xvpickve2gr_w(mask, 4) << 16) | (__lasx_xvpickve2gr_w(mask, 0));
+    }
+   simdjson_inline bool any() const {
+      __m256i v = __lasx_xvmsknz_b(*this);
+      return (0 == __lasx_xvpickve2gr_w(v, 0)) && (0 == __lasx_xvpickve2gr_w(v, 4));
+    }
+    simdjson_inline simd8<bool> operator~() const { return *this ^ true; }
+  };
+
+  template<typename T>
+  struct base8_numeric: base8<T> {
+    static simdjson_inline simd8<T> splat(T _value) {
+      return __lasx_xvreplgr2vr_b(_value);
+    }
+    static simdjson_inline simd8<T> zero() { return __lasx_xvldi(0); }
+    static simdjson_inline simd8<T> load(const T values[32]) {
+      return __lasx_xvld(reinterpret_cast<const __m256i *>(values), 0);
+    }
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    static simdjson_inline simd8<T> repeat_16(
+      T v0,  T v1,  T v2,  T v3,  T v4,  T v5,  T v6,  T v7,
+      T v8,  T v9,  T v10, T v11, T v12, T v13, T v14, T v15
+    ) {
+      return simd8<T>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    simdjson_inline base8_numeric() : base8<T>() {}
+    simdjson_inline base8_numeric(const __m256i _value) : base8<T>(_value) {}
+
+    // Store to array
+    simdjson_inline void store(T dst[32]) const {
+      return __lasx_xvst(*this, reinterpret_cast<__m256i *>(dst), 0);
+    }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<T> operator+(const simd8<T> other) const { return __lasx_xvadd_b(*this, other); }
+    simdjson_inline simd8<T> operator-(const simd8<T> other) const { return __lasx_xvsub_b(*this, other); }
+    simdjson_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
+
+    // Override to distinguish from bool version
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return __lasx_xvshuf_b(lookup_table, lookup_table, *this);
+    }
+
+    // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted as a bitset).
+    // Passing a 0 value for mask would be equivalent to writing out every byte to output.
+    // Only the first 16 - count_ones(mask) bytes of the result are significant but 16 bytes
+    // get written.
+    template<typename L>
+    simdjson_inline void compress(uint32_t mask, L * output) const {
+      using internal::thintable_epi8;
+      using internal::BitsSetTable256mul2;
+      using internal::pshufb_combine_table;
+      // this particular implementation was inspired by haswell
+      // lasx do it in 4 steps, first 8 bytes and then second 8 bytes...
+      uint8_t mask1 = uint8_t(mask); // least significant 8 bits
+      uint8_t mask2 = uint8_t(mask >> 8); // second significant 8 bits
+      uint8_t mask3 = uint8_t(mask >> 16); // ...
+      uint8_t mask4 = uint8_t(mask >> 24); // ...
+      // next line just loads the 64-bit values thintable_epi8[mask{1,2,3,4}]
+      // into a 256-bit register.
+      __m256i shufmask = {int64_t(thintable_epi8[mask1]), int64_t(thintable_epi8[mask2]) + 0x0808080808080808, int64_t(thintable_epi8[mask3]), int64_t(thintable_epi8[mask4]) + 0x0808080808080808};
+      // this is the version "nearly pruned"
+      __m256i pruned = __lasx_xvshuf_b(*this, *this, shufmask);
+      // we still need to put the  pieces back together.
+      // we compute the popcount of the first words:
+      int pop1 = BitsSetTable256mul2[mask1];
+      int pop2 = BitsSetTable256mul2[mask2];
+      int pop3 = BitsSetTable256mul2[mask3];
+
+      // then load the corresponding mask
+      __m256i masklo = __lasx_xvldx(reinterpret_cast<void*>(reinterpret_cast<unsigned long>(pshufb_combine_table)), pop1 * 8);
+      __m256i maskhi = __lasx_xvldx(reinterpret_cast<void*>(reinterpret_cast<unsigned long>(pshufb_combine_table)), pop3 * 8);
+      __m256i compactmask = __lasx_xvpermi_q(maskhi, masklo, 0x20);
+      __m256i answer = __lasx_xvshuf_b(pruned, pruned, compactmask);
+      __lasx_xvst(answer, reinterpret_cast<uint8_t*>(output), 0);
+      uint64_t value3 = __lasx_xvpickve2gr_du(answer, 2);
+      uint64_t value4 = __lasx_xvpickve2gr_du(answer, 3);
+      uint64_t *pos = reinterpret_cast<uint64_t*>(reinterpret_cast<uint8_t*>(output) + 16 - (pop1 + pop2) / 2);
+      pos[0] = value3;
+      pos[1] = value4;
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> : base8_numeric<int8_t> {
+    simdjson_inline simd8() : base8_numeric<int8_t>() {}
+    simdjson_inline simd8(const __m256i _value) : base8_numeric<int8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const int8_t values[32]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15,
+      int8_t v16, int8_t v17, int8_t v18, int8_t v19, int8_t v20, int8_t v21, int8_t v22, int8_t v23,
+      int8_t v24, int8_t v25, int8_t v26, int8_t v27, int8_t v28, int8_t v29, int8_t v30, int8_t v31
+    ) : simd8({
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15,
+      v16,v17,v18,v19,v20,v21,v22,v23,
+      v24,v25,v26,v27,v28,v29,v30,v31
+      }) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return __lasx_xvmax_b(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return __lasx_xvmin_b(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return __lasx_xvslt_b(other, *this); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return __lasx_xvslt_b(*this, other); }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t>: base8_numeric<uint8_t> {
+    simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+    simdjson_inline simd8(const __m256i _value) : base8_numeric<uint8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const uint8_t values[32]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15,
+      uint8_t v16, uint8_t v17, uint8_t v18, uint8_t v19, uint8_t v20, uint8_t v21, uint8_t v22, uint8_t v23,
+      uint8_t v24, uint8_t v25, uint8_t v26, uint8_t v27, uint8_t v28, uint8_t v29, uint8_t v30, uint8_t v31
+    ) : simd8(__m256i(v32u8{
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15,
+      v16,v17,v18,v19,v20,v21,v22,v23,
+      v24,v25,v26,v27,v28,v29,v30,v31
+    })) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15,
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return __lasx_xvsadd_bu(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return __lasx_xvssub_bu(*this, other); }
+
+    // Order-specific operations
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return __lasx_xvmax_bu(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return __lasx_xvmin_bu(other, *this); }
+    // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
+    // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->lt_bits(other).any_bits_set(); }
+
+    // Bit-specific operations
+    simdjson_inline simd8<bool> bits_not_set() const { return *this == uint8_t(0); }
+    simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
+    simdjson_inline bool is_ascii() const {
+      __m256i mask = __lasx_xvmskltz_b(*this);
+      return (0 == __lasx_xvpickve2gr_w(mask, 0)) && (0 == __lasx_xvpickve2gr_w(mask, 4));
+    }
+    simdjson_inline bool bits_not_set_anywhere() const {
+      __m256i v = __lasx_xvmsknz_b(*this);
+      return (0 == __lasx_xvpickve2gr_w(v, 0)) && (0 == __lasx_xvpickve2gr_w(v, 4));
+    }
+    simdjson_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
+    simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const {
+      __m256i v = __lasx_xvmsknz_b(__lasx_xvand_v(*this, bits));
+      return (0 == __lasx_xvpickve2gr_w(v, 0)) && (0 == __lasx_xvpickve2gr_w(v, 4));
+    }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(__lasx_xvsrli_b(*this, N)); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(__lasx_xvslli_b(*this, N)); }
+  };
+
+  template<typename T>
+  struct simd8x64 {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+    static_assert(NUM_CHUNKS == 2, "LASX kernel should use two registers per 64-byte block.");
+    const simd8<T> chunks[NUM_CHUNKS];
+    template<int idx> simd8<uint8_t> get() const { return idx < NUM_CHUNKS ? chunks[idx] : simd8<T>(); }
+
+    simd8x64(const simd8x64<T>& o) = delete; // no copy allowed
+    simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1) : chunks{chunk0, chunk1} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+32)} {}
+
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
+      uint32_t mask1 = uint32_t(mask);
+      uint32_t mask2 = uint32_t(mask >> 32);
+      __m256i zcnt = __lasx_xvpcnt_w(__m256i(v4u64{~mask, 0, 0, 0}));
+      uint64_t zcnt1 = __lasx_xvpickve2gr_wu(zcnt, 0);
+      uint64_t zcnt2 = __lasx_xvpickve2gr_wu(zcnt, 1);
+      // There should be a critical value which processes in scaler is faster.
+      if (zcnt1)
+        this->chunks[0].compress(mask1, output);
+      if (zcnt2)
+        this->chunks[1].compress(mask2, output + zcnt1);
+      return zcnt1 + zcnt2;
+    }
+
+    simdjson_inline void store(T ptr[64]) const {
+      this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
+      this->chunks[1].store(ptr+sizeof(simd8<T>)*1);
+    }
+
+    simdjson_inline uint64_t to_bitmask() const {
+      __m256i mask0 = __lasx_xvmskltz_b(this->chunks[0]);
+      __m256i mask1 = __lasx_xvmskltz_b(this->chunks[1]);
+      __m256i mask_tmp = __lasx_xvpickve_w(mask0, 4);
+      __m256i tmp = __lasx_xvpickve_w(mask1, 4);
+      mask0 = __lasx_xvinsve0_w(mask0, mask1, 1);
+      mask_tmp = __lasx_xvinsve0_w(mask_tmp, tmp, 1);
+      return __lasx_xvpickve2gr_du(__lasx_xvpackev_h(mask_tmp, mask0), 0);
+    }
+
+    simdjson_inline simd8<T> reduce_or() const {
+      return this->chunks[0] | this->chunks[1];
+    }
+
+    simdjson_inline uint64_t eq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] == mask,
+        this->chunks[1] == mask
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+      return  simd8x64<bool>(
+        this->chunks[0] == other.chunks[0],
+        this->chunks[1] == other.chunks[1]
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t lteq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] <= mask,
+        this->chunks[1] <= mask
+      ).to_bitmask();
+    }
+  }; // struct simd8x64<T>
+
+} // namespace simd
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_LASX_SIMD_H
+/* end file simdjson/lasx/simd.h */
+/* including simdjson/lasx/stringparsing_defs.h: #include "simdjson/lasx/stringparsing_defs.h" */
+/* begin file simdjson/lasx/stringparsing_defs.h */
+#ifndef SIMDJSON_LASX_STRINGPARSING_DEFS_H
+#define SIMDJSON_LASX_STRINGPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/simd.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 32;
+  simdjson_inline backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return bs_bits != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+
+  uint32_t bs_bits;
+  uint32_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // this can read up to 31 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  return {
+      static_cast<uint32_t>((v == '\\').to_bitmask()),     // bs_bits
+      static_cast<uint32_t>((v == '"').to_bitmask()), // quote_bits
+  };
+}
+
+
+struct escaping {
+  static constexpr uint32_t BYTES_PROCESSED = 16;
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits != 0; }
+  simdjson_inline int escape_index() { return trailing_zeroes(escape_bits); }
+
+  uint64_t escape_bits;
+}; // struct escaping
+
+
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "escaping finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  simd8<bool> is_quote = (v == '"');
+  simd8<bool> is_backslash = (v == '\\');
+  simd8<bool> is_control = (v < 32);
+  return {
+    static_cast<uint64_t>((is_backslash | is_quote | is_control).to_bitmask())
+  };
+}
+
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_LASX_STRINGPARSING_DEFS_H
+/* end file simdjson/lasx/stringparsing_defs.h */
+
+#define SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT 1
+
+
+/* end file simdjson/lasx/begin.h */
+/* including generic/amalgamated.h for lasx: #include <generic/amalgamated.h> */
+/* begin file generic/amalgamated.h for lasx */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_SRC_GENERIC_DEPENDENCIES_H)
+#error generic/dependencies.h must be included before generic/amalgamated.h!
+#endif
+
+/* including generic/base.h for lasx: #include <generic/base.h> */
+/* begin file generic/base.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+
+struct json_character_block;
+
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_BASE_H
+/* end file generic/base.h for lasx */
+/* including generic/dom_parser_implementation.h for lasx: #include <generic/dom_parser_implementation.h> */
+/* begin file generic/dom_parser_implementation.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// Interface a dom parser implementation must fulfill
+namespace simdjson {
+namespace lasx {
+namespace {
+
+simdjson_inline simd8<uint8_t> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3);
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input);
+
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file generic/dom_parser_implementation.h for lasx */
+/* including generic/json_character_block.h for lasx: #include <generic/json_character_block.h> */
+/* begin file generic/json_character_block.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+
+struct json_character_block {
+  static simdjson_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
+
+  simdjson_inline uint64_t whitespace() const noexcept { return _whitespace; }
+  simdjson_inline uint64_t op() const noexcept { return _op; }
+  simdjson_inline uint64_t scalar() const noexcept { return ~(op() | whitespace()); }
+
+  uint64_t _whitespace;
+  uint64_t _op;
+};
+
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H
+/* end file generic/json_character_block.h for lasx */
+/* end file generic/amalgamated.h for lasx */
+/* including generic/stage1/amalgamated.h for lasx: #include <generic/stage1/amalgamated.h> */
+/* begin file generic/stage1/amalgamated.h for lasx */
+// Stuff other things depend on
+/* including generic/stage1/base.h for lasx: #include <generic/stage1/base.h> */
+/* begin file generic/stage1/base.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_BASE_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace stage1 {
+
+class bit_indexer;
+template<size_t STEP_SIZE>
+struct buf_block_reader;
+struct json_block;
+class json_minifier;
+class json_scanner;
+struct json_string_block;
+class json_string_scanner;
+class json_structural_indexer;
+
+} // namespace stage1
+
+namespace utf8_validation {
+struct utf8_checker;
+} // namespace utf8_validation
+
+using utf8_validation::utf8_checker;
+
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_BASE_H
+/* end file generic/stage1/base.h for lasx */
+/* including generic/stage1/buf_block_reader.h for lasx: #include <generic/stage1/buf_block_reader.h> */
+/* begin file generic/stage1/buf_block_reader.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace stage1 {
+
+// Walks through a buffer in block-sized increments, loading the last part with spaces
+template<size_t STEP_SIZE>
+struct buf_block_reader {
+public:
+  simdjson_inline buf_block_reader(const uint8_t *_buf, size_t _len);
+  simdjson_inline size_t block_index();
+  simdjson_inline bool has_full_block() const;
+  simdjson_inline const uint8_t *full_block() const;
+  /**
+   * Get the last block, padded with spaces.
+   *
+   * There will always be a last block, with at least 1 byte, unless len == 0 (in which case this
+   * function fills the buffer with spaces and returns 0. In particular, if len == STEP_SIZE there
+   * will be 0 full_blocks and 1 remainder block with STEP_SIZE bytes and no spaces for padding.
+   *
+   * @return the number of effective characters in the last block.
+   */
+  simdjson_inline size_t get_remainder(uint8_t *dst) const;
+  simdjson_inline void advance();
+private:
+  const uint8_t *buf;
+  const size_t len;
+  const size_t lenminusstep;
+  size_t idx;
+};
+
+// Routines to print masks and text for debugging bitmask operations
+simdjson_unused static char * format_input_text_64(const uint8_t *text) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    buf[i] = int8_t(text[i]) < ' ' ? '_' : int8_t(text[i]);
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+// Routines to print masks and text for debugging bitmask operations
+simdjson_unused static char * format_input_text(const simd8x64<uint8_t>& in) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  in.store(reinterpret_cast<uint8_t*>(buf));
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] < ' ') { buf[i] = '_'; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+simdjson_unused static char * format_input_text(const simd8x64<uint8_t>& in, uint64_t mask) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  in.store(reinterpret_cast<uint8_t*>(buf));
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] <= ' ') { buf[i] = '_'; }
+    if (!(mask & (size_t(1) << i))) { buf[i] = ' '; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+simdjson_unused static char * format_mask(uint64_t mask) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  for (size_t i=0; i<64; i++) {
+    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
+  }
+  buf[64] = '\0';
+  return buf;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
+
+template<size_t STEP_SIZE>
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
+
+template<size_t STEP_SIZE>
+simdjson_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
+  return idx < lenminusstep;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
+  return &buf[idx];
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
+  if(len == idx) { return 0; } // memcpy(dst, null, 0) will trigger an error with some sanitizers
+  std::memset(dst, 0x20, STEP_SIZE); // std::memset STEP_SIZE because it's more efficient to write out 8 or 16 bytes at once.
+  std::memcpy(dst, buf + idx, len - idx);
+  return len - idx;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline void buf_block_reader<STEP_SIZE>::advance() {
+  idx += STEP_SIZE;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H
+/* end file generic/stage1/buf_block_reader.h for lasx */
+/* including generic/stage1/json_escape_scanner.h for lasx: #include <generic/stage1/json_escape_scanner.h> */
+/* begin file generic/stage1/json_escape_scanner.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_ESCAPE_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_ESCAPE_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace stage1 {
+
+/**
+ * Scans for escape characters in JSON, taking care with multiple backslashes (\\n vs. \n).
+ */
+struct json_escape_scanner {
+  /** The actual escape characters (the backslashes themselves). */
+  uint64_t next_is_escaped = 0ULL;
+
+  struct escaped_and_escape {
+    /**
+     * Mask of escaped characters.
+     *
+     * ```
+     * \n \\n \\\n \\\\n \
+     * 0100100010100101000
+     *  n  \   \ n  \ \
+     * ```
+     */
+    uint64_t escaped;
+    /**
+     * Mask of escape characters.
+     *
+     * ```
+     * \n \\n \\\n \\\\n \
+     * 1001000101001010001
+     * \  \   \ \  \ \   \
+     * ```
+     */
+    uint64_t escape;
+  };
+
+  /**
+   * Get a mask of both escape and escaped characters (the characters following a backslash).
+   *
+   * @param potential_escape A mask of the character that can escape others (but could be
+   *        escaped itself). e.g. block.eq('\\')
+   */
+  simdjson_really_inline escaped_and_escape next(uint64_t backslash) noexcept {
+
+#if !SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+    if (!backslash) { return {next_escaped_without_backslashes(), 0}; }
+#endif
+
+    // |                                | Mask (shows characters instead of 1's) | Depth | Instructions        |
+    // |--------------------------------|----------------------------------------|-------|---------------------|
+    // | string                         | `\\n_\\\n___\\\n___\\\\___\\\\__\\\`   |       |                     |
+    // |                                | `    even   odd    even   odd   odd`   |       |                     |
+    // | potential_escape               | ` \  \\\    \\\    \\\\   \\\\  \\\`   | 1     | 1 (backslash & ~first_is_escaped)
+    // | escape_and_terminal_code       | ` \n \ \n   \ \n   \ \    \ \   \ \`   | 5     | 5 (next_escape_and_terminal_code())
+    // | escaped                        | `\    \ n    \ n    \ \    \ \   \ ` X | 6     | 7 (escape_and_terminal_code ^ (potential_escape | first_is_escaped))
+    // | escape                         | `    \ \    \ \    \ \    \ \   \ \`   | 6     | 8 (escape_and_terminal_code & backslash)
+    // | first_is_escaped               | `\                                 `   | 7 (*) | 9 (escape >> 63) ()
+    //                                                                               (*) this is not needed until the next iteration
+    uint64_t escape_and_terminal_code = next_escape_and_terminal_code(backslash & ~this->next_is_escaped);
+    uint64_t escaped = escape_and_terminal_code ^ (backslash | this->next_is_escaped);
+    uint64_t escape = escape_and_terminal_code & backslash;
+    this->next_is_escaped = escape >> 63;
+    return {escaped, escape};
+  }
+
+private:
+  static constexpr const uint64_t ODD_BITS = 0xAAAAAAAAAAAAAAAAULL;
+
+  simdjson_really_inline uint64_t next_escaped_without_backslashes() noexcept {
+    uint64_t escaped = this->next_is_escaped;
+    this->next_is_escaped = 0;
+    return escaped;
+  }
+
+  /**
+   * Returns a mask of the next escape characters (masking out escaped backslashes), along with
+   * any non-backslash escape codes.
+   *
+   * \n \\n \\\n \\\\n returns:
+   * \n \   \ \n \ \
+   * 11 100 1011 10100
+   *
+   * You are expected to mask out the first bit yourself if the previous block had a trailing
+   * escape.
+   *
+   * & the result with potential_escape to get just the escape characters.
+   * ^ the result with (potential_escape | first_is_escaped) to get escaped characters.
+   */
+  static simdjson_really_inline uint64_t next_escape_and_terminal_code(uint64_t potential_escape) noexcept {
+    // If we were to just shift and mask out any odd bits, we'd actually get a *half* right answer:
+    // any even-aligned backslash runs would be correct! Odd-aligned backslash runs would be
+    // inverted (\\\ would be 010 instead of 101).
+    //
+    // ```
+    // string:              | ____\\\\_\\\\_____ |
+    // maybe_escaped | ODD  |     \ \   \ \      |
+    //               even-aligned ^^^  ^^^^ odd-aligned
+    // ```
+    //
+    // Taking that into account, our basic strategy is:
+    //
+    // 1. Use subtraction to produce a mask with 1's for even-aligned runs and 0's for
+    //    odd-aligned runs.
+    // 2. XOR all odd bits, which masks out the odd bits in even-aligned runs, and brings IN the
+    //    odd bits in odd-aligned runs.
+    // 3. & with backslash to clean up any stray bits.
+    // runs are set to 0, and then XORing with "odd":
+    //
+    // |                                | Mask (shows characters instead of 1's) | Instructions        |
+    // |--------------------------------|----------------------------------------|---------------------|
+    // | string                         | `\\n_\\\n___\\\n___\\\\___\\\\__\\\`   |
+    // |                                | `    even   odd    even   odd   odd`   |
+    // | maybe_escaped                  | `  n  \\n    \\n    \\\_   \\\_  \\` X | 1 (potential_escape << 1)
+    // | maybe_escaped_and_odd          | ` \n_ \\n _ \\\n_ _ \\\__ _\\\_ \\\`   | 1 (maybe_escaped | odd)
+    // | even_series_codes_and_odd      | `  n_\\\  _    n_ _\\\\ _     _    `   | 1 (maybe_escaped_and_odd - potential_escape)
+    // | escape_and_terminal_code       | ` \n \ \n   \ \n   \ \    \ \   \ \`   | 1 (^ odd)
+    //
+
+    // Escaped characters are characters following an escape.
+    uint64_t maybe_escaped = potential_escape << 1;
+
+    // To distinguish odd from even escape sequences, therefore, we turn on any *starting*
+    // escapes that are on an odd byte. (We actually bring in all odd bits, for speed.)
+    // - Odd runs of backslashes are 0000, and the code at the end ("n" in \n or \\n) is 1.
+    // - Odd runs of backslashes are 1111, and the code at the end ("n" in \n or \\n) is 0.
+    // - All other odd bytes are 1, and even bytes are 0.
+    uint64_t maybe_escaped_and_odd_bits     = maybe_escaped | ODD_BITS;
+    uint64_t even_series_codes_and_odd_bits = maybe_escaped_and_odd_bits - potential_escape;
+
+    // Now we flip all odd bytes back with xor. This:
+    // - Makes odd runs of backslashes go from 0000 to 1010
+    // - Makes even runs of backslashes go from 1111 to 1010
+    // - Sets actually-escaped codes to 1 (the n in \n and \\n: \n = 11, \\n = 100)
+    // - Resets all other bytes to 0
+    return even_series_codes_and_odd_bits ^ ODD_BITS;
+  }
+};
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+/* end file generic/stage1/json_escape_scanner.h for lasx */
+/* including generic/stage1/json_string_scanner.h for lasx: #include <generic/stage1/json_string_scanner.h> */
+/* begin file generic/stage1/json_string_scanner.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_escape_scanner.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace stage1 {
+
+struct json_string_block {
+  // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
+  simdjson_really_inline json_string_block(uint64_t escaped, uint64_t quote, uint64_t in_string) :
+  _escaped(escaped), _quote(quote), _in_string(in_string) {}
+
+  // Escaped characters (characters following an escape() character)
+  simdjson_really_inline uint64_t escaped() const { return _escaped; }
+  // Real (non-backslashed) quotes
+  simdjson_really_inline uint64_t quote() const { return _quote; }
+  // Only characters inside the string (not including the quotes)
+  simdjson_really_inline uint64_t string_content() const { return _in_string & ~_quote; }
+  // Return a mask of whether the given characters are inside a string (only works on non-quotes)
+  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
+  // Return a mask of whether the given characters are inside a string (only works on non-quotes)
+  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
+  // Tail of string (everything except the start quote)
+  simdjson_really_inline uint64_t string_tail() const { return _in_string ^ _quote; }
+
+  // escaped characters (backslashed--does not include the hex characters after \u)
+  uint64_t _escaped;
+  // real quotes (non-escaped ones)
+  uint64_t _quote;
+  // string characters (includes start quote but not end quote)
+  uint64_t _in_string;
+};
+
+// Scans blocks for string characters, storing the state necessary to do so
+class json_string_scanner {
+public:
+  simdjson_really_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
+  // Returns either UNCLOSED_STRING or SUCCESS
+  simdjson_really_inline error_code finish();
+
+private:
+  // Scans for escape characters
+  json_escape_scanner escape_scanner{};
+  // Whether the last iteration was still inside a string (all 1's = true, all 0's = false).
+  uint64_t prev_in_string = 0ULL;
+};
+
+//
+// Return a mask of all string characters plus end quotes.
+//
+// prev_escaped is overflow saying whether the next character is escaped.
+// prev_in_string is overflow saying whether we're still in a string.
+//
+// Backslash sequences outside of quotes will be detected in stage 2.
+//
+simdjson_really_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
+  const uint64_t backslash = in.eq('\\');
+  const uint64_t escaped = escape_scanner.next(backslash).escaped;
+  const uint64_t quote = in.eq('"') & ~escaped;
+
+  //
+  // prefix_xor flips on bits inside the string (and flips off the end quote).
+  //
+  // Then we xor with prev_in_string: if we were in a string already, its effect is flipped
+  // (characters inside strings are outside, and characters outside strings are inside).
+  //
+  const uint64_t in_string = prefix_xor(quote) ^ prev_in_string;
+
+  //
+  // Check if we're still in a string at the end of the box so the next block will know
+  //
+  prev_in_string = uint64_t(static_cast<int64_t>(in_string) >> 63);
+
+  // Use ^ to turn the beginning quote off, and the end quote on.
+
+  // We are returning a function-local object so either we get a move constructor
+  // or we get copy elision.
+  return json_string_block(escaped, quote, in_string);
+}
+
+simdjson_really_inline error_code json_string_scanner::finish() {
+  if (prev_in_string) {
+    return UNCLOSED_STRING;
+  }
+  return SUCCESS;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+/* end file generic/stage1/json_string_scanner.h for lasx */
+/* including generic/stage1/utf8_lookup4_algorithm.h for lasx: #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* begin file generic/stage1/utf8_lookup4_algorithm.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace utf8_validation {
+
+using namespace simd;
+
+  simdjson_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+// Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+// Bit 1 = Too Long (ASCII followed by continuation)
+// Bit 2 = Overlong 3-byte
+// Bit 4 = Surrogate
+// Bit 5 = Overlong 2-byte
+// Bit 7 = Two Continuations
+    constexpr const uint8_t TOO_SHORT   = 1<<0; // 11______ 0_______
+                                                // 11______ 11______
+    constexpr const uint8_t TOO_LONG    = 1<<1; // 0_______ 10______
+    constexpr const uint8_t OVERLONG_3  = 1<<2; // 11100000 100_____
+    constexpr const uint8_t SURROGATE   = 1<<4; // 11101101 101_____
+    constexpr const uint8_t OVERLONG_2  = 1<<5; // 1100000_ 10______
+    constexpr const uint8_t TWO_CONTS   = 1<<7; // 10______ 10______
+    constexpr const uint8_t TOO_LARGE   = 1<<3; // 11110100 1001____
+                                                // 11110100 101_____
+                                                // 11110101 1001____
+                                                // 11110101 101_____
+                                                // 1111011_ 1001____
+                                                // 1111011_ 101_____
+                                                // 11111___ 1001____
+                                                // 11111___ 101_____
+    constexpr const uint8_t TOO_LARGE_1000 = 1<<6;
+                                                // 11110101 1000____
+                                                // 1111011_ 1000____
+                                                // 11111___ 1000____
+    constexpr const uint8_t OVERLONG_4  = 1<<6; // 11110000 1000____
+
+    const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      TOO_SHORT,
+      // 1110____ ________ <three byte lead in byte 1>
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4
+    );
+    constexpr const uint8_t CARRY = TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+    const simd8<uint8_t> byte_1_low = (prev1 & 0x0F).lookup_16<uint8_t>(
+      // ____0000 ________
+      CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+      // ____0001 ________
+      CARRY | OVERLONG_2,
+      // ____001_ ________
+      CARRY,
+      CARRY,
+
+      // ____0100 ________
+      CARRY | TOO_LARGE,
+      // ____0101 ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      // ____011_ ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+
+      // ____1___ ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      // ____1101 ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000
+    );
+    const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE  | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE  | TOO_LARGE,
+
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT
+    );
+    return (byte_1_high & byte_1_low & byte_2_high);
+  }
+  simdjson_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
+      const simd8<uint8_t> prev_input, const simd8<uint8_t> sc) {
+    simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+    simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+    simd8<uint8_t> must23 = must_be_2_3_continuation(prev2, prev3);
+    simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+    return must23_80 ^ sc;
+  }
+
+  //
+  // Return nonzero if there are incomplete multibyte characters at the end of the block:
+  // e.g. if there is a 4-byte character, but it's 3 bytes from the end.
+  //
+  simdjson_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+    // If the previous input's last 3 bytes match this, they're too short (they ended at EOF):
+    // ... 1111____ 111_____ 11______
+#if SIMDJSON_IMPLEMENTATION_ICELAKE || (SIMDJSON_IMPLEMENTATION_RVV_VLS && __riscv_v_fixed_vlen >= 512)
+    static const uint8_t max_array[64] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
+    };
+#else
+    static const uint8_t max_array[32] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
+    };
+#endif
+    const simd8<uint8_t> max_value(&max_array[sizeof(max_array)-sizeof(simd8<uint8_t>)]);
+    return input.gt_bits(max_value);
+  }
+
+  struct utf8_checker {
+    // If this is nonzero, there has been a UTF-8 error.
+    simd8<uint8_t> error;
+    // The last input we received
+    simd8<uint8_t> prev_input_block;
+    // Whether the last input we received was incomplete (used for ASCII fast path)
+    simd8<uint8_t> prev_incomplete;
+
+    //
+    // Check whether the current bytes are valid UTF-8.
+    //
+    simdjson_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
+      // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ lead bytes
+      // (2, 3, 4-byte leads become large positive numbers instead of small negative numbers)
+      simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+      simd8<uint8_t> sc = check_special_cases(input, prev1);
+      this->error |= check_multibyte_lengths(input, prev_input, sc);
+    }
+
+    // The only problem that can happen at EOF is that a multibyte character is too short
+    // or a byte value too large in the last bytes: check_special_cases only checks for bytes
+    // too large in the first of two bytes.
+    simdjson_inline void check_eof() {
+      // If the previous block had incomplete UTF-8 characters at the end, an ASCII block can't
+      // possibly finish them.
+      this->error |= this->prev_incomplete;
+    }
+
+    simdjson_inline void check_next_input(const simd8x64<uint8_t>& input) {
+      if(simdjson_likely(is_ascii(input))) {
+        this->error |= this->prev_incomplete;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio, it is not good enough.
+        static_assert((simd8x64<uint8_t>::NUM_CHUNKS == 1)
+                ||(simd8x64<uint8_t>::NUM_CHUNKS == 2)
+                || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+                "We support one, two or four chunks per 64-byte block.");
+        SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 1) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+          this->check_utf8_bytes(input.get<1>(), input.get<0>());
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+          this->check_utf8_bytes(input.get<1>(), input.get<0>());
+          this->check_utf8_bytes(input.get<2>(), input.get<1>());
+          this->check_utf8_bytes(input.get<3>(), input.get<2>());
+        }
+        this->prev_incomplete = is_incomplete(input.get<simd8x64<uint8_t>::NUM_CHUNKS-1>());
+        this->prev_input_block = input.get<simd8x64<uint8_t>::NUM_CHUNKS-1>();
+      }
+    }
+    // do not forget to call check_eof!
+    simdjson_warn_unused simdjson_inline error_code errors() {
+      return this->error.any_bits_set_anywhere() ? error_code::UTF8_ERROR : error_code::SUCCESS;
+    }
+
+  }; // struct utf8_checker
+} // namespace utf8_validation
+
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H
+/* end file generic/stage1/utf8_lookup4_algorithm.h for lasx */
+/* including generic/stage1/json_scanner.h for lasx: #include <generic/stage1/json_scanner.h> */
+/* begin file generic/stage1/json_scanner.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/json_character_block.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_string_scanner.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace stage1 {
+
+/**
+ * A block of scanned json, with information on operators and scalars.
+ *
+ * We seek to identify pseudo-structural characters. Anything that is inside
+ * a string must be omitted (hence  & ~_string.string_tail()).
+ * Otherwise, pseudo-structural characters come in two forms.
+ * 1. We have the structural characters ([,],{,},:, comma). The
+ *    term 'structural character' is from the JSON RFC.
+ * 2. We have the 'scalar pseudo-structural characters'.
+ *    Scalars are quotes, and any character except structural characters and white space.
+ *
+ * To identify the scalar pseudo-structural characters, we must look at what comes
+ * before them: it must be a space, a quote or a structural characters.
+ * Starting with simdjson v0.3, we identify them by
+ * negation: we identify everything that is followed by a non-quote scalar,
+ * and we negate that. Whatever remains must be a 'scalar pseudo-structural character'.
+ */
+struct json_block {
+public:
+  // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
+  simdjson_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  _string(std::move(string)), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
+  simdjson_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  _string(string), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
+
+  /**
+   * The start of structurals.
+   * In simdjson prior to v0.3, these were called the pseudo-structural characters.
+   **/
+  simdjson_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
+  /** All JSON whitespace (i.e. not in a string) */
+  simdjson_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
+
+  // Helpers
+
+  /** Whether the given characters are inside a string (only works on non-quotes) */
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
+  /** Whether the given characters are outside a string (only works on non-quotes) */
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
+
+  // string and escape characters
+  json_string_block _string;
+  // whitespace, structural characters ('operators'), scalars
+  json_character_block _characters;
+  // whether the previous character was a scalar
+  uint64_t _follows_potential_nonquote_scalar;
+private:
+  // Potential structurals (i.e. disregarding strings)
+
+  /**
+   * structural elements ([,],{,},:, comma) plus scalar starts like 123, true and "abc".
+   * They may reside inside a string.
+   **/
+  simdjson_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
+  /**
+   * The start of non-operator runs, like 123, true and "abc".
+   * It main reside inside a string.
+   **/
+  simdjson_inline uint64_t potential_scalar_start() const noexcept {
+    // The term "scalar" refers to anything except structural characters and white space
+    // (so letters, numbers, quotes).
+    // Whenever it is preceded by something that is not a structural element ({,},[,],:, ") nor a white-space
+    // then we know that it is irrelevant structurally.
+    return _characters.scalar() & ~follows_potential_scalar();
+  }
+  /**
+   * Whether the given character is immediately after a non-operator like 123, true.
+   * The characters following a quote are not included.
+   */
+  simdjson_inline uint64_t follows_potential_scalar() const noexcept {
+    // _follows_potential_nonquote_scalar: is defined as marking any character that follows a character
+    // that is not a structural element ({,},[,],:, comma) nor a quote (") and that is not a
+    // white space.
+    // It is understood that within quoted region, anything at all could be marked (irrelevant).
+    return _follows_potential_nonquote_scalar;
+  }
+};
+
+/**
+ * Scans JSON for important bits: structural characters or 'operators', strings, and scalars.
+ *
+ * The scanner starts by calculating two distinct things:
+ * - string characters (taking \" into account)
+ * - structural characters or 'operators' ([]{},:, comma)
+ *   and scalars (runs of non-operators like 123, true and "abc")
+ *
+ * To minimize data dependency (a key component of the scanner's speed), it finds these in parallel:
+ * in particular, the operator/scalar bit will find plenty of things that are actually part of
+ * strings. When we're done, json_block will fuse the two together by masking out tokens that are
+ * part of a string.
+ */
+class json_scanner {
+public:
+  json_scanner() = default;
+  simdjson_inline json_block next(const simd::simd8x64<uint8_t>& in);
+  // Returns either UNCLOSED_STRING or SUCCESS
+  simdjson_warn_unused simdjson_inline error_code finish();
+
+private:
+  // Whether the last character of the previous iteration is part of a scalar token
+  // (anything except whitespace or a structural character/'operator').
+  uint64_t prev_scalar = 0ULL;
+  json_string_scanner string_scanner{};
+};
+
+
+//
+// Check if the current character immediately follows a matching character.
+//
+// For example, this checks for quotes with backslashes in front of them:
+//
+//     const uint64_t backslashed_quote = in.eq('"') & immediately_follows(in.eq('\'), prev_backslash);
+//
+simdjson_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
+  const uint64_t result = match << 1 | overflow;
+  overflow = match >> 63;
+  return result;
+}
+
+simdjson_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
+  json_string_block strings = string_scanner.next(in);
+  // identifies the white-space and the structural characters
+  json_character_block characters = json_character_block::classify(in);
+  // The term "scalar" refers to anything except structural characters and white space
+  // (so letters, numbers, quotes).
+  // We want follows_scalar to mark anything that follows a non-quote scalar (so letters and numbers).
+  //
+  // A terminal quote should either be followed by a structural character (comma, brace, bracket, colon)
+  // or nothing. However, we still want ' "a string"true ' to mark the 't' of 'true' as a potential
+  // pseudo-structural character just like we would if we had  ' "a string" true '; otherwise we
+  // may need to add an extra check when parsing strings.
+  //
+  // Performance: there are many ways to skin this cat.
+  const uint64_t nonquote_scalar = characters.scalar() & ~strings.quote();
+  uint64_t follows_nonquote_scalar = follows(nonquote_scalar, prev_scalar);
+  // We are returning a function-local object so either we get a move constructor
+  // or we get copy elision.
+  return json_block(
+    strings,// strings is a function-local object so either it moves or the copy is elided.
+    characters,
+    follows_nonquote_scalar
+  );
+}
+
+simdjson_warn_unused simdjson_inline error_code json_scanner::finish() {
+  return string_scanner.finish();
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H
+/* end file generic/stage1/json_scanner.h for lasx */
+
+// All other declarations
+/* including generic/stage1/find_next_document_index.h for lasx: #include <generic/stage1/find_next_document_index.h> */
+/* begin file generic/stage1/find_next_document_index.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace stage1 {
+
+/**
+  * This algorithm is used to quickly identify the last structural position that
+  * makes up a complete document.
+  *
+  * It does this by going backwards and finding the last *document boundary* (a
+  * place where one value follows another without a comma between them). If the
+  * last document (the characters after the boundary) has an equal number of
+  * start and end brackets, it is considered complete.
+  *
+  * Simply put, we iterate over the structural characters, starting from
+  * the end. We consider that we found the end of a JSON document when the
+  * first element of the pair is NOT one of these characters: '{' '[' ':' ','
+  * and when the second element is NOT one of these characters: '}' ']' ':' ','.
+  *
+  * This simple comparison works most of the time, but it does not cover cases
+  * where the batch's structural indexes contain a perfect amount of documents.
+  * In such a case, we do not have access to the structural index which follows
+  * the last document, therefore, we do not have access to the second element in
+  * the pair, and that means we cannot identify the last document. To fix this
+  * issue, we keep a count of the open and closed curly/square braces we found
+  * while searching for the pair. When we find a pair AND the count of open and
+  * closed curly/square braces is the same, we know that we just passed a
+  * complete document, therefore the last json buffer location is the end of the
+  * batch.
+  */
+simdjson_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
+  // Variant: do not count separately, just figure out depth
+  if(parser.n_structural_indexes == 0) { return 0; }
+  auto arr_cnt = 0;
+  auto obj_cnt = 0;
+  for (auto i = parser.n_structural_indexes - 1; i > 0; i--) {
+    auto idxb = parser.structural_indexes[i];
+    switch (parser.buf[idxb]) {
+    case ':':
+    case ',':
+      continue;
+    case '}':
+      obj_cnt--;
+      continue;
+    case ']':
+      arr_cnt--;
+      continue;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+    }
+    auto idxa = parser.structural_indexes[i - 1];
+    switch (parser.buf[idxa]) {
+    case '{':
+    case '[':
+    case ':':
+    case ',':
+      continue;
+    }
+    // Last document is complete, so the next document will appear after!
+    if (!arr_cnt && !obj_cnt) {
+      return parser.n_structural_indexes;
+    }
+    // Last document is incomplete; mark the document at i + 1 as the next one
+    return i;
+  }
+  // If we made it to the end, we want to finish counting to see if we have a full document.
+  switch (parser.buf[parser.structural_indexes[0]]) {
+    case '}':
+      obj_cnt--;
+      break;
+    case ']':
+      arr_cnt--;
+      break;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+  }
+  if (!arr_cnt && !obj_cnt) {
+    // We have a complete document.
+    return parser.n_structural_indexes;
+  }
+  return 0;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+/* end file generic/stage1/find_next_document_index.h for lasx */
+/* including generic/stage1/json_minifier.h for lasx: #include <generic/stage1/json_minifier.h> */
+/* begin file generic/stage1/json_minifier.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses in stage1
+// It is intended to be included multiple times and compiled multiple times
+// We assume the file in which it is included already includes
+// "simdjson/stage1.h" (this simplifies amalgation)
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace stage1 {
+
+class json_minifier {
+public:
+  template<size_t STEP_SIZE>
+  static error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept;
+
+private:
+  simdjson_inline json_minifier(uint8_t *_dst)
+  : dst{_dst}
+  {}
+  template<size_t STEP_SIZE>
+  simdjson_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
+  simdjson_warn_unused simdjson_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
+  json_scanner scanner{};
+  uint8_t *dst;
+};
+
+simdjson_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
+  uint64_t mask = block.whitespace();
+  dst += in.compress(mask, dst);
+}
+
+simdjson_warn_unused simdjson_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
+  error_code error = scanner.finish();
+  if (error) { dst_len = 0; return error; }
+  dst_len = dst - dst_start;
+  return SUCCESS;
+}
+
+template<>
+simdjson_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block_buf);
+  simd::simd8x64<uint8_t> in_2(block_buf+64);
+  json_block block_1 = scanner.next(in_1);
+  json_block block_2 = scanner.next(in_2);
+  this->next(in_1, block_1);
+  this->next(in_2, block_2);
+  reader.advance();
+}
+
+template<>
+simdjson_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block_buf);
+  json_block block_1 = scanner.next(in_1);
+  this->next(block_buf, block_1);
+  reader.advance();
+}
+
+template<size_t STEP_SIZE>
+error_code json_minifier::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept {
+  buf_block_reader<STEP_SIZE> reader(buf, len);
+  json_minifier minifier(dst);
+
+  // Index the first n-1 blocks
+  while (reader.has_full_block()) {
+    minifier.step<STEP_SIZE>(reader.full_block(), reader);
+  }
+
+  // Index the last (remainder) block, padded with spaces
+  uint8_t block[STEP_SIZE];
+  size_t remaining_bytes = reader.get_remainder(block);
+  if (remaining_bytes > 0) {
+    // We do not want to write directly to the output stream. Rather, we write
+    // to a local buffer (for safety).
+    uint8_t out_block[STEP_SIZE];
+    uint8_t * const guarded_dst{minifier.dst};
+    minifier.dst = out_block;
+    minifier.step<STEP_SIZE>(block, reader);
+    size_t to_write = minifier.dst - out_block;
+    // In some cases, we could be enticed to consider the padded spaces
+    // as part of the string. This is fine as long as we do not write more
+    // than we consumed.
+    if(to_write > remaining_bytes) { to_write = remaining_bytes; }
+    memcpy(guarded_dst, out_block, to_write);
+    minifier.dst = guarded_dst + to_write;
+  }
+  return minifier.finish(dst, dst_len);
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H
+/* end file generic/stage1/json_minifier.h for lasx */
+/* including generic/stage1/json_structural_indexer.h for lasx: #include <generic/stage1/json_structural_indexer.h> */
+/* begin file generic/stage1/json_structural_indexer.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_string_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_minifier.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/find_next_document_index.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses in stage1
+// It is intended to be included multiple times and compiled multiple times
+// We assume the file in which it is included already includes
+// "simdjson/stage1.h" (this simplifies amalgation)
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace stage1 {
+
+class bit_indexer {
+public:
+  uint32_t *tail;
+
+  simdjson_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
+
+#if SIMDJSON_PREFER_REVERSE_BITS
+  /**
+    * ARM lacks a fast trailing zero instruction, but it has a fast
+    * bit reversal instruction and a fast leading zero instruction.
+    * Thus it may be profitable to reverse the bits (once) and then
+    * to rely on a sequence of instructions that call the leading
+    * zero instruction.
+    *
+    * Performance notes:
+    * The chosen routine is not optimal in terms of data dependency
+    * since zero_leading_bit might require two instructions. However,
+    * it tends to minimize the total number of instructions which is
+    * beneficial.
+    */
+  simdjson_inline void write_index(uint32_t idx, uint64_t& rev_bits, int i) {
+    int lz = leading_zeroes(rev_bits);
+    this->tail[i] = static_cast<uint32_t>(idx) + lz;
+    rev_bits = zero_leading_bit(rev_bits, lz);
+  }
+#else
+  /**
+    * Under recent x64 systems, we often have both a fast trailing zero
+    * instruction and a fast 'clear-lower-bit' instruction so the following
+    * algorithm can be competitive.
+    */
+
+  simdjson_inline void write_index(uint32_t idx, uint64_t& bits, int i) {
+    this->tail[i] = idx + trailing_zeroes(bits);
+    bits = clear_lowest_bit(bits);
+  }
+#endif // SIMDJSON_PREFER_REVERSE_BITS
+
+  template <int START, int N>
+  simdjson_inline int write_indexes(uint32_t idx, uint64_t& bits) {
+    write_index(idx, bits, START);
+    SIMDJSON_IF_CONSTEXPR (N > 1) {
+      write_indexes<(N-1>0?START+1:START), (N-1>=0?N-1:1)>(idx, bits);
+    }
+    return START+N;
+  }
+
+  template <int START, int END, int STEP>
+  simdjson_inline int write_indexes_stepped(uint32_t idx, uint64_t& bits, int cnt) {
+    write_indexes<START, STEP>(idx, bits);
+    SIMDJSON_IF_CONSTEXPR ((START+STEP)  < END) {
+      if (simdjson_unlikely((START+STEP) < cnt)) {
+        write_indexes_stepped<(START+STEP<END?START+STEP:END), END, STEP>(idx, bits, cnt);
+      }
+    }
+    return ((END-START) % STEP) == 0 ? END : (END-START) - ((END-START) % STEP) + STEP;
+  }
+
+  // flatten out values in 'bits' assuming that they are are to have values of idx
+  // plus their position in the bitvector, and store these indexes at
+  // base_ptr[base] incrementing base as we go
+  // will potentially store extra values beyond end of valid bits, so base_ptr
+  // needs to be large enough to handle this
+  //
+  // If the kernel sets SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER, then it
+  // will provide its own version of the code.
+#ifdef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+  simdjson_inline void write(uint32_t idx, uint64_t bits);
+#else
+  simdjson_inline void write(uint32_t idx, uint64_t bits) {
+    // In some instances, the next branch is expensive because it is mispredicted.
+    // Unfortunately, in other cases,
+    // it helps tremendously.
+    if (bits == 0)
+        return;
+
+    int cnt = static_cast<int>(count_ones(bits));
+
+#if SIMDJSON_PREFER_REVERSE_BITS
+    bits = reverse_bits(bits);
+#endif
+#ifdef SIMDJSON_STRUCTURAL_INDEXER_STEP
+    static constexpr const int STEP = SIMDJSON_STRUCTURAL_INDEXER_STEP;
+#else
+    static constexpr const int STEP = 4;
+#endif
+    static constexpr const int STEP_UNTIL = 24;
+
+    write_indexes_stepped<0, STEP_UNTIL, STEP>(idx, bits, cnt);
+    SIMDJSON_IF_CONSTEXPR (STEP_UNTIL < 64) {
+      if (simdjson_unlikely(STEP_UNTIL < cnt)) {
+        for (int i=STEP_UNTIL; i<cnt; i++) {
+          write_index(idx, bits, i);
+        }
+      }
+    }
+
+    this->tail += cnt;
+  }
+#endif // SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+};
+
+class json_structural_indexer {
+public:
+  /**
+   * Find the important bits of JSON in a 128-byte chunk, and add them to structural_indexes.
+   *
+   * @param partial Setting the partial parameter to true allows the find_structural_bits to
+   *   tolerate unclosed strings. The caller should still ensure that the input is valid UTF-8. If
+   *   you are processing substrings, you may want to call on a function like trimmed_length_safe_utf8.
+   */
+  template<size_t STEP_SIZE>
+  static error_code index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept;
+
+private:
+  simdjson_inline json_structural_indexer(uint32_t *structural_indexes);
+  template<size_t STEP_SIZE>
+  simdjson_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
+  simdjson_warn_unused simdjson_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
+
+  json_scanner scanner{};
+  utf8_checker checker{};
+  bit_indexer indexer;
+  uint64_t prev_structurals = 0;
+  uint64_t unescaped_chars_error = 0;
+};
+
+simdjson_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
+
+// Skip the last character if it is partial
+simdjson_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
+  if (simdjson_unlikely(len < 3)) {
+    switch (len) {
+      case 2:
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
+        return len;
+      case 1:
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        return len;
+      case 0:
+        return len;
+    }
+  }
+  if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+  if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 1 byte left
+  if (buf[len-3] >= 0xf0) { return len-3; } // 4-byte characters with only 3 bytes left
+  return len;
+}
+
+//
+// PERF NOTES:
+// We pipe 2 inputs through these stages:
+// 1. Load JSON into registers. This takes a long time and is highly parallelizable, so we load
+//    2 inputs' worth at once so that by the time step 2 is looking for them input, it's available.
+// 2. Scan the JSON for critical data: strings, scalars and operators. This is the critical path.
+//    The output of step 1 depends entirely on this information. These functions don't quite use
+//    up enough CPU: the second half of the functions is highly serial, only using 1 execution core
+//    at a time. The second input's scans has some dependency on the first ones finishing it, but
+//    they can make a lot of progress before they need that information.
+// 3. Step 1 does not use enough capacity, so we run some extra stuff while we're waiting for that
+//    to finish: utf-8 checks and generating the output from the last iteration.
+//
+// The reason we run 2 inputs at a time, is steps 2 and 3 are *still* not enough to soak up all
+// available capacity with just one input. Running 2 at a time seems to give the CPU a good enough
+// workout.
+//
+template<size_t STEP_SIZE>
+error_code json_structural_indexer::index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept {
+  if (simdjson_unlikely(len > parser.capacity())) { return CAPACITY; }
+  // We guard the rest of the code so that we can assume that len > 0 throughout.
+  if (len == 0) { return EMPTY; }
+  if (is_streaming(partial)) {
+    len = trim_partial_utf8(buf, len);
+    // If you end up with an empty window after trimming
+    // the partial UTF-8 bytes, then chances are good that you
+    // have an UTF-8 formatting error.
+    if(len == 0) { return UTF8_ERROR; }
+  }
+  buf_block_reader<STEP_SIZE> reader(buf, len);
+  json_structural_indexer indexer(parser.structural_indexes.get());
+
+  // Read all but the last block
+  while (reader.has_full_block()) {
+    indexer.step<STEP_SIZE>(reader.full_block(), reader);
+  }
+  // Take care of the last block (will always be there unless file is empty which is
+  // not supposed to happen.)
+  uint8_t block[STEP_SIZE];
+  if (simdjson_unlikely(reader.get_remainder(block) == 0)) { return UNEXPECTED_ERROR; }
+  indexer.step<STEP_SIZE>(block, reader);
+  return indexer.finish(parser, reader.block_index(), len, partial);
+}
+
+template<>
+simdjson_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block);
+  simd::simd8x64<uint8_t> in_2(block+64);
+  json_block block_1 = scanner.next(in_1);
+  json_block block_2 = scanner.next(in_2);
+  this->next(in_1, block_1, reader.block_index());
+  this->next(in_2, block_2, reader.block_index()+64);
+  reader.advance();
+}
+
+template<>
+simdjson_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block);
+  json_block block_1 = scanner.next(in_1);
+  this->next(in_1, block_1, reader.block_index());
+  reader.advance();
+}
+
+simdjson_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
+  uint64_t unescaped = in.lteq(0x1F);
+#if SIMDJSON_UTF8VALIDATION
+  checker.check_next_input(in);
+#endif
+  indexer.write(uint32_t(idx-64), prev_structurals); // Output *last* iteration's structurals to the parser
+  prev_structurals = block.structural_start();
+  unescaped_chars_error |= block.non_quote_inside_string(unescaped);
+}
+
+simdjson_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
+  // Write out the final iteration's structurals
+  indexer.write(uint32_t(idx-64), prev_structurals);
+  error_code error = scanner.finish();
+  // We deliberately break down the next expression so that it is
+  // human readable.
+  const bool should_we_exit = is_streaming(partial) ?
+    ((error != SUCCESS) && (error != UNCLOSED_STRING)) // when partial we tolerate UNCLOSED_STRING
+    : (error != SUCCESS); // if partial is false, we must have SUCCESS
+  const bool have_unclosed_string = (error == UNCLOSED_STRING);
+  if (simdjson_unlikely(should_we_exit)) { return error; }
+
+  if (unescaped_chars_error) {
+    return UNESCAPED_CHARS;
+  }
+  parser.n_structural_indexes = uint32_t(indexer.tail - parser.structural_indexes.get());
+  /***
+   * The On-Demand API requires special padding.
+   *
+   * This is related to https://github.com/simdjson/simdjson/issues/906
+   * Basically, we want to make sure that if the parsing continues beyond the last (valid)
+   * structural character, it quickly stops.
+   * Only three structural characters can be repeated without triggering an error in JSON:  [,] and }.
+   * We repeat the padding character (at 'len'). We don't know what it is, but if the parsing
+   * continues, then it must be [,] or }.
+   * Suppose it is ] or }. We backtrack to the first character, what could it be that would
+   * not trigger an error? It could be ] or } but no, because you can't start a document that way.
+   * It can't be a comma, a colon or any simple value. So the only way we could continue is
+   * if the repeated character is [. But if so, the document must start with [. But if the document
+   * starts with [, it should end with ]. If we enforce that rule, then we would get
+   * ][[ which is invalid.
+   *
+   * This is illustrated with the test array_iterate_unclosed_error() on the following input:
+   * R"({ "a": [,,)"
+   **/
+  parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len); // used later in partial == stage1_mode::streaming_final
+  parser.structural_indexes[parser.n_structural_indexes + 1] = uint32_t(len);
+  parser.structural_indexes[parser.n_structural_indexes + 2] = 0;
+  parser.next_structural_index = 0;
+  // a valid JSON file cannot have zero structural indexes - we should have found something
+  if (simdjson_unlikely(parser.n_structural_indexes == 0u)) {
+    return EMPTY;
+  }
+  if (simdjson_unlikely(parser.structural_indexes[parser.n_structural_indexes - 1] > len)) {
+    return UNEXPECTED_ERROR;
+  }
+  if (partial == stage1_mode::streaming_partial) {
+    // If we have an unclosed string, then the last structural
+    // will be the quote and we want to make sure to omit it.
+    if(have_unclosed_string) {
+      parser.n_structural_indexes--;
+      // a valid JSON file cannot have zero structural indexes - we should have found something
+      if (simdjson_unlikely(parser.n_structural_indexes == 0u)) { return CAPACITY; }
+    }
+    // We truncate the input to the end of the last complete document (or zero).
+    auto new_structural_indexes = find_next_document_index(parser);
+    if (new_structural_indexes == 0 && parser.n_structural_indexes > 0) {
+      if(parser.structural_indexes[0] == 0) {
+        // If the buffer is partial and we started at index 0 but the document is
+        // incomplete, it's too big to parse.
+        return CAPACITY;
+      } else {
+        // It is possible that the document could be parsed, we just had a lot
+        // of white space.
+        parser.n_structural_indexes = 0;
+        return EMPTY;
+      }
+    }
+
+    parser.n_structural_indexes = new_structural_indexes;
+  } else if (partial == stage1_mode::streaming_final) {
+    if(have_unclosed_string) { parser.n_structural_indexes--; }
+    // We truncate the input to the end of the last complete document (or zero).
+    // Because partial == stage1_mode::streaming_final, it means that we may
+    // silently ignore trailing garbage. Though it sounds bad, we do it
+    // deliberately because many people who have streams of JSON documents
+    // will truncate them for processing. E.g., imagine that you are uncompressing
+    // the data from a size file or receiving it in chunks from the network. You
+    // may not know where exactly the last document will be. Meanwhile the
+    // document_stream instances allow people to know the JSON documents they are
+    // parsing (see the iterator.source() method).
+    parser.n_structural_indexes = find_next_document_index(parser);
+    // We store the initial n_structural_indexes so that the client can see
+    // whether we used truncation. If initial_n_structural_indexes == parser.n_structural_indexes,
+    // then this will query parser.structural_indexes[parser.n_structural_indexes] which is len,
+    // otherwise, it will copy some prior index.
+    parser.structural_indexes[parser.n_structural_indexes + 1] = parser.structural_indexes[parser.n_structural_indexes];
+    // This next line is critical, do not change it unless you understand what you are
+    // doing.
+    parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len);
+    if (simdjson_unlikely(parser.n_structural_indexes == 0u)) {
+        // We tolerate an unclosed string at the very end of the stream. Indeed, users
+        // often load their data in bulk without being careful and they want us to ignore
+        // the trailing garbage.
+        return EMPTY;
+    }
+  }
+  checker.check_eof();
+  return checker.errors();
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+// Clear CUSTOM_BIT_INDEXER so other implementations can set it if they need to.
+#undef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H
+/* end file generic/stage1/json_structural_indexer.h for lasx */
+/* including generic/stage1/utf8_validator.h for lasx: #include <generic/stage1/utf8_validator.h> */
+/* begin file generic/stage1/utf8_validator.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace stage1 {
+
+/**
+ * Validates that the string is actual UTF-8.
+ */
+template<class checker>
+bool generic_validate_utf8(const uint8_t * input, size_t length) {
+    checker c{};
+    buf_block_reader<64> reader(input, length);
+    while (reader.has_full_block()) {
+      simd::simd8x64<uint8_t> in(reader.full_block());
+      c.check_next_input(in);
+      reader.advance();
+    }
+    uint8_t block[64]{};
+    reader.get_remainder(block);
+    simd::simd8x64<uint8_t> in(block);
+    c.check_next_input(in);
+    reader.advance();
+    c.check_eof();
+    return c.errors() == error_code::SUCCESS;
+}
+
+bool generic_validate_utf8(const char * input, size_t length) {
+    return generic_validate_utf8<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H
+/* end file generic/stage1/utf8_validator.h for lasx */
+/* end file generic/stage1/amalgamated.h for lasx */
+/* including generic/stage2/amalgamated.h for lasx: #include <generic/stage2/amalgamated.h> */
+/* begin file generic/stage2/amalgamated.h for lasx */
+// Stuff other things depend on
+/* including generic/stage2/base.h for lasx: #include <generic/stage2/base.h> */
+/* begin file generic/stage2/base.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_BASE_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace stage2 {
+
+class json_iterator;
+class structural_iterator;
+struct tape_builder;
+struct tape_writer;
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_BASE_H
+/* end file generic/stage2/base.h for lasx */
+/* including generic/stage2/tape_writer.h for lasx: #include <generic/stage2/tape_writer.h> */
+/* begin file generic/stage2/tape_writer.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/internal/tape_type.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace stage2 {
+
+struct tape_writer {
+  /** The next place to write to tape */
+  uint64_t *next_tape_loc;
+
+  /** Write a signed 64-bit value to tape. */
+  simdjson_inline void append_s64(int64_t value) noexcept;
+
+  /** Write an unsigned 64-bit value to tape. */
+  simdjson_inline void append_u64(uint64_t value) noexcept;
+
+  /** Write a double value to tape. */
+  simdjson_inline void append_double(double value) noexcept;
+
+  /** Write a big integer (as string) to tape. src points to first digit, len is byte count. */
+  simdjson_inline void append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept;
+
+  /**
+   * Append a tape entry (an 8-bit type,and 56 bits worth of value).
+   */
+  simdjson_inline void append(uint64_t val, internal::tape_type t) noexcept;
+
+  /**
+   * Skip the current tape entry without writing.
+   *
+   * Used to skip the start of the container, since we'll come back later to fill it in when the
+   * container ends.
+   */
+  simdjson_inline void skip() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a large u64 or i64.
+   */
+  simdjson_inline void skip_large_integer() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a double.
+   */
+  simdjson_inline void skip_double() noexcept;
+
+  /**
+   * Write a value to a known location on tape.
+   *
+   * Used to go back and write out the start of a container after the container ends.
+   */
+  simdjson_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
+
+private:
+  /**
+   * Append both the tape entry, and a supplementary value following it. Used for types that need
+   * all 64 bits, such as double and uint64_t.
+   */
+  template<typename T>
+  simdjson_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
+}; // struct tape_writer
+
+simdjson_inline void tape_writer::append_s64(int64_t value) noexcept {
+  append2(0, value, internal::tape_type::INT64);
+}
+
+simdjson_inline void tape_writer::append_u64(uint64_t value) noexcept {
+  append(0, internal::tape_type::UINT64);
+  *next_tape_loc = value;
+  next_tape_loc++;
+}
+
+/** Write a double value to tape. */
+simdjson_inline void tape_writer::append_double(double value) noexcept {
+  append2(0, value, internal::tape_type::DOUBLE);
+}
+
+simdjson_inline void tape_writer::skip() noexcept {
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::skip_large_integer() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::skip_double() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
+  *next_tape_loc = val | ((uint64_t(char(t))) << 56);
+  next_tape_loc++;
+}
+
+template<typename T>
+simdjson_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
+  append(val, t);
+  static_assert(sizeof(val2) == sizeof(*next_tape_loc), "Type is not 64 bits!");
+  memcpy(next_tape_loc, &val2, sizeof(val2));
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
+  tape_loc = val | ((uint64_t(char(t))) << 56);
+}
+
+simdjson_inline void tape_writer::append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept {
+  // Write to string buffer: [4-byte LE length][digits][null]
+  uint32_t str_len = uint32_t(len);
+  memcpy(string_buf, &str_len, sizeof(uint32_t));
+  memcpy(string_buf + sizeof(uint32_t), src, len);
+  string_buf[sizeof(uint32_t) + len] = 0;
+  // Tape entry: offset into string buffer
+  // The caller must set the offset relative to doc.string_buf base
+  append(0, internal::tape_type::BIGINT); // placeholder offset, caller patches
+  string_buf += sizeof(uint32_t) + len + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+/* end file generic/stage2/tape_writer.h for lasx */
+/* including generic/stage2/logger.h for lasx: #include <generic/stage2/logger.h> */
+/* begin file generic/stage2/logger.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+
+// This is for an internal-only stage 2 specific logger.
+// Set LOG_ENABLED = true to log what stage 2 is doing!
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace logger {
+
+  static constexpr const char * DASHES = "----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------";
+
+#if SIMDJSON_VERBOSE_LOGGING
+  static constexpr const bool LOG_ENABLED = true;
+#else
+  static constexpr const bool LOG_ENABLED = false;
+#endif
+  static constexpr const int LOG_EVENT_LEN = 20;
+  static constexpr const int LOG_BUFFER_LEN = 30;
+  static constexpr const int LOG_SMALL_BUFFER_LEN = 10;
+  static constexpr const int LOG_INDEX_LEN = 5;
+
+  static int log_depth; // Not threadsafe. Log only.
+
+  // Helper to turn unprintable or newline characters into spaces
+  static simdjson_inline char printable_char(char c) {
+    if (c >= 0x20) {
+      return c;
+    } else {
+      return ' ';
+    }
+  }
+
+  // Print the header and set up log_start
+  static simdjson_inline void log_start() {
+    if (LOG_ENABLED) {
+      log_depth = 0;
+      printf("\n");
+      printf("| %-*s | %-*s | %-*s | %-*s | Detail |\n", LOG_EVENT_LEN, "Event", LOG_BUFFER_LEN, "Buffer", LOG_SMALL_BUFFER_LEN, "Next", 5, "Next#");
+      printf("|%.*s|%.*s|%.*s|%.*s|--------|\n", LOG_EVENT_LEN+2, DASHES, LOG_BUFFER_LEN+2, DASHES, LOG_SMALL_BUFFER_LEN+2, DASHES, 5+2, DASHES);
+    }
+  }
+
+  simdjson_unused static simdjson_inline void log_string(const char *message) {
+    if (LOG_ENABLED) {
+      printf("%s\n", message);
+    }
+  }
+
+  // Logs a single line from the stage 2 DOM parser
+  template<typename S>
+  static simdjson_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
+    if (LOG_ENABLED) {
+      printf("| %*s%s%-*s ", log_depth*2, "", title_prefix, LOG_EVENT_LEN - log_depth*2 - int(strlen(title_prefix)), title);
+      auto current_index = structurals.at_beginning() ? nullptr : structurals.next_structural-1;
+      auto next_index = structurals.next_structural;
+      auto current = current_index ? &structurals.buf[*current_index] : reinterpret_cast<const uint8_t*>("                                                       ");
+      auto next = &structurals.buf[*next_index];
+      {
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_BUFFER_LEN;i++) {
+          printf("%c", printable_char(current[i]));
+        }
+        printf(" ");
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_SMALL_BUFFER_LEN;i++) {
+          printf("%c", printable_char(next[i]));
+        }
+        printf(" ");
+      }
+      if (current_index) {
+        printf("| %*u ", LOG_INDEX_LEN, *current_index);
+      } else {
+        printf("| %-*s ", LOG_INDEX_LEN, "");
+      }
+      // printf("| %*u ", LOG_INDEX_LEN, structurals.next_tape_index());
+      printf("| %-s ", detail);
+      printf("|\n");
+    }
+  }
+
+} // namespace logger
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+/* end file generic/stage2/logger.h for lasx */
+
+// All other declarations
+/* including generic/stage2/json_iterator.h for lasx: #include <generic/stage2/json_iterator.h> */
+/* begin file generic/stage2/json_iterator.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/logger.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace stage2 {
+
+class json_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+  uint32_t depth{0};
+
+  /**
+   * Walk the JSON document.
+   *
+   * The visitor receives callbacks when values are encountered. All callbacks pass the iterator as
+   * the first parameter; some callbacks have other parameters as well:
+   *
+   * - visit_document_start() - at the beginning.
+   * - visit_document_end() - at the end (if things were successful).
+   *
+   * - visit_array_start() - at the start `[` of a non-empty array.
+   * - visit_array_end() - at the end `]` of a non-empty array.
+   * - visit_empty_array() - when an empty array is encountered.
+   *
+   * - visit_object_end() - at the start `]` of a non-empty object.
+   * - visit_object_start() - at the end `]` of a non-empty object.
+   * - visit_empty_object() - when an empty object is encountered.
+   * - visit_key(const uint8_t *key) - when a key in an object field is encountered. key is
+   *                                   guaranteed to point at the first quote of the string (`"key"`).
+   * - visit_primitive(const uint8_t *value) - when a value is a string, number, boolean or null.
+   * - visit_root_primitive(iter, uint8_t *value) - when the top-level value is a string, number, boolean or null.
+   *
+   * - increment_count(iter) - each time a value is found in an array or object.
+   */
+  template<bool STREAMING, typename V>
+  simdjson_warn_unused simdjson_inline error_code walk_document(V &visitor) noexcept;
+
+  /**
+   * Create an iterator capable of walking a JSON document.
+   *
+   * The document must have already passed through stage 1.
+   */
+  simdjson_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
+
+  /**
+   * Look at the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *peek() const noexcept;
+  /**
+   * Advance to the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *advance() noexcept;
+  /**
+   * Get the remaining length of the document, from the start of the current token.
+   */
+  simdjson_inline size_t remaining_len() const noexcept;
+  /**
+   * Check if we are at the end of the document.
+   *
+   * If this is true, there are no more tokens.
+   */
+  simdjson_inline bool at_eof() const noexcept;
+  /**
+   * Check if we are at the beginning of the document.
