simdjson.cpp

/* auto-generated on Fri Mar 20 11:47:31 PDT 2020. Do not edit! */
#include "simdjson.h"

/* used for http://dmalloc.com/ Dmalloc - Debug Malloc Library */
#ifdef DMALLOC
#include "dmalloc.h"
#endif

/* begin file src/simdjson.cpp */
/* begin file src/implementation.cpp */
/* begin file src/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_ISADETECTION_H
#define SIMDJSON_ISADETECTION_H

#include <stdint.h>
#include <stdlib.h>
#if defined(_MSC_VER)
#include <intrin.h>
#elif defined(HAVE_GCC_GET_CPUID) && defined(USE_GCC_GET_CPUID)
#include <cpuid.h>
#endif

namespace simdjson {

// Can be found on Intel ISA Reference for CPUID
constexpr uint32_t cpuid_avx2_bit = 1 << 5;      // Bit 5 of EBX for EAX=0x7
constexpr uint32_t cpuid_bmi1_bit = 1 << 3;      // bit 3 of EBX for EAX=0x7
constexpr uint32_t cpuid_bmi2_bit = 1 << 8;      // bit 8 of EBX for EAX=0x7
constexpr uint32_t cpuid_sse42_bit = 1 << 20;    // bit 20 of ECX for EAX=0x1
constexpr uint32_t cpuid_pclmulqdq_bit = 1 << 1; // bit  1 of ECX for EAX=0x1

enum instruction_set {
  DEFAULT = 0x0,
  NEON = 0x1,
  AVX2 = 0x4,
  SSE42 = 0x8,
  PCLMULQDQ = 0x10,
  BMI1 = 0x20,
  BMI2 = 0x40
};

#if defined(__arm__) || defined(__aarch64__) // incl. armel, armhf, arm64

#if defined(__ARM_NEON)

static inline uint32_t detect_supported_architectures() {
  return instruction_set::NEON;
}

#else // ARM without NEON

static inline uint32_t detect_supported_architectures() {
  return instruction_set::DEFAULT;
}

#endif

#else // x86
static inline void cpuid(uint32_t *eax, uint32_t *ebx, uint32_t *ecx,
                         uint32_t *edx) {
#if defined(_MSC_VER)
  int cpu_info[4];
  __cpuid(cpu_info, *eax);
  *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 uint32_t detect_supported_architectures() {
  uint32_t eax, ebx, ecx, edx;
  uint32_t host_isa = 0x0;

  // 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;
  }

  // EBX for EAX=0x1
  eax = 0x1;
  cpuid(&eax, &ebx, &ecx, &edx);

  if (ecx & cpuid_sse42_bit) {
    host_isa |= instruction_set::SSE42;
  }

  if (ecx & cpuid_pclmulqdq_bit) {
    host_isa |= instruction_set::PCLMULQDQ;
  }

  return host_isa;
}

#endif // end SIMD extension detection code

} // namespace simdjson::internal

#endif // SIMDJSON_ISADETECTION_H
/* end file src/isadetection.h */
#include <initializer_list>

// Static array of known implementations. We're hoping these get baked into the executable
// without requiring a static initializer.

#if SIMDJSON_IMPLEMENTATION_HASWELL
/* begin file src/haswell/implementation.h */
#ifndef SIMDJSON_HASWELL_IMPLEMENTATION_H
#define SIMDJSON_HASWELL_IMPLEMENTATION_H

/* isadetection.h already included: #include "isadetection.h" */

namespace simdjson::haswell {

class implementation final : public simdjson::implementation {
public:
  really_inline implementation() : simdjson::implementation(
      "haswell",
      "Intel/AMD AVX2",
      instruction_set::AVX2 | instruction_set::PCLMULQDQ | instruction_set::BMI1 | instruction_set::BMI2
  ) {}
  WARN_UNUSED error_code parse(const uint8_t *buf, size_t len, document::parser &parser) const noexcept final;
  WARN_UNUSED error_code stage1(const uint8_t *buf, size_t len, document::parser &parser, bool streaming) const noexcept final;
  WARN_UNUSED error_code stage2(const uint8_t *buf, size_t len, document::parser &parser) const noexcept final;
  WARN_UNUSED error_code stage2(const uint8_t *buf, size_t len, document::parser &parser, size_t &next_json) const noexcept final;
};

} // namespace simdjson::haswell

#endif // SIMDJSON_HASWELL_IMPLEMENTATION_H
/* end file src/haswell/implementation.h */
namespace simdjson::internal { const haswell::implementation haswell_singleton{}; }
#endif // SIMDJSON_IMPLEMENTATION_HASWELL

#if SIMDJSON_IMPLEMENTATION_WESTMERE
/* begin file src/westmere/implementation.h */
#ifndef SIMDJSON_WESTMERE_IMPLEMENTATION_H
#define SIMDJSON_WESTMERE_IMPLEMENTATION_H

