xref: /third_party/node/deps/simdutf/simdutf.cpp

/* auto-generated on 2023-06-05 08:58:28 -0400. Do not edit! */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf.cpp
/* begin file src/simdutf.cpp */
#include "simdutf.h"
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=implementation.cpp
/* begin file src/implementation.cpp */
#include <initializer_list>
#include <climits>

// Useful for debugging purposes
namespace simdutf {
namespace {

template <typename T>
std::string toBinaryString(T b) {
   std::string binary = "";
   T mask = T(1) << (sizeof(T) * CHAR_BIT - 1);
   while (mask > 0) {
    binary += ((b & mask) == 0) ? '0' : '1';
    mask >>= 1;
  }
  return binary;
}
}
}

// Implementations
// The best choice should always come first!
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/arm64.h
/* begin file src/simdutf/arm64.h */
#ifndef SIMDUTF_ARM64_H
#define SIMDUTF_ARM64_H

#ifdef SIMDUTF_FALLBACK_H
#error "arm64.h must be included before fallback.h"
#endif


#ifndef SIMDUTF_IMPLEMENTATION_ARM64
#define SIMDUTF_IMPLEMENTATION_ARM64 (SIMDUTF_IS_ARM64)
#endif
#define SIMDUTF_CAN_ALWAYS_RUN_ARM64 SIMDUTF_IMPLEMENTATION_ARM64 && SIMDUTF_IS_ARM64



#if SIMDUTF_IMPLEMENTATION_ARM64

namespace simdutf {
/**
 * Implementation for NEON (ARMv8).
 */
namespace arm64 {
} // namespace arm64
} // namespace simdutf

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/arm64/implementation.h
/* begin file src/simdutf/arm64/implementation.h */
#ifndef SIMDUTF_ARM64_IMPLEMENTATION_H
#define SIMDUTF_ARM64_IMPLEMENTATION_H


namespace simdutf {
namespace arm64 {

namespace {
using namespace simdutf;
}

class implementation final : public simdutf::implementation {
public:
  simdutf_really_inline implementation() : simdutf::implementation("arm64", "ARM NEON", internal::instruction_set::NEON) {}
  simdutf_warn_unused int detect_encodings(const char * input, size_t length) const noexcept final;
  simdutf_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_ascii(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf16le(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf16be(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf16le_with_errors(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf16be_with_errors(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf32(const char32_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf32_with_errors(const char32_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf16le(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf16be(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(const char * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(const char * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf32(const char * buf, size_t len, char32_t* utf32_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(const char * buf, size_t len, char32_t* utf32_output) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(const char * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16le_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16be_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf8(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf16le(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf16be(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf16le(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf16be(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16le_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16be_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16le_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16be_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  void change_endianness_utf16(const char16_t * buf, size_t length, char16_t * output) const noexcept final;
  simdutf_warn_unused size_t count_utf16le(const char16_t * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t count_utf16be(const char16_t * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t count_utf8(const char * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf16le(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf16be(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf16le(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf16be(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf16_length_from_utf8(const char * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf32(const char32_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf16_length_from_utf32(const char32_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf8(const char * input, size_t length) const noexcept;
};

} // namespace arm64
} // namespace simdutf

#endif // SIMDUTF_ARM64_IMPLEMENTATION_H
/* end file src/simdutf/arm64/implementation.h */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/arm64/begin.h
/* begin file src/simdutf/arm64/begin.h */
// redefining SIMDUTF_IMPLEMENTATION to "arm64"
// #define SIMDUTF_IMPLEMENTATION arm64
/* end file src/simdutf/arm64/begin.h */

// Declarations
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/arm64/intrinsics.h
/* begin file src/simdutf/arm64/intrinsics.h */
#ifndef SIMDUTF_ARM64_INTRINSICS_H
#define SIMDUTF_ARM64_INTRINSICS_H


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

#endif //  SIMDUTF_ARM64_INTRINSICS_H
/* end file src/simdutf/arm64/intrinsics.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/arm64/bitmanipulation.h
/* begin file src/simdutf/arm64/bitmanipulation.h */
#ifndef SIMDUTF_ARM64_BITMANIPULATION_H
#define SIMDUTF_ARM64_BITMANIPULATION_H

namespace simdutf {
namespace arm64 {
namespace {

/* result might be undefined when input_num is zero */
simdutf_really_inline int count_ones(uint64_t input_num) {
   return vaddv_u8(vcnt_u8(vcreate_u8(input_num)));
}

} // unnamed namespace
} // namespace arm64
} // namespace simdutf

#endif // SIMDUTF_ARM64_BITMANIPULATION_H
/* end file src/simdutf/arm64/bitmanipulation.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/arm64/simd.h
/* begin file src/simdutf/arm64/simd.h */
#ifndef SIMDUTF_ARM64_SIMD_H
#define SIMDUTF_ARM64_SIMD_H

#include <type_traits>


namespace simdutf {
namespace arm64 {
namespace {
namespace simd {

#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
namespace {
// Start of private section with Visual Studio workaround


/**
 * make_uint8x16_t initializes a SIMD register (uint8x16_t).
 * This is needed because, incredibly, the syntax uint8x16_t x = {1,2,3...}
 * is not recognized under Visual Studio! This is a workaround.
 * Using a std::initializer_list<uint8_t>  as a parameter resulted in
 * inefficient code. With the current approach, if the parameters are
 * compile-time constants,
 * GNU GCC compiles it to ldr, the same as uint8x16_t x = {1,2,3...}.
 * You should not use this function except for compile-time constants:
 * it is not efficient.
 */
simdutf_really_inline uint8x16_t make_uint8x16_t(uint8_t x1,  uint8_t x2,  uint8_t x3,  uint8_t x4,
                                         uint8_t x5,  uint8_t x6,  uint8_t x7,  uint8_t x8,
                                         uint8_t x9,  uint8_t x10, uint8_t x11, uint8_t x12,
                                         uint8_t x13, uint8_t x14, uint8_t x15, uint8_t x16) {
  // Doing a load like so end ups generating worse code.
  // uint8_t array[16] = {x1, x2, x3, x4, x5, x6, x7, x8,
  //                     x9, x10,x11,x12,x13,x14,x15,x16};
  // return vld1q_u8(array);
  uint8x16_t x{};
  // incredibly, Visual Studio does not allow x[0] = x1
  x = vsetq_lane_u8(x1, x, 0);
  x = vsetq_lane_u8(x2, x, 1);
  x = vsetq_lane_u8(x3, x, 2);
  x = vsetq_lane_u8(x4, x, 3);
  x = vsetq_lane_u8(x5, x, 4);
  x = vsetq_lane_u8(x6, x, 5);
  x = vsetq_lane_u8(x7, x, 6);
  x = vsetq_lane_u8(x8, x, 7);
  x = vsetq_lane_u8(x9, x, 8);
  x = vsetq_lane_u8(x10, x, 9);
  x = vsetq_lane_u8(x11, x, 10);
  x = vsetq_lane_u8(x12, x, 11);
  x = vsetq_lane_u8(x13, x, 12);
  x = vsetq_lane_u8(x14, x, 13);
  x = vsetq_lane_u8(x15, x, 14);
  x = vsetq_lane_u8(x16, x, 15);
  return x;
}

// We have to do the same work for make_int8x16_t
simdutf_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x3,  int8_t x4,
                                       int8_t x5,  int8_t x6,  int8_t x7,  int8_t x8,
                                       int8_t x9,  int8_t x10, int8_t x11, int8_t x12,
                                       int8_t x13, int8_t x14, int8_t x15, int8_t x16) {
  // Doing a load like so end ups generating worse code.
  // int8_t array[16] = {x1, x2, x3, x4, x5, x6, x7, x8,
  //                     x9, x10,x11,x12,x13,x14,x15,x16};
  // return vld1q_s8(array);
  int8x16_t x{};
  // incredibly, Visual Studio does not allow x[0] = x1
  x = vsetq_lane_s8(x1, x, 0);
  x = vsetq_lane_s8(x2, x, 1);
  x = vsetq_lane_s8(x3, x, 2);
  x = vsetq_lane_s8(x4, x, 3);
  x = vsetq_lane_s8(x5, x, 4);
  x = vsetq_lane_s8(x6, x, 5);
  x = vsetq_lane_s8(x7, x, 6);
  x = vsetq_lane_s8(x8, x, 7);
  x = vsetq_lane_s8(x9, x, 8);
  x = vsetq_lane_s8(x10, x, 9);
  x = vsetq_lane_s8(x11, x, 10);
  x = vsetq_lane_s8(x12, x, 11);
  x = vsetq_lane_s8(x13, x, 12);
  x = vsetq_lane_s8(x14, x, 13);
  x = vsetq_lane_s8(x15, x, 14);
  x = vsetq_lane_s8(x16, x, 15);
  return x;
}

simdutf_really_inline uint8x8_t make_uint8x8_t(uint8_t x1,  uint8_t x2,  uint8_t x3,  uint8_t x4,
                                         uint8_t x5,  uint8_t x6,  uint8_t x7,  uint8_t x8) {
  uint8x8_t x{};
  x = vset_lane_u8(x1, x, 0);
  x = vset_lane_u8(x2, x, 1);
  x = vset_lane_u8(x3, x, 2);
  x = vset_lane_u8(x4, x, 3);
  x = vset_lane_u8(x5, x, 4);
  x = vset_lane_u8(x6, x, 5);
  x = vset_lane_u8(x7, x, 6);
  x = vset_lane_u8(x8, x, 7);
  return x;
}

simdutf_really_inline uint16x8_t make_uint16x8_t(uint16_t x1,  uint16_t x2,  uint16_t x3,  uint16_t x4,
                                       uint16_t x5,  uint16_t x6,  uint16_t x7,  uint16_t x8) {
  uint16x8_t x{};
  x = vsetq_lane_u16(x1, x, 0);
  x = vsetq_lane_u16(x2, x, 1);
  x = vsetq_lane_u16(x3, x, 2);
  x = vsetq_lane_u16(x4, x, 3);
  x = vsetq_lane_u16(x5, x, 4);
  x = vsetq_lane_u16(x6, x, 5);
  x = vsetq_lane_u16(x7, x, 6);
  x = vsetq_lane_u16(x8, x, 7);;
  return x;
}

simdutf_really_inline int16x8_t make_int16x8_t(int16_t x1,  int16_t x2,  int16_t x3,  int16_t x4,
                                       int16_t x5,  int16_t x6,  int16_t x7,  int16_t x8) {
  uint16x8_t x{};
  x = vsetq_lane_s16(x1, x, 0);
  x = vsetq_lane_s16(x2, x, 1);
  x = vsetq_lane_s16(x3, x, 2);
  x = vsetq_lane_s16(x4, x, 3);
  x = vsetq_lane_s16(x5, x, 4);
  x = vsetq_lane_s16(x6, x, 5);
  x = vsetq_lane_s16(x7, x, 6);
  x = vsetq_lane_s16(x8, x, 7);;
  return x;
}


// End of private section with Visual Studio workaround
} // namespace
#endif // SIMDUTF_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
    simdutf_really_inline base_u8(const uint8x16_t _value) : value(_value) {}
    simdutf_really_inline operator const uint8x16_t&() const { return this->value; }
    simdutf_really_inline operator uint8x16_t&() { return this->value; }
    simdutf_really_inline T first() const { return vgetq_lane_u8(*this,0); }
    simdutf_really_inline T last() const { return vgetq_lane_u8(*this,15); }

    // Bit operations
    simdutf_really_inline simd8<T> operator|(const simd8<T> other) const { return vorrq_u8(*this, other); }
    simdutf_really_inline simd8<T> operator&(const simd8<T> other) const { return vandq_u8(*this, other); }
    simdutf_really_inline simd8<T> operator^(const simd8<T> other) const { return veorq_u8(*this, other); }
    simdutf_really_inline simd8<T> bit_andnot(const simd8<T> other) const { return vbicq_u8(*this, other); }
    simdutf_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
    simdutf_really_inline simd8<T>& operator|=(const simd8<T> other) { auto this_cast = static_cast<simd8<T>*>(this); *this_cast = *this_cast | other; return *this_cast; }
    simdutf_really_inline simd8<T>& operator&=(const simd8<T> other) { auto this_cast = static_cast<simd8<T>*>(this); *this_cast = *this_cast & other; return *this_cast; }
    simdutf_really_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 simdutf_really_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return vceqq_u8(lhs, rhs); }

    template<int N=1>
    simdutf_really_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 simdutf_really_inline simd8<bool> splat(bool _value) { return vmovq_n_u8(uint8_t(-(!!_value))); }

    simdutf_really_inline simd8(const uint8x16_t _value) : base_u8<bool>(_value) {}
    // False constructor
    simdutf_really_inline simd8() : simd8(vdupq_n_u8(0)) {}
    // Splat constructor
    simdutf_really_inline simd8(bool _value) : simd8(splat(_value)) {}
    simdutf_really_inline void store(uint8_t dst[16]) const { return vst1q_u8(dst, *this); }

    // 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).
    simdutf_really_inline uint32_t to_bitmask() const {
#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
      const uint8x16_t bit_mask =  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.
    // This method is expected to be faster than none() and is equivalent
    // when the vector register is the result of a comparison, with byte
    // values 0xff and 0x00.
    simdutf_really_inline uint64_t to_bitmask64() const {
      return vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(vreinterpretq_u16_u8(*this), 4)), 0);
    }

    simdutf_really_inline bool any() const { return vmaxvq_u8(*this) != 0; }
    simdutf_really_inline bool none() const { return vmaxvq_u8(*this) == 0; }
    simdutf_really_inline bool all() const { return vminvq_u8(*this) == 0xFF; }


  };

  // Unsigned bytes
  template<>
  struct simd8<uint8_t>: base_u8<uint8_t> {
    static simdutf_really_inline simd8<uint8_t> splat(uint8_t _value) { return vmovq_n_u8(_value); }
    static simdutf_really_inline simd8<uint8_t> zero() { return vdupq_n_u8(0); }
    static simdutf_really_inline simd8<uint8_t> load(const uint8_t* values) { return vld1q_u8(values); }
    simdutf_really_inline simd8(const uint8x16_t _value) : base_u8<uint8_t>(_value) {}
    // Zero constructor
    simdutf_really_inline simd8() : simd8(zero()) {}
    // Array constructor
    simdutf_really_inline simd8(const uint8_t values[16]) : simd8(load(values)) {}
    // Splat constructor
    simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
    // Member-by-member initialization
#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
    simdutf_really_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(make_uint8x16_t(
      v0, v1, v2, v3, v4, v5, v6, v7,
      v8, v9, v10,v11,v12,v13,v14,v15
    )) {}
#else
    simdutf_really_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)
    simdutf_really_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
    simdutf_really_inline void store(uint8_t dst[16]) const { return vst1q_u8(dst, *this); }

    // Saturated math
    simdutf_really_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return vqaddq_u8(*this, other); }
    simdutf_really_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
    simdutf_really_inline simd8<uint8_t> operator+(const simd8<uint8_t> other) const { return vaddq_u8(*this, other); }
    simdutf_really_inline simd8<uint8_t> operator-(const simd8<uint8_t> other) const { return vsubq_u8(*this, other); }
    simdutf_really_inline simd8<uint8_t>& operator+=(const simd8<uint8_t> other) { *this = *this + other; return *this; }
    simdutf_really_inline simd8<uint8_t>& operator-=(const simd8<uint8_t> other) { *this = *this - other; return *this; }

    // Order-specific operations
    simdutf_really_inline uint8_t max_val() const { return vmaxvq_u8(*this); }
    simdutf_really_inline uint8_t min_val() const { return vminvq_u8(*this); }
    simdutf_really_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return vmaxq_u8(*this, other); }
    simdutf_really_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return vminq_u8(*this, other); }
    simdutf_really_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return vcleq_u8(*this, other); }
    simdutf_really_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return vcgeq_u8(*this, other); }
    simdutf_really_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return vcltq_u8(*this, other); }
    simdutf_really_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.
    simdutf_really_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.
    simdutf_really_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return simd8<uint8_t>(*this < other); }

    // Bit-specific operations
    simdutf_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return vtstq_u8(*this, bits); }
    simdutf_really_inline bool is_ascii() const { return this->max_val() < 0b10000000u; }

    simdutf_really_inline bool any_bits_set_anywhere() const { return this->max_val() != 0; }
    simdutf_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return (*this & bits).any_bits_set_anywhere(); }
    template<int N>
    simdutf_really_inline simd8<uint8_t> shr() const { return vshrq_n_u8(*this, N); }
    template<int N>
    simdutf_really_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>
    simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
      return lookup_table.apply_lookup_16_to(*this);
    }


    template<typename L>
    simdutf_really_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>
    simdutf_really_inline simd8<uint8_t> apply_lookup_16_to(const simd8<T> original) const {
      return vqtbl1q_u8(*this, simd8<uint8_t>(original));
    }
  };

  // Signed bytes
  template<>
  struct simd8<int8_t> {
    int8x16_t value;

    static simdutf_really_inline simd8<int8_t> splat(int8_t _value) { return vmovq_n_s8(_value); }
    static simdutf_really_inline simd8<int8_t> zero() { return vdupq_n_s8(0); }
    static simdutf_really_inline simd8<int8_t> load(const int8_t values[16]) { return vld1q_s8(values); }
    template <endianness big_endian>
    simdutf_really_inline void store_ascii_as_utf16(char16_t * p) const {
      uint16x8_t first = vmovl_u8(vget_low_u8 (vreinterpretq_u8_s8(this->value)));
      uint16x8_t second = vmovl_high_u8(vreinterpretq_u8_s8(this->value));
      if (!match_system(big_endian)) {
        #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
        const uint8x16_t swap = make_uint8x16_t(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
        #else
        const uint8x16_t swap = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
        #endif
        first = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(first), swap));
        second = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(second), swap));
      }
      vst1q_u16(reinterpret_cast<uint16_t*>(p), first);
      vst1q_u16(reinterpret_cast<uint16_t*>(p + 8), second);
    }
    simdutf_really_inline void store_ascii_as_utf32(char32_t * p) const {
      vst1q_u32(reinterpret_cast<uint32_t*>(p), vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8 (vreinterpretq_u8_s8(this->value))))));
      vst1q_u32(reinterpret_cast<uint32_t*>(p + 4), vmovl_high_u16(vmovl_u8(vget_low_u8 (vreinterpretq_u8_s8(this->value)))));
      vst1q_u32(reinterpret_cast<uint32_t*>(p + 8), vmovl_u16(vget_low_u16(vmovl_high_u8(vreinterpretq_u8_s8(this->value)))));
      vst1q_u32(reinterpret_cast<uint32_t*>(p + 12), vmovl_high_u16(vmovl_high_u8(vreinterpretq_u8_s8(this->value))));
    }
    // Conversion from/to SIMD register
    simdutf_really_inline simd8(const int8x16_t _value) : value{_value} {}
    simdutf_really_inline operator const int8x16_t&() const { return this->value; }
    simdutf_really_inline operator const uint8x16_t() const { return vreinterpretq_u8_s8(this->value); }
    simdutf_really_inline operator int8x16_t&() { return this->value; }

    // Zero constructor
    simdutf_really_inline simd8() : simd8(zero()) {}
    // Splat constructor
    simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
    // Array constructor
    simdutf_really_inline simd8(const int8_t* values) : simd8(load(values)) {}
    // Member-by-member initialization
#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
    simdutf_really_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(make_int8x16_t(
      v0, v1, v2, v3, v4, v5, v6, v7,
      v8, v9, v10,v11,v12,v13,v14,v15
    )) {}
#else
    simdutf_really_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)
    simdutf_really_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
    simdutf_really_inline void store(int8_t dst[16]) const { return vst1q_s8(dst, value); }
    // 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 SIMDUTF_REGULAR_VISUAL_STUDIO
    simdutf_really_inline explicit simd8(const uint8x16_t other): simd8(vreinterpretq_s8_u8(other)) {}
#endif
    simdutf_really_inline operator simd8<uint8_t>() const { return vreinterpretq_u8_s8(this->value); }

    simdutf_really_inline simd8<int8_t> operator|(const simd8<int8_t> other) const { return vorrq_s8(value, other.value); }
    simdutf_really_inline simd8<int8_t> operator&(const simd8<int8_t> other) const { return vandq_s8(value, other.value); }
    simdutf_really_inline simd8<int8_t> operator^(const simd8<int8_t> other) const { return veorq_s8(value, other.value); }
    simdutf_really_inline simd8<int8_t> bit_andnot(const simd8<int8_t> other) const { return vbicq_s8(value, other.value); }

    // Math
    simdutf_really_inline simd8<int8_t> operator+(const simd8<int8_t> other) const { return vaddq_s8(value, other.value); }
    simdutf_really_inline simd8<int8_t> operator-(const simd8<int8_t> other) const { return vsubq_s8(value, other.value); }
    simdutf_really_inline simd8<int8_t>& operator+=(const simd8<int8_t> other) { *this = *this + other; return *this; }
    simdutf_really_inline simd8<int8_t>& operator-=(const simd8<int8_t> other) { *this = *this - other; return *this; }

    simdutf_really_inline int8_t max_val() const { return vmaxvq_s8(value); }
    simdutf_really_inline int8_t min_val() const { return vminvq_s8(value); }
    simdutf_really_inline bool is_ascii() const { return this->min_val() >= 0; }

    // Order-sensitive comparisons
    simdutf_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return vmaxq_s8(value, other.value); }
    simdutf_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return vminq_s8(value, other.value); }
    simdutf_really_inline simd8<bool> operator>(const simd8<int8_t> other) const { return vcgtq_s8(value, other.value); }
    simdutf_really_inline simd8<bool> operator<(const simd8<int8_t> other) const { return vcltq_s8(value, other.value); }
    simdutf_really_inline simd8<bool> operator==(const simd8<int8_t> other) const { return vceqq_s8(value, other.value); }

    template<int N=1>
    simdutf_really_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>
    simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
      return lookup_table.apply_lookup_16_to(*this);
    }
    template<typename L>
    simdutf_really_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>
    simdutf_really_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.");
    simd8<T> chunks[NUM_CHUNKS];

    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

    simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1, const simd8<T> chunk2, const simd8<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
    simdutf_really_inline simd8x64(const T* ptr) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+sizeof(simd8<T>)/sizeof(T)), simd8<T>::load(ptr+2*sizeof(simd8<T>)/sizeof(T)), simd8<T>::load(ptr+3*sizeof(simd8<T>)/sizeof(T))} {}

    simdutf_really_inline void store(T* ptr) const {
      this->chunks[0].store(ptr+sizeof(simd8<T>)*0/sizeof(T));
      this->chunks[1].store(ptr+sizeof(simd8<T>)*1/sizeof(T));
      this->chunks[2].store(ptr+sizeof(simd8<T>)*2/sizeof(T));
      this->chunks[3].store(ptr+sizeof(simd8<T>)*3/sizeof(T));
    }


    simdutf_really_inline simd8x64<T>& operator |=(const simd8x64<T> &other) {
      this->chunks[0] |= other.chunks[0];
      this->chunks[1] |= other.chunks[1];
      this->chunks[2] |= other.chunks[2];
      this->chunks[3] |= other.chunks[3];
      return *this;
    }

    simdutf_really_inline simd8<T> reduce_or() const {
      return (this->chunks[0] | this->chunks[1]) | (this->chunks[2] | this->chunks[3]);
    }

    simdutf_really_inline bool is_ascii() const {
      return reduce_or().is_ascii();
    }

    template <endianness endian>
    simdutf_really_inline void store_ascii_as_utf16(char16_t * ptr) const {
      this->chunks[0].template store_ascii_as_utf16<endian>(ptr+sizeof(simd8<T>)*0);
      this->chunks[1].template store_ascii_as_utf16<endian>(ptr+sizeof(simd8<T>)*1);
      this->chunks[2].template store_ascii_as_utf16<endian>(ptr+sizeof(simd8<T>)*2);
      this->chunks[3].template store_ascii_as_utf16<endian>(ptr+sizeof(simd8<T>)*3);
    }

    simdutf_really_inline void store_ascii_as_utf32(char32_t * ptr) const {
      this->chunks[0].store_ascii_as_utf32(ptr+sizeof(simd8<T>)*0);
      this->chunks[1].store_ascii_as_utf32(ptr+sizeof(simd8<T>)*1);
      this->chunks[2].store_ascii_as_utf32(ptr+sizeof(simd8<T>)*2);
      this->chunks[3].store_ascii_as_utf32(ptr+sizeof(simd8<T>)*3);
    }

    simdutf_really_inline uint64_t to_bitmask() const {
#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
      const uint8x16_t bit_mask = 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(vandq_u8(uint8x16_t(this->chunks[0]), bit_mask), vandq_u8(uint8x16_t(this->chunks[1]), bit_mask));
      uint8x16_t sum1 = vpaddq_u8(vandq_u8(uint8x16_t(this->chunks[2]), bit_mask), vandq_u8(uint8x16_t(this->chunks[3]), bit_mask));
      sum0 = vpaddq_u8(sum0, sum1);
      sum0 = vpaddq_u8(sum0, sum0);
      return vgetq_lane_u64(vreinterpretq_u64_u8(sum0), 0);
    }

    simdutf_really_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();
  }

  simdutf_really_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();
  }

    simdutf_really_inline uint64_t in_range(const T low, const T high) const {
      const simd8<T> mask_low = simd8<T>::splat(low);
      const simd8<T> mask_high = simd8<T>::splat(high);

      return  simd8x64<bool>(
        (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
        (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
        (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
        (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low)
      ).to_bitmask();
    }
    simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
      const simd8<T> mask_low = simd8<T>::splat(low);
      const simd8<T> mask_high = simd8<T>::splat(high);
      return  simd8x64<bool>(
        (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
        (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low),
        (this->chunks[2] > mask_high) | (this->chunks[2] < mask_low),
        (this->chunks[3] > mask_high) | (this->chunks[3] < mask_low)
      ).to_bitmask();
    }
    simdutf_really_inline uint64_t lt(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();
    }
    simdutf_really_inline uint64_t gt(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();
    }
    simdutf_really_inline uint64_t gteq(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();
    }
    simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
      const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
      return  simd8x64<bool>(
        simd8<uint8_t>(uint8x16_t(this->chunks[0])) >= mask,
        simd8<uint8_t>(uint8x16_t(this->chunks[1])) >= mask,
        simd8<uint8_t>(uint8x16_t(this->chunks[2])) >= mask,
        simd8<uint8_t>(uint8x16_t(this->chunks[3])) >= mask
      ).to_bitmask();
    }
  }; // struct simd8x64<T>
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/arm64/simd16-inl.h
/* begin file src/simdutf/arm64/simd16-inl.h */
template<typename T>
struct simd16;

  template<typename T, typename Mask=simd16<bool>>
  struct base_u16 {
    uint16x8_t value;
    static const int SIZE = sizeof(value);

    // Conversion from/to SIMD register
    simdutf_really_inline base_u16() = default;
    simdutf_really_inline base_u16(const uint16x8_t _value) : value(_value) {}
    simdutf_really_inline operator const uint16x8_t&() const { return this->value; }
    simdutf_really_inline operator uint16x8_t&() { return this->value; }
    // Bit operations
    simdutf_really_inline simd16<T> operator|(const simd16<T> other) const { return vorrq_u16(*this, other); }
    simdutf_really_inline simd16<T> operator&(const simd16<T> other) const { return vandq_u16(*this, other); }
    simdutf_really_inline simd16<T> operator^(const simd16<T> other) const { return veorq_u16(*this, other); }
    simdutf_really_inline simd16<T> bit_andnot(const simd16<T> other) const { return vbicq_u16(*this, other); }
    simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFu; }
    simdutf_really_inline simd16<T>& operator|=(const simd16<T> other) { auto this_cast = static_cast<simd16<T>*>(this); *this_cast = *this_cast | other; return *this_cast; }
    simdutf_really_inline simd16<T>& operator&=(const simd16<T> other) { auto this_cast = static_cast<simd16<T>*>(this); *this_cast = *this_cast & other; return *this_cast; }
    simdutf_really_inline simd16<T>& operator^=(const simd16<T> other) { auto this_cast = static_cast<simd16<T>*>(this); *this_cast = *this_cast ^ other; return *this_cast; }

    friend simdutf_really_inline Mask operator==(const simd16<T> lhs, const simd16<T> rhs) { return vceqq_u16(lhs, rhs); }

    template<int N=1>
    simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
      return vextq_u18(prev_chunk, *this, 8 - N);
    }
  };

template<typename T, typename Mask=simd16<bool>>
struct base16: base_u16<T> {
  typedef uint16_t bitmask_t;
  typedef uint32_t bitmask2_t;

  simdutf_really_inline base16() : base_u16<T>() {}
  simdutf_really_inline base16(const uint16x8_t _value) : base_u16<T>(_value) {}
  template <typename Pointer>
  simdutf_really_inline base16(const Pointer* ptr) : base16(vld1q_u16(ptr)) {}

  static const int SIZE = sizeof(base_u16<T>::value);

  template<int N=1>
  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
    return vextq_u18(prev_chunk, *this, 8 - N);
  }
};

// SIMD byte mask type (returned by things like eq and gt)
template<>
struct simd16<bool>: base16<bool> {
  static simdutf_really_inline simd16<bool> splat(bool _value) { return vmovq_n_u16(uint16_t(-(!!_value))); }

  simdutf_really_inline simd16<bool>() : base16() {}
  simdutf_really_inline simd16<bool>(const uint16x8_t _value) : base16<bool>(_value) {}
  // Splat constructor
  simdutf_really_inline simd16<bool>(bool _value) : base16<bool>(splat(_value)) {}

};

template<typename T>
struct base16_numeric: base16<T> {
  static simdutf_really_inline simd16<T> splat(T _value) { return vmovq_n_u16(_value); }
  static simdutf_really_inline simd16<T> zero() { return vdupq_n_u16(0); }
  static simdutf_really_inline simd16<T> load(const T values[8]) {
    return vld1q_u16(reinterpret_cast<const uint16_t*>(values));
  }

  simdutf_really_inline base16_numeric() : base16<T>() {}
  simdutf_really_inline base16_numeric(const uint16x8_t _value) : base16<T>(_value) {}

  // Store to array
  simdutf_really_inline void store(T dst[8]) const { return vst1q_u16(dst, *this); }

  // Override to distinguish from bool version
  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFu; }

  // Addition/subtraction are the same for signed and unsigned
  simdutf_really_inline simd16<T> operator+(const simd16<T> other) const { return vaddq_u8(*this, other); }
  simdutf_really_inline simd16<T> operator-(const simd16<T> other) const { return vsubq_u8(*this, other); }
  simdutf_really_inline simd16<T>& operator+=(const simd16<T> other) { *this = *this + other; return *static_cast<simd16<T>*>(this); }
  simdutf_really_inline simd16<T>& operator-=(const simd16<T> other) { *this = *this - other; return *static_cast<simd16<T>*>(this); }
};

// Signed words
template<>
struct simd16<int16_t> : base16_numeric<int16_t> {
  simdutf_really_inline simd16() : base16_numeric<int16_t>() {}
#ifndef SIMDUTF_REGULAR_VISUAL_STUDIO
  simdutf_really_inline simd16(const uint16x8_t _value) : base16_numeric<int16_t>(_value) {}
#endif
  simdutf_really_inline simd16(const int16x8_t _value) : base16_numeric<int16_t>(vreinterpretq_u16_s16(_value)) {}

  // Splat constructor
  simdutf_really_inline simd16(int16_t _value) : simd16(splat(_value)) {}
  // Array constructor
  simdutf_really_inline simd16(const int16_t* values) : simd16(load(values)) {}
  simdutf_really_inline simd16(const char16_t* values) : simd16(load(reinterpret_cast<const int16_t*>(values))) {}
  simdutf_really_inline operator simd16<uint16_t>() const;
  simdutf_really_inline operator const uint16x8_t&() const { return this->value; }
  simdutf_really_inline operator const int16x8_t() const { return vreinterpretq_s16_u16(this->value); }

  simdutf_really_inline int16_t max_val() const { return vmaxvq_s16(vreinterpretq_s16_u16(this->value)); }
  simdutf_really_inline int16_t min_val() const { return vminvq_s16(vreinterpretq_s16_u16(this->value)); }
  // Order-sensitive comparisons
  simdutf_really_inline simd16<int16_t> max_val(const simd16<int16_t> other) const { return vmaxq_s16(vreinterpretq_s16_u16(this->value), vreinterpretq_s16_u16(other.value)); }
  simdutf_really_inline simd16<int16_t> min_val(const simd16<int16_t> other) const { return vmaxq_s16(vreinterpretq_s16_u16(this->value), vreinterpretq_s16_u16(other.value)); }
  simdutf_really_inline simd16<bool> operator>(const simd16<int16_t> other) const { return vcgtq_s16(vreinterpretq_s16_u16(this->value), vreinterpretq_s16_u16(other.value)); }
  simdutf_really_inline simd16<bool> operator<(const simd16<int16_t> other) const { return vcltq_s16(vreinterpretq_s16_u16(this->value), vreinterpretq_s16_u16(other.value)); }
};




// Unsigned words
template<>
struct simd16<uint16_t>: base16_numeric<uint16_t>  {
  simdutf_really_inline simd16() : base16_numeric<uint16_t>() {}
  simdutf_really_inline simd16(const uint16x8_t _value) : base16_numeric<uint16_t>(_value) {}

  // Splat constructor
  simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {}
  // Array constructor
  simdutf_really_inline simd16(const uint16_t* values) : simd16(load(values)) {}
  simdutf_really_inline simd16(const char16_t* values) : simd16(load(reinterpret_cast<const uint16_t*>(values))) {}


  simdutf_really_inline int16_t max_val() const { return vmaxvq_u16(*this); }
  simdutf_really_inline int16_t min_val() const { return vminvq_u16(*this); }
  // Saturated math
  simdutf_really_inline simd16<uint16_t> saturating_add(const simd16<uint16_t> other) const { return vqaddq_u16(*this, other); }
  simdutf_really_inline simd16<uint16_t> saturating_sub(const simd16<uint16_t> other) const { return vqsubq_u16(*this, other); }

  // Order-specific operations
  simdutf_really_inline simd16<uint16_t> max_val(const simd16<uint16_t> other) const { return vmaxq_u16(*this, other); }
  simdutf_really_inline simd16<uint16_t> min_val(const simd16<uint16_t> other) const { return vminq_u16(*this, other); }
  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
  simdutf_really_inline simd16<uint16_t> gt_bits(const simd16<uint16_t> other) const { return this->saturating_sub(other); }
  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
  simdutf_really_inline simd16<uint16_t> lt_bits(const simd16<uint16_t> other) const { return other.saturating_sub(*this); }
  simdutf_really_inline simd16<bool> operator<=(const simd16<uint16_t> other) const { return vcleq_u16(*this, other); }
  simdutf_really_inline simd16<bool> operator>=(const simd16<uint16_t> other) const { return vcgeq_u16(*this, other); }
  simdutf_really_inline simd16<bool> operator>(const simd16<uint16_t> other) const { return  vcgtq_u16(*this, other); }
  simdutf_really_inline simd16<bool> operator<(const simd16<uint16_t> other) const { return vcltq_u16(*this, other); }

  // Bit-specific operations
  simdutf_really_inline simd16<bool> bits_not_set() const { return *this == uint16_t(0); }
  template<int N>
  simdutf_really_inline simd16<uint16_t> shr() const { return simd16<uint16_t>(vshrq_n_u16(*this, N)); }
  template<int N>
  simdutf_really_inline simd16<uint16_t> shl() const { return simd16<uint16_t>(vshlq_n_u16(*this, N)); }

  // logical operations
  simdutf_really_inline simd16<uint16_t> operator|(const simd16<uint16_t> other) const { return vorrq_u16(*this, other); }
  simdutf_really_inline simd16<uint16_t> operator&(const simd16<uint16_t> other) const { return vandq_u16(*this, other); }
  simdutf_really_inline simd16<uint16_t> operator^(const simd16<uint16_t> other) const { return veorq_u16(*this, other); }

  // Pack with the unsigned saturation  two uint16_t words into single uint8_t vector
  static simdutf_really_inline simd8<uint8_t> pack(const simd16<uint16_t>& v0, const simd16<uint16_t>& v1) {
    return vqmovn_high_u16(vqmovn_u16(v0), v1);
  }