+   */
+  simdjson_inline bool at_beginning() const noexcept;
+  simdjson_inline uint8_t last_structural() const noexcept;
+
+  /**
+   * Log that a value has been found.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_value(const char *type) const noexcept;
+  /**
+   * Log the start of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_start_value(const char *type) const noexcept;
+  /**
+   * Log the end of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_end_value(const char *type) const noexcept;
+  /**
+   * Log an error.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_error(const char *error) const noexcept;
+
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
+};
+
+template<bool STREAMING, typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::walk_document(V &visitor) noexcept {
+  logger::log_start();
+
+  //
+  // Start the document
+  //
+  if (at_eof()) { return EMPTY; }
+  log_start_value("document");
+  SIMDJSON_TRY( visitor.visit_document_start(*this) );
+
+  //
+  // Read first value
+  //
+  {
+    auto value = advance();
+
+    // Make sure the outer object or array is closed before continuing; otherwise, there are ways we
+    // could get into memory corruption. See https://github.com/simdjson/simdjson/issues/906
+    if (!STREAMING) {
+      switch (*value) {
+        case '{': if (last_structural() != '}') { log_value("starting brace unmatched"); return TAPE_ERROR; }; break;
+        case '[': if (last_structural() != ']') { log_value("starting bracket unmatched"); return TAPE_ERROR; }; break;
+      }
+    }
+
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_root_primitive(*this, value) ); break;
+    }
+  }
+  goto document_end;
+
+//
+// Object parser states
+//
+object_begin:
+  log_start_value("object");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = false;
+  SIMDJSON_TRY( visitor.visit_object_start(*this) );
+
+  {
+    auto key = advance();
+    if (*key != '"') { log_error("Object does not start with a key"); return TAPE_ERROR; }
+    SIMDJSON_TRY( visitor.increment_count(*this) );
+    SIMDJSON_TRY( visitor.visit_key(*this, key) );
+  }
+
+object_field:
+  if (simdjson_unlikely( *advance() != ':' )) { log_error("Missing colon after key in object"); return TAPE_ERROR; }
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+object_continue:
+  switch (*advance()) {
+    case ',':
+      SIMDJSON_TRY( visitor.increment_count(*this) );
+      {
+        auto key = advance();
+        if (simdjson_unlikely( *key != '"' )) { log_error("Key string missing at beginning of field in object"); return TAPE_ERROR; }
+        SIMDJSON_TRY( visitor.visit_key(*this, key) );
+      }
+      goto object_field;
+    case '}': log_end_value("object"); SIMDJSON_TRY( visitor.visit_object_end(*this) ); goto scope_end;
+    default: log_error("No comma between object fields"); return TAPE_ERROR;
+  }
+
+scope_end:
+  depth--;
+  if (depth == 0) { goto document_end; }
+  if (dom_parser.is_array[depth]) { goto array_continue; }
+  goto object_continue;
+
+//
+// Array parser states
+//
+array_begin:
+  log_start_value("array");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = true;
+  SIMDJSON_TRY( visitor.visit_array_start(*this) );
+  SIMDJSON_TRY( visitor.increment_count(*this) );
+
+array_value:
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+array_continue:
+  switch (*advance()) {
+    case ',': SIMDJSON_TRY( visitor.increment_count(*this) ); goto array_value;
+    case ']': log_end_value("array"); SIMDJSON_TRY( visitor.visit_array_end(*this) ); goto scope_end;
+    default: log_error("Missing comma between array values"); return TAPE_ERROR;
+  }
+
+document_end:
+  log_end_value("document");
+  SIMDJSON_TRY( visitor.visit_document_end(*this) );
+
+  dom_parser.next_structural_index = uint32_t(next_structural - &dom_parser.structural_indexes[0]);
+
+  // If we didn't make it to the end, it's an error
+  if ( !STREAMING && dom_parser.next_structural_index != dom_parser.n_structural_indexes ) {
+    log_error("More than one JSON value at the root of the document, or extra characters at the end of the JSON!");
+    return TAPE_ERROR;
+  }
+
+  return SUCCESS;
+
+} // walk_document()
+
+simdjson_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+  : buf{_dom_parser.buf},
+    next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+    dom_parser{_dom_parser} {
+}
+
+simdjson_inline const uint8_t *json_iterator::peek() const noexcept {
+  return &buf[*(next_structural)];
+}
+simdjson_inline const uint8_t *json_iterator::advance() noexcept {
+  return &buf[*(next_structural++)];
+}
+simdjson_inline size_t json_iterator::remaining_len() const noexcept {
+  return dom_parser.len - *(next_structural-1);
+}
+
+simdjson_inline bool json_iterator::at_eof() const noexcept {
+  return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+}
+simdjson_inline bool json_iterator::at_beginning() const noexcept {
+  return next_structural == dom_parser.structural_indexes.get();
+}
+simdjson_inline uint8_t json_iterator::last_structural() const noexcept {
+  return buf[dom_parser.structural_indexes[dom_parser.n_structural_indexes - 1]];
+}
+
+simdjson_inline void json_iterator::log_value(const char *type) const noexcept {
+  logger::log_line(*this, "", type, "");
+}
+
+simdjson_inline void json_iterator::log_start_value(const char *type) const noexcept {
+  logger::log_line(*this, "+", type, "");
+  if (logger::LOG_ENABLED) { logger::log_depth++; }
+}
+
+simdjson_inline void json_iterator::log_end_value(const char *type) const noexcept {
+  if (logger::LOG_ENABLED) { logger::log_depth--; }
+  logger::log_line(*this, "-", type, "");
+}
+
+simdjson_inline void json_iterator::log_error(const char *error) const noexcept {
+  logger::log_line(*this, "", "ERROR", error);
+}
+
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
+  switch (*value) {
+    case '"': return visitor.visit_root_string(*this, value);
+    case 't': return visitor.visit_root_true_atom(*this, value);
+    case 'f': return visitor.visit_root_false_atom(*this, value);
+    case 'n': return visitor.visit_root_null_atom(*this, value);
+    case '-':
+    case '0': case '1': case '2': case '3': case '4':
+    case '5': case '6': case '7': case '8': case '9':
+      return visitor.visit_root_number(*this, value);
+    default:
+      log_error("Document starts with a non-value character");
+      return TAPE_ERROR;
+  }
+}
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
+  // Use the fact that most scalars are going to be either strings or numbers.
+  if(*value == '"') {
+    return visitor.visit_string(*this, value);
+  } else if (((*value - '0')  < 10) || (*value == '-')) {
+    return visitor.visit_number(*this, value);
+  }
+  // true, false, null are uncommon.
+  switch (*value) {
+    case 't': return visitor.visit_true_atom(*this, value);
+    case 'f': return visitor.visit_false_atom(*this, value);
+    case 'n': return visitor.visit_null_atom(*this, value);
+    default:
+      log_error("Non-value found when value was expected!");
+      return TAPE_ERROR;
+  }
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+/* end file generic/stage2/json_iterator.h for lasx */
+/* including generic/stage2/stringparsing.h for lasx: #include <generic/stage2/stringparsing.h> */
+/* begin file generic/stage2/stringparsing.h for lasx */
+#include <cstdint>
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/jsoncharutils.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses
+// It is intended to be included multiple times and compiled multiple times
+
+namespace simdjson {
+namespace lasx {
+namespace {
+/// @private
+namespace stringparsing {
+
+// begin copypasta
+// These chars yield themselves: " \ /
+// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
+// u not handled in this table as it's complex
+static const uint8_t escape_map[256] = {
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
+    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
+    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+};
+
+// handle a unicode codepoint
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr, bool allow_replacement) {
+  // Use the default Unicode Character 'REPLACEMENT CHARACTER' (U+FFFD)
+  constexpr uint32_t substitution_code_point = 0xfffd;
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) != ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+    } else {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+
+      // We have already checked that the high surrogate is valid and
+      // (code_point - 0xd800) < 1024.
+      //
+      // Check that code_point_2 is in the range 0xdc00..0xdfff
+      // and that code_point_2 was parsed from valid hex.
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if (low_bit >> 10) {
+        if(!allow_replacement) { return false; }
+        code_point = substitution_code_point;
+      } else {
+        code_point =  (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+
+    }
+  } else if (code_point >= 0xdc00 && code_point <= 0xdfff) {
+      // If we encounter a low surrogate (not preceded by a high surrogate)
+      // then we have an error.
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+  }
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+// handle a unicode codepoint using the wobbly convention
+// https://simonsapin.github.io/wtf-8/
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint_wobbly(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr) {
+  // It is not ideal that this function is nearly identical to handle_unicode_codepoint.
+  //
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) == ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if ((low_bit >> 10) ==  0) {
+        code_point =
+          (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+    }
+  }
+
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+/**
+ * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+ * must be an unescaped quote terminating the string. It returns the final output
+ * position as pointer. In case of error (e.g., the string has bad escaped codes),
+ * then null_ptr is returned. It is assumed that the output buffer is large
+ * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+ * SIMDJSON_PADDING bytes.
+ */
+simdjson_warn_unused simdjson_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) {
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint(&src, &dst, allow_replacement)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+simdjson_warn_unused simdjson_inline uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) {
+  // It is not ideal that this function is nearly identical to parse_string.
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint_wobbly(&src, &dst)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+} // namespace stringparsing
+
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+/* end file generic/stage2/stringparsing.h for lasx */
+/* including generic/stage2/structural_iterator.h for lasx: #include <generic/stage2/structural_iterator.h> */
+/* begin file generic/stage2/structural_iterator.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace stage2 {
+
+class structural_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+
+  // Start a structural
+  simdjson_inline structural_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+    : buf{_dom_parser.buf},
+      next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+      dom_parser{_dom_parser} {
+  }
+  // Get the buffer position of the current structural character
+  simdjson_inline const uint8_t* current() {
+    return &buf[*(next_structural-1)];
+  }
+  // Get the current structural character
+  simdjson_inline char current_char() {
+    return buf[*(next_structural-1)];
+  }
+  // Get the next structural character without advancing
+  simdjson_inline char peek_next_char() {
+    return buf[*next_structural];
+  }
+  simdjson_inline const uint8_t* peek() {
+    return &buf[*next_structural];
+  }
+  simdjson_inline const uint8_t* advance() {
+    return &buf[*(next_structural++)];
+  }
+  simdjson_inline char advance_char() {
+    return buf[*(next_structural++)];
+  }
+  simdjson_inline size_t remaining_len() {
+    return dom_parser.len - *(next_structural-1);
+  }
+
+  simdjson_inline bool at_end() {
+    return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+  }
+  simdjson_inline bool at_beginning() {
+    return next_structural == dom_parser.structural_indexes.get();
+  }
+};
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H
+/* end file generic/stage2/structural_iterator.h for lasx */
+/* including generic/stage2/tape_builder.h for lasx: #include <generic/stage2/tape_builder.h> */
+/* begin file generic/stage2/tape_builder.h for lasx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/json_iterator.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/stringparsing.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/tape_writer.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/dom/document.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/atomparsing.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/numberparsing.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+
+namespace simdjson {
+namespace lasx {
+namespace {
+namespace stage2 {
+
+struct tape_builder {
+  template<bool STREAMING>
+  simdjson_warn_unused static simdjson_inline error_code parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept;
+
+  /** Called when a non-empty document starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty document ends without error. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_end(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty array starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty array ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_end(json_iterator &iter) noexcept;
+  /** Called when an empty array is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_array(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty object starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_start(json_iterator &iter) noexcept;
+  /**
+   * Called when a key in a field is encountered.
+   *
+   * primitive, visit_object_start, visit_empty_object, visit_array_start, or visit_empty_array
+   * will be called after this with the field value.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
+  /** Called when a non-empty object ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_end(json_iterator &iter) noexcept;
+  /** Called when an empty object is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_object(json_iterator &iter) noexcept;
+
+  /**
+   * Called when a string, number, boolean or null is found.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  /**
+   * Called when a string, number, boolean or null is found at the top level of a document (i.e.
+   * when there is no array or object and the entire document is a single string, number, boolean or
+   * null.
+   *
+   * This is separate from primitive() because simdjson's normal primitive parsing routines assume
+   * there is at least one more token after the value, which is only true in an array or object.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  /** Called each time a new field or element in an array or object is found. */
+  simdjson_warn_unused simdjson_inline error_code increment_count(json_iterator &iter) noexcept;
+
+  /** Next location to write to tape */
+  tape_writer tape;
+private:
+  /** Next write location in the string buf for stage 2 parsing */
+  uint8_t *current_string_buf_loc;
+
+  simdjson_inline tape_builder(dom::document &doc) noexcept;
+
+  simdjson_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
+  simdjson_inline void start_container(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_warn_unused simdjson_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
+  simdjson_inline void on_end_string(uint8_t *dst) noexcept;
+}; // struct tape_builder
+
+template<bool STREAMING>
+simdjson_warn_unused simdjson_inline error_code tape_builder::parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept {
+  dom_parser.doc = &doc;
+  json_iterator iter(dom_parser, STREAMING ? dom_parser.next_structural_index : 0);
+  tape_builder builder(doc);
+  return iter.walk_document<STREAMING>(builder);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_root_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
+  constexpr uint32_t start_tape_index = 0;
+  tape.append(start_tape_index, internal::tape_type::ROOT);
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter), internal::tape_type::ROOT);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
+  return visit_string(iter, key, true);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].count++; // we have a key value pair in the object at parser.dom_parser.depth - 1
+  return SUCCESS;
+}
+
+simdjson_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
+  iter.log_value(key ? "key" : "string");
+  uint8_t *dst = on_start_string(iter);
+  dst = stringparsing::parse_string(value+1, dst, false); // We do not allow replacement when the escape characters are invalid.
+  if (dst == nullptr) {
+    iter.log_error("Invalid escape in string");
+    return STRING_ERROR;
+  }
+  on_end_string(dst);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
+  return visit_string(iter, value);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("number");
+  error_code err = numberparsing::parse_number(value, tape);
+  if (simdjson_unlikely(err == BIGINT_ERROR &&
+      iter.dom_parser._number_as_string)) {
+    // Write big integer to string buffer using the same format as strings.
+    // Scan digits the same way parse_number does (skip optional '-', then digits).
+    const uint8_t *p = value;
+    if (*p == '-') p++;
+    while (numberparsing::is_digit(*p)) p++;
+    size_t len = size_t(p - value);
+    tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::BIGINT);
+    uint8_t *dst = current_string_buf_loc + sizeof(uint32_t);
+    memcpy(dst, value, len);
+    dst += len;
+    on_end_string(dst);
+    return SUCCESS;
+  }
+  return err;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
+  //
+  // We need to make a copy to make sure that the string is space terminated.
+  // This is not about padding the input, which should already padded up
+  // to len + SIMDJSON_PADDING. However, we have no control at this stage
+  // on how the padding was done. What if the input string was padded with nulls?
+  // It is quite common for an input string to have an extra null character (C string).
+  // We do not want to allow 9\0 (where \0 is the null character) inside a JSON
+  // document, but the string "9\0" by itself is fine. So we make a copy and
+  // pad the input with spaces when we know that there is just one input element.
+  // This copy is relatively expensive, but it will almost never be called in
+  // practice unless you are in the strange scenario where you have many JSON
+  // documents made of single atoms.
+  //
+  std::unique_ptr<uint8_t[]>copy(new (std::nothrow) uint8_t[iter.remaining_len() + SIMDJSON_PADDING]);
+  if (copy.get() == nullptr) { return MEMALLOC; }
+  std::memcpy(copy.get(), value, iter.remaining_len());
+  std::memset(copy.get() + iter.remaining_len(), ' ', SIMDJSON_PADDING);
+  error_code error = visit_number(iter, copy.get());
+  return error;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value)) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value, iter.remaining_len())) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value)) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value, iter.remaining_len())) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value)) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value, iter.remaining_len())) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+// private:
+
+simdjson_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
+  return uint32_t(tape.next_tape_loc - iter.dom_parser.doc->tape.get());
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  auto start_index = next_tape_index(iter);
+  tape.append(start_index+2, start);
+  tape.append(start_index, end);
+  return SUCCESS;
+}
+
+simdjson_inline void tape_builder::start_container(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].tape_index = next_tape_index(iter);
+  iter.dom_parser.open_containers[iter.depth].count = 0;
+  tape.skip(); // We don't actually *write* the start element until the end.
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  // Write the ending tape element, pointing at the start location
+  const uint32_t start_tape_index = iter.dom_parser.open_containers[iter.depth].tape_index;
+  tape.append(start_tape_index, end);
+  // Write the start tape element, pointing at the end location (and including count)
+  // count can overflow if it exceeds 24 bits... so we saturate
+  // the convention being that a cnt of 0xffffff or more is undetermined in value (>=  0xffffff).
+  const uint32_t count = iter.dom_parser.open_containers[iter.depth].count;
+  const uint32_t cntsat = count > 0xFFFFFF ? 0xFFFFFF : count;
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter) | (uint64_t(cntsat) << 32), start);
+  return SUCCESS;
+}
+
+simdjson_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
+  // we advance the point, accounting for the fact that we have a NULL termination
+  tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::STRING);
+  return current_string_buf_loc + sizeof(uint32_t);
+}
+
+simdjson_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
+  uint32_t str_length = uint32_t(dst - (current_string_buf_loc + sizeof(uint32_t)));
+  // TODO check for overflow in case someone has a crazy string (>=4GB?)
+  // But only add the overflow check when the document itself exceeds 4GB
+  // Currently unneeded because we refuse to parse docs larger or equal to 4GB.
+  memcpy(current_string_buf_loc, &str_length, sizeof(uint32_t));
+  // NULL termination is still handy if you expect all your strings to
+  // be NULL terminated? It comes at a small cost
+  *dst = 0;
+  current_string_buf_loc = dst + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+/* end file generic/stage2/tape_builder.h for lasx */
+/* end file generic/stage2/amalgamated.h for lasx */
+
+//
+// Stage 1
+//
+namespace simdjson {
+namespace lasx {
+
+simdjson_warn_unused error_code implementation::create_dom_parser_implementation(
+  size_t capacity,
+  size_t max_depth,
+  std::unique_ptr<internal::dom_parser_implementation>& dst
+) const noexcept {
+  dst.reset( new (std::nothrow) dom_parser_implementation() );
+  if (!dst) { return MEMALLOC; }
+  if (auto err = dst->set_capacity(capacity))
+    return err;
+  if (auto err = dst->set_max_depth(max_depth))
+    return err;
+  return SUCCESS;
+}
+
+namespace {
+
+using namespace simd;
+
+simdjson_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
+  // Inspired by haswell.
+  // LASX use low 5 bits as index. For the 6 operators (:,[]{}), the unique-5bits is [6:2].
+  // The ASCII white-space and operators have these values: (char, hex, unique-5bits)
+  // (' ', 20, 00000) ('\t', 09, 01001) ('\n', 0A, 01010) ('\r', 0D, 01101)
+  // (',', 2C, 01011) (':', 3A, 01110) ('[', 5B, 10110) ('{', 7B, 11110) (']', 5D, 10111) ('}', 7D, 11111)
+  const simd8<uint8_t> ws_table = simd8<uint8_t>::repeat_16(
+    ' ', 0, 0, 0, 0, 0, 0, 0, 0, '\t', '\n', 0, 0, '\r', 0, 0
+  );
+  const simd8<uint8_t> op_table_lo = simd8<uint8_t>::repeat_16(
+    1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ',', 0, 0, ':', 0
+  );
+  const simd8<uint8_t> op_table_hi = simd8<uint8_t>::repeat_16(
+    0, 0, 0, 0, 0, 0, '[', ']', 0, 0, 0, 0, 0, 0, '{', '}'
+  );
+  uint64_t ws = in.eq({
+    in.chunks[0].lookup_16(ws_table),
+    in.chunks[1].lookup_16(ws_table),
+  });
+  uint64_t op = in.eq({
+    __lasx_xvshuf_b(op_table_hi, op_table_lo, in.chunks[0].shr<2>()),
+    __lasx_xvshuf_b(op_table_hi, op_table_lo, in.chunks[1].shr<2>()),
+  });
+
+  return { ws, op };
+}
+
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input) {
+  return input.reduce_or().is_ascii();
+}
+
+simdjson_inline simd8<uint8_t> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+    simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-0x80); // Only 111_____ will be >= 0x80
+    simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-0x80); // Only 1111____ will be >= 0x80
+    return is_third_byte | is_fourth_byte;
+}
+
+} // unnamed namespace
+} // namespace lasx
+} // namespace simdjson
+
+//
+// Stage 2
+//
+
+//
+// Implementation-specific overrides
+//
+namespace simdjson {
+namespace lasx {
+
+simdjson_warn_unused error_code implementation::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept {
+  return lasx::stage1::json_minifier::minify<64>(buf, len, dst, dst_len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage1(const uint8_t *_buf, size_t _len, stage1_mode streaming) noexcept {
+  this->buf = _buf;
+  this->len = _len;
+  return lasx::stage1::json_structural_indexer::index<64>(buf, len, *this, streaming);
+}
+
+simdjson_warn_unused bool implementation::validate_utf8(const char *buf, size_t len) const noexcept {
+  return lasx::stage1::generic_validate_utf8(buf,len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<false>(*this, _doc);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2_next(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<true>(*this, _doc);
+}
+
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) const noexcept {
+  return lasx::stringparsing::parse_string(src, dst, allow_replacement);
+}
+
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept {
+  return lasx::stringparsing::parse_wobbly_string(src, dst);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::parse(const uint8_t *_buf, size_t _len, dom::document &_doc) noexcept {
+  auto error = stage1(_buf, _len, stage1_mode::regular);
+  if (error) { return error; }
+  return stage2(_doc);
+}
+
+} // namespace lasx
+} // namespace simdjson
+
+/* including simdjson/lasx/end.h: #include <simdjson/lasx/end.h> */
+/* begin file simdjson/lasx/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#undef SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+/* undefining SIMDJSON_IMPLEMENTATION from "lasx" */
+#undef SIMDJSON_IMPLEMENTATION
+
+
+#if SIMDJSON_CAN_ALWAYS_RUN_LASX
+// nothing needed.
+#else
+SIMDJSON_UNTARGET_REGION
+#endif
+/* end file simdjson/lasx/end.h */
+
+#endif // SIMDJSON_SRC_LASX_CPP
+/* end file lasx.cpp */
+#endif
+#if SIMDJSON_IMPLEMENTATION_LSX
+/* including lsx.cpp: #include <lsx.cpp> */
+/* begin file lsx.cpp */
+#ifndef SIMDJSON_SRC_LSX_CPP
+#define SIMDJSON_SRC_LSX_CPP
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+/* including simdjson/lsx.h: #include <simdjson/lsx.h> */
+/* begin file simdjson/lsx.h */
+#ifndef SIMDJSON_LSX_H
+#define SIMDJSON_LSX_H
+
+/* including simdjson/lsx/begin.h: #include "simdjson/lsx/begin.h" */
+/* begin file simdjson/lsx/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "lsx" */
+#define SIMDJSON_IMPLEMENTATION lsx
+/* including simdjson/lsx/base.h: #include "simdjson/lsx/base.h" */
+/* begin file simdjson/lsx/base.h */
+#ifndef SIMDJSON_LSX_BASE_H
+#define SIMDJSON_LSX_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+/**
+ * Implementation for LSX.
+ */
+namespace lsx {
+
+class implementation;
+
+namespace {
+namespace simd {
+template <typename T> struct simd8;
+template <typename T> struct simd8x64;
+} // namespace simd
+} // unnamed namespace
+
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_LSX_BASE_H
+/* end file simdjson/lsx/base.h */
+/* including simdjson/lsx/intrinsics.h: #include "simdjson/lsx/intrinsics.h" */
+/* begin file simdjson/lsx/intrinsics.h */
+#ifndef SIMDJSON_LSX_INTRINSICS_H
+#define SIMDJSON_LSX_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <lsxintrin.h>
+
+static_assert(sizeof(__m128i) <= simdjson::SIMDJSON_PADDING, "insufficient padding for LoongArch SX");
+
+#endif //  SIMDJSON_LSX_INTRINSICS_H
+/* end file simdjson/lsx/intrinsics.h */
+/* including simdjson/lsx/bitmanipulation.h: #include "simdjson/lsx/bitmanipulation.h" */
+/* begin file simdjson/lsx/bitmanipulation.h */
+#ifndef SIMDJSON_LSX_BITMANIPULATION_H
+#define SIMDJSON_LSX_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/bitmask.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+  return __builtin_ctzll(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return input_num & (input_num-1);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+  return __builtin_clzll(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int count_ones(uint64_t input_num) {
+  return __lsx_vpickve2gr_w(__lsx_vpcnt_d(__m128i(v2u64{input_num, 0})), 0);
+}
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2, uint64_t *result) {
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+}
+
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_LSX_BITMANIPULATION_H
+/* end file simdjson/lsx/bitmanipulation.h */
+/* including simdjson/lsx/bitmask.h: #include "simdjson/lsx/bitmask.h" */
+/* begin file simdjson/lsx/bitmask.h */
+#ifndef SIMDJSON_LSX_BITMASK_H
+#define SIMDJSON_LSX_BITMASK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(uint64_t bitmask) {
+  bitmask ^= bitmask << 1;
+  bitmask ^= bitmask << 2;
+  bitmask ^= bitmask << 4;
+  bitmask ^= bitmask << 8;
+  bitmask ^= bitmask << 16;
+  bitmask ^= bitmask << 32;
+  return bitmask;
+}
+
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif
+/* end file simdjson/lsx/bitmask.h */
+/* including simdjson/lsx/numberparsing_defs.h: #include "simdjson/lsx/numberparsing_defs.h" */
+/* begin file simdjson/lsx/numberparsing_defs.h */
+#ifndef SIMDJSON_LSX_NUMBERPARSING_DEFS_H
+#define SIMDJSON_LSX_NUMBERPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace lsx {
+namespace numberparsing {
+
+// we don't have appropriate instructions, so let us use a scalar function
+// credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  uint64_t val;
+  std::memcpy(&val, chars, sizeof(uint64_t));
+  val = (val & 0x0F0F0F0F0F0F0F0F) * 2561 >> 8;
+  val = (val & 0x00FF00FF00FF00FF) * 6553601 >> 16;
+  return uint32_t((val & 0x0000FFFF0000FFFF) * 42949672960001 >> 32);
+}
+
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace lsx
+} // namespace simdjson
+
+#ifndef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_IS_BIG_ENDIAN
+#define SIMDJSON_SWAR_NUMBER_PARSING 0
+#else
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+#endif
+#endif
+
+#endif // SIMDJSON_LSX_NUMBERPARSING_DEFS_H
+/* end file simdjson/lsx/numberparsing_defs.h */
+/* including simdjson/lsx/simd.h: #include "simdjson/lsx/simd.h" */
+/* begin file simdjson/lsx/simd.h */
+#ifndef SIMDJSON_LSX_SIMD_H
+#define SIMDJSON_LSX_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace simd {
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename Child>
+  struct base {
+    __m128i value;
+
+    // Zero constructor
+    simdjson_inline base() : value{__m128i()} {}
+
+    // Conversion from SIMD register
+    simdjson_inline base(const __m128i _value) : value(_value) {}
+
+    // Conversion to SIMD register
+    simdjson_inline operator const __m128i&() const { return this->value; }
+    simdjson_inline operator __m128i&() { return this->value; }
+    simdjson_inline operator const v16i8&() const { return (v16i8&)this->value; }
+    simdjson_inline operator v16i8&() { return (v16i8&)this->value; }
+
+    // Bit operations
+    simdjson_inline Child operator|(const Child other) const { return __lsx_vor_v(*this, other); }
+    simdjson_inline Child operator&(const Child other) const { return __lsx_vand_v(*this, other); }
+    simdjson_inline Child operator^(const Child other) const { return __lsx_vxor_v(*this, other); }
+    simdjson_inline Child bit_andnot(const Child other) const { return __lsx_vandn_v(other, *this); }
+    simdjson_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+  };
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename T>
+  struct simd8;
+
+  template<typename T, typename Mask=simd8<bool>>
+  struct base8: base<simd8<T>> {
+    simdjson_inline base8() : base<simd8<T>>() {}
+    simdjson_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}
+
+    friend simdjson_really_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return __lsx_vseq_b(lhs, rhs); }
+
+    static const int SIZE = sizeof(base<simd8<T>>::value);
+
+    template<int N=1>
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+      return __lsx_vor_v(__lsx_vbsll_v(*this, N), __lsx_vbsrl_v(prev_chunk, 16 - N));
+    }
+  };
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool>: base8<bool> {
+    static simdjson_inline simd8<bool> splat(bool _value) {
+      return __lsx_vreplgr2vr_b(uint8_t(-(!!_value)));
+    }
+
+    simdjson_inline simd8() : base8() {}
+    simdjson_inline simd8(const __m128i _value) : base8<bool>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+
+    simdjson_inline int to_bitmask() const { return __lsx_vpickve2gr_w(__lsx_vmskltz_b(*this), 0); }
+    simdjson_inline bool any() const { return 0 == __lsx_vpickve2gr_hu(__lsx_vmsknz_b(*this), 0); }
+    simdjson_inline simd8<bool> operator~() const { return *this ^ true; }
+  };
+
+  template<typename T>
+  struct base8_numeric: base8<T> {
+    static simdjson_inline simd8<T> splat(T _value) { return __lsx_vreplgr2vr_b(_value); }
+    static simdjson_inline simd8<T> zero() { return __lsx_vldi(0); }
+    static simdjson_inline simd8<T> load(const T values[16]) {
+      return __lsx_vld(reinterpret_cast<const __m128i *>(values), 0);
+    }
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    static simdjson_inline simd8<T> repeat_16(
+      T v0,  T v1,  T v2,  T v3,  T v4,  T v5,  T v6,  T v7,
+      T v8,  T v9,  T v10, T v11, T v12, T v13, T v14, T v15
+    ) {
+      return simd8<T>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    simdjson_inline base8_numeric() : base8<T>() {}
+    simdjson_inline base8_numeric(const __m128i _value) : base8<T>(_value) {}
+
+    // Store to array
+    simdjson_inline void store(T dst[16]) const {
+      return __lsx_vst(*this, reinterpret_cast<__m128i *>(dst), 0);
+    }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<T> operator+(const simd8<T> other) const { return __lsx_vadd_b(*this, other); }
+    simdjson_inline simd8<T> operator-(const simd8<T> other) const { return __lsx_vsub_b(*this, other); }
+    simdjson_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
+
+    // Override to distinguish from bool version
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return __lsx_vshuf_b(lookup_table, lookup_table, *this);
+    }
+
+    // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted as a bitset).
+    // Passing a 0 value for mask would be equivalent to writing out every byte to output.
+    // Only the first 16 - count_ones(mask) bytes of the result are significant but 16 bytes
+    // get written.
+    template<typename L>
+    simdjson_inline void compress(uint16_t mask, L * output) const {
+      using internal::thintable_epi8;
+      using internal::BitsSetTable256mul2;
+      using internal::pshufb_combine_table;
+      // this particular implementation was inspired by haswell
+      // lsx do it in 2 steps, first 8 bytes and then second 8 bytes...
+      uint8_t mask1 = uint8_t(mask); // least significant 8 bits
+      uint8_t mask2 = uint8_t(mask >> 8); // second least significant 8 bits
+      // next line just loads the 64-bit values thintable_epi8[mask1] and
+      // thintable_epi8[mask2] into a 128-bit register.
+      __m128i shufmask = {int64_t(thintable_epi8[mask1]), int64_t(thintable_epi8[mask2]) + 0x0808080808080808};
+      // this is the version "nearly pruned"
+      __m128i pruned = __lsx_vshuf_b(*this, *this, shufmask);
+      // we still need to put the  pieces back together.
+      // we compute the popcount of the first words:
+      int pop1 = BitsSetTable256mul2[mask1];
+      // then load the corresponding mask
+      __m128i compactmask = __lsx_vldx(reinterpret_cast<void*>(reinterpret_cast<unsigned long>(pshufb_combine_table)), pop1 * 8);
+      __m128i answer = __lsx_vshuf_b(pruned, pruned, compactmask);
+      __lsx_vst(answer, reinterpret_cast<uint8_t*>(output), 0);
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> : base8_numeric<int8_t> {
+    simdjson_inline simd8() : base8_numeric<int8_t>() {}
+    simdjson_inline simd8(const __m128i _value) : base8_numeric<int8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const int8_t values[16]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) : simd8({
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+      }) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return __lsx_vmax_b(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return __lsx_vmin_b(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return __lsx_vslt_b(other, *this); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return __lsx_vslt_b(*this, other); }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t>: base8_numeric<uint8_t> {
+    simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+    simdjson_inline simd8(const __m128i _value) : base8_numeric<uint8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const uint8_t values[16]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) : simd8(__m128i(v16u8{
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    })) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return __lsx_vsadd_bu(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return __lsx_vssub_bu(*this, other); }
+
+    // Order-specific operations
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return __lsx_vmax_bu(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return __lsx_vmin_bu(other, *this); }
+    // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
+    // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->lt_bits(other).any_bits_set(); }
+
+    // Bit-specific operations
+    simdjson_inline simd8<bool> bits_not_set() const { return *this == uint8_t(0); }
+    simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
+    simdjson_inline bool is_ascii() const { return 0 == __lsx_vpickve2gr_w(__lsx_vmskltz_b(*this), 0); }
+    simdjson_inline bool bits_not_set_anywhere() const { return 0 == __lsx_vpickve2gr_hu(__lsx_vmsknz_b(*this), 0); }
+    simdjson_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
+    simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const {
+      return 0 == __lsx_vpickve2gr_hu(__lsx_vmsknz_b(__lsx_vand_v(*this, bits)), 0);
+    }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(__lsx_vsrli_b(*this, N)); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(__lsx_vslli_b(*this, N)); }
+  };
+
+  template<typename T>
+  struct simd8x64 {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+    static_assert(NUM_CHUNKS == 4, "LSX kernel should use four registers per 64-byte block.");
+    const simd8<T> chunks[NUM_CHUNKS];
+    template<int idx> simd8<uint8_t> get() const { return idx < NUM_CHUNKS ? chunks[idx] : simd8<T>(); }
+
+    simd8x64(const simd8x64<T>& o) = delete; // no copy allowed
+    simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1, const simd8<T> chunk2, const simd8<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+16), simd8<T>::load(ptr+32), simd8<T>::load(ptr+48)} {}
+
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
+      uint16_t mask1 = uint16_t(mask);
+      uint16_t mask2 = uint16_t(mask >> 16);
+      uint16_t mask3 = uint16_t(mask >> 32);
+      uint16_t mask4 = uint16_t(mask >> 48);
+      __m128i zcnt = __lsx_vpcnt_h(__m128i(v2u64{~mask, 0}));
+      uint64_t zcnt1 = __lsx_vpickve2gr_hu(zcnt, 0);
+      uint64_t zcnt2 = __lsx_vpickve2gr_hu(zcnt, 1);
+      uint64_t zcnt3 = __lsx_vpickve2gr_hu(zcnt, 2);
+      uint64_t zcnt4 = __lsx_vpickve2gr_hu(zcnt, 3);
+      uint8_t *voutput = reinterpret_cast<uint8_t*>(output);
+      // There should be a critical value which processes in scaler is faster.
+      if (zcnt1)
+        this->chunks[0].compress(mask1, reinterpret_cast<T*>(voutput));
+      voutput += zcnt1;
+      if (zcnt2)
+        this->chunks[1].compress(mask2, reinterpret_cast<T*>(voutput));
+      voutput += zcnt2;
+      if (zcnt3)
+        this->chunks[2].compress(mask3, reinterpret_cast<T*>(voutput));
+      voutput += zcnt3;
+      if (zcnt4)
+        this->chunks[3].compress(mask4, reinterpret_cast<T*>(voutput));
+      voutput += zcnt4;
+      return reinterpret_cast<uint64_t>(voutput) - reinterpret_cast<uint64_t>(output);
+    }
+
+    simdjson_inline void store(T ptr[64]) const {
+      this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
+      this->chunks[1].store(ptr+sizeof(simd8<T>)*1);
+      this->chunks[2].store(ptr+sizeof(simd8<T>)*2);
+      this->chunks[3].store(ptr+sizeof(simd8<T>)*3);
+    }
+
+    simdjson_inline uint64_t to_bitmask() const {
+      __m128i mask1 = __lsx_vmskltz_b(this->chunks[0]);
+      __m128i mask2 = __lsx_vmskltz_b(this->chunks[1]);
+      __m128i mask3 = __lsx_vmskltz_b(this->chunks[2]);
+      __m128i mask4 = __lsx_vmskltz_b(this->chunks[3]);
+      mask1 = __lsx_vilvl_h(mask2, mask1);
+      mask2 = __lsx_vilvl_h(mask4, mask3);
+      return __lsx_vpickve2gr_du(__lsx_vilvl_w(mask2, mask1), 0);
+    }
+
+    simdjson_inline simd8<T> reduce_or() const {
+      return (this->chunks[0] | this->chunks[1]) | (this->chunks[2] | this->chunks[3]);
+    }
+
+    simdjson_inline uint64_t eq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] == mask,
+        this->chunks[1] == mask,
+        this->chunks[2] == mask,
+        this->chunks[3] == mask
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+      return  simd8x64<bool>(
+        this->chunks[0] == other.chunks[0],
+        this->chunks[1] == other.chunks[1],
+        this->chunks[2] == other.chunks[2],
+        this->chunks[3] == other.chunks[3]
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t lteq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] <= mask,
+        this->chunks[1] <= mask,
+        this->chunks[2] <= mask,
+        this->chunks[3] <= mask
+      ).to_bitmask();
+    }
+  }; // struct simd8x64<T>
+
+} // namespace simd
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_LSX_SIMD_H
+/* end file simdjson/lsx/simd.h */
+/* including simdjson/lsx/stringparsing_defs.h: #include "simdjson/lsx/stringparsing_defs.h" */
+/* begin file simdjson/lsx/stringparsing_defs.h */
+#ifndef SIMDJSON_LSX_STRINGPARSING_DEFS_H
+#define SIMDJSON_LSX_STRINGPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/simd.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 32;
+  simdjson_inline backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return bs_bits != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+
+  uint32_t bs_bits;
+  uint32_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // this can read up to 31 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v0(src);
+  simd8<uint8_t> v1(src + sizeof(v0));
+  v0.store(dst);
+  v1.store(dst + sizeof(v0));
+
+  // Getting a 64-bit bitmask is much cheaper than multiple 16-bit bitmasks on LSX; therefore, we
+  // smash them together into a 64-byte mask and get the bitmask from there.
+  uint64_t bs_and_quote = simd8x64<bool>(v0 == '\\', v1 == '\\', v0 == '"', v1 == '"').to_bitmask();
+  return {
+    uint32_t(bs_and_quote),      // bs_bits
+    uint32_t(bs_and_quote >> 32) // quote_bits
+  };
+}
+
+
+struct escaping {
+  static constexpr uint32_t BYTES_PROCESSED = 16;
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits != 0; }
+  simdjson_inline int escape_index() { return trailing_zeroes(escape_bits); }
+
+  uint64_t escape_bits;
+}; // struct escaping
+
+
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "escaping finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  simd8<bool> is_quote = (v == '"');
+  simd8<bool> is_backslash = (v == '\\');
+  simd8<bool> is_control = (v < 32);
+  return {
+    static_cast<uint64_t>((is_backslash | is_quote | is_control).to_bitmask())
+  };
+}
+
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_LSX_STRINGPARSING_DEFS_H
+/* end file simdjson/lsx/stringparsing_defs.h */
+
+#define SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT 1
+/* end file simdjson/lsx/begin.h */
+/* including simdjson/generic/amalgamated.h for lsx: #include "simdjson/generic/amalgamated.h" */
+/* begin file simdjson/generic/amalgamated.h for lsx */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_GENERIC_DEPENDENCIES_H)
+#error simdjson/generic/dependencies.h must be included before simdjson/generic/amalgamated.h!
+#endif
+
+/* including simdjson/generic/base.h for lsx: #include "simdjson/generic/base.h" */
+/* begin file simdjson/generic/base.h for lsx */
+#ifndef SIMDJSON_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): // If we haven't got an implementation yet, we're in the editor, editing a generic file! Just */
+/* amalgamation skipped (editor-only): // use the most advanced one we can so the most possible stuff can be tested. */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation_detection.h" */
+/* amalgamation skipped (editor-only): #if SIMDJSON_IMPLEMENTATION_ICELAKE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_HASWELL */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_WESTMERE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_ARM64 */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_PPC64 */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LASX */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LSX */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_RVV_VLS */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_FALLBACK */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/begin.h" */
+/* amalgamation skipped (editor-only): #else */
+/* amalgamation skipped (editor-only): #error "All possible implementations (including fallback) have been disabled! simdjson will not run." */
+/* amalgamation skipped (editor-only): #endif */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+
+struct open_container;
+class dom_parser_implementation;
+
+/**
+ * The type of a JSON number
+ */
+enum class number_type {
+    floating_point_number=1, /// a binary64 number
+    signed_integer,          /// a signed integer that fits in a 64-bit word using two's complement
+    unsigned_integer,        /// a positive integer larger or equal to 1<<63
+    big_integer              /// a big integer that does not fit in a 64-bit word
+};
+
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_BASE_H
+/* end file simdjson/generic/base.h for lsx */
+/* including simdjson/generic/jsoncharutils.h for lsx: #include "simdjson/generic/jsoncharutils.h" */
+/* begin file simdjson/generic/jsoncharutils.h for lsx */
+#ifndef SIMDJSON_GENERIC_JSONCHARUTILS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_JSONCHARUTILS_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/jsoncharutils_tables.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace jsoncharutils {
+
+// return non-zero if not a structural or whitespace char
+// zero otherwise
+simdjson_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace_negated[c];
+}
+
+simdjson_inline uint32_t is_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace[c];
+}
+
+// returns a value with the high 16 bits set if not valid
+// otherwise returns the conversion of the 4 hex digits at src into the bottom
+// 16 bits of the 32-bit return register
+//
+// see
+// https://lemire.me/blog/2019/04/17/parsing-short-hexadecimal-strings-efficiently/
+static inline uint32_t hex_to_u32_nocheck(
+    const uint8_t *src) { // strictly speaking, static inline is a C-ism
+  uint32_t v1 = internal::digit_to_val32[630 + src[0]];
+  uint32_t v2 = internal::digit_to_val32[420 + src[1]];
+  uint32_t v3 = internal::digit_to_val32[210 + src[2]];
+  uint32_t v4 = internal::digit_to_val32[0 + src[3]];
+  return v1 | v2 | v3 | v4;
+}
+
+// given a code point cp, writes to c
+// the utf-8 code, outputting the length in
+// bytes, if the length is zero, the code point
+// is invalid
+//
+// This can possibly be made faster using pdep
+// and clz and table lookups, but JSON documents
+// have few escaped code points, and the following
+// function looks cheap.
+//
+// Note: we assume that surrogates are treated separately
+//
+simdjson_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
+  if (cp <= 0x7F) {
+    c[0] = uint8_t(cp);
+    return 1; // ascii
+  }
+  if (cp <= 0x7FF) {
+    c[0] = uint8_t((cp >> 6) + 192);
+    c[1] = uint8_t((cp & 63) + 128);
+    return 2; // universal plane
+    //  Surrogates are treated elsewhere...
+    //} //else if (0xd800 <= cp && cp <= 0xdfff) {
+    //  return 0; // surrogates // could put assert here
+  } else if (cp <= 0xFFFF) {
+    c[0] = uint8_t((cp >> 12) + 224);
+    c[1] = uint8_t(((cp >> 6) & 63) + 128);
+    c[2] = uint8_t((cp & 63) + 128);
+    return 3;
+  } else if (cp <= 0x10FFFF) { // if you know you have a valid code point, this
+                               // is not needed
+    c[0] = uint8_t((cp >> 18) + 240);
+    c[1] = uint8_t(((cp >> 12) & 63) + 128);
+    c[2] = uint8_t(((cp >> 6) & 63) + 128);
+    c[3] = uint8_t((cp & 63) + 128);
+    return 4;
+  }
+  // will return 0 when the code point was too large.
+  return 0; // bad r
+}
+
+#if SIMDJSON_IS_32BITS // _umul128 for x86, arm
+// this is a slow emulation routine for 32-bit
+//
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+  return x * (uint64_t)y;
+}
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+  uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
+  uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
+  uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
+  uint64_t adbc_carry = !!(adbc < ad);
+  uint64_t lo = bd + (adbc << 32);
+  *hi = __emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) +
+        (adbc_carry << 32) + !!(lo < bd);
+  return lo;
+}
+#endif
+
+} // namespace jsoncharutils
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_JSONCHARUTILS_H
+/* end file simdjson/generic/jsoncharutils.h for lsx */
+/* including simdjson/generic/atomparsing.h for lsx: #include "simdjson/generic/atomparsing.h" */
+/* begin file simdjson/generic/atomparsing.h for lsx */
+#ifndef SIMDJSON_GENERIC_ATOMPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_ATOMPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace lsx {
+namespace {
+/// @private
+namespace atomparsing {
+
+// The string_to_uint32 is exclusively used to map literal strings to 32-bit values.
+// We use memcpy instead of a pointer cast to avoid undefined behaviors since we cannot
+// be certain that the character pointer will be properly aligned.
+// You might think that using memcpy makes this function expensive, but you'd be wrong.
+// All decent optimizing compilers (GCC, clang, Visual Studio) will compile string_to_uint32("false");
+// to the compile-time constant 1936482662.
+simdjson_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
+
+
+// Again in str4ncmp we use a memcpy to avoid undefined behavior. The memcpy may appear expensive.
+// Yet all decent optimizing compilers will compile memcpy to a single instruction, just about.
+simdjson_warn_unused
+simdjson_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
+  uint32_t srcval; // we want to avoid unaligned 32-bit loads (undefined in C/C++)
+  static_assert(sizeof(uint32_t) <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be larger than 4 bytes");
+  std::memcpy(&srcval, src, sizeof(uint32_t));
+  return srcval ^ string_to_uint32(atom);
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src) {
+  return (str4ncmp(src, "true") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_true_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "true"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src) {
+  return (str4ncmp(src+1, "alse") | jsoncharutils::is_not_structural_or_whitespace(src[5])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
+  if (len > 5) { return is_valid_false_atom(src); }
+  else if (len == 5) { return !str4ncmp(src+1, "alse"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src) {
+  return (str4ncmp(src, "null") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_null_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "null"); }
+  else { return false; }
+}
+
+} // namespace atomparsing
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_ATOMPARSING_H
+/* end file simdjson/generic/atomparsing.h for lsx */
+/* including simdjson/generic/dom_parser_implementation.h for lsx: #include "simdjson/generic/dom_parser_implementation.h" */
+/* begin file simdjson/generic/dom_parser_implementation.h for lsx */
+#ifndef SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/dom_parser_implementation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+
+// expectation: sizeof(open_container) = 64/8.
+struct open_container {
+  uint32_t tape_index; // where, on the tape, does the scope ([,{) begins
+  uint32_t count; // how many elements in the scope
+}; // struct open_container
+
+static_assert(sizeof(open_container) == 64/8, "Open container must be 64 bits");
+
+class dom_parser_implementation final : public internal::dom_parser_implementation {
+public:
+  /** Tape location of each open { or [ */
+  std::unique_ptr<open_container[]> open_containers{};
+  /** Whether each open container is a [ or { */
+  std::unique_ptr<bool[]> is_array{};
+  /** Buffer passed to stage 1 */
+  const uint8_t *buf{};
+  /** Length passed to stage 1 */
+  size_t len{0};
+  /** Document passed to stage 2 */
+  dom::document *doc{};
+
+  inline dom_parser_implementation() noexcept;
+  inline dom_parser_implementation(dom_parser_implementation &&other) noexcept;
+  inline dom_parser_implementation &operator=(dom_parser_implementation &&other) noexcept;
+  dom_parser_implementation(const dom_parser_implementation &) = delete;
+  dom_parser_implementation &operator=(const dom_parser_implementation &) = delete;
+
+  simdjson_warn_unused error_code parse(const uint8_t *buf, size_t len, dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage1(const uint8_t *buf, size_t len, stage1_mode partial) noexcept final;
+  simdjson_warn_unused error_code stage2(dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage2_next(dom::document &doc) noexcept final;
+  simdjson_warn_unused uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) const noexcept final;
+  simdjson_warn_unused uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept final;
+  inline simdjson_warn_unused error_code set_capacity(size_t capacity) noexcept final;
+  inline simdjson_warn_unused error_code set_max_depth(size_t max_depth) noexcept final;
+private:
+  simdjson_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
+
+};
+
+} // namespace lsx
+} // namespace simdjson
+
+namespace simdjson {
+namespace lsx {
+
+inline dom_parser_implementation::dom_parser_implementation() noexcept = default;
+inline dom_parser_implementation::dom_parser_implementation(dom_parser_implementation &&other) noexcept = default;
+inline dom_parser_implementation &dom_parser_implementation::operator=(dom_parser_implementation &&other) noexcept = default;
+
+// Leaving these here so they can be inlined if so desired
+inline simdjson_warn_unused error_code dom_parser_implementation::set_capacity(size_t capacity) noexcept {
+  if(capacity > SIMDJSON_MAXSIZE_BYTES) { return CAPACITY; }
+  // Stage 1 index output
+  size_t max_structures = SIMDJSON_ROUNDUP_N(capacity, 64) + 2 + 7;
+  structural_indexes.reset( new (std::nothrow) uint32_t[max_structures] );
+  if (!structural_indexes) { _capacity = 0; return MEMALLOC; }
+  structural_indexes[0] = 0;
+  n_structural_indexes = 0;
+
+  _capacity = capacity;
+  return SUCCESS;
+}
+
+inline simdjson_warn_unused error_code dom_parser_implementation::set_max_depth(size_t max_depth) noexcept {
+  // Stage 2 stacks
+  open_containers.reset(new (std::nothrow) open_container[max_depth]);
+  is_array.reset(new (std::nothrow) bool[max_depth]);
+  if (!is_array || !open_containers) { _max_depth = 0; return MEMALLOC; }
+
+  _max_depth = max_depth;
+  return SUCCESS;
+}
+
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file simdjson/generic/dom_parser_implementation.h for lsx */
+/* including simdjson/generic/implementation_simdjson_result_base.h for lsx: #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base.h for lsx */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+
+// This is a near copy of include/error.h's implementation_simdjson_result_base, except it doesn't use std::pair
+// so we can avoid inlining errors
+// TODO reconcile these!
+/**
+ * The result of a simdjson operation that could fail.
+ *
+ * Gives the option of reading error codes, or throwing an exception by casting to the desired result.
+ *
+ * This is a base class for implementations that want to add functions to the result type for
+ * chaining.
+ *
+ * Override like:
+ *
+ *   struct simdjson_result<T> : public internal::implementation_simdjson_result_base<T> {
+ *     simdjson_result() noexcept : internal::implementation_simdjson_result_base<T>() {}
+ *     simdjson_result(error_code error) noexcept : internal::implementation_simdjson_result_base<T>(error) {}
+ *     simdjson_result(T &&value) noexcept : internal::implementation_simdjson_result_base<T>(std::forward(value)) {}
+ *     simdjson_result(T &&value, error_code error) noexcept : internal::implementation_simdjson_result_base<T>(value, error) {}
+ *     // Your extra methods here
+ *   }
+ *
+ * Then any method returning simdjson_result<T> will be chainable with your methods.
+ */
+template<typename T>
+struct implementation_simdjson_result_base {
+
+  /**
+   * Create a new empty result with error = UNINITIALIZED.
+   */
+  simdjson_inline implementation_simdjson_result_base() noexcept = default;
+
+  /**
+   * Create a new error result.
+   */
+  simdjson_inline implementation_simdjson_result_base(error_code error) noexcept;
+
+  /**
+   * Create a new successful result.
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value) noexcept;
+
+  /**
+   * Create a new result with both things (use if you don't want to branch when creating the result).
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value, error_code error) noexcept;
+
+  /**
+   * Move the value and the error to the provided variables.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   * @param error The variable to assign the error to. Set to SUCCESS if there is no error.
+   */
+  simdjson_inline void tie(T &value, error_code &error) && noexcept;
+
+  /**
+   * Move the value to the provided variable.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   */
+  simdjson_warn_unused simdjson_inline error_code get(T &value) && noexcept;
+
+  /**
+   * The error.
+   */
+  simdjson_warn_unused simdjson_inline error_code error() const noexcept;
+
+  /**
+   * Whether there is a value.
+   */
+  simdjson_warn_unused simdjson_inline bool has_value() const noexcept;
+
+#if SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T& operator*() &  noexcept(false);
+  simdjson_inline T&& operator*() &&  noexcept(false);
+  /**
+   * Arrow operator to access members of the contained value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T* operator->() noexcept(false);
+  simdjson_inline const T* operator->() const noexcept(false);
+
+  simdjson_inline T& value() & noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& value() && noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& take_value() && noexcept(false);
+
+  /**
+   * Cast to the value (will throw on error).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline operator T&&() && noexcept(false);
+
+
+#endif // SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline const T& value_unsafe() const& noexcept;
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T& value_unsafe() & noexcept;
+  /**
+   * Take the result value (move it). This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T&& value_unsafe() && noexcept;
+
+  using value_type = T;
+  using error_type = error_code;
+
+protected:
+  /** users should never directly access first and second. **/
+  T first{}; /** Users should never directly access 'first'. **/
+  error_code second{UNINITIALIZED}; /** Users should never directly access 'second'. **/
+}; // struct implementation_simdjson_result_base
+
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+/* end file simdjson/generic/implementation_simdjson_result_base.h for lsx */
+/* including simdjson/generic/numberparsing.h for lsx: #include "simdjson/generic/numberparsing.h" */
+/* begin file simdjson/generic/numberparsing.h for lsx */
+#ifndef SIMDJSON_GENERIC_NUMBERPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_NUMBERPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <limits>
+#include <ostream>
+#include <cstring>
+
+namespace simdjson {
+namespace lsx {
+namespace numberparsing {
+
+#ifdef JSON_TEST_NUMBERS
+#define INVALID_NUMBER(SRC) (found_invalid_number((SRC)), NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (found_integer((VALUE), (SRC)), (WRITER).append_s64((VALUE)))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (found_unsigned_integer((VALUE), (SRC)), (WRITER).append_u64((VALUE)))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (found_float((VALUE), (SRC)), (WRITER).append_double((VALUE)))
+#define BIGINT_NUMBER(SRC) (found_invalid_number((SRC)), BIGINT_ERROR)
+#else
+#define INVALID_NUMBER(SRC) (NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (WRITER).append_s64((VALUE))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (WRITER).append_u64((VALUE))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (WRITER).append_double((VALUE))
+#define BIGINT_NUMBER(SRC) (BIGINT_ERROR)
+#endif
+
+namespace {
+
+// Convert a mantissa, an exponent and a sign bit into an ieee64 double.
+// The real_exponent needs to be in [0, 2046] (technically real_exponent = 2047 would be acceptable).
+// The mantissa should be in [0,1<<53). The bit at index (1ULL << 52) while be zeroed.
+simdjson_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
+    double d;
+    mantissa &= ~(1ULL << 52);
+    mantissa |= real_exponent << 52;
+    mantissa |= ((static_cast<uint64_t>(negative)) << 63);
+    std::memcpy(&d, &mantissa, sizeof(d));
+    return d;
+}
+
+// Attempts to compute i * 10^(power) exactly; and if "negative" is
+// true, negate the result.
+// This function will only work in some cases, when it does not work, success is
+// set to false. This should work *most of the time* (like 99% of the time).
+// We assume that power is in the [smallest_power,
+// largest_power] interval: the caller is responsible for this check.
+simdjson_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
+  // we start with a fast path
+  // It was described in
+  // Clinger WD. How to read floating point numbers accurately.
+  // ACM SIGPLAN Notices. 1990
+#ifndef FLT_EVAL_METHOD
+#error "FLT_EVAL_METHOD should be defined, please include cfloat."
+#endif
+#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
+  // We cannot be certain that x/y is rounded to nearest.
+  if (0 <= power && power <= 22 && i <= 9007199254740991)
+#else
+  if (-22 <= power && power <= 22 && i <= 9007199254740991)
+#endif
+  {
+    // convert the integer into a double. This is lossless since
+    // 0 <= i <= 2^53 - 1.
+    d = double(i);
+    //
+    // The general idea is as follows.
+    // If 0 <= s < 2^53 and if 10^0 <= p <= 10^22 then
+    // 1) Both s and p can be represented exactly as 64-bit floating-point
+    // values
+    // (binary64).
+    // 2) Because s and p can be represented exactly as floating-point values,
+    // then s * p
+    // and s / p will produce correctly rounded values.
+    //
+    if (power < 0) {
+      d = d / simdjson::internal::power_of_ten[-power];
+    } else {
+      d = d * simdjson::internal::power_of_ten[power];
+    }
+    if (negative) {
+      d = -d;
+    }
+    return true;
+  }
+  // When 22 < power && power <  22 + 16, we could
+  // hope for another, secondary fast path.  It was
+  // described by David M. Gay in  "Correctly rounded
+  // binary-decimal and decimal-binary conversions." (1990)
+  // If you need to compute i * 10^(22 + x) for x < 16,
+  // first compute i * 10^x, if you know that result is exact
+  // (e.g., when i * 10^x < 2^53),
+  // then you can still proceed and do (i * 10^x) * 10^22.
+  // Is this worth your time?
+  // You need  22 < power *and* power <  22 + 16 *and* (i * 10^(x-22) < 2^53)
+  // for this second fast path to work.
+  // If you you have 22 < power *and* power <  22 + 16, and then you
+  // optimistically compute "i * 10^(x-22)", there is still a chance that you
+  // have wasted your time if i * 10^(x-22) >= 2^53. It makes the use cases of
+  // this optimization maybe less common than we would like. Source:
+  // http://www.exploringbinary.com/fast-path-decimal-to-floating-point-conversion/
+  // also used in RapidJSON: https://rapidjson.org/strtod_8h_source.html
+
+  // The fast path has now failed, so we are failing back on the slower path.
+
+  // In the slow path, we need to adjust i so that it is > 1<<63 which is always
+  // possible, except if i == 0, so we handle i == 0 separately.
+  if(i == 0) {
+    d = negative ? -0.0 : 0.0;
+    return true;
+  }
+
+
+  // The exponent is 1024 + 63 + power
+  //     + floor(log(5**power)/log(2)).
+  // The 1024 comes from the ieee64 standard.
+  // The 63 comes from the fact that we use a 64-bit word.
+  //
+  // Computing floor(log(5**power)/log(2)) could be
+  // slow. Instead we use a fast function.
+  //
+  // For power in (-400,350), we have that
+  // (((152170 + 65536) * power ) >> 16);
+  // is equal to
+  //  floor(log(5**power)/log(2)) + power when power >= 0
+  // and it is equal to
+  //  ceil(log(5**-power)/log(2)) + power when power < 0
+  //
+  // The 65536 is (1<<16) and corresponds to
+  // (65536 * power) >> 16 ---> power
+  //
+  // ((152170 * power ) >> 16) is equal to
+  // floor(log(5**power)/log(2))
+  //
+  // Note that this is not magic: 152170/(1<<16) is
+  // approximately equal to log(5)/log(2).
+  // The 1<<16 value is a power of two; we could use a
+  // larger power of 2 if we wanted to.
+  //
+  int64_t exponent = (((152170 + 65536) * power) >> 16) + 1024 + 63;
+
+
+  // We want the most significant bit of i to be 1. Shift if needed.
+  int lz = leading_zeroes(i);
+  i <<= lz;
+
+
+  // We are going to need to do some 64-bit arithmetic to get a precise product.
+  // We use a table lookup approach.
+  // It is safe because
+  // power >= smallest_power
+  // and power <= largest_power
+  // We recover the mantissa of the power, it has a leading 1. It is always
+  // rounded down.
+  //
+  // We want the most significant 64 bits of the product. We know
+  // this will be non-zero because the most significant bit of i is
+  // 1.
+  const uint32_t index = 2 * uint32_t(power - simdjson::internal::smallest_power);
+  // Optimization: It may be that materializing the index as a variable might confuse some compilers and prevent effective complex-addressing loads. (Done for code clarity.)
+  //
+  // The full_multiplication function computes the 128-bit product of two 64-bit words
+  // with a returned value of type value128 with a "low component" corresponding to the
+  // 64-bit least significant bits of the product and with a "high component" corresponding
+  // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index]);
+#else
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index]);
+#endif
+
+  // Both i and power_of_five_128[index] have their most significant bit set to 1 which
+  // implies that the either the most or the second most significant bit of the product
+  // is 1. We pack values in this manner for efficiency reasons: it maximizes the use
+  // we make of the product. It also makes it easy to reason about the product: there
+  // is 0 or 1 leading zero in the product.
+
+  // Unless the least significant 9 bits of the high (64-bit) part of the full
+  // product are all 1s, then we know that the most significant 55 bits are
+  // exact and no further work is needed. Having 55 bits is necessary because
+  // we need 53 bits for the mantissa but we have to have one rounding bit and
+  // we can waste a bit if the most significant bit of the product is zero.
+  if((firstproduct.high & 0x1FF) == 0x1FF) {
+    // We want to compute i * 5^q, but only care about the top 55 bits at most.
+    // Consider the scenario where q>=0. Then 5^q may not fit in 64-bits. Doing
+    // the full computation is wasteful. So we do what is called a "truncated
+    // multiplication".
+    // We take the most significant 64-bits, and we put them in
+    // power_of_five_128[index]. Usually, that's good enough to approximate i * 5^q
+    // to the desired approximation using one multiplication. Sometimes it does not suffice.
+    // Then we store the next most significant 64 bits in power_of_five_128[index + 1], and
+    // then we get a better approximation to i * 5^q.
+    //
+    // That's for when q>=0. The logic for q<0 is somewhat similar but it is somewhat
+    // more complicated.
+    //
+    // There is an extra layer of complexity in that we need more than 55 bits of
+    // accuracy in the round-to-even scenario.
+    //
+    // The full_multiplication function computes the 128-bit product of two 64-bit words
+    // with a returned value of type value128 with a "low component" corresponding to the
+    // 64-bit least significant bits of the product and with a "high component" corresponding
+    // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index + 1]);
+#else
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index + 1]);
+#endif
+    firstproduct.low += secondproduct.high;
+    if(secondproduct.high > firstproduct.low) { firstproduct.high++; }
+    // As it has been proven by Noble Mushtak and Daniel Lemire in "Fast Number Parsing Without
+    // Fallback" (https://arxiv.org/abs/2212.06644), at this point we are sure that the product
+    // is sufficiently accurate, and more computation is not needed.
+  }
+  uint64_t lower = firstproduct.low;
+  uint64_t upper = firstproduct.high;
+  // The final mantissa should be 53 bits with a leading 1.
+  // We shift it so that it occupies 54 bits with a leading 1.
+  ///////
+  uint64_t upperbit = upper >> 63;
+  uint64_t mantissa = upper >> (upperbit + 9);
+  lz += int(1 ^ upperbit);
+
+  // Here we have mantissa < (1<<54).
+  int64_t real_exponent = exponent - lz;
+  if (simdjson_unlikely(real_exponent <= 0)) { // we have a subnormal?
+    // Here have that real_exponent <= 0 so -real_exponent >= 0
+    if(-real_exponent + 1 >= 64) { // if we have more than 64 bits below the minimum exponent, you have a zero for sure.
+      d = negative ? -0.0 : 0.0;
+      return true;
+    }
+    // next line is safe because -real_exponent + 1 < 0
+    mantissa >>= -real_exponent + 1;
+    // Thankfully, we can't have both "round-to-even" and subnormals because
+    // "round-to-even" only occurs for powers close to 0.
+    mantissa += (mantissa & 1); // round up
+    mantissa >>= 1;
+    // There is a weird scenario where we don't have a subnormal but just.
+    // Suppose we start with 2.2250738585072013e-308, we end up
+    // with 0x3fffffffffffff x 2^-1023-53 which is technically subnormal
+    // whereas 0x40000000000000 x 2^-1023-53  is normal. Now, we need to round
+    // up 0x3fffffffffffff x 2^-1023-53  and once we do, we are no longer
+    // subnormal, but we can only know this after rounding.
+    // So we only declare a subnormal if we are smaller than the threshold.
+    real_exponent = (mantissa < (uint64_t(1) << 52)) ? 0 : 1;
+    d = to_double(mantissa, real_exponent, negative);
+    return true;
+  }
+  // We have to round to even. The "to even" part
+  // is only a problem when we are right in between two floats
+  // which we guard against.
+  // If we have lots of trailing zeros, we may fall right between two
+  // floating-point values.
+  //
+  // The round-to-even cases take the form of a number 2m+1 which is in (2^53,2^54]
+  // times a power of two. That is, it is right between a number with binary significand
+  // m and another number with binary significand m+1; and it must be the case
+  // that it cannot be represented by a float itself.
+  //
+  // We must have that w * 10 ^q == (2m+1) * 2^p for some power of two 2^p.
+  // Recall that 10^q = 5^q * 2^q.
+  // When q >= 0, we must have that (2m+1) is divible by 5^q, so 5^q <= 2^54. We have that
+  //  5^23 <=  2^54 and it is the last power of five to qualify, so q <= 23.
+  // When q<0, we have  w  >=  (2m+1) x 5^{-q}.  We must have that w<2^{64} so
+  // (2m+1) x 5^{-q} < 2^{64}. We have that 2m+1>2^{53}. Hence, we must have
+  // 2^{53} x 5^{-q} < 2^{64}.
+  // Hence we have 5^{-q} < 2^{11}$ or q>= -4.
+  //
+  // We require lower <= 1 and not lower == 0 because we could not prove that
+  // that lower == 0 is implied; but we could prove that lower <= 1 is a necessary and sufficient test.
+  if (simdjson_unlikely((lower <= 1) && (power >= -4) && (power <= 23) && ((mantissa & 3) == 1))) {
+    if((mantissa  << (upperbit + 64 - 53 - 2)) ==  upper) {
+      mantissa &= ~1;             // flip it so that we do not round up
+    }
+  }
+
+  mantissa += mantissa & 1;
+  mantissa >>= 1;
+
+  // Here we have mantissa < (1<<53), unless there was an overflow
+  if (mantissa >= (1ULL << 53)) {
+    //////////
+    // This will happen when parsing values such as 7.2057594037927933e+16
+    ////////
+    mantissa = (1ULL << 52);
+    real_exponent++;
+  }
+  mantissa &= ~(1ULL << 52);
+  // we have to check that real_exponent is in range, otherwise we bail out
+  if (simdjson_unlikely(real_exponent > 2046)) {
+    // We have an infinite value!!! We could actually throw an error here if we could.
+    return false;
+  }
+  d = to_double(mantissa, real_exponent, negative);
+  return true;
+}
+
+// We call a fallback floating-point parser that might be slow. Note
+// it will accept JSON numbers, but the JSON spec. is more restrictive so
+// before you call parse_float_fallback, you need to have validated the input
+// string with the JSON grammar.
+// It will return an error (false) if the parsed number is infinite.
+// The string parsing itself always succeeds. We know that there is at least
+// one digit.
+static bool parse_float_fallback(const uint8_t *ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+static bool parse_float_fallback(const uint8_t *ptr, const uint8_t *end_ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr), reinterpret_cast<const char *>(end_ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+// check quickly whether the next 8 chars are made of digits
+// at a glance, it looks better than Mula's
+// http://0x80.pl/articles/swar-digits-validate.html
+simdjson_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
+  uint64_t val;
+  // this can read up to 7 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(7 <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be bigger than 7");
+  std::memcpy(&val, chars, 8);
+  // a branchy method might be faster:
+  // return (( val & 0xF0F0F0F0F0F0F0F0 ) == 0x3030303030303030)
+  //  && (( (val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0 ) ==
+  //  0x3030303030303030);
+  return (((val & 0xF0F0F0F0F0F0F0F0) |
+           (((val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0) >> 4)) ==
+          0x3333333333333333);
+}
+
+template<typename I>
+SIMDJSON_NO_SANITIZE_UNDEFINED // We deliberately allow overflow here and check later
+simdjson_inline bool parse_digit(const uint8_t c, I &i) {
+  const uint8_t digit = static_cast<uint8_t>(c - '0');
+  if (digit > 9) {
+    return false;
+  }
+  // PERF NOTE: multiplication by 10 is cheaper than arbitrary integer multiplication
+  i = 10 * i + digit; // might overflow, we will handle the overflow later
+  return true;
+}
+
+simdjson_inline bool is_digit(const uint8_t c) {
+  return static_cast<uint8_t>(c - '0') <= 9;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_decimal_after_separator(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
+  // we continue with the fiction that we have an integer. If the
+  // floating point number is representable as x * 10^z for some integer
+  // z that fits in 53 bits, then we will be able to convert back the
+  // the integer into a float in a lossless manner.
+  const uint8_t *const first_after_period = p;
+
+#ifdef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_SWAR_NUMBER_PARSING
+  // this helps if we have lots of decimals!
+  // this turns out to be frequent enough.
+  if (is_made_of_eight_digits_fast(p)) {
+    i = i * 100000000 + parse_eight_digits_unrolled(p);
+    p += 8;
+  }
+#endif // SIMDJSON_SWAR_NUMBER_PARSING
+#endif // #ifdef SIMDJSON_SWAR_NUMBER_PARSING
+  // Unrolling the first digit makes a small difference on some implementations (e.g. westmere)
+  if (parse_digit(*p, i)) { ++p; }
+  while (parse_digit(*p, i)) { p++; }
+  exponent = first_after_period - p;
+  // Decimal without digits (123.) is illegal
+  if (exponent == 0) {
+    return INVALID_NUMBER(src);
+  }
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
+  // Exp Sign: -123.456e[-]78
+  bool neg_exp = ('-' == *p);
+  if (neg_exp || '+' == *p) { p++; } // Skip + as well
+
+  // Exponent: -123.456e-[78]
+  auto start_exp = p;
+  int64_t exp_number = 0;
+  while (parse_digit(*p, exp_number)) { ++p; }
+  // It is possible for parse_digit to overflow.
+  // In particular, it could overflow to INT64_MIN, and we cannot do - INT64_MIN.
+  // Thus we *must* check for possible overflow before we negate exp_number.
+
+  // Performance notes: it may seem like combining the two "simdjson_unlikely checks" below into
+  // a single simdjson_unlikely path would be faster. The reasoning is sound, but the compiler may
+  // not oblige and may, in fact, generate two distinct paths in any case. It might be
+  // possible to do uint64_t(p - start_exp - 1) >= 18 but it could end up trading off
+  // instructions for a simdjson_likely branch, an unconclusive gain.
+
+  // If there were no digits, it's an error.
+  if (simdjson_unlikely(p == start_exp)) {
+    return INVALID_NUMBER(src);
+  }
+  // We have a valid positive exponent in exp_number at this point, except that
+  // it may have overflowed.
+
+  // If there were more than 18 digits, we may have overflowed the integer. We have to do
+  // something!!!!
+  if (simdjson_unlikely(p > start_exp+18)) {
+    // Skip leading zeroes: 1e000000000000000000001 is technically valid and does not overflow
+    while (*start_exp == '0') { start_exp++; }
+    // 19 digits could overflow int64_t and is kind of absurd anyway. We don't
+    // support exponents smaller than -999,999,999,999,999,999 and bigger
+    // than 999,999,999,999,999,999.
+    // We can truncate.
+    // Note that 999999999999999999 is assuredly too large. The maximal ieee64 value before
+    // infinity is ~1.8e308. The smallest subnormal is ~5e-324. So, actually, we could
+    // truncate at 324.
+    // Note that there is no reason to fail per se at this point in time.
+    // E.g., 0e999999999999999999999 is a fine number.
+    if (p > start_exp+18) { exp_number = 999999999999999999; }
+  }
+  // At this point, we know that exp_number is a sane, positive, signed integer.
+  // It is <= 999,999,999,999,999,999. As long as 'exponent' is in
+  // [-8223372036854775808, 8223372036854775808], we won't overflow. Because 'exponent'
+  // is bounded in magnitude by the size of the JSON input, we are fine in this universe.
+  // To sum it up: the next line should never overflow.
+  exponent += (neg_exp ? -exp_number : exp_number);
+  return SUCCESS;
+}
+
+simdjson_inline bool check_if_integer(const uint8_t *const src, size_t max_length) {
+  const uint8_t *const srcend = src + max_length;
+  bool negative = (*src == '-'); // we can always read at least one character after the '-'
+  const uint8_t *p = src + uint8_t(negative);
+  if(p == srcend) { return false; }
+  if(*p == '0') {
+    ++p;
+    if(p == srcend) { return true; }
+    if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+    return true;
+  }
+  while(p != srcend && is_digit(*p)) { ++p; }
+  if(p == srcend) { return true; }
+  if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+  return true;
+}
+
+simdjson_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
+  // It is possible that the integer had an overflow.
+  // We have to handle the case where we have 0.0000somenumber.
+  const uint8_t *start = start_digits;
+  while ((*start == '0') || (*start == '.')) { ++start; }
+  // we over-decrement by one when there is a '.'