/* isadetection.h already included: #include "isadetection.h" */

namespace simdjson::westmere {

class implementation final : public simdjson::implementation {
public:
  really_inline implementation() : simdjson::implementation("westmere", "Intel/AMD SSE4.2", instruction_set::SSE42 | instruction_set::PCLMULQDQ) {}
  WARN_UNUSED error_code parse(const uint8_t *buf, size_t len, document::parser &parser) const noexcept final;
  WARN_UNUSED error_code stage1(const uint8_t *buf, size_t len, document::parser &parser, bool streaming) const noexcept final;
  WARN_UNUSED error_code stage2(const uint8_t *buf, size_t len, document::parser &parser) const noexcept final;
  WARN_UNUSED error_code stage2(const uint8_t *buf, size_t len, document::parser &parser, size_t &next_json) const noexcept final;
};

} // namespace simdjson::westmere

#endif // SIMDJSON_WESTMERE_IMPLEMENTATION_H
/* end file src/westmere/implementation.h */
namespace simdjson::internal { const westmere::implementation westmere_singleton{}; }
#endif // SIMDJSON_IMPLEMENTATION_WESTMERE

#if SIMDJSON_IMPLEMENTATION_ARM64
/* begin file src/arm64/implementation.h */
#ifndef SIMDJSON_ARM64_IMPLEMENTATION_H
#define SIMDJSON_ARM64_IMPLEMENTATION_H

/* isadetection.h already included: #include "isadetection.h" */

namespace simdjson::arm64 {

class implementation final : public simdjson::implementation {
public:
  really_inline implementation() : simdjson::implementation("arm64", "ARM NEON", instruction_set::NEON) {}
  WARN_UNUSED error_code parse(const uint8_t *buf, size_t len, document::parser &parser) const noexcept final;
  WARN_UNUSED error_code stage1(const uint8_t *buf, size_t len, document::parser &parser, bool streaming) const noexcept final;
  WARN_UNUSED error_code stage2(const uint8_t *buf, size_t len, document::parser &parser) const noexcept final;
  WARN_UNUSED error_code stage2(const uint8_t *buf, size_t len, document::parser &parser, size_t &next_json) const noexcept final;
};

} // namespace simdjson::arm64

#endif // SIMDJSON_ARM64_IMPLEMENTATION_H
/* end file src/arm64/implementation.h */
namespace simdjson::internal { const arm64::implementation arm64_singleton{}; }
#endif // SIMDJSON_IMPLEMENTATION_ARM64

#if SIMDJSON_IMPLEMENTATION_FALLBACK
/* begin file src/fallback/implementation.h */
#ifndef SIMDJSON_FALLBACK_IMPLEMENTATION_H
#define SIMDJSON_FALLBACK_IMPLEMENTATION_H

/* isadetection.h already included: #include "isadetection.h" */

namespace simdjson::fallback {

class implementation final : public simdjson::implementation {
public:
  really_inline implementation() : simdjson::implementation(
      "fallback",
      "Generic fallback implementation",
      0
  ) {}
  WARN_UNUSED error_code parse(const uint8_t *buf, size_t len, document::parser &parser) const noexcept final;
  WARN_UNUSED error_code stage1(const uint8_t *buf, size_t len, document::parser &parser, bool streaming) const noexcept final;
  WARN_UNUSED error_code stage2(const uint8_t *buf, size_t len, document::parser &parser) const noexcept final;
  WARN_UNUSED error_code stage2(const uint8_t *buf, size_t len, document::parser &parser, size_t &next_json) const noexcept final;
};

} // namespace simdjson::fallback

#endif // SIMDJSON_FALLBACK_IMPLEMENTATION_H
/* end file src/fallback/implementation.h */
namespace simdjson::internal { const fallback::implementation fallback_singleton{}; }
#endif // SIMDJSON_IMPLEMENTATION_FALLBACK

namespace simdjson::internal {

constexpr const std::initializer_list<const implementation *> available_implementation_pointers {
#if SIMDJSON_IMPLEMENTATION_HASWELL
  &haswell_singleton,
#endif
#if SIMDJSON_IMPLEMENTATION_WESTMERE
  &westmere_singleton,
#endif
#if SIMDJSON_IMPLEMENTATION_ARM64
  &arm64_singleton,
#endif
#if SIMDJSON_IMPLEMENTATION_FALLBACK
  &fallback_singleton,
#endif
}; // available_implementation_pointers