  // Change the endianness
  simdutf_really_inline simd16<uint16_t> swap_bytes() const {
    #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
    const uint8x16_t swap = make_uint8x16_t(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
    #else
    const uint8x16_t swap = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
    #endif
    return vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(*this), swap));
  }
};
simdutf_really_inline simd16<int16_t>::operator simd16<uint16_t>() const { return this->value; }


  template<typename T>
  struct simd16x32 {
    static constexpr int NUM_CHUNKS = 64 / sizeof(simd16<T>);
    static_assert(NUM_CHUNKS == 4, "ARM kernel should use four registers per 64-byte block.");
    simd16<T> chunks[NUM_CHUNKS];

    simd16x32(const simd16x32<T>& o) = delete; // no copy allowed
    simd16x32<T>& operator=(const simd16<T> other) = delete; // no assignment allowed
    simd16x32() = delete; // no default constructor allowed

    simdutf_really_inline simd16x32(const simd16<T> chunk0, const simd16<T> chunk1, const simd16<T> chunk2, const simd16<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
    simdutf_really_inline simd16x32(const T* ptr) : chunks{simd16<T>::load(ptr), simd16<T>::load(ptr+sizeof(simd16<T>)/sizeof(T)), simd16<T>::load(ptr+2*sizeof(simd16<T>)/sizeof(T)), simd16<T>::load(ptr+3*sizeof(simd16<T>)/sizeof(T))} {}

    simdutf_really_inline void store(T* ptr) const {
      this->chunks[0].store(ptr+sizeof(simd16<T>)*0/sizeof(T));
      this->chunks[1].store(ptr+sizeof(simd16<T>)*1/sizeof(T));
      this->chunks[2].store(ptr+sizeof(simd16<T>)*2/sizeof(T));
      this->chunks[3].store(ptr+sizeof(simd16<T>)*3/sizeof(T));
    }

    simdutf_really_inline simd16<T> reduce_or() const {
      return (this->chunks[0] | this->chunks[1]) | (this->chunks[2] | this->chunks[3]);
    }

    simdutf_really_inline bool is_ascii() const {
      return reduce_or().is_ascii();
    }

    simdutf_really_inline void store_ascii_as_utf16(char16_t * ptr) const {
      this->chunks[0].store_ascii_as_utf16(ptr+sizeof(simd16<T>)*0);
      this->chunks[1].store_ascii_as_utf16(ptr+sizeof(simd16<T>)*1);
      this->chunks[2].store_ascii_as_utf16(ptr+sizeof(simd16<T>)*2);
      this->chunks[3].store_ascii_as_utf16(ptr+sizeof(simd16<T>)*3);
    }

    simdutf_really_inline uint64_t to_bitmask() const {
#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
      const uint8x16_t bit_mask = 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(vreinterpretq_u8_u16(this->chunks[0] & vreinterpretq_u16_u8(bit_mask)), vreinterpretq_u8_u16(this->chunks[1] & vreinterpretq_u16_u8(bit_mask)));
      uint8x16_t sum1 = vpaddq_u8(vreinterpretq_u8_u16(this->chunks[2] & vreinterpretq_u16_u8(bit_mask)), vreinterpretq_u8_u16(this->chunks[3] & vreinterpretq_u16_u8(bit_mask)));
      sum0 = vpaddq_u8(sum0, sum1);
      sum0 = vpaddq_u8(sum0, sum0);
      return vgetq_lane_u64(vreinterpretq_u64_u8(sum0), 0);
    }

    simdutf_really_inline void swap_bytes() {
      this->chunks[0] = this->chunks[0].swap_bytes();
      this->chunks[1] = this->chunks[1].swap_bytes();
      this->chunks[2] = this->chunks[2].swap_bytes();
      this->chunks[3] = this->chunks[3].swap_bytes();
    }

    simdutf_really_inline uint64_t eq(const T m) const {
    const simd16<T> mask = simd16<T>::splat(m);
    return  simd16x32<bool>(
      this->chunks[0] == mask,
      this->chunks[1] == mask,
      this->chunks[2] == mask,
      this->chunks[3] == mask
    ).to_bitmask();
  }

  simdutf_really_inline uint64_t lteq(const T m) const {
    const simd16<T> mask = simd16<T>::splat(m);
    return  simd16x32<bool>(
      this->chunks[0] <= mask,
      this->chunks[1] <= mask,
      this->chunks[2] <= mask,
      this->chunks[3] <= mask
    ).to_bitmask();
  }

    simdutf_really_inline uint64_t in_range(const T low, const T high) const {
      const simd16<T> mask_low = simd16<T>::splat(low);
      const simd16<T> mask_high = simd16<T>::splat(high);

      return  simd16x32<bool>(
        (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
        (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
        (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
        (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low)
      ).to_bitmask();
    }
    simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
      const simd16<T> mask_low = simd16<T>::splat(low);
      const simd16<T> mask_high = simd16<T>::splat(high);
      return  simd16x32<bool>(
        (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
        (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low),
        (this->chunks[2] > mask_high) | (this->chunks[2] < mask_low),
        (this->chunks[3] > mask_high) | (this->chunks[3] < mask_low)
      ).to_bitmask();
    }
    simdutf_really_inline uint64_t lt(const T m) const {
      const simd16<T> mask = simd16<T>::splat(m);
      return  simd16x32<bool>(
        this->chunks[0] < mask,
        this->chunks[1] < mask,
        this->chunks[2] < mask,
        this->chunks[3] < mask
      ).to_bitmask();
    }

  }; // struct simd16x32<T>
  template<>
  simdutf_really_inline uint64_t simd16x32<uint16_t>::not_in_range(const uint16_t low, const uint16_t high) const {
      const simd16<uint16_t> mask_low = simd16<uint16_t>::splat(low);
      const simd16<uint16_t> mask_high = simd16<uint16_t>::splat(high);
      simd16x32<uint16_t> x(
        simd16<uint16_t>((this->chunks[0] > mask_high) | (this->chunks[0] < mask_low)),
        simd16<uint16_t>((this->chunks[1] > mask_high) | (this->chunks[1] < mask_low)),
        simd16<uint16_t>((this->chunks[2] > mask_high) | (this->chunks[2] < mask_low)),
        simd16<uint16_t>((this->chunks[3] > mask_high) | (this->chunks[3] < mask_low))
      );
      return  x.to_bitmask();
    }
/* end file src/simdutf/arm64/simd16-inl.h */
} // namespace simd
} // unnamed namespace
} // namespace arm64
} // namespace simdutf

#endif // SIMDUTF_ARM64_SIMD_H
/* end file src/simdutf/arm64/simd.h */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/arm64/end.h
/* begin file src/simdutf/arm64/end.h */
/* end file src/simdutf/arm64/end.h */

#endif // SIMDUTF_IMPLEMENTATION_ARM64

#endif // SIMDUTF_ARM64_H
/* end file src/simdutf/arm64.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/icelake.h
/* begin file src/simdutf/icelake.h */
#ifndef SIMDUTF_ICELAKE_H
#define SIMDUTF_ICELAKE_H



#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 SIMDUTF_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 SIMDUTF_COMPILER_SUPPORTS_VBMI2 1
#endif
#endif

// We allow icelake on x64 as long as the compiler is known to support VBMI2.
#ifndef SIMDUTF_IMPLEMENTATION_ICELAKE
#define SIMDUTF_IMPLEMENTATION_ICELAKE ((SIMDUTF_IS_X86_64) && (SIMDUTF_COMPILER_SUPPORTS_VBMI2))
#endif

// To see why  (__BMI__) && (__LZCNT__) are not part of this next line, see
// https://github.com/simdutf/simdutf/issues/1247
#define SIMDUTF_CAN_ALWAYS_RUN_ICELAKE ((SIMDUTF_IMPLEMENTATION_ICELAKE) && (SIMDUTF_IS_X86_64) && (__AVX2__) && (SIMDUTF_HAS_AVX512F && \
                                         SIMDUTF_HAS_AVX512DQ && \
                                         SIMDUTF_HAS_AVX512VL && \
                                           SIMDUTF_HAS_AVX512VBMI2) && (!SIMDUTF_IS_32BITS))

#if SIMDUTF_IMPLEMENTATION_ICELAKE
#if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
#define SIMDUTF_TARGET_ICELAKE
#else
#define SIMDUTF_TARGET_ICELAKE SIMDUTF_TARGET_REGION("avx512f,avx512dq,avx512cd,avx512bw,avx512vbmi,avx512vbmi2,avx512vl,avx2,bmi,bmi2,pclmul,lzcnt,popcnt")
#endif

namespace simdutf {
namespace icelake {
} // namespace icelake
} // namespace simdutf



//
// These two need to be included outside SIMDUTF_TARGET_REGION
//
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/icelake/intrinsics.h
/* begin file src/simdutf/icelake/intrinsics.h */
#ifndef SIMDUTF_ICELAKE_INTRINSICS_H
#define SIMDUTF_ICELAKE_INTRINSICS_H


#ifdef SIMDUTF_VISUAL_STUDIO
// under clang within visual studio, this will include <x86intrin.h>
#include <intrin.h>  // visual studio or clang
#include <immintrin.h>
#else

#if SIMDUTF_GCC11ORMORE
// We should not get warnings while including <x86intrin.h> yet we do
// under some versions of GCC.
// If the x86intrin.h header has uninitialized values that are problematic,
// it is a GCC issue, we want to ignore these warnigns.
SIMDUTF_DISABLE_GCC_WARNING(-Wuninitialized)
#endif

#include <x86intrin.h> // elsewhere


#if SIMDUTF_GCC11ORMORE
// cancels the suppression of the -Wuninitialized
SIMDUTF_POP_DISABLE_WARNINGS
#endif

#ifndef _tzcnt_u64
#define _tzcnt_u64(x) __tzcnt_u64(x)
#endif // _tzcnt_u64
#endif // SIMDUTF_VISUAL_STUDIO

#ifdef SIMDUTF_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 simdutf, 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 <bmi2intrin.h>  // for _pext_u64, _pdep_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>
// Important: we need the AVX-512 headers:
#include <avx512fintrin.h>
#include <avx512dqintrin.h>
#include <avx512cdintrin.h>
#include <avx512bwintrin.h>
#include <avx512vlintrin.h>
#include <avx512vlbwintrin.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 // SIMDUTF_CLANG_VISUAL_STUDIO



#if defined(__GNUC__) && !defined(__clang__)

#if __GNUC__ == 8
#define SIMDUTF_GCC8 1
#elif __GNUC__ == 9
#define SIMDUTF_GCC9 1
#endif //  __GNUC__ == 8 || __GNUC__ == 9

#endif // defined(__GNUC__) && !defined(__clang__)

#if SIMDUTF_GCC8
#pragma GCC push_options
#pragma GCC target("avx512f")
/**
 * 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));
}
#pragma GCC pop_options
#endif // SIMDUTF_GCC8

#endif // SIMDUTF_HASWELL_INTRINSICS_H
/* end file src/simdutf/icelake/intrinsics.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/icelake/implementation.h
/* begin file src/simdutf/icelake/implementation.h */
#ifndef SIMDUTF_ICELAKE_IMPLEMENTATION_H
#define SIMDUTF_ICELAKE_IMPLEMENTATION_H


namespace simdutf {
namespace icelake {

namespace {
using namespace simdutf;
}

class implementation final : public simdutf::implementation {
public:
  simdutf_really_inline implementation() : simdutf::implementation(
      "icelake",
      "Intel AVX512 (AVX-512BW, AVX-512CD, AVX-512VL, AVX-512VBMI2 extensions)",
      internal::instruction_set::AVX2 | internal::instruction_set::BMI1 | internal::instruction_set::BMI2 | internal::instruction_set::AVX512BW | internal::instruction_set::AVX512CD | internal::instruction_set::AVX512VL | internal::instruction_set::AVX512VBMI2 ) {}
  simdutf_warn_unused int detect_encodings(const char * input, size_t length) const noexcept final;
  simdutf_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_ascii(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf16le(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf16be(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf16le_with_errors(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf16be_with_errors(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf32(const char32_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf32_with_errors(const char32_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf16le(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf16be(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(const char * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(const char * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf32(const char * buf, size_t len, char32_t* utf32_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(const char * buf, size_t len, char32_t* utf32_output) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(const char * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16le_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16be_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf8(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf16le(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf16be(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf16le(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf16be(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16le_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16be_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16le_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16be_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  void change_endianness_utf16(const char16_t * buf, size_t length, char16_t * output) const noexcept final;
  simdutf_warn_unused size_t count_utf16le(const char16_t * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t count_utf16be(const char16_t * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t count_utf8(const char * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf16le(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf16be(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf16le(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf16be(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf16_length_from_utf8(const char * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf32(const char32_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf16_length_from_utf32(const char32_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf8(const char * input, size_t length) const noexcept;
};

} // namespace icelake
} // namespace simdutf

#endif // SIMDUTF_ICELAKE_IMPLEMENTATION_H
/* end file src/simdutf/icelake/implementation.h */

//
// The rest need to be inside the region
//
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/icelake/begin.h
/* begin file src/simdutf/icelake/begin.h */
// redefining SIMDUTF_IMPLEMENTATION to "icelake"
// #define SIMDUTF_IMPLEMENTATION icelake

#if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
// nothing needed.
#else
SIMDUTF_TARGET_ICELAKE
#endif

#if SIMDUTF_GCC11ORMORE // workaround for https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
SIMDUTF_DISABLE_GCC_WARNING(-Wmaybe-uninitialized)
#endif // end of workaround
/* end file src/simdutf/icelake/begin.h */
// Declarations
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/icelake/bitmanipulation.h
/* begin file src/simdutf/icelake/bitmanipulation.h */
#ifndef SIMDUTF_ICELAKE_BITMANIPULATION_H
#define SIMDUTF_ICELAKE_BITMANIPULATION_H

namespace simdutf {
namespace icelake {
namespace {

#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
simdutf_really_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
simdutf_really_inline long long int count_ones(uint64_t input_num) {
  return _popcnt64(input_num);
}
#endif

} // unnamed namespace
} // namespace icelake
} // namespace simdutf

#endif // SIMDUTF_ICELAKE_BITMANIPULATION_H
/* end file src/simdutf/icelake/bitmanipulation.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/icelake/end.h
/* begin file src/simdutf/icelake/end.h */
#if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
// nothing needed.
#else
SIMDUTF_UNTARGET_REGION
#endif


#if SIMDUTF_GCC11ORMORE // workaround for https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
SIMDUTF_POP_DISABLE_WARNINGS
#endif // end of workaround
/* end file src/simdutf/icelake/end.h */



#endif // SIMDUTF_IMPLEMENTATION_ICELAKE
#endif // SIMDUTF_ICELAKE_H
/* end file src/simdutf/icelake.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/haswell.h
/* begin file src/simdutf/haswell.h */
#ifndef SIMDUTF_HASWELL_H
#define SIMDUTF_HASWELL_H

#ifdef SIMDUTF_WESTMERE_H
#error "haswell.h must be included before westmere.h"
#endif
#ifdef SIMDUTF_FALLBACK_H
#error "haswell.h must be included before fallback.h"
#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 SIMDUTF_IMPLEMENTATION_HASWELL
//
// You do not want to restrict it like so: SIMDUTF_IS_X86_64 && __AVX2__
// because we want to rely on *runtime dispatch*.
//
#if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
#define SIMDUTF_IMPLEMENTATION_HASWELL 0
#else
#define SIMDUTF_IMPLEMENTATION_HASWELL (SIMDUTF_IS_X86_64)
#endif

#endif
// To see why  (__BMI__) && (__LZCNT__) are not part of this next line, see
// https://github.com/simdutf/simdutf/issues/1247
#define SIMDUTF_CAN_ALWAYS_RUN_HASWELL ((SIMDUTF_IMPLEMENTATION_HASWELL) && (SIMDUTF_IS_X86_64) && (__AVX2__))

#if SIMDUTF_IMPLEMENTATION_HASWELL

#define SIMDUTF_TARGET_HASWELL SIMDUTF_TARGET_REGION("avx2,bmi,lzcnt,popcnt")

namespace simdutf {
/**
 * Implementation for Haswell (Intel AVX2).
 */
namespace haswell {
} // namespace haswell
} // namespace simdutf

//
// These two need to be included outside SIMDUTF_TARGET_REGION
//
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/haswell/implementation.h
/* begin file src/simdutf/haswell/implementation.h */
#ifndef SIMDUTF_HASWELL_IMPLEMENTATION_H
#define SIMDUTF_HASWELL_IMPLEMENTATION_H


// The constructor may be executed on any host, so we take care not to use SIMDUTF_TARGET_REGION
namespace simdutf {
namespace haswell {

using namespace simdutf;

class implementation final : public simdutf::implementation {
public:
  simdutf_really_inline implementation() : simdutf::implementation(
      "haswell",
      "Intel/AMD AVX2",
      internal::instruction_set::AVX2 | internal::instruction_set::BMI1 | internal::instruction_set::BMI2
  ) {}
  simdutf_warn_unused int detect_encodings(const char * input, size_t length) const noexcept final;
  simdutf_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_ascii(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf16le(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf16be(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf16le_with_errors(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf16be_with_errors(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf32(const char32_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf32_with_errors(const char32_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf16le(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf16be(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(const char * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(const char * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf32(const char * buf, size_t len, char32_t* utf32_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(const char * buf, size_t len, char32_t* utf32_output) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(const char * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16le_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16be_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf8(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf16le(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf16be(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf16le(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf16be(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16le_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16be_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16le_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16be_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  void change_endianness_utf16(const char16_t * buf, size_t length, char16_t * output) const noexcept final;
  simdutf_warn_unused size_t count_utf16le(const char16_t * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t count_utf16be(const char16_t * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t count_utf8(const char * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf16le(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf16be(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf16le(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf16be(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf16_length_from_utf8(const char * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf32(const char32_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf16_length_from_utf32(const char32_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf8(const char * input, size_t length) const noexcept;
};

} // namespace haswell
} // namespace simdutf

#endif // SIMDUTF_HASWELL_IMPLEMENTATION_H
/* end file src/simdutf/haswell/implementation.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/haswell/intrinsics.h
/* begin file src/simdutf/haswell/intrinsics.h */
#ifndef SIMDUTF_HASWELL_INTRINSICS_H
#define SIMDUTF_HASWELL_INTRINSICS_H


#ifdef SIMDUTF_VISUAL_STUDIO
// under clang within visual studio, this will include <x86intrin.h>
#include <intrin.h>  // visual studio or clang
#else

#if SIMDUTF_GCC11ORMORE
// We should not get warnings while including <x86intrin.h> yet we do
// under some versions of GCC.
// If the x86intrin.h header has uninitialized values that are problematic,
// it is a GCC issue, we want to ignore these warnigns.
SIMDUTF_DISABLE_GCC_WARNING(-Wuninitialized)
#endif

#include <x86intrin.h> // elsewhere


#if SIMDUTF_GCC11ORMORE
// cancels the suppression of the -Wuninitialized
SIMDUTF_POP_DISABLE_WARNINGS
#endif

#endif // SIMDUTF_VISUAL_STUDIO

#ifdef SIMDUTF_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 simdutf, 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>
// 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 // SIMDUTF_CLANG_VISUAL_STUDIO

#endif // SIMDUTF_HASWELL_INTRINSICS_H
/* end file src/simdutf/haswell/intrinsics.h */

//
// The rest need to be inside the region
//
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/haswell/begin.h
/* begin file src/simdutf/haswell/begin.h */
// redefining SIMDUTF_IMPLEMENTATION to "haswell"
// #define SIMDUTF_IMPLEMENTATION haswell

#if SIMDUTF_CAN_ALWAYS_RUN_HASWELL
// nothing needed.
#else
SIMDUTF_TARGET_HASWELL
#endif

#if SIMDUTF_GCC11ORMORE // workaround for https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
SIMDUTF_DISABLE_GCC_WARNING(-Wmaybe-uninitialized)
#endif // end of workaround
/* end file src/simdutf/haswell/begin.h */
// Declarations
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/haswell/bitmanipulation.h
/* begin file src/simdutf/haswell/bitmanipulation.h */
#ifndef SIMDUTF_HASWELL_BITMANIPULATION_H
#define SIMDUTF_HASWELL_BITMANIPULATION_H

namespace simdutf {
namespace haswell {
namespace {

#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
simdutf_really_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
simdutf_really_inline long long int count_ones(uint64_t input_num) {
  return _popcnt64(input_num);
}
#endif

} // unnamed namespace
} // namespace haswell
} // namespace simdutf

#endif // SIMDUTF_HASWELL_BITMANIPULATION_H
/* end file src/simdutf/haswell/bitmanipulation.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/haswell/simd.h
/* begin file src/simdutf/haswell/simd.h */
#ifndef SIMDUTF_HASWELL_SIMD_H
#define SIMDUTF_HASWELL_SIMD_H


namespace simdutf {
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
    simdutf_really_inline base() : value{__m256i()} {}

    // Conversion from SIMD register
    simdutf_really_inline base(const __m256i _value) : value(_value) {}
    // Conversion to SIMD register
    simdutf_really_inline operator const __m256i&() const { return this->value; }
    simdutf_really_inline operator __m256i&() { return this->value; }
    template <endianness big_endian>
    simdutf_really_inline void store_ascii_as_utf16(char16_t * ptr) const {
      __m256i first = _mm256_cvtepu8_epi16(_mm256_castsi256_si128(*this));
      __m256i second = _mm256_cvtepu8_epi16(_mm256_extractf128_si256(*this,1));
      if (big_endian) {
        const __m256i swap = _mm256_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14,
                                  17, 16, 19, 18, 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
        first = _mm256_shuffle_epi8(first, swap);
        second = _mm256_shuffle_epi8(second, swap);
      }
      _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr), first);
      _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 16), second);
    }
    simdutf_really_inline void store_ascii_as_utf32(char32_t * ptr) const {
      _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr), _mm256_cvtepu8_epi32(_mm256_castsi256_si128(*this)));
      _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr+8), _mm256_cvtepu8_epi32(_mm256_castsi256_si128(_mm256_srli_si256(*this,8))));
      _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 16), _mm256_cvtepu8_epi32(_mm256_extractf128_si256(*this,1)));
      _mm256_storeu_si256(reinterpret_cast<__m256i *>(ptr + 24), _mm256_cvtepu8_epi32(_mm_srli_si128(_mm256_extractf128_si256(*this,1),8)));
    }
    // Bit operations
    simdutf_really_inline Child operator|(const Child other) const { return _mm256_or_si256(*this, other); }
    simdutf_really_inline Child operator&(const Child other) const { return _mm256_and_si256(*this, other); }
    simdutf_really_inline Child operator^(const Child other) const { return _mm256_xor_si256(*this, other); }
    simdutf_really_inline Child bit_andnot(const Child other) const { return _mm256_andnot_si256(other, *this); }
    simdutf_really_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
    simdutf_really_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
    simdutf_really_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;

    simdutf_really_inline base8() : base<simd8<T>>() {}
    simdutf_really_inline base8(const __m256i _value) : base<simd8<T>>(_value) {}
    simdutf_really_inline T first() const { return _mm256_extract_epi8(*this,0); }
    simdutf_really_inline T last() const { return _mm256_extract_epi8(*this,31); }
    friend simdutf_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>
    simdutf_really_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 simdutf_really_inline simd8<bool> splat(bool _value) { return _mm256_set1_epi8(uint8_t(-(!!_value))); }

    simdutf_really_inline simd8<bool>() : base8() {}
    simdutf_really_inline simd8<bool>(const __m256i _value) : base8<bool>(_value) {}
    // Splat constructor
    simdutf_really_inline simd8<bool>(bool _value) : base8<bool>(splat(_value)) {}

    simdutf_really_inline uint32_t to_bitmask() const { return uint32_t(_mm256_movemask_epi8(*this)); }
    simdutf_really_inline bool any() const { return !_mm256_testz_si256(*this, *this); }
    simdutf_really_inline bool none() const { return _mm256_testz_si256(*this, *this); }
    simdutf_really_inline bool all() const { return static_cast<uint32_t>(_mm256_movemask_epi8(*this)) == 0xFFFFFFFF; }
    simdutf_really_inline simd8<bool> operator~() const { return *this ^ true; }
  };

  template<typename T>
  struct base8_numeric: base8<T> {
    static simdutf_really_inline simd8<T> splat(T _value) { return _mm256_set1_epi8(_value); }
    static simdutf_really_inline simd8<T> zero() { return _mm256_setzero_si256(); }
    static simdutf_really_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 simdutf_really_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
      );
    }

    simdutf_really_inline base8_numeric() : base8<T>() {}
    simdutf_really_inline base8_numeric(const __m256i _value) : base8<T>(_value) {}

    // Store to array
    simdutf_really_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
    simdutf_really_inline simd8<T> operator+(const simd8<T> other) const { return _mm256_add_epi8(*this, other); }
    simdutf_really_inline simd8<T> operator-(const simd8<T> other) const { return _mm256_sub_epi8(*this, other); }
    simdutf_really_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
    simdutf_really_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }

    // Override to distinguish from bool version
    simdutf_really_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>
    simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
      return _mm256_shuffle_epi8(lookup_table, *this);
    }

    template<typename L>
    simdutf_really_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> {
    simdutf_really_inline simd8() : base8_numeric<int8_t>() {}
    simdutf_really_inline simd8(const __m256i _value) : base8_numeric<int8_t>(_value) {}

    // Splat constructor
    simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
    // Array constructor
    simdutf_really_inline simd8(const int8_t values[32]) : simd8(load(values)) {}
    simdutf_really_inline operator simd8<uint8_t>() const;
    // Member-by-member initialization
    simdutf_really_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)
    simdutf_really_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
      );
    }
    simdutf_really_inline bool is_ascii() const { return _mm256_movemask_epi8(*this) == 0; }
    // Order-sensitive comparisons
    simdutf_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm256_max_epi8(*this, other); }
    simdutf_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm256_min_epi8(*this, other); }
    simdutf_really_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm256_cmpgt_epi8(*this, other); }
    simdutf_really_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> {
    simdutf_really_inline simd8() : base8_numeric<uint8_t>() {}
    simdutf_really_inline simd8(const __m256i _value) : base8_numeric<uint8_t>(_value) {}
    // Splat constructor
    simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
    // Array constructor
    simdutf_really_inline simd8(const uint8_t values[32]) : simd8(load(values)) {}
    // Member-by-member initialization
    simdutf_really_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)
    simdutf_really_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
    simdutf_really_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm256_adds_epu8(*this, other); }
    simdutf_really_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm256_subs_epu8(*this, other); }

    // Order-specific operations
    simdutf_really_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm256_max_epu8(*this, other); }
    simdutf_really_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)
    simdutf_really_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)
    simdutf_really_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
    simdutf_really_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
    simdutf_really_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
    simdutf_really_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
    simdutf_really_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->lt_bits(other).any_bits_set(); }

    // Bit-specific operations
    simdutf_really_inline simd8<bool> bits_not_set() const { return *this == uint8_t(0); }
    simdutf_really_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
    simdutf_really_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
    simdutf_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
    simdutf_really_inline bool is_ascii() const { return _mm256_movemask_epi8(*this) == 0; }
    simdutf_really_inline bool bits_not_set_anywhere() const { return _mm256_testz_si256(*this, *this); }
    simdutf_really_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
    simdutf_really_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return _mm256_testz_si256(*this, bits); }
    simdutf_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
    template<int N>
    simdutf_really_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm256_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
    template<int N>
    simdutf_really_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>
    simdutf_really_inline int get_bit() const { return _mm256_movemask_epi8(_mm256_slli_epi16(*this, 7-N)); }
  };
  simdutf_really_inline simd8<int8_t>::operator simd8<uint8_t>() const { return this->value; }


  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.");
    simd8<T> chunks[NUM_CHUNKS];

    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

    simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1) : chunks{chunk0, chunk1} {}
    simdutf_really_inline simd8x64(const T* ptr) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+sizeof(simd8<T>)/sizeof(T))} {}

    simdutf_really_inline void store(T* ptr) const {
      this->chunks[0].store(ptr+sizeof(simd8<T>)*0/sizeof(T));
      this->chunks[1].store(ptr+sizeof(simd8<T>)*1/sizeof(T));
    }

    simdutf_really_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);
    }

    simdutf_really_inline simd8x64<T>& operator|=(const simd8x64<T> &other) {
      this->chunks[0] |= other.chunks[0];
      this->chunks[1] |= other.chunks[1];
      return *this;
    }

    simdutf_really_inline simd8<T> reduce_or() const {
      return this->chunks[0] | this->chunks[1];
    }

    simdutf_really_inline bool is_ascii() const {
      return this->reduce_or().is_ascii();
    }

    template <endianness endian>
    simdutf_really_inline void store_ascii_as_utf16(char16_t * ptr) const {
      this->chunks[0].template store_ascii_as_utf16<endian>(ptr+sizeof(simd8<T>)*0);
      this->chunks[1].template store_ascii_as_utf16<endian>(ptr+sizeof(simd8<T>)*1);
    }

    simdutf_really_inline void store_ascii_as_utf32(char32_t * ptr) const {
      this->chunks[0].store_ascii_as_utf32(ptr+sizeof(simd8<T>)*0);
      this->chunks[1].store_ascii_as_utf32(ptr+sizeof(simd8<T>)*1);
    }

    simdutf_really_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
      );
    }

    simdutf_really_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();
    }

    simdutf_really_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();
    }

    simdutf_really_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();
    }

    simdutf_really_inline uint64_t in_range(const T low, const T high) const {
      const simd8<T> mask_low = simd8<T>::splat(low);
      const simd8<T> mask_high = simd8<T>::splat(high);

      return  simd8x64<bool>(
        (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
        (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
        (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
        (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low)
      ).to_bitmask();
    }
    simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
      const simd8<T> mask_low = simd8<T>::splat(low);
      const simd8<T> mask_high = simd8<T>::splat(high);
      return  simd8x64<bool>(
        (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
        (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low)
      ).to_bitmask();
    }
    simdutf_really_inline uint64_t lt(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();
    }

    simdutf_really_inline uint64_t gt(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();
    }
    simdutf_really_inline uint64_t gteq(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();
    }
    simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
      const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
      return  simd8x64<bool>(
        (simd8<uint8_t>(__m256i(this->chunks[0])) >= mask),
        (simd8<uint8_t>(__m256i(this->chunks[1])) >= mask)
      ).to_bitmask();
    }
  }; // struct simd8x64<T>

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/haswell/simd16-inl.h
/* begin file src/simdutf/haswell/simd16-inl.h */
#ifdef __GNUC__
#if __GNUC__ < 8
#define _mm256_set_m128i(xmm1, xmm2) _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), _mm256_castsi128_si256(xmm2), 2)
#define _mm256_setr_m128i(xmm2, xmm1)  _mm256_permute2f128_si256(_mm256_castsi128_si256(xmm1), _mm256_castsi128_si256(xmm2), 2)
#endif
#endif

template<typename T>
struct simd16;

template<typename T, typename Mask=simd16<bool>>
struct base16: base<simd16<T>> {
  using bitmask_type = uint32_t;

  simdutf_really_inline base16() : base<simd16<T>>() {}
  simdutf_really_inline base16(const __m256i _value) : base<simd16<T>>(_value) {}
  template <typename Pointer>
  simdutf_really_inline base16(const Pointer* ptr) : base16(_mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr))) {}
  friend simdutf_really_inline Mask operator==(const simd16<T> lhs, const simd16<T> rhs) { return _mm256_cmpeq_epi16(lhs, rhs); }

  /// the size of vector in bytes
  static const int SIZE = sizeof(base<simd16<T>>::value);

  /// the number of elements of type T a vector can hold
  static const int ELEMENTS = SIZE / sizeof(T);

  template<int N=1>
  simdutf_really_inline simd16<T> prev(const simd16<T> prev_chunk) const {
    return _mm256_alignr_epi8(*this, prev_chunk, 16 - N);
  }
};

// SIMD byte mask type (returned by things like eq and gt)
template<>
struct simd16<bool>: base16<bool> {
  static simdutf_really_inline simd16<bool> splat(bool _value) { return _mm256_set1_epi16(uint16_t(-(!!_value))); }

  simdutf_really_inline simd16<bool>() : base16() {}
  simdutf_really_inline simd16<bool>(const __m256i _value) : base16<bool>(_value) {}
  // Splat constructor
  simdutf_really_inline simd16<bool>(bool _value) : base16<bool>(splat(_value)) {}

  simdutf_really_inline bitmask_type to_bitmask() const { return _mm256_movemask_epi8(*this); }
  simdutf_really_inline bool any() const { return !_mm256_testz_si256(*this, *this); }
  simdutf_really_inline simd16<bool> operator~() const { return *this ^ true; }
};

template<typename T>
struct base16_numeric: base16<T> {
  static simdutf_really_inline simd16<T> splat(T _value) { return _mm256_set1_epi16(_value); }
  static simdutf_really_inline simd16<T> zero() { return _mm256_setzero_si256(); }
  static simdutf_really_inline simd16<T> load(const T values[8]) {
    return _mm256_loadu_si256(reinterpret_cast<const __m256i *>(values));
  }

  simdutf_really_inline base16_numeric() : base16<T>() {}
  simdutf_really_inline base16_numeric(const __m256i _value) : base16<T>(_value) {}

  // Store to array
  simdutf_really_inline void store(T dst[8]) const { return _mm256_storeu_si256(reinterpret_cast<__m256i *>(dst), *this); }

  // Override to distinguish from bool version
  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFFFu; }

  // Addition/subtraction are the same for signed and unsigned
  simdutf_really_inline simd16<T> operator+(const simd16<T> other) const { return _mm256_add_epi16(*this, other); }
  simdutf_really_inline simd16<T> operator-(const simd16<T> other) const { return _mm256_sub_epi16(*this, other); }
  simdutf_really_inline simd16<T>& operator+=(const simd16<T> other) { *this = *this + other; return *static_cast<simd16<T>*>(this); }
  simdutf_really_inline simd16<T>& operator-=(const simd16<T> other) { *this = *this - other; return *static_cast<simd16<T>*>(this); }
};

// Signed words
template<>
struct simd16<int16_t> : base16_numeric<int16_t> {
  simdutf_really_inline simd16() : base16_numeric<int16_t>() {}
  simdutf_really_inline simd16(const __m256i _value) : base16_numeric<int16_t>(_value) {}
  // Splat constructor
  simdutf_really_inline simd16(int16_t _value) : simd16(splat(_value)) {}
  // Array constructor
  simdutf_really_inline simd16(const int16_t* values) : simd16(load(values)) {}
  simdutf_really_inline simd16(const char16_t* values) : simd16(load(reinterpret_cast<const int16_t*>(values))) {}
  // Order-sensitive comparisons
  simdutf_really_inline simd16<int16_t> max_val(const simd16<int16_t> other) const { return _mm256_max_epi16(*this, other); }
  simdutf_really_inline simd16<int16_t> min_val(const simd16<int16_t> other) const { return _mm256_min_epi16(*this, other); }
  simdutf_really_inline simd16<bool> operator>(const simd16<int16_t> other) const { return _mm256_cmpgt_epi16(*this, other); }
  simdutf_really_inline simd16<bool> operator<(const simd16<int16_t> other) const { return _mm256_cmpgt_epi16(other, *this); }
};

// Unsigned words
template<>
struct simd16<uint16_t>: base16_numeric<uint16_t>  {
  simdutf_really_inline simd16() : base16_numeric<uint16_t>() {}
  simdutf_really_inline simd16(const __m256i _value) : base16_numeric<uint16_t>(_value) {}

  // Splat constructor
  simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {}
  // Array constructor
  simdutf_really_inline simd16(const uint16_t* values) : simd16(load(values)) {}
  simdutf_really_inline simd16(const char16_t* values) : simd16(load(reinterpret_cast<const uint16_t*>(values))) {}

  // Saturated math
  simdutf_really_inline simd16<uint16_t> saturating_add(const simd16<uint16_t> other) const { return _mm256_adds_epu16(*this, other); }
  simdutf_really_inline simd16<uint16_t> saturating_sub(const simd16<uint16_t> other) const { return _mm256_subs_epu16(*this, other); }

  // Order-specific operations
  simdutf_really_inline simd16<uint16_t> max_val(const simd16<uint16_t> other) const { return _mm256_max_epu16(*this, other); }
  simdutf_really_inline simd16<uint16_t> min_val(const simd16<uint16_t> other) const { return _mm256_min_epu16(*this, other); }
  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
  simdutf_really_inline simd16<uint16_t> gt_bits(const simd16<uint16_t> other) const { return this->saturating_sub(other); }
  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
  simdutf_really_inline simd16<uint16_t> lt_bits(const simd16<uint16_t> other) const { return other.saturating_sub(*this); }
  simdutf_really_inline simd16<bool> operator<=(const simd16<uint16_t> other) const { return other.max_val(*this) == other; }
  simdutf_really_inline simd16<bool> operator>=(const simd16<uint16_t> other) const { return other.min_val(*this) == other; }
  simdutf_really_inline simd16<bool> operator>(const simd16<uint16_t> other) const { return this->gt_bits(other).any_bits_set(); }
  simdutf_really_inline simd16<bool> operator<(const simd16<uint16_t> other) const { return this->gt_bits(other).any_bits_set(); }

  // Bit-specific operations
  simdutf_really_inline simd16<bool> bits_not_set() const { return *this == uint16_t(0); }
  simdutf_really_inline simd16<bool> bits_not_set(simd16<uint16_t> bits) const { return (*this & bits).bits_not_set(); }
  simdutf_really_inline simd16<bool> any_bits_set() const { return ~this->bits_not_set(); }
  simdutf_really_inline simd16<bool> any_bits_set(simd16<uint16_t> bits) const { return ~this->bits_not_set(bits); }

  simdutf_really_inline bool bits_not_set_anywhere() const { return _mm256_testz_si256(*this, *this); }
  simdutf_really_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
  simdutf_really_inline bool bits_not_set_anywhere(simd16<uint16_t> bits) const { return _mm256_testz_si256(*this, bits); }
  simdutf_really_inline bool any_bits_set_anywhere(simd16<uint16_t> bits) const { return !bits_not_set_anywhere(bits); }
  template<int N>
  simdutf_really_inline simd16<uint16_t> shr() const { return simd16<uint16_t>(_mm256_srli_epi16(*this, N)); }
  template<int N>
  simdutf_really_inline simd16<uint16_t> shl() const { return simd16<uint16_t>(_mm256_slli_epi16(*this, 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>
  simdutf_really_inline int get_bit() const { return _mm256_movemask_epi8(_mm256_slli_epi16(*this, 15-N)); }

  // Change the endianness
  simdutf_really_inline simd16<uint16_t> swap_bytes() const {
    const __m256i swap = _mm256_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14,
                                  17, 16, 19, 18, 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
    return _mm256_shuffle_epi8(*this, swap);
  }

  // Pack with the unsigned saturation two uint16_t words into single uint8_t vector
  static simdutf_really_inline simd8<uint8_t> pack(const simd16<uint16_t>& v0, const simd16<uint16_t>& v1) {
    // Note: the AVX2 variant of pack operates on 128-bit lanes, thus
    //       we have to shuffle lanes in order to produce bytes in the
    //       correct order.