+  return digit_count - size_t(start - start_digits);
+}
+
+} // unnamed namespace
+
+/** @private */
+static error_code slow_float_parsing(simdjson_unused const uint8_t * src, double* answer) {
+  if (parse_float_fallback(src, answer)) {
+    return SUCCESS;
+  }
+  return INVALID_NUMBER(src);
+}
+
+/** @private */
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
+  // If we frequently had to deal with long strings of digits,
+  // we could extend our code by using a 128-bit integer instead
+  // of a 64-bit integer. However, this is uncommon in practice.
+  //
+  // 9999999999999999999 < 2**64 so we can accommodate 19 digits.
+  // If we have a decimal separator, then digit_count - 1 is the number of digits, but we
+  // may not have a decimal separator!
+  if (simdjson_unlikely(digit_count > 19 && significant_digits(start_digits, digit_count) > 19)) {
+    // Ok, chances are good that we had an overflow!
+    // this is almost never going to get called!!!
+    // we start anew, going slowly!!!
+    // This will happen in the following examples:
+    // 10000000000000000000000000000000000000000000e+308
+    // 3.1415926535897932384626433832795028841971693993751
+    //
+    // NOTE: We do not pass a reference to the to slow_float_parsing. If we passed our writer
+    // reference to it, it would force it to be stored in memory, preventing the compiler from
+    // picking it apart and putting into registers. i.e. if we pass it as reference,
+    // it gets slow.
+    double d;
+    error_code error = slow_float_parsing(src, &d);
+    writer.append_double(d);
+    return error;
+  }
+  // NOTE: it's weird that the simdjson_unlikely() only wraps half the if, but it seems to get slower any other
+  // way we've tried: https://github.com/simdjson/simdjson/pull/990#discussion_r448497331
+  // To future reader: we'd love if someone found a better way, or at least could explain this result!
+  if (simdjson_unlikely(exponent < simdjson::internal::smallest_power) || (exponent > simdjson::internal::largest_power)) {
+    //
+    // Important: smallest_power is such that it leads to a zero value.
+    // Observe that 18446744073709551615e-343 == 0, i.e. (2**64 - 1) e -343 is zero
+    // so something x 10^-343 goes to zero, but not so with  something x 10^-342.
+    static_assert(simdjson::internal::smallest_power <= -342, "smallest_power is not small enough");
+    //
+    if((exponent < simdjson::internal::smallest_power) || (i == 0)) {
+      // E.g. Parse "-0.0e-999" into the same value as "-0.0". See https://en.wikipedia.org/wiki/Signed_zero
+      WRITE_DOUBLE(negative ? -0.0 : 0.0, src, writer);
+      return SUCCESS;
+    } else { // (exponent > largest_power) and (i != 0)
+      // We have, for sure, an infinite value and simdjson refuses to parse infinite values.
+      return INVALID_NUMBER(src);
+    }
+  }
+  double d;
+  if (!compute_float_64(exponent, i, negative, d)) {
+    // we are almost never going to get here.
+    if (!parse_float_fallback(src, &d)) { return INVALID_NUMBER(src); }
+  }
+  WRITE_DOUBLE(d, src, writer);
+  return SUCCESS;
+}
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer);
+
+// for performance analysis, it is sometimes  useful to skip parsing
+#ifdef SIMDJSON_SKIPNUMBERPARSING
+
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const, W &writer) {
+  writer.append_s64(0);        // always write zero
+  return SUCCESS;              // always succeeds
+}
+
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept { return number_type::signed_integer; }
+#else
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer) {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  if (digit_count == 0 || ('0' == *start_digits && digit_count > 1)) { return INVALID_NUMBER(src); }
+
+  //
+  // Handle floats if there is a . or e (or both)
+  //
+  int64_t exponent = 0;
+  bool is_float = false;
+  if ('.' == *p) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_decimal_after_separator(src, p, i, exponent) );
+    digit_count = int(p - start_digits); // used later to guard against overflows
+  }
+  if (('e' == *p) || ('E' == *p)) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_exponent(src, p, exponent) );
+  }
+  if (is_float) {
+    const bool dirty_end = jsoncharutils::is_not_structural_or_whitespace(*p);
+    SIMDJSON_TRY( write_float(src, negative, i, start_digits, digit_count, exponent, writer) );
+    if (dirty_end) { return INVALID_NUMBER(src); }
+    return SUCCESS;
+  }
+
+  // The longest negative 64-bit number is 19 digits.
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  size_t longest_digit_count = negative ? 19 : 20;
+  if (digit_count > longest_digit_count) { return BIGINT_NUMBER(src); }
+  if (digit_count == longest_digit_count) {
+    if (negative) {
+      // Anything negative above INT64_MAX+1 is invalid
+      if (i > uint64_t(INT64_MAX)+1) { return BIGINT_NUMBER(src);  }
+      WRITE_INTEGER(~i+1, src, writer);
+      if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+      return SUCCESS;
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    }  else if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INVALID_NUMBER(src); }
+  }
+
+  // Write unsigned if it does not fit in a signed integer.
+  if (i > uint64_t(INT64_MAX)) {
+    WRITE_UNSIGNED(i, src, writer);
+  } else {
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+    if(i == 0 && negative) {
+      // We have to write -0.0 instead of 0
+      WRITE_DOUBLE(-0.0, src, writer);
+    } else {
+      WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+    }
+#else
+  WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+#endif
+  }
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+  return SUCCESS;
+}
+
+// Inlineable functions
+namespace {
+
+// This table can be used to characterize the final character of an integer
+// string. For JSON structural character and allowable white space characters,
+// we return SUCCESS. For 'e', '.' and 'E', we return INCORRECT_TYPE. Otherwise
+// we return NUMBER_ERROR.
+// Optimization note: we could easily reduce the size of the table by half (to 128)
+// at the cost of an extra branch.
+// Optimization note: we want the values to use at most 8 bits (not, e.g., 32 bits):
+static_assert(error_code(uint8_t(NUMBER_ERROR))== NUMBER_ERROR, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(SUCCESS))== SUCCESS, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(INCORRECT_TYPE))== INCORRECT_TYPE, "bad NUMBER_ERROR cast");
+
+const uint8_t integer_string_finisher[256] = {
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   SUCCESS,      NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, INCORRECT_TYPE,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, SUCCESS,        NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR};
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if ((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
+  const uint8_t *p = src + 1;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    // Note: we use src[1] and not src[0] because src[0] is the quote character in this
+    // instance.
+    if (src[1] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  //
+  // Check for minus sign
+  //
+  if(src == src_end) { return NUMBER_ERROR; }
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = src;
+  uint64_t i = 0;
+  while (parse_digit(*src, i)) { src++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(src - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*src)) {
+  //  return (*src == '.' || *src == 'e' || *src == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(*src != '"') { return NUMBER_ERROR; }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept {
+  return (*src == '-');
+}
+
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) { return true; }
+  return false;
+}
+
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  size_t digit_count = size_t(p - src);
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) {
+    static const uint8_t * smaller_big_integer = reinterpret_cast<const uint8_t *>("9223372036854775808");
+    // We have an integer.
+    if(simdjson_unlikely(digit_count > 20)) {
+      return number_type::big_integer;
+    }
+    // If the number is negative and valid, it must be a signed integer.
+    if(negative) {
+      if (simdjson_unlikely(digit_count > 19)) return number_type::big_integer;
+      if (simdjson_unlikely(digit_count == 19 && memcmp(src, smaller_big_integer, 19) > 0)) {
+        return number_type::big_integer;
+      }
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+      if(digit_count == 1 && src[0] == '0') {
+        // We have to write -0.0 instead of 0
+        return number_type::floating_point_number;
+      }
+#endif
+      return number_type::signed_integer;
+    }
+    // Let us check if we have a big integer (>=2**64).
+    static const uint8_t * two_to_sixtyfour = reinterpret_cast<const uint8_t *>("18446744073709551616");
+    if((digit_count > 20) || (digit_count == 20 && memcmp(src, two_to_sixtyfour, 20) >= 0)) {
+      return number_type::big_integer;
+    }
+    // The number is positive and smaller than 18446744073709551616 (or 2**64).
+    // We want values larger or equal to 9223372036854775808 to be unsigned
+    // integers, and the other values to be signed integers.
+    if((digit_count == 20) || (digit_count >= 19 && memcmp(src, smaller_big_integer, 19) >= 0)) {
+      return number_type::unsigned_integer;
+    }
+    return number_type::signed_integer;
+  }
+  // Hopefully, we have 'e' or 'E' or '.'.
+  return number_type::floating_point_number;
+}
+
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
+  if(src == src_end) { return NUMBER_ERROR; }
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  if(p == src_end) { return NUMBER_ERROR; }
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely((p != src_end) && (*p == '.'))) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if ((p == src_end) || !parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while ((p != src_end) && parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = start_digits-src > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if ((p != src_end) && (*p == 'e' || *p == 'E')) {
+    p++;
+    if(p == src_end) { return NUMBER_ERROR; }
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while ((p != src_end) && parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if ((p != src_end) && jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), src_end, &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+} // unnamed namespace
+#endif // SIMDJSON_SKIPNUMBERPARSING
+
+} // namespace numberparsing
+
+inline std::ostream& operator<<(std::ostream& out, number_type type) noexcept {
+    switch (type) {
+        case number_type::signed_integer: out << "integer in [-9223372036854775808,9223372036854775808)"; break;
+        case number_type::unsigned_integer: out << "unsigned integer in [9223372036854775808,18446744073709551616)"; break;
+        case number_type::floating_point_number: out << "floating-point number (binary64)"; break;
+        case number_type::big_integer: out << "big integer"; break;
+        default: SIMDJSON_UNREACHABLE();
+    }
+    return out;
+}
+
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_NUMBERPARSING_H
+/* end file simdjson/generic/numberparsing.h for lsx */
+
+/* including simdjson/generic/implementation_simdjson_result_base-inl.h for lsx: #include "simdjson/generic/implementation_simdjson_result_base-inl.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base-inl.h for lsx */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+
+//
+// internal::implementation_simdjson_result_base<T> inline implementation
+//
+
+template<typename T>
+simdjson_inline void implementation_simdjson_result_base<T>::tie(T &value, error_code &error) && noexcept {
+  error = this->second;
+  if (!error) {
+    value = std::forward<implementation_simdjson_result_base<T>>(*this).first;
+  }
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::get(T &value) && noexcept {
+  error_code error;
+  std::forward<implementation_simdjson_result_base<T>>(*this).tie(value, error);
+  return error;
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::error() const noexcept {
+  return this->second;
+}
+
+
+template<typename T>
+simdjson_warn_unused simdjson_inline bool implementation_simdjson_result_base<T>::has_value() const noexcept {
+  return this->error() == SUCCESS;
+}
+
+#if SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::operator*() &  noexcept(false) {
+  return this->value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::operator*() &&  noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).value();
+}
+
+template<typename T>
+simdjson_inline T* implementation_simdjson_result_base<T>::operator->() noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+
+template<typename T>
+simdjson_inline const T* implementation_simdjson_result_base<T>::operator->() const noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value() & noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::take_value() && noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::operator T&&() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+#endif // SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline const T& implementation_simdjson_result_base<T>::value_unsafe() const& noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value_unsafe() & noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value_unsafe() && noexcept {
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value, error_code error) noexcept
+    : first{std::forward<T>(value)}, second{error} {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(error_code error) noexcept
+    : implementation_simdjson_result_base(T{}, error) {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value) noexcept
+    : implementation_simdjson_result_base(std::forward<T>(value), SUCCESS) {}
+
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+/* end file simdjson/generic/implementation_simdjson_result_base-inl.h for lsx */
+/* end file simdjson/generic/amalgamated.h for lsx */
+/* including simdjson/lsx/end.h: #include "simdjson/lsx/end.h" */
+/* begin file simdjson/lsx/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#undef SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+/* undefining SIMDJSON_IMPLEMENTATION from "lsx" */
+#undef SIMDJSON_IMPLEMENTATION
+/* end file simdjson/lsx/end.h */
+
+#endif // SIMDJSON_LSX_H
+/* end file simdjson/lsx.h */
+/* including simdjson/lsx/implementation.h: #include <simdjson/lsx/implementation.h> */
+/* begin file simdjson/lsx/implementation.h */
+#ifndef SIMDJSON_LSX_IMPLEMENTATION_H
+#define SIMDJSON_LSX_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/instruction_set.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation("lsx", "LoongArch SX", internal::instruction_set::LSX) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_LSX_IMPLEMENTATION_H
+/* end file simdjson/lsx/implementation.h */
+
+/* including simdjson/lsx/begin.h: #include <simdjson/lsx/begin.h> */
+/* begin file simdjson/lsx/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "lsx" */
+#define SIMDJSON_IMPLEMENTATION lsx
+/* including simdjson/lsx/base.h: #include "simdjson/lsx/base.h" */
+/* begin file simdjson/lsx/base.h */
+#ifndef SIMDJSON_LSX_BASE_H
+#define SIMDJSON_LSX_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+/**
+ * Implementation for LSX.
+ */
+namespace lsx {
+
+class implementation;
+
+namespace {
+namespace simd {
+template <typename T> struct simd8;
+template <typename T> struct simd8x64;
+} // namespace simd
+} // unnamed namespace
+
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_LSX_BASE_H
+/* end file simdjson/lsx/base.h */
+/* including simdjson/lsx/intrinsics.h: #include "simdjson/lsx/intrinsics.h" */
+/* begin file simdjson/lsx/intrinsics.h */
+#ifndef SIMDJSON_LSX_INTRINSICS_H
+#define SIMDJSON_LSX_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <lsxintrin.h>
+
+static_assert(sizeof(__m128i) <= simdjson::SIMDJSON_PADDING, "insufficient padding for LoongArch SX");
+
+#endif //  SIMDJSON_LSX_INTRINSICS_H
+/* end file simdjson/lsx/intrinsics.h */
+/* including simdjson/lsx/bitmanipulation.h: #include "simdjson/lsx/bitmanipulation.h" */
+/* begin file simdjson/lsx/bitmanipulation.h */
+#ifndef SIMDJSON_LSX_BITMANIPULATION_H
+#define SIMDJSON_LSX_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/bitmask.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+  return __builtin_ctzll(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return input_num & (input_num-1);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+  return __builtin_clzll(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int count_ones(uint64_t input_num) {
+  return __lsx_vpickve2gr_w(__lsx_vpcnt_d(__m128i(v2u64{input_num, 0})), 0);
+}
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2, uint64_t *result) {
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+}
+
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_LSX_BITMANIPULATION_H
+/* end file simdjson/lsx/bitmanipulation.h */
+/* including simdjson/lsx/bitmask.h: #include "simdjson/lsx/bitmask.h" */
+/* begin file simdjson/lsx/bitmask.h */
+#ifndef SIMDJSON_LSX_BITMASK_H
+#define SIMDJSON_LSX_BITMASK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(uint64_t bitmask) {
+  bitmask ^= bitmask << 1;
+  bitmask ^= bitmask << 2;
+  bitmask ^= bitmask << 4;
+  bitmask ^= bitmask << 8;
+  bitmask ^= bitmask << 16;
+  bitmask ^= bitmask << 32;
+  return bitmask;
+}
+
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif
+/* end file simdjson/lsx/bitmask.h */
+/* including simdjson/lsx/numberparsing_defs.h: #include "simdjson/lsx/numberparsing_defs.h" */
+/* begin file simdjson/lsx/numberparsing_defs.h */
+#ifndef SIMDJSON_LSX_NUMBERPARSING_DEFS_H
+#define SIMDJSON_LSX_NUMBERPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/intrinsics.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace lsx {
+namespace numberparsing {
+
+// we don't have appropriate instructions, so let us use a scalar function
+// credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  uint64_t val;
+  std::memcpy(&val, chars, sizeof(uint64_t));
+  val = (val & 0x0F0F0F0F0F0F0F0F) * 2561 >> 8;
+  val = (val & 0x00FF00FF00FF00FF) * 6553601 >> 16;
+  return uint32_t((val & 0x0000FFFF0000FFFF) * 42949672960001 >> 32);
+}
+
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace lsx
+} // namespace simdjson
+
+#ifndef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_IS_BIG_ENDIAN
+#define SIMDJSON_SWAR_NUMBER_PARSING 0
+#else
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+#endif
+#endif
+
+#endif // SIMDJSON_LSX_NUMBERPARSING_DEFS_H
+/* end file simdjson/lsx/numberparsing_defs.h */
+/* including simdjson/lsx/simd.h: #include "simdjson/lsx/simd.h" */
+/* begin file simdjson/lsx/simd.h */
+#ifndef SIMDJSON_LSX_SIMD_H
+#define SIMDJSON_LSX_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace simd {
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename Child>
+  struct base {
+    __m128i value;
+
+    // Zero constructor
+    simdjson_inline base() : value{__m128i()} {}
+
+    // Conversion from SIMD register
+    simdjson_inline base(const __m128i _value) : value(_value) {}
+
+    // Conversion to SIMD register
+    simdjson_inline operator const __m128i&() const { return this->value; }
+    simdjson_inline operator __m128i&() { return this->value; }
+    simdjson_inline operator const v16i8&() const { return (v16i8&)this->value; }
+    simdjson_inline operator v16i8&() { return (v16i8&)this->value; }
+
+    // Bit operations
+    simdjson_inline Child operator|(const Child other) const { return __lsx_vor_v(*this, other); }
+    simdjson_inline Child operator&(const Child other) const { return __lsx_vand_v(*this, other); }
+    simdjson_inline Child operator^(const Child other) const { return __lsx_vxor_v(*this, other); }
+    simdjson_inline Child bit_andnot(const Child other) const { return __lsx_vandn_v(other, *this); }
+    simdjson_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+  };
+
+  // Forward-declared so they can be used by splat and friends.
+  template<typename T>
+  struct simd8;
+
+  template<typename T, typename Mask=simd8<bool>>
+  struct base8: base<simd8<T>> {
+    simdjson_inline base8() : base<simd8<T>>() {}
+    simdjson_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}
+
+    friend simdjson_really_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return __lsx_vseq_b(lhs, rhs); }
+
+    static const int SIZE = sizeof(base<simd8<T>>::value);
+
+    template<int N=1>
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+      return __lsx_vor_v(__lsx_vbsll_v(*this, N), __lsx_vbsrl_v(prev_chunk, 16 - N));
+    }
+  };
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool>: base8<bool> {
+    static simdjson_inline simd8<bool> splat(bool _value) {
+      return __lsx_vreplgr2vr_b(uint8_t(-(!!_value)));
+    }
+
+    simdjson_inline simd8() : base8() {}
+    simdjson_inline simd8(const __m128i _value) : base8<bool>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+
+    simdjson_inline int to_bitmask() const { return __lsx_vpickve2gr_w(__lsx_vmskltz_b(*this), 0); }
+    simdjson_inline bool any() const { return 0 == __lsx_vpickve2gr_hu(__lsx_vmsknz_b(*this), 0); }
+    simdjson_inline simd8<bool> operator~() const { return *this ^ true; }
+  };
+
+  template<typename T>
+  struct base8_numeric: base8<T> {
+    static simdjson_inline simd8<T> splat(T _value) { return __lsx_vreplgr2vr_b(_value); }
+    static simdjson_inline simd8<T> zero() { return __lsx_vldi(0); }
+    static simdjson_inline simd8<T> load(const T values[16]) {
+      return __lsx_vld(reinterpret_cast<const __m128i *>(values), 0);
+    }
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    static simdjson_inline simd8<T> repeat_16(
+      T v0,  T v1,  T v2,  T v3,  T v4,  T v5,  T v6,  T v7,
+      T v8,  T v9,  T v10, T v11, T v12, T v13, T v14, T v15
+    ) {
+      return simd8<T>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    simdjson_inline base8_numeric() : base8<T>() {}
+    simdjson_inline base8_numeric(const __m128i _value) : base8<T>(_value) {}
+
+    // Store to array
+    simdjson_inline void store(T dst[16]) const {
+      return __lsx_vst(*this, reinterpret_cast<__m128i *>(dst), 0);
+    }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<T> operator+(const simd8<T> other) const { return __lsx_vadd_b(*this, other); }
+    simdjson_inline simd8<T> operator-(const simd8<T> other) const { return __lsx_vsub_b(*this, other); }
+    simdjson_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
+
+    // Override to distinguish from bool version
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return __lsx_vshuf_b(lookup_table, lookup_table, *this);
+    }
+
+    // Copies to 'output" all bytes corresponding to a 0 in the mask (interpreted as a bitset).
+    // Passing a 0 value for mask would be equivalent to writing out every byte to output.
+    // Only the first 16 - count_ones(mask) bytes of the result are significant but 16 bytes
+    // get written.
+    template<typename L>
+    simdjson_inline void compress(uint16_t mask, L * output) const {
+      using internal::thintable_epi8;
+      using internal::BitsSetTable256mul2;
+      using internal::pshufb_combine_table;
+      // this particular implementation was inspired by haswell
+      // lsx do it in 2 steps, first 8 bytes and then second 8 bytes...
+      uint8_t mask1 = uint8_t(mask); // least significant 8 bits
+      uint8_t mask2 = uint8_t(mask >> 8); // second least significant 8 bits
+      // next line just loads the 64-bit values thintable_epi8[mask1] and
+      // thintable_epi8[mask2] into a 128-bit register.
+      __m128i shufmask = {int64_t(thintable_epi8[mask1]), int64_t(thintable_epi8[mask2]) + 0x0808080808080808};
+      // this is the version "nearly pruned"
+      __m128i pruned = __lsx_vshuf_b(*this, *this, shufmask);
+      // we still need to put the  pieces back together.
+      // we compute the popcount of the first words:
+      int pop1 = BitsSetTable256mul2[mask1];
+      // then load the corresponding mask
+      __m128i compactmask = __lsx_vldx(reinterpret_cast<void*>(reinterpret_cast<unsigned long>(pshufb_combine_table)), pop1 * 8);
+      __m128i answer = __lsx_vshuf_b(pruned, pruned, compactmask);
+      __lsx_vst(answer, reinterpret_cast<uint8_t*>(output), 0);
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> : base8_numeric<int8_t> {
+    simdjson_inline simd8() : base8_numeric<int8_t>() {}
+    simdjson_inline simd8(const __m128i _value) : base8_numeric<int8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const int8_t values[16]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) : simd8({
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+      }) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return __lsx_vmax_b(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return __lsx_vmin_b(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return __lsx_vslt_b(other, *this); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return __lsx_vslt_b(*this, other); }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t>: base8_numeric<uint8_t> {
+    simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+    simdjson_inline simd8(const __m128i _value) : base8_numeric<uint8_t>(_value) {}
+    // Splat constructor
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    // Array constructor
+    simdjson_inline simd8(const uint8_t values[16]) : simd8(load(values)) {}
+    // Member-by-member initialization
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) : simd8(__m128i(v16u8{
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    })) {}
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return __lsx_vsadd_bu(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return __lsx_vssub_bu(*this, other); }
+
+    // Order-specific operations
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return __lsx_vmax_bu(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return __lsx_vmin_bu(other, *this); }
+    // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
+    // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->lt_bits(other).any_bits_set(); }
+
+    // Bit-specific operations
+    simdjson_inline simd8<bool> bits_not_set() const { return *this == uint8_t(0); }
+    simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
+    simdjson_inline bool is_ascii() const { return 0 == __lsx_vpickve2gr_w(__lsx_vmskltz_b(*this), 0); }
+    simdjson_inline bool bits_not_set_anywhere() const { return 0 == __lsx_vpickve2gr_hu(__lsx_vmsknz_b(*this), 0); }
+    simdjson_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
+    simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const {
+      return 0 == __lsx_vpickve2gr_hu(__lsx_vmsknz_b(__lsx_vand_v(*this, bits)), 0);
+    }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(__lsx_vsrli_b(*this, N)); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(__lsx_vslli_b(*this, N)); }
+  };
+
+  template<typename T>
+  struct simd8x64 {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<T>);
+    static_assert(NUM_CHUNKS == 4, "LSX kernel should use four registers per 64-byte block.");
+    const simd8<T> chunks[NUM_CHUNKS];
+    template<int idx> simd8<uint8_t> get() const { return idx < NUM_CHUNKS ? chunks[idx] : simd8<T>(); }
+
+    simd8x64(const simd8x64<T>& o) = delete; // no copy allowed
+    simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1, const simd8<T> chunk2, const simd8<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+16), simd8<T>::load(ptr+32), simd8<T>::load(ptr+48)} {}
+
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
+      uint16_t mask1 = uint16_t(mask);
+      uint16_t mask2 = uint16_t(mask >> 16);
+      uint16_t mask3 = uint16_t(mask >> 32);
+      uint16_t mask4 = uint16_t(mask >> 48);
+      __m128i zcnt = __lsx_vpcnt_h(__m128i(v2u64{~mask, 0}));
+      uint64_t zcnt1 = __lsx_vpickve2gr_hu(zcnt, 0);
+      uint64_t zcnt2 = __lsx_vpickve2gr_hu(zcnt, 1);
+      uint64_t zcnt3 = __lsx_vpickve2gr_hu(zcnt, 2);
+      uint64_t zcnt4 = __lsx_vpickve2gr_hu(zcnt, 3);
+      uint8_t *voutput = reinterpret_cast<uint8_t*>(output);
+      // There should be a critical value which processes in scaler is faster.
+      if (zcnt1)
+        this->chunks[0].compress(mask1, reinterpret_cast<T*>(voutput));
+      voutput += zcnt1;
+      if (zcnt2)
+        this->chunks[1].compress(mask2, reinterpret_cast<T*>(voutput));
+      voutput += zcnt2;
+      if (zcnt3)
+        this->chunks[2].compress(mask3, reinterpret_cast<T*>(voutput));
+      voutput += zcnt3;
+      if (zcnt4)
+        this->chunks[3].compress(mask4, reinterpret_cast<T*>(voutput));
+      voutput += zcnt4;
+      return reinterpret_cast<uint64_t>(voutput) - reinterpret_cast<uint64_t>(output);
+    }
+
+    simdjson_inline void store(T ptr[64]) const {
+      this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
+      this->chunks[1].store(ptr+sizeof(simd8<T>)*1);
+      this->chunks[2].store(ptr+sizeof(simd8<T>)*2);
+      this->chunks[3].store(ptr+sizeof(simd8<T>)*3);
+    }
+
+    simdjson_inline uint64_t to_bitmask() const {
+      __m128i mask1 = __lsx_vmskltz_b(this->chunks[0]);
+      __m128i mask2 = __lsx_vmskltz_b(this->chunks[1]);
+      __m128i mask3 = __lsx_vmskltz_b(this->chunks[2]);
+      __m128i mask4 = __lsx_vmskltz_b(this->chunks[3]);
+      mask1 = __lsx_vilvl_h(mask2, mask1);
+      mask2 = __lsx_vilvl_h(mask4, mask3);
+      return __lsx_vpickve2gr_du(__lsx_vilvl_w(mask2, mask1), 0);
+    }
+
+    simdjson_inline simd8<T> reduce_or() const {
+      return (this->chunks[0] | this->chunks[1]) | (this->chunks[2] | this->chunks[3]);
+    }
+
+    simdjson_inline uint64_t eq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] == mask,
+        this->chunks[1] == mask,
+        this->chunks[2] == mask,
+        this->chunks[3] == mask
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+      return  simd8x64<bool>(
+        this->chunks[0] == other.chunks[0],
+        this->chunks[1] == other.chunks[1],
+        this->chunks[2] == other.chunks[2],
+        this->chunks[3] == other.chunks[3]
+      ).to_bitmask();
+    }
+
+    simdjson_inline uint64_t lteq(const T m) const {
+      const simd8<T> mask = simd8<T>::splat(m);
+      return  simd8x64<bool>(
+        this->chunks[0] <= mask,
+        this->chunks[1] <= mask,
+        this->chunks[2] <= mask,
+        this->chunks[3] <= mask
+      ).to_bitmask();
+    }
+  }; // struct simd8x64<T>
+
+} // namespace simd
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_LSX_SIMD_H
+/* end file simdjson/lsx/simd.h */
+/* including simdjson/lsx/stringparsing_defs.h: #include "simdjson/lsx/stringparsing_defs.h" */
+/* begin file simdjson/lsx/stringparsing_defs.h */
+#ifndef SIMDJSON_LSX_STRINGPARSING_DEFS_H
+#define SIMDJSON_LSX_STRINGPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/simd.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 32;
+  simdjson_inline backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return bs_bits != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+
+  uint32_t bs_bits;
+  uint32_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // this can read up to 31 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v0(src);
+  simd8<uint8_t> v1(src + sizeof(v0));
+  v0.store(dst);
+  v1.store(dst + sizeof(v0));
+
+  // Getting a 64-bit bitmask is much cheaper than multiple 16-bit bitmasks on LSX; therefore, we
+  // smash them together into a 64-byte mask and get the bitmask from there.
+  uint64_t bs_and_quote = simd8x64<bool>(v0 == '\\', v1 == '\\', v0 == '"', v1 == '"').to_bitmask();
+  return {
+    uint32_t(bs_and_quote),      // bs_bits
+    uint32_t(bs_and_quote >> 32) // quote_bits
+  };
+}
+
+
+struct escaping {
+  static constexpr uint32_t BYTES_PROCESSED = 16;
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits != 0; }
+  simdjson_inline int escape_index() { return trailing_zeroes(escape_bits); }
+
+  uint64_t escape_bits;
+}; // struct escaping
+
+
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "escaping finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  simd8<bool> is_quote = (v == '"');
+  simd8<bool> is_backslash = (v == '\\');
+  simd8<bool> is_control = (v < 32);
+  return {
+    static_cast<uint64_t>((is_backslash | is_quote | is_control).to_bitmask())
+  };
+}
+
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_LSX_STRINGPARSING_DEFS_H
+/* end file simdjson/lsx/stringparsing_defs.h */
+
+#define SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT 1
+/* end file simdjson/lsx/begin.h */
+/* including generic/amalgamated.h for lsx: #include <generic/amalgamated.h> */
+/* begin file generic/amalgamated.h for lsx */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_SRC_GENERIC_DEPENDENCIES_H)
+#error generic/dependencies.h must be included before generic/amalgamated.h!
+#endif
+
+/* including generic/base.h for lsx: #include <generic/base.h> */
+/* begin file generic/base.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+
+struct json_character_block;
+
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_BASE_H
+/* end file generic/base.h for lsx */
+/* including generic/dom_parser_implementation.h for lsx: #include <generic/dom_parser_implementation.h> */
+/* begin file generic/dom_parser_implementation.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// Interface a dom parser implementation must fulfill
+namespace simdjson {
+namespace lsx {
+namespace {
+
+simdjson_inline simd8<uint8_t> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3);
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input);
+
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file generic/dom_parser_implementation.h for lsx */
+/* including generic/json_character_block.h for lsx: #include <generic/json_character_block.h> */
+/* begin file generic/json_character_block.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+
+struct json_character_block {
+  static simdjson_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
+
+  simdjson_inline uint64_t whitespace() const noexcept { return _whitespace; }
+  simdjson_inline uint64_t op() const noexcept { return _op; }
+  simdjson_inline uint64_t scalar() const noexcept { return ~(op() | whitespace()); }
+
+  uint64_t _whitespace;
+  uint64_t _op;
+};
+
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H
+/* end file generic/json_character_block.h for lsx */
+/* end file generic/amalgamated.h for lsx */
+/* including generic/stage1/amalgamated.h for lsx: #include <generic/stage1/amalgamated.h> */
+/* begin file generic/stage1/amalgamated.h for lsx */
+// Stuff other things depend on
+/* including generic/stage1/base.h for lsx: #include <generic/stage1/base.h> */
+/* begin file generic/stage1/base.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_BASE_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace stage1 {
+
+class bit_indexer;
+template<size_t STEP_SIZE>
+struct buf_block_reader;
+struct json_block;
+class json_minifier;
+class json_scanner;
+struct json_string_block;
+class json_string_scanner;
+class json_structural_indexer;
+
+} // namespace stage1
+
+namespace utf8_validation {
+struct utf8_checker;
+} // namespace utf8_validation
+
+using utf8_validation::utf8_checker;
+
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_BASE_H
+/* end file generic/stage1/base.h for lsx */
+/* including generic/stage1/buf_block_reader.h for lsx: #include <generic/stage1/buf_block_reader.h> */
+/* begin file generic/stage1/buf_block_reader.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace stage1 {
+
+// Walks through a buffer in block-sized increments, loading the last part with spaces
+template<size_t STEP_SIZE>
+struct buf_block_reader {
+public:
+  simdjson_inline buf_block_reader(const uint8_t *_buf, size_t _len);
+  simdjson_inline size_t block_index();
+  simdjson_inline bool has_full_block() const;
+  simdjson_inline const uint8_t *full_block() const;
+  /**
+   * Get the last block, padded with spaces.
+   *
+   * There will always be a last block, with at least 1 byte, unless len == 0 (in which case this
+   * function fills the buffer with spaces and returns 0. In particular, if len == STEP_SIZE there
+   * will be 0 full_blocks and 1 remainder block with STEP_SIZE bytes and no spaces for padding.
+   *
+   * @return the number of effective characters in the last block.
+   */
+  simdjson_inline size_t get_remainder(uint8_t *dst) const;
+  simdjson_inline void advance();
+private:
+  const uint8_t *buf;
+  const size_t len;
+  const size_t lenminusstep;
+  size_t idx;
+};
+
+// Routines to print masks and text for debugging bitmask operations
+simdjson_unused static char * format_input_text_64(const uint8_t *text) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    buf[i] = int8_t(text[i]) < ' ' ? '_' : int8_t(text[i]);
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+// Routines to print masks and text for debugging bitmask operations
+simdjson_unused static char * format_input_text(const simd8x64<uint8_t>& in) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  in.store(reinterpret_cast<uint8_t*>(buf));
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] < ' ') { buf[i] = '_'; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+simdjson_unused static char * format_input_text(const simd8x64<uint8_t>& in, uint64_t mask) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  in.store(reinterpret_cast<uint8_t*>(buf));
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] <= ' ') { buf[i] = '_'; }
+    if (!(mask & (size_t(1) << i))) { buf[i] = ' '; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+simdjson_unused static char * format_mask(uint64_t mask) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  for (size_t i=0; i<64; i++) {
+    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
+  }
+  buf[64] = '\0';
+  return buf;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
+
+template<size_t STEP_SIZE>
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
+
+template<size_t STEP_SIZE>
+simdjson_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
+  return idx < lenminusstep;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
+  return &buf[idx];
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
+  if(len == idx) { return 0; } // memcpy(dst, null, 0) will trigger an error with some sanitizers
+  std::memset(dst, 0x20, STEP_SIZE); // std::memset STEP_SIZE because it's more efficient to write out 8 or 16 bytes at once.
+  std::memcpy(dst, buf + idx, len - idx);
+  return len - idx;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline void buf_block_reader<STEP_SIZE>::advance() {
+  idx += STEP_SIZE;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H
+/* end file generic/stage1/buf_block_reader.h for lsx */
+/* including generic/stage1/json_escape_scanner.h for lsx: #include <generic/stage1/json_escape_scanner.h> */
+/* begin file generic/stage1/json_escape_scanner.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_ESCAPE_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_ESCAPE_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace stage1 {
+
+/**
+ * Scans for escape characters in JSON, taking care with multiple backslashes (\\n vs. \n).
+ */
+struct json_escape_scanner {
+  /** The actual escape characters (the backslashes themselves). */
+  uint64_t next_is_escaped = 0ULL;
+
+  struct escaped_and_escape {
+    /**
+     * Mask of escaped characters.
+     *
+     * ```
+     * \n \\n \\\n \\\\n \
+     * 0100100010100101000
+     *  n  \   \ n  \ \
+     * ```
+     */
+    uint64_t escaped;
+    /**
+     * Mask of escape characters.
+     *
+     * ```
+     * \n \\n \\\n \\\\n \
+     * 1001000101001010001
+     * \  \   \ \  \ \   \
+     * ```
+     */
+    uint64_t escape;
+  };
+
+  /**
+   * Get a mask of both escape and escaped characters (the characters following a backslash).
+   *
+   * @param potential_escape A mask of the character that can escape others (but could be
+   *        escaped itself). e.g. block.eq('\\')
+   */
+  simdjson_really_inline escaped_and_escape next(uint64_t backslash) noexcept {
+
+#if !SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+    if (!backslash) { return {next_escaped_without_backslashes(), 0}; }
+#endif
+
+    // |                                | Mask (shows characters instead of 1's) | Depth | Instructions        |
+    // |--------------------------------|----------------------------------------|-------|---------------------|
+    // | string                         | `\\n_\\\n___\\\n___\\\\___\\\\__\\\`   |       |                     |
+    // |                                | `    even   odd    even   odd   odd`   |       |                     |
+    // | potential_escape               | ` \  \\\    \\\    \\\\   \\\\  \\\`   | 1     | 1 (backslash & ~first_is_escaped)
+    // | escape_and_terminal_code       | ` \n \ \n   \ \n   \ \    \ \   \ \`   | 5     | 5 (next_escape_and_terminal_code())
+    // | escaped                        | `\    \ n    \ n    \ \    \ \   \ ` X | 6     | 7 (escape_and_terminal_code ^ (potential_escape | first_is_escaped))
+    // | escape                         | `    \ \    \ \    \ \    \ \   \ \`   | 6     | 8 (escape_and_terminal_code & backslash)
+    // | first_is_escaped               | `\                                 `   | 7 (*) | 9 (escape >> 63) ()
+    //                                                                               (*) this is not needed until the next iteration
+    uint64_t escape_and_terminal_code = next_escape_and_terminal_code(backslash & ~this->next_is_escaped);
+    uint64_t escaped = escape_and_terminal_code ^ (backslash | this->next_is_escaped);
+    uint64_t escape = escape_and_terminal_code & backslash;
+    this->next_is_escaped = escape >> 63;
+    return {escaped, escape};
+  }
+
+private:
+  static constexpr const uint64_t ODD_BITS = 0xAAAAAAAAAAAAAAAAULL;
+
+  simdjson_really_inline uint64_t next_escaped_without_backslashes() noexcept {
+    uint64_t escaped = this->next_is_escaped;
+    this->next_is_escaped = 0;
+    return escaped;
+  }
+
+  /**
+   * Returns a mask of the next escape characters (masking out escaped backslashes), along with
+   * any non-backslash escape codes.
+   *
+   * \n \\n \\\n \\\\n returns:
+   * \n \   \ \n \ \
+   * 11 100 1011 10100
+   *
+   * You are expected to mask out the first bit yourself if the previous block had a trailing
+   * escape.
+   *
+   * & the result with potential_escape to get just the escape characters.
+   * ^ the result with (potential_escape | first_is_escaped) to get escaped characters.
+   */
+  static simdjson_really_inline uint64_t next_escape_and_terminal_code(uint64_t potential_escape) noexcept {
+    // If we were to just shift and mask out any odd bits, we'd actually get a *half* right answer:
+    // any even-aligned backslash runs would be correct! Odd-aligned backslash runs would be
+    // inverted (\\\ would be 010 instead of 101).
+    //
+    // ```
+    // string:              | ____\\\\_\\\\_____ |
+    // maybe_escaped | ODD  |     \ \   \ \      |
+    //               even-aligned ^^^  ^^^^ odd-aligned
+    // ```
+    //
+    // Taking that into account, our basic strategy is:
+    //
+    // 1. Use subtraction to produce a mask with 1's for even-aligned runs and 0's for
+    //    odd-aligned runs.
+    // 2. XOR all odd bits, which masks out the odd bits in even-aligned runs, and brings IN the
+    //    odd bits in odd-aligned runs.
+    // 3. & with backslash to clean up any stray bits.
+    // runs are set to 0, and then XORing with "odd":
+    //
+    // |                                | Mask (shows characters instead of 1's) | Instructions        |
+    // |--------------------------------|----------------------------------------|---------------------|
+    // | string                         | `\\n_\\\n___\\\n___\\\\___\\\\__\\\`   |
+    // |                                | `    even   odd    even   odd   odd`   |
+    // | maybe_escaped                  | `  n  \\n    \\n    \\\_   \\\_  \\` X | 1 (potential_escape << 1)
+    // | maybe_escaped_and_odd          | ` \n_ \\n _ \\\n_ _ \\\__ _\\\_ \\\`   | 1 (maybe_escaped | odd)
+    // | even_series_codes_and_odd      | `  n_\\\  _    n_ _\\\\ _     _    `   | 1 (maybe_escaped_and_odd - potential_escape)
+    // | escape_and_terminal_code       | ` \n \ \n   \ \n   \ \    \ \   \ \`   | 1 (^ odd)
+    //
+
+    // Escaped characters are characters following an escape.
+    uint64_t maybe_escaped = potential_escape << 1;
+
+    // To distinguish odd from even escape sequences, therefore, we turn on any *starting*
+    // escapes that are on an odd byte. (We actually bring in all odd bits, for speed.)
+    // - Odd runs of backslashes are 0000, and the code at the end ("n" in \n or \\n) is 1.
+    // - Odd runs of backslashes are 1111, and the code at the end ("n" in \n or \\n) is 0.
+    // - All other odd bytes are 1, and even bytes are 0.
+    uint64_t maybe_escaped_and_odd_bits     = maybe_escaped | ODD_BITS;
+    uint64_t even_series_codes_and_odd_bits = maybe_escaped_and_odd_bits - potential_escape;
+
+    // Now we flip all odd bytes back with xor. This:
+    // - Makes odd runs of backslashes go from 0000 to 1010
+    // - Makes even runs of backslashes go from 1111 to 1010
+    // - Sets actually-escaped codes to 1 (the n in \n and \\n: \n = 11, \\n = 100)
+    // - Resets all other bytes to 0
+    return even_series_codes_and_odd_bits ^ ODD_BITS;
+  }
+};
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+/* end file generic/stage1/json_escape_scanner.h for lsx */
+/* including generic/stage1/json_string_scanner.h for lsx: #include <generic/stage1/json_string_scanner.h> */
+/* begin file generic/stage1/json_string_scanner.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_escape_scanner.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace stage1 {
+
+struct json_string_block {
+  // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
+  simdjson_really_inline json_string_block(uint64_t escaped, uint64_t quote, uint64_t in_string) :
+  _escaped(escaped), _quote(quote), _in_string(in_string) {}
+
+  // Escaped characters (characters following an escape() character)
+  simdjson_really_inline uint64_t escaped() const { return _escaped; }
+  // Real (non-backslashed) quotes
+  simdjson_really_inline uint64_t quote() const { return _quote; }
+  // Only characters inside the string (not including the quotes)
+  simdjson_really_inline uint64_t string_content() const { return _in_string & ~_quote; }
+  // Return a mask of whether the given characters are inside a string (only works on non-quotes)
+  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
+  // Return a mask of whether the given characters are inside a string (only works on non-quotes)
+  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
+  // Tail of string (everything except the start quote)
+  simdjson_really_inline uint64_t string_tail() const { return _in_string ^ _quote; }
+
+  // escaped characters (backslashed--does not include the hex characters after \u)
+  uint64_t _escaped;
+  // real quotes (non-escaped ones)
+  uint64_t _quote;
+  // string characters (includes start quote but not end quote)
+  uint64_t _in_string;
+};
+
+// Scans blocks for string characters, storing the state necessary to do so
+class json_string_scanner {
+public:
+  simdjson_really_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
+  // Returns either UNCLOSED_STRING or SUCCESS
+  simdjson_really_inline error_code finish();
+
+private:
+  // Scans for escape characters
+  json_escape_scanner escape_scanner{};
+  // Whether the last iteration was still inside a string (all 1's = true, all 0's = false).
+  uint64_t prev_in_string = 0ULL;
+};
+
+//
+// Return a mask of all string characters plus end quotes.
+//
+// prev_escaped is overflow saying whether the next character is escaped.
+// prev_in_string is overflow saying whether we're still in a string.
+//
+// Backslash sequences outside of quotes will be detected in stage 2.
+//
+simdjson_really_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
+  const uint64_t backslash = in.eq('\\');
+  const uint64_t escaped = escape_scanner.next(backslash).escaped;
+  const uint64_t quote = in.eq('"') & ~escaped;
+
+  //
+  // prefix_xor flips on bits inside the string (and flips off the end quote).
+  //
+  // Then we xor with prev_in_string: if we were in a string already, its effect is flipped
+  // (characters inside strings are outside, and characters outside strings are inside).
+  //
+  const uint64_t in_string = prefix_xor(quote) ^ prev_in_string;
+
+  //
+  // Check if we're still in a string at the end of the box so the next block will know
+  //
+  prev_in_string = uint64_t(static_cast<int64_t>(in_string) >> 63);
+
+  // Use ^ to turn the beginning quote off, and the end quote on.
+
+  // We are returning a function-local object so either we get a move constructor
+  // or we get copy elision.
+  return json_string_block(escaped, quote, in_string);
+}
+
+simdjson_really_inline error_code json_string_scanner::finish() {
+  if (prev_in_string) {
+    return UNCLOSED_STRING;
+  }
+  return SUCCESS;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+/* end file generic/stage1/json_string_scanner.h for lsx */
+/* including generic/stage1/utf8_lookup4_algorithm.h for lsx: #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* begin file generic/stage1/utf8_lookup4_algorithm.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace utf8_validation {
+
+using namespace simd;
+
+  simdjson_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+// Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+// Bit 1 = Too Long (ASCII followed by continuation)
+// Bit 2 = Overlong 3-byte
+// Bit 4 = Surrogate
+// Bit 5 = Overlong 2-byte
+// Bit 7 = Two Continuations
+    constexpr const uint8_t TOO_SHORT   = 1<<0; // 11______ 0_______
+                                                // 11______ 11______
+    constexpr const uint8_t TOO_LONG    = 1<<1; // 0_______ 10______
+    constexpr const uint8_t OVERLONG_3  = 1<<2; // 11100000 100_____
+    constexpr const uint8_t SURROGATE   = 1<<4; // 11101101 101_____
+    constexpr const uint8_t OVERLONG_2  = 1<<5; // 1100000_ 10______
+    constexpr const uint8_t TWO_CONTS   = 1<<7; // 10______ 10______
+    constexpr const uint8_t TOO_LARGE   = 1<<3; // 11110100 1001____
+                                                // 11110100 101_____
+                                                // 11110101 1001____
+                                                // 11110101 101_____
+                                                // 1111011_ 1001____
+                                                // 1111011_ 101_____
+                                                // 11111___ 1001____
+                                                // 11111___ 101_____
+    constexpr const uint8_t TOO_LARGE_1000 = 1<<6;
+                                                // 11110101 1000____
+                                                // 1111011_ 1000____
+                                                // 11111___ 1000____
+    constexpr const uint8_t OVERLONG_4  = 1<<6; // 11110000 1000____
+
+    const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      TOO_SHORT,
+      // 1110____ ________ <three byte lead in byte 1>
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4
+    );
+    constexpr const uint8_t CARRY = TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+    const simd8<uint8_t> byte_1_low = (prev1 & 0x0F).lookup_16<uint8_t>(
+      // ____0000 ________
+      CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+      // ____0001 ________
+      CARRY | OVERLONG_2,
+      // ____001_ ________
+      CARRY,
+      CARRY,
+
+      // ____0100 ________
+      CARRY | TOO_LARGE,
+      // ____0101 ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      // ____011_ ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+
+      // ____1___ ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      // ____1101 ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000
+    );
+    const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE  | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE  | TOO_LARGE,
+
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT
+    );
+    return (byte_1_high & byte_1_low & byte_2_high);
+  }
+  simdjson_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
+      const simd8<uint8_t> prev_input, const simd8<uint8_t> sc) {
+    simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+    simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+    simd8<uint8_t> must23 = must_be_2_3_continuation(prev2, prev3);
+    simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+    return must23_80 ^ sc;
+  }
+
+  //
+  // Return nonzero if there are incomplete multibyte characters at the end of the block:
+  // e.g. if there is a 4-byte character, but it's 3 bytes from the end.
+  //
+  simdjson_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+    // If the previous input's last 3 bytes match this, they're too short (they ended at EOF):
+    // ... 1111____ 111_____ 11______
+#if SIMDJSON_IMPLEMENTATION_ICELAKE || (SIMDJSON_IMPLEMENTATION_RVV_VLS && __riscv_v_fixed_vlen >= 512)
+    static const uint8_t max_array[64] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
+    };
+#else
+    static const uint8_t max_array[32] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
+    };
+#endif
+    const simd8<uint8_t> max_value(&max_array[sizeof(max_array)-sizeof(simd8<uint8_t>)]);
+    return input.gt_bits(max_value);
+  }
+
+  struct utf8_checker {
+    // If this is nonzero, there has been a UTF-8 error.
+    simd8<uint8_t> error;
+    // The last input we received
+    simd8<uint8_t> prev_input_block;
+    // Whether the last input we received was incomplete (used for ASCII fast path)
+    simd8<uint8_t> prev_incomplete;
+
+    //
+    // Check whether the current bytes are valid UTF-8.
+    //
+    simdjson_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
+      // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ lead bytes
+      // (2, 3, 4-byte leads become large positive numbers instead of small negative numbers)
+      simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+      simd8<uint8_t> sc = check_special_cases(input, prev1);
+      this->error |= check_multibyte_lengths(input, prev_input, sc);
+    }
+
+    // The only problem that can happen at EOF is that a multibyte character is too short
+    // or a byte value too large in the last bytes: check_special_cases only checks for bytes
+    // too large in the first of two bytes.
+    simdjson_inline void check_eof() {
+      // If the previous block had incomplete UTF-8 characters at the end, an ASCII block can't
+      // possibly finish them.
+      this->error |= this->prev_incomplete;
+    }
+
+    simdjson_inline void check_next_input(const simd8x64<uint8_t>& input) {
+      if(simdjson_likely(is_ascii(input))) {
+        this->error |= this->prev_incomplete;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio, it is not good enough.
+        static_assert((simd8x64<uint8_t>::NUM_CHUNKS == 1)
+                ||(simd8x64<uint8_t>::NUM_CHUNKS == 2)
+                || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+                "We support one, two or four chunks per 64-byte block.");
+        SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 1) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+          this->check_utf8_bytes(input.get<1>(), input.get<0>());
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+          this->check_utf8_bytes(input.get<1>(), input.get<0>());
+          this->check_utf8_bytes(input.get<2>(), input.get<1>());
+          this->check_utf8_bytes(input.get<3>(), input.get<2>());
+        }
+        this->prev_incomplete = is_incomplete(input.get<simd8x64<uint8_t>::NUM_CHUNKS-1>());
+        this->prev_input_block = input.get<simd8x64<uint8_t>::NUM_CHUNKS-1>();
+      }
+    }
+    // do not forget to call check_eof!
+    simdjson_warn_unused simdjson_inline error_code errors() {
+      return this->error.any_bits_set_anywhere() ? error_code::UTF8_ERROR : error_code::SUCCESS;
+    }
+
+  }; // struct utf8_checker
+} // namespace utf8_validation
+
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H
+/* end file generic/stage1/utf8_lookup4_algorithm.h for lsx */
+/* including generic/stage1/json_scanner.h for lsx: #include <generic/stage1/json_scanner.h> */
+/* begin file generic/stage1/json_scanner.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/json_character_block.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_string_scanner.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace stage1 {
+
+/**
+ * A block of scanned json, with information on operators and scalars.
+ *
+ * We seek to identify pseudo-structural characters. Anything that is inside
+ * a string must be omitted (hence  & ~_string.string_tail()).
+ * Otherwise, pseudo-structural characters come in two forms.
+ * 1. We have the structural characters ([,],{,},:, comma). The
+ *    term 'structural character' is from the JSON RFC.
+ * 2. We have the 'scalar pseudo-structural characters'.
+ *    Scalars are quotes, and any character except structural characters and white space.
+ *
+ * To identify the scalar pseudo-structural characters, we must look at what comes
+ * before them: it must be a space, a quote or a structural characters.
+ * Starting with simdjson v0.3, we identify them by
+ * negation: we identify everything that is followed by a non-quote scalar,
+ * and we negate that. Whatever remains must be a 'scalar pseudo-structural character'.
+ */
+struct json_block {
+public:
+  // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
+  simdjson_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  _string(std::move(string)), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
+  simdjson_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  _string(string), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
+
+  /**
+   * The start of structurals.
+   * In simdjson prior to v0.3, these were called the pseudo-structural characters.
+   **/
+  simdjson_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
+  /** All JSON whitespace (i.e. not in a string) */
+  simdjson_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
+
+  // Helpers
+
+  /** Whether the given characters are inside a string (only works on non-quotes) */
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
+  /** Whether the given characters are outside a string (only works on non-quotes) */
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
+
+  // string and escape characters
+  json_string_block _string;
+  // whitespace, structural characters ('operators'), scalars
+  json_character_block _characters;
+  // whether the previous character was a scalar
+  uint64_t _follows_potential_nonquote_scalar;
+private:
+  // Potential structurals (i.e. disregarding strings)
+
+  /**
+   * structural elements ([,],{,},:, comma) plus scalar starts like 123, true and "abc".
+   * They may reside inside a string.
+   **/
+  simdjson_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
+  /**
+   * The start of non-operator runs, like 123, true and "abc".
+   * It main reside inside a string.
+   **/
+  simdjson_inline uint64_t potential_scalar_start() const noexcept {
+    // The term "scalar" refers to anything except structural characters and white space
+    // (so letters, numbers, quotes).
+    // Whenever it is preceded by something that is not a structural element ({,},[,],:, ") nor a white-space
+    // then we know that it is irrelevant structurally.
+    return _characters.scalar() & ~follows_potential_scalar();
+  }
+  /**
+   * Whether the given character is immediately after a non-operator like 123, true.
+   * The characters following a quote are not included.
+   */
+  simdjson_inline uint64_t follows_potential_scalar() const noexcept {
+    // _follows_potential_nonquote_scalar: is defined as marking any character that follows a character
+    // that is not a structural element ({,},[,],:, comma) nor a quote (") and that is not a
+    // white space.
+    // It is understood that within quoted region, anything at all could be marked (irrelevant).
+    return _follows_potential_nonquote_scalar;
+  }
+};
+
+/**
+ * Scans JSON for important bits: structural characters or 'operators', strings, and scalars.
+ *
+ * The scanner starts by calculating two distinct things:
+ * - string characters (taking \" into account)
+ * - structural characters or 'operators' ([]{},:, comma)
+ *   and scalars (runs of non-operators like 123, true and "abc")
+ *
+ * To minimize data dependency (a key component of the scanner's speed), it finds these in parallel:
+ * in particular, the operator/scalar bit will find plenty of things that are actually part of
+ * strings. When we're done, json_block will fuse the two together by masking out tokens that are
+ * part of a string.
+ */
+class json_scanner {
+public:
+  json_scanner() = default;
+  simdjson_inline json_block next(const simd::simd8x64<uint8_t>& in);
+  // Returns either UNCLOSED_STRING or SUCCESS
+  simdjson_warn_unused simdjson_inline error_code finish();
+
+private:
+  // Whether the last character of the previous iteration is part of a scalar token
+  // (anything except whitespace or a structural character/'operator').
+  uint64_t prev_scalar = 0ULL;
+  json_string_scanner string_scanner{};
+};
+
+
+//
+// Check if the current character immediately follows a matching character.
+//
+// For example, this checks for quotes with backslashes in front of them:
+//
+//     const uint64_t backslashed_quote = in.eq('"') & immediately_follows(in.eq('\'), prev_backslash);
+//
+simdjson_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
+  const uint64_t result = match << 1 | overflow;
+  overflow = match >> 63;
+  return result;
+}
+
+simdjson_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
+  json_string_block strings = string_scanner.next(in);
+  // identifies the white-space and the structural characters
+  json_character_block characters = json_character_block::classify(in);
+  // The term "scalar" refers to anything except structural characters and white space
+  // (so letters, numbers, quotes).
+  // We want follows_scalar to mark anything that follows a non-quote scalar (so letters and numbers).
+  //
+  // A terminal quote should either be followed by a structural character (comma, brace, bracket, colon)
+  // or nothing. However, we still want ' "a string"true ' to mark the 't' of 'true' as a potential
+  // pseudo-structural character just like we would if we had  ' "a string" true '; otherwise we
+  // may need to add an extra check when parsing strings.
+  //
+  // Performance: there are many ways to skin this cat.
+  const uint64_t nonquote_scalar = characters.scalar() & ~strings.quote();
+  uint64_t follows_nonquote_scalar = follows(nonquote_scalar, prev_scalar);
+  // We are returning a function-local object so either we get a move constructor
+  // or we get copy elision.
+  return json_block(
+    strings,// strings is a function-local object so either it moves or the copy is elided.
+    characters,
+    follows_nonquote_scalar
+  );
+}
+
+simdjson_warn_unused simdjson_inline error_code json_scanner::finish() {
+  return string_scanner.finish();
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H
+/* end file generic/stage1/json_scanner.h for lsx */
+
+// All other declarations
+/* including generic/stage1/find_next_document_index.h for lsx: #include <generic/stage1/find_next_document_index.h> */
+/* begin file generic/stage1/find_next_document_index.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace stage1 {
+
+/**
+  * This algorithm is used to quickly identify the last structural position that
+  * makes up a complete document.
+  *
+  * It does this by going backwards and finding the last *document boundary* (a
+  * place where one value follows another without a comma between them). If the
+  * last document (the characters after the boundary) has an equal number of
+  * start and end brackets, it is considered complete.
+  *
+  * Simply put, we iterate over the structural characters, starting from
+  * the end. We consider that we found the end of a JSON document when the
+  * first element of the pair is NOT one of these characters: '{' '[' ':' ','
+  * and when the second element is NOT one of these characters: '}' ']' ':' ','.
+  *
+  * This simple comparison works most of the time, but it does not cover cases
+  * where the batch's structural indexes contain a perfect amount of documents.
+  * In such a case, we do not have access to the structural index which follows
+  * the last document, therefore, we do not have access to the second element in
+  * the pair, and that means we cannot identify the last document. To fix this
+  * issue, we keep a count of the open and closed curly/square braces we found
+  * while searching for the pair. When we find a pair AND the count of open and
+  * closed curly/square braces is the same, we know that we just passed a
+  * complete document, therefore the last json buffer location is the end of the
+  * batch.
+  */
+simdjson_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
+  // Variant: do not count separately, just figure out depth
+  if(parser.n_structural_indexes == 0) { return 0; }
+  auto arr_cnt = 0;
+  auto obj_cnt = 0;
+  for (auto i = parser.n_structural_indexes - 1; i > 0; i--) {
+    auto idxb = parser.structural_indexes[i];
+    switch (parser.buf[idxb]) {
+    case ':':
+    case ',':
+      continue;
+    case '}':
+      obj_cnt--;
+      continue;
+    case ']':
+      arr_cnt--;
+      continue;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+    }
+    auto idxa = parser.structural_indexes[i - 1];
+    switch (parser.buf[idxa]) {
+    case '{':
+    case '[':
+    case ':':
+    case ',':
+      continue;
+    }
+    // Last document is complete, so the next document will appear after!
+    if (!arr_cnt && !obj_cnt) {
+      return parser.n_structural_indexes;
+    }
+    // Last document is incomplete; mark the document at i + 1 as the next one
+    return i;
+  }
+  // If we made it to the end, we want to finish counting to see if we have a full document.
+  switch (parser.buf[parser.structural_indexes[0]]) {
+    case '}':
+      obj_cnt--;
+      break;
+    case ']':
+      arr_cnt--;
+      break;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+  }
+  if (!arr_cnt && !obj_cnt) {
+    // We have a complete document.
+    return parser.n_structural_indexes;
+  }
+  return 0;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+/* end file generic/stage1/find_next_document_index.h for lsx */
+/* including generic/stage1/json_minifier.h for lsx: #include <generic/stage1/json_minifier.h> */
+/* begin file generic/stage1/json_minifier.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses in stage1
+// It is intended to be included multiple times and compiled multiple times
+// We assume the file in which it is included already includes
+// "simdjson/stage1.h" (this simplifies amalgation)
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace stage1 {
+
+class json_minifier {
+public:
+  template<size_t STEP_SIZE>
+  static error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept;
+
+private:
+  simdjson_inline json_minifier(uint8_t *_dst)
+  : dst{_dst}
+  {}
+  template<size_t STEP_SIZE>
+  simdjson_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
+  simdjson_warn_unused simdjson_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
+  json_scanner scanner{};
+  uint8_t *dst;
+};
+
+simdjson_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
+  uint64_t mask = block.whitespace();
+  dst += in.compress(mask, dst);
+}
+
+simdjson_warn_unused simdjson_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
+  error_code error = scanner.finish();
+  if (error) { dst_len = 0; return error; }
+  dst_len = dst - dst_start;
+  return SUCCESS;
+}
+
+template<>
+simdjson_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block_buf);
+  simd::simd8x64<uint8_t> in_2(block_buf+64);
+  json_block block_1 = scanner.next(in_1);
+  json_block block_2 = scanner.next(in_2);
+  this->next(in_1, block_1);
+  this->next(in_2, block_2);
+  reader.advance();
+}
+
+template<>
+simdjson_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block_buf);
+  json_block block_1 = scanner.next(in_1);
+  this->next(block_buf, block_1);
+  reader.advance();
+}
+
+template<size_t STEP_SIZE>
+error_code json_minifier::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept {
+  buf_block_reader<STEP_SIZE> reader(buf, len);
+  json_minifier minifier(dst);
+
+  // Index the first n-1 blocks
+  while (reader.has_full_block()) {
+    minifier.step<STEP_SIZE>(reader.full_block(), reader);
+  }
+
+  // Index the last (remainder) block, padded with spaces
+  uint8_t block[STEP_SIZE];
+  size_t remaining_bytes = reader.get_remainder(block);
+  if (remaining_bytes > 0) {
+    // We do not want to write directly to the output stream. Rather, we write
+    // to a local buffer (for safety).
+    uint8_t out_block[STEP_SIZE];
+    uint8_t * const guarded_dst{minifier.dst};
+    minifier.dst = out_block;
+    minifier.step<STEP_SIZE>(block, reader);
+    size_t to_write = minifier.dst - out_block;
+    // In some cases, we could be enticed to consider the padded spaces
+    // as part of the string. This is fine as long as we do not write more
+    // than we consumed.
+    if(to_write > remaining_bytes) { to_write = remaining_bytes; }
+    memcpy(guarded_dst, out_block, to_write);
+    minifier.dst = guarded_dst + to_write;
+  }
+  return minifier.finish(dst, dst_len);
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H
+/* end file generic/stage1/json_minifier.h for lsx */
+/* including generic/stage1/json_structural_indexer.h for lsx: #include <generic/stage1/json_structural_indexer.h> */
+/* begin file generic/stage1/json_structural_indexer.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_string_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_minifier.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/find_next_document_index.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses in stage1
+// It is intended to be included multiple times and compiled multiple times
+// We assume the file in which it is included already includes
+// "simdjson/stage1.h" (this simplifies amalgation)
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace stage1 {
+
+class bit_indexer {
+public:
+  uint32_t *tail;
+
+  simdjson_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
+
+#if SIMDJSON_PREFER_REVERSE_BITS
+  /**
+    * ARM lacks a fast trailing zero instruction, but it has a fast
+    * bit reversal instruction and a fast leading zero instruction.
+    * Thus it may be profitable to reverse the bits (once) and then
+    * to rely on a sequence of instructions that call the leading
+    * zero instruction.
+    *
+    * Performance notes:
+    * The chosen routine is not optimal in terms of data dependency
+    * since zero_leading_bit might require two instructions. However,
+    * it tends to minimize the total number of instructions which is
+    * beneficial.
+    */
+  simdjson_inline void write_index(uint32_t idx, uint64_t& rev_bits, int i) {
+    int lz = leading_zeroes(rev_bits);
+    this->tail[i] = static_cast<uint32_t>(idx) + lz;
+    rev_bits = zero_leading_bit(rev_bits, lz);
+  }
+#else
+  /**
+    * Under recent x64 systems, we often have both a fast trailing zero
+    * instruction and a fast 'clear-lower-bit' instruction so the following
+    * algorithm can be competitive.
+    */
+
+  simdjson_inline void write_index(uint32_t idx, uint64_t& bits, int i) {
+    this->tail[i] = idx + trailing_zeroes(bits);
+    bits = clear_lowest_bit(bits);
+  }
+#endif // SIMDJSON_PREFER_REVERSE_BITS
+
+  template <int START, int N>
+  simdjson_inline int write_indexes(uint32_t idx, uint64_t& bits) {
+    write_index(idx, bits, START);
+    SIMDJSON_IF_CONSTEXPR (N > 1) {
+      write_indexes<(N-1>0?START+1:START), (N-1>=0?N-1:1)>(idx, bits);
+    }
+    return START+N;
+  }
+
+  template <int START, int END, int STEP>
+  simdjson_inline int write_indexes_stepped(uint32_t idx, uint64_t& bits, int cnt) {
+    write_indexes<START, STEP>(idx, bits);
+    SIMDJSON_IF_CONSTEXPR ((START+STEP)  < END) {
+      if (simdjson_unlikely((START+STEP) < cnt)) {
+        write_indexes_stepped<(START+STEP<END?START+STEP:END), END, STEP>(idx, bits, cnt);
+      }
+    }
+    return ((END-START) % STEP) == 0 ? END : (END-START) - ((END-START) % STEP) + STEP;
+  }
+
+  // flatten out values in 'bits' assuming that they are are to have values of idx
+  // plus their position in the bitvector, and store these indexes at
+  // base_ptr[base] incrementing base as we go
+  // will potentially store extra values beyond end of valid bits, so base_ptr
+  // needs to be large enough to handle this
+  //
+  // If the kernel sets SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER, then it
+  // will provide its own version of the code.
+#ifdef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+  simdjson_inline void write(uint32_t idx, uint64_t bits);
+#else
+  simdjson_inline void write(uint32_t idx, uint64_t bits) {
+    // In some instances, the next branch is expensive because it is mispredicted.
+    // Unfortunately, in other cases,
+    // it helps tremendously.
+    if (bits == 0)
+        return;
+
+    int cnt = static_cast<int>(count_ones(bits));
+
+#if SIMDJSON_PREFER_REVERSE_BITS
+    bits = reverse_bits(bits);
+#endif
+#ifdef SIMDJSON_STRUCTURAL_INDEXER_STEP
+    static constexpr const int STEP = SIMDJSON_STRUCTURAL_INDEXER_STEP;
+#else
+    static constexpr const int STEP = 4;
+#endif
+    static constexpr const int STEP_UNTIL = 24;
+
+    write_indexes_stepped<0, STEP_UNTIL, STEP>(idx, bits, cnt);
+    SIMDJSON_IF_CONSTEXPR (STEP_UNTIL < 64) {
+      if (simdjson_unlikely(STEP_UNTIL < cnt)) {
+        for (int i=STEP_UNTIL; i<cnt; i++) {
+          write_index(idx, bits, i);
+        }
+      }
+    }
+
+    this->tail += cnt;
+  }
+#endif // SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+};
+
+class json_structural_indexer {
+public:
+  /**
+   * Find the important bits of JSON in a 128-byte chunk, and add them to structural_indexes.
+   *
+   * @param partial Setting the partial parameter to true allows the find_structural_bits to
+   *   tolerate unclosed strings. The caller should still ensure that the input is valid UTF-8. If
+   *   you are processing substrings, you may want to call on a function like trimmed_length_safe_utf8.
+   */
+  template<size_t STEP_SIZE>
+  static error_code index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept;
+
+private:
+  simdjson_inline json_structural_indexer(uint32_t *structural_indexes);
+  template<size_t STEP_SIZE>
+  simdjson_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
+  simdjson_warn_unused simdjson_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
+
+  json_scanner scanner{};
+  utf8_checker checker{};
+  bit_indexer indexer;
+  uint64_t prev_structurals = 0;
+  uint64_t unescaped_chars_error = 0;
+};
+
+simdjson_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
+
+// Skip the last character if it is partial
+simdjson_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
+  if (simdjson_unlikely(len < 3)) {
+    switch (len) {
+      case 2:
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
+        return len;
+      case 1:
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        return len;
+      case 0:
+        return len;
+    }
+  }
+  if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+  if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 1 byte left
+  if (buf[len-3] >= 0xf0) { return len-3; } // 4-byte characters with only 3 bytes left
+  return len;
+}
+
+//
+// PERF NOTES:
+// We pipe 2 inputs through these stages:
+// 1. Load JSON into registers. This takes a long time and is highly parallelizable, so we load
+//    2 inputs' worth at once so that by the time step 2 is looking for them input, it's available.
+// 2. Scan the JSON for critical data: strings, scalars and operators. This is the critical path.
+//    The output of step 1 depends entirely on this information. These functions don't quite use
+//    up enough CPU: the second half of the functions is highly serial, only using 1 execution core
+//    at a time. The second input's scans has some dependency on the first ones finishing it, but
+//    they can make a lot of progress before they need that information.
+// 3. Step 1 does not use enough capacity, so we run some extra stuff while we're waiting for that
+//    to finish: utf-8 checks and generating the output from the last iteration.
+//
+// The reason we run 2 inputs at a time, is steps 2 and 3 are *still* not enough to soak up all
+// available capacity with just one input. Running 2 at a time seems to give the CPU a good enough
+// workout.
+//
+template<size_t STEP_SIZE>
+error_code json_structural_indexer::index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept {
+  if (simdjson_unlikely(len > parser.capacity())) { return CAPACITY; }
+  // We guard the rest of the code so that we can assume that len > 0 throughout.
+  if (len == 0) { return EMPTY; }
+  if (is_streaming(partial)) {
+    len = trim_partial_utf8(buf, len);
+    // If you end up with an empty window after trimming
+    // the partial UTF-8 bytes, then chances are good that you
+    // have an UTF-8 formatting error.
+    if(len == 0) { return UTF8_ERROR; }
+  }
+  buf_block_reader<STEP_SIZE> reader(buf, len);
+  json_structural_indexer indexer(parser.structural_indexes.get());
+
+  // Read all but the last block
+  while (reader.has_full_block()) {
+    indexer.step<STEP_SIZE>(reader.full_block(), reader);
+  }
+  // Take care of the last block (will always be there unless file is empty which is
+  // not supposed to happen.)
+  uint8_t block[STEP_SIZE];
+  if (simdjson_unlikely(reader.get_remainder(block) == 0)) { return UNEXPECTED_ERROR; }
+  indexer.step<STEP_SIZE>(block, reader);
+  return indexer.finish(parser, reader.block_index(), len, partial);
+}
+
+template<>
+simdjson_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block);
+  simd::simd8x64<uint8_t> in_2(block+64);
+  json_block block_1 = scanner.next(in_1);
+  json_block block_2 = scanner.next(in_2);
+  this->next(in_1, block_1, reader.block_index());
+  this->next(in_2, block_2, reader.block_index()+64);
+  reader.advance();
+}
+
+template<>
+simdjson_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block);
+  json_block block_1 = scanner.next(in_1);
+  this->next(in_1, block_1, reader.block_index());
+  reader.advance();
+}
+
+simdjson_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
+  uint64_t unescaped = in.lteq(0x1F);
+#if SIMDJSON_UTF8VALIDATION
+  checker.check_next_input(in);
+#endif
+  indexer.write(uint32_t(idx-64), prev_structurals); // Output *last* iteration's structurals to the parser
+  prev_structurals = block.structural_start();
+  unescaped_chars_error |= block.non_quote_inside_string(unescaped);
+}
+
+simdjson_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
+  // Write out the final iteration's structurals
+  indexer.write(uint32_t(idx-64), prev_structurals);
+  error_code error = scanner.finish();
+  // We deliberately break down the next expression so that it is
+  // human readable.