// So we can return UNSUPPORTED_ARCHITECTURE from the parser when there is no support
class unsupported_implementation final : public implementation {
public:
  WARN_UNUSED error_code parse(const uint8_t *, size_t, document::parser &) const noexcept final {
    return UNSUPPORTED_ARCHITECTURE;
  }
  WARN_UNUSED error_code stage1(const uint8_t *, size_t, document::parser &, bool) const noexcept final {
    return UNSUPPORTED_ARCHITECTURE;
  }
  WARN_UNUSED error_code stage2(const uint8_t *, size_t, document::parser &) const noexcept final {
    return UNSUPPORTED_ARCHITECTURE;
  }
  WARN_UNUSED error_code stage2(const uint8_t *, size_t, document::parser &, size_t &) const noexcept final {
    return UNSUPPORTED_ARCHITECTURE;
  }

  unsupported_implementation() : implementation("unsupported", "Unsupported CPU (no detected SIMD instructions)", 0) {}
};

const unsupported_implementation unsupported_singleton{};

size_t available_implementation_list::size() const noexcept {
  return internal::available_implementation_pointers.size();
}
const implementation * const *available_implementation_list::begin() const noexcept {
  return internal::available_implementation_pointers.begin();
}
const implementation * const *available_implementation_list::end() const noexcept {
  return internal::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 = detect_supported_architectures();
  for (const implementation *impl : internal::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 &unsupported_singleton;
}

const implementation *detect_best_supported_implementation_on_first_use::set_best() const noexcept {
  return active_implementation = available_implementations.detect_best_supported();
}

} // namespace simdjson::internal
/* end file src/fallback/implementation.h */
/* begin file src/jsonminifier.cpp */
#include <cstdint>

#ifndef SIMDJSON_ISSUE384RESOLVED // to avoid tripping users

namespace simdjson {
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,
};

size_t json_minify(const unsigned char *bytes, size_t how_many,
                   unsigned char *out) {
  size_t i = 0, pos = 0;
  uint8_t quote = 0;
  uint8_t nonescape = 1;

  while (i < how_many) {
    unsigned char c = bytes[i];
    uint8_t *meta = jump_table + 3 * c;

    quote = quote ^ (meta[0] & nonescape);
    out[pos] = c;
    pos += meta[2] | quote;

    i += 1;
    nonescape = (~nonescape) | (meta[1]);
  }
  return pos;
}
} // namespace simdjson
#else

//
// This fast code is disabled.
// See issue https://github.com/lemire/simdjson/issues/384
//
/* begin file src/simdprune_tables.h */
#ifndef SIMDJSON_SIMDPRUNE_TABLES_H
#define SIMDJSON_SIMDPRUNE_TABLES_H
#include <cstdint>

namespace simdjson { // table modified and copied from
                     // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetTable
static 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};

static 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, 0x80, 0x00, 0x01, 0x02, 0x03,
    0x04, 0x05, 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, 0x80, 0x80,
    0x00, 0x01, 0x02, 0x03, 0x04, 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e,
    0x0f, 0x80, 0x80, 0x80, 0x00, 0x01, 0x02, 0x03, 0x08, 0x09, 0x0a, 0x0b,
    0x0c, 0x0d, 0x0e, 0x0f, 0x80, 0x80, 0x80, 0x80, 0x00, 0x01, 0x02, 0x08,
    0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, 0x80, 0x80, 0x80, 0x80, 0x80,
    0x00, 0x01, 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e, 0x0f, 0x80, 0x80,
    0x80, 0x80, 0x80, 0x80, 0x00, 0x08, 0x09, 0x0a, 0x0b, 0x0c, 0x0d, 0x0e,
    0x0f, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x08, 0x09, 0x0a, 0x0b,
    0x0c, 0x0d, 0x0e, 0x0f, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
};

// 256 * 8 bytes = 2kB, easily fits in cache.
static 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 simdjson 

#endif // SIMDJSON_SIMDPRUNE_TABLES_H
/* end file src/simdprune_tables.h */
#include <cstring>
#include <x86intrin.h> // currently, there is no runtime dispatch for the minifier

namespace simdjson {

// a straightforward comparison of a mask against input.
static uint64_t cmp_mask_against_input_mini(__m256i input_lo, __m256i input_hi,
                                            __m256i mask) {
  __m256i cmp_res_0 = _mm256_cmpeq_epi8(input_lo, mask);
  uint64_t res_0 = static_cast<uint32_t>(_mm256_movemask_epi8(cmp_res_0));
  __m256i cmp_res_1 = _mm256_cmpeq_epi8(input_hi, mask);
  uint64_t res_1 = _mm256_movemask_epi8(cmp_res_1);
  return res_0 | (res_1 << 32);
}

// Write up to 16 bytes, only the bytes corresponding to a 1-bit are written
// out. credit: Anime Tosho
static __m128i skinnycleanm128(__m128i x, int mask) {
  int mask1 = mask & 0xFF;
  int mask2 = (mask >> 8) & 0xFF;
  __m128i shufmask = _mm_castps_si128(
      _mm_loadh_pi(_mm_castsi128_ps(_mm_loadl_epi64(
                       (const __m128i *)(thintable_epi8 + mask1))),
                   (const __m64 *)(thintable_epi8 + mask2)));
  shufmask =
      _mm_add_epi8(shufmask, _mm_set_epi32(0x08080808, 0x08080808, 0, 0));
  __m128i pruned = _mm_shuffle_epi8(x, shufmask);
  intptr_t popx2 = BitsSetTable256mul2[mask1];
  __m128i compactmask =
      _mm_loadu_si128((const __m128i *)(pshufb_combine_table + popx2 * 8));
  return _mm_shuffle_epi8(pruned, compactmask);
}