    // get the 0th lanes
    const __m128i lo_0 = _mm256_extracti128_si256(v0, 0);
    const __m128i lo_1 = _mm256_extracti128_si256(v1, 0);

    // get the 1st lanes
    const __m128i hi_0 = _mm256_extracti128_si256(v0, 1);
    const __m128i hi_1 = _mm256_extracti128_si256(v1, 1);

    // build new vectors (shuffle lanes)
    const __m256i t0 = _mm256_set_m128i(lo_1, lo_0);
    const __m256i t1 = _mm256_set_m128i(hi_1, hi_0);

    // pack words in linear order from v0 and v1
    return _mm256_packus_epi16(t0, t1);
  }
};


  template<typename T>
  struct simd16x32 {
    static constexpr int NUM_CHUNKS = 64 / sizeof(simd16<T>);
    static_assert(NUM_CHUNKS == 2, "Haswell kernel should use two registers per 64-byte block.");
    simd16<T> chunks[NUM_CHUNKS];

    simd16x32(const simd16x32<T>& o) = delete; // no copy allowed
    simd16x32<T>& operator=(const simd16<T> other) = delete; // no assignment allowed
    simd16x32() = delete; // no default constructor allowed

    simdutf_really_inline simd16x32(const simd16<T> chunk0, const simd16<T> chunk1) : chunks{chunk0, chunk1} {}
    simdutf_really_inline simd16x32(const T* ptr) : chunks{simd16<T>::load(ptr), simd16<T>::load(ptr+sizeof(simd16<T>)/sizeof(T))} {}

    simdutf_really_inline void store(T* ptr) const {
      this->chunks[0].store(ptr+sizeof(simd16<T>)*0/sizeof(T));
      this->chunks[1].store(ptr+sizeof(simd16<T>)*1/sizeof(T));
    }

    simdutf_really_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);
    }

    simdutf_really_inline simd16<T> reduce_or() const {
      return this->chunks[0] | this->chunks[1];
    }

    simdutf_really_inline bool is_ascii() const {
      return this->reduce_or().is_ascii();
    }

    simdutf_really_inline void store_ascii_as_utf16(char16_t * ptr) const {
      this->chunks[0].store_ascii_as_utf16(ptr+sizeof(simd16<T>)*0);
      this->chunks[1].store_ascii_as_utf16(ptr+sizeof(simd16<T>));
    }

    simdutf_really_inline simd16x32<T> bit_or(const T m) const {
      const simd16<T> mask = simd16<T>::splat(m);
      return simd16x32<T>(
        this->chunks[0] | mask,
        this->chunks[1] | mask
      );
    }

    simdutf_really_inline void swap_bytes() {
      this->chunks[0] = this->chunks[0].swap_bytes();
      this->chunks[1] = this->chunks[1].swap_bytes();
    }

    simdutf_really_inline uint64_t eq(const T m) const {
      const simd16<T> mask = simd16<T>::splat(m);
      return  simd16x32<bool>(
        this->chunks[0] == mask,
        this->chunks[1] == mask
      ).to_bitmask();
    }

    simdutf_really_inline uint64_t eq(const simd16x32<uint16_t> &other) const {
      return  simd16x32<bool>(
        this->chunks[0] == other.chunks[0],
        this->chunks[1] == other.chunks[1]
      ).to_bitmask();
    }

    simdutf_really_inline uint64_t lteq(const T m) const {
      const simd16<T> mask = simd16<T>::splat(m);
      return  simd16x32<bool>(
        this->chunks[0] <= mask,
        this->chunks[1] <= mask
      ).to_bitmask();
    }

    simdutf_really_inline uint64_t in_range(const T low, const T high) const {
      const simd16<T> mask_low = simd16<T>::splat(low);
      const simd16<T> mask_high = simd16<T>::splat(high);

      return  simd16x32<bool>(
        (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
        (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
        (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
        (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low)
      ).to_bitmask();
    }
    simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
      const simd16<T> mask_low = simd16<T>::splat(static_cast<T>(low-1));
      const simd16<T> mask_high = simd16<T>::splat(static_cast<T>(high+1));
      return simd16x32<bool>(
        (this->chunks[0] >= mask_high) | (this->chunks[0] <= mask_low),
        (this->chunks[1] >= mask_high) | (this->chunks[1] <= mask_low)
      ).to_bitmask();
    }
    simdutf_really_inline uint64_t lt(const T m) const {
      const simd16<T> mask = simd16<T>::splat(m);
      return  simd16x32<bool>(
        this->chunks[0] < mask,
        this->chunks[1] < mask
      ).to_bitmask();
    }
  }; // struct simd16x32<T>
/* end file src/simdutf/haswell/simd16-inl.h */

} // namespace simd

} // unnamed namespace
} // namespace haswell
} // namespace simdutf

#endif // SIMDUTF_HASWELL_SIMD_H
/* end file src/simdutf/haswell/simd.h */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/haswell/end.h
/* begin file src/simdutf/haswell/end.h */
#if SIMDUTF_CAN_ALWAYS_RUN_HASWELL
// nothing needed.
#else
SIMDUTF_UNTARGET_REGION
#endif


#if SIMDUTF_GCC11ORMORE // workaround for https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
SIMDUTF_POP_DISABLE_WARNINGS
#endif // end of workaround
/* end file src/simdutf/haswell/end.h */

#endif // SIMDUTF_IMPLEMENTATION_HASWELL
#endif // SIMDUTF_HASWELL_COMMON_H
/* end file src/simdutf/haswell.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/westmere.h
/* begin file src/simdutf/westmere.h */
#ifndef SIMDUTF_WESTMERE_H
#define SIMDUTF_WESTMERE_H

#ifdef SIMDUTF_FALLBACK_H
#error "westmere.h must be included before fallback.h"
#endif


// Default Westmere to on if this is x86-64, unless we'll always select Haswell.
#ifndef SIMDUTF_IMPLEMENTATION_WESTMERE
//
// You do not want to set it to (SIMDUTF_IS_X86_64 && !SIMDUTF_REQUIRES_HASWELL)
// because you want to rely on runtime dispatch!
//
#if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE || SIMDUTF_CAN_ALWAYS_RUN_HASWELL
#define SIMDUTF_IMPLEMENTATION_WESTMERE 0
#else
#define SIMDUTF_IMPLEMENTATION_WESTMERE (SIMDUTF_IS_X86_64)
#endif

#endif

#define SIMDUTF_CAN_ALWAYS_RUN_WESTMERE (SIMDUTF_IMPLEMENTATION_WESTMERE && SIMDUTF_IS_X86_64 && __SSE4_2__)

#if SIMDUTF_IMPLEMENTATION_WESTMERE

#define SIMDUTF_TARGET_WESTMERE SIMDUTF_TARGET_REGION("sse4.2,popcnt")

namespace simdutf {
/**
 * Implementation for Westmere (Intel SSE4.2).
 */
namespace westmere {
} // namespace westmere
} // namespace simdutf

//
// These two need to be included outside SIMDUTF_TARGET_REGION
//
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/westmere/implementation.h
/* begin file src/simdutf/westmere/implementation.h */
#ifndef SIMDUTF_WESTMERE_IMPLEMENTATION_H
#define SIMDUTF_WESTMERE_IMPLEMENTATION_H


// The constructor may be executed on any host, so we take care not to use SIMDUTF_TARGET_REGION
namespace simdutf {
namespace westmere {

namespace {
using namespace simdutf;
}

class implementation final : public simdutf::implementation {
public:
  simdutf_really_inline implementation() : simdutf::implementation("westmere", "Intel/AMD SSE4.2", internal::instruction_set::SSE42) {}
  simdutf_warn_unused int detect_encodings(const char * input, size_t length) const noexcept final;
  simdutf_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_ascii(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf16le(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf16be(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf16le_with_errors(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf16be_with_errors(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf32(const char32_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf32_with_errors(const char32_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf16le(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf16be(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(const char * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(const char * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf32(const char * buf, size_t len, char32_t* utf32_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(const char * buf, size_t len, char32_t* utf32_output) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(const char * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16le_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16be_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf8(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf16le(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf16be(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf16le(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf16be(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16le_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16be_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16le_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16be_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  void change_endianness_utf16(const char16_t * buf, size_t length, char16_t * output) const noexcept final;
  simdutf_warn_unused size_t count_utf16le(const char16_t * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t count_utf16be(const char16_t * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t count_utf8(const char * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf16le(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf16be(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf16le(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf16be(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf16_length_from_utf8(const char * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf32(const char32_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf16_length_from_utf32(const char32_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf8(const char * input, size_t length) const noexcept;
};

} // namespace westmere
} // namespace simdutf

#endif // SIMDUTF_WESTMERE_IMPLEMENTATION_H
/* end file src/simdutf/westmere/implementation.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/westmere/intrinsics.h
/* begin file src/simdutf/westmere/intrinsics.h */
#ifndef SIMDUTF_WESTMERE_INTRINSICS_H
#define SIMDUTF_WESTMERE_INTRINSICS_H

#ifdef SIMDUTF_VISUAL_STUDIO
// under clang within visual studio, this will include <x86intrin.h>
#include <intrin.h> // visual studio or clang
#else

#if SIMDUTF_GCC11ORMORE
// We should not get warnings while including <x86intrin.h> yet we do
// under some versions of GCC.
// If the x86intrin.h header has uninitialized values that are problematic,
// it is a GCC issue, we want to ignore these warnigns.
SIMDUTF_DISABLE_GCC_WARNING(-Wuninitialized)
#endif

#include <x86intrin.h> // elsewhere


#if SIMDUTF_GCC11ORMORE
// cancels the suppression of the -Wuninitialized
SIMDUTF_POP_DISABLE_WARNINGS
#endif

#endif // SIMDUTF_VISUAL_STUDIO


#ifdef SIMDUTF_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
#endif



#endif // SIMDUTF_WESTMERE_INTRINSICS_H
/* end file src/simdutf/westmere/intrinsics.h */

//
// The rest need to be inside the region
//
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/westmere/begin.h
/* begin file src/simdutf/westmere/begin.h */
// redefining SIMDUTF_IMPLEMENTATION to "westmere"
// #define SIMDUTF_IMPLEMENTATION westmere

#if SIMDUTF_CAN_ALWAYS_RUN_WESTMERE
// nothing needed.
#else
SIMDUTF_TARGET_WESTMERE
#endif
/* end file src/simdutf/westmere/begin.h */

// Declarations
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/westmere/bitmanipulation.h
/* begin file src/simdutf/westmere/bitmanipulation.h */
#ifndef SIMDUTF_WESTMERE_BITMANIPULATION_H
#define SIMDUTF_WESTMERE_BITMANIPULATION_H

namespace simdutf {
namespace westmere {
namespace {

#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
simdutf_really_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
simdutf_really_inline long long int count_ones(uint64_t input_num) {
  return _popcnt64(input_num);
}
#endif

} // unnamed namespace
} // namespace westmere
} // namespace simdutf

#endif // SIMDUTF_WESTMERE_BITMANIPULATION_H
/* end file src/simdutf/westmere/bitmanipulation.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/westmere/simd.h
/* begin file src/simdutf/westmere/simd.h */
#ifndef SIMDUTF_WESTMERE_SIMD_H
#define SIMDUTF_WESTMERE_SIMD_H

namespace simdutf {
namespace westmere {
namespace {
namespace simd {

  template<typename Child>
  struct base {
    __m128i value;

    // Zero constructor
    simdutf_really_inline base() : value{__m128i()} {}

    // Conversion from SIMD register
    simdutf_really_inline base(const __m128i _value) : value(_value) {}
    // Conversion to SIMD register
    simdutf_really_inline operator const __m128i&() const { return this->value; }
    simdutf_really_inline operator __m128i&() { return this->value; }
    template <endianness big_endian>
    simdutf_really_inline void store_ascii_as_utf16(char16_t * p) const {
      __m128i first = _mm_cvtepu8_epi16(*this);
      __m128i second = _mm_cvtepu8_epi16(_mm_srli_si128(*this,8));
      if (big_endian) {
        const __m128i swap = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
        first = _mm_shuffle_epi8(first, swap);
        second = _mm_shuffle_epi8(second, swap);
      }
      _mm_storeu_si128(reinterpret_cast<__m128i *>(p), first);
      _mm_storeu_si128(reinterpret_cast<__m128i *>(p+8), second);
    }
    simdutf_really_inline void store_ascii_as_utf32(char32_t * p) const {
      _mm_storeu_si128(reinterpret_cast<__m128i *>(p), _mm_cvtepu8_epi32(*this));
      _mm_storeu_si128(reinterpret_cast<__m128i *>(p+4), _mm_cvtepu8_epi32(_mm_srli_si128(*this,4)));
      _mm_storeu_si128(reinterpret_cast<__m128i *>(p+8), _mm_cvtepu8_epi32(_mm_srli_si128(*this,8)));
      _mm_storeu_si128(reinterpret_cast<__m128i *>(p+12), _mm_cvtepu8_epi32(_mm_srli_si128(*this,12)));
    }
    // Bit operations
    simdutf_really_inline Child operator|(const Child other) const { return _mm_or_si128(*this, other); }
    simdutf_really_inline Child operator&(const Child other) const { return _mm_and_si128(*this, other); }
    simdutf_really_inline Child operator^(const Child other) const { return _mm_xor_si128(*this, other); }
    simdutf_really_inline Child bit_andnot(const Child other) const { return _mm_andnot_si128(other, *this); }
    simdutf_really_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
    simdutf_really_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
    simdutf_really_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 uint16_t bitmask_t;
    typedef uint32_t bitmask2_t;

    simdutf_really_inline T first() const { return _mm_extract_epi8(*this,0); }
    simdutf_really_inline T last() const { return _mm_extract_epi8(*this,15); }
    simdutf_really_inline base8() : base<simd8<T>>() {}
    simdutf_really_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}

    friend simdutf_really_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>
    simdutf_really_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 simdutf_really_inline simd8<bool> splat(bool _value) { return _mm_set1_epi8(uint8_t(-(!!_value))); }

    simdutf_really_inline simd8<bool>() : base8() {}
    simdutf_really_inline simd8<bool>(const __m128i _value) : base8<bool>(_value) {}
    // Splat constructor
    simdutf_really_inline simd8<bool>(bool _value) : base8<bool>(splat(_value)) {}

    simdutf_really_inline int to_bitmask() const { return _mm_movemask_epi8(*this); }
    simdutf_really_inline bool any() const { return !_mm_testz_si128(*this, *this); }
    simdutf_really_inline bool none() const { return _mm_testz_si128(*this, *this); }
    simdutf_really_inline bool all() const { return _mm_movemask_epi8(*this) == 0xFFFF; }
    simdutf_really_inline simd8<bool> operator~() const { return *this ^ true; }
  };

  template<typename T>
  struct base8_numeric: base8<T> {
    static simdutf_really_inline simd8<T> splat(T _value) { return _mm_set1_epi8(_value); }
    static simdutf_really_inline simd8<T> zero() { return _mm_setzero_si128(); }
    static simdutf_really_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 simdutf_really_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
      );
    }

    simdutf_really_inline base8_numeric() : base8<T>() {}
    simdutf_really_inline base8_numeric(const __m128i _value) : base8<T>(_value) {}

    // Store to array
    simdutf_really_inline void store(T dst[16]) const { return _mm_storeu_si128(reinterpret_cast<__m128i *>(dst), *this); }

    // Override to distinguish from bool version
    simdutf_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }

    // Addition/subtraction are the same for signed and unsigned
    simdutf_really_inline simd8<T> operator+(const simd8<T> other) const { return _mm_add_epi8(*this, other); }
    simdutf_really_inline simd8<T> operator-(const simd8<T> other) const { return _mm_sub_epi8(*this, other); }
    simdutf_really_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
    simdutf_really_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>
    simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
      return _mm_shuffle_epi8(lookup_table, *this);
    }

    template<typename L>
    simdutf_really_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> {
    simdutf_really_inline simd8() : base8_numeric<int8_t>() {}
    simdutf_really_inline simd8(const __m128i _value) : base8_numeric<int8_t>(_value) {}
    // Splat constructor
    simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
    // Array constructor
    simdutf_really_inline simd8(const int8_t* values) : simd8(load(values)) {}
    // Member-by-member initialization
    simdutf_really_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)
    simdutf_really_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
      );
    }
    simdutf_really_inline operator simd8<uint8_t>() const;
    simdutf_really_inline bool is_ascii() const { return _mm_movemask_epi8(*this) == 0; }

    // Order-sensitive comparisons
    simdutf_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm_max_epi8(*this, other); }
    simdutf_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm_min_epi8(*this, other); }
    simdutf_really_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm_cmpgt_epi8(*this, other); }
    simdutf_really_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>  {
    simdutf_really_inline simd8() : base8_numeric<uint8_t>() {}
    simdutf_really_inline simd8(const __m128i _value) : base8_numeric<uint8_t>(_value) {}

    // Splat constructor
    simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
    // Array constructor
    simdutf_really_inline simd8(const uint8_t* values) : simd8(load(values)) {}
    // Member-by-member initialization
    simdutf_really_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)
    simdutf_really_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
    simdutf_really_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm_adds_epu8(*this, other); }
    simdutf_really_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm_subs_epu8(*this, other); }

    // Order-specific operations
    simdutf_really_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm_max_epu8(*this, other); }
    simdutf_really_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)
    simdutf_really_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)
    simdutf_really_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
    simdutf_really_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
    simdutf_really_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
    simdutf_really_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
    simdutf_really_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }

    // Bit-specific operations
    simdutf_really_inline simd8<bool> bits_not_set() const { return *this == uint8_t(0); }
    simdutf_really_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
    simdutf_really_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
    simdutf_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
    simdutf_really_inline bool is_ascii() const { return _mm_movemask_epi8(*this) == 0; }

    simdutf_really_inline bool bits_not_set_anywhere() const { return _mm_testz_si128(*this, *this); }
    simdutf_really_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
    simdutf_really_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return _mm_testz_si128(*this, bits); }
    simdutf_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
    template<int N>
    simdutf_really_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
    template<int N>
    simdutf_really_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>
    simdutf_really_inline int get_bit() const { return _mm_movemask_epi8(_mm_slli_epi16(*this, 7-N)); }
  };
  simdutf_really_inline simd8<int8_t>::operator simd8<uint8_t>() const { return this->value; }

  // Unsigned bytes
  template<>
  struct simd8<uint16_t>: base<uint16_t> {
    static simdutf_really_inline simd8<uint16_t> splat(uint16_t _value) { return _mm_set1_epi16(_value); }
    static simdutf_really_inline simd8<uint16_t> load(const uint16_t values[8]) {
      return _mm_loadu_si128(reinterpret_cast<const __m128i *>(values));
    }

    simdutf_really_inline simd8() : base<uint16_t>() {}
    simdutf_really_inline simd8(const __m128i _value) : base<uint16_t>(_value) {}
    // Splat constructor
    simdutf_really_inline simd8(uint16_t _value) : simd8(splat(_value)) {}
    // Array constructor
    simdutf_really_inline simd8(const uint16_t* values) : simd8(load(values)) {}
    // Member-by-member initialization
    simdutf_really_inline simd8(
      uint16_t v0,  uint16_t v1,  uint16_t v2,  uint16_t v3,  uint16_t v4,  uint16_t v5,  uint16_t v6,  uint16_t v7
    ) : simd8(_mm_setr_epi16(
      v0, v1, v2, v3, v4, v5, v6, v7
    )) {}

    // Saturated math
    simdutf_really_inline simd8<uint16_t> saturating_add(const simd8<uint16_t> other) const { return _mm_adds_epu16(*this, other); }
    simdutf_really_inline simd8<uint16_t> saturating_sub(const simd8<uint16_t> other) const { return _mm_subs_epu16(*this, other); }

    // Order-specific operations
    simdutf_really_inline simd8<uint16_t> max_val(const simd8<uint16_t> other) const { return _mm_max_epu16(*this, other); }
    simdutf_really_inline simd8<uint16_t> min_val(const simd8<uint16_t> other) const { return _mm_min_epu16(*this, other); }
    // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
    simdutf_really_inline simd8<uint16_t> gt_bits(const simd8<uint16_t> other) const { return this->saturating_sub(other); }
    // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
    simdutf_really_inline simd8<uint16_t> lt_bits(const simd8<uint16_t> other) const { return other.saturating_sub(*this); }
    simdutf_really_inline simd8<bool> operator<=(const simd8<uint16_t> other) const { return other.max_val(*this) == other; }
    simdutf_really_inline simd8<bool> operator>=(const simd8<uint16_t> other) const { return other.min_val(*this) == other; }
    simdutf_really_inline simd8<bool> operator==(const simd8<uint16_t> other) const { return _mm_cmpeq_epi16(*this, other); }
    simdutf_really_inline simd8<bool> operator&(const simd8<uint16_t> other) const { return _mm_and_si128(*this, other); }
    simdutf_really_inline simd8<bool> operator|(const simd8<uint16_t> other) const { return _mm_or_si128(*this, other); }

    // Bit-specific operations
    simdutf_really_inline simd8<bool> bits_not_set() const { return *this == uint16_t(0); }
    simdutf_really_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }

    simdutf_really_inline bool bits_not_set_anywhere() const { return _mm_testz_si128(*this, *this); }
    simdutf_really_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
    simdutf_really_inline bool bits_not_set_anywhere(simd8<uint16_t> bits) const { return _mm_testz_si128(*this, bits); }
    simdutf_really_inline bool any_bits_set_anywhere(simd8<uint16_t> bits) const { return !bits_not_set_anywhere(bits); }
     };
  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.");
    simd8<T> chunks[NUM_CHUNKS];

    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

    simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1, const simd8<T> chunk2, const simd8<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
    simdutf_really_inline simd8x64(const T* ptr) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+sizeof(simd8<T>)/sizeof(T)), simd8<T>::load(ptr+2*sizeof(simd8<T>)/sizeof(T)), simd8<T>::load(ptr+3*sizeof(simd8<T>)/sizeof(T))} {}

    simdutf_really_inline void store(T* ptr) const {
      this->chunks[0].store(ptr+sizeof(simd8<T>)*0/sizeof(T));
      this->chunks[1].store(ptr+sizeof(simd8<T>)*1/sizeof(T));
      this->chunks[2].store(ptr+sizeof(simd8<T>)*2/sizeof(T));
      this->chunks[3].store(ptr+sizeof(simd8<T>)*3/sizeof(T));
    }

    simdutf_really_inline simd8x64<T>& operator |=(const simd8x64<T> &other) {
      this->chunks[0] |= other.chunks[0];
      this->chunks[1] |= other.chunks[1];
      this->chunks[2] |= other.chunks[2];
      this->chunks[3] |= other.chunks[3];
      return *this;
    }

    simdutf_really_inline simd8<T> reduce_or() const {
      return (this->chunks[0] | this->chunks[1]) | (this->chunks[2] | this->chunks[3]);
    }

    simdutf_really_inline bool is_ascii() const {
      return this->reduce_or().is_ascii();
    }

    template <endianness endian>
    simdutf_really_inline void store_ascii_as_utf16(char16_t * ptr) const {
      this->chunks[0].template store_ascii_as_utf16<endian>(ptr+sizeof(simd8<T>)*0);
      this->chunks[1].template store_ascii_as_utf16<endian>(ptr+sizeof(simd8<T>)*1);
      this->chunks[2].template store_ascii_as_utf16<endian>(ptr+sizeof(simd8<T>)*2);
      this->chunks[3].template store_ascii_as_utf16<endian>(ptr+sizeof(simd8<T>)*3);
    }

    simdutf_really_inline void store_ascii_as_utf32(char32_t * ptr) const {
      this->chunks[0].store_ascii_as_utf32(ptr+sizeof(simd8<T>)*0);
      this->chunks[1].store_ascii_as_utf32(ptr+sizeof(simd8<T>)*1);
      this->chunks[2].store_ascii_as_utf32(ptr+sizeof(simd8<T>)*2);
      this->chunks[3].store_ascii_as_utf32(ptr+sizeof(simd8<T>)*3);
    }

    simdutf_really_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);
    }

    simdutf_really_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();
    }

    simdutf_really_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();
    }

    simdutf_really_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();
    }

    simdutf_really_inline uint64_t in_range(const T low, const T high) const {
      const simd8<T> mask_low = simd8<T>::splat(low);
      const simd8<T> mask_high = simd8<T>::splat(high);

      return  simd8x64<bool>(
        (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
        (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
        (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
        (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low)
      ).to_bitmask();
    }
    simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
      const simd8<T> mask_low = simd8<T>::splat(low-1);
      const simd8<T> mask_high = simd8<T>::splat(high+1);
      return simd8x64<bool>(
        (this->chunks[0] >= mask_high) | (this->chunks[0] <= mask_low),
        (this->chunks[1] >= mask_high) | (this->chunks[1] <= mask_low),
        (this->chunks[2] >= mask_high) | (this->chunks[2] <= mask_low),
        (this->chunks[3] >= mask_high) | (this->chunks[3] <= mask_low)
      ).to_bitmask();
    }
    simdutf_really_inline uint64_t lt(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();
    }

    simdutf_really_inline uint64_t gt(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();
    }
    simdutf_really_inline uint64_t gteq(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();
    }
    simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
      const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
      return  simd8x64<bool>(
        simd8<uint8_t>(__m128i(this->chunks[0])) >= mask,
        simd8<uint8_t>(__m128i(this->chunks[1])) >= mask,
        simd8<uint8_t>(__m128i(this->chunks[2])) >= mask,
        simd8<uint8_t>(__m128i(this->chunks[3])) >= mask
      ).to_bitmask();
    }
  }; // struct simd8x64<T>

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/westmere/simd16-inl.h
/* begin file src/simdutf/westmere/simd16-inl.h */
template<typename T>
struct simd16;

template<typename T, typename Mask=simd16<bool>>
struct base16: base<simd16<T>> {
  typedef uint16_t bitmask_t;
  typedef uint32_t bitmask2_t;

  simdutf_really_inline base16() : base<simd16<T>>() {}
  simdutf_really_inline base16(const __m128i _value) : base<simd16<T>>(_value) {}
  template <typename Pointer>
  simdutf_really_inline base16(const Pointer* ptr) : base16(_mm_loadu_si128(reinterpret_cast<const __m128i*>(ptr))) {}

  friend simdutf_really_inline Mask operator==(const simd16<T> lhs, const simd16<T> rhs) { return _mm_cmpeq_epi16(lhs, rhs); }

  static const int SIZE = sizeof(base<simd16<T>>::value);

  template<int N=1>
  simdutf_really_inline simd16<T> prev(const simd16<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 simd16<bool>: base16<bool> {
  static simdutf_really_inline simd16<bool> splat(bool _value) { return _mm_set1_epi16(uint16_t(-(!!_value))); }

  simdutf_really_inline simd16<bool>() : base16() {}
  simdutf_really_inline simd16<bool>(const __m128i _value) : base16<bool>(_value) {}
  // Splat constructor
  simdutf_really_inline simd16<bool>(bool _value) : base16<bool>(splat(_value)) {}

  simdutf_really_inline int to_bitmask() const { return _mm_movemask_epi8(*this); }
  simdutf_really_inline bool any() const { return !_mm_testz_si128(*this, *this); }
  simdutf_really_inline simd16<bool> operator~() const { return *this ^ true; }
};

template<typename T>
struct base16_numeric: base16<T> {
  static simdutf_really_inline simd16<T> splat(T _value) { return _mm_set1_epi16(_value); }
  static simdutf_really_inline simd16<T> zero() { return _mm_setzero_si128(); }
  static simdutf_really_inline simd16<T> load(const T values[8]) {
    return _mm_loadu_si128(reinterpret_cast<const __m128i *>(values));
  }

  simdutf_really_inline base16_numeric() : base16<T>() {}
  simdutf_really_inline base16_numeric(const __m128i _value) : base16<T>(_value) {}

  // Store to array
  simdutf_really_inline void store(T dst[8]) const { return _mm_storeu_si128(reinterpret_cast<__m128i *>(dst), *this); }

  // Override to distinguish from bool version
  simdutf_really_inline simd16<T> operator~() const { return *this ^ 0xFFu; }

  // Addition/subtraction are the same for signed and unsigned
  simdutf_really_inline simd16<T> operator+(const simd16<T> other) const { return _mm_add_epi16(*this, other); }
  simdutf_really_inline simd16<T> operator-(const simd16<T> other) const { return _mm_sub_epi16(*this, other); }
  simdutf_really_inline simd16<T>& operator+=(const simd16<T> other) { *this = *this + other; return *static_cast<simd16<T>*>(this); }
  simdutf_really_inline simd16<T>& operator-=(const simd16<T> other) { *this = *this - other; return *static_cast<simd16<T>*>(this); }
};

// Signed words
template<>
struct simd16<int16_t> : base16_numeric<int16_t> {
  simdutf_really_inline simd16() : base16_numeric<int16_t>() {}
  simdutf_really_inline simd16(const __m128i _value) : base16_numeric<int16_t>(_value) {}
  // Splat constructor
  simdutf_really_inline simd16(int16_t _value) : simd16(splat(_value)) {}
  // Array constructor
  simdutf_really_inline simd16(const int16_t* values) : simd16(load(values)) {}
  simdutf_really_inline simd16(const char16_t* values) : simd16(load(reinterpret_cast<const int16_t*>(values))) {}
  // Member-by-member initialization
  simdutf_really_inline simd16(
    int16_t v0, int16_t v1, int16_t v2, int16_t v3, int16_t v4, int16_t v5, int16_t v6, int16_t v7)
    : simd16(_mm_setr_epi16(v0, v1, v2, v3, v4, v5, v6, v7)) {}
  simdutf_really_inline operator simd16<uint16_t>() const;

  // Order-sensitive comparisons
  simdutf_really_inline simd16<int16_t> max_val(const simd16<int16_t> other) const { return _mm_max_epi16(*this, other); }
  simdutf_really_inline simd16<int16_t> min_val(const simd16<int16_t> other) const { return _mm_min_epi16(*this, other); }
  simdutf_really_inline simd16<bool> operator>(const simd16<int16_t> other) const { return _mm_cmpgt_epi16(*this, other); }
  simdutf_really_inline simd16<bool> operator<(const simd16<int16_t> other) const { return _mm_cmpgt_epi16(other, *this); }
};

// Unsigned words
template<>
struct simd16<uint16_t>: base16_numeric<uint16_t>  {
  simdutf_really_inline simd16() : base16_numeric<uint16_t>() {}
  simdutf_really_inline simd16(const __m128i _value) : base16_numeric<uint16_t>(_value) {}

  // Splat constructor
  simdutf_really_inline simd16(uint16_t _value) : simd16(splat(_value)) {}
  // Array constructor
  simdutf_really_inline simd16(const uint16_t* values) : simd16(load(values)) {}
  simdutf_really_inline simd16(const char16_t* values) : simd16(load(reinterpret_cast<const uint16_t*>(values))) {}
  // Member-by-member initialization
  simdutf_really_inline simd16(
    uint16_t v0, uint16_t v1, uint16_t v2, uint16_t v3, uint16_t v4, uint16_t v5, uint16_t v6, uint16_t v7)
  : simd16(_mm_setr_epi16(v0, v1, v2, v3, v4, v5, v6, v7)) {}
  // Repeat 16 values as many times as necessary (usually for lookup tables)
  simdutf_really_inline static simd16<uint16_t> repeat_16(
    uint16_t v0, uint16_t v1, uint16_t v2, uint16_t v3, uint16_t v4, uint16_t v5, uint16_t v6, uint16_t v7
  ) {
    return simd16<uint16_t>(v0, v1, v2, v3, v4, v5, v6, v7);
  }

  // Saturated math
  simdutf_really_inline simd16<uint16_t> saturating_add(const simd16<uint16_t> other) const { return _mm_adds_epu16(*this, other); }
  simdutf_really_inline simd16<uint16_t> saturating_sub(const simd16<uint16_t> other) const { return _mm_subs_epu16(*this, other); }

  // Order-specific operations
  simdutf_really_inline simd16<uint16_t> max_val(const simd16<uint16_t> other) const { return _mm_max_epu16(*this, other); }
  simdutf_really_inline simd16<uint16_t> min_val(const simd16<uint16_t> other) const { return _mm_min_epu16(*this, other); }
  // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
  simdutf_really_inline simd16<uint16_t> gt_bits(const simd16<uint16_t> other) const { return this->saturating_sub(other); }
  // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
  simdutf_really_inline simd16<uint16_t> lt_bits(const simd16<uint16_t> other) const { return other.saturating_sub(*this); }
  simdutf_really_inline simd16<bool> operator<=(const simd16<uint16_t> other) const { return other.max_val(*this) == other; }
  simdutf_really_inline simd16<bool> operator>=(const simd16<uint16_t> other) const { return other.min_val(*this) == other; }
  simdutf_really_inline simd16<bool> operator>(const simd16<uint16_t> other) const { return this->gt_bits(other).any_bits_set(); }
  simdutf_really_inline simd16<bool> operator<(const simd16<uint16_t> other) const { return this->gt_bits(other).any_bits_set(); }

  // Bit-specific operations
  simdutf_really_inline simd16<bool> bits_not_set() const { return *this == uint16_t(0); }
  simdutf_really_inline simd16<bool> bits_not_set(simd16<uint16_t> bits) const { return (*this & bits).bits_not_set(); }
  simdutf_really_inline simd16<bool> any_bits_set() const { return ~this->bits_not_set(); }
  simdutf_really_inline simd16<bool> any_bits_set(simd16<uint16_t> bits) const { return ~this->bits_not_set(bits); }

  simdutf_really_inline bool bits_not_set_anywhere() const { return _mm_testz_si128(*this, *this); }
  simdutf_really_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
  simdutf_really_inline bool bits_not_set_anywhere(simd16<uint16_t> bits) const { return _mm_testz_si128(*this, bits); }
  simdutf_really_inline bool any_bits_set_anywhere(simd16<uint16_t> bits) const { return !bits_not_set_anywhere(bits); }
  template<int N>
  simdutf_really_inline simd16<uint16_t> shr() const { return simd16<uint16_t>(_mm_srli_epi16(*this, N)); }
  template<int N>
  simdutf_really_inline simd16<uint16_t> shl() const { return simd16<uint16_t>(_mm_slli_epi16(*this, 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>
  simdutf_really_inline int get_bit() const { return _mm_movemask_epi8(_mm_slli_epi16(*this, 7-N)); }

  // Change the endianness
  simdutf_really_inline simd16<uint16_t> swap_bytes() const {
    const __m128i swap = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
    return _mm_shuffle_epi8(*this, swap);
  }