+  const bool should_we_exit = is_streaming(partial) ?
+    ((error != SUCCESS) && (error != UNCLOSED_STRING)) // when partial we tolerate UNCLOSED_STRING
+    : (error != SUCCESS); // if partial is false, we must have SUCCESS
+  const bool have_unclosed_string = (error == UNCLOSED_STRING);
+  if (simdjson_unlikely(should_we_exit)) { return error; }
+
+  if (unescaped_chars_error) {
+    return UNESCAPED_CHARS;
+  }
+  parser.n_structural_indexes = uint32_t(indexer.tail - parser.structural_indexes.get());
+  /***
+   * The On-Demand API requires special padding.
+   *
+   * This is related to https://github.com/simdjson/simdjson/issues/906
+   * Basically, we want to make sure that if the parsing continues beyond the last (valid)
+   * structural character, it quickly stops.
+   * Only three structural characters can be repeated without triggering an error in JSON:  [,] and }.
+   * We repeat the padding character (at 'len'). We don't know what it is, but if the parsing
+   * continues, then it must be [,] or }.
+   * Suppose it is ] or }. We backtrack to the first character, what could it be that would
+   * not trigger an error? It could be ] or } but no, because you can't start a document that way.
+   * It can't be a comma, a colon or any simple value. So the only way we could continue is
+   * if the repeated character is [. But if so, the document must start with [. But if the document
+   * starts with [, it should end with ]. If we enforce that rule, then we would get
+   * ][[ which is invalid.
+   *
+   * This is illustrated with the test array_iterate_unclosed_error() on the following input:
+   * R"({ "a": [,,)"
+   **/
+  parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len); // used later in partial == stage1_mode::streaming_final
+  parser.structural_indexes[parser.n_structural_indexes + 1] = uint32_t(len);
+  parser.structural_indexes[parser.n_structural_indexes + 2] = 0;
+  parser.next_structural_index = 0;
+  // a valid JSON file cannot have zero structural indexes - we should have found something
+  if (simdjson_unlikely(parser.n_structural_indexes == 0u)) {
+    return EMPTY;
+  }
+  if (simdjson_unlikely(parser.structural_indexes[parser.n_structural_indexes - 1] > len)) {
+    return UNEXPECTED_ERROR;
+  }
+  if (partial == stage1_mode::streaming_partial) {
+    // If we have an unclosed string, then the last structural
+    // will be the quote and we want to make sure to omit it.
+    if(have_unclosed_string) {
+      parser.n_structural_indexes--;
+      // a valid JSON file cannot have zero structural indexes - we should have found something
+      if (simdjson_unlikely(parser.n_structural_indexes == 0u)) { return CAPACITY; }
+    }
+    // We truncate the input to the end of the last complete document (or zero).
+    auto new_structural_indexes = find_next_document_index(parser);
+    if (new_structural_indexes == 0 && parser.n_structural_indexes > 0) {
+      if(parser.structural_indexes[0] == 0) {
+        // If the buffer is partial and we started at index 0 but the document is
+        // incomplete, it's too big to parse.
+        return CAPACITY;
+      } else {
+        // It is possible that the document could be parsed, we just had a lot
+        // of white space.
+        parser.n_structural_indexes = 0;
+        return EMPTY;
+      }
+    }
+
+    parser.n_structural_indexes = new_structural_indexes;
+  } else if (partial == stage1_mode::streaming_final) {
+    if(have_unclosed_string) { parser.n_structural_indexes--; }
+    // We truncate the input to the end of the last complete document (or zero).
+    // Because partial == stage1_mode::streaming_final, it means that we may
+    // silently ignore trailing garbage. Though it sounds bad, we do it
+    // deliberately because many people who have streams of JSON documents
+    // will truncate them for processing. E.g., imagine that you are uncompressing
+    // the data from a size file or receiving it in chunks from the network. You
+    // may not know where exactly the last document will be. Meanwhile the
+    // document_stream instances allow people to know the JSON documents they are
+    // parsing (see the iterator.source() method).
+    parser.n_structural_indexes = find_next_document_index(parser);
+    // We store the initial n_structural_indexes so that the client can see
+    // whether we used truncation. If initial_n_structural_indexes == parser.n_structural_indexes,
+    // then this will query parser.structural_indexes[parser.n_structural_indexes] which is len,
+    // otherwise, it will copy some prior index.
+    parser.structural_indexes[parser.n_structural_indexes + 1] = parser.structural_indexes[parser.n_structural_indexes];
+    // This next line is critical, do not change it unless you understand what you are
+    // doing.
+    parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len);
+    if (simdjson_unlikely(parser.n_structural_indexes == 0u)) {
+        // We tolerate an unclosed string at the very end of the stream. Indeed, users
+        // often load their data in bulk without being careful and they want us to ignore
+        // the trailing garbage.
+        return EMPTY;
+    }
+  }
+  checker.check_eof();
+  return checker.errors();
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+// Clear CUSTOM_BIT_INDEXER so other implementations can set it if they need to.
+#undef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H
+/* end file generic/stage1/json_structural_indexer.h for lsx */
+/* including generic/stage1/utf8_validator.h for lsx: #include <generic/stage1/utf8_validator.h> */
+/* begin file generic/stage1/utf8_validator.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace stage1 {
+
+/**
+ * Validates that the string is actual UTF-8.
+ */
+template<class checker>
+bool generic_validate_utf8(const uint8_t * input, size_t length) {
+    checker c{};
+    buf_block_reader<64> reader(input, length);
+    while (reader.has_full_block()) {
+      simd::simd8x64<uint8_t> in(reader.full_block());
+      c.check_next_input(in);
+      reader.advance();
+    }
+    uint8_t block[64]{};
+    reader.get_remainder(block);
+    simd::simd8x64<uint8_t> in(block);
+    c.check_next_input(in);
+    reader.advance();
+    c.check_eof();
+    return c.errors() == error_code::SUCCESS;
+}
+
+bool generic_validate_utf8(const char * input, size_t length) {
+    return generic_validate_utf8<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H
+/* end file generic/stage1/utf8_validator.h for lsx */
+/* end file generic/stage1/amalgamated.h for lsx */
+/* including generic/stage2/amalgamated.h for lsx: #include <generic/stage2/amalgamated.h> */
+/* begin file generic/stage2/amalgamated.h for lsx */
+// Stuff other things depend on
+/* including generic/stage2/base.h for lsx: #include <generic/stage2/base.h> */
+/* begin file generic/stage2/base.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_BASE_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace stage2 {
+
+class json_iterator;
+class structural_iterator;
+struct tape_builder;
+struct tape_writer;
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_BASE_H
+/* end file generic/stage2/base.h for lsx */
+/* including generic/stage2/tape_writer.h for lsx: #include <generic/stage2/tape_writer.h> */
+/* begin file generic/stage2/tape_writer.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/internal/tape_type.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace stage2 {
+
+struct tape_writer {
+  /** The next place to write to tape */
+  uint64_t *next_tape_loc;
+
+  /** Write a signed 64-bit value to tape. */
+  simdjson_inline void append_s64(int64_t value) noexcept;
+
+  /** Write an unsigned 64-bit value to tape. */
+  simdjson_inline void append_u64(uint64_t value) noexcept;
+
+  /** Write a double value to tape. */
+  simdjson_inline void append_double(double value) noexcept;
+
+  /** Write a big integer (as string) to tape. src points to first digit, len is byte count. */
+  simdjson_inline void append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept;
+
+  /**
+   * Append a tape entry (an 8-bit type,and 56 bits worth of value).
+   */
+  simdjson_inline void append(uint64_t val, internal::tape_type t) noexcept;
+
+  /**
+   * Skip the current tape entry without writing.
+   *
+   * Used to skip the start of the container, since we'll come back later to fill it in when the
+   * container ends.
+   */
+  simdjson_inline void skip() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a large u64 or i64.
+   */
+  simdjson_inline void skip_large_integer() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a double.
+   */
+  simdjson_inline void skip_double() noexcept;
+
+  /**
+   * Write a value to a known location on tape.
+   *
+   * Used to go back and write out the start of a container after the container ends.
+   */
+  simdjson_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
+
+private:
+  /**
+   * Append both the tape entry, and a supplementary value following it. Used for types that need
+   * all 64 bits, such as double and uint64_t.
+   */
+  template<typename T>
+  simdjson_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
+}; // struct tape_writer
+
+simdjson_inline void tape_writer::append_s64(int64_t value) noexcept {
+  append2(0, value, internal::tape_type::INT64);
+}
+
+simdjson_inline void tape_writer::append_u64(uint64_t value) noexcept {
+  append(0, internal::tape_type::UINT64);
+  *next_tape_loc = value;
+  next_tape_loc++;
+}
+
+/** Write a double value to tape. */
+simdjson_inline void tape_writer::append_double(double value) noexcept {
+  append2(0, value, internal::tape_type::DOUBLE);
+}
+
+simdjson_inline void tape_writer::skip() noexcept {
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::skip_large_integer() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::skip_double() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
+  *next_tape_loc = val | ((uint64_t(char(t))) << 56);
+  next_tape_loc++;
+}
+
+template<typename T>
+simdjson_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
+  append(val, t);
+  static_assert(sizeof(val2) == sizeof(*next_tape_loc), "Type is not 64 bits!");
+  memcpy(next_tape_loc, &val2, sizeof(val2));
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
+  tape_loc = val | ((uint64_t(char(t))) << 56);
+}
+
+simdjson_inline void tape_writer::append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept {
+  // Write to string buffer: [4-byte LE length][digits][null]
+  uint32_t str_len = uint32_t(len);
+  memcpy(string_buf, &str_len, sizeof(uint32_t));
+  memcpy(string_buf + sizeof(uint32_t), src, len);
+  string_buf[sizeof(uint32_t) + len] = 0;
+  // Tape entry: offset into string buffer
+  // The caller must set the offset relative to doc.string_buf base
+  append(0, internal::tape_type::BIGINT); // placeholder offset, caller patches
+  string_buf += sizeof(uint32_t) + len + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+/* end file generic/stage2/tape_writer.h for lsx */
+/* including generic/stage2/logger.h for lsx: #include <generic/stage2/logger.h> */
+/* begin file generic/stage2/logger.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+
+// This is for an internal-only stage 2 specific logger.
+// Set LOG_ENABLED = true to log what stage 2 is doing!
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace logger {
+
+  static constexpr const char * DASHES = "----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------";
+
+#if SIMDJSON_VERBOSE_LOGGING
+  static constexpr const bool LOG_ENABLED = true;
+#else
+  static constexpr const bool LOG_ENABLED = false;
+#endif
+  static constexpr const int LOG_EVENT_LEN = 20;
+  static constexpr const int LOG_BUFFER_LEN = 30;
+  static constexpr const int LOG_SMALL_BUFFER_LEN = 10;
+  static constexpr const int LOG_INDEX_LEN = 5;
+
+  static int log_depth; // Not threadsafe. Log only.
+
+  // Helper to turn unprintable or newline characters into spaces
+  static simdjson_inline char printable_char(char c) {
+    if (c >= 0x20) {
+      return c;
+    } else {
+      return ' ';
+    }
+  }
+
+  // Print the header and set up log_start
+  static simdjson_inline void log_start() {
+    if (LOG_ENABLED) {
+      log_depth = 0;
+      printf("\n");
+      printf("| %-*s | %-*s | %-*s | %-*s | Detail |\n", LOG_EVENT_LEN, "Event", LOG_BUFFER_LEN, "Buffer", LOG_SMALL_BUFFER_LEN, "Next", 5, "Next#");
+      printf("|%.*s|%.*s|%.*s|%.*s|--------|\n", LOG_EVENT_LEN+2, DASHES, LOG_BUFFER_LEN+2, DASHES, LOG_SMALL_BUFFER_LEN+2, DASHES, 5+2, DASHES);
+    }
+  }
+
+  simdjson_unused static simdjson_inline void log_string(const char *message) {
+    if (LOG_ENABLED) {
+      printf("%s\n", message);
+    }
+  }
+
+  // Logs a single line from the stage 2 DOM parser
+  template<typename S>
+  static simdjson_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
+    if (LOG_ENABLED) {
+      printf("| %*s%s%-*s ", log_depth*2, "", title_prefix, LOG_EVENT_LEN - log_depth*2 - int(strlen(title_prefix)), title);
+      auto current_index = structurals.at_beginning() ? nullptr : structurals.next_structural-1;
+      auto next_index = structurals.next_structural;
+      auto current = current_index ? &structurals.buf[*current_index] : reinterpret_cast<const uint8_t*>("                                                       ");
+      auto next = &structurals.buf[*next_index];
+      {
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_BUFFER_LEN;i++) {
+          printf("%c", printable_char(current[i]));
+        }
+        printf(" ");
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_SMALL_BUFFER_LEN;i++) {
+          printf("%c", printable_char(next[i]));
+        }
+        printf(" ");
+      }
+      if (current_index) {
+        printf("| %*u ", LOG_INDEX_LEN, *current_index);
+      } else {
+        printf("| %-*s ", LOG_INDEX_LEN, "");
+      }
+      // printf("| %*u ", LOG_INDEX_LEN, structurals.next_tape_index());
+      printf("| %-s ", detail);
+      printf("|\n");
+    }
+  }
+
+} // namespace logger
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+/* end file generic/stage2/logger.h for lsx */
+
+// All other declarations
+/* including generic/stage2/json_iterator.h for lsx: #include <generic/stage2/json_iterator.h> */
+/* begin file generic/stage2/json_iterator.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/logger.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace stage2 {
+
+class json_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+  uint32_t depth{0};
+
+  /**
+   * Walk the JSON document.
+   *
+   * The visitor receives callbacks when values are encountered. All callbacks pass the iterator as
+   * the first parameter; some callbacks have other parameters as well:
+   *
+   * - visit_document_start() - at the beginning.
+   * - visit_document_end() - at the end (if things were successful).
+   *
+   * - visit_array_start() - at the start `[` of a non-empty array.
+   * - visit_array_end() - at the end `]` of a non-empty array.
+   * - visit_empty_array() - when an empty array is encountered.
+   *
+   * - visit_object_end() - at the start `]` of a non-empty object.
+   * - visit_object_start() - at the end `]` of a non-empty object.
+   * - visit_empty_object() - when an empty object is encountered.
+   * - visit_key(const uint8_t *key) - when a key in an object field is encountered. key is
+   *                                   guaranteed to point at the first quote of the string (`"key"`).
+   * - visit_primitive(const uint8_t *value) - when a value is a string, number, boolean or null.
+   * - visit_root_primitive(iter, uint8_t *value) - when the top-level value is a string, number, boolean or null.
+   *
+   * - increment_count(iter) - each time a value is found in an array or object.
+   */
+  template<bool STREAMING, typename V>
+  simdjson_warn_unused simdjson_inline error_code walk_document(V &visitor) noexcept;
+
+  /**
+   * Create an iterator capable of walking a JSON document.
+   *
+   * The document must have already passed through stage 1.
+   */
+  simdjson_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
+
+  /**
+   * Look at the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *peek() const noexcept;
+  /**
+   * Advance to the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *advance() noexcept;
+  /**
+   * Get the remaining length of the document, from the start of the current token.
+   */
+  simdjson_inline size_t remaining_len() const noexcept;
+  /**
+   * Check if we are at the end of the document.
+   *
+   * If this is true, there are no more tokens.
+   */
+  simdjson_inline bool at_eof() const noexcept;
+  /**
+   * Check if we are at the beginning of the document.
+   */
+  simdjson_inline bool at_beginning() const noexcept;
+  simdjson_inline uint8_t last_structural() const noexcept;
+
+  /**
+   * Log that a value has been found.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_value(const char *type) const noexcept;
+  /**
+   * Log the start of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_start_value(const char *type) const noexcept;
+  /**
+   * Log the end of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_end_value(const char *type) const noexcept;
+  /**
+   * Log an error.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_error(const char *error) const noexcept;
+
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
+};
+
+template<bool STREAMING, typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::walk_document(V &visitor) noexcept {
+  logger::log_start();
+
+  //
+  // Start the document
+  //
+  if (at_eof()) { return EMPTY; }
+  log_start_value("document");
+  SIMDJSON_TRY( visitor.visit_document_start(*this) );
+
+  //
+  // Read first value
+  //
+  {
+    auto value = advance();
+
+    // Make sure the outer object or array is closed before continuing; otherwise, there are ways we
+    // could get into memory corruption. See https://github.com/simdjson/simdjson/issues/906
+    if (!STREAMING) {
+      switch (*value) {
+        case '{': if (last_structural() != '}') { log_value("starting brace unmatched"); return TAPE_ERROR; }; break;
+        case '[': if (last_structural() != ']') { log_value("starting bracket unmatched"); return TAPE_ERROR; }; break;
+      }
+    }
+
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_root_primitive(*this, value) ); break;
+    }
+  }
+  goto document_end;
+
+//
+// Object parser states
+//
+object_begin:
+  log_start_value("object");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = false;
+  SIMDJSON_TRY( visitor.visit_object_start(*this) );
+
+  {
+    auto key = advance();
+    if (*key != '"') { log_error("Object does not start with a key"); return TAPE_ERROR; }
+    SIMDJSON_TRY( visitor.increment_count(*this) );
+    SIMDJSON_TRY( visitor.visit_key(*this, key) );
+  }
+
+object_field:
+  if (simdjson_unlikely( *advance() != ':' )) { log_error("Missing colon after key in object"); return TAPE_ERROR; }
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+object_continue:
+  switch (*advance()) {
+    case ',':
+      SIMDJSON_TRY( visitor.increment_count(*this) );
+      {
+        auto key = advance();
+        if (simdjson_unlikely( *key != '"' )) { log_error("Key string missing at beginning of field in object"); return TAPE_ERROR; }
+        SIMDJSON_TRY( visitor.visit_key(*this, key) );
+      }
+      goto object_field;
+    case '}': log_end_value("object"); SIMDJSON_TRY( visitor.visit_object_end(*this) ); goto scope_end;
+    default: log_error("No comma between object fields"); return TAPE_ERROR;
+  }
+
+scope_end:
+  depth--;
+  if (depth == 0) { goto document_end; }
+  if (dom_parser.is_array[depth]) { goto array_continue; }
+  goto object_continue;
+
+//
+// Array parser states
+//
+array_begin:
+  log_start_value("array");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = true;
+  SIMDJSON_TRY( visitor.visit_array_start(*this) );
+  SIMDJSON_TRY( visitor.increment_count(*this) );
+
+array_value:
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+array_continue:
+  switch (*advance()) {
+    case ',': SIMDJSON_TRY( visitor.increment_count(*this) ); goto array_value;
+    case ']': log_end_value("array"); SIMDJSON_TRY( visitor.visit_array_end(*this) ); goto scope_end;
+    default: log_error("Missing comma between array values"); return TAPE_ERROR;
+  }
+
+document_end:
+  log_end_value("document");
+  SIMDJSON_TRY( visitor.visit_document_end(*this) );
+
+  dom_parser.next_structural_index = uint32_t(next_structural - &dom_parser.structural_indexes[0]);
+
+  // If we didn't make it to the end, it's an error
+  if ( !STREAMING && dom_parser.next_structural_index != dom_parser.n_structural_indexes ) {
+    log_error("More than one JSON value at the root of the document, or extra characters at the end of the JSON!");
+    return TAPE_ERROR;
+  }
+
+  return SUCCESS;
+
+} // walk_document()
+
+simdjson_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+  : buf{_dom_parser.buf},
+    next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+    dom_parser{_dom_parser} {
+}
+
+simdjson_inline const uint8_t *json_iterator::peek() const noexcept {
+  return &buf[*(next_structural)];
+}
+simdjson_inline const uint8_t *json_iterator::advance() noexcept {
+  return &buf[*(next_structural++)];
+}
+simdjson_inline size_t json_iterator::remaining_len() const noexcept {
+  return dom_parser.len - *(next_structural-1);
+}
+
+simdjson_inline bool json_iterator::at_eof() const noexcept {
+  return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+}
+simdjson_inline bool json_iterator::at_beginning() const noexcept {
+  return next_structural == dom_parser.structural_indexes.get();
+}
+simdjson_inline uint8_t json_iterator::last_structural() const noexcept {
+  return buf[dom_parser.structural_indexes[dom_parser.n_structural_indexes - 1]];
+}
+
+simdjson_inline void json_iterator::log_value(const char *type) const noexcept {
+  logger::log_line(*this, "", type, "");
+}
+
+simdjson_inline void json_iterator::log_start_value(const char *type) const noexcept {
+  logger::log_line(*this, "+", type, "");
+  if (logger::LOG_ENABLED) { logger::log_depth++; }
+}
+
+simdjson_inline void json_iterator::log_end_value(const char *type) const noexcept {
+  if (logger::LOG_ENABLED) { logger::log_depth--; }
+  logger::log_line(*this, "-", type, "");
+}
+
+simdjson_inline void json_iterator::log_error(const char *error) const noexcept {
+  logger::log_line(*this, "", "ERROR", error);
+}
+
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
+  switch (*value) {
+    case '"': return visitor.visit_root_string(*this, value);
+    case 't': return visitor.visit_root_true_atom(*this, value);
+    case 'f': return visitor.visit_root_false_atom(*this, value);
+    case 'n': return visitor.visit_root_null_atom(*this, value);
+    case '-':
+    case '0': case '1': case '2': case '3': case '4':
+    case '5': case '6': case '7': case '8': case '9':
+      return visitor.visit_root_number(*this, value);
+    default:
+      log_error("Document starts with a non-value character");
+      return TAPE_ERROR;
+  }
+}
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
+  // Use the fact that most scalars are going to be either strings or numbers.
+  if(*value == '"') {
+    return visitor.visit_string(*this, value);
+  } else if (((*value - '0')  < 10) || (*value == '-')) {
+    return visitor.visit_number(*this, value);
+  }
+  // true, false, null are uncommon.
+  switch (*value) {
+    case 't': return visitor.visit_true_atom(*this, value);
+    case 'f': return visitor.visit_false_atom(*this, value);
+    case 'n': return visitor.visit_null_atom(*this, value);
+    default:
+      log_error("Non-value found when value was expected!");
+      return TAPE_ERROR;
+  }
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+/* end file generic/stage2/json_iterator.h for lsx */
+/* including generic/stage2/stringparsing.h for lsx: #include <generic/stage2/stringparsing.h> */
+/* begin file generic/stage2/stringparsing.h for lsx */
+#include <cstdint>
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/jsoncharutils.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses
+// It is intended to be included multiple times and compiled multiple times
+
+namespace simdjson {
+namespace lsx {
+namespace {
+/// @private
+namespace stringparsing {
+
+// begin copypasta
+// These chars yield themselves: " \ /
+// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
+// u not handled in this table as it's complex
+static const uint8_t escape_map[256] = {
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
+    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
+    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+};
+
+// handle a unicode codepoint
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr, bool allow_replacement) {
+  // Use the default Unicode Character 'REPLACEMENT CHARACTER' (U+FFFD)
+  constexpr uint32_t substitution_code_point = 0xfffd;
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) != ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+    } else {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+
+      // We have already checked that the high surrogate is valid and
+      // (code_point - 0xd800) < 1024.
+      //
+      // Check that code_point_2 is in the range 0xdc00..0xdfff
+      // and that code_point_2 was parsed from valid hex.
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if (low_bit >> 10) {
+        if(!allow_replacement) { return false; }
+        code_point = substitution_code_point;
+      } else {
+        code_point =  (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+
+    }
+  } else if (code_point >= 0xdc00 && code_point <= 0xdfff) {
+      // If we encounter a low surrogate (not preceded by a high surrogate)
+      // then we have an error.
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+  }
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+// handle a unicode codepoint using the wobbly convention
+// https://simonsapin.github.io/wtf-8/
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint_wobbly(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr) {
+  // It is not ideal that this function is nearly identical to handle_unicode_codepoint.
+  //
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) == ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if ((low_bit >> 10) ==  0) {
+        code_point =
+          (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+    }
+  }
+
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+/**
+ * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+ * must be an unescaped quote terminating the string. It returns the final output
+ * position as pointer. In case of error (e.g., the string has bad escaped codes),
+ * then null_ptr is returned. It is assumed that the output buffer is large
+ * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+ * SIMDJSON_PADDING bytes.
+ */
+simdjson_warn_unused simdjson_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) {
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint(&src, &dst, allow_replacement)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+simdjson_warn_unused simdjson_inline uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) {
+  // It is not ideal that this function is nearly identical to parse_string.
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint_wobbly(&src, &dst)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+} // namespace stringparsing
+
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+/* end file generic/stage2/stringparsing.h for lsx */
+/* including generic/stage2/structural_iterator.h for lsx: #include <generic/stage2/structural_iterator.h> */
+/* begin file generic/stage2/structural_iterator.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace stage2 {
+
+class structural_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+
+  // Start a structural
+  simdjson_inline structural_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+    : buf{_dom_parser.buf},
+      next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+      dom_parser{_dom_parser} {
+  }
+  // Get the buffer position of the current structural character
+  simdjson_inline const uint8_t* current() {
+    return &buf[*(next_structural-1)];
+  }
+  // Get the current structural character
+  simdjson_inline char current_char() {
+    return buf[*(next_structural-1)];
+  }
+  // Get the next structural character without advancing
+  simdjson_inline char peek_next_char() {
+    return buf[*next_structural];
+  }
+  simdjson_inline const uint8_t* peek() {
+    return &buf[*next_structural];
+  }
+  simdjson_inline const uint8_t* advance() {
+    return &buf[*(next_structural++)];
+  }
+  simdjson_inline char advance_char() {
+    return buf[*(next_structural++)];
+  }
+  simdjson_inline size_t remaining_len() {
+    return dom_parser.len - *(next_structural-1);
+  }
+
+  simdjson_inline bool at_end() {
+    return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+  }
+  simdjson_inline bool at_beginning() {
+    return next_structural == dom_parser.structural_indexes.get();
+  }
+};
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H
+/* end file generic/stage2/structural_iterator.h for lsx */
+/* including generic/stage2/tape_builder.h for lsx: #include <generic/stage2/tape_builder.h> */
+/* begin file generic/stage2/tape_builder.h for lsx */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/json_iterator.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/stringparsing.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/tape_writer.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/dom/document.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/atomparsing.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/numberparsing.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+
+namespace simdjson {
+namespace lsx {
+namespace {
+namespace stage2 {
+
+struct tape_builder {
+  template<bool STREAMING>
+  simdjson_warn_unused static simdjson_inline error_code parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept;
+
+  /** Called when a non-empty document starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty document ends without error. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_end(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty array starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty array ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_end(json_iterator &iter) noexcept;
+  /** Called when an empty array is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_array(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty object starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_start(json_iterator &iter) noexcept;
+  /**
+   * Called when a key in a field is encountered.
+   *
+   * primitive, visit_object_start, visit_empty_object, visit_array_start, or visit_empty_array
+   * will be called after this with the field value.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
+  /** Called when a non-empty object ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_end(json_iterator &iter) noexcept;
+  /** Called when an empty object is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_object(json_iterator &iter) noexcept;
+
+  /**
+   * Called when a string, number, boolean or null is found.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  /**
+   * Called when a string, number, boolean or null is found at the top level of a document (i.e.
+   * when there is no array or object and the entire document is a single string, number, boolean or
+   * null.
+   *
+   * This is separate from primitive() because simdjson's normal primitive parsing routines assume
+   * there is at least one more token after the value, which is only true in an array or object.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  /** Called each time a new field or element in an array or object is found. */
+  simdjson_warn_unused simdjson_inline error_code increment_count(json_iterator &iter) noexcept;
+
+  /** Next location to write to tape */
+  tape_writer tape;
+private:
+  /** Next write location in the string buf for stage 2 parsing */
+  uint8_t *current_string_buf_loc;
+
+  simdjson_inline tape_builder(dom::document &doc) noexcept;
+
+  simdjson_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
+  simdjson_inline void start_container(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_warn_unused simdjson_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
+  simdjson_inline void on_end_string(uint8_t *dst) noexcept;
+}; // struct tape_builder
+
+template<bool STREAMING>
+simdjson_warn_unused simdjson_inline error_code tape_builder::parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept {
+  dom_parser.doc = &doc;
+  json_iterator iter(dom_parser, STREAMING ? dom_parser.next_structural_index : 0);
+  tape_builder builder(doc);
+  return iter.walk_document<STREAMING>(builder);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_root_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
+  constexpr uint32_t start_tape_index = 0;
+  tape.append(start_tape_index, internal::tape_type::ROOT);
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter), internal::tape_type::ROOT);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
+  return visit_string(iter, key, true);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].count++; // we have a key value pair in the object at parser.dom_parser.depth - 1
+  return SUCCESS;
+}
+
+simdjson_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
+  iter.log_value(key ? "key" : "string");
+  uint8_t *dst = on_start_string(iter);
+  dst = stringparsing::parse_string(value+1, dst, false); // We do not allow replacement when the escape characters are invalid.
+  if (dst == nullptr) {
+    iter.log_error("Invalid escape in string");
+    return STRING_ERROR;
+  }
+  on_end_string(dst);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
+  return visit_string(iter, value);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("number");
+  error_code err = numberparsing::parse_number(value, tape);
+  if (simdjson_unlikely(err == BIGINT_ERROR &&
+      iter.dom_parser._number_as_string)) {
+    // Write big integer to string buffer using the same format as strings.
+    // Scan digits the same way parse_number does (skip optional '-', then digits).
+    const uint8_t *p = value;
+    if (*p == '-') p++;
+    while (numberparsing::is_digit(*p)) p++;
+    size_t len = size_t(p - value);
+    tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::BIGINT);
+    uint8_t *dst = current_string_buf_loc + sizeof(uint32_t);
+    memcpy(dst, value, len);
+    dst += len;
+    on_end_string(dst);
+    return SUCCESS;
+  }
+  return err;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
+  //
+  // We need to make a copy to make sure that the string is space terminated.
+  // This is not about padding the input, which should already padded up
+  // to len + SIMDJSON_PADDING. However, we have no control at this stage
+  // on how the padding was done. What if the input string was padded with nulls?
+  // It is quite common for an input string to have an extra null character (C string).
+  // We do not want to allow 9\0 (where \0 is the null character) inside a JSON
+  // document, but the string "9\0" by itself is fine. So we make a copy and
+  // pad the input with spaces when we know that there is just one input element.
+  // This copy is relatively expensive, but it will almost never be called in
+  // practice unless you are in the strange scenario where you have many JSON
+  // documents made of single atoms.
+  //
+  std::unique_ptr<uint8_t[]>copy(new (std::nothrow) uint8_t[iter.remaining_len() + SIMDJSON_PADDING]);
+  if (copy.get() == nullptr) { return MEMALLOC; }
+  std::memcpy(copy.get(), value, iter.remaining_len());
+  std::memset(copy.get() + iter.remaining_len(), ' ', SIMDJSON_PADDING);
+  error_code error = visit_number(iter, copy.get());
+  return error;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value)) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value, iter.remaining_len())) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value)) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value, iter.remaining_len())) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value)) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value, iter.remaining_len())) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+// private:
+
+simdjson_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
+  return uint32_t(tape.next_tape_loc - iter.dom_parser.doc->tape.get());
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  auto start_index = next_tape_index(iter);
+  tape.append(start_index+2, start);
+  tape.append(start_index, end);
+  return SUCCESS;
+}
+
+simdjson_inline void tape_builder::start_container(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].tape_index = next_tape_index(iter);
+  iter.dom_parser.open_containers[iter.depth].count = 0;
+  tape.skip(); // We don't actually *write* the start element until the end.
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  // Write the ending tape element, pointing at the start location
+  const uint32_t start_tape_index = iter.dom_parser.open_containers[iter.depth].tape_index;
+  tape.append(start_tape_index, end);
+  // Write the start tape element, pointing at the end location (and including count)
+  // count can overflow if it exceeds 24 bits... so we saturate
+  // the convention being that a cnt of 0xffffff or more is undetermined in value (>=  0xffffff).
+  const uint32_t count = iter.dom_parser.open_containers[iter.depth].count;
+  const uint32_t cntsat = count > 0xFFFFFF ? 0xFFFFFF : count;
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter) | (uint64_t(cntsat) << 32), start);
+  return SUCCESS;
+}
+
+simdjson_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
+  // we advance the point, accounting for the fact that we have a NULL termination
+  tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::STRING);
+  return current_string_buf_loc + sizeof(uint32_t);
+}
+
+simdjson_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
+  uint32_t str_length = uint32_t(dst - (current_string_buf_loc + sizeof(uint32_t)));
+  // TODO check for overflow in case someone has a crazy string (>=4GB?)
+  // But only add the overflow check when the document itself exceeds 4GB
+  // Currently unneeded because we refuse to parse docs larger or equal to 4GB.
+  memcpy(current_string_buf_loc, &str_length, sizeof(uint32_t));
+  // NULL termination is still handy if you expect all your strings to
+  // be NULL terminated? It comes at a small cost
+  *dst = 0;
+  current_string_buf_loc = dst + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+/* end file generic/stage2/tape_builder.h for lsx */
+/* end file generic/stage2/amalgamated.h for lsx */
+
+//
+// Stage 1
+//
+namespace simdjson {
+namespace lsx {
+
+simdjson_warn_unused error_code implementation::create_dom_parser_implementation(
+  size_t capacity,
+  size_t max_depth,
+  std::unique_ptr<internal::dom_parser_implementation>& dst
+) const noexcept {
+  dst.reset( new (std::nothrow) dom_parser_implementation() );
+  if (!dst) { return MEMALLOC; }
+  if (auto err = dst->set_capacity(capacity))
+    return err;
+  if (auto err = dst->set_max_depth(max_depth))
+    return err;
+  return SUCCESS;
+}
+
+namespace {
+
+using namespace simd;
+
+simdjson_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
+  // Inspired by haswell.
+  // LSX use low 5 bits as index. For the 6 operators (:,[]{}), the unique-5bits is [6:2].
+  // The ASCII white-space and operators have these values: (char, hex, unique-5bits)
+  // (' ', 20, 00000) ('\t', 09, 01001) ('\n', 0A, 01010) ('\r', 0D, 01101)
+  // (',', 2C, 01011) (':', 3A, 01110) ('[', 5B, 10110) ('{', 7B, 11110) (']', 5D, 10111) ('}', 7D, 11111)
+  const simd8<uint8_t> ws_table = simd8<uint8_t>::repeat_16(
+    ' ', 0, 0, 0, 0, 0, 0, 0, 0, '\t', '\n', 0, 0, '\r', 0, 0
+  );
+  const simd8<uint8_t> op_table_lo = simd8<uint8_t>::repeat_16(
+    1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ',', 0, 0, ':', 0
+  );
+  const simd8<uint8_t> op_table_hi = simd8<uint8_t>::repeat_16(
+    0, 0, 0, 0, 0, 0, '[', ']', 0, 0, 0, 0, 0, 0, '{', '}'
+  );
+  uint64_t ws = in.eq({
+    in.chunks[0].lookup_16(ws_table),
+    in.chunks[1].lookup_16(ws_table),
+    in.chunks[2].lookup_16(ws_table),
+    in.chunks[3].lookup_16(ws_table)
+  });
+  uint64_t op = in.eq({
+    __lsx_vshuf_b(op_table_hi, op_table_lo, in.chunks[0].shr<2>()),
+    __lsx_vshuf_b(op_table_hi, op_table_lo, in.chunks[1].shr<2>()),
+    __lsx_vshuf_b(op_table_hi, op_table_lo, in.chunks[2].shr<2>()),
+    __lsx_vshuf_b(op_table_hi, op_table_lo, in.chunks[3].shr<2>())
+  });
+
+  return { ws, op };
+}
+
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input) {
+  return input.reduce_or().is_ascii();
+}
+
+simdjson_inline simd8<uint8_t> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+    simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-0x80); // Only 111_____ will be >= 0x80
+    simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-0x80); // Only 1111____ will be >= 0x80
+    return is_third_byte | is_fourth_byte;
+}
+
+} // unnamed namespace
+} // namespace lsx
+} // namespace simdjson
+
+//
+// Stage 2
+//
+
+//
+// Implementation-specific overrides
+//
+namespace simdjson {
+namespace lsx {
+
+simdjson_warn_unused error_code implementation::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept {
+  return lsx::stage1::json_minifier::minify<64>(buf, len, dst, dst_len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage1(const uint8_t *_buf, size_t _len, stage1_mode streaming) noexcept {
+  this->buf = _buf;
+  this->len = _len;
+  return lsx::stage1::json_structural_indexer::index<64>(buf, len, *this, streaming);
+}
+
+simdjson_warn_unused bool implementation::validate_utf8(const char *buf, size_t len) const noexcept {
+  return lsx::stage1::generic_validate_utf8(buf,len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<false>(*this, _doc);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2_next(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<true>(*this, _doc);
+}
+
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) const noexcept {
+  return lsx::stringparsing::parse_string(src, dst, allow_replacement);
+}
+
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept {
+  return lsx::stringparsing::parse_wobbly_string(src, dst);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::parse(const uint8_t *_buf, size_t _len, dom::document &_doc) noexcept {
+  auto error = stage1(_buf, _len, stage1_mode::regular);
+  if (error) { return error; }
+  return stage2(_doc);
+}
+
+} // namespace lsx
+} // namespace simdjson
+
+/* including simdjson/lsx/end.h: #include <simdjson/lsx/end.h> */
+/* begin file simdjson/lsx/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#undef SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+/* undefining SIMDJSON_IMPLEMENTATION from "lsx" */
+#undef SIMDJSON_IMPLEMENTATION
+/* end file simdjson/lsx/end.h */
+
+#endif // SIMDJSON_SRC_LSX_CPP
+/* end file lsx.cpp */
+#endif
+#if SIMDJSON_IMPLEMENTATION_RVV_VLS
+/* including rvv-vls.cpp: #include <rvv-vls.cpp> */
+/* begin file rvv-vls.cpp */
+#ifndef SIMDJSON_SRC_RVV_VLS_CPP
+#define SIMDJSON_SRC_RVV_VLS_CPP
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+/* including simdjson/rvv-vls.h: #include <simdjson/rvv-vls.h> */
+/* begin file simdjson/rvv-vls.h */
+#ifndef SIMDJSON_RVV_VLS_H
+#define SIMDJSON_RVV_VLS_H
+
+
+/* including simdjson/rvv-vls/begin.h: #include "simdjson/rvv-vls/begin.h" */
+/* begin file simdjson/rvv-vls/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "rvv_vls" */
+#define SIMDJSON_IMPLEMENTATION rvv_vls
+/* including simdjson/rvv-vls/base.h: #include "simdjson/rvv-vls/base.h" */
+/* begin file simdjson/rvv-vls/base.h */
+#ifndef SIMDJSON_RVV_VLS_BASE_H
+#define SIMDJSON_RVV_VLS_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+/**
+ * RVV-VLS implementation.
+ */
+namespace rvv_vls {
+
+class implementation;
+
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_RVV_VLS_BASE_H
+/* end file simdjson/rvv-vls/base.h */
+/* including simdjson/rvv-vls/intrinsics.h: #include "simdjson/rvv-vls/intrinsics.h" */
+/* begin file simdjson/rvv-vls/intrinsics.h */
+#ifndef SIMDJSON_RVV_VLS_INTRINSICS_H
+#define SIMDJSON_RVV_VLS_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <riscv_vector.h>
+
+#define simdutf_vrgather_u8m1x2(tbl, idx)                                      \
+  __riscv_vcreate_v_u8m1_u8m2(                                                 \
+      __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m2_u8m1(idx, 0),          \
+                               __riscv_vsetvlmax_e8m1()),                      \
+      __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m2_u8m1(idx, 1),          \
+                               __riscv_vsetvlmax_e8m1()))
+
+#define simdutf_vrgather_u8m1x4(tbl, idx)                                      \
+  __riscv_vcreate_v_u8m1_u8m4(                                                 \
+      __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 0),          \
+                               __riscv_vsetvlmax_e8m1()),                      \
+      __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 1),          \
+                               __riscv_vsetvlmax_e8m1()),                      \
+      __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 2),          \
+                               __riscv_vsetvlmax_e8m1()),                      \
+      __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 3),          \
+                               __riscv_vsetvlmax_e8m1()))
+
+#if __riscv_zbc
+#include <riscv_bitmanip.h>
+#endif
+
+#endif // SIMDJSON_RVV_VLS_INTRINSICS_H
+/* end file simdjson/rvv-vls/intrinsics.h */
+/* including simdjson/rvv-vls/bitmanipulation.h: #include "simdjson/rvv-vls/bitmanipulation.h" */
+/* begin file simdjson/rvv-vls/bitmanipulation.h */
+#ifndef SIMDJSON_RVV_VLS_BITMANIPULATION_H
+#define SIMDJSON_RVV_VLS_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+  return __builtin_ctzll(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return input_num & (input_num-1);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+  return __builtin_clzll(input_num);
+}
+
+simdjson_inline long long int count_ones(uint64_t input_num) {
+  return __builtin_popcountll(input_num);
+}
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2,
+                                uint64_t *result) {
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+}
+
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_RVV_VLS_BITMANIPULATION_H
+/* end file simdjson/rvv-vls/bitmanipulation.h */
+/* including simdjson/rvv-vls/bitmask.h: #include "simdjson/rvv-vls/bitmask.h" */
+/* begin file simdjson/rvv-vls/bitmask.h */
+#ifndef SIMDJSON_RVV_VLS_BITMASK_H
+#define SIMDJSON_RVV_VLS_BITMASK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(uint64_t bitmask) {
+#if __riscv_zbc
+  return __riscv_clmul_64(bitmask, ~(uint64_t)0);
+#elif __riscv_zvbc
+  return __riscv_vmv_x(__riscv_vclmul(__riscv_vmv_s_x_u64m1(bitmask, 1), ~(uint64_t)0, 1));
+#else
+  bitmask ^= bitmask << 1;
+  bitmask ^= bitmask << 2;
+  bitmask ^= bitmask << 4;
+  bitmask ^= bitmask << 8;
+  bitmask ^= bitmask << 16;
+  bitmask ^= bitmask << 32;
+#endif
+  return bitmask;
+}
+
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_RVV_VLS_BITMASK_H
+
+/* end file simdjson/rvv-vls/bitmask.h */
+/* including simdjson/rvv-vls/simd.h: #include "simdjson/rvv-vls/simd.h" */
+/* begin file simdjson/rvv-vls/simd.h */
+#ifndef SIMDJSON_RVV_VLS_SIMD_H
+#define SIMDJSON_RVV_VLS_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace simd {
+
+#if __riscv_v_fixed_vlen >= 512
+  static constexpr size_t VL8 = 512/8;
+  using vint8_t   = vint8m1_t   __attribute__((riscv_rvv_vector_bits(512)));
+  using vuint8_t  = vuint8m1_t  __attribute__((riscv_rvv_vector_bits(512)));
+  using vbool_t    = vbool8_t  __attribute__((riscv_rvv_vector_bits(512/8)));
+  using vbitmask_t = uint64_t;
+#else
+  static constexpr size_t VL8 = __riscv_v_fixed_vlen/8;
+  using vint8_t   = vint8m1_t   __attribute__((riscv_rvv_vector_bits(__riscv_v_fixed_vlen)));
+  using vuint8_t  = vuint8m1_t  __attribute__((riscv_rvv_vector_bits(__riscv_v_fixed_vlen)));
+  using vbool_t    = vbool8_t  __attribute__((riscv_rvv_vector_bits(__riscv_v_fixed_vlen/8)));
+  #if __riscv_v_fixed_vlen == 128
+    using vbitmask_t = uint16_t;
+  #elif __riscv_v_fixed_vlen == 256
+    using vbitmask_t = uint32_t;
+  #endif
+#endif
+
+#if __riscv_v_fixed_vlen == 128
+  using vuint8x64_t = vuint8m4_t __attribute__((riscv_rvv_vector_bits(512)));
+  using vboolx64_t    = vbool2_t  __attribute__((riscv_rvv_vector_bits(512/8)));
+#elif __riscv_v_fixed_vlen == 256
+  using vuint8x64_t = vuint8m2_t __attribute__((riscv_rvv_vector_bits(512)));
+  using vboolx64_t    = vbool4_t  __attribute__((riscv_rvv_vector_bits(512/8)));
+#else
+  using vuint8x64_t = vuint8m1_t __attribute__((riscv_rvv_vector_bits(512)));
+  using vboolx64_t    = vbool8_t  __attribute__((riscv_rvv_vector_bits(512/8)));
+#endif
+
+  template<typename T>
+  struct simd8;
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool> {
+    vbool_t value;
+    using bitmask_t = vbitmask_t;
+    static constexpr int SIZE = sizeof(value);
+
+    simdjson_inline simd8(const vbool_t _value) : value(_value) {}
+    simdjson_inline simd8() : simd8(__riscv_vmclr_m_b8(VL8)) {}
+    simdjson_inline simd8(bool _value) : simd8(splat(_value)) {}
+
+    simdjson_inline operator const vbool_t&() const { return value; }
+    simdjson_inline operator vbool_t&() { return value; }
+
+    static simdjson_inline simd8<bool> splat(bool _value) {
+      return __riscv_vreinterpret_b8(__riscv_vmv_v_x_u64m1(((uint64_t)!_value)-1, 1));
+    }
+
+    simdjson_inline vbitmask_t to_bitmask() const {
+#if __riscv_v_fixed_vlen == 128
+      return __riscv_vmv_x(__riscv_vreinterpret_u16m1(value));
+#elif __riscv_v_fixed_vlen == 256
+      return __riscv_vmv_x(__riscv_vreinterpret_u32m1(value));
+#else
+      return __riscv_vmv_x(__riscv_vreinterpret_u64m1(value));
+#endif
+    }
+
+    // Bit operations
+    simdjson_inline simd8<bool> operator|(const simd8<bool> other) const {  return __riscv_vmor(*this, other, VL8); }
+    simdjson_inline simd8<bool> operator&(const simd8<bool> other) const {  return __riscv_vmand(*this, other, VL8); }
+    simdjson_inline simd8<bool> operator^(const simd8<bool> other) const {  return __riscv_vmxor(*this, other, VL8); }
+    simdjson_inline simd8<bool> bit_andnot(const simd8<bool> other) const { return __riscv_vmandn(other, *this, VL8); }
+    simdjson_inline simd8<bool> operator~() const { return __riscv_vmnot(*this, VL8); }
+    simdjson_inline simd8<bool>& operator|=(const simd8<bool> other) { auto this_cast = static_cast<simd8<bool>*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline simd8<bool>& operator&=(const simd8<bool> other) { auto this_cast = static_cast<simd8<bool>*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline simd8<bool>& operator^=(const simd8<bool> other) { auto this_cast = static_cast<simd8<bool>*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t> {
+
+    vuint8_t value;
+    static constexpr int SIZE = sizeof(value);
+
+    simdjson_inline simd8(const vuint8_t _value) : value(_value) {}
+    simdjson_inline simd8() : simd8(zero()) {}
+    simdjson_inline simd8(const uint8_t values[VL8]) : simd8(load(values)) {}
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    simdjson_inline simd8(simd8<bool> mask) : value(__riscv_vmerge_vxm_u8m1(zero(), -1, (vbool_t)mask, VL8)) {}
+
+    simdjson_inline operator const vuint8_t&() const { return this->value; }
+    simdjson_inline operator vuint8_t&() { return this->value; }
+
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) : simd8(vuint8_t{
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    }) {}
+
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    static simdjson_inline vuint8_t splat(uint8_t _value) { return __riscv_vmv_v_x_u8m1(_value, VL8); }
+    static simdjson_inline vuint8_t zero() { return splat(0); }
+    static simdjson_inline vuint8_t load(const uint8_t values[VL8]) { return __riscv_vle8_v_u8m1(values, VL8); }
+
+    // Bit operations
+    simdjson_inline simd8<uint8_t> operator|(const simd8<uint8_t> other) const { return  __riscv_vor_vv_u8m1(  value, other, VL8); }
+    simdjson_inline simd8<uint8_t> operator&(const simd8<uint8_t> other) const { return  __riscv_vand_vv_u8m1( value, other, VL8); }
+    simdjson_inline simd8<uint8_t> operator^(const simd8<uint8_t> other) const { return  __riscv_vxor_vv_u8m1( value, other, VL8); }
+    simdjson_inline simd8<uint8_t> operator~() const { return __riscv_vnot_v_u8m1(value, VL8); }
+#if __riscv_zvbb
+    simdjson_inline simd8<uint8_t> bit_andnot(const simd8<uint8_t> other) const { return __riscv_vandn_vv_u8m1(other, value, VL8); }
+#else
+    simdjson_inline simd8<uint8_t> bit_andnot(const simd8<uint8_t> other) const { return other & ~*this; }
+#endif
+    simdjson_inline simd8<uint8_t>& operator|=(const simd8<uint8_t> other) { value = *this | other; return *this; }
+    simdjson_inline simd8<uint8_t>& operator&=(const simd8<uint8_t> other) { value = *this & other; return *this; }
+    simdjson_inline simd8<uint8_t>& operator^=(const simd8<uint8_t> other) { value = *this ^ other; return *this; }
+
+    simdjson_inline simd8<bool> operator==(const simd8<uint8_t> other) const { return __riscv_vmseq(value, other, VL8); }
+    simdjson_inline simd8<bool> operator==(uint8_t other) const { return __riscv_vmseq(value, other, VL8); }
+
+    template<int N=1>
+    simdjson_inline simd8<uint8_t> prev(const simd8<uint8_t> prev_chunk) const {
+      return __riscv_vslideup(__riscv_vslidedown(prev_chunk, VL8-N, VL8), value, N, VL8);
+    }
+
+    // Store to array
+    simdjson_inline void store(uint8_t dst[VL8]) const { return __riscv_vse8(dst, value, VL8); }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return __riscv_vsaddu(value, other, VL8); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return __riscv_vssubu(value, other, VL8); }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<uint8_t> operator+(const simd8<uint8_t> other) const { return __riscv_vadd(value, other, VL8); }
+    simdjson_inline simd8<uint8_t> operator-(const simd8<uint8_t> other) const { return __riscv_vsub(value, other, VL8); }
+    simdjson_inline simd8<uint8_t>& operator+=(const simd8<uint8_t> other) { value = *this + other; return *this; }
+    simdjson_inline simd8<uint8_t>& operator-=(const simd8<uint8_t> other) { value = *this - other; return *this; }
+
+    // Order-specific operations
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return __riscv_vmsleu(value, other, VL8); }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return __riscv_vmsgeu(value, other, VL8); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const  { return __riscv_vmsltu(value, other, VL8); }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const  { return __riscv_vmsgtu(value, other, VL8); }
+
+    // Same as >, but instead of guaranteeing all 1's == true, false = 0 and true = nonzero.
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return simd8<uint8_t>(*this > other); }
+    // Same as <, but instead of guaranteeing all 1's == true, false = 0 and true = nonzero.
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return simd8<uint8_t>(*this < other); }
+
+    // Bit-specific operations
+    simdjson_inline bool any_bits_set_anywhere() const {
+      return __riscv_vfirst(__riscv_vmsne(value, 0, VL8), VL8) >= 0;
+    }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return (*this & bits).any_bits_set_anywhere(); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return __riscv_vsrl(value, N, VL8); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return __riscv_vsll(value, N, VL8); }
+
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return __riscv_vrgather(lookup_table, value, VL8);
+    }
+
+    // compress inactive elements, to match AVX-512 behavior
+    template<typename L>
+    simdjson_inline void compress(vbitmask_t mask, L * output) const {
+      mask = (vbitmask_t)~mask;
+#if __riscv_v_fixed_vlen == 128
+      vbool8_t m = __riscv_vreinterpret_b8(__riscv_vmv_s_x_u16m1(mask, 1));
+#elif __riscv_v_fixed_vlen == 256
+      vbool8_t m = __riscv_vreinterpret_b8(__riscv_vmv_s_x_u32m1(mask, 1));
+#else
+      vbool8_t m = __riscv_vreinterpret_b8(__riscv_vmv_s_x_u64m1(mask, 1));
+#endif
+      __riscv_vse8_v_u8m1(output, __riscv_vcompress(value, m, VL8), count_ones(mask));
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> {
+    vint8_t value;
+    static constexpr int SIZE = sizeof(value);
+
+    simdjson_inline simd8(const vint8_t _value) : value(_value) {}
+    simdjson_inline simd8() : simd8(zero()) {}
+    simdjson_inline simd8(const int8_t values[VL8]) : simd8(load(values)) {}
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+
+    simdjson_inline operator const vint8_t&() const { return this->value; }
+    simdjson_inline operator vint8_t&() { return this->value; }
+
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) : simd8(vint8_t{
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    }) {}
+
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    static simdjson_inline vint8_t splat(int8_t _value) { return __riscv_vmv_v_x_i8m1(_value, VL8); }
+    static simdjson_inline vint8_t zero() { return splat(0); }
+    static simdjson_inline vint8_t load(const int8_t values[VL8]) { return __riscv_vle8_v_i8m1(values, VL8); }
+
+
+    simdjson_inline void store(int8_t dst[VL8]) const { return __riscv_vse8(dst, value, VL8); }
+
+    // Explicit conversion to/from unsigned
+    simdjson_inline explicit simd8(const vuint8_t other): simd8(__riscv_vreinterpret_i8m1(other)) {}
+    simdjson_inline explicit operator simd8<uint8_t>() const { return __riscv_vreinterpret_u8m1(value); }
+
+    // Math
+    simdjson_inline simd8<int8_t> operator+(const simd8<int8_t> other) const { return __riscv_vadd(value, other, VL8); }
+    simdjson_inline simd8<int8_t> operator-(const simd8<int8_t> other) const { return __riscv_vsub(value, other, VL8); }
+    simdjson_inline simd8<int8_t>& operator+=(const simd8<int8_t> other) { value = *this + other; return *this; }
+    simdjson_inline simd8<int8_t>& operator-=(const simd8<int8_t> other) { value = *this - other; return *this; }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val( const simd8<int8_t> other) const { return __riscv_vmax( value, other, VL8); }
+    simdjson_inline simd8<int8_t> min_val( const simd8<int8_t> other) const { return __riscv_vmin( value, other, VL8); }
+    simdjson_inline simd8<bool> operator>( const simd8<int8_t> other) const { return __riscv_vmsgt(value, other, VL8); }
+    simdjson_inline simd8<bool> operator<( const simd8<int8_t> other) const { return __riscv_vmslt(value, other, VL8); }
+    simdjson_inline simd8<bool> operator==(const simd8<int8_t> other) const { return __riscv_vmseq(value, other, VL8); }
+
+    template<int N=1>
+    simdjson_inline simd8<int8_t> prev(const simd8<int8_t> prev_chunk) const {
+      return __riscv_vslideup(__riscv_vslidedown(prev_chunk, VL8-N, VL8), value, N, VL8);
+    }
+
+    // Perform a lookup assuming no value is larger than 16
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return __riscv_vrgather(lookup_table, value, VL8);
+    }
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  template<typename T>
+  struct simd8x64;
+  template<>
+  struct simd8x64<uint8_t> {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<uint8_t>);
+    vuint8x64_t value;
+
+#if __riscv_v_fixed_vlen >= 512
+    template<int idx> simd8<uint8_t> get() const { return value; }
+#else
+    template<int idx> simd8<uint8_t> get() const { return __riscv_vget_u8m1(value, idx); }
+#endif
+
+    simdjson_inline operator const vuint8x64_t&() const { return this->value; }
+    simdjson_inline operator vuint8x64_t&() { return this->value; }
+
+    simd8x64(const simd8x64<uint8_t>& o) = delete; // no copy allowed
+    simd8x64<uint8_t>& operator=(const simd8<uint8_t>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+#if __riscv_v_fixed_vlen == 128
+    simdjson_inline simd8x64(const uint8_t *ptr, size_t n = 64) : value(__riscv_vle8_v_u8m4(ptr, n)) {}
+#elif __riscv_v_fixed_vlen == 256
+    simdjson_inline simd8x64(const uint8_t *ptr, size_t n = 64) : value(__riscv_vle8_v_u8m2(ptr, n)) {}
+#else
+    simdjson_inline simd8x64(const uint8_t *ptr, size_t n = 64) : value(__riscv_vle8_v_u8m1(ptr, n)) {}
+#endif
+
+    simdjson_inline void store(uint8_t ptr[64]) const {
+      __riscv_vse8(ptr, value, 64);
+    }
+
+    simdjson_inline bool is_ascii() const {
+#if __riscv_v_fixed_vlen == 128
+      return __riscv_vfirst(__riscv_vmslt(__riscv_vreinterpret_i8m4(value), 0, 64), 64) < 0;
+#elif __riscv_v_fixed_vlen == 256
+      return __riscv_vfirst(__riscv_vmslt(__riscv_vreinterpret_i8m2(value), 0, 64), 64) < 0;
+#else
+      return __riscv_vfirst(__riscv_vmslt(__riscv_vreinterpret_i8m1(value), 0, 64), 64) < 0;
+#endif
+    }
+
+    // compress inactive elements, to match AVX-512 behavior
+    simdjson_inline uint64_t compress(uint64_t mask, uint8_t * output) const {
+      mask = ~mask;
+#if __riscv_v_fixed_vlen == 128
+      vboolx64_t m = __riscv_vreinterpret_b2(__riscv_vmv_s_x_u64m1(mask, 1));
+#elif __riscv_v_fixed_vlen == 256
+      vboolx64_t m = __riscv_vreinterpret_b4(__riscv_vmv_s_x_u64m1(mask, 1));
+#else
+      vboolx64_t m = __riscv_vreinterpret_b8(__riscv_vmv_s_x_u64m1(mask, 1));
+#endif
+      size_t cnt = count_ones(mask);
+      __riscv_vse8(output, __riscv_vcompress(value, m, 64), cnt);
+      return cnt;
+    }
+
+    simdjson_inline uint64_t eq(const uint8_t m) const {
+      return __riscv_vmv_x(__riscv_vreinterpret_u64m1(__riscv_vmseq(value, m, 64)));
+    }
+
+    simdjson_inline uint64_t lteq(const uint8_t m) const {
+      return __riscv_vmv_x(__riscv_vreinterpret_u64m1(__riscv_vmsleu(value, m, 64)));
+    }
+  }; // struct simd8x64<uint8_t>
+
+} // namespace simd
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_RVV_VLS_SIMD_H
+/* end file simdjson/rvv-vls/simd.h */
+/* including simdjson/rvv-vls/stringparsing_defs.h: #include "simdjson/rvv-vls/stringparsing_defs.h" */
+/* begin file simdjson/rvv-vls/stringparsing_defs.h */
+#ifndef SIMDJSON_RVV_VLS_STRINGPARSING_DEFS_H
+#define SIMDJSON_RVV_VLS_STRINGPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint64_t BYTES_PROCESSED = sizeof(simd8<uint8_t>);
+  simdjson_inline backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return ((quote_bits - 1) & bs_bits) != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+
+  uint64_t bs_bits;
+  uint64_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  return { (v == '\\').to_bitmask(), (v == '"').to_bitmask() };
+}
+
+struct escaping {
+  static constexpr uint64_t BYTES_PROCESSED = sizeof(simd8<uint8_t>);
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits != 0; }
+  simdjson_inline int escape_index() { return trailing_zeroes(escape_bits) / 4; }
+
+  uint64_t escape_bits;
+}; // struct escaping
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "escaping finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  return { ((v == '"') | (v == '\\') | (v == 32)).to_bitmask() };
+}
+
+
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_RVV_VLS_STRINGPARSING_DEFS_H
+/* end file simdjson/rvv-vls/stringparsing_defs.h */
+/* including simdjson/rvv-vls/numberparsing_defs.h: #include "simdjson/rvv-vls/numberparsing_defs.h" */
+/* begin file simdjson/rvv-vls/numberparsing_defs.h */
+#ifndef SIMDJSON_RVV_VLS_NUMBERPARSING_DEFS_H
+#define SIMDJSON_RVV_VLS_NUMBERPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+#ifdef JSON_TEST_NUMBERS // for unit testing
+void found_invalid_number(const uint8_t *buf);
+void found_integer(int64_t result, const uint8_t *buf);
+void found_unsigned_integer(uint64_t result, const uint8_t *buf);
+void found_float(double result, const uint8_t *buf);
+#endif
+
+namespace simdjson {
+namespace rvv_vls {
+namespace numberparsing {
+
+// credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const char *chars) {
+  uint64_t val;
+#if __riscv_misaligned_fast
+  memcpy(&val, chars, sizeof(uint64_t));
+#else
+  val = __riscv_vmv_x(__riscv_vreinterpret_u64m1(__riscv_vlmul_ext_u8m1(__riscv_vle8_v_u8mf2((uint8_t*)chars, 8))));
+#endif
+  val = (val & 0x0F0F0F0F0F0F0F0F) * 2561 >> 8;
+  val = (val & 0x00FF00FF00FF00FF) * 6553601 >> 16;
+  return uint32_t((val & 0x0000FFFF0000FFFF) * 42949672960001 >> 32);
+}
+
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  return parse_eight_digits_unrolled(reinterpret_cast<const char *>(chars));
+}
+
+/** @private */
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace rvv_vls
+} // namespace simdjson
+
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+
+#endif // SIMDJSON_RVV_VLS_NUMBERPARSING_DEFS_H
+/* end file simdjson/rvv-vls/numberparsing_defs.h */
+
+#define SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT 1
+/* end file simdjson/rvv-vls/begin.h */
+/* including simdjson/generic/amalgamated.h for rvv_vls: #include "simdjson/generic/amalgamated.h" */
+/* begin file simdjson/generic/amalgamated.h for rvv_vls */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_GENERIC_DEPENDENCIES_H)
+#error simdjson/generic/dependencies.h must be included before simdjson/generic/amalgamated.h!
+#endif
+
+/* including simdjson/generic/base.h for rvv_vls: #include "simdjson/generic/base.h" */
+/* begin file simdjson/generic/base.h for rvv_vls */
+#ifndef SIMDJSON_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): // If we haven't got an implementation yet, we're in the editor, editing a generic file! Just */
+/* amalgamation skipped (editor-only): // use the most advanced one we can so the most possible stuff can be tested. */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation_detection.h" */
+/* amalgamation skipped (editor-only): #if SIMDJSON_IMPLEMENTATION_ICELAKE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_HASWELL */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_WESTMERE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_ARM64 */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_PPC64 */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LASX */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LSX */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_RVV_VLS */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_FALLBACK */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/begin.h" */
+/* amalgamation skipped (editor-only): #else */
+/* amalgamation skipped (editor-only): #error "All possible implementations (including fallback) have been disabled! simdjson will not run." */
+/* amalgamation skipped (editor-only): #endif */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+
+struct open_container;
+class dom_parser_implementation;
+
+/**
+ * The type of a JSON number
+ */
+enum class number_type {
+    floating_point_number=1, /// a binary64 number
+    signed_integer,          /// a signed integer that fits in a 64-bit word using two's complement
+    unsigned_integer,        /// a positive integer larger or equal to 1<<63
+    big_integer              /// a big integer that does not fit in a 64-bit word
+};
+
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_BASE_H
+/* end file simdjson/generic/base.h for rvv_vls */
+/* including simdjson/generic/jsoncharutils.h for rvv_vls: #include "simdjson/generic/jsoncharutils.h" */
+/* begin file simdjson/generic/jsoncharutils.h for rvv_vls */
+#ifndef SIMDJSON_GENERIC_JSONCHARUTILS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_JSONCHARUTILS_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/jsoncharutils_tables.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace jsoncharutils {
+
+// return non-zero if not a structural or whitespace char
+// zero otherwise
+simdjson_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace_negated[c];
+}
+
+simdjson_inline uint32_t is_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace[c];
+}
+
+// returns a value with the high 16 bits set if not valid
+// otherwise returns the conversion of the 4 hex digits at src into the bottom
+// 16 bits of the 32-bit return register
+//
+// see
+// https://lemire.me/blog/2019/04/17/parsing-short-hexadecimal-strings-efficiently/
+static inline uint32_t hex_to_u32_nocheck(
+    const uint8_t *src) { // strictly speaking, static inline is a C-ism
+  uint32_t v1 = internal::digit_to_val32[630 + src[0]];
+  uint32_t v2 = internal::digit_to_val32[420 + src[1]];
+  uint32_t v3 = internal::digit_to_val32[210 + src[2]];
+  uint32_t v4 = internal::digit_to_val32[0 + src[3]];
+  return v1 | v2 | v3 | v4;
+}
+
+// given a code point cp, writes to c
+// the utf-8 code, outputting the length in
+// bytes, if the length is zero, the code point
+// is invalid
+//
+// This can possibly be made faster using pdep
+// and clz and table lookups, but JSON documents
+// have few escaped code points, and the following
+// function looks cheap.
+//
+// Note: we assume that surrogates are treated separately
+//
+simdjson_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
+  if (cp <= 0x7F) {
+    c[0] = uint8_t(cp);
+    return 1; // ascii
+  }
+  if (cp <= 0x7FF) {
+    c[0] = uint8_t((cp >> 6) + 192);
+    c[1] = uint8_t((cp & 63) + 128);
+    return 2; // universal plane
+    //  Surrogates are treated elsewhere...
+    //} //else if (0xd800 <= cp && cp <= 0xdfff) {
+    //  return 0; // surrogates // could put assert here
+  } else if (cp <= 0xFFFF) {
+    c[0] = uint8_t((cp >> 12) + 224);
+    c[1] = uint8_t(((cp >> 6) & 63) + 128);
+    c[2] = uint8_t((cp & 63) + 128);
+    return 3;
+  } else if (cp <= 0x10FFFF) { // if you know you have a valid code point, this
+                               // is not needed
+    c[0] = uint8_t((cp >> 18) + 240);
+    c[1] = uint8_t(((cp >> 12) & 63) + 128);
+    c[2] = uint8_t(((cp >> 6) & 63) + 128);
+    c[3] = uint8_t((cp & 63) + 128);
+    return 4;
+  }
+  // will return 0 when the code point was too large.
+  return 0; // bad r
+}
+
+#if SIMDJSON_IS_32BITS // _umul128 for x86, arm
+// this is a slow emulation routine for 32-bit
+//
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+  return x * (uint64_t)y;
+}
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+  uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
+  uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
+  uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
+  uint64_t adbc_carry = !!(adbc < ad);
+  uint64_t lo = bd + (adbc << 32);
+  *hi = __emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) +
+        (adbc_carry << 32) + !!(lo < bd);
+  return lo;
+}
+#endif
+
+} // namespace jsoncharutils
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_JSONCHARUTILS_H
+/* end file simdjson/generic/jsoncharutils.h for rvv_vls */
+/* including simdjson/generic/atomparsing.h for rvv_vls: #include "simdjson/generic/atomparsing.h" */
+/* begin file simdjson/generic/atomparsing.h for rvv_vls */
+#ifndef SIMDJSON_GENERIC_ATOMPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_ATOMPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+/// @private
+namespace atomparsing {
+
+// The string_to_uint32 is exclusively used to map literal strings to 32-bit values.
+// We use memcpy instead of a pointer cast to avoid undefined behaviors since we cannot
+// be certain that the character pointer will be properly aligned.
+// You might think that using memcpy makes this function expensive, but you'd be wrong.
+// All decent optimizing compilers (GCC, clang, Visual Studio) will compile string_to_uint32("false");
+// to the compile-time constant 1936482662.
+simdjson_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
+
+
+// Again in str4ncmp we use a memcpy to avoid undefined behavior. The memcpy may appear expensive.
+// Yet all decent optimizing compilers will compile memcpy to a single instruction, just about.
+simdjson_warn_unused
+simdjson_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
+  uint32_t srcval; // we want to avoid unaligned 32-bit loads (undefined in C/C++)
+  static_assert(sizeof(uint32_t) <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be larger than 4 bytes");
+  std::memcpy(&srcval, src, sizeof(uint32_t));
+  return srcval ^ string_to_uint32(atom);
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src) {
+  return (str4ncmp(src, "true") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_true_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "true"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src) {
+  return (str4ncmp(src+1, "alse") | jsoncharutils::is_not_structural_or_whitespace(src[5])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
+  if (len > 5) { return is_valid_false_atom(src); }
+  else if (len == 5) { return !str4ncmp(src+1, "alse"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src) {
+  return (str4ncmp(src, "null") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_null_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "null"); }
+  else { return false; }
+}
+
+} // namespace atomparsing
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_ATOMPARSING_H
+/* end file simdjson/generic/atomparsing.h for rvv_vls */
+/* including simdjson/generic/dom_parser_implementation.h for rvv_vls: #include "simdjson/generic/dom_parser_implementation.h" */
+/* begin file simdjson/generic/dom_parser_implementation.h for rvv_vls */
+#ifndef SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/dom_parser_implementation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+
+// expectation: sizeof(open_container) = 64/8.