// take input from buf and remove useless whitespace, input and output can be
// the same, result is null terminated, return the string length (minus the null
// termination)
size_t json_minify(const uint8_t *buf, size_t len, uint8_t *out) {
  // Useful constant masks
  const uint64_t even_bits = 0x5555555555555555ULL;
  const uint64_t odd_bits = ~even_bits;
  uint8_t *initout(out);
  uint64_t prev_iter_ends_odd_backslash =
      0ULL;                               // either 0 or 1, but a 64-bit value
  uint64_t prev_iter_inside_quote = 0ULL; // either all zeros or all ones
  size_t idx = 0;
  if (len >= 64) {
    size_t avx_len = len - 63;

    for (; idx < avx_len; idx += 64) {
      __m256i input_lo =
          _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buf + idx + 0));
      __m256i input_hi =
          _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buf + idx + 32));
      uint64_t bs_bits = cmp_mask_against_input_mini(input_lo, input_hi,
                                                     _mm256_set1_epi8('\\'));
      uint64_t start_edges = bs_bits & ~(bs_bits << 1);
      uint64_t even_start_mask = even_bits ^ prev_iter_ends_odd_backslash;
      uint64_t even_starts = start_edges & even_start_mask;
      uint64_t odd_starts = start_edges & ~even_start_mask;
      uint64_t even_carries = bs_bits + even_starts;
      uint64_t odd_carries;
      bool iter_ends_odd_backslash =
          add_overflow(bs_bits, odd_starts, &odd_carries);
      odd_carries |= prev_iter_ends_odd_backslash;
      prev_iter_ends_odd_backslash = iter_ends_odd_backslash ? 0x1ULL : 0x0ULL;
      uint64_t even_carry_ends = even_carries & ~bs_bits;
      uint64_t odd_carry_ends = odd_carries & ~bs_bits;
      uint64_t even_start_odd_end = even_carry_ends & odd_bits;
      uint64_t odd_start_even_end = odd_carry_ends & even_bits;
      uint64_t odd_ends = even_start_odd_end | odd_start_even_end;
      uint64_t quote_bits = cmp_mask_against_input_mini(input_lo, input_hi,
                                                        _mm256_set1_epi8('"'));
      quote_bits = quote_bits & ~odd_ends;
      uint64_t quote_mask = _mm_cvtsi128_si64(_mm_clmulepi64_si128(
          _mm_set_epi64x(0ULL, quote_bits), _mm_set1_epi8(0xFF), 0));
      quote_mask ^= prev_iter_inside_quote;
      prev_iter_inside_quote = static_cast<uint64_t>(
          static_cast<int64_t>(quote_mask) >>
          63); // might be undefined behavior, should be fully defined in C++20,
               // ok according to John Regher from Utah University
      const __m256i low_nibble_mask = _mm256_setr_epi8(
          //  0                           9  a   b  c  d
          16, 0, 0, 0, 0, 0, 0, 0, 0, 8, 12, 1, 2, 9, 0, 0, 16, 0, 0, 0, 0, 0,
          0, 0, 0, 8, 12, 1, 2, 9, 0, 0);
      const __m256i high_nibble_mask = _mm256_setr_epi8(
          //  0     2   3     5     7
          8, 0, 18, 4, 0, 1, 0, 1, 0, 0, 0, 3, 2, 1, 0, 0, 8, 0, 18, 4, 0, 1, 0,
          1, 0, 0, 0, 3, 2, 1, 0, 0);
      __m256i whitespace_shufti_mask = _mm256_set1_epi8(0x18);
      __m256i v_lo = _mm256_and_si256(
          _mm256_shuffle_epi8(low_nibble_mask, input_lo),
          _mm256_shuffle_epi8(high_nibble_mask,
                              _mm256_and_si256(_mm256_srli_epi32(input_lo, 4),
                                               _mm256_set1_epi8(0x7f))));

      __m256i v_hi = _mm256_and_si256(
          _mm256_shuffle_epi8(low_nibble_mask, input_hi),
          _mm256_shuffle_epi8(high_nibble_mask,
                              _mm256_and_si256(_mm256_srli_epi32(input_hi, 4),
                                               _mm256_set1_epi8(0x7f))));
      __m256i tmp_ws_lo = _mm256_cmpeq_epi8(
          _mm256_and_si256(v_lo, whitespace_shufti_mask), _mm256_set1_epi8(0));
      __m256i tmp_ws_hi = _mm256_cmpeq_epi8(
          _mm256_and_si256(v_hi, whitespace_shufti_mask), _mm256_set1_epi8(0));

      uint64_t ws_res_0 =
          static_cast<uint32_t>(_mm256_movemask_epi8(tmp_ws_lo));
      uint64_t ws_res_1 = _mm256_movemask_epi8(tmp_ws_hi);
      uint64_t whitespace = ~(ws_res_0 | (ws_res_1 << 32));
      whitespace &= ~quote_mask;

      uint64_t non_whitespace = ~whitespace;