  // Pack with the unsigned saturation  two uint16_t words into single uint8_t vector
  static simdutf_really_inline simd8<uint8_t> pack(const simd16<uint16_t>& v0, const simd16<uint16_t>& v1) {
    return _mm_packus_epi16(v0, v1);
  }
};
simdutf_really_inline simd16<int16_t>::operator simd16<uint16_t>() const { return this->value; }

template<typename T>
  struct simd16x32 {
    static constexpr int NUM_CHUNKS = 64 / sizeof(simd16<T>);
    static_assert(NUM_CHUNKS == 4, "Westmere kernel should use four registers per 64-byte block.");
    simd16<T> chunks[NUM_CHUNKS];

    simd16x32(const simd16x32<T>& o) = delete; // no copy allowed
    simd16x32<T>& operator=(const simd16<T> other) = delete; // no assignment allowed
    simd16x32() = delete; // no default constructor allowed

    simdutf_really_inline simd16x32(const simd16<T> chunk0, const simd16<T> chunk1, const simd16<T> chunk2, const simd16<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
    simdutf_really_inline simd16x32(const T* ptr) : chunks{simd16<T>::load(ptr), simd16<T>::load(ptr+sizeof(simd16<T>)/sizeof(T)), simd16<T>::load(ptr+2*sizeof(simd16<T>)/sizeof(T)), simd16<T>::load(ptr+3*sizeof(simd16<T>)/sizeof(T))} {}

    simdutf_really_inline void store(T* ptr) const {
      this->chunks[0].store(ptr+sizeof(simd16<T>)*0/sizeof(T));
      this->chunks[1].store(ptr+sizeof(simd16<T>)*1/sizeof(T));
      this->chunks[2].store(ptr+sizeof(simd16<T>)*2/sizeof(T));
      this->chunks[3].store(ptr+sizeof(simd16<T>)*3/sizeof(T));
    }

    simdutf_really_inline simd16<T> reduce_or() const {
      return (this->chunks[0] | this->chunks[1]) | (this->chunks[2] | this->chunks[3]);
    }

    simdutf_really_inline bool is_ascii() const {
      return this->reduce_or().is_ascii();
    }

    simdutf_really_inline void store_ascii_as_utf16(char16_t * ptr) const {
      this->chunks[0].store_ascii_as_utf16(ptr+sizeof(simd16<T>)*0);
      this->chunks[1].store_ascii_as_utf16(ptr+sizeof(simd16<T>)*1);
      this->chunks[2].store_ascii_as_utf16(ptr+sizeof(simd16<T>)*2);
      this->chunks[3].store_ascii_as_utf16(ptr+sizeof(simd16<T>)*3);
    }

    simdutf_really_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);
    }

    simdutf_really_inline void swap_bytes() {
      this->chunks[0] = this->chunks[0].swap_bytes();
      this->chunks[1] = this->chunks[1].swap_bytes();
      this->chunks[2] = this->chunks[2].swap_bytes();
      this->chunks[3] = this->chunks[3].swap_bytes();
    }

    simdutf_really_inline uint64_t eq(const T m) const {
      const simd16<T> mask = simd16<T>::splat(m);
      return  simd16x32<bool>(
        this->chunks[0] == mask,
        this->chunks[1] == mask,
        this->chunks[2] == mask,
        this->chunks[3] == mask
      ).to_bitmask();
    }

    simdutf_really_inline uint64_t eq(const simd16x32<uint16_t> &other) const {
      return  simd16x32<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();
    }

    simdutf_really_inline uint64_t lteq(const T m) const {
      const simd16<T> mask = simd16<T>::splat(m);
      return  simd16x32<bool>(
        this->chunks[0] <= mask,
        this->chunks[1] <= mask,
        this->chunks[2] <= mask,
        this->chunks[3] <= mask
      ).to_bitmask();
    }

    simdutf_really_inline uint64_t in_range(const T low, const T high) const {
      const simd16<T> mask_low = simd16<T>::splat(low);
      const simd16<T> mask_high = simd16<T>::splat(high);

      return  simd16x32<bool>(
        (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
        (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
        (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
        (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low)
      ).to_bitmask();
    }
    simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
      const simd16<T> mask_low = simd16<T>::splat(static_cast<T>(low-1));
      const simd16<T> mask_high = simd16<T>::splat(static_cast<T>(high+1));
      return simd16x32<bool>(
        (this->chunks[0] >= mask_high) | (this->chunks[0] <= mask_low),
        (this->chunks[1] >= mask_high) | (this->chunks[1] <= mask_low),
        (this->chunks[2] >= mask_high) | (this->chunks[2] <= mask_low),
        (this->chunks[3] >= mask_high) | (this->chunks[3] <= mask_low)
      ).to_bitmask();
    }
    simdutf_really_inline uint64_t lt(const T m) const {
      const simd16<T> mask = simd16<T>::splat(m);
      return  simd16x32<bool>(
        this->chunks[0] < mask,
        this->chunks[1] < mask,
        this->chunks[2] < mask,
        this->chunks[3] < mask
      ).to_bitmask();
    }
  }; // struct simd16x32<T>
/* end file src/simdutf/westmere/simd16-inl.h */

} // namespace simd
} // unnamed namespace
} // namespace westmere
} // namespace simdutf

#endif // SIMDUTF_WESTMERE_SIMD_INPUT_H
/* end file src/simdutf/westmere/simd.h */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/westmere/end.h
/* begin file src/simdutf/westmere/end.h */
#if SIMDUTF_CAN_ALWAYS_RUN_WESTMERE
// nothing needed.
#else
SIMDUTF_UNTARGET_REGION
#endif

/* end file src/simdutf/westmere/end.h */

#endif // SIMDUTF_IMPLEMENTATION_WESTMERE
#endif // SIMDUTF_WESTMERE_COMMON_H
/* end file src/simdutf/westmere.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/ppc64.h
/* begin file src/simdutf/ppc64.h */
#ifndef SIMDUTF_PPC64_H
#define SIMDUTF_PPC64_H

#ifdef SIMDUTF_FALLBACK_H
#error "ppc64.h must be included before fallback.h"
#endif


#ifndef SIMDUTF_IMPLEMENTATION_PPC64
#define SIMDUTF_IMPLEMENTATION_PPC64 (SIMDUTF_IS_PPC64)
#endif
#define SIMDUTF_CAN_ALWAYS_RUN_PPC64 SIMDUTF_IMPLEMENTATION_PPC64 && SIMDUTF_IS_PPC64



#if SIMDUTF_IMPLEMENTATION_PPC64

namespace simdutf {
/**
 * Implementation for ALTIVEC (PPC64).
 */
namespace ppc64 {
} // namespace ppc64
} // namespace simdutf

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/ppc64/implementation.h
/* begin file src/simdutf/ppc64/implementation.h */
#ifndef SIMDUTF_PPC64_IMPLEMENTATION_H
#define SIMDUTF_PPC64_IMPLEMENTATION_H


namespace simdutf {
namespace ppc64 {

namespace {
using namespace simdutf;
} // namespace

class implementation final : public simdutf::implementation {
public:
  simdutf_really_inline implementation()
      : simdutf::implementation("ppc64", "PPC64 ALTIVEC",
                                 internal::instruction_set::ALTIVEC) {}
  simdutf_warn_unused int detect_encodings(const char * input, size_t length) const noexcept final;
  simdutf_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_ascii(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf16le(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf16be(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf16le_with_errors(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf16be_with_errors(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf32(const char32_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf32_with_errors(const char32_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf16le(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf16be(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(const char * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(const char * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf32(const char * buf, size_t len, char32_t* utf32_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(const char * buf, size_t len, char32_t* utf32_output) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(const char * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16le_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16be_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf8(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf16le(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf16be(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf16le(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf16be(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16le_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16be_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16le_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16be_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  void change_endianness_utf16(const char16_t * buf, size_t length, char16_t * output) const noexcept final;
  simdutf_warn_unused size_t count_utf16le(const char16_t * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t count_utf16be(const char16_t * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t count_utf8(const char * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf16le(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf16be(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf16le(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf16be(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf16_length_from_utf8(const char * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf32(const char32_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf16_length_from_utf32(const char32_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf8(const char * input, size_t length) const noexcept;
};

} // namespace ppc64
} // namespace simdutf

#endif // SIMDUTF_PPC64_IMPLEMENTATION_H
/* end file src/simdutf/ppc64/implementation.h */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/ppc64/begin.h
/* begin file src/simdutf/ppc64/begin.h */
// redefining SIMDUTF_IMPLEMENTATION to "ppc64"
// #define SIMDUTF_IMPLEMENTATION ppc64
/* end file src/simdutf/ppc64/begin.h */

// Declarations
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/ppc64/intrinsics.h
/* begin file src/simdutf/ppc64/intrinsics.h */
#ifndef SIMDUTF_PPC64_INTRINSICS_H
#define SIMDUTF_PPC64_INTRINSICS_H


// 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

#endif //  SIMDUTF_PPC64_INTRINSICS_H
/* end file src/simdutf/ppc64/intrinsics.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/ppc64/bitmanipulation.h
/* begin file src/simdutf/ppc64/bitmanipulation.h */
#ifndef SIMDUTF_PPC64_BITMANIPULATION_H
#define SIMDUTF_PPC64_BITMANIPULATION_H

namespace simdutf {
namespace ppc64 {
namespace {

#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
simdutf_really_inline int 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
simdutf_really_inline int count_ones(uint64_t input_num) {
  return __builtin_popcountll(input_num);
}
#endif

} // unnamed namespace
} // namespace ppc64
} // namespace simdutf

#endif // SIMDUTF_PPC64_BITMANIPULATION_H
/* end file src/simdutf/ppc64/bitmanipulation.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/ppc64/simd.h
/* begin file src/simdutf/ppc64/simd.h */
#ifndef SIMDUTF_PPC64_SIMD_H
#define SIMDUTF_PPC64_SIMD_H

#include <type_traits>

namespace simdutf {
namespace ppc64 {
namespace {
namespace simd {

using __m128i = __vector unsigned char;

template <typename Child> struct base {
  __m128i value;

  // Zero constructor
  simdutf_really_inline base() : value{__m128i()} {}

  // Conversion from SIMD register
  simdutf_really_inline base(const __m128i _value) : value(_value) {}

  // Conversion to SIMD register
  simdutf_really_inline operator const __m128i &() const {
    return this->value;
  }
  simdutf_really_inline operator __m128i &() { return this->value; }

  // Bit operations
  simdutf_really_inline Child operator|(const Child other) const {
    return vec_or(this->value, (__m128i)other);
  }
  simdutf_really_inline Child operator&(const Child other) const {
    return vec_and(this->value, (__m128i)other);
  }
  simdutf_really_inline Child operator^(const Child other) const {
    return vec_xor(this->value, (__m128i)other);
  }
  simdutf_really_inline Child bit_andnot(const Child other) const {
    return vec_andc(this->value, (__m128i)other);
  }
  simdutf_really_inline Child &operator|=(const Child other) {
    auto this_cast = static_cast<Child*>(this);
    *this_cast = *this_cast | other;
    return *this_cast;
  }
  simdutf_really_inline Child &operator&=(const Child other) {
    auto this_cast = static_cast<Child*>(this);
    *this_cast = *this_cast & other;
    return *this_cast;
  }
  simdutf_really_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 uint16_t bitmask_t;
  typedef uint32_t bitmask2_t;

  simdutf_really_inline base8() : base<simd8<T>>() {}
  simdutf_really_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}

  friend simdutf_really_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>
  simdutf_really_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 simdutf_really_inline simd8<bool> splat(bool _value) {
    return (__m128i)vec_splats((unsigned char)(-(!!_value)));
  }

  simdutf_really_inline simd8<bool>() : base8() {}
  simdutf_really_inline simd8<bool>(const __m128i _value)
      : base8<bool>(_value) {}
  // Splat constructor
  simdutf_really_inline simd8<bool>(bool _value)
      : base8<bool>(splat(_value)) {}

  simdutf_really_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
  }
  simdutf_really_inline bool any() const {
    return !vec_all_eq(this->value, (__m128i)vec_splats(0));
  }
  simdutf_really_inline simd8<bool> operator~() const {
    return this->value ^ (__m128i)splat(true);
  }
};

template <typename T> struct base8_numeric : base8<T> {
  static simdutf_really_inline simd8<T> splat(T value) {
    (void)value;
    return (__m128i)vec_splats(value);
  }
  static simdutf_really_inline simd8<T> zero() { return splat(0); }
  static simdutf_really_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 simdutf_really_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);
  }

  simdutf_really_inline base8_numeric() : base8<T>() {}
  simdutf_really_inline base8_numeric(const __m128i _value)
      : base8<T>(_value) {}

  // Store to array
  simdutf_really_inline void store(T dst[16]) const {
    vec_vsx_st(this->value, 0, reinterpret_cast<__m128i *>(dst));
  }

  // Override to distinguish from bool version
  simdutf_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }

  // Addition/subtraction are the same for signed and unsigned
  simdutf_really_inline simd8<T> operator+(const simd8<T> other) const {
    return (__m128i)((__m128i)this->value + (__m128i)other);
  }
  simdutf_really_inline simd8<T> operator-(const simd8<T> other) const {
    return (__m128i)((__m128i)this->value - (__m128i)other);
  }
  simdutf_really_inline simd8<T> &operator+=(const simd8<T> other) {
    *this = *this + other;
    return *static_cast<simd8<T> *>(this);
  }
  simdutf_really_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>
  simdutf_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
    return (__m128i)vec_perm((__m128i)lookup_table, (__m128i)lookup_table, this->value);
  }

  template <typename L>
  simdutf_really_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> {
  simdutf_really_inline simd8() : base8_numeric<int8_t>() {}
  simdutf_really_inline simd8(const __m128i _value)
      : base8_numeric<int8_t>(_value) {}

  // Splat constructor
  simdutf_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
  // Array constructor
  simdutf_really_inline simd8(const int8_t *values) : simd8(load(values)) {}
  // Member-by-member initialization
  simdutf_really_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)
  simdutf_really_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
  simdutf_really_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);
  }
  simdutf_really_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);
  }
  simdutf_really_inline simd8<bool>
  operator>(const simd8<int8_t> other) const {
    return (__m128i)vec_cmpgt((__vector signed char)this->value,
                              (__vector signed char)(__m128i)other);
  }
  simdutf_really_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> {
  simdutf_really_inline simd8() : base8_numeric<uint8_t>() {}
  simdutf_really_inline simd8(const __m128i _value)
      : base8_numeric<uint8_t>(_value) {}
  // Splat constructor
  simdutf_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
  // Array constructor
  simdutf_really_inline simd8(const uint8_t *values) : simd8(load(values)) {}
  // Member-by-member initialization
  simdutf_really_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)
  simdutf_really_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
  simdutf_really_inline simd8<uint8_t>
  saturating_add(const simd8<uint8_t> other) const {
    return (__m128i)vec_adds(this->value, (__m128i)other);
  }
  simdutf_really_inline simd8<uint8_t>
  saturating_sub(const simd8<uint8_t> other) const {
    return (__m128i)vec_subs(this->value, (__m128i)other);
  }

  // Order-specific operations
  simdutf_really_inline simd8<uint8_t>
  max_val(const simd8<uint8_t> other) const {
    return (__m128i)vec_max(this->value, (__m128i)other);
  }
  simdutf_really_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)
  simdutf_really_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)
  simdutf_really_inline simd8<uint8_t>
  lt_bits(const simd8<uint8_t> other) const {
    return other.saturating_sub(*this);
  }
  simdutf_really_inline simd8<bool>
  operator<=(const simd8<uint8_t> other) const {
    return other.max_val(*this) == other;
  }
  simdutf_really_inline simd8<bool>
  operator>=(const simd8<uint8_t> other) const {
    return other.min_val(*this) == other;
  }
  simdutf_really_inline simd8<bool>
  operator>(const simd8<uint8_t> other) const {
    return this->gt_bits(other).any_bits_set();
  }
  simdutf_really_inline simd8<bool>
  operator<(const simd8<uint8_t> other) const {
    return this->gt_bits(other).any_bits_set();
  }

  // Bit-specific operations
  simdutf_really_inline simd8<bool> bits_not_set() const {
    return (__m128i)vec_cmpeq(this->value, (__m128i)vec_splats(uint8_t(0)));
  }
  simdutf_really_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const {
    return (*this & bits).bits_not_set();
  }
  simdutf_really_inline simd8<bool> any_bits_set() const {
    return ~this->bits_not_set();
  }
  simdutf_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
    return ~this->bits_not_set(bits);
  }

  simdutf_really_inline bool is_ascii() const {
      return this->saturating_sub(0b01111111u).bits_not_set_anywhere();
  }

  simdutf_really_inline bool bits_not_set_anywhere() const {
    return vec_all_eq(this->value, (__m128i)vec_splats(0));
  }
  simdutf_really_inline bool any_bits_set_anywhere() const {
    return !bits_not_set_anywhere();
  }
  simdutf_really_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));
  }
  simdutf_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const {
    return !bits_not_set_anywhere(bits);
  }
  template <int N> simdutf_really_inline simd8<uint8_t> shr() const {
    return simd8<uint8_t>(
        (__m128i)vec_sr(this->value, (__m128i)vec_splat_u8(N)));
  }
  template <int N> simdutf_really_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.");
  simd8<T> chunks[NUM_CHUNKS];

  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

  simdutf_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1,
                                  const simd8<T> chunk2, const simd8<T> chunk3)
      : chunks{chunk0, chunk1, chunk2, chunk3} {}

  simdutf_really_inline simd8x64(const T* ptr) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+sizeof(simd8<T>)/sizeof(T)), simd8<T>::load(ptr+2*sizeof(simd8<T>)/sizeof(T)), simd8<T>::load(ptr+3*sizeof(simd8<T>)/sizeof(T))} {}

  simdutf_really_inline void store(T* ptr) const {
    this->chunks[0].store(ptr + sizeof(simd8<T>) * 0/sizeof(T));
    this->chunks[1].store(ptr + sizeof(simd8<T>) * 1/sizeof(T));
    this->chunks[2].store(ptr + sizeof(simd8<T>) * 2/sizeof(T));
    this->chunks[3].store(ptr + sizeof(simd8<T>) * 3/sizeof(T));
  }


  simdutf_really_inline simd8x64<T>& operator |=(const simd8x64<T> &other) {
      this->chunks[0] |= other.chunks[0];
      this->chunks[1] |= other.chunks[1];
      this->chunks[2] |= other.chunks[2];
      this->chunks[3] |= other.chunks[3];
      return *this;
    }

  simdutf_really_inline simd8<T> reduce_or() const {
    return (this->chunks[0] | this->chunks[1]) |
           (this->chunks[2] | this->chunks[3]);
  }


  simdutf_really_inline bool is_ascii() const {
    return input.reduce_or().is_ascii();
  }

  simdutf_really_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);
  }

  simdutf_really_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();
  }

  simdutf_really_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();
  }

  simdutf_really_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();
  }

  simdutf_really_inline uint64_t in_range(const T low, const T high) const {
      const simd8<T> mask_low = simd8<T>::splat(low);
      const simd8<T> mask_high = simd8<T>::splat(high);

      return  simd8x64<bool>(
        (this->chunks[0] <= mask_high) & (this->chunks[0] >= mask_low),
        (this->chunks[1] <= mask_high) & (this->chunks[1] >= mask_low),
        (this->chunks[2] <= mask_high) & (this->chunks[2] >= mask_low),
        (this->chunks[3] <= mask_high) & (this->chunks[3] >= mask_low)
      ).to_bitmask();
  }
  simdutf_really_inline uint64_t not_in_range(const T low, const T high) const {
      const simd8<T> mask_low = simd8<T>::splat(low);
      const simd8<T> mask_high = simd8<T>::splat(high);
      return  simd8x64<bool>(
        (this->chunks[0] > mask_high) | (this->chunks[0] < mask_low),
        (this->chunks[1] > mask_high) | (this->chunks[1] < mask_low),
        (this->chunks[2] > mask_high) | (this->chunks[2] < mask_low),
        (this->chunks[3] > mask_high) | (this->chunks[3] < mask_low)
      ).to_bitmask();
  }
  simdutf_really_inline uint64_t lt(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();
  }

  simdutf_really_inline uint64_t gt(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();
  }
  simdutf_really_inline uint64_t gteq(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();
  }
  simdutf_really_inline uint64_t gteq_unsigned(const uint8_t m) const {
      const simd8<uint8_t> mask = simd8<uint8_t>::splat(m);
      return  simd8x64<bool>(
        simd8<uint8_t>(this->chunks[0]) >= mask,
        simd8<uint8_t>(this->chunks[1]) >= mask,
        simd8<uint8_t>(this->chunks[2]) >= mask,
        simd8<uint8_t>(this->chunks[3]) >= mask
      ).to_bitmask();
  }
}; // struct simd8x64<T>

} // namespace simd
} // unnamed namespace
} // namespace ppc64
} // namespace simdutf

#endif // SIMDUTF_PPC64_SIMD_INPUT_H
/* end file src/simdutf/ppc64/simd.h */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/ppc64/end.h
/* begin file src/simdutf/ppc64/end.h */
/* end file src/simdutf/ppc64/end.h */

#endif // SIMDUTF_IMPLEMENTATION_PPC64

#endif // SIMDUTF_PPC64_H
/* end file src/simdutf/ppc64.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/fallback.h
/* begin file src/simdutf/fallback.h */
#ifndef SIMDUTF_FALLBACK_H
#define SIMDUTF_FALLBACK_H


// Note that fallback.h is always imported last.

// Default Fallback to on unless a builtin implementation has already been selected.
#ifndef SIMDUTF_IMPLEMENTATION_FALLBACK
#if SIMDUTF_CAN_ALWAYS_RUN_ARM64 || SIMDUTF_CAN_ALWAYS_RUN_ICELAKE || SIMDUTF_CAN_ALWAYS_RUN_HASWELL || SIMDUTF_CAN_ALWAYS_RUN_WESTMERE || SIMDUTF_CAN_ALWAYS_RUN_PPC64
#define SIMDUTF_IMPLEMENTATION_FALLBACK 0
#else
#define SIMDUTF_IMPLEMENTATION_FALLBACK 1
#endif
#endif

#define SIMDUTF_CAN_ALWAYS_RUN_FALLBACK (SIMDUTF_IMPLEMENTATION_FALLBACK)

#if SIMDUTF_IMPLEMENTATION_FALLBACK

namespace simdutf {
/**
 * Fallback implementation (runs on any machine).
 */
namespace fallback {
} // namespace fallback
} // namespace simdutf

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/fallback/implementation.h
/* begin file src/simdutf/fallback/implementation.h */
#ifndef SIMDUTF_FALLBACK_IMPLEMENTATION_H
#define SIMDUTF_FALLBACK_IMPLEMENTATION_H


namespace simdutf {
namespace fallback {

namespace {
using namespace simdutf;
}

class implementation final : public simdutf::implementation {
public:
  simdutf_really_inline implementation() : simdutf::implementation(
      "fallback",
      "Generic fallback implementation",
      0
  ) {}
  simdutf_warn_unused int detect_encodings(const char * input, size_t length) const noexcept final;
  simdutf_warn_unused bool validate_utf8(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf8_with_errors(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_ascii(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_ascii_with_errors(const char *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf16le(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf16be(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf16le_with_errors(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf16be_with_errors(const char16_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused bool validate_utf32(const char32_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused result validate_utf32_with_errors(const char32_t *buf, size_t len) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf16le(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf16be(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(const char * buf, size_t len, char16_t* utf16_output) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(const char * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(const char * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf8_to_utf32(const char * buf, size_t len, char32_t* utf32_output) const noexcept final;
  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(const char * buf, size_t len, char32_t* utf32_output) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(const char * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16le_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16be_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf8(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(const char32_t * buf, size_t len, char* utf8_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf16le(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf32_to_utf16be(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf16le(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf32_to_utf16be(const char32_t * buf, size_t len, char16_t* utf16_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16le_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_utf16be_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16le_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  simdutf_warn_unused size_t convert_valid_utf16be_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) const noexcept final;
  void change_endianness_utf16(const char16_t * buf, size_t length, char16_t * output) const noexcept final;
  simdutf_warn_unused size_t count_utf16le(const char16_t * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t count_utf16be(const char16_t * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t count_utf8(const char * buf, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf16le(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf16be(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf16le(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf16be(const char16_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf16_length_from_utf8(const char * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf8_length_from_utf32(const char32_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf16_length_from_utf32(const char32_t * input, size_t length) const noexcept;
  simdutf_warn_unused size_t utf32_length_from_utf8(const char * input, size_t length) const noexcept;
};

} // namespace fallback
} // namespace simdutf

#endif // SIMDUTF_FALLBACK_IMPLEMENTATION_H
/* end file src/simdutf/fallback/implementation.h */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/fallback/begin.h
/* begin file src/simdutf/fallback/begin.h */
// redefining SIMDUTF_IMPLEMENTATION to "fallback"
// #define SIMDUTF_IMPLEMENTATION fallback
/* end file src/simdutf/fallback/begin.h */

// Declarations
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/fallback/bitmanipulation.h
/* begin file src/simdutf/fallback/bitmanipulation.h */
#ifndef SIMDUTF_FALLBACK_BITMANIPULATION_H
#define SIMDUTF_FALLBACK_BITMANIPULATION_H

#include <limits>

namespace simdutf {
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

} // unnamed namespace
} // namespace fallback
} // namespace simdutf

#endif // SIMDUTF_FALLBACK_BITMANIPULATION_H
/* end file src/simdutf/fallback/bitmanipulation.h */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/fallback/end.h
/* begin file src/simdutf/fallback/end.h */
/* end file src/simdutf/fallback/end.h */

#endif // SIMDUTF_IMPLEMENTATION_FALLBACK
#endif // SIMDUTF_FALLBACK_H
/* end file src/simdutf/fallback.h */

namespace simdutf {
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);
}

simdutf_warn_unused encoding_type implementation::autodetect_encoding(const char * input, size_t length) const noexcept {
    // If there is a BOM, then we trust it.
    auto bom_encoding = simdutf::BOM::check_bom(input, length);
    if(bom_encoding != encoding_type::unspecified) { return bom_encoding; }
    // UTF8 is common, it includes ASCII, and is commonly represented
    // without a BOM, so if it fits, go with that. Note that it is still
    // possible to get it wrong, we are only 'guessing'. If some has UTF-16
    // data without a BOM, it could pass as UTF-8.
    //
    // An interesting twist might be to check for UTF-16 ASCII first (every
    // other byte is zero).
    if(validate_utf8(input, length)) { return encoding_type::UTF8; }
    // The next most common encoding that might appear without BOM is probably
    // UTF-16LE, so try that next.
    if((length % 2) == 0) {
      // important: we need to divide by two
      if(validate_utf16le(reinterpret_cast<const char16_t*>(input), length/2)) { return encoding_type::UTF16_LE; }
    }
    if((length % 4) == 0) {
      if(validate_utf32(reinterpret_cast<const char32_t*>(input), length/4)) { return encoding_type::UTF32_LE; }
    }
    return encoding_type::unspecified;
}

namespace internal {

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


#if SIMDUTF_IMPLEMENTATION_ICELAKE
const icelake::implementation icelake_singleton{};
#endif
#if SIMDUTF_IMPLEMENTATION_HASWELL
const haswell::implementation haswell_singleton{};
#endif
#if SIMDUTF_IMPLEMENTATION_WESTMERE
const westmere::implementation westmere_singleton{};
#endif
#if SIMDUTF_IMPLEMENTATION_ARM64
const arm64::implementation arm64_singleton{};
#endif
#if SIMDUTF_IMPLEMENTATION_PPC64
const ppc64::implementation ppc64_singleton{};
#endif
#if SIMDUTF_IMPLEMENTATION_FALLBACK
const fallback::implementation fallback_singleton{};
#endif

/**
 * @private Detects best supported implementation on first use, and sets it
 */
class detect_best_supported_implementation_on_first_use final : public implementation {
public:
  const std::string &name() const noexcept final { return set_best()->name(); }
  const std::string &description() const noexcept final { return set_best()->description(); }
  uint32_t required_instruction_sets() const noexcept final { return set_best()->required_instruction_sets(); }

  simdutf_warn_unused int detect_encodings(const char * input, size_t length) const noexcept override {
    return set_best()->detect_encodings(input, length);
  }

  simdutf_warn_unused bool validate_utf8(const char * buf, size_t len) const noexcept final override {
    return set_best()->validate_utf8(buf, len);
  }

  simdutf_warn_unused result validate_utf8_with_errors(const char * buf, size_t len) const noexcept final override {
    return set_best()->validate_utf8_with_errors(buf, len);
  }

  simdutf_warn_unused bool validate_ascii(const char * buf, size_t len) const noexcept final override {
    return set_best()->validate_ascii(buf, len);
  }

  simdutf_warn_unused result validate_ascii_with_errors(const char * buf, size_t len) const noexcept final override {
    return set_best()->validate_ascii_with_errors(buf, len);
  }

  simdutf_warn_unused bool validate_utf16le(const char16_t * buf, size_t len) const noexcept final override {
    return set_best()->validate_utf16le(buf, len);
  }

  simdutf_warn_unused bool validate_utf16be(const char16_t * buf, size_t len) const noexcept final override {
    return set_best()->validate_utf16be(buf, len);
  }

  simdutf_warn_unused result validate_utf16le_with_errors(const char16_t * buf, size_t len) const noexcept final override {
    return set_best()->validate_utf16le_with_errors(buf, len);
  }

  simdutf_warn_unused result validate_utf16be_with_errors(const char16_t * buf, size_t len) const noexcept final override {
    return set_best()->validate_utf16be_with_errors(buf, len);
  }

  simdutf_warn_unused bool validate_utf32(const char32_t * buf, size_t len) const noexcept final override {
    return set_best()->validate_utf32(buf, len);
  }

  simdutf_warn_unused result validate_utf32_with_errors(const char32_t * buf, size_t len) const noexcept final override {
    return set_best()->validate_utf32_with_errors(buf, len);
  }

  simdutf_warn_unused size_t convert_utf8_to_utf16le(const char * buf, size_t len, char16_t* utf16_output) const noexcept final override {
    return set_best()->convert_utf8_to_utf16le(buf, len, utf16_output);
  }

  simdutf_warn_unused size_t convert_utf8_to_utf16be(const char * buf, size_t len, char16_t* utf16_output) const noexcept final override {
    return set_best()->convert_utf8_to_utf16be(buf, len, utf16_output);
  }

  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(const char * buf, size_t len, char16_t* utf16_output) const noexcept final override {
    return set_best()->convert_utf8_to_utf16le_with_errors(buf, len, utf16_output);
  }

  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(const char * buf, size_t len, char16_t* utf16_output) const noexcept final override {
    return set_best()->convert_utf8_to_utf16be_with_errors(buf, len, utf16_output);
  }

  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(const char * buf, size_t len, char16_t* utf16_output) const noexcept final override {
    return set_best()->convert_valid_utf8_to_utf16le(buf, len, utf16_output);
  }

  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(const char * buf, size_t len, char16_t* utf16_output) const noexcept final override {
    return set_best()->convert_valid_utf8_to_utf16be(buf, len, utf16_output);
  }

  simdutf_warn_unused size_t convert_utf8_to_utf32(const char * buf, size_t len, char32_t* utf32_output) const noexcept final override {
    return set_best()->convert_utf8_to_utf32(buf, len, utf32_output);
  }

  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(const char * buf, size_t len, char32_t* utf32_output) const noexcept final override {
    return set_best()->convert_utf8_to_utf32_with_errors(buf, len, utf32_output);
  }

  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(const char * buf, size_t len, char32_t* utf32_output) const noexcept final override {
    return set_best()->convert_valid_utf8_to_utf32(buf, len, utf32_output);
  }

  simdutf_warn_unused size_t convert_utf16le_to_utf8(const char16_t * buf, size_t len, char* utf8_output) const noexcept final override {
    return set_best()->convert_utf16le_to_utf8(buf, len, utf8_output);
  }

  simdutf_warn_unused size_t convert_utf16be_to_utf8(const char16_t * buf, size_t len, char* utf8_output) const noexcept final override {
    return set_best()->convert_utf16be_to_utf8(buf, len, utf8_output);
  }

  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_output) const noexcept final override {
    return set_best()->convert_utf16le_to_utf8_with_errors(buf, len, utf8_output);
  }

  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_output) const noexcept final override {
    return set_best()->convert_utf16be_to_utf8_with_errors(buf, len, utf8_output);
  }

  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(const char16_t * buf, size_t len, char* utf8_output) const noexcept final override {
    return set_best()->convert_valid_utf16le_to_utf8(buf, len, utf8_output);
  }

  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(const char16_t * buf, size_t len, char* utf8_output) const noexcept final override {
    return set_best()->convert_valid_utf16be_to_utf8(buf, len, utf8_output);
  }

  simdutf_warn_unused size_t convert_utf32_to_utf8(const char32_t * buf, size_t len, char* utf8_output) const noexcept final override {
    return set_best()->convert_utf32_to_utf8(buf, len, utf8_output);
  }

  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(const char32_t * buf, size_t len, char* utf8_output) const noexcept final override {
    return set_best()->convert_utf32_to_utf8_with_errors(buf, len, utf8_output);
  }

  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(const char32_t * buf, size_t len, char* utf8_output) const noexcept final override {
    return set_best()->convert_valid_utf32_to_utf8(buf, len, utf8_output);
  }

  simdutf_warn_unused size_t convert_utf32_to_utf16le(const char32_t * buf, size_t len, char16_t* utf16_output) const noexcept final override {
    return set_best()->convert_utf32_to_utf16le(buf, len, utf16_output);
  }

  simdutf_warn_unused size_t convert_utf32_to_utf16be(const char32_t * buf, size_t len, char16_t* utf16_output) const noexcept final override {
    return set_best()->convert_utf32_to_utf16be(buf, len, utf16_output);
  }

  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(const char32_t * buf, size_t len, char16_t* utf16_output) const noexcept final override {
    return set_best()->convert_utf32_to_utf16le_with_errors(buf, len, utf16_output);
  }

  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(const char32_t * buf, size_t len, char16_t* utf16_output) const noexcept final override {
    return set_best()->convert_utf32_to_utf16be_with_errors(buf, len, utf16_output);
  }

  simdutf_warn_unused size_t convert_valid_utf32_to_utf16le(const char32_t * buf, size_t len, char16_t* utf16_output) const noexcept final override {
    return set_best()->convert_valid_utf32_to_utf16le(buf, len, utf16_output);
  }

  simdutf_warn_unused size_t convert_valid_utf32_to_utf16be(const char32_t * buf, size_t len, char16_t* utf16_output) const noexcept final override {
    return set_best()->convert_valid_utf32_to_utf16be(buf, len, utf16_output);
  }

  simdutf_warn_unused size_t convert_utf16le_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_output) const noexcept final override {
    return set_best()->convert_utf16le_to_utf32(buf, len, utf32_output);
  }

  simdutf_warn_unused size_t convert_utf16be_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_output) const noexcept final override {
    return set_best()->convert_utf16be_to_utf32(buf, len, utf32_output);
  }

  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_output) const noexcept final override {
    return set_best()->convert_utf16le_to_utf32_with_errors(buf, len, utf32_output);
  }