+struct open_container {
+  uint32_t tape_index; // where, on the tape, does the scope ([,{) begins
+  uint32_t count; // how many elements in the scope
+}; // struct open_container
+
+static_assert(sizeof(open_container) == 64/8, "Open container must be 64 bits");
+
+class dom_parser_implementation final : public internal::dom_parser_implementation {
+public:
+  /** Tape location of each open { or [ */
+  std::unique_ptr<open_container[]> open_containers{};
+  /** Whether each open container is a [ or { */
+  std::unique_ptr<bool[]> is_array{};
+  /** Buffer passed to stage 1 */
+  const uint8_t *buf{};
+  /** Length passed to stage 1 */
+  size_t len{0};
+  /** Document passed to stage 2 */
+  dom::document *doc{};
+
+  inline dom_parser_implementation() noexcept;
+  inline dom_parser_implementation(dom_parser_implementation &&other) noexcept;
+  inline dom_parser_implementation &operator=(dom_parser_implementation &&other) noexcept;
+  dom_parser_implementation(const dom_parser_implementation &) = delete;
+  dom_parser_implementation &operator=(const dom_parser_implementation &) = delete;
+
+  simdjson_warn_unused error_code parse(const uint8_t *buf, size_t len, dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage1(const uint8_t *buf, size_t len, stage1_mode partial) noexcept final;
+  simdjson_warn_unused error_code stage2(dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage2_next(dom::document &doc) noexcept final;
+  simdjson_warn_unused uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) const noexcept final;
+  simdjson_warn_unused uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept final;
+  inline simdjson_warn_unused error_code set_capacity(size_t capacity) noexcept final;
+  inline simdjson_warn_unused error_code set_max_depth(size_t max_depth) noexcept final;
+private:
+  simdjson_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
+
+};
+
+} // namespace rvv_vls
+} // namespace simdjson
+
+namespace simdjson {
+namespace rvv_vls {
+
+inline dom_parser_implementation::dom_parser_implementation() noexcept = default;
+inline dom_parser_implementation::dom_parser_implementation(dom_parser_implementation &&other) noexcept = default;
+inline dom_parser_implementation &dom_parser_implementation::operator=(dom_parser_implementation &&other) noexcept = default;
+
+// Leaving these here so they can be inlined if so desired
+inline simdjson_warn_unused error_code dom_parser_implementation::set_capacity(size_t capacity) noexcept {
+  if(capacity > SIMDJSON_MAXSIZE_BYTES) { return CAPACITY; }
+  // Stage 1 index output
+  size_t max_structures = SIMDJSON_ROUNDUP_N(capacity, 64) + 2 + 7;
+  structural_indexes.reset( new (std::nothrow) uint32_t[max_structures] );
+  if (!structural_indexes) { _capacity = 0; return MEMALLOC; }
+  structural_indexes[0] = 0;
+  n_structural_indexes = 0;
+
+  _capacity = capacity;
+  return SUCCESS;
+}
+
+inline simdjson_warn_unused error_code dom_parser_implementation::set_max_depth(size_t max_depth) noexcept {
+  // Stage 2 stacks
+  open_containers.reset(new (std::nothrow) open_container[max_depth]);
+  is_array.reset(new (std::nothrow) bool[max_depth]);
+  if (!is_array || !open_containers) { _max_depth = 0; return MEMALLOC; }
+
+  _max_depth = max_depth;
+  return SUCCESS;
+}
+
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file simdjson/generic/dom_parser_implementation.h for rvv_vls */
+/* including simdjson/generic/implementation_simdjson_result_base.h for rvv_vls: #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base.h for rvv_vls */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+
+// This is a near copy of include/error.h's implementation_simdjson_result_base, except it doesn't use std::pair
+// so we can avoid inlining errors
+// TODO reconcile these!
+/**
+ * The result of a simdjson operation that could fail.
+ *
+ * Gives the option of reading error codes, or throwing an exception by casting to the desired result.
+ *
+ * This is a base class for implementations that want to add functions to the result type for
+ * chaining.
+ *
+ * Override like:
+ *
+ *   struct simdjson_result<T> : public internal::implementation_simdjson_result_base<T> {
+ *     simdjson_result() noexcept : internal::implementation_simdjson_result_base<T>() {}
+ *     simdjson_result(error_code error) noexcept : internal::implementation_simdjson_result_base<T>(error) {}
+ *     simdjson_result(T &&value) noexcept : internal::implementation_simdjson_result_base<T>(std::forward(value)) {}
+ *     simdjson_result(T &&value, error_code error) noexcept : internal::implementation_simdjson_result_base<T>(value, error) {}
+ *     // Your extra methods here
+ *   }
+ *
+ * Then any method returning simdjson_result<T> will be chainable with your methods.
+ */
+template<typename T>
+struct implementation_simdjson_result_base {
+
+  /**
+   * Create a new empty result with error = UNINITIALIZED.
+   */
+  simdjson_inline implementation_simdjson_result_base() noexcept = default;
+
+  /**
+   * Create a new error result.
+   */
+  simdjson_inline implementation_simdjson_result_base(error_code error) noexcept;
+
+  /**
+   * Create a new successful result.
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value) noexcept;
+
+  /**
+   * Create a new result with both things (use if you don't want to branch when creating the result).
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value, error_code error) noexcept;
+
+  /**
+   * Move the value and the error to the provided variables.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   * @param error The variable to assign the error to. Set to SUCCESS if there is no error.
+   */
+  simdjson_inline void tie(T &value, error_code &error) && noexcept;
+
+  /**
+   * Move the value to the provided variable.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   */
+  simdjson_warn_unused simdjson_inline error_code get(T &value) && noexcept;
+
+  /**
+   * The error.
+   */
+  simdjson_warn_unused simdjson_inline error_code error() const noexcept;
+
+  /**
+   * Whether there is a value.
+   */
+  simdjson_warn_unused simdjson_inline bool has_value() const noexcept;
+
+#if SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T& operator*() &  noexcept(false);
+  simdjson_inline T&& operator*() &&  noexcept(false);
+  /**
+   * Arrow operator to access members of the contained value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T* operator->() noexcept(false);
+  simdjson_inline const T* operator->() const noexcept(false);
+
+  simdjson_inline T& value() & noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& value() && noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& take_value() && noexcept(false);
+
+  /**
+   * Cast to the value (will throw on error).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline operator T&&() && noexcept(false);
+
+
+#endif // SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline const T& value_unsafe() const& noexcept;
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T& value_unsafe() & noexcept;
+  /**
+   * Take the result value (move it). This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T&& value_unsafe() && noexcept;
+
+  using value_type = T;
+  using error_type = error_code;
+
+protected:
+  /** users should never directly access first and second. **/
+  T first{}; /** Users should never directly access 'first'. **/
+  error_code second{UNINITIALIZED}; /** Users should never directly access 'second'. **/
+}; // struct implementation_simdjson_result_base
+
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+/* end file simdjson/generic/implementation_simdjson_result_base.h for rvv_vls */
+/* including simdjson/generic/numberparsing.h for rvv_vls: #include "simdjson/generic/numberparsing.h" */
+/* begin file simdjson/generic/numberparsing.h for rvv_vls */
+#ifndef SIMDJSON_GENERIC_NUMBERPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_NUMBERPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <limits>
+#include <ostream>
+#include <cstring>
+
+namespace simdjson {
+namespace rvv_vls {
+namespace numberparsing {
+
+#ifdef JSON_TEST_NUMBERS
+#define INVALID_NUMBER(SRC) (found_invalid_number((SRC)), NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (found_integer((VALUE), (SRC)), (WRITER).append_s64((VALUE)))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (found_unsigned_integer((VALUE), (SRC)), (WRITER).append_u64((VALUE)))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (found_float((VALUE), (SRC)), (WRITER).append_double((VALUE)))
+#define BIGINT_NUMBER(SRC) (found_invalid_number((SRC)), BIGINT_ERROR)
+#else
+#define INVALID_NUMBER(SRC) (NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (WRITER).append_s64((VALUE))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (WRITER).append_u64((VALUE))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (WRITER).append_double((VALUE))
+#define BIGINT_NUMBER(SRC) (BIGINT_ERROR)
+#endif
+
+namespace {
+
+// Convert a mantissa, an exponent and a sign bit into an ieee64 double.
+// The real_exponent needs to be in [0, 2046] (technically real_exponent = 2047 would be acceptable).
+// The mantissa should be in [0,1<<53). The bit at index (1ULL << 52) while be zeroed.
+simdjson_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
+    double d;
+    mantissa &= ~(1ULL << 52);
+    mantissa |= real_exponent << 52;
+    mantissa |= ((static_cast<uint64_t>(negative)) << 63);
+    std::memcpy(&d, &mantissa, sizeof(d));
+    return d;
+}
+
+// Attempts to compute i * 10^(power) exactly; and if "negative" is
+// true, negate the result.
+// This function will only work in some cases, when it does not work, success is
+// set to false. This should work *most of the time* (like 99% of the time).
+// We assume that power is in the [smallest_power,
+// largest_power] interval: the caller is responsible for this check.
+simdjson_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
+  // we start with a fast path
+  // It was described in
+  // Clinger WD. How to read floating point numbers accurately.
+  // ACM SIGPLAN Notices. 1990
+#ifndef FLT_EVAL_METHOD
+#error "FLT_EVAL_METHOD should be defined, please include cfloat."
+#endif
+#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
+  // We cannot be certain that x/y is rounded to nearest.
+  if (0 <= power && power <= 22 && i <= 9007199254740991)
+#else
+  if (-22 <= power && power <= 22 && i <= 9007199254740991)
+#endif
+  {
+    // convert the integer into a double. This is lossless since
+    // 0 <= i <= 2^53 - 1.
+    d = double(i);
+    //
+    // The general idea is as follows.
+    // If 0 <= s < 2^53 and if 10^0 <= p <= 10^22 then
+    // 1) Both s and p can be represented exactly as 64-bit floating-point
+    // values
+    // (binary64).
+    // 2) Because s and p can be represented exactly as floating-point values,
+    // then s * p
+    // and s / p will produce correctly rounded values.
+    //
+    if (power < 0) {
+      d = d / simdjson::internal::power_of_ten[-power];
+    } else {
+      d = d * simdjson::internal::power_of_ten[power];
+    }
+    if (negative) {
+      d = -d;
+    }
+    return true;
+  }
+  // When 22 < power && power <  22 + 16, we could
+  // hope for another, secondary fast path.  It was
+  // described by David M. Gay in  "Correctly rounded
+  // binary-decimal and decimal-binary conversions." (1990)
+  // If you need to compute i * 10^(22 + x) for x < 16,
+  // first compute i * 10^x, if you know that result is exact
+  // (e.g., when i * 10^x < 2^53),
+  // then you can still proceed and do (i * 10^x) * 10^22.
+  // Is this worth your time?
+  // You need  22 < power *and* power <  22 + 16 *and* (i * 10^(x-22) < 2^53)
+  // for this second fast path to work.
+  // If you you have 22 < power *and* power <  22 + 16, and then you
+  // optimistically compute "i * 10^(x-22)", there is still a chance that you
+  // have wasted your time if i * 10^(x-22) >= 2^53. It makes the use cases of
+  // this optimization maybe less common than we would like. Source:
+  // http://www.exploringbinary.com/fast-path-decimal-to-floating-point-conversion/
+  // also used in RapidJSON: https://rapidjson.org/strtod_8h_source.html
+
+  // The fast path has now failed, so we are failing back on the slower path.
+
+  // In the slow path, we need to adjust i so that it is > 1<<63 which is always
+  // possible, except if i == 0, so we handle i == 0 separately.
+  if(i == 0) {
+    d = negative ? -0.0 : 0.0;
+    return true;
+  }
+
+
+  // The exponent is 1024 + 63 + power
+  //     + floor(log(5**power)/log(2)).
+  // The 1024 comes from the ieee64 standard.
+  // The 63 comes from the fact that we use a 64-bit word.
+  //
+  // Computing floor(log(5**power)/log(2)) could be
+  // slow. Instead we use a fast function.
+  //
+  // For power in (-400,350), we have that
+  // (((152170 + 65536) * power ) >> 16);
+  // is equal to
+  //  floor(log(5**power)/log(2)) + power when power >= 0
+  // and it is equal to
+  //  ceil(log(5**-power)/log(2)) + power when power < 0
+  //
+  // The 65536 is (1<<16) and corresponds to
+  // (65536 * power) >> 16 ---> power
+  //
+  // ((152170 * power ) >> 16) is equal to
+  // floor(log(5**power)/log(2))
+  //
+  // Note that this is not magic: 152170/(1<<16) is
+  // approximately equal to log(5)/log(2).
+  // The 1<<16 value is a power of two; we could use a
+  // larger power of 2 if we wanted to.
+  //
+  int64_t exponent = (((152170 + 65536) * power) >> 16) + 1024 + 63;
+
+
+  // We want the most significant bit of i to be 1. Shift if needed.
+  int lz = leading_zeroes(i);
+  i <<= lz;
+
+
+  // We are going to need to do some 64-bit arithmetic to get a precise product.
+  // We use a table lookup approach.
+  // It is safe because
+  // power >= smallest_power
+  // and power <= largest_power
+  // We recover the mantissa of the power, it has a leading 1. It is always
+  // rounded down.
+  //
+  // We want the most significant 64 bits of the product. We know
+  // this will be non-zero because the most significant bit of i is
+  // 1.
+  const uint32_t index = 2 * uint32_t(power - simdjson::internal::smallest_power);
+  // Optimization: It may be that materializing the index as a variable might confuse some compilers and prevent effective complex-addressing loads. (Done for code clarity.)
+  //
+  // The full_multiplication function computes the 128-bit product of two 64-bit words
+  // with a returned value of type value128 with a "low component" corresponding to the
+  // 64-bit least significant bits of the product and with a "high component" corresponding
+  // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index]);
+#else
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index]);
+#endif
+
+  // Both i and power_of_five_128[index] have their most significant bit set to 1 which
+  // implies that the either the most or the second most significant bit of the product
+  // is 1. We pack values in this manner for efficiency reasons: it maximizes the use
+  // we make of the product. It also makes it easy to reason about the product: there
+  // is 0 or 1 leading zero in the product.
+
+  // Unless the least significant 9 bits of the high (64-bit) part of the full
+  // product are all 1s, then we know that the most significant 55 bits are
+  // exact and no further work is needed. Having 55 bits is necessary because
+  // we need 53 bits for the mantissa but we have to have one rounding bit and
+  // we can waste a bit if the most significant bit of the product is zero.
+  if((firstproduct.high & 0x1FF) == 0x1FF) {
+    // We want to compute i * 5^q, but only care about the top 55 bits at most.
+    // Consider the scenario where q>=0. Then 5^q may not fit in 64-bits. Doing
+    // the full computation is wasteful. So we do what is called a "truncated
+    // multiplication".
+    // We take the most significant 64-bits, and we put them in
+    // power_of_five_128[index]. Usually, that's good enough to approximate i * 5^q
+    // to the desired approximation using one multiplication. Sometimes it does not suffice.
+    // Then we store the next most significant 64 bits in power_of_five_128[index + 1], and
+    // then we get a better approximation to i * 5^q.
+    //
+    // That's for when q>=0. The logic for q<0 is somewhat similar but it is somewhat
+    // more complicated.
+    //
+    // There is an extra layer of complexity in that we need more than 55 bits of
+    // accuracy in the round-to-even scenario.
+    //
+    // The full_multiplication function computes the 128-bit product of two 64-bit words
+    // with a returned value of type value128 with a "low component" corresponding to the
+    // 64-bit least significant bits of the product and with a "high component" corresponding
+    // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index + 1]);
+#else
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index + 1]);
+#endif
+    firstproduct.low += secondproduct.high;
+    if(secondproduct.high > firstproduct.low) { firstproduct.high++; }
+    // As it has been proven by Noble Mushtak and Daniel Lemire in "Fast Number Parsing Without
+    // Fallback" (https://arxiv.org/abs/2212.06644), at this point we are sure that the product
+    // is sufficiently accurate, and more computation is not needed.
+  }
+  uint64_t lower = firstproduct.low;
+  uint64_t upper = firstproduct.high;
+  // The final mantissa should be 53 bits with a leading 1.
+  // We shift it so that it occupies 54 bits with a leading 1.
+  ///////
+  uint64_t upperbit = upper >> 63;
+  uint64_t mantissa = upper >> (upperbit + 9);
+  lz += int(1 ^ upperbit);
+
+  // Here we have mantissa < (1<<54).
+  int64_t real_exponent = exponent - lz;
+  if (simdjson_unlikely(real_exponent <= 0)) { // we have a subnormal?
+    // Here have that real_exponent <= 0 so -real_exponent >= 0
+    if(-real_exponent + 1 >= 64) { // if we have more than 64 bits below the minimum exponent, you have a zero for sure.
+      d = negative ? -0.0 : 0.0;
+      return true;
+    }
+    // next line is safe because -real_exponent + 1 < 0
+    mantissa >>= -real_exponent + 1;
+    // Thankfully, we can't have both "round-to-even" and subnormals because
+    // "round-to-even" only occurs for powers close to 0.
+    mantissa += (mantissa & 1); // round up
+    mantissa >>= 1;
+    // There is a weird scenario where we don't have a subnormal but just.
+    // Suppose we start with 2.2250738585072013e-308, we end up
+    // with 0x3fffffffffffff x 2^-1023-53 which is technically subnormal
+    // whereas 0x40000000000000 x 2^-1023-53  is normal. Now, we need to round
+    // up 0x3fffffffffffff x 2^-1023-53  and once we do, we are no longer
+    // subnormal, but we can only know this after rounding.
+    // So we only declare a subnormal if we are smaller than the threshold.
+    real_exponent = (mantissa < (uint64_t(1) << 52)) ? 0 : 1;
+    d = to_double(mantissa, real_exponent, negative);
+    return true;
+  }
+  // We have to round to even. The "to even" part
+  // is only a problem when we are right in between two floats
+  // which we guard against.
+  // If we have lots of trailing zeros, we may fall right between two
+  // floating-point values.
+  //
+  // The round-to-even cases take the form of a number 2m+1 which is in (2^53,2^54]
+  // times a power of two. That is, it is right between a number with binary significand
+  // m and another number with binary significand m+1; and it must be the case
+  // that it cannot be represented by a float itself.
+  //
+  // We must have that w * 10 ^q == (2m+1) * 2^p for some power of two 2^p.
+  // Recall that 10^q = 5^q * 2^q.
+  // When q >= 0, we must have that (2m+1) is divible by 5^q, so 5^q <= 2^54. We have that
+  //  5^23 <=  2^54 and it is the last power of five to qualify, so q <= 23.
+  // When q<0, we have  w  >=  (2m+1) x 5^{-q}.  We must have that w<2^{64} so
+  // (2m+1) x 5^{-q} < 2^{64}. We have that 2m+1>2^{53}. Hence, we must have
+  // 2^{53} x 5^{-q} < 2^{64}.
+  // Hence we have 5^{-q} < 2^{11}$ or q>= -4.
+  //
+  // We require lower <= 1 and not lower == 0 because we could not prove that
+  // that lower == 0 is implied; but we could prove that lower <= 1 is a necessary and sufficient test.
+  if (simdjson_unlikely((lower <= 1) && (power >= -4) && (power <= 23) && ((mantissa & 3) == 1))) {
+    if((mantissa  << (upperbit + 64 - 53 - 2)) ==  upper) {
+      mantissa &= ~1;             // flip it so that we do not round up
+    }
+  }
+
+  mantissa += mantissa & 1;
+  mantissa >>= 1;
+
+  // Here we have mantissa < (1<<53), unless there was an overflow
+  if (mantissa >= (1ULL << 53)) {
+    //////////
+    // This will happen when parsing values such as 7.2057594037927933e+16
+    ////////
+    mantissa = (1ULL << 52);
+    real_exponent++;
+  }
+  mantissa &= ~(1ULL << 52);
+  // we have to check that real_exponent is in range, otherwise we bail out
+  if (simdjson_unlikely(real_exponent > 2046)) {
+    // We have an infinite value!!! We could actually throw an error here if we could.
+    return false;
+  }
+  d = to_double(mantissa, real_exponent, negative);
+  return true;
+}
+
+// We call a fallback floating-point parser that might be slow. Note
+// it will accept JSON numbers, but the JSON spec. is more restrictive so
+// before you call parse_float_fallback, you need to have validated the input
+// string with the JSON grammar.
+// It will return an error (false) if the parsed number is infinite.
+// The string parsing itself always succeeds. We know that there is at least
+// one digit.
+static bool parse_float_fallback(const uint8_t *ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+static bool parse_float_fallback(const uint8_t *ptr, const uint8_t *end_ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr), reinterpret_cast<const char *>(end_ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+// check quickly whether the next 8 chars are made of digits
+// at a glance, it looks better than Mula's
+// http://0x80.pl/articles/swar-digits-validate.html
+simdjson_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
+  uint64_t val;
+  // this can read up to 7 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(7 <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be bigger than 7");
+  std::memcpy(&val, chars, 8);
+  // a branchy method might be faster:
+  // return (( val & 0xF0F0F0F0F0F0F0F0 ) == 0x3030303030303030)
+  //  && (( (val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0 ) ==
+  //  0x3030303030303030);
+  return (((val & 0xF0F0F0F0F0F0F0F0) |
+           (((val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0) >> 4)) ==
+          0x3333333333333333);
+}
+
+template<typename I>
+SIMDJSON_NO_SANITIZE_UNDEFINED // We deliberately allow overflow here and check later
+simdjson_inline bool parse_digit(const uint8_t c, I &i) {
+  const uint8_t digit = static_cast<uint8_t>(c - '0');
+  if (digit > 9) {
+    return false;
+  }
+  // PERF NOTE: multiplication by 10 is cheaper than arbitrary integer multiplication
+  i = 10 * i + digit; // might overflow, we will handle the overflow later
+  return true;
+}
+
+simdjson_inline bool is_digit(const uint8_t c) {
+  return static_cast<uint8_t>(c - '0') <= 9;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_decimal_after_separator(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
+  // we continue with the fiction that we have an integer. If the
+  // floating point number is representable as x * 10^z for some integer
+  // z that fits in 53 bits, then we will be able to convert back the
+  // the integer into a float in a lossless manner.
+  const uint8_t *const first_after_period = p;
+
+#ifdef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_SWAR_NUMBER_PARSING
+  // this helps if we have lots of decimals!
+  // this turns out to be frequent enough.
+  if (is_made_of_eight_digits_fast(p)) {
+    i = i * 100000000 + parse_eight_digits_unrolled(p);
+    p += 8;
+  }
+#endif // SIMDJSON_SWAR_NUMBER_PARSING
+#endif // #ifdef SIMDJSON_SWAR_NUMBER_PARSING
+  // Unrolling the first digit makes a small difference on some implementations (e.g. westmere)
+  if (parse_digit(*p, i)) { ++p; }
+  while (parse_digit(*p, i)) { p++; }
+  exponent = first_after_period - p;
+  // Decimal without digits (123.) is illegal
+  if (exponent == 0) {
+    return INVALID_NUMBER(src);
+  }
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
+  // Exp Sign: -123.456e[-]78
+  bool neg_exp = ('-' == *p);
+  if (neg_exp || '+' == *p) { p++; } // Skip + as well
+
+  // Exponent: -123.456e-[78]
+  auto start_exp = p;
+  int64_t exp_number = 0;
+  while (parse_digit(*p, exp_number)) { ++p; }
+  // It is possible for parse_digit to overflow.
+  // In particular, it could overflow to INT64_MIN, and we cannot do - INT64_MIN.
+  // Thus we *must* check for possible overflow before we negate exp_number.
+
+  // Performance notes: it may seem like combining the two "simdjson_unlikely checks" below into
+  // a single simdjson_unlikely path would be faster. The reasoning is sound, but the compiler may
+  // not oblige and may, in fact, generate two distinct paths in any case. It might be
+  // possible to do uint64_t(p - start_exp - 1) >= 18 but it could end up trading off
+  // instructions for a simdjson_likely branch, an unconclusive gain.
+
+  // If there were no digits, it's an error.
+  if (simdjson_unlikely(p == start_exp)) {
+    return INVALID_NUMBER(src);
+  }
+  // We have a valid positive exponent in exp_number at this point, except that
+  // it may have overflowed.
+
+  // If there were more than 18 digits, we may have overflowed the integer. We have to do
+  // something!!!!
+  if (simdjson_unlikely(p > start_exp+18)) {
+    // Skip leading zeroes: 1e000000000000000000001 is technically valid and does not overflow
+    while (*start_exp == '0') { start_exp++; }
+    // 19 digits could overflow int64_t and is kind of absurd anyway. We don't
+    // support exponents smaller than -999,999,999,999,999,999 and bigger
+    // than 999,999,999,999,999,999.
+    // We can truncate.
+    // Note that 999999999999999999 is assuredly too large. The maximal ieee64 value before
+    // infinity is ~1.8e308. The smallest subnormal is ~5e-324. So, actually, we could
+    // truncate at 324.
+    // Note that there is no reason to fail per se at this point in time.
+    // E.g., 0e999999999999999999999 is a fine number.
+    if (p > start_exp+18) { exp_number = 999999999999999999; }
+  }
+  // At this point, we know that exp_number is a sane, positive, signed integer.
+  // It is <= 999,999,999,999,999,999. As long as 'exponent' is in
+  // [-8223372036854775808, 8223372036854775808], we won't overflow. Because 'exponent'
+  // is bounded in magnitude by the size of the JSON input, we are fine in this universe.
+  // To sum it up: the next line should never overflow.
+  exponent += (neg_exp ? -exp_number : exp_number);
+  return SUCCESS;
+}
+
+simdjson_inline bool check_if_integer(const uint8_t *const src, size_t max_length) {
+  const uint8_t *const srcend = src + max_length;
+  bool negative = (*src == '-'); // we can always read at least one character after the '-'
+  const uint8_t *p = src + uint8_t(negative);
+  if(p == srcend) { return false; }
+  if(*p == '0') {
+    ++p;
+    if(p == srcend) { return true; }
+    if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+    return true;
+  }
+  while(p != srcend && is_digit(*p)) { ++p; }
+  if(p == srcend) { return true; }
+  if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+  return true;
+}
+
+simdjson_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
+  // It is possible that the integer had an overflow.
+  // We have to handle the case where we have 0.0000somenumber.
+  const uint8_t *start = start_digits;
+  while ((*start == '0') || (*start == '.')) { ++start; }
+  // we over-decrement by one when there is a '.'
+  return digit_count - size_t(start - start_digits);
+}
+
+} // unnamed namespace
+
+/** @private */
+static error_code slow_float_parsing(simdjson_unused const uint8_t * src, double* answer) {
+  if (parse_float_fallback(src, answer)) {
+    return SUCCESS;
+  }
+  return INVALID_NUMBER(src);
+}
+
+/** @private */
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
+  // If we frequently had to deal with long strings of digits,
+  // we could extend our code by using a 128-bit integer instead
+  // of a 64-bit integer. However, this is uncommon in practice.
+  //
+  // 9999999999999999999 < 2**64 so we can accommodate 19 digits.
+  // If we have a decimal separator, then digit_count - 1 is the number of digits, but we
+  // may not have a decimal separator!
+  if (simdjson_unlikely(digit_count > 19 && significant_digits(start_digits, digit_count) > 19)) {
+    // Ok, chances are good that we had an overflow!
+    // this is almost never going to get called!!!
+    // we start anew, going slowly!!!
+    // This will happen in the following examples:
+    // 10000000000000000000000000000000000000000000e+308
+    // 3.1415926535897932384626433832795028841971693993751
+    //
+    // NOTE: We do not pass a reference to the to slow_float_parsing. If we passed our writer
+    // reference to it, it would force it to be stored in memory, preventing the compiler from
+    // picking it apart and putting into registers. i.e. if we pass it as reference,
+    // it gets slow.
+    double d;
+    error_code error = slow_float_parsing(src, &d);
+    writer.append_double(d);
+    return error;
+  }
+  // NOTE: it's weird that the simdjson_unlikely() only wraps half the if, but it seems to get slower any other
+  // way we've tried: https://github.com/simdjson/simdjson/pull/990#discussion_r448497331
+  // To future reader: we'd love if someone found a better way, or at least could explain this result!
+  if (simdjson_unlikely(exponent < simdjson::internal::smallest_power) || (exponent > simdjson::internal::largest_power)) {
+    //
+    // Important: smallest_power is such that it leads to a zero value.
+    // Observe that 18446744073709551615e-343 == 0, i.e. (2**64 - 1) e -343 is zero
+    // so something x 10^-343 goes to zero, but not so with  something x 10^-342.
+    static_assert(simdjson::internal::smallest_power <= -342, "smallest_power is not small enough");
+    //
+    if((exponent < simdjson::internal::smallest_power) || (i == 0)) {
+      // E.g. Parse "-0.0e-999" into the same value as "-0.0". See https://en.wikipedia.org/wiki/Signed_zero
+      WRITE_DOUBLE(negative ? -0.0 : 0.0, src, writer);
+      return SUCCESS;
+    } else { // (exponent > largest_power) and (i != 0)
+      // We have, for sure, an infinite value and simdjson refuses to parse infinite values.
+      return INVALID_NUMBER(src);
+    }
+  }
+  double d;
+  if (!compute_float_64(exponent, i, negative, d)) {
+    // we are almost never going to get here.
+    if (!parse_float_fallback(src, &d)) { return INVALID_NUMBER(src); }
+  }
+  WRITE_DOUBLE(d, src, writer);
+  return SUCCESS;
+}
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer);
+
+// for performance analysis, it is sometimes  useful to skip parsing
+#ifdef SIMDJSON_SKIPNUMBERPARSING
+
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const, W &writer) {
+  writer.append_s64(0);        // always write zero
+  return SUCCESS;              // always succeeds
+}
+
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept { return number_type::signed_integer; }
+#else
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer) {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  if (digit_count == 0 || ('0' == *start_digits && digit_count > 1)) { return INVALID_NUMBER(src); }
+
+  //
+  // Handle floats if there is a . or e (or both)
+  //
+  int64_t exponent = 0;
+  bool is_float = false;
+  if ('.' == *p) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_decimal_after_separator(src, p, i, exponent) );
+    digit_count = int(p - start_digits); // used later to guard against overflows
+  }
+  if (('e' == *p) || ('E' == *p)) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_exponent(src, p, exponent) );
+  }
+  if (is_float) {
+    const bool dirty_end = jsoncharutils::is_not_structural_or_whitespace(*p);
+    SIMDJSON_TRY( write_float(src, negative, i, start_digits, digit_count, exponent, writer) );
+    if (dirty_end) { return INVALID_NUMBER(src); }
+    return SUCCESS;
+  }
+
+  // The longest negative 64-bit number is 19 digits.
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  size_t longest_digit_count = negative ? 19 : 20;
+  if (digit_count > longest_digit_count) { return BIGINT_NUMBER(src); }
+  if (digit_count == longest_digit_count) {
+    if (negative) {
+      // Anything negative above INT64_MAX+1 is invalid
+      if (i > uint64_t(INT64_MAX)+1) { return BIGINT_NUMBER(src);  }
+      WRITE_INTEGER(~i+1, src, writer);
+      if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+      return SUCCESS;
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    }  else if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INVALID_NUMBER(src); }
+  }
+
+  // Write unsigned if it does not fit in a signed integer.
+  if (i > uint64_t(INT64_MAX)) {
+    WRITE_UNSIGNED(i, src, writer);
+  } else {
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+    if(i == 0 && negative) {
+      // We have to write -0.0 instead of 0
+      WRITE_DOUBLE(-0.0, src, writer);
+    } else {
+      WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+    }
+#else
+  WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+#endif
+  }
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+  return SUCCESS;
+}
+
+// Inlineable functions
+namespace {
+
+// This table can be used to characterize the final character of an integer
+// string. For JSON structural character and allowable white space characters,
+// we return SUCCESS. For 'e', '.' and 'E', we return INCORRECT_TYPE. Otherwise
+// we return NUMBER_ERROR.
+// Optimization note: we could easily reduce the size of the table by half (to 128)
+// at the cost of an extra branch.
+// Optimization note: we want the values to use at most 8 bits (not, e.g., 32 bits):
+static_assert(error_code(uint8_t(NUMBER_ERROR))== NUMBER_ERROR, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(SUCCESS))== SUCCESS, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(INCORRECT_TYPE))== INCORRECT_TYPE, "bad NUMBER_ERROR cast");
+
+const uint8_t integer_string_finisher[256] = {
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   SUCCESS,      NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, INCORRECT_TYPE,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, SUCCESS,        NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR};
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if ((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
+  const uint8_t *p = src + 1;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    // Note: we use src[1] and not src[0] because src[0] is the quote character in this
+    // instance.
+    if (src[1] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  //
+  // Check for minus sign
+  //
+  if(src == src_end) { return NUMBER_ERROR; }
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = src;
+  uint64_t i = 0;
+  while (parse_digit(*src, i)) { src++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(src - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*src)) {
+  //  return (*src == '.' || *src == 'e' || *src == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(*src != '"') { return NUMBER_ERROR; }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept {
+  return (*src == '-');
+}
+
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) { return true; }
+  return false;
+}
+
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  size_t digit_count = size_t(p - src);
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) {
+    static const uint8_t * smaller_big_integer = reinterpret_cast<const uint8_t *>("9223372036854775808");
+    // We have an integer.
+    if(simdjson_unlikely(digit_count > 20)) {
+      return number_type::big_integer;
+    }
+    // If the number is negative and valid, it must be a signed integer.
+    if(negative) {
+      if (simdjson_unlikely(digit_count > 19)) return number_type::big_integer;
+      if (simdjson_unlikely(digit_count == 19 && memcmp(src, smaller_big_integer, 19) > 0)) {
+        return number_type::big_integer;
+      }
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+      if(digit_count == 1 && src[0] == '0') {
+        // We have to write -0.0 instead of 0
+        return number_type::floating_point_number;
+      }
+#endif
+      return number_type::signed_integer;
+    }
+    // Let us check if we have a big integer (>=2**64).
+    static const uint8_t * two_to_sixtyfour = reinterpret_cast<const uint8_t *>("18446744073709551616");
+    if((digit_count > 20) || (digit_count == 20 && memcmp(src, two_to_sixtyfour, 20) >= 0)) {
+      return number_type::big_integer;
+    }
+    // The number is positive and smaller than 18446744073709551616 (or 2**64).
+    // We want values larger or equal to 9223372036854775808 to be unsigned
+    // integers, and the other values to be signed integers.
+    if((digit_count == 20) || (digit_count >= 19 && memcmp(src, smaller_big_integer, 19) >= 0)) {
+      return number_type::unsigned_integer;
+    }
+    return number_type::signed_integer;
+  }
+  // Hopefully, we have 'e' or 'E' or '.'.
+  return number_type::floating_point_number;
+}
+
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
+  if(src == src_end) { return NUMBER_ERROR; }
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  if(p == src_end) { return NUMBER_ERROR; }
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely((p != src_end) && (*p == '.'))) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if ((p == src_end) || !parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while ((p != src_end) && parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = start_digits-src > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if ((p != src_end) && (*p == 'e' || *p == 'E')) {
+    p++;
+    if(p == src_end) { return NUMBER_ERROR; }
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while ((p != src_end) && parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if ((p != src_end) && jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), src_end, &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+} // unnamed namespace
+#endif // SIMDJSON_SKIPNUMBERPARSING
+
+} // namespace numberparsing
+
+inline std::ostream& operator<<(std::ostream& out, number_type type) noexcept {
+    switch (type) {
+        case number_type::signed_integer: out << "integer in [-9223372036854775808,9223372036854775808)"; break;
+        case number_type::unsigned_integer: out << "unsigned integer in [9223372036854775808,18446744073709551616)"; break;
+        case number_type::floating_point_number: out << "floating-point number (binary64)"; break;
+        case number_type::big_integer: out << "big integer"; break;
+        default: SIMDJSON_UNREACHABLE();
+    }
+    return out;
+}
+
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_NUMBERPARSING_H
+/* end file simdjson/generic/numberparsing.h for rvv_vls */
+
+/* including simdjson/generic/implementation_simdjson_result_base-inl.h for rvv_vls: #include "simdjson/generic/implementation_simdjson_result_base-inl.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base-inl.h for rvv_vls */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+
+//
+// internal::implementation_simdjson_result_base<T> inline implementation
+//
+
+template<typename T>
+simdjson_inline void implementation_simdjson_result_base<T>::tie(T &value, error_code &error) && noexcept {
+  error = this->second;
+  if (!error) {
+    value = std::forward<implementation_simdjson_result_base<T>>(*this).first;
+  }
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::get(T &value) && noexcept {
+  error_code error;
+  std::forward<implementation_simdjson_result_base<T>>(*this).tie(value, error);
+  return error;
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::error() const noexcept {
+  return this->second;
+}
+
+
+template<typename T>
+simdjson_warn_unused simdjson_inline bool implementation_simdjson_result_base<T>::has_value() const noexcept {
+  return this->error() == SUCCESS;
+}
+
+#if SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::operator*() &  noexcept(false) {
+  return this->value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::operator*() &&  noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).value();
+}
+
+template<typename T>
+simdjson_inline T* implementation_simdjson_result_base<T>::operator->() noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+
+template<typename T>
+simdjson_inline const T* implementation_simdjson_result_base<T>::operator->() const noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value() & noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::take_value() && noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::operator T&&() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+#endif // SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline const T& implementation_simdjson_result_base<T>::value_unsafe() const& noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value_unsafe() & noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value_unsafe() && noexcept {
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value, error_code error) noexcept
+    : first{std::forward<T>(value)}, second{error} {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(error_code error) noexcept
+    : implementation_simdjson_result_base(T{}, error) {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value) noexcept
+    : implementation_simdjson_result_base(std::forward<T>(value), SUCCESS) {}
+
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+/* end file simdjson/generic/implementation_simdjson_result_base-inl.h for rvv_vls */
+/* end file simdjson/generic/amalgamated.h for rvv_vls */
+/* including simdjson/rvv-vls/end.h: #include "simdjson/rvv-vls/end.h" */
+/* begin file simdjson/rvv-vls/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+/* undefining SIMDJSON_IMPLEMENTATION from "rvv_vls" */
+#undef SIMDJSON_IMPLEMENTATION
+/* end file simdjson/rvv-vls/end.h */
+
+#endif // SIMDJSON_RVV_VLS_H
+/* end file simdjson/rvv-vls.h */
+/* including simdjson/rvv-vls/implementation.h: #include <simdjson/rvv-vls/implementation.h> */
+/* begin file simdjson/rvv-vls/implementation.h */
+#ifndef SIMDJSON_RVV_VLS_IMPLEMENTATION_H
+#define SIMDJSON_RVV_VLS_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation(
+      "rvv_vls",
+      "RISC-V V extension",
+      0
+  ) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<simdjson::internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_RVV_VLS_IMPLEMENTATION_H
+/* end file simdjson/rvv-vls/implementation.h */
+
+/* including simdjson/rvv-vls/begin.h: #include <simdjson/rvv-vls/begin.h> */
+/* begin file simdjson/rvv-vls/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "rvv_vls" */
+#define SIMDJSON_IMPLEMENTATION rvv_vls
+/* including simdjson/rvv-vls/base.h: #include "simdjson/rvv-vls/base.h" */
+/* begin file simdjson/rvv-vls/base.h */
+#ifndef SIMDJSON_RVV_VLS_BASE_H
+#define SIMDJSON_RVV_VLS_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+/**
+ * RVV-VLS implementation.
+ */
+namespace rvv_vls {
+
+class implementation;
+
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_RVV_VLS_BASE_H
+/* end file simdjson/rvv-vls/base.h */
+/* including simdjson/rvv-vls/intrinsics.h: #include "simdjson/rvv-vls/intrinsics.h" */
+/* begin file simdjson/rvv-vls/intrinsics.h */
+#ifndef SIMDJSON_RVV_VLS_INTRINSICS_H
+#define SIMDJSON_RVV_VLS_INTRINSICS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <riscv_vector.h>
+
+#define simdutf_vrgather_u8m1x2(tbl, idx)                                      \
+  __riscv_vcreate_v_u8m1_u8m2(                                                 \
+      __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m2_u8m1(idx, 0),          \
+                               __riscv_vsetvlmax_e8m1()),                      \
+      __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m2_u8m1(idx, 1),          \
+                               __riscv_vsetvlmax_e8m1()))
+
+#define simdutf_vrgather_u8m1x4(tbl, idx)                                      \
+  __riscv_vcreate_v_u8m1_u8m4(                                                 \
+      __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 0),          \
+                               __riscv_vsetvlmax_e8m1()),                      \
+      __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 1),          \
+                               __riscv_vsetvlmax_e8m1()),                      \
+      __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 2),          \
+                               __riscv_vsetvlmax_e8m1()),                      \
+      __riscv_vrgather_vv_u8m1(tbl, __riscv_vget_v_u8m4_u8m1(idx, 3),          \
+                               __riscv_vsetvlmax_e8m1()))
+
+#if __riscv_zbc
+#include <riscv_bitmanip.h>
+#endif
+
+#endif // SIMDJSON_RVV_VLS_INTRINSICS_H
+/* end file simdjson/rvv-vls/intrinsics.h */
+/* including simdjson/rvv-vls/bitmanipulation.h: #include "simdjson/rvv-vls/bitmanipulation.h" */
+/* begin file simdjson/rvv-vls/bitmanipulation.h */
+#ifndef SIMDJSON_RVV_VLS_BITMANIPULATION_H
+#define SIMDJSON_RVV_VLS_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+
+// We sometimes call trailing_zero on inputs that are zero,
+// but the algorithms do not end up using the returned value.
+// Sadly, sanitizers are not smart enough to figure it out.
+SIMDJSON_NO_SANITIZE_UNDEFINED
+// This function can be used safely even if not all bytes have been
+// initialized.
+// See issue https://github.com/simdjson/simdjson/issues/1965
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+  return __builtin_ctzll(input_num);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+  return input_num & (input_num-1);
+}
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+  return __builtin_clzll(input_num);
+}
+
+simdjson_inline long long int count_ones(uint64_t input_num) {
+  return __builtin_popcountll(input_num);
+}
+
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2,
+                                uint64_t *result) {
+  return __builtin_uaddll_overflow(value1, value2,
+                                   reinterpret_cast<unsigned long long *>(result));
+}
+
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_RVV_VLS_BITMANIPULATION_H
+/* end file simdjson/rvv-vls/bitmanipulation.h */
+/* including simdjson/rvv-vls/bitmask.h: #include "simdjson/rvv-vls/bitmask.h" */
+/* begin file simdjson/rvv-vls/bitmask.h */
+#ifndef SIMDJSON_RVV_VLS_BITMASK_H
+#define SIMDJSON_RVV_VLS_BITMASK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/intrinsics.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+
+//
+// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
+//
+// For example, prefix_xor(00100100) == 00011100
+//
+simdjson_inline uint64_t prefix_xor(uint64_t bitmask) {
+#if __riscv_zbc
+  return __riscv_clmul_64(bitmask, ~(uint64_t)0);
+#elif __riscv_zvbc
+  return __riscv_vmv_x(__riscv_vclmul(__riscv_vmv_s_x_u64m1(bitmask, 1), ~(uint64_t)0, 1));
+#else
+  bitmask ^= bitmask << 1;
+  bitmask ^= bitmask << 2;
+  bitmask ^= bitmask << 4;
+  bitmask ^= bitmask << 8;
+  bitmask ^= bitmask << 16;
+  bitmask ^= bitmask << 32;
+#endif
+  return bitmask;
+}
+
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_RVV_VLS_BITMASK_H
+
+/* end file simdjson/rvv-vls/bitmask.h */
+/* including simdjson/rvv-vls/simd.h: #include "simdjson/rvv-vls/simd.h" */
+/* begin file simdjson/rvv-vls/simd.h */
+#ifndef SIMDJSON_RVV_VLS_SIMD_H
+#define SIMDJSON_RVV_VLS_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace simd {
+
+#if __riscv_v_fixed_vlen >= 512
+  static constexpr size_t VL8 = 512/8;
+  using vint8_t   = vint8m1_t   __attribute__((riscv_rvv_vector_bits(512)));
+  using vuint8_t  = vuint8m1_t  __attribute__((riscv_rvv_vector_bits(512)));
+  using vbool_t    = vbool8_t  __attribute__((riscv_rvv_vector_bits(512/8)));
+  using vbitmask_t = uint64_t;
+#else
+  static constexpr size_t VL8 = __riscv_v_fixed_vlen/8;
+  using vint8_t   = vint8m1_t   __attribute__((riscv_rvv_vector_bits(__riscv_v_fixed_vlen)));
+  using vuint8_t  = vuint8m1_t  __attribute__((riscv_rvv_vector_bits(__riscv_v_fixed_vlen)));
+  using vbool_t    = vbool8_t  __attribute__((riscv_rvv_vector_bits(__riscv_v_fixed_vlen/8)));
+  #if __riscv_v_fixed_vlen == 128
+    using vbitmask_t = uint16_t;
+  #elif __riscv_v_fixed_vlen == 256
+    using vbitmask_t = uint32_t;
+  #endif
+#endif
+
+#if __riscv_v_fixed_vlen == 128
+  using vuint8x64_t = vuint8m4_t __attribute__((riscv_rvv_vector_bits(512)));
+  using vboolx64_t    = vbool2_t  __attribute__((riscv_rvv_vector_bits(512/8)));
+#elif __riscv_v_fixed_vlen == 256
+  using vuint8x64_t = vuint8m2_t __attribute__((riscv_rvv_vector_bits(512)));
+  using vboolx64_t    = vbool4_t  __attribute__((riscv_rvv_vector_bits(512/8)));
+#else
+  using vuint8x64_t = vuint8m1_t __attribute__((riscv_rvv_vector_bits(512)));
+  using vboolx64_t    = vbool8_t  __attribute__((riscv_rvv_vector_bits(512/8)));
+#endif
+
+  template<typename T>
+  struct simd8;
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool> {
+    vbool_t value;
+    using bitmask_t = vbitmask_t;
+    static constexpr int SIZE = sizeof(value);
+
+    simdjson_inline simd8(const vbool_t _value) : value(_value) {}
+    simdjson_inline simd8() : simd8(__riscv_vmclr_m_b8(VL8)) {}
+    simdjson_inline simd8(bool _value) : simd8(splat(_value)) {}
+
+    simdjson_inline operator const vbool_t&() const { return value; }
+    simdjson_inline operator vbool_t&() { return value; }
+
+    static simdjson_inline simd8<bool> splat(bool _value) {
+      return __riscv_vreinterpret_b8(__riscv_vmv_v_x_u64m1(((uint64_t)!_value)-1, 1));
+    }
+
+    simdjson_inline vbitmask_t to_bitmask() const {
+#if __riscv_v_fixed_vlen == 128
+      return __riscv_vmv_x(__riscv_vreinterpret_u16m1(value));
+#elif __riscv_v_fixed_vlen == 256
+      return __riscv_vmv_x(__riscv_vreinterpret_u32m1(value));
+#else
+      return __riscv_vmv_x(__riscv_vreinterpret_u64m1(value));
+#endif
+    }
+
+    // Bit operations
+    simdjson_inline simd8<bool> operator|(const simd8<bool> other) const {  return __riscv_vmor(*this, other, VL8); }
+    simdjson_inline simd8<bool> operator&(const simd8<bool> other) const {  return __riscv_vmand(*this, other, VL8); }
+    simdjson_inline simd8<bool> operator^(const simd8<bool> other) const {  return __riscv_vmxor(*this, other, VL8); }
+    simdjson_inline simd8<bool> bit_andnot(const simd8<bool> other) const { return __riscv_vmandn(other, *this, VL8); }
+    simdjson_inline simd8<bool> operator~() const { return __riscv_vmnot(*this, VL8); }
+    simdjson_inline simd8<bool>& operator|=(const simd8<bool> other) { auto this_cast = static_cast<simd8<bool>*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline simd8<bool>& operator&=(const simd8<bool> other) { auto this_cast = static_cast<simd8<bool>*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline simd8<bool>& operator^=(const simd8<bool> other) { auto this_cast = static_cast<simd8<bool>*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t> {
+
+    vuint8_t value;
+    static constexpr int SIZE = sizeof(value);
+
+    simdjson_inline simd8(const vuint8_t _value) : value(_value) {}
+    simdjson_inline simd8() : simd8(zero()) {}
+    simdjson_inline simd8(const uint8_t values[VL8]) : simd8(load(values)) {}
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    simdjson_inline simd8(simd8<bool> mask) : value(__riscv_vmerge_vxm_u8m1(zero(), -1, (vbool_t)mask, VL8)) {}
+
+    simdjson_inline operator const vuint8_t&() const { return this->value; }
+    simdjson_inline operator vuint8_t&() { return this->value; }
+
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) : simd8(vuint8_t{
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    }) {}
+
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    static simdjson_inline vuint8_t splat(uint8_t _value) { return __riscv_vmv_v_x_u8m1(_value, VL8); }
+    static simdjson_inline vuint8_t zero() { return splat(0); }
+    static simdjson_inline vuint8_t load(const uint8_t values[VL8]) { return __riscv_vle8_v_u8m1(values, VL8); }
+
+    // Bit operations
+    simdjson_inline simd8<uint8_t> operator|(const simd8<uint8_t> other) const { return  __riscv_vor_vv_u8m1(  value, other, VL8); }
+    simdjson_inline simd8<uint8_t> operator&(const simd8<uint8_t> other) const { return  __riscv_vand_vv_u8m1( value, other, VL8); }
+    simdjson_inline simd8<uint8_t> operator^(const simd8<uint8_t> other) const { return  __riscv_vxor_vv_u8m1( value, other, VL8); }
+    simdjson_inline simd8<uint8_t> operator~() const { return __riscv_vnot_v_u8m1(value, VL8); }
+#if __riscv_zvbb
+    simdjson_inline simd8<uint8_t> bit_andnot(const simd8<uint8_t> other) const { return __riscv_vandn_vv_u8m1(other, value, VL8); }
+#else
+    simdjson_inline simd8<uint8_t> bit_andnot(const simd8<uint8_t> other) const { return other & ~*this; }
+#endif
+    simdjson_inline simd8<uint8_t>& operator|=(const simd8<uint8_t> other) { value = *this | other; return *this; }
+    simdjson_inline simd8<uint8_t>& operator&=(const simd8<uint8_t> other) { value = *this & other; return *this; }
+    simdjson_inline simd8<uint8_t>& operator^=(const simd8<uint8_t> other) { value = *this ^ other; return *this; }
+
+    simdjson_inline simd8<bool> operator==(const simd8<uint8_t> other) const { return __riscv_vmseq(value, other, VL8); }
+    simdjson_inline simd8<bool> operator==(uint8_t other) const { return __riscv_vmseq(value, other, VL8); }
+
+    template<int N=1>
+    simdjson_inline simd8<uint8_t> prev(const simd8<uint8_t> prev_chunk) const {
+      return __riscv_vslideup(__riscv_vslidedown(prev_chunk, VL8-N, VL8), value, N, VL8);
+    }
+
+    // Store to array
+    simdjson_inline void store(uint8_t dst[VL8]) const { return __riscv_vse8(dst, value, VL8); }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return __riscv_vsaddu(value, other, VL8); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return __riscv_vssubu(value, other, VL8); }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<uint8_t> operator+(const simd8<uint8_t> other) const { return __riscv_vadd(value, other, VL8); }
+    simdjson_inline simd8<uint8_t> operator-(const simd8<uint8_t> other) const { return __riscv_vsub(value, other, VL8); }
+    simdjson_inline simd8<uint8_t>& operator+=(const simd8<uint8_t> other) { value = *this + other; return *this; }
+    simdjson_inline simd8<uint8_t>& operator-=(const simd8<uint8_t> other) { value = *this - other; return *this; }
+
+    // Order-specific operations
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return __riscv_vmsleu(value, other, VL8); }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return __riscv_vmsgeu(value, other, VL8); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const  { return __riscv_vmsltu(value, other, VL8); }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const  { return __riscv_vmsgtu(value, other, VL8); }
+
+    // Same as >, but instead of guaranteeing all 1's == true, false = 0 and true = nonzero.
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return simd8<uint8_t>(*this > other); }
+    // Same as <, but instead of guaranteeing all 1's == true, false = 0 and true = nonzero.
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return simd8<uint8_t>(*this < other); }
+
+    // Bit-specific operations
+    simdjson_inline bool any_bits_set_anywhere() const {
+      return __riscv_vfirst(__riscv_vmsne(value, 0, VL8), VL8) >= 0;
+    }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return (*this & bits).any_bits_set_anywhere(); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return __riscv_vsrl(value, N, VL8); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return __riscv_vsll(value, N, VL8); }
+
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return __riscv_vrgather(lookup_table, value, VL8);
+    }
+
+    // compress inactive elements, to match AVX-512 behavior
+    template<typename L>
+    simdjson_inline void compress(vbitmask_t mask, L * output) const {
+      mask = (vbitmask_t)~mask;
+#if __riscv_v_fixed_vlen == 128
+      vbool8_t m = __riscv_vreinterpret_b8(__riscv_vmv_s_x_u16m1(mask, 1));
+#elif __riscv_v_fixed_vlen == 256
+      vbool8_t m = __riscv_vreinterpret_b8(__riscv_vmv_s_x_u32m1(mask, 1));
+#else
+      vbool8_t m = __riscv_vreinterpret_b8(__riscv_vmv_s_x_u64m1(mask, 1));
+#endif
+      __riscv_vse8_v_u8m1(output, __riscv_vcompress(value, m, VL8), count_ones(mask));
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> {
+    vint8_t value;
+    static constexpr int SIZE = sizeof(value);
+
+    simdjson_inline simd8(const vint8_t _value) : value(_value) {}
+    simdjson_inline simd8() : simd8(zero()) {}
+    simdjson_inline simd8(const int8_t values[VL8]) : simd8(load(values)) {}
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+
+    simdjson_inline operator const vint8_t&() const { return this->value; }
+    simdjson_inline operator vint8_t&() { return this->value; }
+
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) : simd8(vint8_t{
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    }) {}
+
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    static simdjson_inline vint8_t splat(int8_t _value) { return __riscv_vmv_v_x_i8m1(_value, VL8); }
+    static simdjson_inline vint8_t zero() { return splat(0); }
+    static simdjson_inline vint8_t load(const int8_t values[VL8]) { return __riscv_vle8_v_i8m1(values, VL8); }
+
+
+    simdjson_inline void store(int8_t dst[VL8]) const { return __riscv_vse8(dst, value, VL8); }
+
+    // Explicit conversion to/from unsigned
+    simdjson_inline explicit simd8(const vuint8_t other): simd8(__riscv_vreinterpret_i8m1(other)) {}
+    simdjson_inline explicit operator simd8<uint8_t>() const { return __riscv_vreinterpret_u8m1(value); }
+
+    // Math
+    simdjson_inline simd8<int8_t> operator+(const simd8<int8_t> other) const { return __riscv_vadd(value, other, VL8); }
+    simdjson_inline simd8<int8_t> operator-(const simd8<int8_t> other) const { return __riscv_vsub(value, other, VL8); }
+    simdjson_inline simd8<int8_t>& operator+=(const simd8<int8_t> other) { value = *this + other; return *this; }
+    simdjson_inline simd8<int8_t>& operator-=(const simd8<int8_t> other) { value = *this - other; return *this; }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val( const simd8<int8_t> other) const { return __riscv_vmax( value, other, VL8); }
+    simdjson_inline simd8<int8_t> min_val( const simd8<int8_t> other) const { return __riscv_vmin( value, other, VL8); }
+    simdjson_inline simd8<bool> operator>( const simd8<int8_t> other) const { return __riscv_vmsgt(value, other, VL8); }
+    simdjson_inline simd8<bool> operator<( const simd8<int8_t> other) const { return __riscv_vmslt(value, other, VL8); }
+    simdjson_inline simd8<bool> operator==(const simd8<int8_t> other) const { return __riscv_vmseq(value, other, VL8); }
+
+    template<int N=1>
+    simdjson_inline simd8<int8_t> prev(const simd8<int8_t> prev_chunk) const {
+      return __riscv_vslideup(__riscv_vslidedown(prev_chunk, VL8-N, VL8), value, N, VL8);
+    }
+
+    // Perform a lookup assuming no value is larger than 16
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return __riscv_vrgather(lookup_table, value, VL8);
+    }
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  template<typename T>
+  struct simd8x64;
+  template<>
+  struct simd8x64<uint8_t> {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<uint8_t>);
+    vuint8x64_t value;
+
+#if __riscv_v_fixed_vlen >= 512
+    template<int idx> simd8<uint8_t> get() const { return value; }
+#else
+    template<int idx> simd8<uint8_t> get() const { return __riscv_vget_u8m1(value, idx); }
+#endif
+
+    simdjson_inline operator const vuint8x64_t&() const { return this->value; }
+    simdjson_inline operator vuint8x64_t&() { return this->value; }
+
+    simd8x64(const simd8x64<uint8_t>& o) = delete; // no copy allowed
+    simd8x64<uint8_t>& operator=(const simd8<uint8_t>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+#if __riscv_v_fixed_vlen == 128
+    simdjson_inline simd8x64(const uint8_t *ptr, size_t n = 64) : value(__riscv_vle8_v_u8m4(ptr, n)) {}
+#elif __riscv_v_fixed_vlen == 256
+    simdjson_inline simd8x64(const uint8_t *ptr, size_t n = 64) : value(__riscv_vle8_v_u8m2(ptr, n)) {}
+#else
+    simdjson_inline simd8x64(const uint8_t *ptr, size_t n = 64) : value(__riscv_vle8_v_u8m1(ptr, n)) {}
+#endif
+
+    simdjson_inline void store(uint8_t ptr[64]) const {
+      __riscv_vse8(ptr, value, 64);
+    }
+
+    simdjson_inline bool is_ascii() const {
+#if __riscv_v_fixed_vlen == 128
+      return __riscv_vfirst(__riscv_vmslt(__riscv_vreinterpret_i8m4(value), 0, 64), 64) < 0;
+#elif __riscv_v_fixed_vlen == 256
+      return __riscv_vfirst(__riscv_vmslt(__riscv_vreinterpret_i8m2(value), 0, 64), 64) < 0;
+#else
+      return __riscv_vfirst(__riscv_vmslt(__riscv_vreinterpret_i8m1(value), 0, 64), 64) < 0;
+#endif
+    }
+
+    // compress inactive elements, to match AVX-512 behavior
+    simdjson_inline uint64_t compress(uint64_t mask, uint8_t * output) const {
+      mask = ~mask;
+#if __riscv_v_fixed_vlen == 128
+      vboolx64_t m = __riscv_vreinterpret_b2(__riscv_vmv_s_x_u64m1(mask, 1));
+#elif __riscv_v_fixed_vlen == 256
+      vboolx64_t m = __riscv_vreinterpret_b4(__riscv_vmv_s_x_u64m1(mask, 1));
+#else
+      vboolx64_t m = __riscv_vreinterpret_b8(__riscv_vmv_s_x_u64m1(mask, 1));
+#endif
+      size_t cnt = count_ones(mask);
+      __riscv_vse8(output, __riscv_vcompress(value, m, 64), cnt);
+      return cnt;
+    }
+
+    simdjson_inline uint64_t eq(const uint8_t m) const {
+      return __riscv_vmv_x(__riscv_vreinterpret_u64m1(__riscv_vmseq(value, m, 64)));
+    }
+
+    simdjson_inline uint64_t lteq(const uint8_t m) const {
+      return __riscv_vmv_x(__riscv_vreinterpret_u64m1(__riscv_vmsleu(value, m, 64)));
+    }
+  }; // struct simd8x64<uint8_t>
+
+} // namespace simd
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_RVV_VLS_SIMD_H
+/* end file simdjson/rvv-vls/simd.h */
+/* including simdjson/rvv-vls/stringparsing_defs.h: #include "simdjson/rvv-vls/stringparsing_defs.h" */
+/* begin file simdjson/rvv-vls/stringparsing_defs.h */
+#ifndef SIMDJSON_RVV_VLS_STRINGPARSING_DEFS_H
+#define SIMDJSON_RVV_VLS_STRINGPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+
+using namespace simd;
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint64_t BYTES_PROCESSED = sizeof(simd8<uint8_t>);
+  simdjson_inline backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return ((quote_bits - 1) & bs_bits) != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+
+  uint64_t bs_bits;
+  uint64_t quote_bits;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  return { (v == '\\').to_bitmask(), (v == '"').to_bitmask() };
+}
+
+struct escaping {
+  static constexpr uint64_t BYTES_PROCESSED = sizeof(simd8<uint8_t>);
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits != 0; }
+  simdjson_inline int escape_index() { return trailing_zeroes(escape_bits) / 4; }
+
+  uint64_t escape_bits;
+}; // struct escaping
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "escaping finder must process fewer than SIMDJSON_PADDING bytes");
+  simd8<uint8_t> v(src);
+  v.store(dst);
+  return { ((v == '"') | (v == '\\') | (v == 32)).to_bitmask() };
+}
+
+
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_RVV_VLS_STRINGPARSING_DEFS_H
+/* end file simdjson/rvv-vls/stringparsing_defs.h */
+/* including simdjson/rvv-vls/numberparsing_defs.h: #include "simdjson/rvv-vls/numberparsing_defs.h" */
+/* begin file simdjson/rvv-vls/numberparsing_defs.h */
+#ifndef SIMDJSON_RVV_VLS_NUMBERPARSING_DEFS_H
+#define SIMDJSON_RVV_VLS_NUMBERPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+#ifdef JSON_TEST_NUMBERS // for unit testing
+void found_invalid_number(const uint8_t *buf);
+void found_integer(int64_t result, const uint8_t *buf);
+void found_unsigned_integer(uint64_t result, const uint8_t *buf);
+void found_float(double result, const uint8_t *buf);
+#endif
+
+namespace simdjson {
+namespace rvv_vls {
+namespace numberparsing {
+
+// credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const char *chars) {
+  uint64_t val;
+#if __riscv_misaligned_fast
+  memcpy(&val, chars, sizeof(uint64_t));
+#else
+  val = __riscv_vmv_x(__riscv_vreinterpret_u64m1(__riscv_vlmul_ext_u8m1(__riscv_vle8_v_u8mf2((uint8_t*)chars, 8))));
+#endif
+  val = (val & 0x0F0F0F0F0F0F0F0F) * 2561 >> 8;
+  val = (val & 0x00FF00FF00FF00FF) * 6553601 >> 16;
+  return uint32_t((val & 0x0000FFFF0000FFFF) * 42949672960001 >> 32);
+}
+
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  return parse_eight_digits_unrolled(reinterpret_cast<const char *>(chars));
+}
+
+/** @private */
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace rvv_vls
+} // namespace simdjson
+
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+
+#endif // SIMDJSON_RVV_VLS_NUMBERPARSING_DEFS_H
+/* end file simdjson/rvv-vls/numberparsing_defs.h */
+
+#define SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT 1
+/* end file simdjson/rvv-vls/begin.h */
+/* including simdjson/rvv-vls/simd.h: #include <simdjson/rvv-vls/simd.h> */
+/* begin file simdjson/rvv-vls/simd.h */
+#ifndef SIMDJSON_RVV_VLS_SIMD_H
+#define SIMDJSON_RVV_VLS_SIMD_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/bitmanipulation.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/simdprune_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace simd {
+
+#if __riscv_v_fixed_vlen >= 512
+  static constexpr size_t VL8 = 512/8;
+  using vint8_t   = vint8m1_t   __attribute__((riscv_rvv_vector_bits(512)));
+  using vuint8_t  = vuint8m1_t  __attribute__((riscv_rvv_vector_bits(512)));
+  using vbool_t    = vbool8_t  __attribute__((riscv_rvv_vector_bits(512/8)));
+  using vbitmask_t = uint64_t;
+#else
+  static constexpr size_t VL8 = __riscv_v_fixed_vlen/8;
+  using vint8_t   = vint8m1_t   __attribute__((riscv_rvv_vector_bits(__riscv_v_fixed_vlen)));
+  using vuint8_t  = vuint8m1_t  __attribute__((riscv_rvv_vector_bits(__riscv_v_fixed_vlen)));
+  using vbool_t    = vbool8_t  __attribute__((riscv_rvv_vector_bits(__riscv_v_fixed_vlen/8)));
+  #if __riscv_v_fixed_vlen == 128
+    using vbitmask_t = uint16_t;
+  #elif __riscv_v_fixed_vlen == 256
+    using vbitmask_t = uint32_t;
+  #endif
+#endif
+
+#if __riscv_v_fixed_vlen == 128
+  using vuint8x64_t = vuint8m4_t __attribute__((riscv_rvv_vector_bits(512)));
+  using vboolx64_t    = vbool2_t  __attribute__((riscv_rvv_vector_bits(512/8)));
+#elif __riscv_v_fixed_vlen == 256
+  using vuint8x64_t = vuint8m2_t __attribute__((riscv_rvv_vector_bits(512)));
+  using vboolx64_t    = vbool4_t  __attribute__((riscv_rvv_vector_bits(512/8)));
+#else
+  using vuint8x64_t = vuint8m1_t __attribute__((riscv_rvv_vector_bits(512)));
+  using vboolx64_t    = vbool8_t  __attribute__((riscv_rvv_vector_bits(512/8)));
+#endif
+
+  template<typename T>
+  struct simd8;
+
+  // SIMD byte mask type (returned by things like eq and gt)
+  template<>
+  struct simd8<bool> {
+    vbool_t value;
+    using bitmask_t = vbitmask_t;
+    static constexpr int SIZE = sizeof(value);
+
+    simdjson_inline simd8(const vbool_t _value) : value(_value) {}
+    simdjson_inline simd8() : simd8(__riscv_vmclr_m_b8(VL8)) {}
+    simdjson_inline simd8(bool _value) : simd8(splat(_value)) {}
+
+    simdjson_inline operator const vbool_t&() const { return value; }
+    simdjson_inline operator vbool_t&() { return value; }
+
+    static simdjson_inline simd8<bool> splat(bool _value) {
+      return __riscv_vreinterpret_b8(__riscv_vmv_v_x_u64m1(((uint64_t)!_value)-1, 1));
+    }
+
+    simdjson_inline vbitmask_t to_bitmask() const {
+#if __riscv_v_fixed_vlen == 128
+      return __riscv_vmv_x(__riscv_vreinterpret_u16m1(value));
+#elif __riscv_v_fixed_vlen == 256
+      return __riscv_vmv_x(__riscv_vreinterpret_u32m1(value));
+#else
+      return __riscv_vmv_x(__riscv_vreinterpret_u64m1(value));
+#endif
+    }
+
+    // Bit operations
+    simdjson_inline simd8<bool> operator|(const simd8<bool> other) const {  return __riscv_vmor(*this, other, VL8); }
+    simdjson_inline simd8<bool> operator&(const simd8<bool> other) const {  return __riscv_vmand(*this, other, VL8); }
+    simdjson_inline simd8<bool> operator^(const simd8<bool> other) const {  return __riscv_vmxor(*this, other, VL8); }
+    simdjson_inline simd8<bool> bit_andnot(const simd8<bool> other) const { return __riscv_vmandn(other, *this, VL8); }
+    simdjson_inline simd8<bool> operator~() const { return __riscv_vmnot(*this, VL8); }
+    simdjson_inline simd8<bool>& operator|=(const simd8<bool> other) { auto this_cast = static_cast<simd8<bool>*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline simd8<bool>& operator&=(const simd8<bool> other) { auto this_cast = static_cast<simd8<bool>*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline simd8<bool>& operator^=(const simd8<bool> other) { auto this_cast = static_cast<simd8<bool>*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+  };
+
+  // Unsigned bytes
+  template<>
+  struct simd8<uint8_t> {
+
+    vuint8_t value;
+    static constexpr int SIZE = sizeof(value);
+
+    simdjson_inline simd8(const vuint8_t _value) : value(_value) {}
+    simdjson_inline simd8() : simd8(zero()) {}
+    simdjson_inline simd8(const uint8_t values[VL8]) : simd8(load(values)) {}
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    simdjson_inline simd8(simd8<bool> mask) : value(__riscv_vmerge_vxm_u8m1(zero(), -1, (vbool_t)mask, VL8)) {}
+
+    simdjson_inline operator const vuint8_t&() const { return this->value; }
+    simdjson_inline operator vuint8_t&() { return this->value; }
+
+    simdjson_inline simd8(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) : simd8(vuint8_t{
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    }) {}
+
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<uint8_t> repeat_16(
+      uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
+      uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
+    ) {
+      return simd8<uint8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    static simdjson_inline vuint8_t splat(uint8_t _value) { return __riscv_vmv_v_x_u8m1(_value, VL8); }
+    static simdjson_inline vuint8_t zero() { return splat(0); }
+    static simdjson_inline vuint8_t load(const uint8_t values[VL8]) { return __riscv_vle8_v_u8m1(values, VL8); }
+
+    // Bit operations
+    simdjson_inline simd8<uint8_t> operator|(const simd8<uint8_t> other) const { return  __riscv_vor_vv_u8m1(  value, other, VL8); }
+    simdjson_inline simd8<uint8_t> operator&(const simd8<uint8_t> other) const { return  __riscv_vand_vv_u8m1( value, other, VL8); }
+    simdjson_inline simd8<uint8_t> operator^(const simd8<uint8_t> other) const { return  __riscv_vxor_vv_u8m1( value, other, VL8); }
+    simdjson_inline simd8<uint8_t> operator~() const { return __riscv_vnot_v_u8m1(value, VL8); }
+#if __riscv_zvbb
+    simdjson_inline simd8<uint8_t> bit_andnot(const simd8<uint8_t> other) const { return __riscv_vandn_vv_u8m1(other, value, VL8); }
+#else
+    simdjson_inline simd8<uint8_t> bit_andnot(const simd8<uint8_t> other) const { return other & ~*this; }
+#endif
+    simdjson_inline simd8<uint8_t>& operator|=(const simd8<uint8_t> other) { value = *this | other; return *this; }
+    simdjson_inline simd8<uint8_t>& operator&=(const simd8<uint8_t> other) { value = *this & other; return *this; }
+    simdjson_inline simd8<uint8_t>& operator^=(const simd8<uint8_t> other) { value = *this ^ other; return *this; }
+
+    simdjson_inline simd8<bool> operator==(const simd8<uint8_t> other) const { return __riscv_vmseq(value, other, VL8); }
+    simdjson_inline simd8<bool> operator==(uint8_t other) const { return __riscv_vmseq(value, other, VL8); }
+
+    template<int N=1>
+    simdjson_inline simd8<uint8_t> prev(const simd8<uint8_t> prev_chunk) const {
+      return __riscv_vslideup(__riscv_vslidedown(prev_chunk, VL8-N, VL8), value, N, VL8);
+    }
+
+    // Store to array
+    simdjson_inline void store(uint8_t dst[VL8]) const { return __riscv_vse8(dst, value, VL8); }
+
+    // Saturated math
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return __riscv_vsaddu(value, other, VL8); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return __riscv_vssubu(value, other, VL8); }
+
+    // Addition/subtraction are the same for signed and unsigned
+    simdjson_inline simd8<uint8_t> operator+(const simd8<uint8_t> other) const { return __riscv_vadd(value, other, VL8); }
+    simdjson_inline simd8<uint8_t> operator-(const simd8<uint8_t> other) const { return __riscv_vsub(value, other, VL8); }
+    simdjson_inline simd8<uint8_t>& operator+=(const simd8<uint8_t> other) { value = *this + other; return *this; }
+    simdjson_inline simd8<uint8_t>& operator-=(const simd8<uint8_t> other) { value = *this - other; return *this; }
+
+    // Order-specific operations
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return __riscv_vmsleu(value, other, VL8); }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return __riscv_vmsgeu(value, other, VL8); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const  { return __riscv_vmsltu(value, other, VL8); }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const  { return __riscv_vmsgtu(value, other, VL8); }
+
+    // Same as >, but instead of guaranteeing all 1's == true, false = 0 and true = nonzero.
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return simd8<uint8_t>(*this > other); }
+    // Same as <, but instead of guaranteeing all 1's == true, false = 0 and true = nonzero.