      __m128i x1 = _mm256_extracti128_si256(input_lo, 0);
      __m128i x2 = _mm256_extracti128_si256(input_lo, 1);
      __m128i x3 = _mm256_extracti128_si256(input_hi, 0);
      __m128i x4 = _mm256_extracti128_si256(input_hi, 1);

      int mask1 = non_whitespace & 0xFFFF;
      int mask2 = (non_whitespace >> 16) & 0xFFFF;
      int mask3 = (non_whitespace >> 32) & 0xFFFF;
      int mask4 = (non_whitespace >> 48) & 0xFFFF;

      x1 = skinnycleanm128(x1, mask1);
      x2 = skinnycleanm128(x2, mask2);
      x3 = skinnycleanm128(x3, mask3);
      x4 = skinnycleanm128(x4, mask4);
      int pop1 = hamming(non_whitespace & 0xFFFF);
      int pop2 = hamming(non_whitespace & UINT64_C(0xFFFFFFFF));
      int pop3 = hamming(non_whitespace & UINT64_C(0xFFFFFFFFFFFF));
      int pop4 = hamming(non_whitespace);
      _mm_storeu_si128(reinterpret_cast<__m128i *>(out), x1);
      _mm_storeu_si128(reinterpret_cast<__m128i *>(out + pop1), x2);
      _mm_storeu_si128(reinterpret_cast<__m128i *>(out + pop2), x3);
      _mm_storeu_si128(reinterpret_cast<__m128i *>(out + pop3), x4);
      out += pop4;
    }
  }
  // we finish off the job... copying and pasting the code is not ideal here,
  // but it gets the job done.
  if (idx < len) {
    uint8_t buffer[64];
    memset(buffer, 0, 64);
    memcpy(buffer, buf + idx, len - idx);
    __m256i input_lo =
        _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buffer));
    __m256i input_hi =
        _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buffer + 32));
    uint64_t bs_bits =
        cmp_mask_against_input_mini(input_lo, input_hi, _mm256_set1_epi8('\\'));
    uint64_t start_edges = bs_bits & ~(bs_bits << 1);
    uint64_t even_start_mask = even_bits ^ prev_iter_ends_odd_backslash;
    uint64_t even_starts = start_edges & even_start_mask;
    uint64_t odd_starts = start_edges & ~even_start_mask;
    uint64_t even_carries = bs_bits + even_starts;
    uint64_t odd_carries;
    // bool iter_ends_odd_backslash =
    add_overflow(bs_bits, odd_starts, &odd_carries);
    odd_carries |= prev_iter_ends_odd_backslash;
    // prev_iter_ends_odd_backslash = iter_ends_odd_backslash ? 0x1ULL : 0x0ULL;
    // // we never use it
    uint64_t even_carry_ends = even_carries & ~bs_bits;
    uint64_t odd_carry_ends = odd_carries & ~bs_bits;
    uint64_t even_start_odd_end = even_carry_ends & odd_bits;
    uint64_t odd_start_even_end = odd_carry_ends & even_bits;
    uint64_t odd_ends = even_start_odd_end | odd_start_even_end;
    uint64_t quote_bits =
        cmp_mask_against_input_mini(input_lo, input_hi, _mm256_set1_epi8('"'));
    quote_bits = quote_bits & ~odd_ends;
    uint64_t quote_mask = _mm_cvtsi128_si64(_mm_clmulepi64_si128(
        _mm_set_epi64x(0ULL, quote_bits), _mm_set1_epi8(0xFF), 0));
    quote_mask ^= prev_iter_inside_quote;
    // prev_iter_inside_quote = (uint64_t)((int64_t)quote_mask >> 63);// we
    // don't need this anymore

    __m256i mask_20 = _mm256_set1_epi8(0x20); // c==32
    __m256i mask_70 =
        _mm256_set1_epi8(0x70); // adding 0x70 does not check low 4-bits
    // but moves any value >= 16 above 128

    __m256i lut_cntrl = _mm256_setr_epi8(
        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xFF, 0xFF, 0x00,
        0x00, 0xFF, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
        0x00, 0xFF, 0xFF, 0x00, 0x00, 0xFF, 0x00, 0x00);