  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_output) const noexcept final override {
    return set_best()->convert_utf16be_to_utf32_with_errors(buf, len, utf32_output);
  }

  simdutf_warn_unused size_t convert_valid_utf16le_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_output) const noexcept final override {
    return set_best()->convert_valid_utf16le_to_utf32(buf, len, utf32_output);
  }

  simdutf_warn_unused size_t convert_valid_utf16be_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_output) const noexcept final override {
    return set_best()->convert_valid_utf16be_to_utf32(buf, len, utf32_output);
  }

  void change_endianness_utf16(const char16_t * buf, size_t len, char16_t * output) const noexcept final override {
    set_best()->change_endianness_utf16(buf, len, output);
  }

  simdutf_warn_unused size_t count_utf16le(const char16_t * buf, size_t len) const noexcept final override {
    return set_best()->count_utf16le(buf, len);
  }

  simdutf_warn_unused size_t count_utf16be(const char16_t * buf, size_t len) const noexcept final override {
    return set_best()->count_utf16be(buf, len);
  }

  simdutf_warn_unused size_t count_utf8(const char * buf, size_t len) const noexcept final override {
    return set_best()->count_utf8(buf, len);
  }

  simdutf_warn_unused size_t utf8_length_from_utf16le(const char16_t * buf, size_t len) const noexcept override {
    return set_best()->utf8_length_from_utf16le(buf, len);
  }

  simdutf_warn_unused size_t utf8_length_from_utf16be(const char16_t * buf, size_t len) const noexcept override {
    return set_best()->utf8_length_from_utf16be(buf, len);
  }

  simdutf_warn_unused size_t utf32_length_from_utf16le(const char16_t * buf, size_t len) const noexcept override {
    return set_best()->utf32_length_from_utf16le(buf, len);
  }

  simdutf_warn_unused size_t utf32_length_from_utf16be(const char16_t * buf, size_t len) const noexcept override {
    return set_best()->utf32_length_from_utf16be(buf, len);
  }

  simdutf_warn_unused size_t utf16_length_from_utf8(const char * buf, size_t len) const noexcept override {
    return set_best()->utf16_length_from_utf8(buf, len);
  }

  simdutf_warn_unused size_t utf8_length_from_utf32(const char32_t * buf, size_t len) const noexcept override {
    return set_best()->utf8_length_from_utf32(buf, len);
  }

  simdutf_warn_unused size_t utf16_length_from_utf32(const char32_t * buf, size_t len) const noexcept override {
    return set_best()->utf16_length_from_utf32(buf, len);
  }

  simdutf_warn_unused size_t utf32_length_from_utf8(const char * buf, size_t len) const noexcept override {
    return set_best()->utf32_length_from_utf8(buf, len);
  }

  simdutf_really_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;
};


const std::initializer_list<const implementation *> available_implementation_pointers {
#if SIMDUTF_IMPLEMENTATION_ICELAKE
  &icelake_singleton,
#endif
#if SIMDUTF_IMPLEMENTATION_HASWELL
  &haswell_singleton,
#endif
#if SIMDUTF_IMPLEMENTATION_WESTMERE
  &westmere_singleton,
#endif
#if SIMDUTF_IMPLEMENTATION_ARM64
  &arm64_singleton,
#endif
#if SIMDUTF_IMPLEMENTATION_PPC64
  &ppc64_singleton,
#endif
#if SIMDUTF_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:
  simdutf_warn_unused int detect_encodings(const char *, size_t) const noexcept override {
    return encoding_type::unspecified;
  }

  simdutf_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.
  }

  simdutf_warn_unused result validate_utf8_with_errors(const char *, size_t) const noexcept final override {
    return result(error_code::OTHER, 0);
  }

  simdutf_warn_unused bool validate_ascii(const char *, size_t) const noexcept final override {
    return false;
  }

  simdutf_warn_unused result validate_ascii_with_errors(const char *, size_t) const noexcept final override {
    return result(error_code::OTHER, 0);
  }

  simdutf_warn_unused bool validate_utf16le(const char16_t*, size_t) const noexcept final override {
    return false;
  }

  simdutf_warn_unused bool validate_utf16be(const char16_t*, size_t) const noexcept final override {
    return false;
  }

  simdutf_warn_unused result validate_utf16le_with_errors(const char16_t*, size_t) const noexcept final override {
    return result(error_code::OTHER, 0);
  }

  simdutf_warn_unused result validate_utf16be_with_errors(const char16_t*, size_t) const noexcept final override {
    return result(error_code::OTHER, 0);
  }

  simdutf_warn_unused bool validate_utf32(const char32_t*, size_t) const noexcept final override {
    return false;
  }

  simdutf_warn_unused result validate_utf32_with_errors(const char32_t*, size_t) const noexcept final override {
    return result(error_code::OTHER, 0);
  }

  simdutf_warn_unused size_t convert_utf8_to_utf16le(const char*, size_t, char16_t*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused size_t convert_utf8_to_utf16be(const char*, size_t, char16_t*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(const char*, size_t, char16_t*) const noexcept final override {
    return result(error_code::OTHER, 0);
  }

  simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(const char*, size_t, char16_t*) const noexcept final override {
    return result(error_code::OTHER, 0);
  }

  simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(const char*, size_t, char16_t*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(const char*, size_t, char16_t*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused size_t convert_utf8_to_utf32(const char*, size_t, char32_t*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused result convert_utf8_to_utf32_with_errors(const char*, size_t, char32_t*) const noexcept final override {
    return result(error_code::OTHER, 0);
  }

  simdutf_warn_unused size_t convert_valid_utf8_to_utf32(const char*, size_t, char32_t*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused size_t convert_utf16le_to_utf8(const char16_t*, size_t, char*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused size_t convert_utf16be_to_utf8(const char16_t*, size_t, char*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(const char16_t*, size_t, char*) const noexcept final override {
    return result(error_code::OTHER, 0);
  }

  simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(const char16_t*, size_t, char*) const noexcept final override {
    return result(error_code::OTHER, 0);
  }

  simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(const char16_t*, size_t, char*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(const char16_t*, size_t, char*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused size_t convert_utf32_to_utf8(const char32_t*, size_t, char*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused result convert_utf32_to_utf8_with_errors(const char32_t*, size_t, char*) const noexcept final override {
    return result(error_code::OTHER, 0);
  }

  simdutf_warn_unused size_t convert_valid_utf32_to_utf8(const char32_t*, size_t, char*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused size_t convert_utf32_to_utf16le(const char32_t*, size_t, char16_t*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused size_t convert_utf32_to_utf16be(const char32_t*, size_t, char16_t*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(const char32_t*, size_t, char16_t*) const noexcept final override {
    return result(error_code::OTHER, 0);
  }

  simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(const char32_t*, size_t, char16_t*) const noexcept final override {
    return result(error_code::OTHER, 0);
  }

  simdutf_warn_unused size_t convert_valid_utf32_to_utf16le(const char32_t*, size_t, char16_t*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused size_t convert_valid_utf32_to_utf16be(const char32_t*, size_t, char16_t*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused size_t convert_utf16le_to_utf32(const char16_t*, size_t, char32_t*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused size_t convert_utf16be_to_utf32(const char16_t*, size_t, char32_t*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(const char16_t*, size_t, char32_t*) const noexcept final override {
    return result(error_code::OTHER, 0);
  }

  simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(const char16_t*, size_t, char32_t*) const noexcept final override {
    return result(error_code::OTHER, 0);
  }

  simdutf_warn_unused size_t convert_valid_utf16le_to_utf32(const char16_t*, size_t, char32_t*) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused size_t convert_valid_utf16be_to_utf32(const char16_t*, size_t, char32_t*) const noexcept final override {
    return 0;
  }

  void change_endianness_utf16(const char16_t *, size_t, char16_t *) const noexcept final override {

  }

  simdutf_warn_unused size_t count_utf16le(const char16_t *, size_t) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused size_t count_utf16be(const char16_t *, size_t) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused size_t count_utf8(const char *, size_t) const noexcept final override {
    return 0;
  }

  simdutf_warn_unused size_t utf8_length_from_utf16le(const char16_t *, size_t) const noexcept override {
    return 0;
  }

  simdutf_warn_unused size_t utf8_length_from_utf16be(const char16_t *, size_t) const noexcept override {
    return 0;
  }

  simdutf_warn_unused size_t utf32_length_from_utf16le(const char16_t *, size_t) const noexcept override {
    return 0;
  }

  simdutf_warn_unused size_t utf32_length_from_utf16be(const char16_t *, size_t) const noexcept override {
    return 0;
  }

  simdutf_warn_unused size_t utf16_length_from_utf8(const char *, size_t) const noexcept override {
    return 0;
  }

  simdutf_warn_unused size_t utf8_length_from_utf32(const char32_t *, size_t) const noexcept override {
    return 0;
  }

  simdutf_warn_unused size_t utf16_length_from_utf32(const char32_t *, size_t) const noexcept override {
    return 0;
  }

  simdutf_warn_unused size_t utf32_length_from_utf8(const char *, size_t) const noexcept override {
    return 0;
  }

  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 = internal::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; // this should never happen?
}

const implementation *detect_best_supported_implementation_on_first_use::set_best() const noexcept {
  SIMDUTF_PUSH_DISABLE_WARNINGS
  SIMDUTF_DISABLE_DEPRECATED_WARNING // Disable CRT_SECURE warning on MSVC: manually verified this is safe
  char *force_implementation_name = getenv("SIMDUTF_FORCE_IMPLEMENTATION");
  SIMDUTF_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() = &unsupported_singleton;
    }
  }
  return get_active_implementation() = get_available_implementations().detect_best_supported();
}

} // namespace internal



/**
 * The list of available implementations compiled into simdutf.
 */
SIMDUTF_DLLIMPORTEXPORT const internal::available_implementation_list& get_available_implementations() {
  static const internal::available_implementation_list available_implementations{};
  return available_implementations;
}

/**
  * The active implementation.
  */
SIMDUTF_DLLIMPORTEXPORT internal::atomic_ptr<const implementation>& get_active_implementation() {
    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;
}

simdutf_warn_unused bool validate_utf8(const char *buf, size_t len) noexcept {
  return get_active_implementation()->validate_utf8(buf, len);
}
simdutf_warn_unused result validate_utf8_with_errors(const char *buf, size_t len) noexcept {
  return get_active_implementation()->validate_utf8_with_errors(buf, len);
}
simdutf_warn_unused bool validate_ascii(const char *buf, size_t len) noexcept {
  return get_active_implementation()->validate_ascii(buf, len);
}
simdutf_warn_unused result validate_ascii_with_errors(const char *buf, size_t len) noexcept {
  return get_active_implementation()->validate_ascii_with_errors(buf, len);
}
simdutf_warn_unused size_t convert_utf8_to_utf16(const char * input, size_t length, char16_t* utf16_output) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return convert_utf8_to_utf16be(input, length, utf16_output);
  #else
  return convert_utf8_to_utf16le(input, length, utf16_output);
  #endif
}
simdutf_warn_unused size_t convert_utf8_to_utf16le(const char * input, size_t length, char16_t* utf16_output) noexcept {
  return get_active_implementation()->convert_utf8_to_utf16le(input, length, utf16_output);
}
simdutf_warn_unused size_t convert_utf8_to_utf16be(const char * input, size_t length, char16_t* utf16_output) noexcept {
  return get_active_implementation()->convert_utf8_to_utf16be(input, length, utf16_output);
}
simdutf_warn_unused result convert_utf8_to_utf16_with_errors(const char * input, size_t length, char16_t* utf16_output) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return convert_utf8_to_utf16be_with_errors(input, length, utf16_output);
  #else
  return convert_utf8_to_utf16le_with_errors(input, length, utf16_output);
  #endif
}
simdutf_warn_unused result convert_utf8_to_utf16le_with_errors(const char * input, size_t length, char16_t* utf16_output) noexcept {
  return get_active_implementation()->convert_utf8_to_utf16le_with_errors(input, length, utf16_output);
}
simdutf_warn_unused result convert_utf8_to_utf16be_with_errors(const char * input, size_t length, char16_t* utf16_output) noexcept {
  return get_active_implementation()->convert_utf8_to_utf16be_with_errors(input, length, utf16_output);
}
simdutf_warn_unused size_t convert_utf8_to_utf32(const char * input, size_t length, char32_t* utf32_output) noexcept {
  return get_active_implementation()->convert_utf8_to_utf32(input, length, utf32_output);
}
simdutf_warn_unused result convert_utf8_to_utf32_with_errors(const char * input, size_t length, char32_t* utf32_output) noexcept {
  return get_active_implementation()->convert_utf8_to_utf32_with_errors(input, length, utf32_output);
}
simdutf_warn_unused bool validate_utf16(const char16_t * buf, size_t len) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return validate_utf16be(buf, len);
  #else
  return validate_utf16le(buf, len);
  #endif
}
simdutf_warn_unused bool validate_utf16le(const char16_t * buf, size_t len) noexcept {
  return get_active_implementation()->validate_utf16le(buf, len);
}
simdutf_warn_unused bool validate_utf16be(const char16_t * buf, size_t len) noexcept {
  return get_active_implementation()->validate_utf16be(buf, len);
}
simdutf_warn_unused result validate_utf16_with_errors(const char16_t * buf, size_t len) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return validate_utf16be_with_errors(buf, len);
  #else
  return validate_utf16le_with_errors(buf, len);
  #endif
}
simdutf_warn_unused result validate_utf16le_with_errors(const char16_t * buf, size_t len) noexcept {
  return get_active_implementation()->validate_utf16le_with_errors(buf, len);
}
simdutf_warn_unused result validate_utf16be_with_errors(const char16_t * buf, size_t len) noexcept {
  return get_active_implementation()->validate_utf16be_with_errors(buf, len);
}
simdutf_warn_unused bool validate_utf32(const char32_t * buf, size_t len) noexcept {
  return get_active_implementation()->validate_utf32(buf, len);
}
simdutf_warn_unused result validate_utf32_with_errors(const char32_t * buf, size_t len) noexcept {
  return get_active_implementation()->validate_utf32_with_errors(buf, len);
}
simdutf_warn_unused size_t convert_valid_utf8_to_utf16(const char * input, size_t length, char16_t* utf16_buffer) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return convert_valid_utf8_to_utf16be(input, length, utf16_buffer);
  #else
  return convert_valid_utf8_to_utf16le(input, length, utf16_buffer);
  #endif
}
simdutf_warn_unused size_t convert_valid_utf8_to_utf16le(const char * input, size_t length, char16_t* utf16_buffer) noexcept {
  return get_active_implementation()->convert_valid_utf8_to_utf16le(input, length, utf16_buffer);
}
simdutf_warn_unused size_t convert_valid_utf8_to_utf16be(const char * input, size_t length, char16_t* utf16_buffer) noexcept {
  return get_active_implementation()->convert_valid_utf8_to_utf16be(input, length, utf16_buffer);
}
simdutf_warn_unused size_t convert_valid_utf8_to_utf32(const char * input, size_t length, char32_t* utf32_buffer) noexcept {
  return get_active_implementation()->convert_valid_utf8_to_utf32(input, length, utf32_buffer);
}
simdutf_warn_unused size_t convert_utf16_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return convert_utf16be_to_utf8(buf, len, utf8_buffer);
  #else
  return convert_utf16le_to_utf8(buf, len, utf8_buffer);
  #endif
}
simdutf_warn_unused size_t convert_utf16le_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) noexcept {
  return get_active_implementation()->convert_utf16le_to_utf8(buf, len, utf8_buffer);
}
simdutf_warn_unused size_t convert_utf16be_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) noexcept {
  return get_active_implementation()->convert_utf16be_to_utf8(buf, len, utf8_buffer);
}
simdutf_warn_unused result convert_utf16_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_buffer) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return convert_utf16be_to_utf8_with_errors(buf, len, utf8_buffer);
  #else
  return convert_utf16le_to_utf8_with_errors(buf, len, utf8_buffer);
  #endif
}
simdutf_warn_unused result convert_utf16le_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_buffer) noexcept {
  return get_active_implementation()->convert_utf16le_to_utf8_with_errors(buf, len, utf8_buffer);
}
simdutf_warn_unused result convert_utf16be_to_utf8_with_errors(const char16_t * buf, size_t len, char* utf8_buffer) noexcept {
  return get_active_implementation()->convert_utf16be_to_utf8_with_errors(buf, len, utf8_buffer);
}
simdutf_warn_unused size_t convert_valid_utf16_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return convert_valid_utf16be_to_utf8(buf, len, utf8_buffer);
  #else
  return convert_valid_utf16le_to_utf8(buf, len, utf8_buffer);
  #endif
}
simdutf_warn_unused size_t convert_valid_utf16le_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) noexcept {
  return get_active_implementation()->convert_valid_utf16le_to_utf8(buf, len, utf8_buffer);
}
simdutf_warn_unused size_t convert_valid_utf16be_to_utf8(const char16_t * buf, size_t len, char* utf8_buffer) noexcept {
  return get_active_implementation()->convert_valid_utf16be_to_utf8(buf, len, utf8_buffer);
}
simdutf_warn_unused size_t convert_utf32_to_utf8(const char32_t * buf, size_t len, char* utf8_buffer) noexcept {
  return get_active_implementation()->convert_utf32_to_utf8(buf, len, utf8_buffer);
}
simdutf_warn_unused result convert_utf32_to_utf8_with_errors(const char32_t * buf, size_t len, char* utf8_buffer) noexcept {
  return get_active_implementation()->convert_utf32_to_utf8_with_errors(buf, len, utf8_buffer);
}
simdutf_warn_unused size_t convert_valid_utf32_to_utf8(const char32_t * buf, size_t len, char* utf8_buffer) noexcept {
  return get_active_implementation()->convert_valid_utf32_to_utf8(buf, len, utf8_buffer);
}
simdutf_warn_unused size_t convert_utf32_to_utf16(const char32_t * buf, size_t len, char16_t* utf16_buffer) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return convert_utf32_to_utf16be(buf, len, utf16_buffer);
  #else
  return convert_utf32_to_utf16le(buf, len, utf16_buffer);
  #endif
}
simdutf_warn_unused size_t convert_utf32_to_utf16le(const char32_t * buf, size_t len, char16_t* utf16_buffer) noexcept {
  return get_active_implementation()->convert_utf32_to_utf16le(buf, len, utf16_buffer);
}
simdutf_warn_unused size_t convert_utf32_to_utf16be(const char32_t * buf, size_t len, char16_t* utf16_buffer) noexcept {
  return get_active_implementation()->convert_utf32_to_utf16be(buf, len, utf16_buffer);
}
simdutf_warn_unused result convert_utf32_to_utf16_with_errors(const char32_t * buf, size_t len, char16_t* utf16_buffer) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return convert_utf32_to_utf16be_with_errors(buf, len, utf16_buffer);
  #else
  return convert_utf32_to_utf16le_with_errors(buf, len, utf16_buffer);
  #endif
}
simdutf_warn_unused result convert_utf32_to_utf16le_with_errors(const char32_t * buf, size_t len, char16_t* utf16_buffer) noexcept {
  return get_active_implementation()->convert_utf32_to_utf16le_with_errors(buf, len, utf16_buffer);
}
simdutf_warn_unused result convert_utf32_to_utf16be_with_errors(const char32_t * buf, size_t len, char16_t* utf16_buffer) noexcept {
  return get_active_implementation()->convert_utf32_to_utf16be_with_errors(buf, len, utf16_buffer);
}
simdutf_warn_unused size_t convert_valid_utf32_to_utf16(const char32_t * buf, size_t len, char16_t* utf16_buffer) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return convert_valid_utf32_to_utf16be(buf, len, utf16_buffer);
  #else
  return convert_valid_utf32_to_utf16le(buf, len, utf16_buffer);
  #endif
}
simdutf_warn_unused size_t convert_valid_utf32_to_utf16le(const char32_t * buf, size_t len, char16_t* utf16_buffer) noexcept {
  return get_active_implementation()->convert_valid_utf32_to_utf16le(buf, len, utf16_buffer);
}
simdutf_warn_unused size_t convert_valid_utf32_to_utf16be(const char32_t * buf, size_t len, char16_t* utf16_buffer) noexcept {
  return get_active_implementation()->convert_valid_utf32_to_utf16be(buf, len, utf16_buffer);
}
simdutf_warn_unused size_t convert_utf16_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return convert_utf16be_to_utf32(buf, len, utf32_buffer);
  #else
  return convert_utf16le_to_utf32(buf, len, utf32_buffer);
  #endif
}
simdutf_warn_unused size_t convert_utf16le_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) noexcept {
  return get_active_implementation()->convert_utf16le_to_utf32(buf, len, utf32_buffer);
}
simdutf_warn_unused size_t convert_utf16be_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) noexcept {
  return get_active_implementation()->convert_utf16be_to_utf32(buf, len, utf32_buffer);
}
simdutf_warn_unused result convert_utf16_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_buffer) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return convert_utf16be_to_utf32_with_errors(buf, len, utf32_buffer);
  #else
  return convert_utf16le_to_utf32_with_errors(buf, len, utf32_buffer);
  #endif
}
simdutf_warn_unused result convert_utf16le_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_buffer) noexcept {
  return get_active_implementation()->convert_utf16le_to_utf32_with_errors(buf, len, utf32_buffer);
}
simdutf_warn_unused result convert_utf16be_to_utf32_with_errors(const char16_t * buf, size_t len, char32_t* utf32_buffer) noexcept {
  return get_active_implementation()->convert_utf16be_to_utf32_with_errors(buf, len, utf32_buffer);
}
simdutf_warn_unused size_t convert_valid_utf16_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return convert_valid_utf16be_to_utf32(buf, len, utf32_buffer);
  #else
  return convert_valid_utf16le_to_utf32(buf, len, utf32_buffer);
  #endif
}
simdutf_warn_unused size_t convert_valid_utf16le_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) noexcept {
  return get_active_implementation()->convert_valid_utf16le_to_utf32(buf, len, utf32_buffer);
}
simdutf_warn_unused size_t convert_valid_utf16be_to_utf32(const char16_t * buf, size_t len, char32_t* utf32_buffer) noexcept {
  return get_active_implementation()->convert_valid_utf16be_to_utf32(buf, len, utf32_buffer);
}
void change_endianness_utf16(const char16_t * input, size_t length, char16_t * output) noexcept {
  get_active_implementation()->change_endianness_utf16(input, length, output);
}
simdutf_warn_unused size_t count_utf16(const char16_t * input, size_t length) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return count_utf16be(input, length);
  #else
  return count_utf16le(input, length);
  #endif
}
simdutf_warn_unused size_t count_utf16le(const char16_t * input, size_t length) noexcept {
  return get_active_implementation()->count_utf16le(input, length);
}
simdutf_warn_unused size_t count_utf16be(const char16_t * input, size_t length) noexcept {
  return get_active_implementation()->count_utf16be(input, length);
}
simdutf_warn_unused size_t count_utf8(const char * input, size_t length) noexcept {
  return get_active_implementation()->count_utf8(input, length);
}
simdutf_warn_unused size_t utf8_length_from_utf16(const char16_t * input, size_t length) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return utf8_length_from_utf16be(input, length);
  #else
  return utf8_length_from_utf16le(input, length);
  #endif
}
simdutf_warn_unused size_t utf8_length_from_utf16le(const char16_t * input, size_t length) noexcept {
  return get_active_implementation()->utf8_length_from_utf16le(input, length);
}
simdutf_warn_unused size_t utf8_length_from_utf16be(const char16_t * input, size_t length) noexcept {
  return get_active_implementation()->utf8_length_from_utf16be(input, length);
}
simdutf_warn_unused size_t utf32_length_from_utf16(const char16_t * input, size_t length) noexcept {
  #if SIMDUTF_IS_BIG_ENDIAN
  return utf32_length_from_utf16be(input, length);
  #else
  return utf32_length_from_utf16le(input, length);
  #endif
}
simdutf_warn_unused size_t utf32_length_from_utf16le(const char16_t * input, size_t length) noexcept {
  return get_active_implementation()->utf32_length_from_utf16le(input, length);
}
simdutf_warn_unused size_t utf32_length_from_utf16be(const char16_t * input, size_t length) noexcept {
  return get_active_implementation()->utf32_length_from_utf16be(input, length);
}
simdutf_warn_unused size_t utf16_length_from_utf8(const char * input, size_t length) noexcept {
  return get_active_implementation()->utf16_length_from_utf8(input, length);
}
simdutf_warn_unused size_t utf8_length_from_utf32(const char32_t * input, size_t length) noexcept {
  return get_active_implementation()->utf8_length_from_utf32(input, length);
}
simdutf_warn_unused size_t utf16_length_from_utf32(const char32_t * input, size_t length) noexcept {
  return get_active_implementation()->utf16_length_from_utf32(input, length);
}
simdutf_warn_unused size_t utf32_length_from_utf8(const char * input, size_t length) noexcept {
  return get_active_implementation()->utf32_length_from_utf8(input, length);
}
simdutf_warn_unused simdutf::encoding_type autodetect_encoding(const char * buf, size_t length) noexcept {
  return get_active_implementation()->autodetect_encoding(buf, length);
}
simdutf_warn_unused int detect_encodings(const char * buf, size_t length) noexcept {
  return get_active_implementation()->detect_encodings(buf, length);
}

const implementation * builtin_implementation() {
  static const implementation * builtin_impl = get_available_implementations()[SIMDUTF_STRINGIFY(SIMDUTF_BUILTIN_IMPLEMENTATION)];
  return builtin_impl;
}


} // namespace simdutf

/* end file src/implementation.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=encoding_types.cpp
/* begin file src/encoding_types.cpp */

namespace simdutf {
bool match_system(endianness e) {
#if SIMDUTF_IS_BIG_ENDIAN
    return e == endianness::BIG;
#else
    return e == endianness::LITTLE;
#endif
}

std::string to_string(encoding_type bom) {
  switch (bom) {
      case UTF16_LE:     return "UTF16 little-endian";
      case UTF16_BE:     return "UTF16 big-endian";
      case UTF32_LE:     return "UTF32 little-endian";
      case UTF32_BE:     return "UTF32 big-endian";
      case UTF8:         return "UTF8";
      case unspecified:  return "unknown";
      default:           return "error";
  }
}

namespace BOM {
// Note that BOM for UTF8 is discouraged.
encoding_type check_bom(const uint8_t* byte, size_t length) {
        if (length >= 2 && byte[0] == 0xff and byte[1] == 0xfe) {
            if (length >= 4 && byte[2] == 0x00 and byte[3] == 0x0) {
                return encoding_type::UTF32_LE;
            } else {
                return encoding_type::UTF16_LE;
            }
        } else if (length >= 2 && byte[0] == 0xfe and byte[1] == 0xff) {
            return encoding_type::UTF16_BE;
        } else if (length >= 4 && byte[0] == 0x00 and byte[1] == 0x00 and byte[2] == 0xfe and byte[3] == 0xff) {
            return encoding_type::UTF32_BE;
        } else if (length >= 4 && byte[0] == 0xef and byte[1] == 0xbb and byte[3] == 0xbf) {
            return encoding_type::UTF8;
        }
        return encoding_type::unspecified;
    }

encoding_type check_bom(const char* byte, size_t length) {
      return check_bom(reinterpret_cast<const uint8_t*>(byte), length);
 }

 size_t bom_byte_size(encoding_type bom) {
        switch (bom) {
            case UTF16_LE:     return 2;
            case UTF16_BE:     return 2;
            case UTF32_LE:     return 4;
            case UTF32_BE:     return 4;
            case UTF8:         return 3;
            case unspecified:  return 0;
            default:           return 0;
        }
}

}
}
/* end file src/encoding_types.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=error.cpp
/* begin file src/error.cpp */
namespace simdutf {

  simdutf_really_inline result::result() : error{error_code::SUCCESS}, count{0} {};

  simdutf_really_inline result::result(error_code _err, size_t _pos) : error{_err}, count{_pos} {};

}
/* end file src/error.cpp */
// The large tables should be included once and they
// should not depend on a kernel.
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=tables/utf8_to_utf16_tables.h
/* begin file src/tables/utf8_to_utf16_tables.h */
#ifndef SIMDUTF_UTF8_TO_UTF16_TABLES_H
#define SIMDUTF_UTF8_TO_UTF16_TABLES_H
#include <cstdint>

namespace simdutf {
namespace {
namespace tables {
namespace utf8_to_utf16 {
/**
 * utf8bigindex uses about 8 kB
 * shufutf8 uses about 3344 B
 *
 * So we use a bit over 11 kB. It would be
 * easy to save about 4 kB by only
 * storing the index in utf8bigindex, and
 * deriving the consumed bytes otherwise.
 * However, this may come at a significant (10% to 20%)
 * performance penalty.
 */