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return simd8<uint8_t>(*this < other); }
+
+    // Bit-specific operations
+    simdjson_inline bool any_bits_set_anywhere() const {
+      return __riscv_vfirst(__riscv_vmsne(value, 0, VL8), VL8) >= 0;
+    }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return (*this & bits).any_bits_set_anywhere(); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shr() const { return __riscv_vsrl(value, N, VL8); }
+    template<int N>
+    simdjson_inline simd8<uint8_t> shl() const { return __riscv_vsll(value, N, VL8); }
+
+
+    // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return __riscv_vrgather(lookup_table, value, VL8);
+    }
+
+    // compress inactive elements, to match AVX-512 behavior
+    template<typename L>
+    simdjson_inline void compress(vbitmask_t mask, L * output) const {
+      mask = (vbitmask_t)~mask;
+#if __riscv_v_fixed_vlen == 128
+      vbool8_t m = __riscv_vreinterpret_b8(__riscv_vmv_s_x_u16m1(mask, 1));
+#elif __riscv_v_fixed_vlen == 256
+      vbool8_t m = __riscv_vreinterpret_b8(__riscv_vmv_s_x_u32m1(mask, 1));
+#else
+      vbool8_t m = __riscv_vreinterpret_b8(__riscv_vmv_s_x_u64m1(mask, 1));
+#endif
+      __riscv_vse8_v_u8m1(output, __riscv_vcompress(value, m, VL8), count_ones(mask));
+    }
+
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  // Signed bytes
+  template<>
+  struct simd8<int8_t> {
+    vint8_t value;
+    static constexpr int SIZE = sizeof(value);
+
+    simdjson_inline simd8(const vint8_t _value) : value(_value) {}
+    simdjson_inline simd8() : simd8(zero()) {}
+    simdjson_inline simd8(const int8_t values[VL8]) : simd8(load(values)) {}
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+
+    simdjson_inline operator const vint8_t&() const { return this->value; }
+    simdjson_inline operator vint8_t&() { return this->value; }
+
+    simdjson_inline simd8(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) : simd8(vint8_t{
+      v0, v1, v2, v3, v4, v5, v6, v7,
+      v8, v9, v10,v11,v12,v13,v14,v15
+    }) {}
+
+    // Repeat 16 values as many times as necessary (usually for lookup tables)
+    simdjson_inline static simd8<int8_t> repeat_16(
+      int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
+      int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
+    ) {
+      return simd8<int8_t>(
+        v0, v1, v2, v3, v4, v5, v6, v7,
+        v8, v9, v10,v11,v12,v13,v14,v15
+      );
+    }
+
+    static simdjson_inline vint8_t splat(int8_t _value) { return __riscv_vmv_v_x_i8m1(_value, VL8); }
+    static simdjson_inline vint8_t zero() { return splat(0); }
+    static simdjson_inline vint8_t load(const int8_t values[VL8]) { return __riscv_vle8_v_i8m1(values, VL8); }
+
+
+    simdjson_inline void store(int8_t dst[VL8]) const { return __riscv_vse8(dst, value, VL8); }
+
+    // Explicit conversion to/from unsigned
+    simdjson_inline explicit simd8(const vuint8_t other): simd8(__riscv_vreinterpret_i8m1(other)) {}
+    simdjson_inline explicit operator simd8<uint8_t>() const { return __riscv_vreinterpret_u8m1(value); }
+
+    // Math
+    simdjson_inline simd8<int8_t> operator+(const simd8<int8_t> other) const { return __riscv_vadd(value, other, VL8); }
+    simdjson_inline simd8<int8_t> operator-(const simd8<int8_t> other) const { return __riscv_vsub(value, other, VL8); }
+    simdjson_inline simd8<int8_t>& operator+=(const simd8<int8_t> other) { value = *this + other; return *this; }
+    simdjson_inline simd8<int8_t>& operator-=(const simd8<int8_t> other) { value = *this - other; return *this; }
+
+    // Order-sensitive comparisons
+    simdjson_inline simd8<int8_t> max_val( const simd8<int8_t> other) const { return __riscv_vmax( value, other, VL8); }
+    simdjson_inline simd8<int8_t> min_val( const simd8<int8_t> other) const { return __riscv_vmin( value, other, VL8); }
+    simdjson_inline simd8<bool> operator>( const simd8<int8_t> other) const { return __riscv_vmsgt(value, other, VL8); }
+    simdjson_inline simd8<bool> operator<( const simd8<int8_t> other) const { return __riscv_vmslt(value, other, VL8); }
+    simdjson_inline simd8<bool> operator==(const simd8<int8_t> other) const { return __riscv_vmseq(value, other, VL8); }
+
+    template<int N=1>
+    simdjson_inline simd8<int8_t> prev(const simd8<int8_t> prev_chunk) const {
+      return __riscv_vslideup(__riscv_vslidedown(prev_chunk, VL8-N, VL8), value, N, VL8);
+    }
+
+    // Perform a lookup assuming no value is larger than 16
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+      return __riscv_vrgather(lookup_table, value, VL8);
+    }
+    template<typename L>
+    simdjson_inline simd8<L> lookup_16(
+        L replace0,  L replace1,  L replace2,  L replace3,
+        L replace4,  L replace5,  L replace6,  L replace7,
+        L replace8,  L replace9,  L replace10, L replace11,
+        L replace12, L replace13, L replace14, L replace15) const {
+      return lookup_16(simd8<L>::repeat_16(
+        replace0,  replace1,  replace2,  replace3,
+        replace4,  replace5,  replace6,  replace7,
+        replace8,  replace9,  replace10, replace11,
+        replace12, replace13, replace14, replace15
+      ));
+    }
+  };
+
+  template<typename T>
+  struct simd8x64;
+  template<>
+  struct simd8x64<uint8_t> {
+    static constexpr int NUM_CHUNKS = 64 / sizeof(simd8<uint8_t>);
+    vuint8x64_t value;
+
+#if __riscv_v_fixed_vlen >= 512
+    template<int idx> simd8<uint8_t> get() const { return value; }
+#else
+    template<int idx> simd8<uint8_t> get() const { return __riscv_vget_u8m1(value, idx); }
+#endif
+
+    simdjson_inline operator const vuint8x64_t&() const { return this->value; }
+    simdjson_inline operator vuint8x64_t&() { return this->value; }
+
+    simd8x64(const simd8x64<uint8_t>& o) = delete; // no copy allowed
+    simd8x64<uint8_t>& operator=(const simd8<uint8_t>& other) = delete; // no assignment allowed
+    simd8x64() = delete; // no default constructor allowed
+
+#if __riscv_v_fixed_vlen == 128
+    simdjson_inline simd8x64(const uint8_t *ptr, size_t n = 64) : value(__riscv_vle8_v_u8m4(ptr, n)) {}
+#elif __riscv_v_fixed_vlen == 256
+    simdjson_inline simd8x64(const uint8_t *ptr, size_t n = 64) : value(__riscv_vle8_v_u8m2(ptr, n)) {}
+#else
+    simdjson_inline simd8x64(const uint8_t *ptr, size_t n = 64) : value(__riscv_vle8_v_u8m1(ptr, n)) {}
+#endif
+
+    simdjson_inline void store(uint8_t ptr[64]) const {
+      __riscv_vse8(ptr, value, 64);
+    }
+
+    simdjson_inline bool is_ascii() const {
+#if __riscv_v_fixed_vlen == 128
+      return __riscv_vfirst(__riscv_vmslt(__riscv_vreinterpret_i8m4(value), 0, 64), 64) < 0;
+#elif __riscv_v_fixed_vlen == 256
+      return __riscv_vfirst(__riscv_vmslt(__riscv_vreinterpret_i8m2(value), 0, 64), 64) < 0;
+#else
+      return __riscv_vfirst(__riscv_vmslt(__riscv_vreinterpret_i8m1(value), 0, 64), 64) < 0;
+#endif
+    }
+
+    // compress inactive elements, to match AVX-512 behavior
+    simdjson_inline uint64_t compress(uint64_t mask, uint8_t * output) const {
+      mask = ~mask;
+#if __riscv_v_fixed_vlen == 128
+      vboolx64_t m = __riscv_vreinterpret_b2(__riscv_vmv_s_x_u64m1(mask, 1));
+#elif __riscv_v_fixed_vlen == 256
+      vboolx64_t m = __riscv_vreinterpret_b4(__riscv_vmv_s_x_u64m1(mask, 1));
+#else
+      vboolx64_t m = __riscv_vreinterpret_b8(__riscv_vmv_s_x_u64m1(mask, 1));
+#endif
+      size_t cnt = count_ones(mask);
+      __riscv_vse8(output, __riscv_vcompress(value, m, 64), cnt);
+      return cnt;
+    }
+
+    simdjson_inline uint64_t eq(const uint8_t m) const {
+      return __riscv_vmv_x(__riscv_vreinterpret_u64m1(__riscv_vmseq(value, m, 64)));
+    }
+
+    simdjson_inline uint64_t lteq(const uint8_t m) const {
+      return __riscv_vmv_x(__riscv_vreinterpret_u64m1(__riscv_vmsleu(value, m, 64)));
+    }
+  }; // struct simd8x64<uint8_t>
+
+} // namespace simd
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_RVV_VLS_SIMD_H
+/* end file simdjson/rvv-vls/simd.h */
+/* including generic/amalgamated.h for rvv_vls: #include <generic/amalgamated.h> */
+/* begin file generic/amalgamated.h for rvv_vls */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_SRC_GENERIC_DEPENDENCIES_H)
+#error generic/dependencies.h must be included before generic/amalgamated.h!
+#endif
+
+/* including generic/base.h for rvv_vls: #include <generic/base.h> */
+/* begin file generic/base.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+
+struct json_character_block;
+
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_BASE_H
+/* end file generic/base.h for rvv_vls */
+/* including generic/dom_parser_implementation.h for rvv_vls: #include <generic/dom_parser_implementation.h> */
+/* begin file generic/dom_parser_implementation.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// Interface a dom parser implementation must fulfill
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+
+simdjson_inline simd8<uint8_t> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3);
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input);
+
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file generic/dom_parser_implementation.h for rvv_vls */
+/* including generic/json_character_block.h for rvv_vls: #include <generic/json_character_block.h> */
+/* begin file generic/json_character_block.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+
+struct json_character_block {
+  static simdjson_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
+
+  simdjson_inline uint64_t whitespace() const noexcept { return _whitespace; }
+  simdjson_inline uint64_t op() const noexcept { return _op; }
+  simdjson_inline uint64_t scalar() const noexcept { return ~(op() | whitespace()); }
+
+  uint64_t _whitespace;
+  uint64_t _op;
+};
+
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_JSON_CHARACTER_BLOCK_H
+/* end file generic/json_character_block.h for rvv_vls */
+/* end file generic/amalgamated.h for rvv_vls */
+/* including generic/stage1/amalgamated.h for rvv_vls: #include <generic/stage1/amalgamated.h> */
+/* begin file generic/stage1/amalgamated.h for rvv_vls */
+// Stuff other things depend on
+/* including generic/stage1/base.h for rvv_vls: #include <generic/stage1/base.h> */
+/* begin file generic/stage1/base.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_BASE_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace stage1 {
+
+class bit_indexer;
+template<size_t STEP_SIZE>
+struct buf_block_reader;
+struct json_block;
+class json_minifier;
+class json_scanner;
+struct json_string_block;
+class json_string_scanner;
+class json_structural_indexer;
+
+} // namespace stage1
+
+namespace utf8_validation {
+struct utf8_checker;
+} // namespace utf8_validation
+
+using utf8_validation::utf8_checker;
+
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_BASE_H
+/* end file generic/stage1/base.h for rvv_vls */
+/* including generic/stage1/buf_block_reader.h for rvv_vls: #include <generic/stage1/buf_block_reader.h> */
+/* begin file generic/stage1/buf_block_reader.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace stage1 {
+
+// Walks through a buffer in block-sized increments, loading the last part with spaces
+template<size_t STEP_SIZE>
+struct buf_block_reader {
+public:
+  simdjson_inline buf_block_reader(const uint8_t *_buf, size_t _len);
+  simdjson_inline size_t block_index();
+  simdjson_inline bool has_full_block() const;
+  simdjson_inline const uint8_t *full_block() const;
+  /**
+   * Get the last block, padded with spaces.
+   *
+   * There will always be a last block, with at least 1 byte, unless len == 0 (in which case this
+   * function fills the buffer with spaces and returns 0. In particular, if len == STEP_SIZE there
+   * will be 0 full_blocks and 1 remainder block with STEP_SIZE bytes and no spaces for padding.
+   *
+   * @return the number of effective characters in the last block.
+   */
+  simdjson_inline size_t get_remainder(uint8_t *dst) const;
+  simdjson_inline void advance();
+private:
+  const uint8_t *buf;
+  const size_t len;
+  const size_t lenminusstep;
+  size_t idx;
+};
+
+// Routines to print masks and text for debugging bitmask operations
+simdjson_unused static char * format_input_text_64(const uint8_t *text) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    buf[i] = int8_t(text[i]) < ' ' ? '_' : int8_t(text[i]);
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+// Routines to print masks and text for debugging bitmask operations
+simdjson_unused static char * format_input_text(const simd8x64<uint8_t>& in) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  in.store(reinterpret_cast<uint8_t*>(buf));
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] < ' ') { buf[i] = '_'; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+simdjson_unused static char * format_input_text(const simd8x64<uint8_t>& in, uint64_t mask) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  in.store(reinterpret_cast<uint8_t*>(buf));
+  for (size_t i=0; i<sizeof(simd8x64<uint8_t>); i++) {
+    if (buf[i] <= ' ') { buf[i] = '_'; }
+    if (!(mask & (size_t(1) << i))) { buf[i] = ' '; }
+  }
+  buf[sizeof(simd8x64<uint8_t>)] = '\0';
+  return buf;
+}
+
+simdjson_unused static char * format_mask(uint64_t mask) {
+  static char buf[sizeof(simd8x64<uint8_t>) + 1];
+  for (size_t i=0; i<64; i++) {
+    buf[i] = (mask & (size_t(1) << i)) ? 'X' : ' ';
+  }
+  buf[64] = '\0';
+  return buf;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
+
+template<size_t STEP_SIZE>
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
+
+template<size_t STEP_SIZE>
+simdjson_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
+  return idx < lenminusstep;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
+  return &buf[idx];
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
+  if(len == idx) { return 0; } // memcpy(dst, null, 0) will trigger an error with some sanitizers
+  std::memset(dst, 0x20, STEP_SIZE); // std::memset STEP_SIZE because it's more efficient to write out 8 or 16 bytes at once.
+  std::memcpy(dst, buf + idx, len - idx);
+  return len - idx;
+}
+
+template<size_t STEP_SIZE>
+simdjson_inline void buf_block_reader<STEP_SIZE>::advance() {
+  idx += STEP_SIZE;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_BUF_BLOCK_READER_H
+/* end file generic/stage1/buf_block_reader.h for rvv_vls */
+/* including generic/stage1/json_escape_scanner.h for rvv_vls: #include <generic/stage1/json_escape_scanner.h> */
+/* begin file generic/stage1/json_escape_scanner.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_ESCAPE_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_ESCAPE_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace stage1 {
+
+/**
+ * Scans for escape characters in JSON, taking care with multiple backslashes (\\n vs. \n).
+ */
+struct json_escape_scanner {
+  /** The actual escape characters (the backslashes themselves). */
+  uint64_t next_is_escaped = 0ULL;
+
+  struct escaped_and_escape {
+    /**
+     * Mask of escaped characters.
+     *
+     * ```
+     * \n \\n \\\n \\\\n \
+     * 0100100010100101000
+     *  n  \   \ n  \ \
+     * ```
+     */
+    uint64_t escaped;
+    /**
+     * Mask of escape characters.
+     *
+     * ```
+     * \n \\n \\\n \\\\n \
+     * 1001000101001010001
+     * \  \   \ \  \ \   \
+     * ```
+     */
+    uint64_t escape;
+  };
+
+  /**
+   * Get a mask of both escape and escaped characters (the characters following a backslash).
+   *
+   * @param potential_escape A mask of the character that can escape others (but could be
+   *        escaped itself). e.g. block.eq('\\')
+   */
+  simdjson_really_inline escaped_and_escape next(uint64_t backslash) noexcept {
+
+#if !SIMDJSON_SKIP_BACKSLASH_SHORT_CIRCUIT
+    if (!backslash) { return {next_escaped_without_backslashes(), 0}; }
+#endif
+
+    // |                                | Mask (shows characters instead of 1's) | Depth | Instructions        |
+    // |--------------------------------|----------------------------------------|-------|---------------------|
+    // | string                         | `\\n_\\\n___\\\n___\\\\___\\\\__\\\`   |       |                     |
+    // |                                | `    even   odd    even   odd   odd`   |       |                     |
+    // | potential_escape               | ` \  \\\    \\\    \\\\   \\\\  \\\`   | 1     | 1 (backslash & ~first_is_escaped)
+    // | escape_and_terminal_code       | ` \n \ \n   \ \n   \ \    \ \   \ \`   | 5     | 5 (next_escape_and_terminal_code())
+    // | escaped                        | `\    \ n    \ n    \ \    \ \   \ ` X | 6     | 7 (escape_and_terminal_code ^ (potential_escape | first_is_escaped))
+    // | escape                         | `    \ \    \ \    \ \    \ \   \ \`   | 6     | 8 (escape_and_terminal_code & backslash)
+    // | first_is_escaped               | `\                                 `   | 7 (*) | 9 (escape >> 63) ()
+    //                                                                               (*) this is not needed until the next iteration
+    uint64_t escape_and_terminal_code = next_escape_and_terminal_code(backslash & ~this->next_is_escaped);
+    uint64_t escaped = escape_and_terminal_code ^ (backslash | this->next_is_escaped);
+    uint64_t escape = escape_and_terminal_code & backslash;
+    this->next_is_escaped = escape >> 63;
+    return {escaped, escape};
+  }
+
+private:
+  static constexpr const uint64_t ODD_BITS = 0xAAAAAAAAAAAAAAAAULL;
+
+  simdjson_really_inline uint64_t next_escaped_without_backslashes() noexcept {
+    uint64_t escaped = this->next_is_escaped;
+    this->next_is_escaped = 0;
+    return escaped;
+  }
+
+  /**
+   * Returns a mask of the next escape characters (masking out escaped backslashes), along with
+   * any non-backslash escape codes.
+   *
+   * \n \\n \\\n \\\\n returns:
+   * \n \   \ \n \ \
+   * 11 100 1011 10100
+   *
+   * You are expected to mask out the first bit yourself if the previous block had a trailing
+   * escape.
+   *
+   * & the result with potential_escape to get just the escape characters.
+   * ^ the result with (potential_escape | first_is_escaped) to get escaped characters.
+   */
+  static simdjson_really_inline uint64_t next_escape_and_terminal_code(uint64_t potential_escape) noexcept {
+    // If we were to just shift and mask out any odd bits, we'd actually get a *half* right answer:
+    // any even-aligned backslash runs would be correct! Odd-aligned backslash runs would be
+    // inverted (\\\ would be 010 instead of 101).
+    //
+    // ```
+    // string:              | ____\\\\_\\\\_____ |
+    // maybe_escaped | ODD  |     \ \   \ \      |
+    //               even-aligned ^^^  ^^^^ odd-aligned
+    // ```
+    //
+    // Taking that into account, our basic strategy is:
+    //
+    // 1. Use subtraction to produce a mask with 1's for even-aligned runs and 0's for
+    //    odd-aligned runs.
+    // 2. XOR all odd bits, which masks out the odd bits in even-aligned runs, and brings IN the
+    //    odd bits in odd-aligned runs.
+    // 3. & with backslash to clean up any stray bits.
+    // runs are set to 0, and then XORing with "odd":
+    //
+    // |                                | Mask (shows characters instead of 1's) | Instructions        |
+    // |--------------------------------|----------------------------------------|---------------------|
+    // | string                         | `\\n_\\\n___\\\n___\\\\___\\\\__\\\`   |
+    // |                                | `    even   odd    even   odd   odd`   |
+    // | maybe_escaped                  | `  n  \\n    \\n    \\\_   \\\_  \\` X | 1 (potential_escape << 1)
+    // | maybe_escaped_and_odd          | ` \n_ \\n _ \\\n_ _ \\\__ _\\\_ \\\`   | 1 (maybe_escaped | odd)
+    // | even_series_codes_and_odd      | `  n_\\\  _    n_ _\\\\ _     _    `   | 1 (maybe_escaped_and_odd - potential_escape)
+    // | escape_and_terminal_code       | ` \n \ \n   \ \n   \ \    \ \   \ \`   | 1 (^ odd)
+    //
+
+    // Escaped characters are characters following an escape.
+    uint64_t maybe_escaped = potential_escape << 1;
+
+    // To distinguish odd from even escape sequences, therefore, we turn on any *starting*
+    // escapes that are on an odd byte. (We actually bring in all odd bits, for speed.)
+    // - Odd runs of backslashes are 0000, and the code at the end ("n" in \n or \\n) is 1.
+    // - Odd runs of backslashes are 1111, and the code at the end ("n" in \n or \\n) is 0.
+    // - All other odd bytes are 1, and even bytes are 0.
+    uint64_t maybe_escaped_and_odd_bits     = maybe_escaped | ODD_BITS;
+    uint64_t even_series_codes_and_odd_bits = maybe_escaped_and_odd_bits - potential_escape;
+
+    // Now we flip all odd bytes back with xor. This:
+    // - Makes odd runs of backslashes go from 0000 to 1010
+    // - Makes even runs of backslashes go from 1111 to 1010
+    // - Sets actually-escaped codes to 1 (the n in \n and \\n: \n = 11, \\n = 100)
+    // - Resets all other bytes to 0
+    return even_series_codes_and_odd_bits ^ ODD_BITS;
+  }
+};
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+/* end file generic/stage1/json_escape_scanner.h for rvv_vls */
+/* including generic/stage1/json_string_scanner.h for rvv_vls: #include <generic/stage1/json_string_scanner.h> */
+/* begin file generic/stage1/json_string_scanner.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_escape_scanner.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace stage1 {
+
+struct json_string_block {
+  // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
+  simdjson_really_inline json_string_block(uint64_t escaped, uint64_t quote, uint64_t in_string) :
+  _escaped(escaped), _quote(quote), _in_string(in_string) {}
+
+  // Escaped characters (characters following an escape() character)
+  simdjson_really_inline uint64_t escaped() const { return _escaped; }
+  // Real (non-backslashed) quotes
+  simdjson_really_inline uint64_t quote() const { return _quote; }
+  // Only characters inside the string (not including the quotes)
+  simdjson_really_inline uint64_t string_content() const { return _in_string & ~_quote; }
+  // Return a mask of whether the given characters are inside a string (only works on non-quotes)
+  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
+  // Return a mask of whether the given characters are inside a string (only works on non-quotes)
+  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
+  // Tail of string (everything except the start quote)
+  simdjson_really_inline uint64_t string_tail() const { return _in_string ^ _quote; }
+
+  // escaped characters (backslashed--does not include the hex characters after \u)
+  uint64_t _escaped;
+  // real quotes (non-escaped ones)
+  uint64_t _quote;
+  // string characters (includes start quote but not end quote)
+  uint64_t _in_string;
+};
+
+// Scans blocks for string characters, storing the state necessary to do so
+class json_string_scanner {
+public:
+  simdjson_really_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
+  // Returns either UNCLOSED_STRING or SUCCESS
+  simdjson_really_inline error_code finish();
+
+private:
+  // Scans for escape characters
+  json_escape_scanner escape_scanner{};
+  // Whether the last iteration was still inside a string (all 1's = true, all 0's = false).
+  uint64_t prev_in_string = 0ULL;
+};
+
+//
+// Return a mask of all string characters plus end quotes.
+//
+// prev_escaped is overflow saying whether the next character is escaped.
+// prev_in_string is overflow saying whether we're still in a string.
+//
+// Backslash sequences outside of quotes will be detected in stage 2.
+//
+simdjson_really_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
+  const uint64_t backslash = in.eq('\\');
+  const uint64_t escaped = escape_scanner.next(backslash).escaped;
+  const uint64_t quote = in.eq('"') & ~escaped;
+
+  //
+  // prefix_xor flips on bits inside the string (and flips off the end quote).
+  //
+  // Then we xor with prev_in_string: if we were in a string already, its effect is flipped
+  // (characters inside strings are outside, and characters outside strings are inside).
+  //
+  const uint64_t in_string = prefix_xor(quote) ^ prev_in_string;
+
+  //
+  // Check if we're still in a string at the end of the box so the next block will know
+  //
+  prev_in_string = uint64_t(static_cast<int64_t>(in_string) >> 63);
+
+  // Use ^ to turn the beginning quote off, and the end quote on.
+
+  // We are returning a function-local object so either we get a move constructor
+  // or we get copy elision.
+  return json_string_block(escaped, quote, in_string);
+}
+
+simdjson_really_inline error_code json_string_scanner::finish() {
+  if (prev_in_string) {
+    return UNCLOSED_STRING;
+  }
+  return SUCCESS;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRING_SCANNER_H
+/* end file generic/stage1/json_string_scanner.h for rvv_vls */
+/* including generic/stage1/utf8_lookup4_algorithm.h for rvv_vls: #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* begin file generic/stage1/utf8_lookup4_algorithm.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace utf8_validation {
+
+using namespace simd;
+
+  simdjson_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+// Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
+// Bit 1 = Too Long (ASCII followed by continuation)
+// Bit 2 = Overlong 3-byte
+// Bit 4 = Surrogate
+// Bit 5 = Overlong 2-byte
+// Bit 7 = Two Continuations
+    constexpr const uint8_t TOO_SHORT   = 1<<0; // 11______ 0_______
+                                                // 11______ 11______
+    constexpr const uint8_t TOO_LONG    = 1<<1; // 0_______ 10______
+    constexpr const uint8_t OVERLONG_3  = 1<<2; // 11100000 100_____
+    constexpr const uint8_t SURROGATE   = 1<<4; // 11101101 101_____
+    constexpr const uint8_t OVERLONG_2  = 1<<5; // 1100000_ 10______
+    constexpr const uint8_t TWO_CONTS   = 1<<7; // 10______ 10______
+    constexpr const uint8_t TOO_LARGE   = 1<<3; // 11110100 1001____
+                                                // 11110100 101_____
+                                                // 11110101 1001____
+                                                // 11110101 101_____
+                                                // 1111011_ 1001____
+                                                // 1111011_ 101_____
+                                                // 11111___ 1001____
+                                                // 11111___ 101_____
+    constexpr const uint8_t TOO_LARGE_1000 = 1<<6;
+                                                // 11110101 1000____
+                                                // 1111011_ 1000____
+                                                // 11111___ 1000____
+    constexpr const uint8_t OVERLONG_4  = 1<<6; // 11110000 1000____
+
+    const simd8<uint8_t> byte_1_high = prev1.shr<4>().lookup_16<uint8_t>(
+      // 0_______ ________ <ASCII in byte 1>
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      TOO_LONG, TOO_LONG, TOO_LONG, TOO_LONG,
+      // 10______ ________ <continuation in byte 1>
+      TWO_CONTS, TWO_CONTS, TWO_CONTS, TWO_CONTS,
+      // 1100____ ________ <two byte lead in byte 1>
+      TOO_SHORT | OVERLONG_2,
+      // 1101____ ________ <two byte lead in byte 1>
+      TOO_SHORT,
+      // 1110____ ________ <three byte lead in byte 1>
+      TOO_SHORT | OVERLONG_3 | SURROGATE,
+      // 1111____ ________ <four+ byte lead in byte 1>
+      TOO_SHORT | TOO_LARGE | TOO_LARGE_1000 | OVERLONG_4
+    );
+    constexpr const uint8_t CARRY = TOO_SHORT | TOO_LONG | TWO_CONTS; // These all have ____ in byte 1 .
+    const simd8<uint8_t> byte_1_low = (prev1 & 0x0F).lookup_16<uint8_t>(
+      // ____0000 ________
+      CARRY | OVERLONG_3 | OVERLONG_2 | OVERLONG_4,
+      // ____0001 ________
+      CARRY | OVERLONG_2,
+      // ____001_ ________
+      CARRY,
+      CARRY,
+
+      // ____0100 ________
+      CARRY | TOO_LARGE,
+      // ____0101 ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      // ____011_ ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+
+      // ____1___ ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      // ____1101 ________
+      CARRY | TOO_LARGE | TOO_LARGE_1000 | SURROGATE,
+      CARRY | TOO_LARGE | TOO_LARGE_1000,
+      CARRY | TOO_LARGE | TOO_LARGE_1000
+    );
+    const simd8<uint8_t> byte_2_high = input.shr<4>().lookup_16<uint8_t>(
+      // ________ 0_______ <ASCII in byte 2>
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT,
+
+      // ________ 1000____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE_1000 | OVERLONG_4,
+      // ________ 1001____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | OVERLONG_3 | TOO_LARGE,
+      // ________ 101_____
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE  | TOO_LARGE,
+      TOO_LONG | OVERLONG_2 | TWO_CONTS | SURROGATE  | TOO_LARGE,
+
+      // ________ 11______
+      TOO_SHORT, TOO_SHORT, TOO_SHORT, TOO_SHORT
+    );
+    return (byte_1_high & byte_1_low & byte_2_high);
+  }
+  simdjson_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
+      const simd8<uint8_t> prev_input, const simd8<uint8_t> sc) {
+    simd8<uint8_t> prev2 = input.prev<2>(prev_input);
+    simd8<uint8_t> prev3 = input.prev<3>(prev_input);
+    simd8<uint8_t> must23 = must_be_2_3_continuation(prev2, prev3);
+    simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
+    return must23_80 ^ sc;
+  }
+
+  //
+  // Return nonzero if there are incomplete multibyte characters at the end of the block:
+  // e.g. if there is a 4-byte character, but it's 3 bytes from the end.
+  //
+  simdjson_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+    // If the previous input's last 3 bytes match this, they're too short (they ended at EOF):
+    // ... 1111____ 111_____ 11______
+#if SIMDJSON_IMPLEMENTATION_ICELAKE || (SIMDJSON_IMPLEMENTATION_RVV_VLS && __riscv_v_fixed_vlen >= 512)
+    static const uint8_t max_array[64] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
+    };
+#else
+    static const uint8_t max_array[32] = {
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 255, 255, 255,
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
+    };
+#endif
+    const simd8<uint8_t> max_value(&max_array[sizeof(max_array)-sizeof(simd8<uint8_t>)]);
+    return input.gt_bits(max_value);
+  }
+
+  struct utf8_checker {
+    // If this is nonzero, there has been a UTF-8 error.
+    simd8<uint8_t> error;
+    // The last input we received
+    simd8<uint8_t> prev_input_block;
+    // Whether the last input we received was incomplete (used for ASCII fast path)
+    simd8<uint8_t> prev_incomplete;
+
+    //
+    // Check whether the current bytes are valid UTF-8.
+    //
+    simdjson_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
+      // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ lead bytes
+      // (2, 3, 4-byte leads become large positive numbers instead of small negative numbers)
+      simd8<uint8_t> prev1 = input.prev<1>(prev_input);
+      simd8<uint8_t> sc = check_special_cases(input, prev1);
+      this->error |= check_multibyte_lengths(input, prev_input, sc);
+    }
+
+    // The only problem that can happen at EOF is that a multibyte character is too short
+    // or a byte value too large in the last bytes: check_special_cases only checks for bytes
+    // too large in the first of two bytes.
+    simdjson_inline void check_eof() {
+      // If the previous block had incomplete UTF-8 characters at the end, an ASCII block can't
+      // possibly finish them.
+      this->error |= this->prev_incomplete;
+    }
+
+    simdjson_inline void check_next_input(const simd8x64<uint8_t>& input) {
+      if(simdjson_likely(is_ascii(input))) {
+        this->error |= this->prev_incomplete;
+      } else {
+        // you might think that a for-loop would work, but under Visual Studio, it is not good enough.
+        static_assert((simd8x64<uint8_t>::NUM_CHUNKS == 1)
+                ||(simd8x64<uint8_t>::NUM_CHUNKS == 2)
+                || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
+                "We support one, two or four chunks per 64-byte block.");
+        SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 1) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+          this->check_utf8_bytes(input.get<1>(), input.get<0>());
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+          this->check_utf8_bytes(input.get<0>(), this->prev_input_block);
+          this->check_utf8_bytes(input.get<1>(), input.get<0>());
+          this->check_utf8_bytes(input.get<2>(), input.get<1>());
+          this->check_utf8_bytes(input.get<3>(), input.get<2>());
+        }
+        this->prev_incomplete = is_incomplete(input.get<simd8x64<uint8_t>::NUM_CHUNKS-1>());
+        this->prev_input_block = input.get<simd8x64<uint8_t>::NUM_CHUNKS-1>();
+      }
+    }
+    // do not forget to call check_eof!
+    simdjson_warn_unused simdjson_inline error_code errors() {
+      return this->error.any_bits_set_anywhere() ? error_code::UTF8_ERROR : error_code::SUCCESS;
+    }
+
+  }; // struct utf8_checker
+} // namespace utf8_validation
+
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_UTF8_LOOKUP4_ALGORITHM_H
+/* end file generic/stage1/utf8_lookup4_algorithm.h for rvv_vls */
+/* including generic/stage1/json_scanner.h for rvv_vls: #include <generic/stage1/json_scanner.h> */
+/* begin file generic/stage1/json_scanner.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/json_character_block.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_string_scanner.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace stage1 {
+
+/**
+ * A block of scanned json, with information on operators and scalars.
+ *
+ * We seek to identify pseudo-structural characters. Anything that is inside
+ * a string must be omitted (hence  & ~_string.string_tail()).
+ * Otherwise, pseudo-structural characters come in two forms.
+ * 1. We have the structural characters ([,],{,},:, comma). The
+ *    term 'structural character' is from the JSON RFC.
+ * 2. We have the 'scalar pseudo-structural characters'.
+ *    Scalars are quotes, and any character except structural characters and white space.
+ *
+ * To identify the scalar pseudo-structural characters, we must look at what comes
+ * before them: it must be a space, a quote or a structural characters.
+ * Starting with simdjson v0.3, we identify them by
+ * negation: we identify everything that is followed by a non-quote scalar,
+ * and we negate that. Whatever remains must be a 'scalar pseudo-structural character'.
+ */
+struct json_block {
+public:
+  // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
+  simdjson_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  _string(std::move(string)), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
+  simdjson_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  _string(string), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
+
+  /**
+   * The start of structurals.
+   * In simdjson prior to v0.3, these were called the pseudo-structural characters.
+   **/
+  simdjson_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
+  /** All JSON whitespace (i.e. not in a string) */
+  simdjson_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
+
+  // Helpers
+
+  /** Whether the given characters are inside a string (only works on non-quotes) */
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
+  /** Whether the given characters are outside a string (only works on non-quotes) */
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
+
+  // string and escape characters
+  json_string_block _string;
+  // whitespace, structural characters ('operators'), scalars
+  json_character_block _characters;
+  // whether the previous character was a scalar
+  uint64_t _follows_potential_nonquote_scalar;
+private:
+  // Potential structurals (i.e. disregarding strings)
+
+  /**
+   * structural elements ([,],{,},:, comma) plus scalar starts like 123, true and "abc".
+   * They may reside inside a string.
+   **/
+  simdjson_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
+  /**
+   * The start of non-operator runs, like 123, true and "abc".
+   * It main reside inside a string.
+   **/
+  simdjson_inline uint64_t potential_scalar_start() const noexcept {
+    // The term "scalar" refers to anything except structural characters and white space
+    // (so letters, numbers, quotes).
+    // Whenever it is preceded by something that is not a structural element ({,},[,],:, ") nor a white-space
+    // then we know that it is irrelevant structurally.
+    return _characters.scalar() & ~follows_potential_scalar();
+  }
+  /**
+   * Whether the given character is immediately after a non-operator like 123, true.
+   * The characters following a quote are not included.
+   */
+  simdjson_inline uint64_t follows_potential_scalar() const noexcept {
+    // _follows_potential_nonquote_scalar: is defined as marking any character that follows a character
+    // that is not a structural element ({,},[,],:, comma) nor a quote (") and that is not a
+    // white space.
+    // It is understood that within quoted region, anything at all could be marked (irrelevant).
+    return _follows_potential_nonquote_scalar;
+  }
+};
+
+/**
+ * Scans JSON for important bits: structural characters or 'operators', strings, and scalars.
+ *
+ * The scanner starts by calculating two distinct things:
+ * - string characters (taking \" into account)
+ * - structural characters or 'operators' ([]{},:, comma)
+ *   and scalars (runs of non-operators like 123, true and "abc")
+ *
+ * To minimize data dependency (a key component of the scanner's speed), it finds these in parallel:
+ * in particular, the operator/scalar bit will find plenty of things that are actually part of
+ * strings. When we're done, json_block will fuse the two together by masking out tokens that are
+ * part of a string.
+ */
+class json_scanner {
+public:
+  json_scanner() = default;
+  simdjson_inline json_block next(const simd::simd8x64<uint8_t>& in);
+  // Returns either UNCLOSED_STRING or SUCCESS
+  simdjson_warn_unused simdjson_inline error_code finish();
+
+private:
+  // Whether the last character of the previous iteration is part of a scalar token
+  // (anything except whitespace or a structural character/'operator').
+  uint64_t prev_scalar = 0ULL;
+  json_string_scanner string_scanner{};
+};
+
+
+//
+// Check if the current character immediately follows a matching character.
+//
+// For example, this checks for quotes with backslashes in front of them:
+//
+//     const uint64_t backslashed_quote = in.eq('"') & immediately_follows(in.eq('\'), prev_backslash);
+//
+simdjson_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
+  const uint64_t result = match << 1 | overflow;
+  overflow = match >> 63;
+  return result;
+}
+
+simdjson_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
+  json_string_block strings = string_scanner.next(in);
+  // identifies the white-space and the structural characters
+  json_character_block characters = json_character_block::classify(in);
+  // The term "scalar" refers to anything except structural characters and white space
+  // (so letters, numbers, quotes).
+  // We want follows_scalar to mark anything that follows a non-quote scalar (so letters and numbers).
+  //
+  // A terminal quote should either be followed by a structural character (comma, brace, bracket, colon)
+  // or nothing. However, we still want ' "a string"true ' to mark the 't' of 'true' as a potential
+  // pseudo-structural character just like we would if we had  ' "a string" true '; otherwise we
+  // may need to add an extra check when parsing strings.
+  //
+  // Performance: there are many ways to skin this cat.
+  const uint64_t nonquote_scalar = characters.scalar() & ~strings.quote();
+  uint64_t follows_nonquote_scalar = follows(nonquote_scalar, prev_scalar);
+  // We are returning a function-local object so either we get a move constructor
+  // or we get copy elision.
+  return json_block(
+    strings,// strings is a function-local object so either it moves or the copy is elided.
+    characters,
+    follows_nonquote_scalar
+  );
+}
+
+simdjson_warn_unused simdjson_inline error_code json_scanner::finish() {
+  return string_scanner.finish();
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_SCANNER_H
+/* end file generic/stage1/json_scanner.h for rvv_vls */
+
+// All other declarations
+/* including generic/stage1/find_next_document_index.h for rvv_vls: #include <generic/stage1/find_next_document_index.h> */
+/* begin file generic/stage1/find_next_document_index.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace stage1 {
+
+/**
+  * This algorithm is used to quickly identify the last structural position that
+  * makes up a complete document.
+  *
+  * It does this by going backwards and finding the last *document boundary* (a
+  * place where one value follows another without a comma between them). If the
+  * last document (the characters after the boundary) has an equal number of
+  * start and end brackets, it is considered complete.
+  *
+  * Simply put, we iterate over the structural characters, starting from
+  * the end. We consider that we found the end of a JSON document when the
+  * first element of the pair is NOT one of these characters: '{' '[' ':' ','
+  * and when the second element is NOT one of these characters: '}' ']' ':' ','.
+  *
+  * This simple comparison works most of the time, but it does not cover cases
+  * where the batch's structural indexes contain a perfect amount of documents.
+  * In such a case, we do not have access to the structural index which follows
+  * the last document, therefore, we do not have access to the second element in
+  * the pair, and that means we cannot identify the last document. To fix this
+  * issue, we keep a count of the open and closed curly/square braces we found
+  * while searching for the pair. When we find a pair AND the count of open and
+  * closed curly/square braces is the same, we know that we just passed a
+  * complete document, therefore the last json buffer location is the end of the
+  * batch.
+  */
+simdjson_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
+  // Variant: do not count separately, just figure out depth
+  if(parser.n_structural_indexes == 0) { return 0; }
+  auto arr_cnt = 0;
+  auto obj_cnt = 0;
+  for (auto i = parser.n_structural_indexes - 1; i > 0; i--) {
+    auto idxb = parser.structural_indexes[i];
+    switch (parser.buf[idxb]) {
+    case ':':
+    case ',':
+      continue;
+    case '}':
+      obj_cnt--;
+      continue;
+    case ']':
+      arr_cnt--;
+      continue;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+    }
+    auto idxa = parser.structural_indexes[i - 1];
+    switch (parser.buf[idxa]) {
+    case '{':
+    case '[':
+    case ':':
+    case ',':
+      continue;
+    }
+    // Last document is complete, so the next document will appear after!
+    if (!arr_cnt && !obj_cnt) {
+      return parser.n_structural_indexes;
+    }
+    // Last document is incomplete; mark the document at i + 1 as the next one
+    return i;
+  }
+  // If we made it to the end, we want to finish counting to see if we have a full document.
+  switch (parser.buf[parser.structural_indexes[0]]) {
+    case '}':
+      obj_cnt--;
+      break;
+    case ']':
+      arr_cnt--;
+      break;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+  }
+  if (!arr_cnt && !obj_cnt) {
+    // We have a complete document.
+    return parser.n_structural_indexes;
+  }
+  return 0;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+/* end file generic/stage1/find_next_document_index.h for rvv_vls */
+/* including generic/stage1/json_minifier.h for rvv_vls: #include <generic/stage1/json_minifier.h> */
+/* begin file generic/stage1/json_minifier.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses in stage1
+// It is intended to be included multiple times and compiled multiple times
+// We assume the file in which it is included already includes
+// "simdjson/stage1.h" (this simplifies amalgation)
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace stage1 {
+
+class json_minifier {
+public:
+  template<size_t STEP_SIZE>
+  static error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept;
+
+private:
+  simdjson_inline json_minifier(uint8_t *_dst)
+  : dst{_dst}
+  {}
+  template<size_t STEP_SIZE>
+  simdjson_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
+  simdjson_warn_unused simdjson_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
+  json_scanner scanner{};
+  uint8_t *dst;
+};
+
+simdjson_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
+  uint64_t mask = block.whitespace();
+  dst += in.compress(mask, dst);
+}
+
+simdjson_warn_unused simdjson_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
+  error_code error = scanner.finish();
+  if (error) { dst_len = 0; return error; }
+  dst_len = dst - dst_start;
+  return SUCCESS;
+}
+
+template<>
+simdjson_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block_buf);
+  simd::simd8x64<uint8_t> in_2(block_buf+64);
+  json_block block_1 = scanner.next(in_1);
+  json_block block_2 = scanner.next(in_2);
+  this->next(in_1, block_1);
+  this->next(in_2, block_2);
+  reader.advance();
+}
+
+template<>
+simdjson_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block_buf);
+  json_block block_1 = scanner.next(in_1);
+  this->next(block_buf, block_1);
+  reader.advance();
+}
+
+template<size_t STEP_SIZE>
+error_code json_minifier::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept {
+  buf_block_reader<STEP_SIZE> reader(buf, len);
+  json_minifier minifier(dst);
+
+  // Index the first n-1 blocks
+  while (reader.has_full_block()) {
+    minifier.step<STEP_SIZE>(reader.full_block(), reader);
+  }
+
+  // Index the last (remainder) block, padded with spaces
+  uint8_t block[STEP_SIZE];
+  size_t remaining_bytes = reader.get_remainder(block);
+  if (remaining_bytes > 0) {
+    // We do not want to write directly to the output stream. Rather, we write
+    // to a local buffer (for safety).
+    uint8_t out_block[STEP_SIZE];
+    uint8_t * const guarded_dst{minifier.dst};
+    minifier.dst = out_block;
+    minifier.step<STEP_SIZE>(block, reader);
+    size_t to_write = minifier.dst - out_block;
+    // In some cases, we could be enticed to consider the padded spaces
+    // as part of the string. This is fine as long as we do not write more
+    // than we consumed.
+    if(to_write > remaining_bytes) { to_write = remaining_bytes; }
+    memcpy(guarded_dst, out_block, to_write);
+    minifier.dst = guarded_dst + to_write;
+  }
+  return minifier.finish(dst, dst_len);
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_MINIFIER_H
+/* end file generic/stage1/json_minifier.h for rvv_vls */
+/* including generic/stage1/json_structural_indexer.h for rvv_vls: #include <generic/stage1/json_structural_indexer.h> */
+/* begin file generic/stage1/json_structural_indexer.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_string_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_scanner.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/json_minifier.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/find_next_document_index.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses in stage1
+// It is intended to be included multiple times and compiled multiple times
+// We assume the file in which it is included already includes
+// "simdjson/stage1.h" (this simplifies amalgation)
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace stage1 {
+
+class bit_indexer {
+public:
+  uint32_t *tail;
+
+  simdjson_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
+
+#if SIMDJSON_PREFER_REVERSE_BITS
+  /**
+    * ARM lacks a fast trailing zero instruction, but it has a fast
+    * bit reversal instruction and a fast leading zero instruction.
+    * Thus it may be profitable to reverse the bits (once) and then
+    * to rely on a sequence of instructions that call the leading
+    * zero instruction.
+    *
+    * Performance notes:
+    * The chosen routine is not optimal in terms of data dependency
+    * since zero_leading_bit might require two instructions. However,
+    * it tends to minimize the total number of instructions which is
+    * beneficial.
+    */
+  simdjson_inline void write_index(uint32_t idx, uint64_t& rev_bits, int i) {
+    int lz = leading_zeroes(rev_bits);
+    this->tail[i] = static_cast<uint32_t>(idx) + lz;
+    rev_bits = zero_leading_bit(rev_bits, lz);
+  }
+#else
+  /**
+    * Under recent x64 systems, we often have both a fast trailing zero
+    * instruction and a fast 'clear-lower-bit' instruction so the following
+    * algorithm can be competitive.
+    */
+
+  simdjson_inline void write_index(uint32_t idx, uint64_t& bits, int i) {
+    this->tail[i] = idx + trailing_zeroes(bits);
+    bits = clear_lowest_bit(bits);
+  }
+#endif // SIMDJSON_PREFER_REVERSE_BITS
+
+  template <int START, int N>
+  simdjson_inline int write_indexes(uint32_t idx, uint64_t& bits) {
+    write_index(idx, bits, START);
+    SIMDJSON_IF_CONSTEXPR (N > 1) {
+      write_indexes<(N-1>0?START+1:START), (N-1>=0?N-1:1)>(idx, bits);
+    }
+    return START+N;
+  }
+
+  template <int START, int END, int STEP>
+  simdjson_inline int write_indexes_stepped(uint32_t idx, uint64_t& bits, int cnt) {
+    write_indexes<START, STEP>(idx, bits);
+    SIMDJSON_IF_CONSTEXPR ((START+STEP)  < END) {
+      if (simdjson_unlikely((START+STEP) < cnt)) {
+        write_indexes_stepped<(START+STEP<END?START+STEP:END), END, STEP>(idx, bits, cnt);
+      }
+    }
+    return ((END-START) % STEP) == 0 ? END : (END-START) - ((END-START) % STEP) + STEP;
+  }
+
+  // flatten out values in 'bits' assuming that they are are to have values of idx
+  // plus their position in the bitvector, and store these indexes at
+  // base_ptr[base] incrementing base as we go
+  // will potentially store extra values beyond end of valid bits, so base_ptr
+  // needs to be large enough to handle this
+  //
+  // If the kernel sets SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER, then it
+  // will provide its own version of the code.
+#ifdef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+  simdjson_inline void write(uint32_t idx, uint64_t bits);
+#else
+  simdjson_inline void write(uint32_t idx, uint64_t bits) {
+    // In some instances, the next branch is expensive because it is mispredicted.
+    // Unfortunately, in other cases,
+    // it helps tremendously.
+    if (bits == 0)
+        return;
+
+    int cnt = static_cast<int>(count_ones(bits));
+
+#if SIMDJSON_PREFER_REVERSE_BITS
+    bits = reverse_bits(bits);
+#endif
+#ifdef SIMDJSON_STRUCTURAL_INDEXER_STEP
+    static constexpr const int STEP = SIMDJSON_STRUCTURAL_INDEXER_STEP;
+#else
+    static constexpr const int STEP = 4;
+#endif
+    static constexpr const int STEP_UNTIL = 24;
+
+    write_indexes_stepped<0, STEP_UNTIL, STEP>(idx, bits, cnt);
+    SIMDJSON_IF_CONSTEXPR (STEP_UNTIL < 64) {
+      if (simdjson_unlikely(STEP_UNTIL < cnt)) {
+        for (int i=STEP_UNTIL; i<cnt; i++) {
+          write_index(idx, bits, i);
+        }
+      }
+    }
+
+    this->tail += cnt;
+  }
+#endif // SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+};
+
+class json_structural_indexer {
+public:
+  /**
+   * Find the important bits of JSON in a 128-byte chunk, and add them to structural_indexes.
+   *
+   * @param partial Setting the partial parameter to true allows the find_structural_bits to
+   *   tolerate unclosed strings. The caller should still ensure that the input is valid UTF-8. If
+   *   you are processing substrings, you may want to call on a function like trimmed_length_safe_utf8.
+   */
+  template<size_t STEP_SIZE>
+  static error_code index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept;
+
+private:
+  simdjson_inline json_structural_indexer(uint32_t *structural_indexes);
+  template<size_t STEP_SIZE>
+  simdjson_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
+  simdjson_warn_unused simdjson_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
+
+  json_scanner scanner{};
+  utf8_checker checker{};
+  bit_indexer indexer;
+  uint64_t prev_structurals = 0;
+  uint64_t unescaped_chars_error = 0;
+};
+
+simdjson_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
+
+// Skip the last character if it is partial
+simdjson_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
+  if (simdjson_unlikely(len < 3)) {
+    switch (len) {
+      case 2:
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
+        return len;
+      case 1:
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        return len;
+      case 0:
+        return len;
+    }
+  }
+  if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+  if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 1 byte left
+  if (buf[len-3] >= 0xf0) { return len-3; } // 4-byte characters with only 3 bytes left
+  return len;
+}
+
+//
+// PERF NOTES:
+// We pipe 2 inputs through these stages:
+// 1. Load JSON into registers. This takes a long time and is highly parallelizable, so we load
+//    2 inputs' worth at once so that by the time step 2 is looking for them input, it's available.
+// 2. Scan the JSON for critical data: strings, scalars and operators. This is the critical path.
+//    The output of step 1 depends entirely on this information. These functions don't quite use
+//    up enough CPU: the second half of the functions is highly serial, only using 1 execution core
+//    at a time. The second input's scans has some dependency on the first ones finishing it, but
+//    they can make a lot of progress before they need that information.
+// 3. Step 1 does not use enough capacity, so we run some extra stuff while we're waiting for that
+//    to finish: utf-8 checks and generating the output from the last iteration.
+//
+// The reason we run 2 inputs at a time, is steps 2 and 3 are *still* not enough to soak up all
+// available capacity with just one input. Running 2 at a time seems to give the CPU a good enough
+// workout.
+//
+template<size_t STEP_SIZE>
+error_code json_structural_indexer::index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept {
+  if (simdjson_unlikely(len > parser.capacity())) { return CAPACITY; }
+  // We guard the rest of the code so that we can assume that len > 0 throughout.
+  if (len == 0) { return EMPTY; }
+  if (is_streaming(partial)) {
+    len = trim_partial_utf8(buf, len);
+    // If you end up with an empty window after trimming
+    // the partial UTF-8 bytes, then chances are good that you
+    // have an UTF-8 formatting error.
+    if(len == 0) { return UTF8_ERROR; }
+  }
+  buf_block_reader<STEP_SIZE> reader(buf, len);
+  json_structural_indexer indexer(parser.structural_indexes.get());
+
+  // Read all but the last block
+  while (reader.has_full_block()) {
+    indexer.step<STEP_SIZE>(reader.full_block(), reader);
+  }
+  // Take care of the last block (will always be there unless file is empty which is
+  // not supposed to happen.)
+  uint8_t block[STEP_SIZE];
+  if (simdjson_unlikely(reader.get_remainder(block) == 0)) { return UNEXPECTED_ERROR; }
+  indexer.step<STEP_SIZE>(block, reader);
+  return indexer.finish(parser, reader.block_index(), len, partial);
+}
+
+template<>
+simdjson_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block);
+  simd::simd8x64<uint8_t> in_2(block+64);
+  json_block block_1 = scanner.next(in_1);
+  json_block block_2 = scanner.next(in_2);
+  this->next(in_1, block_1, reader.block_index());
+  this->next(in_2, block_2, reader.block_index()+64);
+  reader.advance();
+}
+
+template<>
+simdjson_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
+  simd::simd8x64<uint8_t> in_1(block);
+  json_block block_1 = scanner.next(in_1);
+  this->next(in_1, block_1, reader.block_index());
+  reader.advance();
+}
+
+simdjson_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
+  uint64_t unescaped = in.lteq(0x1F);
+#if SIMDJSON_UTF8VALIDATION
+  checker.check_next_input(in);
+#endif
+  indexer.write(uint32_t(idx-64), prev_structurals); // Output *last* iteration's structurals to the parser
+  prev_structurals = block.structural_start();
+  unescaped_chars_error |= block.non_quote_inside_string(unescaped);
+}
+
+simdjson_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
+  // Write out the final iteration's structurals
+  indexer.write(uint32_t(idx-64), prev_structurals);
+  error_code error = scanner.finish();
+  // We deliberately break down the next expression so that it is
+  // human readable.
+  const bool should_we_exit = is_streaming(partial) ?
+    ((error != SUCCESS) && (error != UNCLOSED_STRING)) // when partial we tolerate UNCLOSED_STRING
+    : (error != SUCCESS); // if partial is false, we must have SUCCESS
+  const bool have_unclosed_string = (error == UNCLOSED_STRING);
+  if (simdjson_unlikely(should_we_exit)) { return error; }
+
+  if (unescaped_chars_error) {
+    return UNESCAPED_CHARS;
+  }
+  parser.n_structural_indexes = uint32_t(indexer.tail - parser.structural_indexes.get());
+  /***
+   * The On-Demand API requires special padding.
+   *
+   * This is related to https://github.com/simdjson/simdjson/issues/906
+   * Basically, we want to make sure that if the parsing continues beyond the last (valid)
+   * structural character, it quickly stops.
+   * Only three structural characters can be repeated without triggering an error in JSON:  [,] and }.
+   * We repeat the padding character (at 'len'). We don't know what it is, but if the parsing
+   * continues, then it must be [,] or }.
+   * Suppose it is ] or }. We backtrack to the first character, what could it be that would
+   * not trigger an error? It could be ] or } but no, because you can't start a document that way.
+   * It can't be a comma, a colon or any simple value. So the only way we could continue is
+   * if the repeated character is [. But if so, the document must start with [. But if the document
+   * starts with [, it should end with ]. If we enforce that rule, then we would get
+   * ][[ which is invalid.
+   *
+   * This is illustrated with the test array_iterate_unclosed_error() on the following input:
+   * R"({ "a": [,,)"
+   **/
+  parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len); // used later in partial == stage1_mode::streaming_final
+  parser.structural_indexes[parser.n_structural_indexes + 1] = uint32_t(len);
+  parser.structural_indexes[parser.n_structural_indexes + 2] = 0;
+  parser.next_structural_index = 0;
+  // a valid JSON file cannot have zero structural indexes - we should have found something
+  if (simdjson_unlikely(parser.n_structural_indexes == 0u)) {
+    return EMPTY;
+  }
+  if (simdjson_unlikely(parser.structural_indexes[parser.n_structural_indexes - 1] > len)) {
+    return UNEXPECTED_ERROR;
+  }
+  if (partial == stage1_mode::streaming_partial) {
+    // If we have an unclosed string, then the last structural
+    // will be the quote and we want to make sure to omit it.
+    if(have_unclosed_string) {
+      parser.n_structural_indexes--;
+      // a valid JSON file cannot have zero structural indexes - we should have found something
+      if (simdjson_unlikely(parser.n_structural_indexes == 0u)) { return CAPACITY; }
+    }
+    // We truncate the input to the end of the last complete document (or zero).
+    auto new_structural_indexes = find_next_document_index(parser);
+    if (new_structural_indexes == 0 && parser.n_structural_indexes > 0) {
+      if(parser.structural_indexes[0] == 0) {
+        // If the buffer is partial and we started at index 0 but the document is
+        // incomplete, it's too big to parse.
+        return CAPACITY;
+      } else {
+        // It is possible that the document could be parsed, we just had a lot
+        // of white space.
+        parser.n_structural_indexes = 0;
+        return EMPTY;
+      }
+    }
+
+    parser.n_structural_indexes = new_structural_indexes;
+  } else if (partial == stage1_mode::streaming_final) {
+    if(have_unclosed_string) { parser.n_structural_indexes--; }
+    // We truncate the input to the end of the last complete document (or zero).
+    // Because partial == stage1_mode::streaming_final, it means that we may
+    // silently ignore trailing garbage. Though it sounds bad, we do it
+    // deliberately because many people who have streams of JSON documents
+    // will truncate them for processing. E.g., imagine that you are uncompressing
+    // the data from a size file or receiving it in chunks from the network. You
+    // may not know where exactly the last document will be. Meanwhile the
+    // document_stream instances allow people to know the JSON documents they are
+    // parsing (see the iterator.source() method).
+    parser.n_structural_indexes = find_next_document_index(parser);
+    // We store the initial n_structural_indexes so that the client can see
+    // whether we used truncation. If initial_n_structural_indexes == parser.n_structural_indexes,
+    // then this will query parser.structural_indexes[parser.n_structural_indexes] which is len,
+    // otherwise, it will copy some prior index.
+    parser.structural_indexes[parser.n_structural_indexes + 1] = parser.structural_indexes[parser.n_structural_indexes];
+    // This next line is critical, do not change it unless you understand what you are
+    // doing.
+    parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len);
+    if (simdjson_unlikely(parser.n_structural_indexes == 0u)) {
+        // We tolerate an unclosed string at the very end of the stream. Indeed, users
+        // often load their data in bulk without being careful and they want us to ignore
+        // the trailing garbage.
+        return EMPTY;
+    }
+  }
+  checker.check_eof();
+  return checker.errors();
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+// Clear CUSTOM_BIT_INDEXER so other implementations can set it if they need to.
+#undef SIMDJSON_GENERIC_JSON_STRUCTURAL_INDEXER_CUSTOM_BIT_INDEXER
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_JSON_STRUCTURAL_INDEXER_H
+/* end file generic/stage1/json_structural_indexer.h for rvv_vls */
+/* including generic/stage1/utf8_validator.h for rvv_vls: #include <generic/stage1/utf8_validator.h> */
+/* begin file generic/stage1/utf8_validator.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/buf_block_reader.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage1/utf8_lookup4_algorithm.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace stage1 {
+
+/**
+ * Validates that the string is actual UTF-8.
+ */
+template<class checker>
+bool generic_validate_utf8(const uint8_t * input, size_t length) {
+    checker c{};
+    buf_block_reader<64> reader(input, length);
+    while (reader.has_full_block()) {
+      simd::simd8x64<uint8_t> in(reader.full_block());
+      c.check_next_input(in);
+      reader.advance();
+    }
+    uint8_t block[64]{};
+    reader.get_remainder(block);
+    simd::simd8x64<uint8_t> in(block);
+    c.check_next_input(in);
+    reader.advance();
+    c.check_eof();
+    return c.errors() == error_code::SUCCESS;
+}
+
+bool generic_validate_utf8(const char * input, size_t length) {
+    return generic_validate_utf8<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_UTF8_VALIDATOR_H
+/* end file generic/stage1/utf8_validator.h for rvv_vls */
+/* end file generic/stage1/amalgamated.h for rvv_vls */
+/* including generic/stage2/amalgamated.h for rvv_vls: #include <generic/stage2/amalgamated.h> */
+/* begin file generic/stage2/amalgamated.h for rvv_vls */
+// Stuff other things depend on
+/* including generic/stage2/base.h for rvv_vls: #include <generic/stage2/base.h> */
+/* begin file generic/stage2/base.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_BASE_H */
+/* amalgamation skipped (editor-only): #include <generic/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace stage2 {
+
+class json_iterator;
+class structural_iterator;
+struct tape_builder;
+struct tape_writer;
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_BASE_H
+/* end file generic/stage2/base.h for rvv_vls */
+/* including generic/stage2/tape_writer.h for rvv_vls: #include <generic/stage2/tape_writer.h> */
+/* begin file generic/stage2/tape_writer.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/internal/tape_type.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace stage2 {
+
+struct tape_writer {
+  /** The next place to write to tape */
+  uint64_t *next_tape_loc;
+
+  /** Write a signed 64-bit value to tape. */
+  simdjson_inline void append_s64(int64_t value) noexcept;
+
+  /** Write an unsigned 64-bit value to tape. */
+  simdjson_inline void append_u64(uint64_t value) noexcept;
+
+  /** Write a double value to tape. */
+  simdjson_inline void append_double(double value) noexcept;
+
+  /** Write a big integer (as string) to tape. src points to first digit, len is byte count. */
+  simdjson_inline void append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept;
+
+  /**
+   * Append a tape entry (an 8-bit type,and 56 bits worth of value).
+   */
+  simdjson_inline void append(uint64_t val, internal::tape_type t) noexcept;
+
+  /**
+   * Skip the current tape entry without writing.
+   *
+   * Used to skip the start of the container, since we'll come back later to fill it in when the
+   * container ends.
+   */
+  simdjson_inline void skip() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a large u64 or i64.
+   */
+  simdjson_inline void skip_large_integer() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a double.
+   */
+  simdjson_inline void skip_double() noexcept;
+
+  /**
+   * Write a value to a known location on tape.
+   *
+   * Used to go back and write out the start of a container after the container ends.
+   */
+  simdjson_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
+
+private:
+  /**
+   * Append both the tape entry, and a supplementary value following it. Used for types that need
+   * all 64 bits, such as double and uint64_t.
+   */
+  template<typename T>
+  simdjson_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
+}; // struct tape_writer
+
+simdjson_inline void tape_writer::append_s64(int64_t value) noexcept {
+  append2(0, value, internal::tape_type::INT64);
+}
+
+simdjson_inline void tape_writer::append_u64(uint64_t value) noexcept {
+  append(0, internal::tape_type::UINT64);
+  *next_tape_loc = value;
+  next_tape_loc++;
+}
+
+/** Write a double value to tape. */
+simdjson_inline void tape_writer::append_double(double value) noexcept {
+  append2(0, value, internal::tape_type::DOUBLE);
+}
+
+simdjson_inline void tape_writer::skip() noexcept {
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::skip_large_integer() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::skip_double() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
+  *next_tape_loc = val | ((uint64_t(char(t))) << 56);
+  next_tape_loc++;
+}
+
+template<typename T>
+simdjson_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
+  append(val, t);
+  static_assert(sizeof(val2) == sizeof(*next_tape_loc), "Type is not 64 bits!");
+  memcpy(next_tape_loc, &val2, sizeof(val2));
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
+  tape_loc = val | ((uint64_t(char(t))) << 56);
+}
+
+simdjson_inline void tape_writer::append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept {
+  // Write to string buffer: [4-byte LE length][digits][null]
+  uint32_t str_len = uint32_t(len);
+  memcpy(string_buf, &str_len, sizeof(uint32_t));
+  memcpy(string_buf + sizeof(uint32_t), src, len);
+  string_buf[sizeof(uint32_t) + len] = 0;
+  // Tape entry: offset into string buffer
+  // The caller must set the offset relative to doc.string_buf base
+  append(0, internal::tape_type::BIGINT); // placeholder offset, caller patches
+  string_buf += sizeof(uint32_t) + len + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+/* end file generic/stage2/tape_writer.h for rvv_vls */
+/* including generic/stage2/logger.h for rvv_vls: #include <generic/stage2/logger.h> */
+/* begin file generic/stage2/logger.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+
+// This is for an internal-only stage 2 specific logger.
+// Set LOG_ENABLED = true to log what stage 2 is doing!
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace logger {
+
+  static constexpr const char * DASHES = "----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------";
+
+#if SIMDJSON_VERBOSE_LOGGING
+  static constexpr const bool LOG_ENABLED = true;
+#else
+  static constexpr const bool LOG_ENABLED = false;
+#endif
+  static constexpr const int LOG_EVENT_LEN = 20;
+  static constexpr const int LOG_BUFFER_LEN = 30;
+  static constexpr const int LOG_SMALL_BUFFER_LEN = 10;
+  static constexpr const int LOG_INDEX_LEN = 5;
+
+  static int log_depth; // Not threadsafe. Log only.
+
+  // Helper to turn unprintable or newline characters into spaces
+  static simdjson_inline char printable_char(char c) {
+    if (c >= 0x20) {
+      return c;
+    } else {
+      return ' ';
+    }
+  }
+
+  // Print the header and set up log_start
+  static simdjson_inline void log_start() {
+    if (LOG_ENABLED) {
+      log_depth = 0;
+      printf("\n");
+      printf("| %-*s | %-*s | %-*s | %-*s | Detail |\n", LOG_EVENT_LEN, "Event", LOG_BUFFER_LEN, "Buffer", LOG_SMALL_BUFFER_LEN, "Next", 5, "Next#");
+      printf("|%.*s|%.*s|%.*s|%.*s|--------|\n", LOG_EVENT_LEN+2, DASHES, LOG_BUFFER_LEN+2, DASHES, LOG_SMALL_BUFFER_LEN+2, DASHES, 5+2, DASHES);
+    }
+  }
+
+  simdjson_unused static simdjson_inline void log_string(const char *message) {
+    if (LOG_ENABLED) {
+      printf("%s\n", message);
+    }
+  }
+
+  // Logs a single line from the stage 2 DOM parser
+  template<typename S>
+  static simdjson_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
+    if (LOG_ENABLED) {
+      printf("| %*s%s%-*s ", log_depth*2, "", title_prefix, LOG_EVENT_LEN - log_depth*2 - int(strlen(title_prefix)), title);
+      auto current_index = structurals.at_beginning() ? nullptr : structurals.next_structural-1;
+      auto next_index = structurals.next_structural;
+      auto current = current_index ? &structurals.buf[*current_index] : reinterpret_cast<const uint8_t*>("                                                       ");
+      auto next = &structurals.buf[*next_index];
+      {
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_BUFFER_LEN;i++) {
+          printf("%c", printable_char(current[i]));
+        }
+        printf(" ");
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_SMALL_BUFFER_LEN;i++) {
+          printf("%c", printable_char(next[i]));
+        }
+        printf(" ");
+      }
+      if (current_index) {
+        printf("| %*u ", LOG_INDEX_LEN, *current_index);
+      } else {
+        printf("| %-*s ", LOG_INDEX_LEN, "");
+      }
+      // printf("| %*u ", LOG_INDEX_LEN, structurals.next_tape_index());
+      printf("| %-s ", detail);
+      printf("|\n");
+    }
+  }
+
+} // namespace logger
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+/* end file generic/stage2/logger.h for rvv_vls */
+
+// All other declarations
+/* including generic/stage2/json_iterator.h for rvv_vls: #include <generic/stage2/json_iterator.h> */
+/* begin file generic/stage2/json_iterator.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/logger.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace stage2 {
+
+class json_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+  uint32_t depth{0};
+
+  /**
+   * Walk the JSON document.
+   *
+   * The visitor receives callbacks when values are encountered. All callbacks pass the iterator as
+   * the first parameter; some callbacks have other parameters as well:
+   *
+   * - visit_document_start() - at the beginning.
+   * - visit_document_end() - at the end (if things were successful).
+   *
+   * - visit_array_start() - at the start `[` of a non-empty array.
+   * - visit_array_end() - at the end `]` of a non-empty array.
+   * - visit_empty_array() - when an empty array is encountered.
+   *
+   * - visit_object_end() - at the start `]` of a non-empty object.
+   * - visit_object_start() - at the end `]` of a non-empty object.
+   * - visit_empty_object() - when an empty object is encountered.
+   * - visit_key(const uint8_t *key) - when a key in an object field is encountered. key is
+   *                                   guaranteed to point at the first quote of the string (`"key"`).
+   * - visit_primitive(const uint8_t *value) - when a value is a string, number, boolean or null.
+   * - visit_root_primitive(iter, uint8_t *value) - when the top-level value is a string, number, boolean or null.
+   *
+   * - increment_count(iter) - each time a value is found in an array or object.
+   */
+  template<bool STREAMING, typename V>
+  simdjson_warn_unused simdjson_inline error_code walk_document(V &visitor) noexcept;
+
+  /**
+   * Create an iterator capable of walking a JSON document.
+   *
+   * The document must have already passed through stage 1.
+   */
+  simdjson_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
+
+  /**
+   * Look at the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *peek() const noexcept;
+  /**
+   * Advance to the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *advance() noexcept;
+  /**
+   * Get the remaining length of the document, from the start of the current token.
+   */
+  simdjson_inline size_t remaining_len() const noexcept;
+  /**
+   * Check if we are at the end of the document.
+   *
+   * If this is true, there are no more tokens.
+   */
+  simdjson_inline bool at_eof() const noexcept;
+  /**
+   * Check if we are at the beginning of the document.