    __m256i tmp_ws_lo = _mm256_or_si256(
        _mm256_cmpeq_epi8(mask_20, input_lo),
        _mm256_shuffle_epi8(lut_cntrl, _mm256_adds_epu8(mask_70, input_lo)));
    __m256i tmp_ws_hi = _mm256_or_si256(
        _mm256_cmpeq_epi8(mask_20, input_hi),
        _mm256_shuffle_epi8(lut_cntrl, _mm256_adds_epu8(mask_70, input_hi)));
    uint64_t ws_res_0 = static_cast<uint32_t>(_mm256_movemask_epi8(tmp_ws_lo));
    uint64_t ws_res_1 = _mm256_movemask_epi8(tmp_ws_hi);
    uint64_t whitespace = (ws_res_0 | (ws_res_1 << 32));
    whitespace &= ~quote_mask;

    if (len - idx < 64) {
      whitespace |= UINT64_C(0xFFFFFFFFFFFFFFFF) << (len - idx);
    }
    uint64_t non_whitespace = ~whitespace;

    int mask1 = non_whitespace & 0xFFFF;
    int mask2 = (non_whitespace >> 16) & 0xFFFF;
    int mask3 = (non_whitespace >> 32) & 0xFFFF;
    int mask4 = (non_whitespace >> 48) & 0xFFFF;

    x1 = skinnycleanm128(x1, mask1);
    x2 = skinnycleanm128(x2, mask2);
    x3 = skinnycleanm128(x3, mask3);
    x4 = skinnycleanm128(x4, mask4);
    int pop1 = hamming(non_whitespace & 0xFFFF);
    int pop2 = hamming(non_whitespace & UINT64_C(0xFFFFFFFF));
    int pop3 = hamming(non_whitespace & UINT64_C(0xFFFFFFFFFFFF));
    int pop4 = hamming(non_whitespace);
    _mm_storeu_si128(reinterpret_cast<__m128i *>(out), x1);
    _mm_storeu_si128(reinterpret_cast<__m128i *>(out + pop1), x2);
    _mm_storeu_si128(reinterpret_cast<__m128i *>(out + pop2), x3);
    _mm_storeu_si128(reinterpret_cast<__m128i *>(out + pop3), x4);
    out += pop4;
  }
  *out = '\0'; // NULL termination
  return out - initout;
}

size_t oldjson_minify(const uint8_t *buf, size_t len, uint8_t *out) {
  // Useful constant masks
  const uint64_t even_bits = 0x5555555555555555ULL;
  const uint64_t odd_bits = ~even_bits;
  uint8_t *initout(out);
  uint64_t prev_iter_ends_odd_backslash =
      0ULL;                               // either 0 or 1, but a 64-bit value
  uint64_t prev_iter_inside_quote = 0ULL; // either all zeros or all ones
  size_t idx = 0;
  if (len >= 64) {
    size_t avx_len = len - 63;

    for (; idx < avx_len; idx += 64) {
      __m256i input_lo =
          _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buf + idx + 0));
      __m256i input_hi =
          _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buf + idx + 32));
      uint64_t bs_bits = cmp_mask_against_input_mini(input_lo, input_hi,
                                                     _mm256_set1_epi8('\\'));
      uint64_t start_edges = bs_bits & ~(bs_bits << 1);
      uint64_t even_start_mask = even_bits ^ prev_iter_ends_odd_backslash;
      uint64_t even_starts = start_edges & even_start_mask;
      uint64_t odd_starts = start_edges & ~even_start_mask;
      uint64_t even_carries = bs_bits + even_starts;
      uint64_t odd_carries;
      bool iter_ends_odd_backslash =
          add_overflow(bs_bits, odd_starts, &odd_carries);
      odd_carries |= prev_iter_ends_odd_backslash;
      prev_iter_ends_odd_backslash = iter_ends_odd_backslash ? 0x1ULL : 0x0ULL;
      uint64_t even_carry_ends = even_carries & ~bs_bits;
      uint64_t odd_carry_ends = odd_carries & ~bs_bits;
      uint64_t even_start_odd_end = even_carry_ends & odd_bits;
      uint64_t odd_start_even_end = odd_carry_ends & even_bits;
      uint64_t odd_ends = even_start_odd_end | odd_start_even_end;
      uint64_t quote_bits = cmp_mask_against_input_mini(input_lo, input_hi,
                                                        _mm256_set1_epi8('"'));
      quote_bits = quote_bits & ~odd_ends;
      uint64_t quote_mask = _mm_cvtsi128_si64(_mm_clmulepi64_si128(
          _mm_set_epi64x(0ULL, quote_bits), _mm_set1_epi8(0xFF), 0));
      quote_mask ^= prev_iter_inside_quote;
      prev_iter_inside_quote = static_cast<uint64_t>(
          static_cast<int64_t>(quote_mask) >>
          63); // might be undefined behavior, should be fully defined in C++20,
               // ok according to John Regher from Utah University
      const __m256i low_nibble_mask = _mm256_setr_epi8(
          //  0                           9  a   b  c  d
          16, 0, 0, 0, 0, 0, 0, 0, 0, 8, 12, 1, 2, 9, 0, 0, 16, 0, 0, 0, 0, 0,
          0, 0, 0, 8, 12, 1, 2, 9, 0, 0);
      const __m256i high_nibble_mask = _mm256_setr_epi8(
          //  0     2   3     5     7
          8, 0, 18, 4, 0, 1, 0, 1, 0, 0, 0, 3, 2, 1, 0, 0, 8, 0, 18, 4, 0, 1, 0,
          1, 0, 0, 0, 3, 2, 1, 0, 0);
      __m256i whitespace_shufti_mask = _mm256_set1_epi8(0x18);
      __m256i v_lo = _mm256_and_si256(
          _mm256_shuffle_epi8(low_nibble_mask, input_lo),
          _mm256_shuffle_epi8(high_nibble_mask,
                              _mm256_and_si256(_mm256_srli_epi32(input_lo, 4),
                                               _mm256_set1_epi8(0x7f))));