const uint8_t shufutf8[209][16] =
{	{0, 255, 1, 255, 2, 255, 3, 255, 4, 255, 5, 255, 0, 0, 0, 0},
 	{0, 255, 1, 255, 2, 255, 3, 255, 4, 255, 6, 5, 0, 0, 0, 0},
 	{0, 255, 1, 255, 2, 255, 3, 255, 5, 4, 6, 255, 0, 0, 0, 0},
 	{0, 255, 1, 255, 2, 255, 3, 255, 5, 4, 7, 6, 0, 0, 0, 0},
 	{0, 255, 1, 255, 2, 255, 4, 3, 5, 255, 6, 255, 0, 0, 0, 0},
 	{0, 255, 1, 255, 2, 255, 4, 3, 5, 255, 7, 6, 0, 0, 0, 0},
 	{0, 255, 1, 255, 2, 255, 4, 3, 6, 5, 7, 255, 0, 0, 0, 0},
 	{0, 255, 1, 255, 2, 255, 4, 3, 6, 5, 8, 7, 0, 0, 0, 0},
 	{0, 255, 1, 255, 3, 2, 4, 255, 5, 255, 6, 255, 0, 0, 0, 0},
 	{0, 255, 1, 255, 3, 2, 4, 255, 5, 255, 7, 6, 0, 0, 0, 0},
 	{0, 255, 1, 255, 3, 2, 4, 255, 6, 5, 7, 255, 0, 0, 0, 0},
 	{0, 255, 1, 255, 3, 2, 4, 255, 6, 5, 8, 7, 0, 0, 0, 0},
 	{0, 255, 1, 255, 3, 2, 5, 4, 6, 255, 7, 255, 0, 0, 0, 0},
 	{0, 255, 1, 255, 3, 2, 5, 4, 6, 255, 8, 7, 0, 0, 0, 0},
 	{0, 255, 1, 255, 3, 2, 5, 4, 7, 6, 8, 255, 0, 0, 0, 0},
 	{0, 255, 1, 255, 3, 2, 5, 4, 7, 6, 9, 8, 0, 0, 0, 0},
 	{0, 255, 2, 1, 3, 255, 4, 255, 5, 255, 6, 255, 0, 0, 0, 0},
 	{0, 255, 2, 1, 3, 255, 4, 255, 5, 255, 7, 6, 0, 0, 0, 0},
 	{0, 255, 2, 1, 3, 255, 4, 255, 6, 5, 7, 255, 0, 0, 0, 0},
 	{0, 255, 2, 1, 3, 255, 4, 255, 6, 5, 8, 7, 0, 0, 0, 0},
 	{0, 255, 2, 1, 3, 255, 5, 4, 6, 255, 7, 255, 0, 0, 0, 0},
 	{0, 255, 2, 1, 3, 255, 5, 4, 6, 255, 8, 7, 0, 0, 0, 0},
 	{0, 255, 2, 1, 3, 255, 5, 4, 7, 6, 8, 255, 0, 0, 0, 0},
 	{0, 255, 2, 1, 3, 255, 5, 4, 7, 6, 9, 8, 0, 0, 0, 0},
 	{0, 255, 2, 1, 4, 3, 5, 255, 6, 255, 7, 255, 0, 0, 0, 0},
 	{0, 255, 2, 1, 4, 3, 5, 255, 6, 255, 8, 7, 0, 0, 0, 0},
 	{0, 255, 2, 1, 4, 3, 5, 255, 7, 6, 8, 255, 0, 0, 0, 0},
 	{0, 255, 2, 1, 4, 3, 5, 255, 7, 6, 9, 8, 0, 0, 0, 0},
 	{0, 255, 2, 1, 4, 3, 6, 5, 7, 255, 8, 255, 0, 0, 0, 0},
 	{0, 255, 2, 1, 4, 3, 6, 5, 7, 255, 9, 8, 0, 0, 0, 0},
 	{0, 255, 2, 1, 4, 3, 6, 5, 8, 7, 9, 255, 0, 0, 0, 0},
 	{0, 255, 2, 1, 4, 3, 6, 5, 8, 7, 10, 9, 0, 0, 0, 0},
 	{1, 0, 2, 255, 3, 255, 4, 255, 5, 255, 6, 255, 0, 0, 0, 0},
 	{1, 0, 2, 255, 3, 255, 4, 255, 5, 255, 7, 6, 0, 0, 0, 0},
 	{1, 0, 2, 255, 3, 255, 4, 255, 6, 5, 7, 255, 0, 0, 0, 0},
 	{1, 0, 2, 255, 3, 255, 4, 255, 6, 5, 8, 7, 0, 0, 0, 0},
 	{1, 0, 2, 255, 3, 255, 5, 4, 6, 255, 7, 255, 0, 0, 0, 0},
 	{1, 0, 2, 255, 3, 255, 5, 4, 6, 255, 8, 7, 0, 0, 0, 0},
 	{1, 0, 2, 255, 3, 255, 5, 4, 7, 6, 8, 255, 0, 0, 0, 0},
 	{1, 0, 2, 255, 3, 255, 5, 4, 7, 6, 9, 8, 0, 0, 0, 0},
 	{1, 0, 2, 255, 4, 3, 5, 255, 6, 255, 7, 255, 0, 0, 0, 0},
 	{1, 0, 2, 255, 4, 3, 5, 255, 6, 255, 8, 7, 0, 0, 0, 0},
 	{1, 0, 2, 255, 4, 3, 5, 255, 7, 6, 8, 255, 0, 0, 0, 0},
 	{1, 0, 2, 255, 4, 3, 5, 255, 7, 6, 9, 8, 0, 0, 0, 0},
 	{1, 0, 2, 255, 4, 3, 6, 5, 7, 255, 8, 255, 0, 0, 0, 0},
 	{1, 0, 2, 255, 4, 3, 6, 5, 7, 255, 9, 8, 0, 0, 0, 0},
 	{1, 0, 2, 255, 4, 3, 6, 5, 8, 7, 9, 255, 0, 0, 0, 0},
 	{1, 0, 2, 255, 4, 3, 6, 5, 8, 7, 10, 9, 0, 0, 0, 0},
 	{1, 0, 3, 2, 4, 255, 5, 255, 6, 255, 7, 255, 0, 0, 0, 0},
 	{1, 0, 3, 2, 4, 255, 5, 255, 6, 255, 8, 7, 0, 0, 0, 0},
 	{1, 0, 3, 2, 4, 255, 5, 255, 7, 6, 8, 255, 0, 0, 0, 0},
 	{1, 0, 3, 2, 4, 255, 5, 255, 7, 6, 9, 8, 0, 0, 0, 0},
 	{1, 0, 3, 2, 4, 255, 6, 5, 7, 255, 8, 255, 0, 0, 0, 0},
 	{1, 0, 3, 2, 4, 255, 6, 5, 7, 255, 9, 8, 0, 0, 0, 0},
 	{1, 0, 3, 2, 4, 255, 6, 5, 8, 7, 9, 255, 0, 0, 0, 0},
 	{1, 0, 3, 2, 4, 255, 6, 5, 8, 7, 10, 9, 0, 0, 0, 0},
 	{1, 0, 3, 2, 5, 4, 6, 255, 7, 255, 8, 255, 0, 0, 0, 0},
 	{1, 0, 3, 2, 5, 4, 6, 255, 7, 255, 9, 8, 0, 0, 0, 0},
 	{1, 0, 3, 2, 5, 4, 6, 255, 8, 7, 9, 255, 0, 0, 0, 0},
 	{1, 0, 3, 2, 5, 4, 6, 255, 8, 7, 10, 9, 0, 0, 0, 0},
 	{1, 0, 3, 2, 5, 4, 7, 6, 8, 255, 9, 255, 0, 0, 0, 0},
 	{1, 0, 3, 2, 5, 4, 7, 6, 8, 255, 10, 9, 0, 0, 0, 0},
 	{1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 10, 255, 0, 0, 0, 0},
 	{1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 0, 0, 0, 0},
 	{0, 255, 255, 255, 1, 255, 255, 255, 2, 255, 255, 255, 3, 255, 255, 255},
 	{0, 255, 255, 255, 1, 255, 255, 255, 2, 255, 255, 255, 4, 3, 255, 255},
 	{0, 255, 255, 255, 1, 255, 255, 255, 2, 255, 255, 255, 5, 4, 3, 255},
 	{0, 255, 255, 255, 1, 255, 255, 255, 3, 2, 255, 255, 4, 255, 255, 255},
 	{0, 255, 255, 255, 1, 255, 255, 255, 3, 2, 255, 255, 5, 4, 255, 255},
 	{0, 255, 255, 255, 1, 255, 255, 255, 3, 2, 255, 255, 6, 5, 4, 255},
 	{0, 255, 255, 255, 1, 255, 255, 255, 4, 3, 2, 255, 5, 255, 255, 255},
 	{0, 255, 255, 255, 1, 255, 255, 255, 4, 3, 2, 255, 6, 5, 255, 255},
 	{0, 255, 255, 255, 1, 255, 255, 255, 4, 3, 2, 255, 7, 6, 5, 255},
 	{0, 255, 255, 255, 2, 1, 255, 255, 3, 255, 255, 255, 4, 255, 255, 255},
 	{0, 255, 255, 255, 2, 1, 255, 255, 3, 255, 255, 255, 5, 4, 255, 255},
 	{0, 255, 255, 255, 2, 1, 255, 255, 3, 255, 255, 255, 6, 5, 4, 255},
 	{0, 255, 255, 255, 2, 1, 255, 255, 4, 3, 255, 255, 5, 255, 255, 255},
 	{0, 255, 255, 255, 2, 1, 255, 255, 4, 3, 255, 255, 6, 5, 255, 255},
 	{0, 255, 255, 255, 2, 1, 255, 255, 4, 3, 255, 255, 7, 6, 5, 255},
 	{0, 255, 255, 255, 2, 1, 255, 255, 5, 4, 3, 255, 6, 255, 255, 255},
 	{0, 255, 255, 255, 2, 1, 255, 255, 5, 4, 3, 255, 7, 6, 255, 255},
 	{0, 255, 255, 255, 2, 1, 255, 255, 5, 4, 3, 255, 8, 7, 6, 255},
 	{0, 255, 255, 255, 3, 2, 1, 255, 4, 255, 255, 255, 5, 255, 255, 255},
 	{0, 255, 255, 255, 3, 2, 1, 255, 4, 255, 255, 255, 6, 5, 255, 255},
 	{0, 255, 255, 255, 3, 2, 1, 255, 4, 255, 255, 255, 7, 6, 5, 255},
 	{0, 255, 255, 255, 3, 2, 1, 255, 5, 4, 255, 255, 6, 255, 255, 255},
 	{0, 255, 255, 255, 3, 2, 1, 255, 5, 4, 255, 255, 7, 6, 255, 255},
 	{0, 255, 255, 255, 3, 2, 1, 255, 5, 4, 255, 255, 8, 7, 6, 255},
 	{0, 255, 255, 255, 3, 2, 1, 255, 6, 5, 4, 255, 7, 255, 255, 255},
 	{0, 255, 255, 255, 3, 2, 1, 255, 6, 5, 4, 255, 8, 7, 255, 255},
 	{0, 255, 255, 255, 3, 2, 1, 255, 6, 5, 4, 255, 9, 8, 7, 255},
 	{1, 0, 255, 255, 2, 255, 255, 255, 3, 255, 255, 255, 4, 255, 255, 255},
 	{1, 0, 255, 255, 2, 255, 255, 255, 3, 255, 255, 255, 5, 4, 255, 255},
 	{1, 0, 255, 255, 2, 255, 255, 255, 3, 255, 255, 255, 6, 5, 4, 255},
 	{1, 0, 255, 255, 2, 255, 255, 255, 4, 3, 255, 255, 5, 255, 255, 255},
 	{1, 0, 255, 255, 2, 255, 255, 255, 4, 3, 255, 255, 6, 5, 255, 255},
 	{1, 0, 255, 255, 2, 255, 255, 255, 4, 3, 255, 255, 7, 6, 5, 255},
 	{1, 0, 255, 255, 2, 255, 255, 255, 5, 4, 3, 255, 6, 255, 255, 255},
 	{1, 0, 255, 255, 2, 255, 255, 255, 5, 4, 3, 255, 7, 6, 255, 255},
 	{1, 0, 255, 255, 2, 255, 255, 255, 5, 4, 3, 255, 8, 7, 6, 255},
 	{1, 0, 255, 255, 3, 2, 255, 255, 4, 255, 255, 255, 5, 255, 255, 255},
 	{1, 0, 255, 255, 3, 2, 255, 255, 4, 255, 255, 255, 6, 5, 255, 255},
 	{1, 0, 255, 255, 3, 2, 255, 255, 4, 255, 255, 255, 7, 6, 5, 255},
 	{1, 0, 255, 255, 3, 2, 255, 255, 5, 4, 255, 255, 6, 255, 255, 255},
 	{1, 0, 255, 255, 3, 2, 255, 255, 5, 4, 255, 255, 7, 6, 255, 255},
 	{1, 0, 255, 255, 3, 2, 255, 255, 5, 4, 255, 255, 8, 7, 6, 255},
 	{1, 0, 255, 255, 3, 2, 255, 255, 6, 5, 4, 255, 7, 255, 255, 255},
 	{1, 0, 255, 255, 3, 2, 255, 255, 6, 5, 4, 255, 8, 7, 255, 255},
 	{1, 0, 255, 255, 3, 2, 255, 255, 6, 5, 4, 255, 9, 8, 7, 255},
 	{1, 0, 255, 255, 4, 3, 2, 255, 5, 255, 255, 255, 6, 255, 255, 255},
 	{1, 0, 255, 255, 4, 3, 2, 255, 5, 255, 255, 255, 7, 6, 255, 255},
 	{1, 0, 255, 255, 4, 3, 2, 255, 5, 255, 255, 255, 8, 7, 6, 255},
 	{1, 0, 255, 255, 4, 3, 2, 255, 6, 5, 255, 255, 7, 255, 255, 255},
 	{1, 0, 255, 255, 4, 3, 2, 255, 6, 5, 255, 255, 8, 7, 255, 255},
 	{1, 0, 255, 255, 4, 3, 2, 255, 6, 5, 255, 255, 9, 8, 7, 255},
 	{1, 0, 255, 255, 4, 3, 2, 255, 7, 6, 5, 255, 8, 255, 255, 255},
 	{1, 0, 255, 255, 4, 3, 2, 255, 7, 6, 5, 255, 9, 8, 255, 255},
 	{1, 0, 255, 255, 4, 3, 2, 255, 7, 6, 5, 255, 10, 9, 8, 255},
 	{2, 1, 0, 255, 3, 255, 255, 255, 4, 255, 255, 255, 5, 255, 255, 255},
 	{2, 1, 0, 255, 3, 255, 255, 255, 4, 255, 255, 255, 6, 5, 255, 255},
 	{2, 1, 0, 255, 3, 255, 255, 255, 4, 255, 255, 255, 7, 6, 5, 255},
 	{2, 1, 0, 255, 3, 255, 255, 255, 5, 4, 255, 255, 6, 255, 255, 255},
 	{2, 1, 0, 255, 3, 255, 255, 255, 5, 4, 255, 255, 7, 6, 255, 255},
 	{2, 1, 0, 255, 3, 255, 255, 255, 5, 4, 255, 255, 8, 7, 6, 255},
 	{2, 1, 0, 255, 3, 255, 255, 255, 6, 5, 4, 255, 7, 255, 255, 255},
 	{2, 1, 0, 255, 3, 255, 255, 255, 6, 5, 4, 255, 8, 7, 255, 255},
 	{2, 1, 0, 255, 3, 255, 255, 255, 6, 5, 4, 255, 9, 8, 7, 255},
 	{2, 1, 0, 255, 4, 3, 255, 255, 5, 255, 255, 255, 6, 255, 255, 255},
 	{2, 1, 0, 255, 4, 3, 255, 255, 5, 255, 255, 255, 7, 6, 255, 255},
 	{2, 1, 0, 255, 4, 3, 255, 255, 5, 255, 255, 255, 8, 7, 6, 255},
 	{2, 1, 0, 255, 4, 3, 255, 255, 6, 5, 255, 255, 7, 255, 255, 255},
 	{2, 1, 0, 255, 4, 3, 255, 255, 6, 5, 255, 255, 8, 7, 255, 255},
 	{2, 1, 0, 255, 4, 3, 255, 255, 6, 5, 255, 255, 9, 8, 7, 255},
 	{2, 1, 0, 255, 4, 3, 255, 255, 7, 6, 5, 255, 8, 255, 255, 255},
 	{2, 1, 0, 255, 4, 3, 255, 255, 7, 6, 5, 255, 9, 8, 255, 255},
 	{2, 1, 0, 255, 4, 3, 255, 255, 7, 6, 5, 255, 10, 9, 8, 255},
 	{2, 1, 0, 255, 5, 4, 3, 255, 6, 255, 255, 255, 7, 255, 255, 255},
 	{2, 1, 0, 255, 5, 4, 3, 255, 6, 255, 255, 255, 8, 7, 255, 255},
 	{2, 1, 0, 255, 5, 4, 3, 255, 6, 255, 255, 255, 9, 8, 7, 255},
 	{2, 1, 0, 255, 5, 4, 3, 255, 7, 6, 255, 255, 8, 255, 255, 255},
 	{2, 1, 0, 255, 5, 4, 3, 255, 7, 6, 255, 255, 9, 8, 255, 255},
 	{2, 1, 0, 255, 5, 4, 3, 255, 7, 6, 255, 255, 10, 9, 8, 255},
 	{2, 1, 0, 255, 5, 4, 3, 255, 8, 7, 6, 255, 9, 255, 255, 255},
 	{2, 1, 0, 255, 5, 4, 3, 255, 8, 7, 6, 255, 10, 9, 255, 255},
 	{2, 1, 0, 255, 5, 4, 3, 255, 8, 7, 6, 255, 11, 10, 9, 255},
 	{0, 255, 255, 255, 1, 255, 255, 255, 2, 255, 255, 255, 0, 0, 0, 0},
 	{0, 255, 255, 255, 1, 255, 255, 255, 3, 2, 255, 255, 0, 0, 0, 0},
 	{0, 255, 255, 255, 1, 255, 255, 255, 4, 3, 2, 255, 0, 0, 0, 0},
 	{0, 255, 255, 255, 1, 255, 255, 255, 5, 4, 3, 2, 0, 0, 0, 0},
 	{0, 255, 255, 255, 2, 1, 255, 255, 3, 255, 255, 255, 0, 0, 0, 0},
 	{0, 255, 255, 255, 2, 1, 255, 255, 4, 3, 255, 255, 0, 0, 0, 0},
 	{0, 255, 255, 255, 2, 1, 255, 255, 5, 4, 3, 255, 0, 0, 0, 0},
 	{0, 255, 255, 255, 2, 1, 255, 255, 6, 5, 4, 3, 0, 0, 0, 0},
 	{0, 255, 255, 255, 3, 2, 1, 255, 4, 255, 255, 255, 0, 0, 0, 0},
 	{0, 255, 255, 255, 3, 2, 1, 255, 5, 4, 255, 255, 0, 0, 0, 0},
 	{0, 255, 255, 255, 3, 2, 1, 255, 6, 5, 4, 255, 0, 0, 0, 0},
 	{0, 255, 255, 255, 3, 2, 1, 255, 7, 6, 5, 4, 0, 0, 0, 0},
 	{0, 255, 255, 255, 4, 3, 2, 1, 5, 255, 255, 255, 0, 0, 0, 0},
 	{0, 255, 255, 255, 4, 3, 2, 1, 6, 5, 255, 255, 0, 0, 0, 0},
 	{0, 255, 255, 255, 4, 3, 2, 1, 7, 6, 5, 255, 0, 0, 0, 0},
 	{0, 255, 255, 255, 4, 3, 2, 1, 8, 7, 6, 5, 0, 0, 0, 0},
 	{1, 0, 255, 255, 2, 255, 255, 255, 3, 255, 255, 255, 0, 0, 0, 0},
 	{1, 0, 255, 255, 2, 255, 255, 255, 4, 3, 255, 255, 0, 0, 0, 0},
 	{1, 0, 255, 255, 2, 255, 255, 255, 5, 4, 3, 255, 0, 0, 0, 0},
 	{1, 0, 255, 255, 2, 255, 255, 255, 6, 5, 4, 3, 0, 0, 0, 0},
 	{1, 0, 255, 255, 3, 2, 255, 255, 4, 255, 255, 255, 0, 0, 0, 0},
 	{1, 0, 255, 255, 3, 2, 255, 255, 5, 4, 255, 255, 0, 0, 0, 0},
 	{1, 0, 255, 255, 3, 2, 255, 255, 6, 5, 4, 255, 0, 0, 0, 0},
 	{1, 0, 255, 255, 3, 2, 255, 255, 7, 6, 5, 4, 0, 0, 0, 0},
 	{1, 0, 255, 255, 4, 3, 2, 255, 5, 255, 255, 255, 0, 0, 0, 0},
 	{1, 0, 255, 255, 4, 3, 2, 255, 6, 5, 255, 255, 0, 0, 0, 0},
 	{1, 0, 255, 255, 4, 3, 2, 255, 7, 6, 5, 255, 0, 0, 0, 0},
 	{1, 0, 255, 255, 4, 3, 2, 255, 8, 7, 6, 5, 0, 0, 0, 0},
 	{1, 0, 255, 255, 5, 4, 3, 2, 6, 255, 255, 255, 0, 0, 0, 0},
 	{1, 0, 255, 255, 5, 4, 3, 2, 7, 6, 255, 255, 0, 0, 0, 0},
 	{1, 0, 255, 255, 5, 4, 3, 2, 8, 7, 6, 255, 0, 0, 0, 0},
 	{1, 0, 255, 255, 5, 4, 3, 2, 9, 8, 7, 6, 0, 0, 0, 0},
 	{2, 1, 0, 255, 3, 255, 255, 255, 4, 255, 255, 255, 0, 0, 0, 0},
 	{2, 1, 0, 255, 3, 255, 255, 255, 5, 4, 255, 255, 0, 0, 0, 0},
 	{2, 1, 0, 255, 3, 255, 255, 255, 6, 5, 4, 255, 0, 0, 0, 0},
 	{2, 1, 0, 255, 3, 255, 255, 255, 7, 6, 5, 4, 0, 0, 0, 0},
 	{2, 1, 0, 255, 4, 3, 255, 255, 5, 255, 255, 255, 0, 0, 0, 0},
 	{2, 1, 0, 255, 4, 3, 255, 255, 6, 5, 255, 255, 0, 0, 0, 0},
 	{2, 1, 0, 255, 4, 3, 255, 255, 7, 6, 5, 255, 0, 0, 0, 0},
 	{2, 1, 0, 255, 4, 3, 255, 255, 8, 7, 6, 5, 0, 0, 0, 0},
 	{2, 1, 0, 255, 5, 4, 3, 255, 6, 255, 255, 255, 0, 0, 0, 0},
 	{2, 1, 0, 255, 5, 4, 3, 255, 7, 6, 255, 255, 0, 0, 0, 0},
 	{2, 1, 0, 255, 5, 4, 3, 255, 8, 7, 6, 255, 0, 0, 0, 0},
 	{2, 1, 0, 255, 5, 4, 3, 255, 9, 8, 7, 6, 0, 0, 0, 0},
 	{2, 1, 0, 255, 6, 5, 4, 3, 7, 255, 255, 255, 0, 0, 0, 0},
 	{2, 1, 0, 255, 6, 5, 4, 3, 8, 7, 255, 255, 0, 0, 0, 0},
 	{2, 1, 0, 255, 6, 5, 4, 3, 9, 8, 7, 255, 0, 0, 0, 0},
 	{2, 1, 0, 255, 6, 5, 4, 3, 10, 9, 8, 7, 0, 0, 0, 0},
 	{3, 2, 1, 0, 4, 255, 255, 255, 5, 255, 255, 255, 0, 0, 0, 0},
 	{3, 2, 1, 0, 4, 255, 255, 255, 6, 5, 255, 255, 0, 0, 0, 0},
 	{3, 2, 1, 0, 4, 255, 255, 255, 7, 6, 5, 255, 0, 0, 0, 0},
 	{3, 2, 1, 0, 4, 255, 255, 255, 8, 7, 6, 5, 0, 0, 0, 0},
 	{3, 2, 1, 0, 5, 4, 255, 255, 6, 255, 255, 255, 0, 0, 0, 0},
 	{3, 2, 1, 0, 5, 4, 255, 255, 7, 6, 255, 255, 0, 0, 0, 0},
 	{3, 2, 1, 0, 5, 4, 255, 255, 8, 7, 6, 255, 0, 0, 0, 0},
 	{3, 2, 1, 0, 5, 4, 255, 255, 9, 8, 7, 6, 0, 0, 0, 0},
 	{3, 2, 1, 0, 6, 5, 4, 255, 7, 255, 255, 255, 0, 0, 0, 0},
 	{3, 2, 1, 0, 6, 5, 4, 255, 8, 7, 255, 255, 0, 0, 0, 0},
 	{3, 2, 1, 0, 6, 5, 4, 255, 9, 8, 7, 255, 0, 0, 0, 0},
 	{3, 2, 1, 0, 6, 5, 4, 255, 10, 9, 8, 7, 0, 0, 0, 0},
 	{3, 2, 1, 0, 7, 6, 5, 4, 8, 255, 255, 255, 0, 0, 0, 0},
 	{3, 2, 1, 0, 7, 6, 5, 4, 9, 8, 255, 255, 0, 0, 0, 0},
 	{3, 2, 1, 0, 7, 6, 5, 4, 10, 9, 8, 255, 0, 0, 0, 0},
 	{3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 0, 0, 0, 0}};
/* number of two bytes : 64 */
/* number of two + three bytes : 145 */
/* number of two + three + four bytes : 209 */
const uint8_t utf8bigindex[4096][2] =
{	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{146, 4},
 	{209, 12},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{147, 5},
 	{209, 12},
 	{150, 5},
 	{162, 5},
 	{65, 5},
 	{209, 12},
 	{153, 5},
 	{165, 5},
 	{67, 5},
 	{177, 5},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{148, 6},
 	{209, 12},
 	{151, 6},
 	{163, 6},
 	{66, 6},
 	{209, 12},
 	{154, 6},
 	{166, 6},
 	{68, 6},
 	{178, 6},
 	{74, 6},
 	{92, 6},
 	{64, 4},
 	{209, 12},
 	{157, 6},
 	{169, 6},
 	{70, 6},
 	{181, 6},
 	{76, 6},
 	{94, 6},
 	{65, 5},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{67, 5},
 	{118, 6},
 	{73, 5},
 	{91, 5},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{209, 12},
 	{155, 7},
 	{167, 7},
 	{69, 7},
 	{179, 7},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{170, 7},
 	{71, 7},
 	{182, 7},
 	{77, 7},
 	{95, 7},
 	{65, 5},
 	{194, 7},
 	{83, 7},
 	{101, 7},
 	{67, 5},
 	{119, 7},
 	{73, 5},
 	{91, 5},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{185, 7},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{103, 7},
 	{68, 6},
 	{121, 7},
 	{74, 6},
 	{92, 6},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{76, 6},
 	{94, 6},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{156, 8},
 	{168, 8},
 	{146, 4},
 	{180, 8},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{159, 8},
 	{171, 8},
 	{72, 8},
 	{183, 8},
 	{78, 8},
 	{96, 8},
 	{65, 5},
 	{195, 8},
 	{84, 8},
 	{102, 8},
 	{67, 5},
 	{120, 8},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{174, 8},
 	{148, 6},
 	{186, 8},
 	{80, 8},
 	{98, 8},
 	{66, 6},
 	{198, 8},
 	{86, 8},
 	{104, 8},
 	{68, 6},
 	{122, 8},
 	{74, 6},
 	{92, 6},
 	{3, 8},
 	{209, 12},
 	{157, 6},
 	{110, 8},
 	{70, 6},
 	{128, 8},
 	{76, 6},
 	{94, 6},
 	{5, 8},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{9, 8},
 	{118, 6},
 	{17, 8},
 	{33, 8},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{189, 8},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{201, 8},
 	{88, 8},
 	{106, 8},
 	{69, 7},
 	{124, 8},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{112, 8},
 	{71, 7},
 	{130, 8},
 	{77, 7},
 	{95, 7},
 	{6, 8},
 	{194, 7},
 	{83, 7},
 	{101, 7},
 	{10, 8},
 	{119, 7},
 	{18, 8},
 	{34, 8},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{136, 8},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{103, 7},
 	{12, 8},
 	{121, 7},
 	{20, 8},
 	{36, 8},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{24, 8},
 	{40, 8},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{48, 8},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{146, 4},
 	{209, 12},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{160, 9},
 	{172, 9},
 	{147, 5},
 	{184, 9},
 	{150, 5},
 	{162, 5},
 	{65, 5},
 	{196, 9},
 	{153, 5},
 	{165, 5},
 	{67, 5},
 	{177, 5},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{175, 9},
 	{148, 6},
 	{187, 9},
 	{81, 9},
 	{99, 9},
 	{66, 6},
 	{199, 9},
 	{87, 9},
 	{105, 9},
 	{68, 6},
 	{123, 9},
 	{74, 6},
 	{92, 6},
 	{64, 4},
 	{209, 12},
 	{157, 6},
 	{111, 9},
 	{70, 6},
 	{129, 9},
 	{76, 6},
 	{94, 6},
 	{65, 5},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{67, 5},
 	{118, 6},
 	{73, 5},
 	{91, 5},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{190, 9},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{202, 9},
 	{89, 9},
 	{107, 9},
 	{69, 7},
 	{125, 9},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{113, 9},
 	{71, 7},
 	{131, 9},
 	{77, 7},
 	{95, 7},
 	{7, 9},
 	{194, 7},
 	{83, 7},
 	{101, 7},
 	{11, 9},
 	{119, 7},
 	{19, 9},
 	{35, 9},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{137, 9},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{103, 7},
 	{13, 9},
 	{121, 7},
 	{21, 9},
 	{37, 9},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{25, 9},
 	{41, 9},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{49, 9},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{205, 9},
 	{156, 8},
 	{168, 8},
 	{146, 4},
 	{180, 8},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{159, 8},
 	{115, 9},
 	{72, 8},
 	{133, 9},
 	{78, 8},
 	{96, 8},
 	{65, 5},
 	{195, 8},
 	{84, 8},
 	{102, 8},
 	{67, 5},
 	{120, 8},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{174, 8},
 	{148, 6},
 	{139, 9},
 	{80, 8},
 	{98, 8},
 	{66, 6},
 	{198, 8},
 	{86, 8},
 	{104, 8},
 	{14, 9},
 	{122, 8},
 	{22, 9},
 	{38, 9},
 	{3, 8},
 	{209, 12},
 	{157, 6},
 	{110, 8},
 	{70, 6},
 	{128, 8},
 	{26, 9},
 	{42, 9},
 	{5, 8},
 	{193, 6},
 	{82, 6},
 	{50, 9},
 	{9, 8},
 	{118, 6},
 	{17, 8},
 	{33, 8},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{189, 8},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{201, 8},
 	{88, 8},
 	{106, 8},
 	{69, 7},
 	{124, 8},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{112, 8},
 	{71, 7},
 	{130, 8},
 	{28, 9},
 	{44, 9},
 	{6, 8},
 	{194, 7},
 	{83, 7},
 	{52, 9},
 	{10, 8},
 	{119, 7},
 	{18, 8},
 	{34, 8},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{136, 8},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{56, 9},
 	{12, 8},
 	{121, 7},
 	{20, 8},
 	{36, 8},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{24, 8},
 	{40, 8},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{48, 8},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{146, 4},
 	{209, 12},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{147, 5},
 	{209, 12},
 	{150, 5},
 	{162, 5},
 	{65, 5},
 	{209, 12},
 	{153, 5},
 	{165, 5},
 	{67, 5},
 	{177, 5},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{176, 10},
 	{148, 6},
 	{188, 10},
 	{151, 6},
 	{163, 6},
 	{66, 6},
 	{200, 10},
 	{154, 6},
 	{166, 6},
 	{68, 6},
 	{178, 6},
 	{74, 6},
 	{92, 6},
 	{64, 4},
 	{209, 12},
 	{157, 6},
 	{169, 6},
 	{70, 6},
 	{181, 6},
 	{76, 6},
 	{94, 6},
 	{65, 5},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{67, 5},
 	{118, 6},
 	{73, 5},
 	{91, 5},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{191, 10},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{203, 10},
 	{90, 10},
 	{108, 10},
 	{69, 7},
 	{126, 10},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{114, 10},
 	{71, 7},
 	{132, 10},
 	{77, 7},
 	{95, 7},
 	{65, 5},
 	{194, 7},
 	{83, 7},
 	{101, 7},
 	{67, 5},
 	{119, 7},
 	{73, 5},
 	{91, 5},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{138, 10},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{103, 7},
 	{68, 6},
 	{121, 7},
 	{74, 6},
 	{92, 6},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{76, 6},
 	{94, 6},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{206, 10},
 	{156, 8},
 	{168, 8},
 	{146, 4},
 	{180, 8},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{159, 8},
 	{116, 10},
 	{72, 8},
 	{134, 10},
 	{78, 8},
 	{96, 8},
 	{65, 5},
 	{195, 8},
 	{84, 8},
 	{102, 8},
 	{67, 5},
 	{120, 8},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{174, 8},
 	{148, 6},
 	{140, 10},
 	{80, 8},
 	{98, 8},
 	{66, 6},
 	{198, 8},
 	{86, 8},
 	{104, 8},
 	{15, 10},
 	{122, 8},
 	{23, 10},
 	{39, 10},
 	{3, 8},
 	{209, 12},
 	{157, 6},
 	{110, 8},
 	{70, 6},
 	{128, 8},
 	{27, 10},
 	{43, 10},
 	{5, 8},
 	{193, 6},
 	{82, 6},
 	{51, 10},
 	{9, 8},
 	{118, 6},
 	{17, 8},
 	{33, 8},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{189, 8},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{201, 8},
 	{88, 8},
 	{106, 8},
 	{69, 7},
 	{124, 8},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{112, 8},
 	{71, 7},
 	{130, 8},
 	{29, 10},
 	{45, 10},
 	{6, 8},
 	{194, 7},
 	{83, 7},
 	{53, 10},
 	{10, 8},
 	{119, 7},
 	{18, 8},
 	{34, 8},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{136, 8},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{57, 10},
 	{12, 8},
 	{121, 7},
 	{20, 8},
 	{36, 8},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{24, 8},
 	{40, 8},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{48, 8},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{146, 4},
 	{209, 12},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{160, 9},
 	{172, 9},
 	{147, 5},
 	{184, 9},
 	{150, 5},
 	{162, 5},
 	{65, 5},
 	{196, 9},
 	{153, 5},
 	{165, 5},
 	{67, 5},
 	{177, 5},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{175, 9},
 	{148, 6},
 	{142, 10},
 	{81, 9},
 	{99, 9},
 	{66, 6},
 	{199, 9},
 	{87, 9},
 	{105, 9},
 	{68, 6},
 	{123, 9},
 	{74, 6},
 	{92, 6},
 	{64, 4},
 	{209, 12},
 	{157, 6},
 	{111, 9},
 	{70, 6},
 	{129, 9},
 	{76, 6},
 	{94, 6},
 	{65, 5},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{67, 5},
 	{118, 6},
 	{73, 5},
 	{91, 5},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{190, 9},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{202, 9},
 	{89, 9},
 	{107, 9},
 	{69, 7},
 	{125, 9},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{113, 9},
 	{71, 7},
 	{131, 9},
 	{30, 10},
 	{46, 10},
 	{7, 9},
 	{194, 7},
 	{83, 7},
 	{54, 10},
 	{11, 9},
 	{119, 7},
 	{19, 9},
 	{35, 9},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{137, 9},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{58, 10},
 	{13, 9},
 	{121, 7},
 	{21, 9},
 	{37, 9},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{25, 9},
 	{41, 9},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{49, 9},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{205, 9},
 	{156, 8},
 	{168, 8},
 	{146, 4},
 	{180, 8},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{159, 8},
 	{115, 9},
 	{72, 8},
 	{133, 9},
 	{78, 8},
 	{96, 8},
 	{65, 5},
 	{195, 8},
 	{84, 8},
 	{102, 8},
 	{67, 5},
 	{120, 8},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{174, 8},
 	{148, 6},
 	{139, 9},
 	{80, 8},
 	{98, 8},
 	{66, 6},
 	{198, 8},
 	{86, 8},
 	{60, 10},
 	{14, 9},
 	{122, 8},
 	{22, 9},
 	{38, 9},
 	{3, 8},
 	{209, 12},
 	{157, 6},
 	{110, 8},
 	{70, 6},
 	{128, 8},
 	{26, 9},
 	{42, 9},
 	{5, 8},
 	{193, 6},
 	{82, 6},
 	{50, 9},
 	{9, 8},
 	{118, 6},
 	{17, 8},
 	{33, 8},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{189, 8},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{201, 8},
 	{88, 8},
 	{106, 8},
 	{69, 7},
 	{124, 8},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{112, 8},
 	{71, 7},
 	{130, 8},
 	{28, 9},
 	{44, 9},
 	{6, 8},
 	{194, 7},
 	{83, 7},
 	{52, 9},
 	{10, 8},
 	{119, 7},
 	{18, 8},
 	{34, 8},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{136, 8},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{56, 9},
 	{12, 8},
 	{121, 7},
 	{20, 8},
 	{36, 8},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{24, 8},
 	{40, 8},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{48, 8},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{146, 4},
 	{209, 12},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{147, 5},
 	{209, 12},
 	{150, 5},
 	{162, 5},
 	{65, 5},
 	{209, 12},
 	{153, 5},
 	{165, 5},
 	{67, 5},
 	{177, 5},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{148, 6},
 	{209, 12},
 	{151, 6},
 	{163, 6},
 	{66, 6},
 	{209, 12},
 	{154, 6},
 	{166, 6},
 	{68, 6},
 	{178, 6},
 	{74, 6},
 	{92, 6},
 	{64, 4},
 	{209, 12},
 	{157, 6},
 	{169, 6},
 	{70, 6},
 	{181, 6},
 	{76, 6},
 	{94, 6},
 	{65, 5},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{67, 5},
 	{118, 6},
 	{73, 5},
 	{91, 5},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{192, 11},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{204, 11},
 	{155, 7},
 	{167, 7},
 	{69, 7},
 	{179, 7},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{170, 7},
 	{71, 7},
 	{182, 7},
 	{77, 7},
 	{95, 7},
 	{65, 5},
 	{194, 7},
 	{83, 7},
 	{101, 7},
 	{67, 5},
 	{119, 7},
 	{73, 5},
 	{91, 5},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{185, 7},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{103, 7},
 	{68, 6},
 	{121, 7},
 	{74, 6},
 	{92, 6},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{76, 6},
 	{94, 6},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{207, 11},
 	{156, 8},
 	{168, 8},
 	{146, 4},
 	{180, 8},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{159, 8},
 	{117, 11},
 	{72, 8},
 	{135, 11},
 	{78, 8},
 	{96, 8},
 	{65, 5},
 	{195, 8},
 	{84, 8},
 	{102, 8},
 	{67, 5},
 	{120, 8},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{174, 8},
 	{148, 6},
 	{141, 11},
 	{80, 8},
 	{98, 8},
 	{66, 6},
 	{198, 8},
 	{86, 8},
 	{104, 8},
 	{68, 6},
 	{122, 8},
 	{74, 6},
 	{92, 6},
 	{3, 8},
 	{209, 12},
 	{157, 6},
 	{110, 8},
 	{70, 6},
 	{128, 8},
 	{76, 6},
 	{94, 6},
 	{5, 8},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{9, 8},
 	{118, 6},
 	{17, 8},
 	{33, 8},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{189, 8},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{201, 8},
 	{88, 8},
 	{106, 8},
 	{69, 7},
 	{124, 8},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{112, 8},
 	{71, 7},
 	{130, 8},
 	{77, 7},
 	{95, 7},
 	{6, 8},
 	{194, 7},
 	{83, 7},
 	{101, 7},
 	{10, 8},
 	{119, 7},
 	{18, 8},
 	{34, 8},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{136, 8},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{103, 7},
 	{12, 8},
 	{121, 7},
 	{20, 8},
 	{36, 8},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{24, 8},
 	{40, 8},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{48, 8},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{146, 4},
 	{209, 12},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{160, 9},
 	{172, 9},
 	{147, 5},
 	{184, 9},
 	{150, 5},
 	{162, 5},
 	{65, 5},
 	{196, 9},
 	{153, 5},
 	{165, 5},
 	{67, 5},
 	{177, 5},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{175, 9},
 	{148, 6},
 	{143, 11},
 	{81, 9},
 	{99, 9},
 	{66, 6},
 	{199, 9},
 	{87, 9},
 	{105, 9},
 	{68, 6},
 	{123, 9},
 	{74, 6},
 	{92, 6},
 	{64, 4},
 	{209, 12},
 	{157, 6},
 	{111, 9},
 	{70, 6},
 	{129, 9},
 	{76, 6},
 	{94, 6},
 	{65, 5},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{67, 5},
 	{118, 6},
 	{73, 5},
 	{91, 5},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{190, 9},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{202, 9},
 	{89, 9},
 	{107, 9},
 	{69, 7},
 	{125, 9},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{113, 9},
 	{71, 7},
 	{131, 9},
 	{31, 11},
 	{47, 11},
 	{7, 9},
 	{194, 7},
 	{83, 7},
 	{55, 11},
 	{11, 9},
 	{119, 7},
 	{19, 9},
 	{35, 9},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{137, 9},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{59, 11},
 	{13, 9},
 	{121, 7},
 	{21, 9},
 	{37, 9},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{25, 9},
 	{41, 9},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{49, 9},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{205, 9},