+   */
+  simdjson_inline bool at_beginning() const noexcept;
+  simdjson_inline uint8_t last_structural() const noexcept;
+
+  /**
+   * Log that a value has been found.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_value(const char *type) const noexcept;
+  /**
+   * Log the start of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_start_value(const char *type) const noexcept;
+  /**
+   * Log the end of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_end_value(const char *type) const noexcept;
+  /**
+   * Log an error.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_error(const char *error) const noexcept;
+
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
+};
+
+template<bool STREAMING, typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::walk_document(V &visitor) noexcept {
+  logger::log_start();
+
+  //
+  // Start the document
+  //
+  if (at_eof()) { return EMPTY; }
+  log_start_value("document");
+  SIMDJSON_TRY( visitor.visit_document_start(*this) );
+
+  //
+  // Read first value
+  //
+  {
+    auto value = advance();
+
+    // Make sure the outer object or array is closed before continuing; otherwise, there are ways we
+    // could get into memory corruption. See https://github.com/simdjson/simdjson/issues/906
+    if (!STREAMING) {
+      switch (*value) {
+        case '{': if (last_structural() != '}') { log_value("starting brace unmatched"); return TAPE_ERROR; }; break;
+        case '[': if (last_structural() != ']') { log_value("starting bracket unmatched"); return TAPE_ERROR; }; break;
+      }
+    }
+
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_root_primitive(*this, value) ); break;
+    }
+  }
+  goto document_end;
+
+//
+// Object parser states
+//
+object_begin:
+  log_start_value("object");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = false;
+  SIMDJSON_TRY( visitor.visit_object_start(*this) );
+
+  {
+    auto key = advance();
+    if (*key != '"') { log_error("Object does not start with a key"); return TAPE_ERROR; }
+    SIMDJSON_TRY( visitor.increment_count(*this) );
+    SIMDJSON_TRY( visitor.visit_key(*this, key) );
+  }
+
+object_field:
+  if (simdjson_unlikely( *advance() != ':' )) { log_error("Missing colon after key in object"); return TAPE_ERROR; }
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+object_continue:
+  switch (*advance()) {
+    case ',':
+      SIMDJSON_TRY( visitor.increment_count(*this) );
+      {
+        auto key = advance();
+        if (simdjson_unlikely( *key != '"' )) { log_error("Key string missing at beginning of field in object"); return TAPE_ERROR; }
+        SIMDJSON_TRY( visitor.visit_key(*this, key) );
+      }
+      goto object_field;
+    case '}': log_end_value("object"); SIMDJSON_TRY( visitor.visit_object_end(*this) ); goto scope_end;
+    default: log_error("No comma between object fields"); return TAPE_ERROR;
+  }
+
+scope_end:
+  depth--;
+  if (depth == 0) { goto document_end; }
+  if (dom_parser.is_array[depth]) { goto array_continue; }
+  goto object_continue;
+
+//
+// Array parser states
+//
+array_begin:
+  log_start_value("array");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = true;
+  SIMDJSON_TRY( visitor.visit_array_start(*this) );
+  SIMDJSON_TRY( visitor.increment_count(*this) );
+
+array_value:
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+array_continue:
+  switch (*advance()) {
+    case ',': SIMDJSON_TRY( visitor.increment_count(*this) ); goto array_value;
+    case ']': log_end_value("array"); SIMDJSON_TRY( visitor.visit_array_end(*this) ); goto scope_end;
+    default: log_error("Missing comma between array values"); return TAPE_ERROR;
+  }
+
+document_end:
+  log_end_value("document");
+  SIMDJSON_TRY( visitor.visit_document_end(*this) );
+
+  dom_parser.next_structural_index = uint32_t(next_structural - &dom_parser.structural_indexes[0]);
+
+  // If we didn't make it to the end, it's an error
+  if ( !STREAMING && dom_parser.next_structural_index != dom_parser.n_structural_indexes ) {
+    log_error("More than one JSON value at the root of the document, or extra characters at the end of the JSON!");
+    return TAPE_ERROR;
+  }
+
+  return SUCCESS;
+
+} // walk_document()
+
+simdjson_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+  : buf{_dom_parser.buf},
+    next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+    dom_parser{_dom_parser} {
+}
+
+simdjson_inline const uint8_t *json_iterator::peek() const noexcept {
+  return &buf[*(next_structural)];
+}
+simdjson_inline const uint8_t *json_iterator::advance() noexcept {
+  return &buf[*(next_structural++)];
+}
+simdjson_inline size_t json_iterator::remaining_len() const noexcept {
+  return dom_parser.len - *(next_structural-1);
+}
+
+simdjson_inline bool json_iterator::at_eof() const noexcept {
+  return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+}
+simdjson_inline bool json_iterator::at_beginning() const noexcept {
+  return next_structural == dom_parser.structural_indexes.get();
+}
+simdjson_inline uint8_t json_iterator::last_structural() const noexcept {
+  return buf[dom_parser.structural_indexes[dom_parser.n_structural_indexes - 1]];
+}
+
+simdjson_inline void json_iterator::log_value(const char *type) const noexcept {
+  logger::log_line(*this, "", type, "");
+}
+
+simdjson_inline void json_iterator::log_start_value(const char *type) const noexcept {
+  logger::log_line(*this, "+", type, "");
+  if (logger::LOG_ENABLED) { logger::log_depth++; }
+}
+
+simdjson_inline void json_iterator::log_end_value(const char *type) const noexcept {
+  if (logger::LOG_ENABLED) { logger::log_depth--; }
+  logger::log_line(*this, "-", type, "");
+}
+
+simdjson_inline void json_iterator::log_error(const char *error) const noexcept {
+  logger::log_line(*this, "", "ERROR", error);
+}
+
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
+  switch (*value) {
+    case '"': return visitor.visit_root_string(*this, value);
+    case 't': return visitor.visit_root_true_atom(*this, value);
+    case 'f': return visitor.visit_root_false_atom(*this, value);
+    case 'n': return visitor.visit_root_null_atom(*this, value);
+    case '-':
+    case '0': case '1': case '2': case '3': case '4':
+    case '5': case '6': case '7': case '8': case '9':
+      return visitor.visit_root_number(*this, value);
+    default:
+      log_error("Document starts with a non-value character");
+      return TAPE_ERROR;
+  }
+}
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
+  // Use the fact that most scalars are going to be either strings or numbers.
+  if(*value == '"') {
+    return visitor.visit_string(*this, value);
+  } else if (((*value - '0')  < 10) || (*value == '-')) {
+    return visitor.visit_number(*this, value);
+  }
+  // true, false, null are uncommon.
+  switch (*value) {
+    case 't': return visitor.visit_true_atom(*this, value);
+    case 'f': return visitor.visit_false_atom(*this, value);
+    case 'n': return visitor.visit_null_atom(*this, value);
+    default:
+      log_error("Non-value found when value was expected!");
+      return TAPE_ERROR;
+  }
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+/* end file generic/stage2/json_iterator.h for rvv_vls */
+/* including generic/stage2/stringparsing.h for rvv_vls: #include <generic/stage2/stringparsing.h> */
+/* begin file generic/stage2/stringparsing.h for rvv_vls */
+#include <cstdint>
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/jsoncharutils.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses
+// It is intended to be included multiple times and compiled multiple times
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+/// @private
+namespace stringparsing {
+
+// begin copypasta
+// These chars yield themselves: " \ /
+// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
+// u not handled in this table as it's complex
+static const uint8_t escape_map[256] = {
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
+    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
+    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+};
+
+// handle a unicode codepoint
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr, bool allow_replacement) {
+  // Use the default Unicode Character 'REPLACEMENT CHARACTER' (U+FFFD)
+  constexpr uint32_t substitution_code_point = 0xfffd;
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) != ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+    } else {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+
+      // We have already checked that the high surrogate is valid and
+      // (code_point - 0xd800) < 1024.
+      //
+      // Check that code_point_2 is in the range 0xdc00..0xdfff
+      // and that code_point_2 was parsed from valid hex.
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if (low_bit >> 10) {
+        if(!allow_replacement) { return false; }
+        code_point = substitution_code_point;
+      } else {
+        code_point =  (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+
+    }
+  } else if (code_point >= 0xdc00 && code_point <= 0xdfff) {
+      // If we encounter a low surrogate (not preceded by a high surrogate)
+      // then we have an error.
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+  }
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+// handle a unicode codepoint using the wobbly convention
+// https://simonsapin.github.io/wtf-8/
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint_wobbly(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr) {
+  // It is not ideal that this function is nearly identical to handle_unicode_codepoint.
+  //
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) == ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if ((low_bit >> 10) ==  0) {
+        code_point =
+          (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+    }
+  }
+
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+/**
+ * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+ * must be an unescaped quote terminating the string. It returns the final output
+ * position as pointer. In case of error (e.g., the string has bad escaped codes),
+ * then null_ptr is returned. It is assumed that the output buffer is large
+ * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+ * SIMDJSON_PADDING bytes.
+ */
+simdjson_warn_unused simdjson_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) {
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint(&src, &dst, allow_replacement)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+simdjson_warn_unused simdjson_inline uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) {
+  // It is not ideal that this function is nearly identical to parse_string.
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint_wobbly(&src, &dst)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+} // namespace stringparsing
+
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+/* end file generic/stage2/stringparsing.h for rvv_vls */
+/* including generic/stage2/structural_iterator.h for rvv_vls: #include <generic/stage2/structural_iterator.h> */
+/* begin file generic/stage2/structural_iterator.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace stage2 {
+
+class structural_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+
+  // Start a structural
+  simdjson_inline structural_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+    : buf{_dom_parser.buf},
+      next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+      dom_parser{_dom_parser} {
+  }
+  // Get the buffer position of the current structural character
+  simdjson_inline const uint8_t* current() {
+    return &buf[*(next_structural-1)];
+  }
+  // Get the current structural character
+  simdjson_inline char current_char() {
+    return buf[*(next_structural-1)];
+  }
+  // Get the next structural character without advancing
+  simdjson_inline char peek_next_char() {
+    return buf[*next_structural];
+  }
+  simdjson_inline const uint8_t* peek() {
+    return &buf[*next_structural];
+  }
+  simdjson_inline const uint8_t* advance() {
+    return &buf[*(next_structural++)];
+  }
+  simdjson_inline char advance_char() {
+    return buf[*(next_structural++)];
+  }
+  simdjson_inline size_t remaining_len() {
+    return dom_parser.len - *(next_structural-1);
+  }
+
+  simdjson_inline bool at_end() {
+    return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+  }
+  simdjson_inline bool at_beginning() {
+    return next_structural == dom_parser.structural_indexes.get();
+  }
+};
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRUCTURAL_ITERATOR_H
+/* end file generic/stage2/structural_iterator.h for rvv_vls */
+/* including generic/stage2/tape_builder.h for rvv_vls: #include <generic/stage2/tape_builder.h> */
+/* begin file generic/stage2/tape_builder.h for rvv_vls */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/json_iterator.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/stringparsing.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/tape_writer.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/dom/document.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/atomparsing.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/numberparsing.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+
+namespace simdjson {
+namespace rvv_vls {
+namespace {
+namespace stage2 {
+
+struct tape_builder {
+  template<bool STREAMING>
+  simdjson_warn_unused static simdjson_inline error_code parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept;
+
+  /** Called when a non-empty document starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty document ends without error. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_end(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty array starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty array ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_end(json_iterator &iter) noexcept;
+  /** Called when an empty array is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_array(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty object starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_start(json_iterator &iter) noexcept;
+  /**
+   * Called when a key in a field is encountered.
+   *
+   * primitive, visit_object_start, visit_empty_object, visit_array_start, or visit_empty_array
+   * will be called after this with the field value.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
+  /** Called when a non-empty object ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_end(json_iterator &iter) noexcept;
+  /** Called when an empty object is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_object(json_iterator &iter) noexcept;
+
+  /**
+   * Called when a string, number, boolean or null is found.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  /**
+   * Called when a string, number, boolean or null is found at the top level of a document (i.e.
+   * when there is no array or object and the entire document is a single string, number, boolean or
+   * null.
+   *
+   * This is separate from primitive() because simdjson's normal primitive parsing routines assume
+   * there is at least one more token after the value, which is only true in an array or object.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  /** Called each time a new field or element in an array or object is found. */
+  simdjson_warn_unused simdjson_inline error_code increment_count(json_iterator &iter) noexcept;
+
+  /** Next location to write to tape */
+  tape_writer tape;
+private:
+  /** Next write location in the string buf for stage 2 parsing */
+  uint8_t *current_string_buf_loc;
+
+  simdjson_inline tape_builder(dom::document &doc) noexcept;
+
+  simdjson_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
+  simdjson_inline void start_container(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_warn_unused simdjson_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
+  simdjson_inline void on_end_string(uint8_t *dst) noexcept;
+}; // struct tape_builder
+
+template<bool STREAMING>
+simdjson_warn_unused simdjson_inline error_code tape_builder::parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept {
+  dom_parser.doc = &doc;
+  json_iterator iter(dom_parser, STREAMING ? dom_parser.next_structural_index : 0);
+  tape_builder builder(doc);
+  return iter.walk_document<STREAMING>(builder);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_root_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
+  constexpr uint32_t start_tape_index = 0;
+  tape.append(start_tape_index, internal::tape_type::ROOT);
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter), internal::tape_type::ROOT);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
+  return visit_string(iter, key, true);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].count++; // we have a key value pair in the object at parser.dom_parser.depth - 1
+  return SUCCESS;
+}
+
+simdjson_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
+  iter.log_value(key ? "key" : "string");
+  uint8_t *dst = on_start_string(iter);
+  dst = stringparsing::parse_string(value+1, dst, false); // We do not allow replacement when the escape characters are invalid.
+  if (dst == nullptr) {
+    iter.log_error("Invalid escape in string");
+    return STRING_ERROR;
+  }
+  on_end_string(dst);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
+  return visit_string(iter, value);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("number");
+  error_code err = numberparsing::parse_number(value, tape);
+  if (simdjson_unlikely(err == BIGINT_ERROR &&
+      iter.dom_parser._number_as_string)) {
+    // Write big integer to string buffer using the same format as strings.
+    // Scan digits the same way parse_number does (skip optional '-', then digits).
+    const uint8_t *p = value;
+    if (*p == '-') p++;
+    while (numberparsing::is_digit(*p)) p++;
+    size_t len = size_t(p - value);
+    tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::BIGINT);
+    uint8_t *dst = current_string_buf_loc + sizeof(uint32_t);
+    memcpy(dst, value, len);
+    dst += len;
+    on_end_string(dst);
+    return SUCCESS;
+  }
+  return err;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
+  //
+  // We need to make a copy to make sure that the string is space terminated.
+  // This is not about padding the input, which should already padded up
+  // to len + SIMDJSON_PADDING. However, we have no control at this stage
+  // on how the padding was done. What if the input string was padded with nulls?
+  // It is quite common for an input string to have an extra null character (C string).
+  // We do not want to allow 9\0 (where \0 is the null character) inside a JSON
+  // document, but the string "9\0" by itself is fine. So we make a copy and
+  // pad the input with spaces when we know that there is just one input element.
+  // This copy is relatively expensive, but it will almost never be called in
+  // practice unless you are in the strange scenario where you have many JSON
+  // documents made of single atoms.
+  //
+  std::unique_ptr<uint8_t[]>copy(new (std::nothrow) uint8_t[iter.remaining_len() + SIMDJSON_PADDING]);
+  if (copy.get() == nullptr) { return MEMALLOC; }
+  std::memcpy(copy.get(), value, iter.remaining_len());
+  std::memset(copy.get() + iter.remaining_len(), ' ', SIMDJSON_PADDING);
+  error_code error = visit_number(iter, copy.get());
+  return error;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value)) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value, iter.remaining_len())) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value)) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value, iter.remaining_len())) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value)) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value, iter.remaining_len())) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+// private:
+
+simdjson_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
+  return uint32_t(tape.next_tape_loc - iter.dom_parser.doc->tape.get());
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  auto start_index = next_tape_index(iter);
+  tape.append(start_index+2, start);
+  tape.append(start_index, end);
+  return SUCCESS;
+}
+
+simdjson_inline void tape_builder::start_container(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].tape_index = next_tape_index(iter);
+  iter.dom_parser.open_containers[iter.depth].count = 0;
+  tape.skip(); // We don't actually *write* the start element until the end.
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  // Write the ending tape element, pointing at the start location
+  const uint32_t start_tape_index = iter.dom_parser.open_containers[iter.depth].tape_index;
+  tape.append(start_tape_index, end);
+  // Write the start tape element, pointing at the end location (and including count)
+  // count can overflow if it exceeds 24 bits... so we saturate
+  // the convention being that a cnt of 0xffffff or more is undetermined in value (>=  0xffffff).
+  const uint32_t count = iter.dom_parser.open_containers[iter.depth].count;
+  const uint32_t cntsat = count > 0xFFFFFF ? 0xFFFFFF : count;
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter) | (uint64_t(cntsat) << 32), start);
+  return SUCCESS;
+}
+
+simdjson_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
+  // we advance the point, accounting for the fact that we have a NULL termination
+  tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::STRING);
+  return current_string_buf_loc + sizeof(uint32_t);
+}
+
+simdjson_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
+  uint32_t str_length = uint32_t(dst - (current_string_buf_loc + sizeof(uint32_t)));
+  // TODO check for overflow in case someone has a crazy string (>=4GB?)
+  // But only add the overflow check when the document itself exceeds 4GB
+  // Currently unneeded because we refuse to parse docs larger or equal to 4GB.
+  memcpy(current_string_buf_loc, &str_length, sizeof(uint32_t));
+  // NULL termination is still handy if you expect all your strings to
+  // be NULL terminated? It comes at a small cost
+  *dst = 0;
+  current_string_buf_loc = dst + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+/* end file generic/stage2/tape_builder.h for rvv_vls */
+/* end file generic/stage2/amalgamated.h for rvv_vls */
+
+//
+// Stage 1
+//
+namespace simdjson {
+namespace rvv_vls {
+
+simdjson_warn_unused error_code implementation::create_dom_parser_implementation(
+  size_t capacity,
+  size_t max_depth,
+  std::unique_ptr<internal::dom_parser_implementation>& dst
+) const noexcept {
+  dst.reset( new (std::nothrow) dom_parser_implementation() );
+  if (!dst) { return MEMALLOC; }
+  if (auto err = dst->set_capacity(capacity))
+    return err;
+  if (auto err = dst->set_max_depth(max_depth))
+    return err;
+  return SUCCESS;
+}
+
+namespace {
+
+using namespace simd;
+
+simdjson_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
+  static const uint8_t wsTable[16] = { ' ', 100, 100, 100, 17, 100, 113, 2, 100, '\t', '\n', 112, 100, '\r', 100, 100 };
+  static const uint8_t opTable[16] = { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, ':', '{', ',', '}', 0, 0 };
+  vuint8_t vws = __riscv_vle8_v_u8m1(wsTable, 16);
+  vuint8_t vop = __riscv_vle8_v_u8m1(opTable, 16);
+  vuint8x64_t lo = __riscv_vand(in, 15, 64);
+  vuint8x64_t curl = __riscv_vor(in, 0x20, 64);
+
+#if __riscv_v_fixed_vlen == 128
+  vboolx64_t mws = __riscv_vmseq(simdutf_vrgather_u8m1x4(vws, lo), in, 64);
+  vboolx64_t mop = __riscv_vmseq(simdutf_vrgather_u8m1x4(vop, lo), curl, 64);
+#elif __riscv_v_fixed_vlen == 256
+  vboolx64_t mws = __riscv_vmseq(simdutf_vrgather_u8m1x2(vws, lo), in, 64);
+  vboolx64_t mop = __riscv_vmseq(simdutf_vrgather_u8m1x2(vop, lo), curl, 64);
+#else
+  vboolx64_t mws = __riscv_vmseq(__riscv_vrgather(vws, lo, 64), in, 64);
+  vboolx64_t mop = __riscv_vmseq(__riscv_vrgather(vop, lo, 64), curl, 64);
+#endif
+
+  return {
+      __riscv_vmv_x(__riscv_vreinterpret_u64m1(mws)),
+      __riscv_vmv_x(__riscv_vreinterpret_u64m1(mop))
+  };
+}
+
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input) {
+  return input.is_ascii();
+}
+
+simdjson_inline simd8<uint8_t> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+    simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-0x80); // Only 111_____ will be >= 0x80
+    simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-0x80); // Only 1111____ will be >= 0x80
+    return is_third_byte | is_fourth_byte;
+}
+
+} // unnamed namespace
+} // namespace rvv_vls
+} // namespace simdjson
+
+//
+// Stage 2
+//
+
+//
+// Implementation-specific overrides
+//
+namespace simdjson {
+namespace rvv_vls {
+
+simdjson_warn_unused error_code implementation::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept {
+  return rvv_vls::stage1::json_minifier::minify<64>(buf, len, dst, dst_len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage1(const uint8_t *_buf, size_t _len, stage1_mode streaming) noexcept {
+  this->buf = _buf;
+  this->len = _len;
+  return rvv_vls::stage1::json_structural_indexer::index<64>(buf, len, *this, streaming);
+}
+
+simdjson_warn_unused bool implementation::validate_utf8(const char *buf, size_t len) const noexcept {
+  return rvv_vls::stage1::generic_validate_utf8(buf,len);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<false>(*this, _doc);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2_next(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<true>(*this, _doc);
+}
+
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) const noexcept {
+  return rvv_vls::stringparsing::parse_string(src, dst, allow_replacement);
+}
+
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept {
+  return rvv_vls::stringparsing::parse_wobbly_string(src, dst);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::parse(const uint8_t *_buf, size_t _len, dom::document &_doc) noexcept {
+  auto error = stage1(_buf, _len, stage1_mode::regular);
+  if (error) { return error; }
+  return stage2(_doc);
+}
+
+} // namespace rvv_vls
+} // namespace simdjson
+
+/* including simdjson/rvv-vls/end.h: #include <simdjson/rvv-vls/end.h> */
+/* begin file simdjson/rvv-vls/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+/* undefining SIMDJSON_IMPLEMENTATION from "rvv_vls" */
+#undef SIMDJSON_IMPLEMENTATION
+/* end file simdjson/rvv-vls/end.h */
+
+#endif // SIMDJSON_SRC_RVV_VLS_CPP
+/* end file rvv-vls.cpp */
+#endif
+#if SIMDJSON_IMPLEMENTATION_FALLBACK
+/* including fallback.cpp: #include <fallback.cpp> */
+/* begin file fallback.cpp */
+#ifndef SIMDJSON_SRC_FALLBACK_CPP
+#define SIMDJSON_SRC_FALLBACK_CPP
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include <base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+/* including simdjson/fallback.h: #include <simdjson/fallback.h> */
+/* begin file simdjson/fallback.h */
+#ifndef SIMDJSON_FALLBACK_H
+#define SIMDJSON_FALLBACK_H
+
+/* including simdjson/fallback/begin.h: #include "simdjson/fallback/begin.h" */
+/* begin file simdjson/fallback/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "fallback" */
+#define SIMDJSON_IMPLEMENTATION fallback
+/* including simdjson/fallback/base.h: #include "simdjson/fallback/base.h" */
+/* begin file simdjson/fallback/base.h */
+#ifndef SIMDJSON_FALLBACK_BASE_H
+#define SIMDJSON_FALLBACK_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+/**
+ * Fallback implementation (runs on any machine).
+ */
+namespace fallback {
+
+class implementation;
+
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_FALLBACK_BASE_H
+/* end file simdjson/fallback/base.h */
+/* including simdjson/fallback/bitmanipulation.h: #include "simdjson/fallback/bitmanipulation.h" */
+/* begin file simdjson/fallback/bitmanipulation.h */
+#ifndef SIMDJSON_FALLBACK_BITMANIPULATION_H
+#define SIMDJSON_FALLBACK_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace fallback {
+namespace {
+
+#if defined(_MSC_VER) && !defined(_M_ARM64) && !defined(_M_X64)
+static inline unsigned char _BitScanForward64(unsigned long* ret, uint64_t x) {
+  unsigned long x0 = (unsigned long)x, top, bottom;
+  _BitScanForward(&top, (unsigned long)(x >> 32));
+  _BitScanForward(&bottom, x0);
+  *ret = x0 ? bottom : 32 + top;
+  return x != 0;
+}
+static unsigned char _BitScanReverse64(unsigned long* ret, uint64_t x) {
+  unsigned long x1 = (unsigned long)(x >> 32), top, bottom;
+  _BitScanReverse(&top, x1);
+  _BitScanReverse(&bottom, (unsigned long)x);
+  *ret = x1 ? top + 32 : bottom;
+  return x != 0;
+}
+#endif
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+#ifdef _MSC_VER
+  unsigned long leading_zero = 0;
+  // Search the mask data from most significant bit (MSB)
+  // to least significant bit (LSB) for a set bit (1).
+  if (_BitScanReverse64(&leading_zero, input_num))
+    return (int)(63 - leading_zero);
+  else
+    return 64;
+#else
+  return __builtin_clzll(input_num);
+#endif// _MSC_VER
+}
+
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+#ifdef _MSC_VER
+  unsigned long trailing_zero = 0;
+  // Search the mask data from least significant bit (LSB)
+  // to most significant bit (MSB) for a set bit (1).
+  if (_BitScanForward64(&trailing_zero, input_num))
+    return (int)trailing_zero;
+  else    return 64;
+#else
+  return __builtin_ctzll(input_num);
+#endif// _MSC_VER
+}
+
+} // unnamed namespace
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_FALLBACK_BITMANIPULATION_H
+/* end file simdjson/fallback/bitmanipulation.h */
+/* including simdjson/fallback/stringparsing_defs.h: #include "simdjson/fallback/stringparsing_defs.h" */
+/* begin file simdjson/fallback/stringparsing_defs.h */
+#ifndef SIMDJSON_FALLBACK_STRINGPARSING_DEFS_H
+#define SIMDJSON_FALLBACK_STRINGPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace fallback {
+namespace {
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 1;
+  simdjson_inline backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() { return c == '"'; }
+  simdjson_inline bool has_backslash() { return c == '\\'; }
+  simdjson_inline int quote_index() { return c == '"' ? 0 : 1; }
+  simdjson_inline int backslash_index() { return c == '\\' ? 0 : 1; }
+
+  uint8_t c;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // store to dest unconditionally - we can overwrite the bits we don't like later
+  dst[0] = src[0];
+  return { src[0] };
+}
+
+
+struct escaping {
+  static constexpr uint32_t BYTES_PROCESSED = 1;
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits; }
+  simdjson_inline int escape_index() { return 0; }
+
+  bool escape_bits;
+}; // struct escaping
+
+
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  dst[0] = src[0];
+  return { (src[0] == '\\') || (src[0] == '"') || (src[0] < 32) };
+}
+
+} // unnamed namespace
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_FALLBACK_STRINGPARSING_DEFS_H
+/* end file simdjson/fallback/stringparsing_defs.h */
+/* including simdjson/fallback/numberparsing_defs.h: #include "simdjson/fallback/numberparsing_defs.h" */
+/* begin file simdjson/fallback/numberparsing_defs.h */
+#ifndef SIMDJSON_FALLBACK_NUMBERPARSING_DEFS_H
+#define SIMDJSON_FALLBACK_NUMBERPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+#ifdef JSON_TEST_NUMBERS // for unit testing
+void found_invalid_number(const uint8_t *buf);
+void found_integer(int64_t result, const uint8_t *buf);
+void found_unsigned_integer(uint64_t result, const uint8_t *buf);
+void found_float(double result, const uint8_t *buf);
+#endif
+
+namespace simdjson {
+namespace fallback {
+namespace numberparsing {
+
+// credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const char *chars) {
+  uint64_t val;
+  memcpy(&val, chars, sizeof(uint64_t));
+  val = (val & 0x0F0F0F0F0F0F0F0F) * 2561 >> 8;
+  val = (val & 0x00FF00FF00FF00FF) * 6553601 >> 16;
+  return uint32_t((val & 0x0000FFFF0000FFFF) * 42949672960001 >> 32);
+}
+
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  return parse_eight_digits_unrolled(reinterpret_cast<const char *>(chars));
+}
+
+#if SIMDJSON_IS_32BITS // _umul128 for x86, arm
+// this is a slow emulation routine for 32-bit
+//
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+  return x * (uint64_t)y;
+}
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+  uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
+  uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
+  uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
+  uint64_t adbc_carry = !!(adbc < ad);
+  uint64_t lo = bd + (adbc << 32);
+  *hi = __emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) +
+        (adbc_carry << 32) + !!(lo < bd);
+  return lo;
+}
+#endif
+
+/** @private */
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+#if SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+#if SIMDJSON_IS_ARM64
+  // ARM64 has native support for 64-bit multiplications, no need to emultate
+  answer.high = __umulh(value1, value2);
+  answer.low = value1 * value2;
+#else
+  answer.low = _umul128(value1, value2, &answer.high); // _umul128 not available on ARM64
+#endif // SIMDJSON_IS_ARM64
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+#endif
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace fallback
+} // namespace simdjson
+
+#ifndef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_IS_BIG_ENDIAN
+#define SIMDJSON_SWAR_NUMBER_PARSING 0
+#else
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+#endif
+#endif
+
+#endif // SIMDJSON_FALLBACK_NUMBERPARSING_DEFS_H
+/* end file simdjson/fallback/numberparsing_defs.h */
+/* end file simdjson/fallback/begin.h */
+/* including simdjson/generic/amalgamated.h for fallback: #include "simdjson/generic/amalgamated.h" */
+/* begin file simdjson/generic/amalgamated.h for fallback */
+#if defined(SIMDJSON_CONDITIONAL_INCLUDE) && !defined(SIMDJSON_GENERIC_DEPENDENCIES_H)
+#error simdjson/generic/dependencies.h must be included before simdjson/generic/amalgamated.h!
+#endif
+
+/* including simdjson/generic/base.h for fallback: #include "simdjson/generic/base.h" */
+/* begin file simdjson/generic/base.h for fallback */
+#ifndef SIMDJSON_GENERIC_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): // If we haven't got an implementation yet, we're in the editor, editing a generic file! Just */
+/* amalgamation skipped (editor-only): // use the most advanced one we can so the most possible stuff can be tested. */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation_detection.h" */
+/* amalgamation skipped (editor-only): #if SIMDJSON_IMPLEMENTATION_ICELAKE */
+/* amalgamation skipped (editor-only): #include "simdjson/icelake/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_HASWELL */
+/* amalgamation skipped (editor-only): #include "simdjson/haswell/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_WESTMERE */
+/* amalgamation skipped (editor-only): #include "simdjson/westmere/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_ARM64 */
+/* amalgamation skipped (editor-only): #include "simdjson/arm64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_PPC64 */
+/* amalgamation skipped (editor-only): #include "simdjson/ppc64/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LASX */
+/* amalgamation skipped (editor-only): #include "simdjson/lasx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_LSX */
+/* amalgamation skipped (editor-only): #include "simdjson/lsx/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_RVV_VLS */
+/* amalgamation skipped (editor-only): #include "simdjson/rvv-vls/begin.h" */
+/* amalgamation skipped (editor-only): #elif SIMDJSON_IMPLEMENTATION_FALLBACK */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/begin.h" */
+/* amalgamation skipped (editor-only): #else */
+/* amalgamation skipped (editor-only): #error "All possible implementations (including fallback) have been disabled! simdjson will not run." */
+/* amalgamation skipped (editor-only): #endif */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_IMPLEMENTATION */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace fallback {
+
+struct open_container;
+class dom_parser_implementation;
+
+/**
+ * The type of a JSON number
+ */
+enum class number_type {
+    floating_point_number=1, /// a binary64 number
+    signed_integer,          /// a signed integer that fits in a 64-bit word using two's complement
+    unsigned_integer,        /// a positive integer larger or equal to 1<<63
+    big_integer              /// a big integer that does not fit in a 64-bit word
+};
+
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_BASE_H
+/* end file simdjson/generic/base.h for fallback */
+/* including simdjson/generic/jsoncharutils.h for fallback: #include "simdjson/generic/jsoncharutils.h" */
+/* begin file simdjson/generic/jsoncharutils.h for fallback */
+#ifndef SIMDJSON_GENERIC_JSONCHARUTILS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_JSONCHARUTILS_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/jsoncharutils_tables.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace fallback {
+namespace {
+namespace jsoncharutils {
+
+// return non-zero if not a structural or whitespace char
+// zero otherwise
+simdjson_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace_negated[c];
+}
+
+simdjson_inline uint32_t is_structural_or_whitespace(uint8_t c) {
+  return internal::structural_or_whitespace[c];
+}
+
+// returns a value with the high 16 bits set if not valid
+// otherwise returns the conversion of the 4 hex digits at src into the bottom
+// 16 bits of the 32-bit return register
+//
+// see
+// https://lemire.me/blog/2019/04/17/parsing-short-hexadecimal-strings-efficiently/
+static inline uint32_t hex_to_u32_nocheck(
+    const uint8_t *src) { // strictly speaking, static inline is a C-ism
+  uint32_t v1 = internal::digit_to_val32[630 + src[0]];
+  uint32_t v2 = internal::digit_to_val32[420 + src[1]];
+  uint32_t v3 = internal::digit_to_val32[210 + src[2]];
+  uint32_t v4 = internal::digit_to_val32[0 + src[3]];
+  return v1 | v2 | v3 | v4;
+}
+
+// given a code point cp, writes to c
+// the utf-8 code, outputting the length in
+// bytes, if the length is zero, the code point
+// is invalid
+//
+// This can possibly be made faster using pdep
+// and clz and table lookups, but JSON documents
+// have few escaped code points, and the following
+// function looks cheap.
+//
+// Note: we assume that surrogates are treated separately
+//
+simdjson_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
+  if (cp <= 0x7F) {
+    c[0] = uint8_t(cp);
+    return 1; // ascii
+  }
+  if (cp <= 0x7FF) {
+    c[0] = uint8_t((cp >> 6) + 192);
+    c[1] = uint8_t((cp & 63) + 128);
+    return 2; // universal plane
+    //  Surrogates are treated elsewhere...
+    //} //else if (0xd800 <= cp && cp <= 0xdfff) {
+    //  return 0; // surrogates // could put assert here
+  } else if (cp <= 0xFFFF) {
+    c[0] = uint8_t((cp >> 12) + 224);
+    c[1] = uint8_t(((cp >> 6) & 63) + 128);
+    c[2] = uint8_t((cp & 63) + 128);
+    return 3;
+  } else if (cp <= 0x10FFFF) { // if you know you have a valid code point, this
+                               // is not needed
+    c[0] = uint8_t((cp >> 18) + 240);
+    c[1] = uint8_t(((cp >> 12) & 63) + 128);
+    c[2] = uint8_t(((cp >> 6) & 63) + 128);
+    c[3] = uint8_t((cp & 63) + 128);
+    return 4;
+  }
+  // will return 0 when the code point was too large.
+  return 0; // bad r
+}
+
+#if SIMDJSON_IS_32BITS // _umul128 for x86, arm
+// this is a slow emulation routine for 32-bit
+//
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+  return x * (uint64_t)y;
+}
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+  uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
+  uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
+  uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
+  uint64_t adbc_carry = !!(adbc < ad);
+  uint64_t lo = bd + (adbc << 32);
+  *hi = __emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) +
+        (adbc_carry << 32) + !!(lo < bd);
+  return lo;
+}
+#endif
+
+} // namespace jsoncharutils
+} // unnamed namespace
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_JSONCHARUTILS_H
+/* end file simdjson/generic/jsoncharutils.h for fallback */
+/* including simdjson/generic/atomparsing.h for fallback: #include "simdjson/generic/atomparsing.h" */
+/* begin file simdjson/generic/atomparsing.h for fallback */
+#ifndef SIMDJSON_GENERIC_ATOMPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_ATOMPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace fallback {
+namespace {
+/// @private
+namespace atomparsing {
+
+// The string_to_uint32 is exclusively used to map literal strings to 32-bit values.
+// We use memcpy instead of a pointer cast to avoid undefined behaviors since we cannot
+// be certain that the character pointer will be properly aligned.
+// You might think that using memcpy makes this function expensive, but you'd be wrong.
+// All decent optimizing compilers (GCC, clang, Visual Studio) will compile string_to_uint32("false");
+// to the compile-time constant 1936482662.
+simdjson_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
+
+
+// Again in str4ncmp we use a memcpy to avoid undefined behavior. The memcpy may appear expensive.
+// Yet all decent optimizing compilers will compile memcpy to a single instruction, just about.
+simdjson_warn_unused
+simdjson_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
+  uint32_t srcval; // we want to avoid unaligned 32-bit loads (undefined in C/C++)
+  static_assert(sizeof(uint32_t) <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be larger than 4 bytes");
+  std::memcpy(&srcval, src, sizeof(uint32_t));
+  return srcval ^ string_to_uint32(atom);
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src) {
+  return (str4ncmp(src, "true") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_true_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "true"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src) {
+  return (str4ncmp(src+1, "alse") | jsoncharutils::is_not_structural_or_whitespace(src[5])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
+  if (len > 5) { return is_valid_false_atom(src); }
+  else if (len == 5) { return !str4ncmp(src+1, "alse"); }
+  else { return false; }
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src) {
+  return (str4ncmp(src, "null") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
+}
+
+simdjson_warn_unused
+simdjson_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
+  if (len > 4) { return is_valid_null_atom(src); }
+  else if (len == 4) { return !str4ncmp(src, "null"); }
+  else { return false; }
+}
+
+} // namespace atomparsing
+} // unnamed namespace
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_ATOMPARSING_H
+/* end file simdjson/generic/atomparsing.h for fallback */
+/* including simdjson/generic/dom_parser_implementation.h for fallback: #include "simdjson/generic/dom_parser_implementation.h" */
+/* begin file simdjson/generic/dom_parser_implementation.h for fallback */
+#ifndef SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/dom_parser_implementation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace fallback {
+
+// expectation: sizeof(open_container) = 64/8.
+struct open_container {
+  uint32_t tape_index; // where, on the tape, does the scope ([,{) begins
+  uint32_t count; // how many elements in the scope
+}; // struct open_container
+
+static_assert(sizeof(open_container) == 64/8, "Open container must be 64 bits");
+
+class dom_parser_implementation final : public internal::dom_parser_implementation {
+public:
+  /** Tape location of each open { or [ */
+  std::unique_ptr<open_container[]> open_containers{};
+  /** Whether each open container is a [ or { */
+  std::unique_ptr<bool[]> is_array{};
+  /** Buffer passed to stage 1 */
+  const uint8_t *buf{};
+  /** Length passed to stage 1 */
+  size_t len{0};
+  /** Document passed to stage 2 */
+  dom::document *doc{};
+
+  inline dom_parser_implementation() noexcept;
+  inline dom_parser_implementation(dom_parser_implementation &&other) noexcept;
+  inline dom_parser_implementation &operator=(dom_parser_implementation &&other) noexcept;
+  dom_parser_implementation(const dom_parser_implementation &) = delete;
+  dom_parser_implementation &operator=(const dom_parser_implementation &) = delete;
+
+  simdjson_warn_unused error_code parse(const uint8_t *buf, size_t len, dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage1(const uint8_t *buf, size_t len, stage1_mode partial) noexcept final;
+  simdjson_warn_unused error_code stage2(dom::document &doc) noexcept final;
+  simdjson_warn_unused error_code stage2_next(dom::document &doc) noexcept final;
+  simdjson_warn_unused uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) const noexcept final;
+  simdjson_warn_unused uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept final;
+  inline simdjson_warn_unused error_code set_capacity(size_t capacity) noexcept final;
+  inline simdjson_warn_unused error_code set_max_depth(size_t max_depth) noexcept final;
+private:
+  simdjson_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
+
+};
+
+} // namespace fallback
+} // namespace simdjson
+
+namespace simdjson {
+namespace fallback {
+
+inline dom_parser_implementation::dom_parser_implementation() noexcept = default;
+inline dom_parser_implementation::dom_parser_implementation(dom_parser_implementation &&other) noexcept = default;
+inline dom_parser_implementation &dom_parser_implementation::operator=(dom_parser_implementation &&other) noexcept = default;
+
+// Leaving these here so they can be inlined if so desired
+inline simdjson_warn_unused error_code dom_parser_implementation::set_capacity(size_t capacity) noexcept {
+  if(capacity > SIMDJSON_MAXSIZE_BYTES) { return CAPACITY; }
+  // Stage 1 index output
+  size_t max_structures = SIMDJSON_ROUNDUP_N(capacity, 64) + 2 + 7;
+  structural_indexes.reset( new (std::nothrow) uint32_t[max_structures] );
+  if (!structural_indexes) { _capacity = 0; return MEMALLOC; }
+  structural_indexes[0] = 0;
+  n_structural_indexes = 0;
+
+  _capacity = capacity;
+  return SUCCESS;
+}
+
+inline simdjson_warn_unused error_code dom_parser_implementation::set_max_depth(size_t max_depth) noexcept {
+  // Stage 2 stacks
+  open_containers.reset(new (std::nothrow) open_container[max_depth]);
+  is_array.reset(new (std::nothrow) bool[max_depth]);
+  if (!is_array || !open_containers) { _max_depth = 0; return MEMALLOC; }
+
+  _max_depth = max_depth;
+  return SUCCESS;
+}
+
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_DOM_PARSER_IMPLEMENTATION_H
+/* end file simdjson/generic/dom_parser_implementation.h for fallback */
+/* including simdjson/generic/implementation_simdjson_result_base.h for fallback: #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base.h for fallback */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace fallback {
+
+// This is a near copy of include/error.h's implementation_simdjson_result_base, except it doesn't use std::pair
+// so we can avoid inlining errors
+// TODO reconcile these!
+/**
+ * The result of a simdjson operation that could fail.
+ *
+ * Gives the option of reading error codes, or throwing an exception by casting to the desired result.
+ *
+ * This is a base class for implementations that want to add functions to the result type for
+ * chaining.
+ *
+ * Override like:
+ *
+ *   struct simdjson_result<T> : public internal::implementation_simdjson_result_base<T> {
+ *     simdjson_result() noexcept : internal::implementation_simdjson_result_base<T>() {}
+ *     simdjson_result(error_code error) noexcept : internal::implementation_simdjson_result_base<T>(error) {}
+ *     simdjson_result(T &&value) noexcept : internal::implementation_simdjson_result_base<T>(std::forward(value)) {}
+ *     simdjson_result(T &&value, error_code error) noexcept : internal::implementation_simdjson_result_base<T>(value, error) {}
+ *     // Your extra methods here
+ *   }
+ *
+ * Then any method returning simdjson_result<T> will be chainable with your methods.
+ */
+template<typename T>
+struct implementation_simdjson_result_base {
+
+  /**
+   * Create a new empty result with error = UNINITIALIZED.
+   */
+  simdjson_inline implementation_simdjson_result_base() noexcept = default;
+
+  /**
+   * Create a new error result.
+   */
+  simdjson_inline implementation_simdjson_result_base(error_code error) noexcept;
+
+  /**
+   * Create a new successful result.
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value) noexcept;
+
+  /**
+   * Create a new result with both things (use if you don't want to branch when creating the result).
+   */
+  simdjson_inline implementation_simdjson_result_base(T &&value, error_code error) noexcept;
+
+  /**
+   * Move the value and the error to the provided variables.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   * @param error The variable to assign the error to. Set to SUCCESS if there is no error.
+   */
+  simdjson_inline void tie(T &value, error_code &error) && noexcept;
+
+  /**
+   * Move the value to the provided variable.
+   *
+   * @param value The variable to assign the value to. May not be set if there is an error.
+   */
+  simdjson_warn_unused simdjson_inline error_code get(T &value) && noexcept;
+
+  /**
+   * The error.
+   */
+  simdjson_warn_unused simdjson_inline error_code error() const noexcept;
+
+  /**
+   * Whether there is a value.
+   */
+  simdjson_warn_unused simdjson_inline bool has_value() const noexcept;
+
+#if SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T& operator*() &  noexcept(false);
+  simdjson_inline T&& operator*() &&  noexcept(false);
+  /**
+   * Arrow operator to access members of the contained value.
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T* operator->() noexcept(false);
+  simdjson_inline const T* operator->() const noexcept(false);
+
+  simdjson_inline T& value() & noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& value() && noexcept(false);
+
+  /**
+   * Take the result value (move it).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline T&& take_value() && noexcept(false);
+
+  /**
+   * Cast to the value (will throw on error).
+   *
+   * @throw simdjson_error if there was an error.
+   */
+  simdjson_inline operator T&&() && noexcept(false);
+
+
+#endif // SIMDJSON_EXCEPTIONS
+
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline const T& value_unsafe() const& noexcept;
+  /**
+   * Get the result value. This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T& value_unsafe() & noexcept;
+  /**
+   * Take the result value (move it). This function is safe if and only
+   * the error() method returns a value that evaluates to false.
+   */
+  simdjson_inline T&& value_unsafe() && noexcept;
+
+  using value_type = T;
+  using error_type = error_code;
+
+protected:
+  /** users should never directly access first and second. **/
+  T first{}; /** Users should never directly access 'first'. **/
+  error_code second{UNINITIALIZED}; /** Users should never directly access 'second'. **/
+}; // struct implementation_simdjson_result_base
+
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_H
+/* end file simdjson/generic/implementation_simdjson_result_base.h for fallback */
+/* including simdjson/generic/numberparsing.h for fallback: #include "simdjson/generic/numberparsing.h" */
+/* begin file simdjson/generic/numberparsing.h for fallback */
+#ifndef SIMDJSON_GENERIC_NUMBERPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_NUMBERPARSING_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/jsoncharutils.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <limits>
+#include <ostream>
+#include <cstring>
+
+namespace simdjson {
+namespace fallback {
+namespace numberparsing {
+
+#ifdef JSON_TEST_NUMBERS
+#define INVALID_NUMBER(SRC) (found_invalid_number((SRC)), NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (found_integer((VALUE), (SRC)), (WRITER).append_s64((VALUE)))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (found_unsigned_integer((VALUE), (SRC)), (WRITER).append_u64((VALUE)))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (found_float((VALUE), (SRC)), (WRITER).append_double((VALUE)))
+#define BIGINT_NUMBER(SRC) (found_invalid_number((SRC)), BIGINT_ERROR)
+#else
+#define INVALID_NUMBER(SRC) (NUMBER_ERROR)
+#define WRITE_INTEGER(VALUE, SRC, WRITER) (WRITER).append_s64((VALUE))
+#define WRITE_UNSIGNED(VALUE, SRC, WRITER) (WRITER).append_u64((VALUE))
+#define WRITE_DOUBLE(VALUE, SRC, WRITER) (WRITER).append_double((VALUE))
+#define BIGINT_NUMBER(SRC) (BIGINT_ERROR)
+#endif
+
+namespace {
+
+// Convert a mantissa, an exponent and a sign bit into an ieee64 double.
+// The real_exponent needs to be in [0, 2046] (technically real_exponent = 2047 would be acceptable).
+// The mantissa should be in [0,1<<53). The bit at index (1ULL << 52) while be zeroed.
+simdjson_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
+    double d;
+    mantissa &= ~(1ULL << 52);
+    mantissa |= real_exponent << 52;
+    mantissa |= ((static_cast<uint64_t>(negative)) << 63);
+    std::memcpy(&d, &mantissa, sizeof(d));
+    return d;
+}
+
+// Attempts to compute i * 10^(power) exactly; and if "negative" is
+// true, negate the result.
+// This function will only work in some cases, when it does not work, success is
+// set to false. This should work *most of the time* (like 99% of the time).
+// We assume that power is in the [smallest_power,
+// largest_power] interval: the caller is responsible for this check.
+simdjson_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
+  // we start with a fast path
+  // It was described in
+  // Clinger WD. How to read floating point numbers accurately.
+  // ACM SIGPLAN Notices. 1990
+#ifndef FLT_EVAL_METHOD
+#error "FLT_EVAL_METHOD should be defined, please include cfloat."
+#endif
+#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
+  // We cannot be certain that x/y is rounded to nearest.
+  if (0 <= power && power <= 22 && i <= 9007199254740991)
+#else
+  if (-22 <= power && power <= 22 && i <= 9007199254740991)
+#endif
+  {
+    // convert the integer into a double. This is lossless since
+    // 0 <= i <= 2^53 - 1.
+    d = double(i);
+    //
+    // The general idea is as follows.
+    // If 0 <= s < 2^53 and if 10^0 <= p <= 10^22 then
+    // 1) Both s and p can be represented exactly as 64-bit floating-point
+    // values
+    // (binary64).
+    // 2) Because s and p can be represented exactly as floating-point values,
+    // then s * p
+    // and s / p will produce correctly rounded values.
+    //
+    if (power < 0) {
+      d = d / simdjson::internal::power_of_ten[-power];
+    } else {
+      d = d * simdjson::internal::power_of_ten[power];
+    }
+    if (negative) {
+      d = -d;
+    }
+    return true;
+  }
+  // When 22 < power && power <  22 + 16, we could
+  // hope for another, secondary fast path.  It was
+  // described by David M. Gay in  "Correctly rounded
+  // binary-decimal and decimal-binary conversions." (1990)
+  // If you need to compute i * 10^(22 + x) for x < 16,
+  // first compute i * 10^x, if you know that result is exact
+  // (e.g., when i * 10^x < 2^53),
+  // then you can still proceed and do (i * 10^x) * 10^22.
+  // Is this worth your time?
+  // You need  22 < power *and* power <  22 + 16 *and* (i * 10^(x-22) < 2^53)
+  // for this second fast path to work.
+  // If you you have 22 < power *and* power <  22 + 16, and then you
+  // optimistically compute "i * 10^(x-22)", there is still a chance that you
+  // have wasted your time if i * 10^(x-22) >= 2^53. It makes the use cases of
+  // this optimization maybe less common than we would like. Source:
+  // http://www.exploringbinary.com/fast-path-decimal-to-floating-point-conversion/
+  // also used in RapidJSON: https://rapidjson.org/strtod_8h_source.html
+
+  // The fast path has now failed, so we are failing back on the slower path.
+
+  // In the slow path, we need to adjust i so that it is > 1<<63 which is always
+  // possible, except if i == 0, so we handle i == 0 separately.
+  if(i == 0) {
+    d = negative ? -0.0 : 0.0;
+    return true;
+  }
+
+
+  // The exponent is 1024 + 63 + power
+  //     + floor(log(5**power)/log(2)).
+  // The 1024 comes from the ieee64 standard.
+  // The 63 comes from the fact that we use a 64-bit word.
+  //
+  // Computing floor(log(5**power)/log(2)) could be
+  // slow. Instead we use a fast function.
+  //
+  // For power in (-400,350), we have that
+  // (((152170 + 65536) * power ) >> 16);
+  // is equal to
+  //  floor(log(5**power)/log(2)) + power when power >= 0
+  // and it is equal to
+  //  ceil(log(5**-power)/log(2)) + power when power < 0
+  //
+  // The 65536 is (1<<16) and corresponds to
+  // (65536 * power) >> 16 ---> power
+  //
+  // ((152170 * power ) >> 16) is equal to
+  // floor(log(5**power)/log(2))
+  //
+  // Note that this is not magic: 152170/(1<<16) is
+  // approximately equal to log(5)/log(2).
+  // The 1<<16 value is a power of two; we could use a
+  // larger power of 2 if we wanted to.
+  //
+  int64_t exponent = (((152170 + 65536) * power) >> 16) + 1024 + 63;
+
+
+  // We want the most significant bit of i to be 1. Shift if needed.
+  int lz = leading_zeroes(i);
+  i <<= lz;
+
+
+  // We are going to need to do some 64-bit arithmetic to get a precise product.
+  // We use a table lookup approach.
+  // It is safe because
+  // power >= smallest_power
+  // and power <= largest_power
+  // We recover the mantissa of the power, it has a leading 1. It is always
+  // rounded down.
+  //
+  // We want the most significant 64 bits of the product. We know
+  // this will be non-zero because the most significant bit of i is
+  // 1.
+  const uint32_t index = 2 * uint32_t(power - simdjson::internal::smallest_power);
+  // Optimization: It may be that materializing the index as a variable might confuse some compilers and prevent effective complex-addressing loads. (Done for code clarity.)
+  //
+  // The full_multiplication function computes the 128-bit product of two 64-bit words
+  // with a returned value of type value128 with a "low component" corresponding to the
+  // 64-bit least significant bits of the product and with a "high component" corresponding
+  // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index]);
+#else
+  simdjson::internal::value128 firstproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index]);
+#endif
+
+  // Both i and power_of_five_128[index] have their most significant bit set to 1 which
+  // implies that the either the most or the second most significant bit of the product
+  // is 1. We pack values in this manner for efficiency reasons: it maximizes the use
+  // we make of the product. It also makes it easy to reason about the product: there
+  // is 0 or 1 leading zero in the product.
+
+  // Unless the least significant 9 bits of the high (64-bit) part of the full
+  // product are all 1s, then we know that the most significant 55 bits are
+  // exact and no further work is needed. Having 55 bits is necessary because
+  // we need 53 bits for the mantissa but we have to have one rounding bit and
+  // we can waste a bit if the most significant bit of the product is zero.
+  if((firstproduct.high & 0x1FF) == 0x1FF) {
+    // We want to compute i * 5^q, but only care about the top 55 bits at most.
+    // Consider the scenario where q>=0. Then 5^q may not fit in 64-bits. Doing
+    // the full computation is wasteful. So we do what is called a "truncated
+    // multiplication".
+    // We take the most significant 64-bits, and we put them in
+    // power_of_five_128[index]. Usually, that's good enough to approximate i * 5^q
+    // to the desired approximation using one multiplication. Sometimes it does not suffice.
+    // Then we store the next most significant 64 bits in power_of_five_128[index + 1], and
+    // then we get a better approximation to i * 5^q.
+    //
+    // That's for when q>=0. The logic for q<0 is somewhat similar but it is somewhat
+    // more complicated.
+    //
+    // There is an extra layer of complexity in that we need more than 55 bits of
+    // accuracy in the round-to-even scenario.
+    //
+    // The full_multiplication function computes the 128-bit product of two 64-bit words
+    // with a returned value of type value128 with a "low component" corresponding to the
+    // 64-bit least significant bits of the product and with a "high component" corresponding
+    // to the 64-bit most significant bits of the product.
+#if SIMDJSON_STATIC_REFLECTION
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::powers_template<>::power_of_five_128[index + 1]);
+#else
+    simdjson::internal::value128 secondproduct = full_multiplication(i, simdjson::internal::power_of_five_128[index + 1]);
+#endif
+    firstproduct.low += secondproduct.high;
+    if(secondproduct.high > firstproduct.low) { firstproduct.high++; }
+    // As it has been proven by Noble Mushtak and Daniel Lemire in "Fast Number Parsing Without
+    // Fallback" (https://arxiv.org/abs/2212.06644), at this point we are sure that the product
+    // is sufficiently accurate, and more computation is not needed.
+  }
+  uint64_t lower = firstproduct.low;
+  uint64_t upper = firstproduct.high;
+  // The final mantissa should be 53 bits with a leading 1.
+  // We shift it so that it occupies 54 bits with a leading 1.
+  ///////
+  uint64_t upperbit = upper >> 63;
+  uint64_t mantissa = upper >> (upperbit + 9);
+  lz += int(1 ^ upperbit);
+
+  // Here we have mantissa < (1<<54).
+  int64_t real_exponent = exponent - lz;
+  if (simdjson_unlikely(real_exponent <= 0)) { // we have a subnormal?
+    // Here have that real_exponent <= 0 so -real_exponent >= 0
+    if(-real_exponent + 1 >= 64) { // if we have more than 64 bits below the minimum exponent, you have a zero for sure.
+      d = negative ? -0.0 : 0.0;
+      return true;
+    }
+    // next line is safe because -real_exponent + 1 < 0
+    mantissa >>= -real_exponent + 1;
+    // Thankfully, we can't have both "round-to-even" and subnormals because
+    // "round-to-even" only occurs for powers close to 0.
+    mantissa += (mantissa & 1); // round up
+    mantissa >>= 1;
+    // There is a weird scenario where we don't have a subnormal but just.
+    // Suppose we start with 2.2250738585072013e-308, we end up
+    // with 0x3fffffffffffff x 2^-1023-53 which is technically subnormal
+    // whereas 0x40000000000000 x 2^-1023-53  is normal. Now, we need to round
+    // up 0x3fffffffffffff x 2^-1023-53  and once we do, we are no longer
+    // subnormal, but we can only know this after rounding.
+    // So we only declare a subnormal if we are smaller than the threshold.
+    real_exponent = (mantissa < (uint64_t(1) << 52)) ? 0 : 1;
+    d = to_double(mantissa, real_exponent, negative);
+    return true;
+  }
+  // We have to round to even. The "to even" part
+  // is only a problem when we are right in between two floats
+  // which we guard against.
+  // If we have lots of trailing zeros, we may fall right between two
+  // floating-point values.
+  //
+  // The round-to-even cases take the form of a number 2m+1 which is in (2^53,2^54]
+  // times a power of two. That is, it is right between a number with binary significand
+  // m and another number with binary significand m+1; and it must be the case
+  // that it cannot be represented by a float itself.
+  //
+  // We must have that w * 10 ^q == (2m+1) * 2^p for some power of two 2^p.
+  // Recall that 10^q = 5^q * 2^q.
+  // When q >= 0, we must have that (2m+1) is divible by 5^q, so 5^q <= 2^54. We have that
+  //  5^23 <=  2^54 and it is the last power of five to qualify, so q <= 23.
+  // When q<0, we have  w  >=  (2m+1) x 5^{-q}.  We must have that w<2^{64} so
+  // (2m+1) x 5^{-q} < 2^{64}. We have that 2m+1>2^{53}. Hence, we must have
+  // 2^{53} x 5^{-q} < 2^{64}.
+  // Hence we have 5^{-q} < 2^{11}$ or q>= -4.
+  //
+  // We require lower <= 1 and not lower == 0 because we could not prove that
+  // that lower == 0 is implied; but we could prove that lower <= 1 is a necessary and sufficient test.
+  if (simdjson_unlikely((lower <= 1) && (power >= -4) && (power <= 23) && ((mantissa & 3) == 1))) {
+    if((mantissa  << (upperbit + 64 - 53 - 2)) ==  upper) {
+      mantissa &= ~1;             // flip it so that we do not round up
+    }
+  }
+
+  mantissa += mantissa & 1;
+  mantissa >>= 1;
+
+  // Here we have mantissa < (1<<53), unless there was an overflow
+  if (mantissa >= (1ULL << 53)) {
+    //////////
+    // This will happen when parsing values such as 7.2057594037927933e+16
+    ////////
+    mantissa = (1ULL << 52);
+    real_exponent++;
+  }
+  mantissa &= ~(1ULL << 52);
+  // we have to check that real_exponent is in range, otherwise we bail out
+  if (simdjson_unlikely(real_exponent > 2046)) {
+    // We have an infinite value!!! We could actually throw an error here if we could.
+    return false;
+  }
+  d = to_double(mantissa, real_exponent, negative);
+  return true;
+}
+
+// We call a fallback floating-point parser that might be slow. Note
+// it will accept JSON numbers, but the JSON spec. is more restrictive so
+// before you call parse_float_fallback, you need to have validated the input
+// string with the JSON grammar.
+// It will return an error (false) if the parsed number is infinite.
+// The string parsing itself always succeeds. We know that there is at least
+// one digit.
+static bool parse_float_fallback(const uint8_t *ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+static bool parse_float_fallback(const uint8_t *ptr, const uint8_t *end_ptr, double *outDouble) {
+  *outDouble = simdjson::internal::from_chars(reinterpret_cast<const char *>(ptr), reinterpret_cast<const char *>(end_ptr));
+  // We do not accept infinite values.
+
+  // Detecting finite values in a portable manner is ridiculously hard, ideally
+  // we would want to do:
+  // return !std::isfinite(*outDouble);
+  // but that mysteriously fails under legacy/old libc++ libraries, see
+  // https://github.com/simdjson/simdjson/issues/1286
+  //
+  // Therefore, fall back to this solution (the extra parens are there
+  // to handle that max may be a macro on windows).
+  return !(*outDouble > (std::numeric_limits<double>::max)() || *outDouble < std::numeric_limits<double>::lowest());
+}
+
+// check quickly whether the next 8 chars are made of digits
+// at a glance, it looks better than Mula's
+// http://0x80.pl/articles/swar-digits-validate.html
+simdjson_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
+  uint64_t val;
+  // this can read up to 7 bytes beyond the buffer size, but we require
+  // SIMDJSON_PADDING of padding
+  static_assert(7 <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be bigger than 7");
+  std::memcpy(&val, chars, 8);
+  // a branchy method might be faster:
+  // return (( val & 0xF0F0F0F0F0F0F0F0 ) == 0x3030303030303030)
+  //  && (( (val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0 ) ==
+  //  0x3030303030303030);
+  return (((val & 0xF0F0F0F0F0F0F0F0) |
+           (((val + 0x0606060606060606) & 0xF0F0F0F0F0F0F0F0) >> 4)) ==
+          0x3333333333333333);
+}
+
+template<typename I>
+SIMDJSON_NO_SANITIZE_UNDEFINED // We deliberately allow overflow here and check later
+simdjson_inline bool parse_digit(const uint8_t c, I &i) {
+  const uint8_t digit = static_cast<uint8_t>(c - '0');
+  if (digit > 9) {
+    return false;
+  }
+  // PERF NOTE: multiplication by 10 is cheaper than arbitrary integer multiplication
+  i = 10 * i + digit; // might overflow, we will handle the overflow later
+  return true;
+}
+
+simdjson_inline bool is_digit(const uint8_t c) {
+  return static_cast<uint8_t>(c - '0') <= 9;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_decimal_after_separator(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
+  // we continue with the fiction that we have an integer. If the
+  // floating point number is representable as x * 10^z for some integer
+  // z that fits in 53 bits, then we will be able to convert back the
+  // the integer into a float in a lossless manner.
+  const uint8_t *const first_after_period = p;
+
+#ifdef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_SWAR_NUMBER_PARSING
+  // this helps if we have lots of decimals!
+  // this turns out to be frequent enough.
+  if (is_made_of_eight_digits_fast(p)) {
+    i = i * 100000000 + parse_eight_digits_unrolled(p);
+    p += 8;
+  }
+#endif // SIMDJSON_SWAR_NUMBER_PARSING
+#endif // #ifdef SIMDJSON_SWAR_NUMBER_PARSING
+  // Unrolling the first digit makes a small difference on some implementations (e.g. westmere)
+  if (parse_digit(*p, i)) { ++p; }
+  while (parse_digit(*p, i)) { p++; }
+  exponent = first_after_period - p;
+  // Decimal without digits (123.) is illegal
+  if (exponent == 0) {
+    return INVALID_NUMBER(src);
+  }
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
+  // Exp Sign: -123.456e[-]78
+  bool neg_exp = ('-' == *p);
+  if (neg_exp || '+' == *p) { p++; } // Skip + as well
+
+  // Exponent: -123.456e-[78]
+  auto start_exp = p;
+  int64_t exp_number = 0;
+  while (parse_digit(*p, exp_number)) { ++p; }
+  // It is possible for parse_digit to overflow.
+  // In particular, it could overflow to INT64_MIN, and we cannot do - INT64_MIN.
+  // Thus we *must* check for possible overflow before we negate exp_number.
+
+  // Performance notes: it may seem like combining the two "simdjson_unlikely checks" below into
+  // a single simdjson_unlikely path would be faster. The reasoning is sound, but the compiler may
+  // not oblige and may, in fact, generate two distinct paths in any case. It might be
+  // possible to do uint64_t(p - start_exp - 1) >= 18 but it could end up trading off
+  // instructions for a simdjson_likely branch, an unconclusive gain.
+
+  // If there were no digits, it's an error.
+  if (simdjson_unlikely(p == start_exp)) {
+    return INVALID_NUMBER(src);
+  }
+  // We have a valid positive exponent in exp_number at this point, except that
+  // it may have overflowed.
+
+  // If there were more than 18 digits, we may have overflowed the integer. We have to do
+  // something!!!!
+  if (simdjson_unlikely(p > start_exp+18)) {
+    // Skip leading zeroes: 1e000000000000000000001 is technically valid and does not overflow
+    while (*start_exp == '0') { start_exp++; }
+    // 19 digits could overflow int64_t and is kind of absurd anyway. We don't
+    // support exponents smaller than -999,999,999,999,999,999 and bigger
+    // than 999,999,999,999,999,999.
+    // We can truncate.
+    // Note that 999999999999999999 is assuredly too large. The maximal ieee64 value before
+    // infinity is ~1.8e308. The smallest subnormal is ~5e-324. So, actually, we could
+    // truncate at 324.
+    // Note that there is no reason to fail per se at this point in time.
+    // E.g., 0e999999999999999999999 is a fine number.
+    if (p > start_exp+18) { exp_number = 999999999999999999; }
+  }
+  // At this point, we know that exp_number is a sane, positive, signed integer.
+  // It is <= 999,999,999,999,999,999. As long as 'exponent' is in
+  // [-8223372036854775808, 8223372036854775808], we won't overflow. Because 'exponent'
+  // is bounded in magnitude by the size of the JSON input, we are fine in this universe.
+  // To sum it up: the next line should never overflow.
+  exponent += (neg_exp ? -exp_number : exp_number);
+  return SUCCESS;
+}
+
+simdjson_inline bool check_if_integer(const uint8_t *const src, size_t max_length) {
+  const uint8_t *const srcend = src + max_length;
+  bool negative = (*src == '-'); // we can always read at least one character after the '-'
+  const uint8_t *p = src + uint8_t(negative);
+  if(p == srcend) { return false; }
+  if(*p == '0') {
+    ++p;
+    if(p == srcend) { return true; }
+    if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+    return true;
+  }
+  while(p != srcend && is_digit(*p)) { ++p; }
+  if(p == srcend) { return true; }
+  if(jsoncharutils::is_not_structural_or_whitespace(*p)) { return false; }
+  return true;
+}
+
+simdjson_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
+  // It is possible that the integer had an overflow.
+  // We have to handle the case where we have 0.0000somenumber.
+  const uint8_t *start = start_digits;
+  while ((*start == '0') || (*start == '.')) { ++start; }
+  // we over-decrement by one when there is a '.'
+  return digit_count - size_t(start - start_digits);
+}
+
+} // unnamed namespace
+
+/** @private */
+static error_code slow_float_parsing(simdjson_unused const uint8_t * src, double* answer) {
+  if (parse_float_fallback(src, answer)) {
+    return SUCCESS;
+  }
+  return INVALID_NUMBER(src);
+}
+
+/** @private */
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
+  // If we frequently had to deal with long strings of digits,
+  // we could extend our code by using a 128-bit integer instead
+  // of a 64-bit integer. However, this is uncommon in practice.
+  //
+  // 9999999999999999999 < 2**64 so we can accommodate 19 digits.
+  // If we have a decimal separator, then digit_count - 1 is the number of digits, but we
+  // may not have a decimal separator!
+  if (simdjson_unlikely(digit_count > 19 && significant_digits(start_digits, digit_count) > 19)) {
+    // Ok, chances are good that we had an overflow!
+    // this is almost never going to get called!!!
+    // we start anew, going slowly!!!
+    // This will happen in the following examples:
+    // 10000000000000000000000000000000000000000000e+308
+    // 3.1415926535897932384626433832795028841971693993751
+    //
+    // NOTE: We do not pass a reference to the to slow_float_parsing. If we passed our writer
+    // reference to it, it would force it to be stored in memory, preventing the compiler from
+    // picking it apart and putting into registers. i.e. if we pass it as reference,
+    // it gets slow.
+    double d;
+    error_code error = slow_float_parsing(src, &d);
+    writer.append_double(d);
+    return error;
+  }
+  // NOTE: it's weird that the simdjson_unlikely() only wraps half the if, but it seems to get slower any other
+  // way we've tried: https://github.com/simdjson/simdjson/pull/990#discussion_r448497331
+  // To future reader: we'd love if someone found a better way, or at least could explain this result!
+  if (simdjson_unlikely(exponent < simdjson::internal::smallest_power) || (exponent > simdjson::internal::largest_power)) {
+    //
+    // Important: smallest_power is such that it leads to a zero value.
+    // Observe that 18446744073709551615e-343 == 0, i.e. (2**64 - 1) e -343 is zero
+    // so something x 10^-343 goes to zero, but not so with  something x 10^-342.
+    static_assert(simdjson::internal::smallest_power <= -342, "smallest_power is not small enough");
+    //
+    if((exponent < simdjson::internal::smallest_power) || (i == 0)) {
+      // E.g. Parse "-0.0e-999" into the same value as "-0.0". See https://en.wikipedia.org/wiki/Signed_zero
+      WRITE_DOUBLE(negative ? -0.0 : 0.0, src, writer);
+      return SUCCESS;
+    } else { // (exponent > largest_power) and (i != 0)
+      // We have, for sure, an infinite value and simdjson refuses to parse infinite values.
+      return INVALID_NUMBER(src);
+    }
+  }
+  double d;
+  if (!compute_float_64(exponent, i, negative, d)) {
+    // we are almost never going to get here.
+    if (!parse_float_fallback(src, &d)) { return INVALID_NUMBER(src); }
+  }
+  WRITE_DOUBLE(d, src, writer);
+  return SUCCESS;
+}
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer);
+
+// for performance analysis, it is sometimes  useful to skip parsing
+#ifdef SIMDJSON_SKIPNUMBERPARSING
+
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const, W &writer) {
+  writer.append_s64(0);        // always write zero
+  return SUCCESS;              // always succeeds
+}
+
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept { return number_type::signed_integer; }
+#else
+
+// parse the number at src
+// define JSON_TEST_NUMBERS for unit testing
+//
+// It is assumed that the number is followed by a structural ({,},],[) character
+// or a white space character. If that is not the case (e.g., when the JSON
+// document is made of a single number), then it is necessary to copy the
+// content and append a space before calling this function.
+//
+// Our objective is accurate parsing (ULP of 0) at high speed.
+template<typename W>
+simdjson_warn_unused simdjson_inline error_code parse_number(const uint8_t *const src, W &writer) {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  if (digit_count == 0 || ('0' == *start_digits && digit_count > 1)) { return INVALID_NUMBER(src); }
+
+  //
+  // Handle floats if there is a . or e (or both)
+  //
+  int64_t exponent = 0;
+  bool is_float = false;
+  if ('.' == *p) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_decimal_after_separator(src, p, i, exponent) );
+    digit_count = int(p - start_digits); // used later to guard against overflows
+  }
+  if (('e' == *p) || ('E' == *p)) {
+    is_float = true;
+    ++p;
+    SIMDJSON_TRY( parse_exponent(src, p, exponent) );
+  }
+  if (is_float) {
+    const bool dirty_end = jsoncharutils::is_not_structural_or_whitespace(*p);
+    SIMDJSON_TRY( write_float(src, negative, i, start_digits, digit_count, exponent, writer) );
+    if (dirty_end) { return INVALID_NUMBER(src); }
+    return SUCCESS;
+  }
+
+  // The longest negative 64-bit number is 19 digits.
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  size_t longest_digit_count = negative ? 19 : 20;
+  if (digit_count > longest_digit_count) { return BIGINT_NUMBER(src); }
+  if (digit_count == longest_digit_count) {
+    if (negative) {
+      // Anything negative above INT64_MAX+1 is invalid
+      if (i > uint64_t(INT64_MAX)+1) { return BIGINT_NUMBER(src);  }
+      WRITE_INTEGER(~i+1, src, writer);
+      if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+      return SUCCESS;
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    }  else if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INVALID_NUMBER(src); }
+  }
+
+  // Write unsigned if it does not fit in a signed integer.
+  if (i > uint64_t(INT64_MAX)) {
+    WRITE_UNSIGNED(i, src, writer);
+  } else {
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+    if(i == 0 && negative) {
+      // We have to write -0.0 instead of 0
+      WRITE_DOUBLE(-0.0, src, writer);
+    } else {
+      WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+    }
+#else
+  WRITE_INTEGER(negative ? (~i+1) : i, src, writer);
+#endif
+  }
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return INVALID_NUMBER(src); }
+  return SUCCESS;
+}
+
+// Inlineable functions
+namespace {
+
+// This table can be used to characterize the final character of an integer
+// string. For JSON structural character and allowable white space characters,
+// we return SUCCESS. For 'e', '.' and 'E', we return INCORRECT_TYPE. Otherwise
+// we return NUMBER_ERROR.
+// Optimization note: we could easily reduce the size of the table by half (to 128)
+// at the cost of an extra branch.