      __m256i v_hi = _mm256_and_si256(
          _mm256_shuffle_epi8(low_nibble_mask, input_hi),
          _mm256_shuffle_epi8(high_nibble_mask,
                              _mm256_and_si256(_mm256_srli_epi32(input_hi, 4),
                                               _mm256_set1_epi8(0x7f))));
      __m256i tmp_ws_lo = _mm256_cmpeq_epi8(
          _mm256_and_si256(v_lo, whitespace_shufti_mask), _mm256_set1_epi8(0));
      __m256i tmp_ws_hi = _mm256_cmpeq_epi8(
          _mm256_and_si256(v_hi, whitespace_shufti_mask), _mm256_set1_epi8(0));

      uint64_t ws_res_0 =
          static_cast<uint32_t>(_mm256_movemask_epi8(tmp_ws_lo));
      uint64_t ws_res_1 = _mm256_movemask_epi8(tmp_ws_hi);
      uint64_t whitespace = ~(ws_res_0 | (ws_res_1 << 32));
      whitespace &= ~quote_mask;
      int mask1 = whitespace & 0xFFFF;
      int mask2 = (whitespace >> 16) & 0xFFFF;
      int mask3 = (whitespace >> 32) & 0xFFFF;
      int mask4 = (whitespace >> 48) & 0xFFFF;
      int pop1 = hamming((~whitespace) & 0xFFFF);
      int pop2 = hamming((~whitespace) & UINT64_C(0xFFFFFFFF));
      int pop3 = hamming((~whitespace) & UINT64_C(0xFFFFFFFFFFFF));
      int pop4 = hamming((~whitespace));
      __m128i x1 = _mm256_extracti128_si256(input_lo, 0);
      __m128i x2 = _mm256_extracti128_si256(input_lo, 1);
      __m128i x3 = _mm256_extracti128_si256(input_hi, 0);
      __m128i x4 = _mm256_extracti128_si256(input_hi, 1);
      x1 = skinnycleanm128(x1, mask1);
      x2 = skinnycleanm128(x2, mask2);
      x3 = skinnycleanm128(x3, mask3);
      x4 = skinnycleanm128(x4, mask4);
      _mm_storeu_si128(reinterpret_cast<__m128i *>(out), x1);
      _mm_storeu_si128(reinterpret_cast<__m128i *>(out + pop1), x2);
      _mm_storeu_si128(reinterpret_cast<__m128i *>(out + pop2), x3);
      _mm_storeu_si128(reinterpret_cast<__m128i *>(out + pop3), x4);
      out += pop4;
    }
  }
  // we finish off the job... copying and pasting the code is not ideal here,
  // but it gets the job done.
  if (idx < len) {
    uint8_t buffer[64];
    memset(buffer, 0, 64);
    memcpy(buffer, buf + idx, len - idx);
    __m256i input_lo =
        _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buffer));
    __m256i input_hi =
        _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buffer + 32));
    uint64_t bs_bits =
        cmp_mask_against_input_mini(input_lo, input_hi, _mm256_set1_epi8('\\'));
    uint64_t start_edges = bs_bits & ~(bs_bits << 1);
    uint64_t even_start_mask = even_bits ^ prev_iter_ends_odd_backslash;
    uint64_t even_starts = start_edges & even_start_mask;
    uint64_t odd_starts = start_edges & ~even_start_mask;
    uint64_t even_carries = bs_bits + even_starts;
    uint64_t odd_carries;
    // bool iter_ends_odd_backslash =
    add_overflow(bs_bits, odd_starts, &odd_carries);
    odd_carries |= prev_iter_ends_odd_backslash;
    // prev_iter_ends_odd_backslash = iter_ends_odd_backslash ? 0x1ULL : 0x0ULL;
    // // we never use it
    uint64_t even_carry_ends = even_carries & ~bs_bits;
    uint64_t odd_carry_ends = odd_carries & ~bs_bits;
    uint64_t even_start_odd_end = even_carry_ends & odd_bits;
    uint64_t odd_start_even_end = odd_carry_ends & even_bits;
    uint64_t odd_ends = even_start_odd_end | odd_start_even_end;
    uint64_t quote_bits =
        cmp_mask_against_input_mini(input_lo, input_hi, _mm256_set1_epi8('"'));
    quote_bits = quote_bits & ~odd_ends;
    uint64_t quote_mask = _mm_cvtsi128_si64(_mm_clmulepi64_si128(
        _mm_set_epi64x(0ULL, quote_bits), _mm_set1_epi8(0xFF), 0));
    quote_mask ^= prev_iter_inside_quote;
    // prev_iter_inside_quote = (uint64_t)((int64_t)quote_mask >> 63);// we
    // don't need this anymore