 	{156, 8},
 	{168, 8},
 	{146, 4},
 	{180, 8},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{159, 8},
 	{115, 9},
 	{72, 8},
 	{133, 9},
 	{78, 8},
 	{96, 8},
 	{65, 5},
 	{195, 8},
 	{84, 8},
 	{102, 8},
 	{67, 5},
 	{120, 8},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{174, 8},
 	{148, 6},
 	{139, 9},
 	{80, 8},
 	{98, 8},
 	{66, 6},
 	{198, 8},
 	{86, 8},
 	{61, 11},
 	{14, 9},
 	{122, 8},
 	{22, 9},
 	{38, 9},
 	{3, 8},
 	{209, 12},
 	{157, 6},
 	{110, 8},
 	{70, 6},
 	{128, 8},
 	{26, 9},
 	{42, 9},
 	{5, 8},
 	{193, 6},
 	{82, 6},
 	{50, 9},
 	{9, 8},
 	{118, 6},
 	{17, 8},
 	{33, 8},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{189, 8},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{201, 8},
 	{88, 8},
 	{106, 8},
 	{69, 7},
 	{124, 8},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{112, 8},
 	{71, 7},
 	{130, 8},
 	{28, 9},
 	{44, 9},
 	{6, 8},
 	{194, 7},
 	{83, 7},
 	{52, 9},
 	{10, 8},
 	{119, 7},
 	{18, 8},
 	{34, 8},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{136, 8},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{56, 9},
 	{12, 8},
 	{121, 7},
 	{20, 8},
 	{36, 8},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{24, 8},
 	{40, 8},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{48, 8},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{146, 4},
 	{209, 12},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{147, 5},
 	{209, 12},
 	{150, 5},
 	{162, 5},
 	{65, 5},
 	{209, 12},
 	{153, 5},
 	{165, 5},
 	{67, 5},
 	{177, 5},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{176, 10},
 	{148, 6},
 	{188, 10},
 	{151, 6},
 	{163, 6},
 	{66, 6},
 	{200, 10},
 	{154, 6},
 	{166, 6},
 	{68, 6},
 	{178, 6},
 	{74, 6},
 	{92, 6},
 	{64, 4},
 	{209, 12},
 	{157, 6},
 	{169, 6},
 	{70, 6},
 	{181, 6},
 	{76, 6},
 	{94, 6},
 	{65, 5},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{67, 5},
 	{118, 6},
 	{73, 5},
 	{91, 5},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{191, 10},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{203, 10},
 	{90, 10},
 	{108, 10},
 	{69, 7},
 	{126, 10},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{114, 10},
 	{71, 7},
 	{132, 10},
 	{77, 7},
 	{95, 7},
 	{65, 5},
 	{194, 7},
 	{83, 7},
 	{101, 7},
 	{67, 5},
 	{119, 7},
 	{73, 5},
 	{91, 5},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{138, 10},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{103, 7},
 	{68, 6},
 	{121, 7},
 	{74, 6},
 	{92, 6},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{76, 6},
 	{94, 6},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{206, 10},
 	{156, 8},
 	{168, 8},
 	{146, 4},
 	{180, 8},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{159, 8},
 	{116, 10},
 	{72, 8},
 	{134, 10},
 	{78, 8},
 	{96, 8},
 	{65, 5},
 	{195, 8},
 	{84, 8},
 	{102, 8},
 	{67, 5},
 	{120, 8},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{174, 8},
 	{148, 6},
 	{140, 10},
 	{80, 8},
 	{98, 8},
 	{66, 6},
 	{198, 8},
 	{86, 8},
 	{62, 11},
 	{15, 10},
 	{122, 8},
 	{23, 10},
 	{39, 10},
 	{3, 8},
 	{209, 12},
 	{157, 6},
 	{110, 8},
 	{70, 6},
 	{128, 8},
 	{27, 10},
 	{43, 10},
 	{5, 8},
 	{193, 6},
 	{82, 6},
 	{51, 10},
 	{9, 8},
 	{118, 6},
 	{17, 8},
 	{33, 8},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{189, 8},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{201, 8},
 	{88, 8},
 	{106, 8},
 	{69, 7},
 	{124, 8},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{112, 8},
 	{71, 7},
 	{130, 8},
 	{29, 10},
 	{45, 10},
 	{6, 8},
 	{194, 7},
 	{83, 7},
 	{53, 10},
 	{10, 8},
 	{119, 7},
 	{18, 8},
 	{34, 8},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{136, 8},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{57, 10},
 	{12, 8},
 	{121, 7},
 	{20, 8},
 	{36, 8},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{24, 8},
 	{40, 8},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{48, 8},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{146, 4},
 	{209, 12},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{160, 9},
 	{172, 9},
 	{147, 5},
 	{184, 9},
 	{150, 5},
 	{162, 5},
 	{65, 5},
 	{196, 9},
 	{153, 5},
 	{165, 5},
 	{67, 5},
 	{177, 5},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{175, 9},
 	{148, 6},
 	{142, 10},
 	{81, 9},
 	{99, 9},
 	{66, 6},
 	{199, 9},
 	{87, 9},
 	{105, 9},
 	{68, 6},
 	{123, 9},
 	{74, 6},
 	{92, 6},
 	{64, 4},
 	{209, 12},
 	{157, 6},
 	{111, 9},
 	{70, 6},
 	{129, 9},
 	{76, 6},
 	{94, 6},
 	{65, 5},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{67, 5},
 	{118, 6},
 	{73, 5},
 	{91, 5},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{190, 9},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{202, 9},
 	{89, 9},
 	{107, 9},
 	{69, 7},
 	{125, 9},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{113, 9},
 	{71, 7},
 	{131, 9},
 	{30, 10},
 	{46, 10},
 	{7, 9},
 	{194, 7},
 	{83, 7},
 	{54, 10},
 	{11, 9},
 	{119, 7},
 	{19, 9},
 	{35, 9},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{137, 9},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{58, 10},
 	{13, 9},
 	{121, 7},
 	{21, 9},
 	{37, 9},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{25, 9},
 	{41, 9},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{49, 9},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{205, 9},
 	{156, 8},
 	{168, 8},
 	{146, 4},
 	{180, 8},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{159, 8},
 	{115, 9},
 	{72, 8},
 	{133, 9},
 	{78, 8},
 	{96, 8},
 	{65, 5},
 	{195, 8},
 	{84, 8},
 	{102, 8},
 	{67, 5},
 	{120, 8},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{174, 8},
 	{148, 6},
 	{139, 9},
 	{80, 8},
 	{98, 8},
 	{66, 6},
 	{198, 8},
 	{86, 8},
 	{60, 10},
 	{14, 9},
 	{122, 8},
 	{22, 9},
 	{38, 9},
 	{3, 8},
 	{209, 12},
 	{157, 6},
 	{110, 8},
 	{70, 6},
 	{128, 8},
 	{26, 9},
 	{42, 9},
 	{5, 8},
 	{193, 6},
 	{82, 6},
 	{50, 9},
 	{9, 8},
 	{118, 6},
 	{17, 8},
 	{33, 8},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{189, 8},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{201, 8},
 	{88, 8},
 	{106, 8},
 	{69, 7},
 	{124, 8},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{112, 8},
 	{71, 7},
 	{130, 8},
 	{28, 9},
 	{44, 9},
 	{6, 8},
 	{194, 7},
 	{83, 7},
 	{52, 9},
 	{10, 8},
 	{119, 7},
 	{18, 8},
 	{34, 8},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{136, 8},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{56, 9},
 	{12, 8},
 	{121, 7},
 	{20, 8},
 	{36, 8},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{24, 8},
 	{40, 8},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{48, 8},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{146, 4},
 	{209, 12},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{147, 5},
 	{209, 12},
 	{150, 5},
 	{162, 5},
 	{65, 5},
 	{209, 12},
 	{153, 5},
 	{165, 5},
 	{67, 5},
 	{177, 5},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{148, 6},
 	{209, 12},
 	{151, 6},
 	{163, 6},
 	{66, 6},
 	{209, 12},
 	{154, 6},
 	{166, 6},
 	{68, 6},
 	{178, 6},
 	{74, 6},
 	{92, 6},
 	{64, 4},
 	{209, 12},
 	{157, 6},
 	{169, 6},
 	{70, 6},
 	{181, 6},
 	{76, 6},
 	{94, 6},
 	{65, 5},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{67, 5},
 	{118, 6},
 	{73, 5},
 	{91, 5},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{209, 12},
 	{155, 7},
 	{167, 7},
 	{69, 7},
 	{179, 7},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{170, 7},
 	{71, 7},
 	{182, 7},
 	{77, 7},
 	{95, 7},
 	{65, 5},
 	{194, 7},
 	{83, 7},
 	{101, 7},
 	{67, 5},
 	{119, 7},
 	{73, 5},
 	{91, 5},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{185, 7},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{103, 7},
 	{68, 6},
 	{121, 7},
 	{74, 6},
 	{92, 6},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{76, 6},
 	{94, 6},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{208, 12},
 	{156, 8},
 	{168, 8},
 	{146, 4},
 	{180, 8},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{159, 8},
 	{171, 8},
 	{72, 8},
 	{183, 8},
 	{78, 8},
 	{96, 8},
 	{65, 5},
 	{195, 8},
 	{84, 8},
 	{102, 8},
 	{67, 5},
 	{120, 8},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{174, 8},
 	{148, 6},
 	{186, 8},
 	{80, 8},
 	{98, 8},
 	{66, 6},
 	{198, 8},
 	{86, 8},
 	{104, 8},
 	{68, 6},
 	{122, 8},
 	{74, 6},
 	{92, 6},
 	{3, 8},
 	{209, 12},
 	{157, 6},
 	{110, 8},
 	{70, 6},
 	{128, 8},
 	{76, 6},
 	{94, 6},
 	{5, 8},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{9, 8},
 	{118, 6},
 	{17, 8},
 	{33, 8},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{189, 8},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{201, 8},
 	{88, 8},
 	{106, 8},
 	{69, 7},
 	{124, 8},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{112, 8},
 	{71, 7},
 	{130, 8},
 	{77, 7},
 	{95, 7},
 	{6, 8},
 	{194, 7},
 	{83, 7},
 	{101, 7},
 	{10, 8},
 	{119, 7},
 	{18, 8},
 	{34, 8},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{136, 8},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{103, 7},
 	{12, 8},
 	{121, 7},
 	{20, 8},
 	{36, 8},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{24, 8},
 	{40, 8},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{48, 8},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{146, 4},
 	{209, 12},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{160, 9},
 	{172, 9},
 	{147, 5},
 	{184, 9},
 	{150, 5},
 	{162, 5},
 	{65, 5},
 	{196, 9},
 	{153, 5},
 	{165, 5},
 	{67, 5},
 	{177, 5},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{175, 9},
 	{148, 6},
 	{144, 12},
 	{81, 9},
 	{99, 9},
 	{66, 6},
 	{199, 9},
 	{87, 9},
 	{105, 9},
 	{68, 6},
 	{123, 9},
 	{74, 6},
 	{92, 6},
 	{64, 4},
 	{209, 12},
 	{157, 6},
 	{111, 9},
 	{70, 6},
 	{129, 9},
 	{76, 6},
 	{94, 6},
 	{65, 5},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{67, 5},
 	{118, 6},
 	{73, 5},
 	{91, 5},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{190, 9},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{202, 9},
 	{89, 9},
 	{107, 9},
 	{69, 7},
 	{125, 9},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{113, 9},
 	{71, 7},
 	{131, 9},
 	{77, 7},
 	{95, 7},
 	{7, 9},
 	{194, 7},
 	{83, 7},
 	{101, 7},
 	{11, 9},
 	{119, 7},
 	{19, 9},
 	{35, 9},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{137, 9},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{103, 7},
 	{13, 9},
 	{121, 7},
 	{21, 9},
 	{37, 9},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{25, 9},
 	{41, 9},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{49, 9},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{205, 9},
 	{156, 8},
 	{168, 8},
 	{146, 4},
 	{180, 8},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{159, 8},
 	{115, 9},
 	{72, 8},
 	{133, 9},
 	{78, 8},
 	{96, 8},
 	{65, 5},
 	{195, 8},
 	{84, 8},
 	{102, 8},
 	{67, 5},
 	{120, 8},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{174, 8},
 	{148, 6},
 	{139, 9},
 	{80, 8},
 	{98, 8},
 	{66, 6},
 	{198, 8},
 	{86, 8},
 	{104, 8},
 	{14, 9},
 	{122, 8},
 	{22, 9},
 	{38, 9},
 	{3, 8},
 	{209, 12},
 	{157, 6},
 	{110, 8},
 	{70, 6},
 	{128, 8},
 	{26, 9},
 	{42, 9},
 	{5, 8},
 	{193, 6},
 	{82, 6},
 	{50, 9},
 	{9, 8},
 	{118, 6},
 	{17, 8},
 	{33, 8},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{189, 8},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{201, 8},
 	{88, 8},
 	{106, 8},
 	{69, 7},
 	{124, 8},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{112, 8},
 	{71, 7},
 	{130, 8},
 	{28, 9},
 	{44, 9},
 	{6, 8},
 	{194, 7},
 	{83, 7},
 	{52, 9},
 	{10, 8},
 	{119, 7},
 	{18, 8},
 	{34, 8},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{136, 8},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{56, 9},
 	{12, 8},
 	{121, 7},
 	{20, 8},
 	{36, 8},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{24, 8},
 	{40, 8},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{48, 8},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{146, 4},
 	{209, 12},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{147, 5},
 	{209, 12},
 	{150, 5},
 	{162, 5},
 	{65, 5},
 	{209, 12},
 	{153, 5},
 	{165, 5},
 	{67, 5},
 	{177, 5},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{176, 10},
 	{148, 6},
 	{188, 10},
 	{151, 6},
 	{163, 6},
 	{66, 6},
 	{200, 10},
 	{154, 6},
 	{166, 6},
 	{68, 6},
 	{178, 6},
 	{74, 6},
 	{92, 6},
 	{64, 4},
 	{209, 12},
 	{157, 6},
 	{169, 6},
 	{70, 6},
 	{181, 6},
 	{76, 6},
 	{94, 6},
 	{65, 5},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{67, 5},
 	{118, 6},
 	{73, 5},
 	{91, 5},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{191, 10},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{203, 10},
 	{90, 10},
 	{108, 10},
 	{69, 7},
 	{126, 10},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{114, 10},
 	{71, 7},
 	{132, 10},
 	{77, 7},
 	{95, 7},
 	{65, 5},
 	{194, 7},
 	{83, 7},
 	{101, 7},
 	{67, 5},
 	{119, 7},
 	{73, 5},
 	{91, 5},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{138, 10},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{103, 7},
 	{68, 6},
 	{121, 7},
 	{74, 6},
 	{92, 6},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{76, 6},
 	{94, 6},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{206, 10},
 	{156, 8},
 	{168, 8},
 	{146, 4},
 	{180, 8},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{159, 8},
 	{116, 10},
 	{72, 8},
 	{134, 10},
 	{78, 8},
 	{96, 8},
 	{65, 5},
 	{195, 8},
 	{84, 8},
 	{102, 8},
 	{67, 5},
 	{120, 8},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{174, 8},
 	{148, 6},
 	{140, 10},
 	{80, 8},
 	{98, 8},
 	{66, 6},
 	{198, 8},
 	{86, 8},
 	{63, 12},
 	{15, 10},
 	{122, 8},
 	{23, 10},
 	{39, 10},
 	{3, 8},
 	{209, 12},
 	{157, 6},
 	{110, 8},
 	{70, 6},
 	{128, 8},
 	{27, 10},
 	{43, 10},
 	{5, 8},
 	{193, 6},
 	{82, 6},
 	{51, 10},
 	{9, 8},
 	{118, 6},
 	{17, 8},
 	{33, 8},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{189, 8},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{201, 8},
 	{88, 8},
 	{106, 8},
 	{69, 7},
 	{124, 8},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{112, 8},
 	{71, 7},
 	{130, 8},
 	{29, 10},
 	{45, 10},
 	{6, 8},
 	{194, 7},
 	{83, 7},
 	{53, 10},
 	{10, 8},
 	{119, 7},
 	{18, 8},
 	{34, 8},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{136, 8},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{57, 10},
 	{12, 8},
 	{121, 7},
 	{20, 8},
 	{36, 8},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{24, 8},
 	{40, 8},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{48, 8},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{146, 4},
 	{209, 12},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{160, 9},
 	{172, 9},
 	{147, 5},
 	{184, 9},
 	{150, 5},
 	{162, 5},
 	{65, 5},
 	{196, 9},
 	{153, 5},
 	{165, 5},
 	{67, 5},
 	{177, 5},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{175, 9},
 	{148, 6},
 	{142, 10},
 	{81, 9},
 	{99, 9},
 	{66, 6},
 	{199, 9},
 	{87, 9},
 	{105, 9},
 	{68, 6},
 	{123, 9},
 	{74, 6},
 	{92, 6},
 	{64, 4},
 	{209, 12},
 	{157, 6},
 	{111, 9},
 	{70, 6},
 	{129, 9},
 	{76, 6},
 	{94, 6},
 	{65, 5},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{67, 5},
 	{118, 6},
 	{73, 5},
 	{91, 5},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{190, 9},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{202, 9},
 	{89, 9},
 	{107, 9},
 	{69, 7},
 	{125, 9},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{113, 9},
 	{71, 7},
 	{131, 9},
 	{30, 10},
 	{46, 10},
 	{7, 9},
 	{194, 7},
 	{83, 7},
 	{54, 10},
 	{11, 9},
 	{119, 7},
 	{19, 9},
 	{35, 9},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{137, 9},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{58, 10},
 	{13, 9},
 	{121, 7},
 	{21, 9},
 	{37, 9},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{25, 9},
 	{41, 9},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{49, 9},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{205, 9},
 	{156, 8},
 	{168, 8},
 	{146, 4},
 	{180, 8},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{159, 8},
 	{115, 9},
 	{72, 8},
 	{133, 9},
 	{78, 8},
 	{96, 8},
 	{65, 5},
 	{195, 8},
 	{84, 8},
 	{102, 8},
 	{67, 5},
 	{120, 8},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{174, 8},
 	{148, 6},
 	{139, 9},
 	{80, 8},
 	{98, 8},
 	{66, 6},
 	{198, 8},
 	{86, 8},
 	{60, 10},
 	{14, 9},
 	{122, 8},
 	{22, 9},
 	{38, 9},
 	{3, 8},
 	{209, 12},
 	{157, 6},
 	{110, 8},
 	{70, 6},
 	{128, 8},
 	{26, 9},
 	{42, 9},
 	{5, 8},
 	{193, 6},
 	{82, 6},
 	{50, 9},
 	{9, 8},
 	{118, 6},
 	{17, 8},
 	{33, 8},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{189, 8},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{201, 8},
 	{88, 8},
 	{106, 8},
 	{69, 7},
 	{124, 8},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{112, 8},
 	{71, 7},
 	{130, 8},
 	{28, 9},
 	{44, 9},
 	{6, 8},
 	{194, 7},
 	{83, 7},
 	{52, 9},
 	{10, 8},
 	{119, 7},
 	{18, 8},
 	{34, 8},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{136, 8},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{56, 9},
 	{12, 8},
 	{121, 7},
 	{20, 8},
 	{36, 8},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{24, 8},
 	{40, 8},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{48, 8},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{146, 4},
 	{209, 12},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{147, 5},
 	{209, 12},
 	{150, 5},
 	{162, 5},
 	{65, 5},
 	{209, 12},
 	{153, 5},
 	{165, 5},
 	{67, 5},
 	{177, 5},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{148, 6},
 	{209, 12},
 	{151, 6},
 	{163, 6},
 	{66, 6},
 	{209, 12},
 	{154, 6},
 	{166, 6},
 	{68, 6},
 	{178, 6},
 	{74, 6},
 	{92, 6},
 	{64, 4},
 	{209, 12},
 	{157, 6},
 	{169, 6},
 	{70, 6},
 	{181, 6},
 	{76, 6},
 	{94, 6},
 	{65, 5},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{67, 5},
 	{118, 6},
 	{73, 5},
 	{91, 5},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{192, 11},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{204, 11},
 	{155, 7},
 	{167, 7},
 	{69, 7},
 	{179, 7},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{170, 7},
 	{71, 7},
 	{182, 7},
 	{77, 7},
 	{95, 7},
 	{65, 5},
 	{194, 7},
 	{83, 7},
 	{101, 7},
 	{67, 5},
 	{119, 7},
 	{73, 5},
 	{91, 5},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{185, 7},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{103, 7},
 	{68, 6},
 	{121, 7},
 	{74, 6},
 	{92, 6},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{76, 6},
 	{94, 6},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{207, 11},
 	{156, 8},
 	{168, 8},
 	{146, 4},
 	{180, 8},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{159, 8},
 	{117, 11},
 	{72, 8},
 	{135, 11},
 	{78, 8},
 	{96, 8},
 	{65, 5},
 	{195, 8},
 	{84, 8},
 	{102, 8},
 	{67, 5},
 	{120, 8},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{174, 8},
 	{148, 6},
 	{141, 11},
 	{80, 8},
 	{98, 8},
 	{66, 6},
 	{198, 8},
 	{86, 8},
 	{104, 8},
 	{68, 6},
 	{122, 8},
 	{74, 6},
 	{92, 6},
 	{3, 8},
 	{209, 12},
 	{157, 6},
 	{110, 8},
 	{70, 6},
 	{128, 8},
 	{76, 6},
 	{94, 6},
 	{5, 8},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{9, 8},
 	{118, 6},
 	{17, 8},
 	{33, 8},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{189, 8},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{201, 8},
 	{88, 8},
 	{106, 8},
 	{69, 7},
 	{124, 8},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{112, 8},
 	{71, 7},
 	{130, 8},
 	{77, 7},
 	{95, 7},
 	{6, 8},
 	{194, 7},
 	{83, 7},
 	{101, 7},
 	{10, 8},
 	{119, 7},
 	{18, 8},
 	{34, 8},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{136, 8},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{103, 7},
 	{12, 8},
 	{121, 7},
 	{20, 8},
 	{36, 8},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{24, 8},
 	{40, 8},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{48, 8},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{146, 4},
 	{209, 12},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{160, 9},
 	{172, 9},
 	{147, 5},
 	{184, 9},
 	{150, 5},
 	{162, 5},
 	{65, 5},
 	{196, 9},
 	{153, 5},
 	{165, 5},
 	{67, 5},
 	{177, 5},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{175, 9},
 	{148, 6},
 	{143, 11},
 	{81, 9},
 	{99, 9},
 	{66, 6},
 	{199, 9},
 	{87, 9},
 	{105, 9},
 	{68, 6},
 	{123, 9},
 	{74, 6},
 	{92, 6},
 	{64, 4},
 	{209, 12},
 	{157, 6},
 	{111, 9},
 	{70, 6},
 	{129, 9},
 	{76, 6},
 	{94, 6},
 	{65, 5},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{67, 5},
 	{118, 6},
 	{73, 5},
 	{91, 5},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{190, 9},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{202, 9},
 	{89, 9},
 	{107, 9},
 	{69, 7},
 	{125, 9},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{113, 9},
 	{71, 7},
 	{131, 9},
 	{31, 11},
 	{47, 11},
 	{7, 9},
 	{194, 7},
 	{83, 7},
 	{55, 11},
 	{11, 9},
 	{119, 7},
 	{19, 9},
 	{35, 9},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{137, 9},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{59, 11},
 	{13, 9},
 	{121, 7},
 	{21, 9},
 	{37, 9},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{25, 9},
 	{41, 9},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{49, 9},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{205, 9},
 	{156, 8},
 	{168, 8},
 	{146, 4},
 	{180, 8},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{159, 8},
 	{115, 9},
 	{72, 8},
 	{133, 9},
 	{78, 8},
 	{96, 8},
 	{65, 5},
 	{195, 8},
 	{84, 8},
 	{102, 8},
 	{67, 5},
 	{120, 8},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{174, 8},
 	{148, 6},
 	{139, 9},
 	{80, 8},
 	{98, 8},
 	{66, 6},
 	{198, 8},
 	{86, 8},
 	{61, 11},
 	{14, 9},
 	{122, 8},
 	{22, 9},
 	{38, 9},
 	{3, 8},
 	{209, 12},
 	{157, 6},
 	{110, 8},
 	{70, 6},
 	{128, 8},
 	{26, 9},
 	{42, 9},
 	{5, 8},
 	{193, 6},
 	{82, 6},
 	{50, 9},
 	{9, 8},
 	{118, 6},
 	{17, 8},
 	{33, 8},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{189, 8},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{201, 8},
 	{88, 8},
 	{106, 8},
 	{69, 7},
 	{124, 8},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{112, 8},
 	{71, 7},
 	{130, 8},
 	{28, 9},
 	{44, 9},
 	{6, 8},
 	{194, 7},
 	{83, 7},
 	{52, 9},
 	{10, 8},
 	{119, 7},
 	{18, 8},
 	{34, 8},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{136, 8},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{56, 9},
 	{12, 8},
 	{121, 7},
 	{20, 8},
 	{36, 8},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{24, 8},
 	{40, 8},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{48, 8},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{146, 4},
 	{209, 12},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{147, 5},
 	{209, 12},
 	{150, 5},
 	{162, 5},
 	{65, 5},
 	{209, 12},
 	{153, 5},
 	{165, 5},
 	{67, 5},
 	{177, 5},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{176, 10},
 	{148, 6},
 	{188, 10},
 	{151, 6},
 	{163, 6},
 	{66, 6},
 	{200, 10},
 	{154, 6},
 	{166, 6},
 	{68, 6},
 	{178, 6},
 	{74, 6},
 	{92, 6},
 	{64, 4},
 	{209, 12},
 	{157, 6},
 	{169, 6},
 	{70, 6},
 	{181, 6},
 	{76, 6},
 	{94, 6},
 	{65, 5},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{67, 5},
 	{118, 6},
 	{73, 5},
 	{91, 5},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{191, 10},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{203, 10},
 	{90, 10},
 	{108, 10},
 	{69, 7},
 	{126, 10},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{114, 10},
 	{71, 7},
 	{132, 10},
 	{77, 7},
 	{95, 7},
 	{65, 5},
 	{194, 7},
 	{83, 7},
 	{101, 7},
 	{67, 5},
 	{119, 7},
 	{73, 5},
 	{91, 5},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{138, 10},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{103, 7},
 	{68, 6},
 	{121, 7},
 	{74, 6},
 	{92, 6},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{76, 6},
 	{94, 6},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{206, 10},
 	{156, 8},
 	{168, 8},
 	{146, 4},
 	{180, 8},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{159, 8},
 	{116, 10},
 	{72, 8},
 	{134, 10},
 	{78, 8},
 	{96, 8},
 	{65, 5},
 	{195, 8},
 	{84, 8},
 	{102, 8},
 	{67, 5},
 	{120, 8},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{174, 8},
 	{148, 6},
 	{140, 10},
 	{80, 8},
 	{98, 8},
 	{66, 6},
 	{198, 8},
 	{86, 8},
 	{62, 11},
 	{15, 10},
 	{122, 8},
 	{23, 10},
 	{39, 10},
 	{3, 8},
 	{209, 12},
 	{157, 6},
 	{110, 8},
 	{70, 6},
 	{128, 8},
 	{27, 10},
 	{43, 10},
 	{5, 8},
 	{193, 6},
 	{82, 6},
 	{51, 10},
 	{9, 8},
 	{118, 6},
 	{17, 8},
 	{33, 8},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{189, 8},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{201, 8},
 	{88, 8},
 	{106, 8},
 	{69, 7},
 	{124, 8},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{112, 8},
 	{71, 7},
 	{130, 8},
 	{29, 10},
 	{45, 10},
 	{6, 8},
 	{194, 7},
 	{83, 7},
 	{53, 10},
 	{10, 8},
 	{119, 7},
 	{18, 8},
 	{34, 8},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{136, 8},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{57, 10},
 	{12, 8},
 	{121, 7},
 	{20, 8},
 	{36, 8},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{24, 8},
 	{40, 8},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{48, 8},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{146, 4},
 	{209, 12},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{160, 9},
 	{172, 9},
 	{147, 5},
 	{184, 9},
 	{150, 5},
 	{162, 5},
 	{65, 5},
 	{196, 9},
 	{153, 5},
 	{165, 5},
 	{67, 5},
 	{177, 5},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{175, 9},
 	{148, 6},
 	{142, 10},
 	{81, 9},
 	{99, 9},
 	{66, 6},
 	{199, 9},
 	{87, 9},
 	{105, 9},
 	{68, 6},
 	{123, 9},
 	{74, 6},
 	{92, 6},
 	{64, 4},
 	{209, 12},
 	{157, 6},
 	{111, 9},
 	{70, 6},
 	{129, 9},
 	{76, 6},
 	{94, 6},
 	{65, 5},
 	{193, 6},
 	{82, 6},
 	{100, 6},
 	{67, 5},
 	{118, 6},
 	{73, 5},
 	{91, 5},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{190, 9},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{202, 9},
 	{89, 9},
 	{107, 9},
 	{69, 7},
 	{125, 9},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{113, 9},
 	{71, 7},
 	{131, 9},
 	{30, 10},
 	{46, 10},
 	{7, 9},
 	{194, 7},
 	{83, 7},
 	{54, 10},
 	{11, 9},
 	{119, 7},
 	{19, 9},
 	{35, 9},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{137, 9},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{58, 10},
 	{13, 9},
 	{121, 7},
 	{21, 9},
 	{37, 9},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{25, 9},
 	{41, 9},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{49, 9},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{145, 3},
 	{205, 9},
 	{156, 8},
 	{168, 8},
 	{146, 4},
 	{180, 8},
 	{149, 4},
 	{161, 4},
 	{64, 4},
 	{209, 12},
 	{159, 8},
 	{115, 9},
 	{72, 8},
 	{133, 9},
 	{78, 8},
 	{96, 8},
 	{65, 5},
 	{195, 8},
 	{84, 8},
 	{102, 8},
 	{67, 5},
 	{120, 8},
 	{73, 5},
 	{91, 5},
 	{64, 4},
 	{209, 12},
 	{209, 12},
 	{174, 8},
 	{148, 6},
 	{139, 9},
 	{80, 8},
 	{98, 8},
 	{66, 6},
 	{198, 8},
 	{86, 8},
 	{60, 10},
 	{14, 9},
 	{122, 8},
 	{22, 9},
 	{38, 9},
 	{3, 8},
 	{209, 12},
 	{157, 6},
 	{110, 8},
 	{70, 6},
 	{128, 8},
 	{26, 9},
 	{42, 9},
 	{5, 8},
 	{193, 6},
 	{82, 6},
 	{50, 9},
 	{9, 8},
 	{118, 6},
 	{17, 8},
 	{33, 8},
 	{0, 6},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{209, 12},
 	{189, 8},
 	{152, 7},
 	{164, 7},
 	{145, 3},
 	{201, 8},
 	{88, 8},
 	{106, 8},
 	{69, 7},
 	{124, 8},
 	{75, 7},
 	{93, 7},
 	{64, 4},
 	{209, 12},
 	{158, 7},
 	{112, 8},
 	{71, 7},
 	{130, 8},
 	{28, 9},
 	{44, 9},
 	{6, 8},
 	{194, 7},
 	{83, 7},
 	{52, 9},
 	{10, 8},
 	{119, 7},
 	{18, 8},
 	{34, 8},
 	{1, 7},
 	{209, 12},
 	{209, 12},
 	{173, 7},
 	{148, 6},
 	{136, 8},
 	{79, 7},
 	{97, 7},
 	{66, 6},
 	{197, 7},
 	{85, 7},
 	{56, 9},
 	{12, 8},
 	{121, 7},
 	{20, 8},
 	{36, 8},
 	{2, 7},
 	{209, 12},
 	{157, 6},
 	{109, 7},
 	{70, 6},
 	{127, 7},
 	{24, 8},
 	{40, 8},
 	{4, 7},
 	{193, 6},
 	{82, 6},
 	{48, 8},
 	{8, 7},
 	{118, 6},
 	{16, 7},
 	{32, 7},
 	{0, 6}};
} // utf8_to_utf16 namespace
} // tables namespace
} // unnamed namespace
} // namespace simdutf