+// Optimization note: we want the values to use at most 8 bits (not, e.g., 32 bits):
+static_assert(error_code(uint8_t(NUMBER_ERROR))== NUMBER_ERROR, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(SUCCESS))== SUCCESS, "bad NUMBER_ERROR cast");
+static_assert(error_code(uint8_t(INCORRECT_TYPE))== INCORRECT_TYPE, "bad NUMBER_ERROR cast");
+
+const uint8_t integer_string_finisher[256] = {
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   SUCCESS,      NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, SUCCESS,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, INCORRECT_TYPE,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, SUCCESS,        NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, INCORRECT_TYPE, NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, SUCCESS,      NUMBER_ERROR,
+    SUCCESS,      NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR, NUMBER_ERROR,   NUMBER_ERROR, NUMBER_ERROR, NUMBER_ERROR,
+    NUMBER_ERROR};
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  const uint8_t *p = src;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if ((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    if (src[0] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from 0 to 18,446,744,073,709,551,615
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
+  const uint8_t *p = src + 1;
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // The longest positive 64-bit number is 20 digits.
+  // We do it this way so we don't trigger this branch unless we must.
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > 20))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > 20)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  if (digit_count == 20) {
+    // Positive overflow check:
+    // - A 20 digit number starting with 2-9 is overflow, because 18,446,744,073,709,551,615 is the
+    //   biggest uint64_t.
+    // - A 20 digit number starting with 1 is overflow if it is less than INT64_MAX.
+    //   If we got here, it's a 20 digit number starting with the digit "1".
+    // - If a 20 digit number starting with 1 overflowed (i*10+digit), the result will be smaller
+    //   than 1,553,255,926,290,448,384.
+    // - That is smaller than the smallest possible 20-digit number the user could write:
+    //   10,000,000,000,000,000,000.
+    // - Therefore, if the number is positive and lower than that, it's overflow.
+    // - The value we are looking at is less than or equal to INT64_MAX.
+    //
+    // Note: we use src[1] and not src[0] because src[0] is the quote character in this
+    // instance.
+    if (src[1] != uint8_t('1') || i <= uint64_t(INT64_MAX)) { return INCORRECT_TYPE; }
+  }
+
+  return i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while (parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+  //
+  // Check for minus sign
+  //
+  if(src == src_end) { return NUMBER_ERROR; }
+  bool negative = (*src == '-');
+  const uint8_t *p = src + uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = p;
+  uint64_t i = 0;
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(p - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
+  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if((p != src_end) && integer_string_finisher[*p] != SUCCESS) { return error_code(integer_string_finisher[*p]); }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+// Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
+  const uint8_t *const start_digits = src;
+  uint64_t i = 0;
+  while (parse_digit(*src, i)) { src++; }
+
+  // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
+  // Optimization note: size_t is expected to be unsigned.
+  size_t digit_count = size_t(src - start_digits);
+  // We go from
+  // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
+  // so we can never represent numbers that have more than 19 digits.
+  size_t longest_digit_count = 19;
+  // Optimization note: the compiler can probably merge
+  // ((digit_count == 0) || (digit_count > longest_digit_count))
+  // into a single  branch since digit_count is unsigned.
+  if ((digit_count == 0) || (digit_count > longest_digit_count)) { return INCORRECT_TYPE; }
+  // Here digit_count > 0.
+  if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
+  // We can do the following...
+  // if (!jsoncharutils::is_structural_or_whitespace(*src)) {
+  //  return (*src == '.' || *src == 'e' || *src == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // }
+  // as a single table lookup:
+  if(*src != '"') { return NUMBER_ERROR; }
+  // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
+  // Performance note: This check is only needed when digit_count == longest_digit_count but it is
+  // so cheap that we might as well always make it.
+  if(i > uint64_t(INT64_MAX) + uint64_t(negative)) { return INCORRECT_TYPE; }
+  return negative ? (~i+1) : i;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept {
+  return (*src == '-');
+}
+
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) { return true; }
+  return false;
+}
+
+simdjson_unused simdjson_inline simdjson_result<number_type> get_number_type(const uint8_t * src) noexcept {
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+  const uint8_t *p = src;
+  while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
+  size_t digit_count = size_t(p - src);
+  if ( p == src ) { return NUMBER_ERROR; }
+  if (jsoncharutils::is_structural_or_whitespace(*p)) {
+    static const uint8_t * smaller_big_integer = reinterpret_cast<const uint8_t *>("9223372036854775808");
+    // We have an integer.
+    if(simdjson_unlikely(digit_count > 20)) {
+      return number_type::big_integer;
+    }
+    // If the number is negative and valid, it must be a signed integer.
+    if(negative) {
+      if (simdjson_unlikely(digit_count > 19)) return number_type::big_integer;
+      if (simdjson_unlikely(digit_count == 19 && memcmp(src, smaller_big_integer, 19) > 0)) {
+        return number_type::big_integer;
+      }
+#if SIMDJSON_MINUS_ZERO_AS_FLOAT
+      if(digit_count == 1 && src[0] == '0') {
+        // We have to write -0.0 instead of 0
+        return number_type::floating_point_number;
+      }
+#endif
+      return number_type::signed_integer;
+    }
+    // Let us check if we have a big integer (>=2**64).
+    static const uint8_t * two_to_sixtyfour = reinterpret_cast<const uint8_t *>("18446744073709551616");
+    if((digit_count > 20) || (digit_count == 20 && memcmp(src, two_to_sixtyfour, 20) >= 0)) {
+      return number_type::big_integer;
+    }
+    // The number is positive and smaller than 18446744073709551616 (or 2**64).
+    // We want values larger or equal to 9223372036854775808 to be unsigned
+    // integers, and the other values to be signed integers.
+    if((digit_count == 20) || (digit_count >= 19 && memcmp(src, smaller_big_integer, 19) >= 0)) {
+      return number_type::unsigned_integer;
+    }
+    return number_type::signed_integer;
+  }
+  // Hopefully, we have 'e' or 'E' or '.'.
+  return number_type::floating_point_number;
+}
+
+// Never read at src_end or beyond
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
+  if(src == src_end) { return NUMBER_ERROR; }
+  //
+  // Check for minus sign
+  //
+  bool negative = (*src == '-');
+  src += uint8_t(negative);
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  if(p == src_end) { return NUMBER_ERROR; }
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while ((p != src_end) && parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely((p != src_end) && (*p == '.'))) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if ((p == src_end) || !parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while ((p != src_end) && parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = start_digits-src > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if ((p != src_end) && (*p == 'e' || *p == 'E')) {
+    p++;
+    if(p == src_end) { return NUMBER_ERROR; }
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while ((p != src_end) && parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if ((p != src_end) && jsoncharutils::is_not_structural_or_whitespace(*p)) { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), src_end, &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
+  //
+  // Check for minus sign
+  //
+  bool negative = (*(src + 1) == '-');
+  src += uint8_t(negative) + 1;
+
+  //
+  // Parse the integer part.
+  //
+  uint64_t i = 0;
+  const uint8_t *p = src;
+  p += parse_digit(*p, i);
+  bool leading_zero = (i == 0);
+  while (parse_digit(*p, i)) { p++; }
+  // no integer digits, or 0123 (zero must be solo)
+  if ( p == src ) { return INCORRECT_TYPE; }
+  if ( (leading_zero && p != src+1)) { return NUMBER_ERROR; }
+
+  //
+  // Parse the decimal part.
+  //
+  int64_t exponent = 0;
+  bool overflow;
+  if (simdjson_likely(*p == '.')) {
+    p++;
+    const uint8_t *start_decimal_digits = p;
+    if (!parse_digit(*p, i)) { return NUMBER_ERROR; } // no decimal digits
+    p++;
+    while (parse_digit(*p, i)) { p++; }
+    exponent = -(p - start_decimal_digits);
+
+    // Overflow check. More than 19 digits (minus the decimal) may be overflow.
+    overflow = p-src-1 > 19;
+    if (simdjson_unlikely(overflow && leading_zero)) {
+      // Skip leading 0.00000 and see if it still overflows
+      const uint8_t *start_digits = src + 2;
+      while (*start_digits == '0') { start_digits++; }
+      overflow = p-start_digits > 19;
+    }
+  } else {
+    overflow = p-src > 19;
+  }
+
+  //
+  // Parse the exponent
+  //
+  if (*p == 'e' || *p == 'E') {
+    p++;
+    bool exp_neg = *p == '-';
+    p += exp_neg || *p == '+';
+
+    uint64_t exp = 0;
+    const uint8_t *start_exp_digits = p;
+    while (parse_digit(*p, exp)) { p++; }
+    // no exp digits, or 20+ exp digits
+    if (p-start_exp_digits == 0 || p-start_exp_digits > 19) { return NUMBER_ERROR; }
+
+    exponent += exp_neg ? 0-exp : exp;
+  }
+
+  if (*p != '"') { return NUMBER_ERROR; }
+
+  overflow = overflow || exponent < simdjson::internal::smallest_power || exponent > simdjson::internal::largest_power;
+
+  //
+  // Assemble (or slow-parse) the float
+  //
+  double d;
+  if (simdjson_likely(!overflow)) {
+    if (compute_float_64(exponent, i, negative, d)) { return d; }
+  }
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
+    return NUMBER_ERROR;
+  }
+  return d;
+}
+
+} // unnamed namespace
+#endif // SIMDJSON_SKIPNUMBERPARSING
+
+} // namespace numberparsing
+
+inline std::ostream& operator<<(std::ostream& out, number_type type) noexcept {
+    switch (type) {
+        case number_type::signed_integer: out << "integer in [-9223372036854775808,9223372036854775808)"; break;
+        case number_type::unsigned_integer: out << "unsigned integer in [9223372036854775808,18446744073709551616)"; break;
+        case number_type::floating_point_number: out << "floating-point number (binary64)"; break;
+        case number_type::big_integer: out << "big integer"; break;
+        default: SIMDJSON_UNREACHABLE();
+    }
+    return out;
+}
+
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_NUMBERPARSING_H
+/* end file simdjson/generic/numberparsing.h for fallback */
+
+/* including simdjson/generic/implementation_simdjson_result_base-inl.h for fallback: #include "simdjson/generic/implementation_simdjson_result_base-inl.h" */
+/* begin file simdjson/generic/implementation_simdjson_result_base-inl.h for fallback */
+#ifndef SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/generic/implementation_simdjson_result_base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace fallback {
+
+//
+// internal::implementation_simdjson_result_base<T> inline implementation
+//
+
+template<typename T>
+simdjson_inline void implementation_simdjson_result_base<T>::tie(T &value, error_code &error) && noexcept {
+  error = this->second;
+  if (!error) {
+    value = std::forward<implementation_simdjson_result_base<T>>(*this).first;
+  }
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::get(T &value) && noexcept {
+  error_code error;
+  std::forward<implementation_simdjson_result_base<T>>(*this).tie(value, error);
+  return error;
+}
+
+template<typename T>
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::error() const noexcept {
+  return this->second;
+}
+
+
+template<typename T>
+simdjson_warn_unused simdjson_inline bool implementation_simdjson_result_base<T>::has_value() const noexcept {
+  return this->error() == SUCCESS;
+}
+
+#if SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::operator*() &  noexcept(false) {
+  return this->value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::operator*() &&  noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).value();
+}
+
+template<typename T>
+simdjson_inline T* implementation_simdjson_result_base<T>::operator->() noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+
+template<typename T>
+simdjson_inline const T* implementation_simdjson_result_base<T>::operator->() const noexcept(false) {
+  if (this->error()) { throw simdjson_error(this->error()); }
+  return &this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value() & noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::take_value() && noexcept(false) {
+  if (error()) { throw simdjson_error(error()); }
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::operator T&&() && noexcept(false) {
+  return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
+}
+
+#endif // SIMDJSON_EXCEPTIONS
+
+template<typename T>
+simdjson_inline const T& implementation_simdjson_result_base<T>::value_unsafe() const& noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T& implementation_simdjson_result_base<T>::value_unsafe() & noexcept {
+  return this->first;
+}
+
+template<typename T>
+simdjson_inline T&& implementation_simdjson_result_base<T>::value_unsafe() && noexcept {
+  return std::forward<T>(this->first);
+}
+
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value, error_code error) noexcept
+    : first{std::forward<T>(value)}, second{error} {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(error_code error) noexcept
+    : implementation_simdjson_result_base(T{}, error) {}
+template<typename T>
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value) noexcept
+    : implementation_simdjson_result_base(std::forward<T>(value), SUCCESS) {}
+
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_GENERIC_IMPLEMENTATION_SIMDJSON_RESULT_BASE_INL_H
+/* end file simdjson/generic/implementation_simdjson_result_base-inl.h for fallback */
+/* end file simdjson/generic/amalgamated.h for fallback */
+/* including simdjson/fallback/end.h: #include "simdjson/fallback/end.h" */
+/* begin file simdjson/fallback/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+/* undefining SIMDJSON_IMPLEMENTATION from "fallback" */
+#undef SIMDJSON_IMPLEMENTATION
+/* end file simdjson/fallback/end.h */
+
+#endif // SIMDJSON_FALLBACK_H
+/* end file simdjson/fallback.h */
+/* including simdjson/fallback/implementation.h: #include <simdjson/fallback/implementation.h> */
+/* begin file simdjson/fallback/implementation.h */
+#ifndef SIMDJSON_FALLBACK_IMPLEMENTATION_H
+#define SIMDJSON_FALLBACK_IMPLEMENTATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/implementation.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace fallback {
+
+/**
+ * @private
+ */
+class implementation final : public simdjson::implementation {
+public:
+  simdjson_inline implementation() : simdjson::implementation(
+      "fallback",
+      "Generic fallback implementation",
+      0
+  ) {}
+  simdjson_warn_unused error_code create_dom_parser_implementation(
+    size_t capacity,
+    size_t max_length,
+    std::unique_ptr<simdjson::internal::dom_parser_implementation>& dst
+  ) const noexcept final;
+  simdjson_warn_unused error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept final;
+  simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
+};
+
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_FALLBACK_IMPLEMENTATION_H
+/* end file simdjson/fallback/implementation.h */
+
+/* including simdjson/fallback/begin.h: #include <simdjson/fallback/begin.h> */
+/* begin file simdjson/fallback/begin.h */
+/* defining SIMDJSON_IMPLEMENTATION to "fallback" */
+#define SIMDJSON_IMPLEMENTATION fallback
+/* including simdjson/fallback/base.h: #include "simdjson/fallback/base.h" */
+/* begin file simdjson/fallback/base.h */
+#ifndef SIMDJSON_FALLBACK_BASE_H
+#define SIMDJSON_FALLBACK_BASE_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+/**
+ * Fallback implementation (runs on any machine).
+ */
+namespace fallback {
+
+class implementation;
+
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_FALLBACK_BASE_H
+/* end file simdjson/fallback/base.h */
+/* including simdjson/fallback/bitmanipulation.h: #include "simdjson/fallback/bitmanipulation.h" */
+/* begin file simdjson/fallback/bitmanipulation.h */
+#ifndef SIMDJSON_FALLBACK_BITMANIPULATION_H
+#define SIMDJSON_FALLBACK_BITMANIPULATION_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace fallback {
+namespace {
+
+#if defined(_MSC_VER) && !defined(_M_ARM64) && !defined(_M_X64)
+static inline unsigned char _BitScanForward64(unsigned long* ret, uint64_t x) {
+  unsigned long x0 = (unsigned long)x, top, bottom;
+  _BitScanForward(&top, (unsigned long)(x >> 32));
+  _BitScanForward(&bottom, x0);
+  *ret = x0 ? bottom : 32 + top;
+  return x != 0;
+}
+static unsigned char _BitScanReverse64(unsigned long* ret, uint64_t x) {
+  unsigned long x1 = (unsigned long)(x >> 32), top, bottom;
+  _BitScanReverse(&top, x1);
+  _BitScanReverse(&bottom, (unsigned long)x);
+  *ret = x1 ? top + 32 : bottom;
+  return x != 0;
+}
+#endif
+
+/* result might be undefined when input_num is zero */
+simdjson_inline int leading_zeroes(uint64_t input_num) {
+#ifdef _MSC_VER
+  unsigned long leading_zero = 0;
+  // Search the mask data from most significant bit (MSB)
+  // to least significant bit (LSB) for a set bit (1).
+  if (_BitScanReverse64(&leading_zero, input_num))
+    return (int)(63 - leading_zero);
+  else
+    return 64;
+#else
+  return __builtin_clzll(input_num);
+#endif// _MSC_VER
+}
+
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
+#ifdef _MSC_VER
+  unsigned long trailing_zero = 0;
+  // Search the mask data from least significant bit (LSB)
+  // to most significant bit (MSB) for a set bit (1).
+  if (_BitScanForward64(&trailing_zero, input_num))
+    return (int)trailing_zero;
+  else    return 64;
+#else
+  return __builtin_ctzll(input_num);
+#endif// _MSC_VER
+}
+
+} // unnamed namespace
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_FALLBACK_BITMANIPULATION_H
+/* end file simdjson/fallback/bitmanipulation.h */
+/* including simdjson/fallback/stringparsing_defs.h: #include "simdjson/fallback/stringparsing_defs.h" */
+/* begin file simdjson/fallback/stringparsing_defs.h */
+#ifndef SIMDJSON_FALLBACK_STRINGPARSING_DEFS_H
+#define SIMDJSON_FALLBACK_STRINGPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace fallback {
+namespace {
+
+// Holds backslashes and quotes locations.
+struct backslash_and_quote {
+public:
+  static constexpr uint32_t BYTES_PROCESSED = 1;
+  simdjson_inline backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_quote_first() { return c == '"'; }
+  simdjson_inline bool has_backslash() { return c == '\\'; }
+  simdjson_inline int quote_index() { return c == '"' ? 0 : 1; }
+  simdjson_inline int backslash_index() { return c == '\\' ? 0 : 1; }
+
+  uint8_t c;
+}; // struct backslash_and_quote
+
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  // store to dest unconditionally - we can overwrite the bits we don't like later
+  dst[0] = src[0];
+  return { src[0] };
+}
+
+
+struct escaping {
+  static constexpr uint32_t BYTES_PROCESSED = 1;
+  simdjson_inline static escaping copy_and_find(const uint8_t *src, uint8_t *dst);
+
+  simdjson_inline bool has_escape() { return escape_bits; }
+  simdjson_inline int escape_index() { return 0; }
+
+  bool escape_bits;
+}; // struct escaping
+
+
+
+simdjson_inline escaping escaping::copy_and_find(const uint8_t *src, uint8_t *dst) {
+  dst[0] = src[0];
+  return { (src[0] == '\\') || (src[0] == '"') || (src[0] < 32) };
+}
+
+} // unnamed namespace
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_FALLBACK_STRINGPARSING_DEFS_H
+/* end file simdjson/fallback/stringparsing_defs.h */
+/* including simdjson/fallback/numberparsing_defs.h: #include "simdjson/fallback/numberparsing_defs.h" */
+/* begin file simdjson/fallback/numberparsing_defs.h */
+#ifndef SIMDJSON_FALLBACK_NUMBERPARSING_DEFS_H
+#define SIMDJSON_FALLBACK_NUMBERPARSING_DEFS_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/base.h" */
+/* amalgamation skipped (editor-only): #include "simdjson/internal/numberparsing_tables.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+#ifdef JSON_TEST_NUMBERS // for unit testing
+void found_invalid_number(const uint8_t *buf);
+void found_integer(int64_t result, const uint8_t *buf);
+void found_unsigned_integer(uint64_t result, const uint8_t *buf);
+void found_float(double result, const uint8_t *buf);
+#endif
+
+namespace simdjson {
+namespace fallback {
+namespace numberparsing {
+
+// credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const char *chars) {
+  uint64_t val;
+  memcpy(&val, chars, sizeof(uint64_t));
+  val = (val & 0x0F0F0F0F0F0F0F0F) * 2561 >> 8;
+  val = (val & 0x00FF00FF00FF00FF) * 6553601 >> 16;
+  return uint32_t((val & 0x0000FFFF0000FFFF) * 42949672960001 >> 32);
+}
+
+/** @private */
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+  return parse_eight_digits_unrolled(reinterpret_cast<const char *>(chars));
+}
+
+#if SIMDJSON_IS_32BITS // _umul128 for x86, arm
+// this is a slow emulation routine for 32-bit
+//
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+  return x * (uint64_t)y;
+}
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+  uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
+  uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
+  uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
+  uint64_t adbc_carry = !!(adbc < ad);
+  uint64_t lo = bd + (adbc << 32);
+  *hi = __emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) +
+        (adbc_carry << 32) + !!(lo < bd);
+  return lo;
+}
+#endif
+
+/** @private */
+simdjson_inline internal::value128 full_multiplication(uint64_t value1, uint64_t value2) {
+  internal::value128 answer;
+#if SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+#if SIMDJSON_IS_ARM64
+  // ARM64 has native support for 64-bit multiplications, no need to emultate
+  answer.high = __umulh(value1, value2);
+  answer.low = value1 * value2;
+#else
+  answer.low = _umul128(value1, value2, &answer.high); // _umul128 not available on ARM64
+#endif // SIMDJSON_IS_ARM64
+#else // SIMDJSON_REGULAR_VISUAL_STUDIO || SIMDJSON_IS_32BITS
+  __uint128_t r = (static_cast<__uint128_t>(value1)) * value2;
+  answer.low = uint64_t(r);
+  answer.high = uint64_t(r >> 64);
+#endif
+  return answer;
+}
+
+} // namespace numberparsing
+} // namespace fallback
+} // namespace simdjson
+
+#ifndef SIMDJSON_SWAR_NUMBER_PARSING
+#if SIMDJSON_IS_BIG_ENDIAN
+#define SIMDJSON_SWAR_NUMBER_PARSING 0
+#else
+#define SIMDJSON_SWAR_NUMBER_PARSING 1
+#endif
+#endif
+
+#endif // SIMDJSON_FALLBACK_NUMBERPARSING_DEFS_H
+/* end file simdjson/fallback/numberparsing_defs.h */
+/* end file simdjson/fallback/begin.h */
+/* including generic/stage1/find_next_document_index.h for fallback: #include <generic/stage1/find_next_document_index.h> */
+/* begin file generic/stage1/find_next_document_index.h for fallback */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H */
+/* amalgamation skipped (editor-only): #include <generic/stage1/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace fallback {
+namespace {
+namespace stage1 {
+
+/**
+  * This algorithm is used to quickly identify the last structural position that
+  * makes up a complete document.
+  *
+  * It does this by going backwards and finding the last *document boundary* (a
+  * place where one value follows another without a comma between them). If the
+  * last document (the characters after the boundary) has an equal number of
+  * start and end brackets, it is considered complete.
+  *
+  * Simply put, we iterate over the structural characters, starting from
+  * the end. We consider that we found the end of a JSON document when the
+  * first element of the pair is NOT one of these characters: '{' '[' ':' ','
+  * and when the second element is NOT one of these characters: '}' ']' ':' ','.
+  *
+  * This simple comparison works most of the time, but it does not cover cases
+  * where the batch's structural indexes contain a perfect amount of documents.
+  * In such a case, we do not have access to the structural index which follows
+  * the last document, therefore, we do not have access to the second element in
+  * the pair, and that means we cannot identify the last document. To fix this
+  * issue, we keep a count of the open and closed curly/square braces we found
+  * while searching for the pair. When we find a pair AND the count of open and
+  * closed curly/square braces is the same, we know that we just passed a
+  * complete document, therefore the last json buffer location is the end of the
+  * batch.
+  */
+simdjson_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
+  // Variant: do not count separately, just figure out depth
+  if(parser.n_structural_indexes == 0) { return 0; }
+  auto arr_cnt = 0;
+  auto obj_cnt = 0;
+  for (auto i = parser.n_structural_indexes - 1; i > 0; i--) {
+    auto idxb = parser.structural_indexes[i];
+    switch (parser.buf[idxb]) {
+    case ':':
+    case ',':
+      continue;
+    case '}':
+      obj_cnt--;
+      continue;
+    case ']':
+      arr_cnt--;
+      continue;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+    }
+    auto idxa = parser.structural_indexes[i - 1];
+    switch (parser.buf[idxa]) {
+    case '{':
+    case '[':
+    case ':':
+    case ',':
+      continue;
+    }
+    // Last document is complete, so the next document will appear after!
+    if (!arr_cnt && !obj_cnt) {
+      return parser.n_structural_indexes;
+    }
+    // Last document is incomplete; mark the document at i + 1 as the next one
+    return i;
+  }
+  // If we made it to the end, we want to finish counting to see if we have a full document.
+  switch (parser.buf[parser.structural_indexes[0]]) {
+    case '}':
+      obj_cnt--;
+      break;
+    case ']':
+      arr_cnt--;
+      break;
+    case '{':
+      obj_cnt++;
+      break;
+    case '[':
+      arr_cnt++;
+      break;
+  }
+  if (!arr_cnt && !obj_cnt) {
+    // We have a complete document.
+    return parser.n_structural_indexes;
+  }
+  return 0;
+}
+
+} // namespace stage1
+} // unnamed namespace
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE1_FIND_NEXT_DOCUMENT_INDEX_H
+/* end file generic/stage1/find_next_document_index.h for fallback */
+/* including generic/stage2/stringparsing.h for fallback: #include <generic/stage2/stringparsing.h> */
+/* begin file generic/stage2/stringparsing.h for fallback */
+#include <cstdint>
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/jsoncharutils.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+// This file contains the common code every implementation uses
+// It is intended to be included multiple times and compiled multiple times
+
+namespace simdjson {
+namespace fallback {
+namespace {
+/// @private
+namespace stringparsing {
+
+// begin copypasta
+// These chars yield themselves: " \ /
+// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
+// u not handled in this table as it's complex
+static const uint8_t escape_map[256] = {
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
+    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
+    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+};
+
+// handle a unicode codepoint
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr, bool allow_replacement) {
+  // Use the default Unicode Character 'REPLACEMENT CHARACTER' (U+FFFD)
+  constexpr uint32_t substitution_code_point = 0xfffd;
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) != ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+    } else {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+
+      // We have already checked that the high surrogate is valid and
+      // (code_point - 0xd800) < 1024.
+      //
+      // Check that code_point_2 is in the range 0xdc00..0xdfff
+      // and that code_point_2 was parsed from valid hex.
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if (low_bit >> 10) {
+        if(!allow_replacement) { return false; }
+        code_point = substitution_code_point;
+      } else {
+        code_point =  (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+
+    }
+  } else if (code_point >= 0xdc00 && code_point <= 0xdfff) {
+      // If we encounter a low surrogate (not preceded by a high surrogate)
+      // then we have an error.
+      if(!allow_replacement) { return false; }
+      code_point = substitution_code_point;
+  }
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+// handle a unicode codepoint using the wobbly convention
+// https://simonsapin.github.io/wtf-8/
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint_wobbly(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr) {
+  // It is not ideal that this function is nearly identical to handle_unicode_codepoint.
+  //
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion is not valid; we defer the check for this to inside the
+  // multilingual plane check.
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) == ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+      uint32_t low_bit = code_point_2 - 0xdc00;
+      if ((low_bit >> 10) ==  0) {
+        code_point =
+          (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+        *src_ptr += 6;
+      }
+    }
+  }
+
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+
+/**
+ * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+ * must be an unescaped quote terminating the string. It returns the final output
+ * position as pointer. In case of error (e.g., the string has bad escaped codes),
+ * then null_ptr is returned. It is assumed that the output buffer is large
+ * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+ * SIMDJSON_PADDING bytes.
+ */
+simdjson_warn_unused simdjson_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst, bool allow_replacement) {
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint(&src, &dst, allow_replacement)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+simdjson_warn_unused simdjson_inline uint8_t *parse_wobbly_string(const uint8_t *src, uint8_t *dst) {
+  // It is not ideal that this function is nearly identical to parse_string.
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto b = backslash_and_quote{};
+    auto bs_quote = b.copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint_wobbly(&src, &dst)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+}
+
+} // namespace stringparsing
+
+} // unnamed namespace
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_STRINGPARSING_H
+/* end file generic/stage2/stringparsing.h for fallback */
+/* including generic/stage2/logger.h for fallback: #include <generic/stage2/logger.h> */
+/* begin file generic/stage2/logger.h for fallback */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+
+// This is for an internal-only stage 2 specific logger.
+// Set LOG_ENABLED = true to log what stage 2 is doing!
+namespace simdjson {
+namespace fallback {
+namespace {
+namespace logger {
+
+  static constexpr const char * DASHES = "----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------";
+
+#if SIMDJSON_VERBOSE_LOGGING
+  static constexpr const bool LOG_ENABLED = true;
+#else
+  static constexpr const bool LOG_ENABLED = false;
+#endif
+  static constexpr const int LOG_EVENT_LEN = 20;
+  static constexpr const int LOG_BUFFER_LEN = 30;
+  static constexpr const int LOG_SMALL_BUFFER_LEN = 10;
+  static constexpr const int LOG_INDEX_LEN = 5;
+
+  static int log_depth; // Not threadsafe. Log only.
+
+  // Helper to turn unprintable or newline characters into spaces
+  static simdjson_inline char printable_char(char c) {
+    if (c >= 0x20) {
+      return c;
+    } else {
+      return ' ';
+    }
+  }
+
+  // Print the header and set up log_start
+  static simdjson_inline void log_start() {
+    if (LOG_ENABLED) {
+      log_depth = 0;
+      printf("\n");
+      printf("| %-*s | %-*s | %-*s | %-*s | Detail |\n", LOG_EVENT_LEN, "Event", LOG_BUFFER_LEN, "Buffer", LOG_SMALL_BUFFER_LEN, "Next", 5, "Next#");
+      printf("|%.*s|%.*s|%.*s|%.*s|--------|\n", LOG_EVENT_LEN+2, DASHES, LOG_BUFFER_LEN+2, DASHES, LOG_SMALL_BUFFER_LEN+2, DASHES, 5+2, DASHES);
+    }
+  }
+
+  simdjson_unused static simdjson_inline void log_string(const char *message) {
+    if (LOG_ENABLED) {
+      printf("%s\n", message);
+    }
+  }
+
+  // Logs a single line from the stage 2 DOM parser
+  template<typename S>
+  static simdjson_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
+    if (LOG_ENABLED) {
+      printf("| %*s%s%-*s ", log_depth*2, "", title_prefix, LOG_EVENT_LEN - log_depth*2 - int(strlen(title_prefix)), title);
+      auto current_index = structurals.at_beginning() ? nullptr : structurals.next_structural-1;
+      auto next_index = structurals.next_structural;
+      auto current = current_index ? &structurals.buf[*current_index] : reinterpret_cast<const uint8_t*>("                                                       ");
+      auto next = &structurals.buf[*next_index];
+      {
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_BUFFER_LEN;i++) {
+          printf("%c", printable_char(current[i]));
+        }
+        printf(" ");
+        // Print the next N characters in the buffer.
+        printf("| ");
+        // Otherwise, print the characters starting from the buffer position.
+        // Print spaces for unprintable or newline characters.
+        for (int i=0;i<LOG_SMALL_BUFFER_LEN;i++) {
+          printf("%c", printable_char(next[i]));
+        }
+        printf(" ");
+      }
+      if (current_index) {
+        printf("| %*u ", LOG_INDEX_LEN, *current_index);
+      } else {
+        printf("| %-*s ", LOG_INDEX_LEN, "");
+      }
+      // printf("| %*u ", LOG_INDEX_LEN, structurals.next_tape_index());
+      printf("| %-s ", detail);
+      printf("|\n");
+    }
+  }
+
+} // namespace logger
+} // unnamed namespace
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_LOGGER_H
+/* end file generic/stage2/logger.h for fallback */
+/* including generic/stage2/json_iterator.h for fallback: #include <generic/stage2/json_iterator.h> */
+/* begin file generic/stage2/json_iterator.h for fallback */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/logger.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+namespace simdjson {
+namespace fallback {
+namespace {
+namespace stage2 {
+
+class json_iterator {
+public:
+  const uint8_t* const buf;
+  uint32_t *next_structural;
+  dom_parser_implementation &dom_parser;
+  uint32_t depth{0};
+
+  /**
+   * Walk the JSON document.
+   *
+   * The visitor receives callbacks when values are encountered. All callbacks pass the iterator as
+   * the first parameter; some callbacks have other parameters as well:
+   *
+   * - visit_document_start() - at the beginning.
+   * - visit_document_end() - at the end (if things were successful).
+   *
+   * - visit_array_start() - at the start `[` of a non-empty array.
+   * - visit_array_end() - at the end `]` of a non-empty array.
+   * - visit_empty_array() - when an empty array is encountered.
+   *
+   * - visit_object_end() - at the start `]` of a non-empty object.
+   * - visit_object_start() - at the end `]` of a non-empty object.
+   * - visit_empty_object() - when an empty object is encountered.
+   * - visit_key(const uint8_t *key) - when a key in an object field is encountered. key is
+   *                                   guaranteed to point at the first quote of the string (`"key"`).
+   * - visit_primitive(const uint8_t *value) - when a value is a string, number, boolean or null.
+   * - visit_root_primitive(iter, uint8_t *value) - when the top-level value is a string, number, boolean or null.
+   *
+   * - increment_count(iter) - each time a value is found in an array or object.
+   */
+  template<bool STREAMING, typename V>
+  simdjson_warn_unused simdjson_inline error_code walk_document(V &visitor) noexcept;
+
+  /**
+   * Create an iterator capable of walking a JSON document.
+   *
+   * The document must have already passed through stage 1.
+   */
+  simdjson_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
+
+  /**
+   * Look at the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *peek() const noexcept;
+  /**
+   * Advance to the next token.
+   *
+   * Tokens can be strings, numbers, booleans, null, or operators (`[{]},:`)).
+   *
+   * They may include invalid JSON as well (such as `1.2.3` or `ture`).
+   */
+  simdjson_inline const uint8_t *advance() noexcept;
+  /**
+   * Get the remaining length of the document, from the start of the current token.
+   */
+  simdjson_inline size_t remaining_len() const noexcept;
+  /**
+   * Check if we are at the end of the document.
+   *
+   * If this is true, there are no more tokens.
+   */
+  simdjson_inline bool at_eof() const noexcept;
+  /**
+   * Check if we are at the beginning of the document.
+   */
+  simdjson_inline bool at_beginning() const noexcept;
+  simdjson_inline uint8_t last_structural() const noexcept;
+
+  /**
+   * Log that a value has been found.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_value(const char *type) const noexcept;
+  /**
+   * Log the start of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_start_value(const char *type) const noexcept;
+  /**
+   * Log the end of a multipart value.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_end_value(const char *type) const noexcept;
+  /**
+   * Log an error.
+   *
+   * Set LOG_ENABLED=true in logger.h to see logging.
+   */
+  simdjson_inline void log_error(const char *error) const noexcept;
+
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
+  template<typename V>
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
+};
+
+template<bool STREAMING, typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::walk_document(V &visitor) noexcept {
+  logger::log_start();
+
+  //
+  // Start the document
+  //
+  if (at_eof()) { return EMPTY; }
+  log_start_value("document");
+  SIMDJSON_TRY( visitor.visit_document_start(*this) );
+
+  //
+  // Read first value
+  //
+  {
+    auto value = advance();
+
+    // Make sure the outer object or array is closed before continuing; otherwise, there are ways we
+    // could get into memory corruption. See https://github.com/simdjson/simdjson/issues/906
+    if (!STREAMING) {
+      switch (*value) {
+        case '{': if (last_structural() != '}') { log_value("starting brace unmatched"); return TAPE_ERROR; }; break;
+        case '[': if (last_structural() != ']') { log_value("starting bracket unmatched"); return TAPE_ERROR; }; break;
+      }
+    }
+
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_root_primitive(*this, value) ); break;
+    }
+  }
+  goto document_end;
+
+//
+// Object parser states
+//
+object_begin:
+  log_start_value("object");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = false;
+  SIMDJSON_TRY( visitor.visit_object_start(*this) );
+
+  {
+    auto key = advance();
+    if (*key != '"') { log_error("Object does not start with a key"); return TAPE_ERROR; }
+    SIMDJSON_TRY( visitor.increment_count(*this) );
+    SIMDJSON_TRY( visitor.visit_key(*this, key) );
+  }
+
+object_field:
+  if (simdjson_unlikely( *advance() != ':' )) { log_error("Missing colon after key in object"); return TAPE_ERROR; }
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+object_continue:
+  switch (*advance()) {
+    case ',':
+      SIMDJSON_TRY( visitor.increment_count(*this) );
+      {
+        auto key = advance();
+        if (simdjson_unlikely( *key != '"' )) { log_error("Key string missing at beginning of field in object"); return TAPE_ERROR; }
+        SIMDJSON_TRY( visitor.visit_key(*this, key) );
+      }
+      goto object_field;
+    case '}': log_end_value("object"); SIMDJSON_TRY( visitor.visit_object_end(*this) ); goto scope_end;
+    default: log_error("No comma between object fields"); return TAPE_ERROR;
+  }
+
+scope_end:
+  depth--;
+  if (depth == 0) { goto document_end; }
+  if (dom_parser.is_array[depth]) { goto array_continue; }
+  goto object_continue;
+
+//
+// Array parser states
+//
+array_begin:
+  log_start_value("array");
+  depth++;
+  if (depth >= dom_parser.max_depth()) { log_error("Exceeded max depth!"); return DEPTH_ERROR; }
+  dom_parser.is_array[depth] = true;
+  SIMDJSON_TRY( visitor.visit_array_start(*this) );
+  SIMDJSON_TRY( visitor.increment_count(*this) );
+
+array_value:
+  {
+    auto value = advance();
+    switch (*value) {
+      case '{': if (*peek() == '}') { advance(); log_value("empty object"); SIMDJSON_TRY( visitor.visit_empty_object(*this) ); break; } goto object_begin;
+      case '[': if (*peek() == ']') { advance(); log_value("empty array"); SIMDJSON_TRY( visitor.visit_empty_array(*this) ); break; } goto array_begin;
+      default: SIMDJSON_TRY( visitor.visit_primitive(*this, value) ); break;
+    }
+  }
+
+array_continue:
+  switch (*advance()) {
+    case ',': SIMDJSON_TRY( visitor.increment_count(*this) ); goto array_value;
+    case ']': log_end_value("array"); SIMDJSON_TRY( visitor.visit_array_end(*this) ); goto scope_end;
+    default: log_error("Missing comma between array values"); return TAPE_ERROR;
+  }
+
+document_end:
+  log_end_value("document");
+  SIMDJSON_TRY( visitor.visit_document_end(*this) );
+
+  dom_parser.next_structural_index = uint32_t(next_structural - &dom_parser.structural_indexes[0]);
+
+  // If we didn't make it to the end, it's an error
+  if ( !STREAMING && dom_parser.next_structural_index != dom_parser.n_structural_indexes ) {
+    log_error("More than one JSON value at the root of the document, or extra characters at the end of the JSON!");
+    return TAPE_ERROR;
+  }
+
+  return SUCCESS;
+
+} // walk_document()
+
+simdjson_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+  : buf{_dom_parser.buf},
+    next_structural{&_dom_parser.structural_indexes[start_structural_index]},
+    dom_parser{_dom_parser} {
+}
+
+simdjson_inline const uint8_t *json_iterator::peek() const noexcept {
+  return &buf[*(next_structural)];
+}
+simdjson_inline const uint8_t *json_iterator::advance() noexcept {
+  return &buf[*(next_structural++)];
+}
+simdjson_inline size_t json_iterator::remaining_len() const noexcept {
+  return dom_parser.len - *(next_structural-1);
+}
+
+simdjson_inline bool json_iterator::at_eof() const noexcept {
+  return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
+}
+simdjson_inline bool json_iterator::at_beginning() const noexcept {
+  return next_structural == dom_parser.structural_indexes.get();
+}
+simdjson_inline uint8_t json_iterator::last_structural() const noexcept {
+  return buf[dom_parser.structural_indexes[dom_parser.n_structural_indexes - 1]];
+}
+
+simdjson_inline void json_iterator::log_value(const char *type) const noexcept {
+  logger::log_line(*this, "", type, "");
+}
+
+simdjson_inline void json_iterator::log_start_value(const char *type) const noexcept {
+  logger::log_line(*this, "+", type, "");
+  if (logger::LOG_ENABLED) { logger::log_depth++; }
+}
+
+simdjson_inline void json_iterator::log_end_value(const char *type) const noexcept {
+  if (logger::LOG_ENABLED) { logger::log_depth--; }
+  logger::log_line(*this, "-", type, "");
+}
+
+simdjson_inline void json_iterator::log_error(const char *error) const noexcept {
+  logger::log_line(*this, "", "ERROR", error);
+}
+
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
+  switch (*value) {
+    case '"': return visitor.visit_root_string(*this, value);
+    case 't': return visitor.visit_root_true_atom(*this, value);
+    case 'f': return visitor.visit_root_false_atom(*this, value);
+    case 'n': return visitor.visit_root_null_atom(*this, value);
+    case '-':
+    case '0': case '1': case '2': case '3': case '4':
+    case '5': case '6': case '7': case '8': case '9':
+      return visitor.visit_root_number(*this, value);
+    default:
+      log_error("Document starts with a non-value character");
+      return TAPE_ERROR;
+  }
+}
+template<typename V>
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
+  // Use the fact that most scalars are going to be either strings or numbers.
+  if(*value == '"') {
+    return visitor.visit_string(*this, value);
+  } else if (((*value - '0')  < 10) || (*value == '-')) {
+    return visitor.visit_number(*this, value);
+  }
+  // true, false, null are uncommon.
+  switch (*value) {
+    case 't': return visitor.visit_true_atom(*this, value);
+    case 'f': return visitor.visit_false_atom(*this, value);
+    case 'n': return visitor.visit_null_atom(*this, value);
+    default:
+      log_error("Non-value found when value was expected!");
+      return TAPE_ERROR;
+  }
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_JSON_ITERATOR_H
+/* end file generic/stage2/json_iterator.h for fallback */
+/* including generic/stage2/tape_writer.h for fallback: #include <generic/stage2/tape_writer.h> */
+/* begin file generic/stage2/tape_writer.h for fallback */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/internal/tape_type.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+#include <cstring>
+
+namespace simdjson {
+namespace fallback {
+namespace {
+namespace stage2 {
+
+struct tape_writer {
+  /** The next place to write to tape */
+  uint64_t *next_tape_loc;
+
+  /** Write a signed 64-bit value to tape. */
+  simdjson_inline void append_s64(int64_t value) noexcept;
+
+  /** Write an unsigned 64-bit value to tape. */
+  simdjson_inline void append_u64(uint64_t value) noexcept;
+
+  /** Write a double value to tape. */
+  simdjson_inline void append_double(double value) noexcept;
+
+  /** Write a big integer (as string) to tape. src points to first digit, len is byte count. */
+  simdjson_inline void append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept;
+
+  /**
+   * Append a tape entry (an 8-bit type,and 56 bits worth of value).
+   */
+  simdjson_inline void append(uint64_t val, internal::tape_type t) noexcept;
+
+  /**
+   * Skip the current tape entry without writing.
+   *
+   * Used to skip the start of the container, since we'll come back later to fill it in when the
+   * container ends.
+   */
+  simdjson_inline void skip() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a large u64 or i64.
+   */
+  simdjson_inline void skip_large_integer() noexcept;
+
+  /**
+   * Skip the number of tape entries necessary to write a double.
+   */
+  simdjson_inline void skip_double() noexcept;
+
+  /**
+   * Write a value to a known location on tape.
+   *
+   * Used to go back and write out the start of a container after the container ends.
+   */
+  simdjson_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
+
+private:
+  /**
+   * Append both the tape entry, and a supplementary value following it. Used for types that need
+   * all 64 bits, such as double and uint64_t.
+   */
+  template<typename T>
+  simdjson_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
+}; // struct tape_writer
+
+simdjson_inline void tape_writer::append_s64(int64_t value) noexcept {
+  append2(0, value, internal::tape_type::INT64);
+}
+
+simdjson_inline void tape_writer::append_u64(uint64_t value) noexcept {
+  append(0, internal::tape_type::UINT64);
+  *next_tape_loc = value;
+  next_tape_loc++;
+}
+
+/** Write a double value to tape. */
+simdjson_inline void tape_writer::append_double(double value) noexcept {
+  append2(0, value, internal::tape_type::DOUBLE);
+}
+
+simdjson_inline void tape_writer::skip() noexcept {
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::skip_large_integer() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::skip_double() noexcept {
+  next_tape_loc += 2;
+}
+
+simdjson_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
+  *next_tape_loc = val | ((uint64_t(char(t))) << 56);
+  next_tape_loc++;
+}
+
+template<typename T>
+simdjson_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
+  append(val, t);
+  static_assert(sizeof(val2) == sizeof(*next_tape_loc), "Type is not 64 bits!");
+  memcpy(next_tape_loc, &val2, sizeof(val2));
+  next_tape_loc++;
+}
+
+simdjson_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
+  tape_loc = val | ((uint64_t(char(t))) << 56);
+}
+
+simdjson_inline void tape_writer::append_bigint(const uint8_t *src, size_t len, uint8_t *&string_buf) noexcept {
+  // Write to string buffer: [4-byte LE length][digits][null]
+  uint32_t str_len = uint32_t(len);
+  memcpy(string_buf, &str_len, sizeof(uint32_t));
+  memcpy(string_buf + sizeof(uint32_t), src, len);
+  string_buf[sizeof(uint32_t) + len] = 0;
+  // Tape entry: offset into string buffer
+  // The caller must set the offset relative to doc.string_buf base
+  append(0, internal::tape_type::BIGINT); // placeholder offset, caller patches
+  string_buf += sizeof(uint32_t) + len + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_WRITER_H
+/* end file generic/stage2/tape_writer.h for fallback */
+/* including generic/stage2/tape_builder.h for fallback: #include <generic/stage2/tape_builder.h> */
+/* begin file generic/stage2/tape_builder.h for fallback */
+#ifndef SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #define SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H */
+/* amalgamation skipped (editor-only): #include <generic/stage2/base.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/json_iterator.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/stringparsing.h> */
+/* amalgamation skipped (editor-only): #include <generic/stage2/tape_writer.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/dom/document.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/atomparsing.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/dom_parser_implementation.h> */
+/* amalgamation skipped (editor-only): #include <simdjson/generic/numberparsing.h> */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+
+namespace simdjson {
+namespace fallback {
+namespace {
+namespace stage2 {
+
+struct tape_builder {
+  template<bool STREAMING>
+  simdjson_warn_unused static simdjson_inline error_code parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept;
+
+  /** Called when a non-empty document starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty document ends without error. */
+  simdjson_warn_unused simdjson_inline error_code visit_document_end(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty array starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_start(json_iterator &iter) noexcept;
+  /** Called when a non-empty array ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_array_end(json_iterator &iter) noexcept;
+  /** Called when an empty array is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_array(json_iterator &iter) noexcept;
+
+  /** Called when a non-empty object starts. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_start(json_iterator &iter) noexcept;
+  /**
+   * Called when a key in a field is encountered.
+   *
+   * primitive, visit_object_start, visit_empty_object, visit_array_start, or visit_empty_array
+   * will be called after this with the field value.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
+  /** Called when a non-empty object ends. */
+  simdjson_warn_unused simdjson_inline error_code visit_object_end(json_iterator &iter) noexcept;
+  /** Called when an empty object is found. */
+  simdjson_warn_unused simdjson_inline error_code visit_empty_object(json_iterator &iter) noexcept;
+
+  /**
+   * Called when a string, number, boolean or null is found.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  /**
+   * Called when a string, number, boolean or null is found at the top level of a document (i.e.
+   * when there is no array or object and the entire document is a single string, number, boolean or
+   * null.
+   *
+   * This is separate from primitive() because simdjson's normal primitive parsing routines assume
+   * there is at least one more token after the value, which is only true in an array or object.
+   */
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  simdjson_warn_unused simdjson_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+
+  /** Called each time a new field or element in an array or object is found. */
+  simdjson_warn_unused simdjson_inline error_code increment_count(json_iterator &iter) noexcept;
+
+  /** Next location to write to tape */
+  tape_writer tape;
+private:
+  /** Next write location in the string buf for stage 2 parsing */
+  uint8_t *current_string_buf_loc;
+
+  simdjson_inline tape_builder(dom::document &doc) noexcept;
+
+  simdjson_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
+  simdjson_inline void start_container(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_warn_unused simdjson_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
+  simdjson_inline void on_end_string(uint8_t *dst) noexcept;
+}; // struct tape_builder
+
+template<bool STREAMING>
+simdjson_warn_unused simdjson_inline error_code tape_builder::parse_document(
+    dom_parser_implementation &dom_parser,
+    dom::document &doc) noexcept {
+  dom_parser.doc = &doc;
+  json_iterator iter(dom_parser, STREAMING ? dom_parser.next_structural_index : 0);
+  tape_builder builder(doc);
+  return iter.walk_document<STREAMING>(builder);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_root_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+  return iter.visit_primitive(*this, value);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
+  return empty_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
+  start_container(iter);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
+  return end_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
+  constexpr uint32_t start_tape_index = 0;
+  tape.append(start_tape_index, internal::tape_type::ROOT);
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter), internal::tape_type::ROOT);
+  return SUCCESS;
+}
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
+  return visit_string(iter, key, true);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].count++; // we have a key value pair in the object at parser.dom_parser.depth - 1
+  return SUCCESS;
+}
+
+simdjson_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
+  iter.log_value(key ? "key" : "string");
+  uint8_t *dst = on_start_string(iter);
+  dst = stringparsing::parse_string(value+1, dst, false); // We do not allow replacement when the escape characters are invalid.
+  if (dst == nullptr) {
+    iter.log_error("Invalid escape in string");
+    return STRING_ERROR;
+  }
+  on_end_string(dst);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
+  return visit_string(iter, value);
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("number");
+  error_code err = numberparsing::parse_number(value, tape);
+  if (simdjson_unlikely(err == BIGINT_ERROR &&
+      iter.dom_parser._number_as_string)) {
+    // Write big integer to string buffer using the same format as strings.
+    // Scan digits the same way parse_number does (skip optional '-', then digits).
+    const uint8_t *p = value;
+    if (*p == '-') p++;
+    while (numberparsing::is_digit(*p)) p++;
+    size_t len = size_t(p - value);
+    tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::BIGINT);
+    uint8_t *dst = current_string_buf_loc + sizeof(uint32_t);
+    memcpy(dst, value, len);
+    dst += len;
+    on_end_string(dst);
+    return SUCCESS;
+  }
+  return err;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
+  //
+  // We need to make a copy to make sure that the string is space terminated.
+  // This is not about padding the input, which should already padded up
+  // to len + SIMDJSON_PADDING. However, we have no control at this stage
+  // on how the padding was done. What if the input string was padded with nulls?
+  // It is quite common for an input string to have an extra null character (C string).
+  // We do not want to allow 9\0 (where \0 is the null character) inside a JSON
+  // document, but the string "9\0" by itself is fine. So we make a copy and
+  // pad the input with spaces when we know that there is just one input element.
+  // This copy is relatively expensive, but it will almost never be called in
+  // practice unless you are in the strange scenario where you have many JSON
+  // documents made of single atoms.
+  //
+  std::unique_ptr<uint8_t[]>copy(new (std::nothrow) uint8_t[iter.remaining_len() + SIMDJSON_PADDING]);
+  if (copy.get() == nullptr) { return MEMALLOC; }
+  std::memcpy(copy.get(), value, iter.remaining_len());
+  std::memset(copy.get() + iter.remaining_len(), ' ', SIMDJSON_PADDING);
+  error_code error = visit_number(iter, copy.get());
+  return error;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value)) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("true");
+  if (!atomparsing::is_valid_true_atom(value, iter.remaining_len())) { return T_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::TRUE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value)) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("false");
+  if (!atomparsing::is_valid_false_atom(value, iter.remaining_len())) { return F_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::FALSE_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value)) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+  iter.log_value("null");
+  if (!atomparsing::is_valid_null_atom(value, iter.remaining_len())) { return N_ATOM_ERROR; }
+  tape.append(0, internal::tape_type::NULL_VALUE);
+  return SUCCESS;
+}
+
+// private:
+
+simdjson_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
+  return uint32_t(tape.next_tape_loc - iter.dom_parser.doc->tape.get());
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  auto start_index = next_tape_index(iter);
+  tape.append(start_index+2, start);
+  tape.append(start_index, end);
+  return SUCCESS;
+}
+
+simdjson_inline void tape_builder::start_container(json_iterator &iter) noexcept {
+  iter.dom_parser.open_containers[iter.depth].tape_index = next_tape_index(iter);
+  iter.dom_parser.open_containers[iter.depth].count = 0;
+  tape.skip(); // We don't actually *write* the start element until the end.
+}
+
+simdjson_warn_unused simdjson_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+  // Write the ending tape element, pointing at the start location
+  const uint32_t start_tape_index = iter.dom_parser.open_containers[iter.depth].tape_index;
+  tape.append(start_tape_index, end);
+  // Write the start tape element, pointing at the end location (and including count)
+  // count can overflow if it exceeds 24 bits... so we saturate
+  // the convention being that a cnt of 0xffffff or more is undetermined in value (>=  0xffffff).
+  const uint32_t count = iter.dom_parser.open_containers[iter.depth].count;
+  const uint32_t cntsat = count > 0xFFFFFF ? 0xFFFFFF : count;
+  tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter) | (uint64_t(cntsat) << 32), start);
+  return SUCCESS;
+}
+
+simdjson_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
+  // we advance the point, accounting for the fact that we have a NULL termination
+  tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::STRING);
+  return current_string_buf_loc + sizeof(uint32_t);
+}
+
+simdjson_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
+  uint32_t str_length = uint32_t(dst - (current_string_buf_loc + sizeof(uint32_t)));
+  // TODO check for overflow in case someone has a crazy string (>=4GB?)
+  // But only add the overflow check when the document itself exceeds 4GB
+  // Currently unneeded because we refuse to parse docs larger or equal to 4GB.
+  memcpy(current_string_buf_loc, &str_length, sizeof(uint32_t));
+  // NULL termination is still handy if you expect all your strings to
+  // be NULL terminated? It comes at a small cost
+  *dst = 0;
+  current_string_buf_loc = dst + 1;
+}
+
+} // namespace stage2
+} // unnamed namespace
+} // namespace fallback
+} // namespace simdjson
+
+#endif // SIMDJSON_SRC_GENERIC_STAGE2_TAPE_BUILDER_H
+/* end file generic/stage2/tape_builder.h for fallback */
+
+//
+// Stage 1
+//
+
+namespace simdjson {
+namespace fallback {
+
+simdjson_warn_unused error_code implementation::create_dom_parser_implementation(
+  size_t capacity,
+  size_t max_depth,
+  std::unique_ptr<internal::dom_parser_implementation>& dst
+) const noexcept {
+  dst.reset( new (std::nothrow) fallback::dom_parser_implementation() );
+  if (!dst) { return MEMALLOC; }
+  if (auto err = dst->set_capacity(capacity))
+    return err;
+  if (auto err = dst->set_max_depth(max_depth))
+    return err;
+  return SUCCESS;
+}
+
+namespace {
+namespace stage1 {
+
+class structural_scanner {
+public:
+
+simdjson_inline structural_scanner(dom_parser_implementation &_parser, stage1_mode _partial)
+  : buf{_parser.buf},
+    next_structural_index{_parser.structural_indexes.get()},
+    parser{_parser},
+    len{static_cast<uint32_t>(_parser.len)},
+    partial{_partial} {
+}
+
+simdjson_inline void add_structural() {
+  *next_structural_index = idx;
+  next_structural_index++;
+}
+
+simdjson_inline bool is_continuation(uint8_t c) {
+  return (c & 0xc0) == 0x80;
+}
+
+simdjson_inline void validate_utf8_character() {
+  // Continuation
+  if (simdjson_unlikely((buf[idx] & 0x40) == 0)) {
+    // extra continuation
+    error = UTF8_ERROR;
+    idx++;
+    return;
+  }
+
+  // 2-byte
+  if ((buf[idx] & 0x20) == 0) {
+    // missing continuation
+    if (simdjson_unlikely(idx+1 > len || !is_continuation(buf[idx+1]))) {
+      if (idx+1 > len && is_streaming(partial)) { idx = len; return; }
+      error = UTF8_ERROR;
+      idx++;
+      return;
+    }
+    // overlong: 1100000_ 10______
+    if (buf[idx] <= 0xc1) { error = UTF8_ERROR; }
+    idx += 2;
+    return;
+  }
+
+  // 3-byte
+  if ((buf[idx] & 0x10) == 0) {
+    // missing continuation
+    if (simdjson_unlikely(idx+2 > len || !is_continuation(buf[idx+1]) || !is_continuation(buf[idx+2]))) {
+      if (idx+2 > len && is_streaming(partial)) { idx = len; return; }
+      error = UTF8_ERROR;
+      idx++;
+      return;
+    }
+    // overlong: 11100000 100_____ ________
+    if (buf[idx] == 0xe0 && buf[idx+1] <= 0x9f) { error = UTF8_ERROR; }
+    // surrogates: U+D800-U+DFFF 11101101 101_____
+    if (buf[idx] == 0xed && buf[idx+1] >= 0xa0) { error = UTF8_ERROR; }
+    idx += 3;
+    return;
+  }
+
+  // 4-byte
+  // missing continuation
+  if (simdjson_unlikely(idx+3 > len || !is_continuation(buf[idx+1]) || !is_continuation(buf[idx+2]) || !is_continuation(buf[idx+3]))) {
+    if (idx+2 > len && is_streaming(partial)) { idx = len; return; }
+    error = UTF8_ERROR;
+    idx++;
+    return;
+  }
+  // overlong: 11110000 1000____ ________ ________
+  if (buf[idx] == 0xf0 && buf[idx+1] <= 0x8f) { error = UTF8_ERROR; }
+  // too large: > U+10FFFF:
+  // 11110100 (1001|101_)____
+  // 1111(1___|011_|0101) 10______
+  // also includes 5, 6, 7 and 8 byte characters:
+  // 11111___
+  if (buf[idx] == 0xf4 && buf[idx+1] >= 0x90) { error = UTF8_ERROR; }
+  if (buf[idx] >= 0xf5) { error = UTF8_ERROR; }
+  idx += 4;
+}
+
+static const uint8_t CHAR_TYPE_SPACE     = 1 << 0;
+static const uint8_t CHAR_TYPE_OPERATOR  = 1 << 1;
+static const uint8_t CHAR_TYPE_ESC_ASCII = 1 << 2;
+static const uint8_t CHAR_TYPE_NON_ASCII = 1 << 3;
+
+const uint8_t char_table[256] = {
+  0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04,
+  0x04, 0x05, 0x05, 0x04, 0x04, 0x05, 0x04, 0x04,
+  0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04,
+  0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04,
+  0x01, 0x00, 0x04, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x02, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x02, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x02, 0x04, 0x02, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
+  0x00, 0x00, 0x00, 0x02, 0x00, 0x02, 0x00, 0x00,
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08,
+  0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08
+};
+
+simdjson_inline bool char_is_type(uint8_t c, uint8_t type) {
+  return (char_table[c] & type);
+}
+
+simdjson_inline bool char_is_space(uint8_t c) {
+  return char_is_type(c, CHAR_TYPE_SPACE);
+}
+
+simdjson_inline bool char_is_operator(uint8_t c) {
+  return char_is_type(c, CHAR_TYPE_OPERATOR);
+}
+
+simdjson_inline bool char_is_space_or_operator(uint8_t c) {
+  return char_is_type(c, CHAR_TYPE_SPACE | CHAR_TYPE_OPERATOR);
+}
+
+simdjson_inline bool char_is_ascii_stop(uint8_t c) {
+  return char_is_type(c, CHAR_TYPE_ESC_ASCII | CHAR_TYPE_NON_ASCII);
+}
+
+// Returns true if the string is unclosed.
+simdjson_inline bool validate_string() {
+  idx++; // skip first quote
+  while (idx < len) {
+    do {
+      if (char_is_ascii_stop(buf[idx])) { break; }
+      idx++;
+    } while (idx < len);
+    if (idx >= len) { return true; }
+    if (buf[idx] == '"') {
+      return false;
+    }
+    if (buf[idx] == '\\') {
+      idx += 2;
+    } else if (simdjson_unlikely(buf[idx] & 0x80)) {
+      validate_utf8_character();
+    } else {
+      if (buf[idx] < 0x20) { error = UNESCAPED_CHARS; }
+      idx++;
+    }
+  }
+  if (idx >= len) { return true; }
+  return false;
+}
+
+//
+// Parse the entire input in STEP_SIZE-byte chunks.
+//
+simdjson_warn_unused simdjson_inline error_code scan() {
+  bool unclosed_string = false;
+  for (;idx<len;idx++) {
+    do {
+      if (!char_is_space(buf[idx])) { break; }
+      idx++;
+    } while (idx < len);
+    if (idx >= len) { break; }
+    // String
+    if (buf[idx] == '"') {
+      add_structural();
+      unclosed_string |= validate_string();
+    // Operator
+    } else if (char_is_operator(buf[idx])) {
+      add_structural();
+    // Primitive or invalid character (invalid characters will be checked in stage 2)
+    } else {
+      // Anything else, add the structural and go until we find the next one
+      add_structural();
+      while (idx+1<len && !char_is_space_or_operator(buf[idx+1])) {
+        idx++;
+      };
+    }
+  }
+  // We pad beyond.
+  // https://github.com/simdjson/simdjson/issues/906
+  // See json_structural_indexer.h for an explanation.
+  *next_structural_index = len; // assumed later in partial == stage1_mode::streaming_final
+  next_structural_index[1] = len;
+  next_structural_index[2] = 0;
+  parser.n_structural_indexes = uint32_t(next_structural_index - parser.structural_indexes.get());
+  if (simdjson_unlikely(parser.n_structural_indexes == 0)) { return EMPTY; }
+  parser.next_structural_index = 0;
+  if (partial == stage1_mode::streaming_partial) {
+    if(unclosed_string) {
+      parser.n_structural_indexes--;
+      if (simdjson_unlikely(parser.n_structural_indexes == 0)) { return CAPACITY; }
+    }
+    // We truncate the input to the end of the last complete document (or zero).
+    auto new_structural_indexes = find_next_document_index(parser);
+    if (new_structural_indexes == 0 && parser.n_structural_indexes > 0) {
+      if(parser.structural_indexes[0] == 0) {
+        // If the buffer is partial and we started at index 0 but the document is
+        // incomplete, it's too big to parse.
+        return CAPACITY;
+      } else {
+        // It is possible that the document could be parsed, we just had a lot
+        // of white space.
+        parser.n_structural_indexes = 0;
+        return EMPTY;
+      }
+    }
+    parser.n_structural_indexes = new_structural_indexes;
+  } else if(partial == stage1_mode::streaming_final) {
+    if(unclosed_string) { parser.n_structural_indexes--; }
+    // We truncate the input to the end of the last complete document (or zero).
+    // Because partial == stage1_mode::streaming_final, it means that we may
+    // silently ignore trailing garbage. Though it sounds bad, we do it
+    // deliberately because many people who have streams of JSON documents
+    // will truncate them for processing. E.g., imagine that you are uncompressing
+    // the data from a size file or receiving it in chunks from the network. You
+    // may not know where exactly the last document will be. Meanwhile the
+    // document_stream instances allow people to know the JSON documents they are
+    // parsing (see the iterator.source() method).
+    parser.n_structural_indexes = find_next_document_index(parser);
+    // We store the initial n_structural_indexes so that the client can see
+    // whether we used truncation. If initial_n_structural_indexes == parser.n_structural_indexes,
+    // then this will query parser.structural_indexes[parser.n_structural_indexes] which is len,
+    // otherwise, it will copy some prior index.
+    parser.structural_indexes[parser.n_structural_indexes + 1] = parser.structural_indexes[parser.n_structural_indexes];
+    // This next line is critical, do not change it unless you understand what you are
+    // doing.
+    parser.structural_indexes[parser.n_structural_indexes] = uint32_t(len);
+    if (parser.n_structural_indexes == 0) { return EMPTY; }
+  } else if(unclosed_string) { error = UNCLOSED_STRING; }
+  return error;
+}
+
+private:
+  const uint8_t *buf;
+  uint32_t *next_structural_index;
+  dom_parser_implementation &parser;
+  uint32_t len;
+  uint32_t idx{0};
+  error_code error{SUCCESS};
+  stage1_mode partial;
+}; // structural_scanner
+
+} // namespace stage1
+} // unnamed namespace
+
+simdjson_warn_unused error_code dom_parser_implementation::stage1(const uint8_t *_buf, size_t _len, stage1_mode partial) noexcept {
+  this->buf = _buf;
+  this->len = _len;
+  stage1::structural_scanner scanner(*this, partial);
+  return scanner.scan();
+}
+
+// big table for the minifier
+static uint8_t jump_table[256 * 3] = {
+    0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0,
+    1, 1, 0, 1, 0, 0, 1, 0, 0, 1, 1, 0, 1, 1, 0, 1, 0, 0, 1, 1, 0, 1, 1, 0, 1,
+    1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1,
+    0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 0, 0,
+    1, 1, 1, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1,
+    1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1,
+    0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0,
+    1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1,
+    1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1,
+    0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0,
+    1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1,
+    1, 0, 0, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1,
+    0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0,
+    1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1,
+    1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1,
+    0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0,
+    1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1,
+    1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1,
+    0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0,
+    1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1,
+    1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1,
+    0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0,
+    1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1,
+    1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1,
+    0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0,
+    1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1,
+    1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1,
+    0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0,
+    1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1,
+    1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1,
+    0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1, 0, 1, 1,
+};
+
+simdjson_warn_unused error_code implementation::minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) const noexcept {
+  size_t i = 0, pos = 0;
+  uint8_t quote = 0;
+  uint8_t nonescape = 1;
+
+  while (i < len) {
+    unsigned char c = buf[i];
+    uint8_t *meta = jump_table + 3 * c;
+
+    quote = quote ^ (meta[0] & nonescape);
+    dst[pos] = c;
+    pos += meta[2] | quote;
+
+    i += 1;
+    nonescape = uint8_t(~nonescape) | (meta[1]);
+  }
+  dst_len = pos; // we intentionally do not work with a reference
+  // for fear of aliasing
+  return quote ? UNCLOSED_STRING : SUCCESS;
+}
+
+// credit: based on code from Google Fuchsia (Apache Licensed)
+simdjson_warn_unused bool implementation::validate_utf8(const char *buf, size_t len) const noexcept {
+  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
+  uint64_t pos = 0;
+  uint32_t code_point = 0;
+  while (pos < len) {
+    // check of the next 8 bytes are ascii.
+    uint64_t next_pos = pos + 16;
+    if (next_pos <= len) { // if it is safe to read 8 more bytes, check that they are ascii
+      uint64_t v1;
+      memcpy(&v1, data + pos, sizeof(uint64_t));
+      uint64_t v2;
+      memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
+      uint64_t v{v1 | v2};
+      if ((v & 0x8080808080808080) == 0) {
+        pos = next_pos;
+        continue;
+      }
+    }
+    unsigned char byte = data[pos];
+    if (byte < 0x80) {
+      pos++;
+      continue;
+    } else if ((byte & 0xe0) == 0xc0) {
+      next_pos = pos + 2;
+      if (next_pos > len) { return false; }
+      if ((data[pos + 1] & 0xc0) != 0x80) { return false; }
+      // range check
+      code_point = (byte & 0x1f) << 6 | (data[pos + 1] & 0x3f);
+      if (code_point < 0x80 || 0x7ff < code_point) { return false; }
+    } else if ((byte & 0xf0) == 0xe0) {
+      next_pos = pos + 3;
+      if (next_pos > len) { return false; }
+      if ((data[pos + 1] & 0xc0) != 0x80) { return false; }
+      if ((data[pos + 2] & 0xc0) != 0x80) { return false; }
+      // range check
+      code_point = (byte & 0x0f) << 12 |
+                   (data[pos + 1] & 0x3f) << 6 |
+                   (data[pos + 2] & 0x3f);
+      if (code_point < 0x800 || 0xffff < code_point ||
+          (0xd7ff < code_point && code_point < 0xe000)) {
+        return false;
+      }
+    } else if ((byte & 0xf8) == 0xf0) { // 0b11110000
+      next_pos = pos + 4;
+      if (next_pos > len) { return false; }
+      if ((data[pos + 1] & 0xc0) != 0x80) { return false; }
+      if ((data[pos + 2] & 0xc0) != 0x80) { return false; }
+      if ((data[pos + 3] & 0xc0) != 0x80) { return false; }
+      // range check
+      code_point =
+          (byte & 0x07) << 18 | (data[pos + 1] & 0x3f) << 12 |
+          (data[pos + 2] & 0x3f) << 6 | (data[pos + 3] & 0x3f);
+      if (code_point <= 0xffff || 0x10ffff < code_point) { return false; }
+    } else {
+      // we may have a continuation
+      return false;
+    }
+    pos = next_pos;
+  }
+  return true;
+}
+
+} // namespace fallback
+} // namespace simdjson
+
+//
+// Stage 2
+//
+
+namespace simdjson {
+namespace fallback {
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<false>(*this, _doc);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::stage2_next(dom::document &_doc) noexcept {
+  return stage2::tape_builder::parse_document<true>(*this, _doc);
+}
+
+SIMDJSON_NO_SANITIZE_MEMORY
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_string(const uint8_t *src, uint8_t *dst, bool replacement_char) const noexcept {
+  return fallback::stringparsing::parse_string(src, dst, replacement_char);
+}
+
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_wobbly_string(const uint8_t *src, uint8_t *dst) const noexcept {
+  return fallback::stringparsing::parse_wobbly_string(src, dst);
+}
+
+simdjson_warn_unused error_code dom_parser_implementation::parse(const uint8_t *_buf, size_t _len, dom::document &_doc) noexcept {
+  auto error = stage1(_buf, _len, stage1_mode::regular);
+  if (error) { return error; }
+  return stage2(_doc);
+}
+
+} // namespace fallback
+} // namespace simdjson
+
+/* including simdjson/fallback/end.h: #include <simdjson/fallback/end.h> */
+/* begin file simdjson/fallback/end.h */
+/* amalgamation skipped (editor-only): #ifndef SIMDJSON_CONDITIONAL_INCLUDE */
+/* amalgamation skipped (editor-only): #include "simdjson/fallback/base.h" */
+/* amalgamation skipped (editor-only): #endif // SIMDJSON_CONDITIONAL_INCLUDE */
+
+/* undefining SIMDJSON_IMPLEMENTATION from "fallback" */
+#undef SIMDJSON_IMPLEMENTATION
+/* end file simdjson/fallback/end.h */
+
+#endif // SIMDJSON_SRC_FALLBACK_CPP
+/* end file fallback.cpp */
+#endif
+/* undefining SIMDJSON_CONDITIONAL_INCLUDE */
+#undef SIMDJSON_CONDITIONAL_INCLUDE
+
+SIMDJSON_POP_DISABLE_UNUSED_WARNINGS
+
+/* end file simdjson.cpp */