    __m256i mask_20 = _mm256_set1_epi8(0x20); // c==32
    __m256i mask_70 =
        _mm256_set1_epi8(0x70); // adding 0x70 does not check low 4-bits
    // but moves any value >= 16 above 128

    __m256i lut_cntrl = _mm256_setr_epi8(
        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0xFF, 0xFF, 0x00,
        0x00, 0xFF, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00,
        0x00, 0xFF, 0xFF, 0x00, 0x00, 0xFF, 0x00, 0x00);

    __m256i tmp_ws_lo = _mm256_or_si256(
        _mm256_cmpeq_epi8(mask_20, input_lo),
        _mm256_shuffle_epi8(lut_cntrl, _mm256_adds_epu8(mask_70, input_lo)));
    __m256i tmp_ws_hi = _mm256_or_si256(
        _mm256_cmpeq_epi8(mask_20, input_hi),
        _mm256_shuffle_epi8(lut_cntrl, _mm256_adds_epu8(mask_70, input_hi)));
    uint64_t ws_res_0 = static_cast<uint32_t>(_mm256_movemask_epi8(tmp_ws_lo));
    uint64_t ws_res_1 = _mm256_movemask_epi8(tmp_ws_hi);
    uint64_t whitespace = (ws_res_0 | (ws_res_1 << 32));
    whitespace &= ~quote_mask;

    if (len - idx < 64) {
      whitespace |= UINT64_C(0xFFFFFFFFFFFFFFFF) << (len - idx);
    }
    int mask1 = whitespace & 0xFFFF;
    int mask2 = (whitespace >> 16) & 0xFFFF;
    int mask3 = (whitespace >> 32) & 0xFFFF;
    int mask4 = (whitespace >> 48) & 0xFFFF;
    int pop1 = hamming((~whitespace) & 0xFFFF);
    int pop2 = hamming((~whitespace) & UINT64_C(0xFFFFFFFF));
    int pop3 = hamming((~whitespace) & UINT64_C(0xFFFFFFFFFFFF));
    int pop4 = hamming((~whitespace));
    __m128i x1 = _mm256_extracti128_si256(input_lo, 0);
    __m128i x2 = _mm256_extracti128_si256(input_lo, 1);
    __m128i x3 = _mm256_extracti128_si256(input_hi, 0);
    __m128i x4 = _mm256_extracti128_si256(input_hi, 1);
    x1 = skinnycleanm128(x1, mask1);
    x2 = skinnycleanm128(x2, mask2);
    x3 = skinnycleanm128(x3, mask3);
    x4 = skinnycleanm128(x4, mask4);
    _mm_storeu_si128(reinterpret_cast<__m128i *>(buffer), x1);
    _mm_storeu_si128(reinterpret_cast<__m128i *>(buffer + pop1), x2);
    _mm_storeu_si128(reinterpret_cast<__m128i *>(buffer + pop2), x3);
    _mm_storeu_si128(reinterpret_cast<__m128i *>(buffer + pop3), x4);
    memcpy(out, buffer, pop4);
    out += pop4;
  }
  *out = '\0'; // NULL termination
  return out - initout;
}

} // namespace simdjson
#endif
/* end file src/simdprune_tables.h */
/* begin file src/stage1_find_marks.cpp */
#if SIMDJSON_IMPLEMENTATION_ARM64
/* begin file src/arm64/stage1_find_marks.h */
#ifndef SIMDJSON_ARM64_STAGE1_FIND_MARKS_H
#define SIMDJSON_ARM64_STAGE1_FIND_MARKS_H

/* begin file src/arm64/bitmask.h */
#ifndef SIMDJSON_ARM64_BITMASK_H
#define SIMDJSON_ARM64_BITMASK_H


/* begin file src/arm64/intrinsics.h */
#ifndef SIMDJSON_ARM64_INTRINSICS_H
#define SIMDJSON_ARM64_INTRINSICS_H


// This should be the correct header whether
// you use visual studio or other compilers.
#include <arm_neon.h>

#endif //  SIMDJSON_ARM64_INTRINSICS_H
/* end file src/arm64/intrinsics.h */

namespace simdjson::arm64 {

//
// Perform a "cumulative bitwise xor," flipping bits each time a 1 is encountered.
//
// For example, prefix_xor(00100100) == 00011100
//
really_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;