#endif // SIMDUTF_UTF8_TO_UTF16_TABLES_H
/* end file src/tables/utf8_to_utf16_tables.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=tables/utf16_to_utf8_tables.h
/* begin file src/tables/utf16_to_utf8_tables.h */
// file generated by scripts/sse_convert_utf16_to_utf8.py
#ifndef SIMDUTF_UTF16_TO_UTF8_TABLES_H
#define SIMDUTF_UTF16_TO_UTF8_TABLES_H

namespace simdutf {
namespace {
namespace tables {
namespace utf16_to_utf8 {

  // 1 byte for length, 16 bytes for mask
  const uint8_t pack_1_2_utf8_bytes[256][17] = {
    {16,1,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14},
    {15,0,3,2,5,4,7,6,9,8,11,10,13,12,15,14,0x80},
    {15,1,0,3,2,5,4,7,6,8,11,10,13,12,15,14,0x80},
    {14,0,3,2,5,4,7,6,8,11,10,13,12,15,14,0x80,0x80},
    {15,1,0,2,5,4,7,6,9,8,11,10,13,12,15,14,0x80},
    {14,0,2,5,4,7,6,9,8,11,10,13,12,15,14,0x80,0x80},
    {14,1,0,2,5,4,7,6,8,11,10,13,12,15,14,0x80,0x80},
    {13,0,2,5,4,7,6,8,11,10,13,12,15,14,0x80,0x80,0x80},
    {15,1,0,3,2,5,4,7,6,9,8,10,13,12,15,14,0x80},
    {14,0,3,2,5,4,7,6,9,8,10,13,12,15,14,0x80,0x80},
    {14,1,0,3,2,5,4,7,6,8,10,13,12,15,14,0x80,0x80},
    {13,0,3,2,5,4,7,6,8,10,13,12,15,14,0x80,0x80,0x80},
    {14,1,0,2,5,4,7,6,9,8,10,13,12,15,14,0x80,0x80},
    {13,0,2,5,4,7,6,9,8,10,13,12,15,14,0x80,0x80,0x80},
    {13,1,0,2,5,4,7,6,8,10,13,12,15,14,0x80,0x80,0x80},
    {12,0,2,5,4,7,6,8,10,13,12,15,14,0x80,0x80,0x80,0x80},
    {15,1,0,3,2,4,7,6,9,8,11,10,13,12,15,14,0x80},
    {14,0,3,2,4,7,6,9,8,11,10,13,12,15,14,0x80,0x80},
    {14,1,0,3,2,4,7,6,8,11,10,13,12,15,14,0x80,0x80},
    {13,0,3,2,4,7,6,8,11,10,13,12,15,14,0x80,0x80,0x80},
    {14,1,0,2,4,7,6,9,8,11,10,13,12,15,14,0x80,0x80},
    {13,0,2,4,7,6,9,8,11,10,13,12,15,14,0x80,0x80,0x80},
    {13,1,0,2,4,7,6,8,11,10,13,12,15,14,0x80,0x80,0x80},
    {12,0,2,4,7,6,8,11,10,13,12,15,14,0x80,0x80,0x80,0x80},
    {14,1,0,3,2,4,7,6,9,8,10,13,12,15,14,0x80,0x80},
    {13,0,3,2,4,7,6,9,8,10,13,12,15,14,0x80,0x80,0x80},
    {13,1,0,3,2,4,7,6,8,10,13,12,15,14,0x80,0x80,0x80},
    {12,0,3,2,4,7,6,8,10,13,12,15,14,0x80,0x80,0x80,0x80},
    {13,1,0,2,4,7,6,9,8,10,13,12,15,14,0x80,0x80,0x80},
    {12,0,2,4,7,6,9,8,10,13,12,15,14,0x80,0x80,0x80,0x80},
    {12,1,0,2,4,7,6,8,10,13,12,15,14,0x80,0x80,0x80,0x80},
    {11,0,2,4,7,6,8,10,13,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {15,1,0,3,2,5,4,7,6,9,8,11,10,12,15,14,0x80},
    {14,0,3,2,5,4,7,6,9,8,11,10,12,15,14,0x80,0x80},
    {14,1,0,3,2,5,4,7,6,8,11,10,12,15,14,0x80,0x80},
    {13,0,3,2,5,4,7,6,8,11,10,12,15,14,0x80,0x80,0x80},
    {14,1,0,2,5,4,7,6,9,8,11,10,12,15,14,0x80,0x80},
    {13,0,2,5,4,7,6,9,8,11,10,12,15,14,0x80,0x80,0x80},
    {13,1,0,2,5,4,7,6,8,11,10,12,15,14,0x80,0x80,0x80},
    {12,0,2,5,4,7,6,8,11,10,12,15,14,0x80,0x80,0x80,0x80},
    {14,1,0,3,2,5,4,7,6,9,8,10,12,15,14,0x80,0x80},
    {13,0,3,2,5,4,7,6,9,8,10,12,15,14,0x80,0x80,0x80},
    {13,1,0,3,2,5,4,7,6,8,10,12,15,14,0x80,0x80,0x80},
    {12,0,3,2,5,4,7,6,8,10,12,15,14,0x80,0x80,0x80,0x80},
    {13,1,0,2,5,4,7,6,9,8,10,12,15,14,0x80,0x80,0x80},
    {12,0,2,5,4,7,6,9,8,10,12,15,14,0x80,0x80,0x80,0x80},
    {12,1,0,2,5,4,7,6,8,10,12,15,14,0x80,0x80,0x80,0x80},
    {11,0,2,5,4,7,6,8,10,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {14,1,0,3,2,4,7,6,9,8,11,10,12,15,14,0x80,0x80},
    {13,0,3,2,4,7,6,9,8,11,10,12,15,14,0x80,0x80,0x80},
    {13,1,0,3,2,4,7,6,8,11,10,12,15,14,0x80,0x80,0x80},
    {12,0,3,2,4,7,6,8,11,10,12,15,14,0x80,0x80,0x80,0x80},
    {13,1,0,2,4,7,6,9,8,11,10,12,15,14,0x80,0x80,0x80},
    {12,0,2,4,7,6,9,8,11,10,12,15,14,0x80,0x80,0x80,0x80},
    {12,1,0,2,4,7,6,8,11,10,12,15,14,0x80,0x80,0x80,0x80},
    {11,0,2,4,7,6,8,11,10,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {13,1,0,3,2,4,7,6,9,8,10,12,15,14,0x80,0x80,0x80},
    {12,0,3,2,4,7,6,9,8,10,12,15,14,0x80,0x80,0x80,0x80},
    {12,1,0,3,2,4,7,6,8,10,12,15,14,0x80,0x80,0x80,0x80},
    {11,0,3,2,4,7,6,8,10,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {12,1,0,2,4,7,6,9,8,10,12,15,14,0x80,0x80,0x80,0x80},
    {11,0,2,4,7,6,9,8,10,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,2,4,7,6,8,10,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,2,4,7,6,8,10,12,15,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {15,1,0,3,2,5,4,6,9,8,11,10,13,12,15,14,0x80},
    {14,0,3,2,5,4,6,9,8,11,10,13,12,15,14,0x80,0x80},
    {14,1,0,3,2,5,4,6,8,11,10,13,12,15,14,0x80,0x80},
    {13,0,3,2,5,4,6,8,11,10,13,12,15,14,0x80,0x80,0x80},
    {14,1,0,2,5,4,6,9,8,11,10,13,12,15,14,0x80,0x80},
    {13,0,2,5,4,6,9,8,11,10,13,12,15,14,0x80,0x80,0x80},
    {13,1,0,2,5,4,6,8,11,10,13,12,15,14,0x80,0x80,0x80},
    {12,0,2,5,4,6,8,11,10,13,12,15,14,0x80,0x80,0x80,0x80},
    {14,1,0,3,2,5,4,6,9,8,10,13,12,15,14,0x80,0x80},
    {13,0,3,2,5,4,6,9,8,10,13,12,15,14,0x80,0x80,0x80},
    {13,1,0,3,2,5,4,6,8,10,13,12,15,14,0x80,0x80,0x80},
    {12,0,3,2,5,4,6,8,10,13,12,15,14,0x80,0x80,0x80,0x80},
    {13,1,0,2,5,4,6,9,8,10,13,12,15,14,0x80,0x80,0x80},
    {12,0,2,5,4,6,9,8,10,13,12,15,14,0x80,0x80,0x80,0x80},
    {12,1,0,2,5,4,6,8,10,13,12,15,14,0x80,0x80,0x80,0x80},
    {11,0,2,5,4,6,8,10,13,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {14,1,0,3,2,4,6,9,8,11,10,13,12,15,14,0x80,0x80},
    {13,0,3,2,4,6,9,8,11,10,13,12,15,14,0x80,0x80,0x80},
    {13,1,0,3,2,4,6,8,11,10,13,12,15,14,0x80,0x80,0x80},
    {12,0,3,2,4,6,8,11,10,13,12,15,14,0x80,0x80,0x80,0x80},
    {13,1,0,2,4,6,9,8,11,10,13,12,15,14,0x80,0x80,0x80},
    {12,0,2,4,6,9,8,11,10,13,12,15,14,0x80,0x80,0x80,0x80},
    {12,1,0,2,4,6,8,11,10,13,12,15,14,0x80,0x80,0x80,0x80},
    {11,0,2,4,6,8,11,10,13,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {13,1,0,3,2,4,6,9,8,10,13,12,15,14,0x80,0x80,0x80},
    {12,0,3,2,4,6,9,8,10,13,12,15,14,0x80,0x80,0x80,0x80},
    {12,1,0,3,2,4,6,8,10,13,12,15,14,0x80,0x80,0x80,0x80},
    {11,0,3,2,4,6,8,10,13,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {12,1,0,2,4,6,9,8,10,13,12,15,14,0x80,0x80,0x80,0x80},
    {11,0,2,4,6,9,8,10,13,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,2,4,6,8,10,13,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,2,4,6,8,10,13,12,15,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {14,1,0,3,2,5,4,6,9,8,11,10,12,15,14,0x80,0x80},
    {13,0,3,2,5,4,6,9,8,11,10,12,15,14,0x80,0x80,0x80},
    {13,1,0,3,2,5,4,6,8,11,10,12,15,14,0x80,0x80,0x80},
    {12,0,3,2,5,4,6,8,11,10,12,15,14,0x80,0x80,0x80,0x80},
    {13,1,0,2,5,4,6,9,8,11,10,12,15,14,0x80,0x80,0x80},
    {12,0,2,5,4,6,9,8,11,10,12,15,14,0x80,0x80,0x80,0x80},
    {12,1,0,2,5,4,6,8,11,10,12,15,14,0x80,0x80,0x80,0x80},
    {11,0,2,5,4,6,8,11,10,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {13,1,0,3,2,5,4,6,9,8,10,12,15,14,0x80,0x80,0x80},
    {12,0,3,2,5,4,6,9,8,10,12,15,14,0x80,0x80,0x80,0x80},
    {12,1,0,3,2,5,4,6,8,10,12,15,14,0x80,0x80,0x80,0x80},
    {11,0,3,2,5,4,6,8,10,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {12,1,0,2,5,4,6,9,8,10,12,15,14,0x80,0x80,0x80,0x80},
    {11,0,2,5,4,6,9,8,10,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,2,5,4,6,8,10,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,2,5,4,6,8,10,12,15,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {13,1,0,3,2,4,6,9,8,11,10,12,15,14,0x80,0x80,0x80},
    {12,0,3,2,4,6,9,8,11,10,12,15,14,0x80,0x80,0x80,0x80},
    {12,1,0,3,2,4,6,8,11,10,12,15,14,0x80,0x80,0x80,0x80},
    {11,0,3,2,4,6,8,11,10,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {12,1,0,2,4,6,9,8,11,10,12,15,14,0x80,0x80,0x80,0x80},
    {11,0,2,4,6,9,8,11,10,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,2,4,6,8,11,10,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,2,4,6,8,11,10,12,15,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {12,1,0,3,2,4,6,9,8,10,12,15,14,0x80,0x80,0x80,0x80},
    {11,0,3,2,4,6,9,8,10,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,3,2,4,6,8,10,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,3,2,4,6,8,10,12,15,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,2,4,6,9,8,10,12,15,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,2,4,6,9,8,10,12,15,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {10,1,0,2,4,6,8,10,12,15,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,0,2,4,6,8,10,12,15,14,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {15,1,0,3,2,5,4,7,6,9,8,11,10,13,12,14,0x80},
    {14,0,3,2,5,4,7,6,9,8,11,10,13,12,14,0x80,0x80},
    {14,1,0,3,2,5,4,7,6,8,11,10,13,12,14,0x80,0x80},
    {13,0,3,2,5,4,7,6,8,11,10,13,12,14,0x80,0x80,0x80},
    {14,1,0,2,5,4,7,6,9,8,11,10,13,12,14,0x80,0x80},
    {13,0,2,5,4,7,6,9,8,11,10,13,12,14,0x80,0x80,0x80},
    {13,1,0,2,5,4,7,6,8,11,10,13,12,14,0x80,0x80,0x80},
    {12,0,2,5,4,7,6,8,11,10,13,12,14,0x80,0x80,0x80,0x80},
    {14,1,0,3,2,5,4,7,6,9,8,10,13,12,14,0x80,0x80},
    {13,0,3,2,5,4,7,6,9,8,10,13,12,14,0x80,0x80,0x80},
    {13,1,0,3,2,5,4,7,6,8,10,13,12,14,0x80,0x80,0x80},
    {12,0,3,2,5,4,7,6,8,10,13,12,14,0x80,0x80,0x80,0x80},
    {13,1,0,2,5,4,7,6,9,8,10,13,12,14,0x80,0x80,0x80},
    {12,0,2,5,4,7,6,9,8,10,13,12,14,0x80,0x80,0x80,0x80},
    {12,1,0,2,5,4,7,6,8,10,13,12,14,0x80,0x80,0x80,0x80},
    {11,0,2,5,4,7,6,8,10,13,12,14,0x80,0x80,0x80,0x80,0x80},
    {14,1,0,3,2,4,7,6,9,8,11,10,13,12,14,0x80,0x80},
    {13,0,3,2,4,7,6,9,8,11,10,13,12,14,0x80,0x80,0x80},
    {13,1,0,3,2,4,7,6,8,11,10,13,12,14,0x80,0x80,0x80},
    {12,0,3,2,4,7,6,8,11,10,13,12,14,0x80,0x80,0x80,0x80},
    {13,1,0,2,4,7,6,9,8,11,10,13,12,14,0x80,0x80,0x80},
    {12,0,2,4,7,6,9,8,11,10,13,12,14,0x80,0x80,0x80,0x80},
    {12,1,0,2,4,7,6,8,11,10,13,12,14,0x80,0x80,0x80,0x80},
    {11,0,2,4,7,6,8,11,10,13,12,14,0x80,0x80,0x80,0x80,0x80},
    {13,1,0,3,2,4,7,6,9,8,10,13,12,14,0x80,0x80,0x80},
    {12,0,3,2,4,7,6,9,8,10,13,12,14,0x80,0x80,0x80,0x80},
    {12,1,0,3,2,4,7,6,8,10,13,12,14,0x80,0x80,0x80,0x80},
    {11,0,3,2,4,7,6,8,10,13,12,14,0x80,0x80,0x80,0x80,0x80},
    {12,1,0,2,4,7,6,9,8,10,13,12,14,0x80,0x80,0x80,0x80},
    {11,0,2,4,7,6,9,8,10,13,12,14,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,2,4,7,6,8,10,13,12,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,2,4,7,6,8,10,13,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {14,1,0,3,2,5,4,7,6,9,8,11,10,12,14,0x80,0x80},
    {13,0,3,2,5,4,7,6,9,8,11,10,12,14,0x80,0x80,0x80},
    {13,1,0,3,2,5,4,7,6,8,11,10,12,14,0x80,0x80,0x80},
    {12,0,3,2,5,4,7,6,8,11,10,12,14,0x80,0x80,0x80,0x80},
    {13,1,0,2,5,4,7,6,9,8,11,10,12,14,0x80,0x80,0x80},
    {12,0,2,5,4,7,6,9,8,11,10,12,14,0x80,0x80,0x80,0x80},
    {12,1,0,2,5,4,7,6,8,11,10,12,14,0x80,0x80,0x80,0x80},
    {11,0,2,5,4,7,6,8,11,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {13,1,0,3,2,5,4,7,6,9,8,10,12,14,0x80,0x80,0x80},
    {12,0,3,2,5,4,7,6,9,8,10,12,14,0x80,0x80,0x80,0x80},
    {12,1,0,3,2,5,4,7,6,8,10,12,14,0x80,0x80,0x80,0x80},
    {11,0,3,2,5,4,7,6,8,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {12,1,0,2,5,4,7,6,9,8,10,12,14,0x80,0x80,0x80,0x80},
    {11,0,2,5,4,7,6,9,8,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,2,5,4,7,6,8,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,2,5,4,7,6,8,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {13,1,0,3,2,4,7,6,9,8,11,10,12,14,0x80,0x80,0x80},
    {12,0,3,2,4,7,6,9,8,11,10,12,14,0x80,0x80,0x80,0x80},
    {12,1,0,3,2,4,7,6,8,11,10,12,14,0x80,0x80,0x80,0x80},
    {11,0,3,2,4,7,6,8,11,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {12,1,0,2,4,7,6,9,8,11,10,12,14,0x80,0x80,0x80,0x80},
    {11,0,2,4,7,6,9,8,11,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,2,4,7,6,8,11,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,2,4,7,6,8,11,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {12,1,0,3,2,4,7,6,9,8,10,12,14,0x80,0x80,0x80,0x80},
    {11,0,3,2,4,7,6,9,8,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,3,2,4,7,6,8,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,3,2,4,7,6,8,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,2,4,7,6,9,8,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,2,4,7,6,9,8,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {10,1,0,2,4,7,6,8,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,0,2,4,7,6,8,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {14,1,0,3,2,5,4,6,9,8,11,10,13,12,14,0x80,0x80},
    {13,0,3,2,5,4,6,9,8,11,10,13,12,14,0x80,0x80,0x80},
    {13,1,0,3,2,5,4,6,8,11,10,13,12,14,0x80,0x80,0x80},
    {12,0,3,2,5,4,6,8,11,10,13,12,14,0x80,0x80,0x80,0x80},
    {13,1,0,2,5,4,6,9,8,11,10,13,12,14,0x80,0x80,0x80},
    {12,0,2,5,4,6,9,8,11,10,13,12,14,0x80,0x80,0x80,0x80},
    {12,1,0,2,5,4,6,8,11,10,13,12,14,0x80,0x80,0x80,0x80},
    {11,0,2,5,4,6,8,11,10,13,12,14,0x80,0x80,0x80,0x80,0x80},
    {13,1,0,3,2,5,4,6,9,8,10,13,12,14,0x80,0x80,0x80},
    {12,0,3,2,5,4,6,9,8,10,13,12,14,0x80,0x80,0x80,0x80},
    {12,1,0,3,2,5,4,6,8,10,13,12,14,0x80,0x80,0x80,0x80},
    {11,0,3,2,5,4,6,8,10,13,12,14,0x80,0x80,0x80,0x80,0x80},
    {12,1,0,2,5,4,6,9,8,10,13,12,14,0x80,0x80,0x80,0x80},
    {11,0,2,5,4,6,9,8,10,13,12,14,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,2,5,4,6,8,10,13,12,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,2,5,4,6,8,10,13,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {13,1,0,3,2,4,6,9,8,11,10,13,12,14,0x80,0x80,0x80},
    {12,0,3,2,4,6,9,8,11,10,13,12,14,0x80,0x80,0x80,0x80},
    {12,1,0,3,2,4,6,8,11,10,13,12,14,0x80,0x80,0x80,0x80},
    {11,0,3,2,4,6,8,11,10,13,12,14,0x80,0x80,0x80,0x80,0x80},
    {12,1,0,2,4,6,9,8,11,10,13,12,14,0x80,0x80,0x80,0x80},
    {11,0,2,4,6,9,8,11,10,13,12,14,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,2,4,6,8,11,10,13,12,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,2,4,6,8,11,10,13,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {12,1,0,3,2,4,6,9,8,10,13,12,14,0x80,0x80,0x80,0x80},
    {11,0,3,2,4,6,9,8,10,13,12,14,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,3,2,4,6,8,10,13,12,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,3,2,4,6,8,10,13,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,2,4,6,9,8,10,13,12,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,2,4,6,9,8,10,13,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {10,1,0,2,4,6,8,10,13,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,0,2,4,6,8,10,13,12,14,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {13,1,0,3,2,5,4,6,9,8,11,10,12,14,0x80,0x80,0x80},
    {12,0,3,2,5,4,6,9,8,11,10,12,14,0x80,0x80,0x80,0x80},
    {12,1,0,3,2,5,4,6,8,11,10,12,14,0x80,0x80,0x80,0x80},
    {11,0,3,2,5,4,6,8,11,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {12,1,0,2,5,4,6,9,8,11,10,12,14,0x80,0x80,0x80,0x80},
    {11,0,2,5,4,6,9,8,11,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,2,5,4,6,8,11,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,2,5,4,6,8,11,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {12,1,0,3,2,5,4,6,9,8,10,12,14,0x80,0x80,0x80,0x80},
    {11,0,3,2,5,4,6,9,8,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,3,2,5,4,6,8,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,3,2,5,4,6,8,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,2,5,4,6,9,8,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,2,5,4,6,9,8,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {10,1,0,2,5,4,6,8,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,0,2,5,4,6,8,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {12,1,0,3,2,4,6,9,8,11,10,12,14,0x80,0x80,0x80,0x80},
    {11,0,3,2,4,6,9,8,11,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,3,2,4,6,8,11,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,3,2,4,6,8,11,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,2,4,6,9,8,11,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,2,4,6,9,8,11,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {10,1,0,2,4,6,8,11,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,0,2,4,6,8,11,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {11,1,0,3,2,4,6,9,8,10,12,14,0x80,0x80,0x80,0x80,0x80},
    {10,0,3,2,4,6,9,8,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {10,1,0,3,2,4,6,8,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,0,3,2,4,6,8,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {10,1,0,2,4,6,9,8,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,0,2,4,6,9,8,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,1,0,2,4,6,8,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,0,2,4,6,8,10,12,14,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80}
  };

  // 1 byte for length, 16 bytes for mask
  const uint8_t pack_1_2_3_utf8_bytes[256][17] = {
    {12,2,3,1,6,7,5,10,11,9,14,15,13,0x80,0x80,0x80,0x80},
    {9,6,7,5,10,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {11,3,1,6,7,5,10,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80},
    {10,0,6,7,5,10,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,2,3,1,10,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,10,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,3,1,10,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,0,10,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {11,2,3,1,7,5,10,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80},
    {8,7,5,10,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {10,3,1,7,5,10,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,0,7,5,10,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {10,2,3,1,4,10,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,4,10,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,3,1,4,10,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,0,4,10,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,2,3,1,6,7,5,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,6,7,5,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,3,1,6,7,5,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,0,6,7,5,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,2,3,1,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,3,1,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,0,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,2,3,1,7,5,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,7,5,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,3,1,7,5,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,0,7,5,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,2,3,1,4,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,4,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,3,1,4,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,0,4,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {11,2,3,1,6,7,5,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80},
    {8,6,7,5,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {10,3,1,6,7,5,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,0,6,7,5,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,2,3,1,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,3,1,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,0,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {10,2,3,1,7,5,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,7,5,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,3,1,7,5,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,0,7,5,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,2,3,1,4,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,4,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,3,1,4,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,0,4,11,9,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {10,2,3,1,6,7,5,8,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,6,7,5,8,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,3,1,6,7,5,8,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,0,6,7,5,8,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,2,3,1,8,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,8,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,3,1,8,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,0,8,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,2,3,1,7,5,8,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,7,5,8,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,3,1,7,5,8,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,0,7,5,8,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,2,3,1,4,8,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,4,8,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,3,1,4,8,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,0,4,8,14,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,2,3,1,6,7,5,10,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,6,7,5,10,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,3,1,6,7,5,10,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,0,6,7,5,10,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,2,3,1,10,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,10,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,3,1,10,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,0,10,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,2,3,1,7,5,10,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,7,5,10,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,3,1,7,5,10,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,0,7,5,10,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,2,3,1,4,10,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,4,10,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,3,1,4,10,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,0,4,10,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,2,3,1,6,7,5,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,6,7,5,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,3,1,6,7,5,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,0,6,7,5,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,2,3,1,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {0,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {2,3,1,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {1,0,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,2,3,1,7,5,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {2,7,5,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,3,1,7,5,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,0,7,5,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,2,3,1,4,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {1,4,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,3,1,4,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {2,0,4,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,2,3,1,6,7,5,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,6,7,5,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,3,1,6,7,5,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,0,6,7,5,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,2,3,1,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {2,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,3,1,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,0,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,2,3,1,7,5,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,7,5,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,3,1,7,5,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,0,7,5,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,2,3,1,4,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,4,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,3,1,4,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,0,4,11,9,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,2,3,1,6,7,5,8,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,6,7,5,8,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,3,1,6,7,5,8,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,0,6,7,5,8,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,2,3,1,8,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {1,8,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,3,1,8,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {2,0,8,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,2,3,1,7,5,8,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,7,5,8,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,3,1,7,5,8,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,0,7,5,8,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,2,3,1,4,8,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {2,4,8,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,3,1,4,8,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,0,4,8,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {11,2,3,1,6,7,5,10,11,9,15,13,0x80,0x80,0x80,0x80,0x80},
    {8,6,7,5,10,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {10,3,1,6,7,5,10,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,0,6,7,5,10,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,2,3,1,10,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,10,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,3,1,10,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,0,10,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {10,2,3,1,7,5,10,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,7,5,10,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,3,1,7,5,10,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,0,7,5,10,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,2,3,1,4,10,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,4,10,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,3,1,4,10,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,0,4,10,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,2,3,1,6,7,5,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,6,7,5,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,3,1,6,7,5,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,0,6,7,5,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,2,3,1,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {2,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,3,1,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,0,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,2,3,1,7,5,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,7,5,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,3,1,7,5,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,0,7,5,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,2,3,1,4,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,4,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,3,1,4,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,0,4,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {10,2,3,1,6,7,5,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,6,7,5,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,3,1,6,7,5,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,0,6,7,5,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,2,3,1,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,3,1,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,0,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,2,3,1,7,5,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,7,5,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,3,1,7,5,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,0,7,5,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,2,3,1,4,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,4,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,3,1,4,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,0,4,11,9,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,2,3,1,6,7,5,8,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,6,7,5,8,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,3,1,6,7,5,8,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,0,6,7,5,8,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,2,3,1,8,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,8,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,3,1,8,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,0,8,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,2,3,1,7,5,8,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,7,5,8,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,3,1,7,5,8,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,0,7,5,8,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,2,3,1,4,8,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,4,8,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,3,1,4,8,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,0,4,8,15,13,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {10,2,3,1,6,7,5,10,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,6,7,5,10,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,3,1,6,7,5,10,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,0,6,7,5,10,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,2,3,1,10,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,10,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,3,1,10,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,0,10,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,2,3,1,7,5,10,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,7,5,10,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,3,1,7,5,10,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,0,7,5,10,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,2,3,1,4,10,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,4,10,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,3,1,4,10,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,0,4,10,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,2,3,1,6,7,5,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,6,7,5,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,3,1,6,7,5,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,0,6,7,5,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,2,3,1,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {1,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,3,1,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {2,0,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,2,3,1,7,5,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,7,5,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,3,1,7,5,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,0,7,5,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,2,3,1,4,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {2,4,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,3,1,4,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,0,4,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {9,2,3,1,6,7,5,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,6,7,5,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,3,1,6,7,5,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,0,6,7,5,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,2,3,1,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,3,1,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,0,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,2,3,1,7,5,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,7,5,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,3,1,7,5,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,0,7,5,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,2,3,1,4,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,4,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,3,1,4,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,0,4,11,9,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {8,2,3,1,6,7,5,8,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,6,7,5,8,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,3,1,6,7,5,8,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,0,6,7,5,8,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,2,3,1,8,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {2,8,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,3,1,8,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,0,8,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {7,2,3,1,7,5,8,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,7,5,8,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,3,1,7,5,8,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,0,7,5,8,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {6,2,3,1,4,8,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {3,4,8,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {5,3,1,4,8,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80},
    {4,0,4,8,12,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80,0x80}
  };

} // utf16_to_utf8 namespace
} // tables namespace
} // unnamed namespace
} // namespace simdutf

#endif // SIMDUTF_UTF16_TO_UTF8_TABLES_H
/* end file src/tables/utf16_to_utf8_tables.h */
// End of tables.

// The scalar routines should be included once.
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=scalar/ascii.h
/* begin file src/scalar/ascii.h */
#ifndef SIMDUTF_ASCII_H
#define SIMDUTF_ASCII_H

namespace simdutf {
namespace scalar {
namespace {
namespace ascii {
#if SIMDUTF_IMPLEMENTATION_FALLBACK
// Only used by the fallback kernel.
inline simdutf_warn_unused bool validate(const char *buf, size_t len) noexcept {
    const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
    uint64_t pos = 0;
    // process in blocks of 16 bytes when possible
    for (;pos + 16 < len; pos += 16) {
        uint64_t v1;
        std::memcpy(&v1, data + pos, sizeof(uint64_t));
        uint64_t v2;
        std::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
        uint64_t v{v1 | v2};
        if ((v & 0x8080808080808080) != 0) { return false; }
    }
    // process the tail byte-by-byte
    for (;pos < len; pos ++) {
        if (data[pos] >= 0b10000000) { return false; }
    }
    return true;
}
#endif

inline simdutf_warn_unused result validate_with_errors(const char *buf, size_t len) noexcept {
    const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
    size_t pos = 0;
    // process in blocks of 16 bytes when possible
    for (;pos + 16 < len; pos += 16) {
        uint64_t v1;
        std::memcpy(&v1, data + pos, sizeof(uint64_t));
        uint64_t v2;
        std::memcpy(&v2, data + pos + sizeof(uint64_t), sizeof(uint64_t));
        uint64_t v{v1 | v2};
        if ((v & 0x8080808080808080) != 0) {
            for (;pos < len; pos ++) {
                if (data[pos] >= 0b10000000) { return result(error_code::TOO_LARGE, pos); }
            }
        }
    }
    // process the tail byte-by-byte
    for (;pos < len; pos ++) {
        if (data[pos] >= 0b10000000) { return result(error_code::TOO_LARGE, pos); }
    }
    return result(error_code::SUCCESS, pos);
}

} // ascii namespace
} // unnamed namespace
} // namespace scalar
} // namespace simdutf

#endif
/* end file src/scalar/ascii.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=scalar/utf8.h
/* begin file src/scalar/utf8.h */
#ifndef SIMDUTF_UTF8_H
#define SIMDUTF_UTF8_H

namespace simdutf {
namespace scalar {
namespace {
namespace utf8 {
#if SIMDUTF_IMPLEMENTATION_FALLBACK
// only used by the fallback kernel.
// credit: based on code from Google Fuchsia (Apache Licensed)
inline simdutf_warn_unused bool validate(const char *buf, size_t len) 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;
      std::memcpy(&v1, data + pos, sizeof(uint64_t));
      uint64_t v2;
      std::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];

    while (byte < 0b10000000) {
      if (++pos == len) { return true; }
      byte = data[pos];
    }

    if ((byte & 0b11100000) == 0b11000000) {
      next_pos = pos + 2;
      if (next_pos > len) { return false; }
      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return false; }
      // range check
      code_point = (byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
      if ((code_point < 0x80) || (0x7ff < code_point)) { return false; }
    } else if ((byte & 0b11110000) == 0b11100000) {
      next_pos = pos + 3;
      if (next_pos > len) { return false; }
      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return false; }
      if ((data[pos + 2] & 0b11000000) != 0b10000000) { return false; }
      // range check
      code_point = (byte & 0b00001111) << 12 |
                   (data[pos + 1] & 0b00111111) << 6 |
                   (data[pos + 2] & 0b00111111);
      if ((code_point < 0x800) || (0xffff < code_point) ||
          (0xd7ff < code_point && code_point < 0xe000)) {
        return false;
      }
    } else if ((byte & 0b11111000) == 0b11110000) { // 0b11110000
      next_pos = pos + 4;
      if (next_pos > len) { return false; }
      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return false; }
      if ((data[pos + 2] & 0b11000000) != 0b10000000) { return false; }
      if ((data[pos + 3] & 0b11000000) != 0b10000000) { return false; }
      // range check
      code_point =
          (byte & 0b00000111) << 18 | (data[pos + 1] & 0b00111111) << 12 |
          (data[pos + 2] & 0b00111111) << 6 | (data[pos + 3] & 0b00111111);
      if (code_point <= 0xffff || 0x10ffff < code_point) { return false; }
    } else {
      // we may have a continuation
      return false;
    }
    pos = next_pos;
  }
  return true;
}
#endif

inline simdutf_warn_unused result validate_with_errors(const char *buf, size_t len) noexcept {
  const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
  size_t pos = 0;
  uint32_t code_point = 0;
  while (pos < len) {
    // check of the next 8 bytes are ascii.
    size_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;
      std::memcpy(&v1, data + pos, sizeof(uint64_t));
      uint64_t v2;
      std::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];

    while (byte < 0b10000000) {
      if (++pos == len) { return result(error_code::SUCCESS, len); }
      byte = data[pos];
    }

    if ((byte & 0b11100000) == 0b11000000) {
      next_pos = pos + 2;
      if (next_pos > len) { return result(error_code::TOO_SHORT, pos); }
      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }
      // range check
      code_point = (byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
      if ((code_point < 0x80) || (0x7ff < code_point)) { return result(error_code::OVERLONG, pos); }
    } else if ((byte & 0b11110000) == 0b11100000) {
      next_pos = pos + 3;
      if (next_pos > len) { return result(error_code::TOO_SHORT, pos); }
      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }
      if ((data[pos + 2] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }
      // range check
      code_point = (byte & 0b00001111) << 12 |
                   (data[pos + 1] & 0b00111111) << 6 |
                   (data[pos + 2] & 0b00111111);
      if ((code_point < 0x800) || (0xffff < code_point)) { return result(error_code::OVERLONG, pos);}
      if (0xd7ff < code_point && code_point < 0xe000) { return result(error_code::SURROGATE, pos); }
    } else if ((byte & 0b11111000) == 0b11110000) { // 0b11110000
      next_pos = pos + 4;
      if (next_pos > len) { return result(error_code::TOO_SHORT, pos); }
      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }
      if ((data[pos + 2] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }
      if ((data[pos + 3] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }
      // range check
      code_point =
          (byte & 0b00000111) << 18 | (data[pos + 1] & 0b00111111) << 12 |
          (data[pos + 2] & 0b00111111) << 6 | (data[pos + 3] & 0b00111111);
      if (code_point <= 0xffff) { return result(error_code::OVERLONG, pos); }
      if (0x10ffff < code_point) { return result(error_code::TOO_LARGE, pos); }
    } else {
      // we either have too many continuation bytes or an invalid leading byte
      if ((byte & 0b11000000) == 0b10000000) { return result(error_code::TOO_LONG, pos); }
      else { return result(error_code::HEADER_BITS, pos); }
    }
    pos = next_pos;
  }
  return result(error_code::SUCCESS, len);
}

// Finds the previous leading byte and validates with errors from there
// Used to pinpoint the location of an error when an invalid chunk is detected
inline simdutf_warn_unused result rewind_and_validate_with_errors(const char *buf, size_t len) noexcept {
  size_t extra_len{0};
  // A leading byte cannot be further than 4 bytes away
  for(int i = 0; i < 5; i++) {
    unsigned char byte = *buf;
    if ((byte & 0b11000000) != 0b10000000) {
      break;
    } else {
      buf--;
      extra_len++;
    }
  }

  result res = validate_with_errors(buf, len + extra_len);
  res.count -= extra_len;
  return res;
}

inline size_t count_code_points(const char* buf, size_t len) {
    const int8_t * p = reinterpret_cast<const int8_t *>(buf);
    size_t counter{0};
    for(size_t i = 0; i < len; i++) {
        // -65 is 0b10111111, anything larger in two-complement's should start a new code point.
        if(p[i] > -65) { counter++; }
    }
    return counter;
}

inline size_t utf16_length_from_utf8(const char* buf, size_t len) {
    const int8_t * p = reinterpret_cast<const int8_t *>(buf);
    size_t counter{0};
    for(size_t i = 0; i < len; i++) {
        if(p[i] > -65) { counter++; }
        if(uint8_t(p[i]) >= 240) { counter++; }
    }
    return counter;
}

} // utf8 namespace
} // unnamed namespace
} // namespace scalar
} // namespace simdutf

#endif
/* end file src/scalar/utf8.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=scalar/utf16.h
/* begin file src/scalar/utf16.h */
#ifndef SIMDUTF_UTF16_H
#define SIMDUTF_UTF16_H

namespace simdutf {
namespace scalar {
namespace {
namespace utf16 {

inline simdutf_warn_unused uint16_t swap_bytes(const uint16_t word) {
  return uint16_t((word >> 8) | (word << 8));
}

template <endianness big_endian>
inline simdutf_warn_unused bool validate(const char16_t *buf, size_t len) noexcept {
  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
  uint64_t pos = 0;
  while (pos < len) {
    uint16_t word = !match_system(big_endian) ? swap_bytes(data[pos]) : data[pos];
    if((word &0xF800) == 0xD800) {
        if(pos + 1 >= len) { return false; }
        uint16_t diff = uint16_t(word - 0xD800);
        if(diff > 0x3FF) { return false; }
        uint16_t next_word = !match_system(big_endian) ? swap_bytes(data[pos + 1]) : data[pos + 1];
        uint16_t diff2 = uint16_t(next_word - 0xDC00);
        if(diff2 > 0x3FF) { return false; }
        pos += 2;
    } else {
        pos++;
    }
  }
  return true;
}

template <endianness big_endian>
inline simdutf_warn_unused result validate_with_errors(const char16_t *buf, size_t len) noexcept {
  const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
  size_t pos = 0;
  while (pos < len) {
    uint16_t word = !match_system(big_endian) ? swap_bytes(data[pos]) : data[pos];
    if((word & 0xF800) == 0xD800) {
        if(pos + 1 >= len) { return result(error_code::SURROGATE, pos); }
        uint16_t diff = uint16_t(word - 0xD800);
        if(diff > 0x3FF) { return result(error_code::SURROGATE, pos); }
        uint16_t next_word = !match_system(big_endian) ? swap_bytes(data[pos + 1]) : data[pos + 1];
        uint16_t diff2 = uint16_t(next_word - 0xDC00);
        if(diff2 > 0x3FF) { return result(error_code::SURROGATE, pos); }
        pos += 2;
    } else {
        pos++;
    }
  }
  return result(error_code::SUCCESS, pos);
}

template <endianness big_endian>
inline size_t count_code_points(const char16_t* buf, size_t len) {
  // We are not BOM aware.
  const uint16_t * p = reinterpret_cast<const uint16_t *>(buf);
  size_t counter{0};
  for(size_t i = 0; i < len; i++) {
    uint16_t word = !match_system(big_endian) ? swap_bytes(p[i]) : p[i];
    counter += ((word & 0xFC00) != 0xDC00);
  }
  return counter;
}

template <endianness big_endian>
inline size_t utf8_length_from_utf16(const char16_t* buf, size_t len) {
  // We are not BOM aware.
  const uint16_t * p = reinterpret_cast<const uint16_t *>(buf);
  size_t counter{0};
  for(size_t i = 0; i < len; i++) {
    uint16_t word = !match_system(big_endian) ? swap_bytes(p[i]) : p[i];
    /** ASCII **/
    if(word <= 0x7F) { counter++; }
    /** two-byte **/
    else if (word <= 0x7FF) { counter += 2; }
    /** three-byte **/
    else if((word <= 0xD7FF) || (word >= 0xE000)) { counter += 3; }
    /** surrogates -- 4 bytes **/
    else { counter += 2; }
  }
  return counter;
}

template <endianness big_endian>
inline size_t utf32_length_from_utf16(const char16_t* buf, size_t len) {
  // We are not BOM aware.
  const uint16_t * p = reinterpret_cast<const uint16_t *>(buf);
  size_t counter{0};
  for(size_t i = 0; i < len; i++) {
    uint16_t word = !match_system(big_endian) ? swap_bytes(p[i]) : p[i];
    counter += ((word & 0xFC00) != 0xDC00);
  }
  return counter;
}

simdutf_really_inline void change_endianness_utf16(const char16_t* in, size_t size, char16_t* out) {
  const uint16_t * input = reinterpret_cast<const uint16_t *>(in);
  uint16_t * output = reinterpret_cast<uint16_t *>(out);
  for (size_t i = 0; i < size; i++) {
    *output++ = uint16_t(input[i] >> 8 | input[i] << 8);
  }
}

} // utf16 namespace
} // unnamed namespace
} // namespace scalar
} // namespace simdutf

#endif
/* end file src/scalar/utf16.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=scalar/utf32.h
/* begin file src/scalar/utf32.h */
#ifndef SIMDUTF_UTF32_H
#define SIMDUTF_UTF32_H

namespace simdutf {
namespace scalar {
namespace {
namespace utf32 {

inline simdutf_warn_unused bool validate(const char32_t *buf, size_t len) noexcept {
  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
  uint64_t pos = 0;
  for(;pos < len; pos++) {
    uint32_t word = data[pos];
    if(word > 0x10FFFF || (word >= 0xD800 && word <= 0xDFFF)) {
        return false;
    }
  }
  return true;
}

inline simdutf_warn_unused result validate_with_errors(const char32_t *buf, size_t len) noexcept {
  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
  size_t pos = 0;
  for(;pos < len; pos++) {
    uint32_t word = data[pos];
    if(word > 0x10FFFF) {
        return result(error_code::TOO_LARGE, pos);
    }
    if(word >= 0xD800 && word <= 0xDFFF) {
        return result(error_code::SURROGATE, pos);
    }
  }
  return result(error_code::SUCCESS, pos);
}

inline size_t utf8_length_from_utf32(const char32_t* buf, size_t len) {
  // We are not BOM aware.
  const uint32_t * p = reinterpret_cast<const uint32_t *>(buf);
  size_t counter{0};
  for(size_t i = 0; i < len; i++) {
    /** ASCII **/
    if(p[i] <= 0x7F) { counter++; }
    /** two-byte **/
    else if(p[i] <= 0x7FF) { counter += 2; }
    /** three-byte **/
    else if(p[i] <= 0xFFFF) { counter += 3; }
    /** four-bytes **/
    else { counter += 4; }
  }
  return counter;
}

inline size_t utf16_length_from_utf32(const char32_t* buf, size_t len) {
  // We are not BOM aware.
  const uint32_t * p = reinterpret_cast<const uint32_t *>(buf);
  size_t counter{0};
  for(size_t i = 0; i < len; i++) {
    /** non-surrogate word **/
    if(p[i] <= 0xFFFF) { counter++; }
    /** surrogate pair **/
    else { counter += 2; }
  }
  return counter;
}

} // utf32 namespace
} // unnamed namespace
} // namespace scalar
} // namespace simdutf

#endif
/* end file src/scalar/utf32.h */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=scalar/utf32_to_utf8/valid_utf32_to_utf8.h
/* begin file src/scalar/utf32_to_utf8/valid_utf32_to_utf8.h */
#ifndef SIMDUTF_VALID_UTF32_TO_UTF8_H
#define SIMDUTF_VALID_UTF32_TO_UTF8_H

namespace simdutf {
namespace scalar {
namespace {
namespace utf32_to_utf8 {

#if SIMDUTF_IMPLEMENTATION_FALLBACK || SIMDUTF_IMPLEMENTATION_PPC64
// only used by the fallback and POWER kernel
inline size_t convert_valid(const char32_t* buf, size_t len, char* utf8_output) {
	const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
  size_t pos = 0;
  char* start{utf8_output};
  while (pos < len) {
    // try to convert the next block of 2 ASCII characters
    if (pos + 2 <= len) { // if it is safe to read 8 more bytes, check that they are ascii
      uint64_t v;
      ::memcpy(&v, data + pos, sizeof(uint64_t));
      if ((v & 0xFFFFFF80FFFFFF80) == 0) {
        *utf8_output++ = char(buf[pos]);
				*utf8_output++ = char(buf[pos+1]);
        pos += 2;
        continue;
      }
    }
    uint32_t word = data[pos];
    if((word & 0xFFFFFF80)==0) {
      // will generate one UTF-8 bytes
      *utf8_output++ = char(word);
      pos++;
    } else if((word & 0xFFFFF800)==0) {
      // will generate two UTF-8 bytes
      // we have 0b110XXXXX 0b10XXXXXX
      *utf8_output++ = char((word>>6) | 0b11000000);
      *utf8_output++ = char((word & 0b111111) | 0b10000000);
      pos++;
    } else if((word & 0xFFFF0000)==0) {
      // will generate three UTF-8 bytes
      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
      *utf8_output++ = char((word>>12) | 0b11100000);
      *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
      *utf8_output++ = char((word & 0b111111) | 0b10000000);
      pos++;
    } else {
      // will generate four UTF-8 bytes
      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
      *utf8_output++ = char((word>>18) | 0b11110000);
      *utf8_output++ = char(((word>>12) & 0b111111) | 0b10000000);
      *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
      *utf8_output++ = char((word & 0b111111) | 0b10000000);
      pos ++;
    }
  }
  return utf8_output - start;
}
#endif // SIMDUTF_IMPLEMENTATION_FALLBACK || SIMDUTF_IMPLEMENTATION_PPC64

} // utf32_to_utf8 namespace
} // unnamed namespace
} // namespace scalar
} // namespace simdutf

#endif
/* end file src/scalar/utf32_to_utf8/valid_utf32_to_utf8.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=scalar/utf32_to_utf8/utf32_to_utf8.h
/* begin file src/scalar/utf32_to_utf8/utf32_to_utf8.h */
#ifndef SIMDUTF_UTF32_TO_UTF8_H
#define SIMDUTF_UTF32_TO_UTF8_H

namespace simdutf {
namespace scalar {
namespace {
namespace utf32_to_utf8 {

inline size_t convert(const char32_t* buf, size_t len, char* utf8_output) {
  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
  size_t pos = 0;
  char* start{utf8_output};
  while (pos < len) {
    // try to convert the next block of 2 ASCII characters
    if (pos + 2 <= len) { // if it is safe to read 8 more bytes, check that they are ascii
      uint64_t v;
      ::memcpy(&v, data + pos, sizeof(uint64_t));
      if ((v & 0xFFFFFF80FFFFFF80) == 0) {
        *utf8_output++ = char(buf[pos]);
				*utf8_output++ = char(buf[pos+1]);
        pos += 2;
        continue;
      }
    }
    uint32_t word = data[pos];
    if((word & 0xFFFFFF80)==0) {
      // will generate one UTF-8 bytes
      *utf8_output++ = char(word);
      pos++;
    } else if((word & 0xFFFFF800)==0) {
      // will generate two UTF-8 bytes
      // we have 0b110XXXXX 0b10XXXXXX
      *utf8_output++ = char((word>>6) | 0b11000000);
      *utf8_output++ = char((word & 0b111111) | 0b10000000);
      pos++;
    } else if((word & 0xFFFF0000)==0) {
      // will generate three UTF-8 bytes
      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
			if (word >= 0xD800 && word <= 0xDFFF) { return 0; }
      *utf8_output++ = char((word>>12) | 0b11100000);
      *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
      *utf8_output++ = char((word & 0b111111) | 0b10000000);
      pos++;
    } else {
      // will generate four UTF-8 bytes
      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
			if (word > 0x10FFFF) { return 0; }
      *utf8_output++ = char((word>>18) | 0b11110000);
      *utf8_output++ = char(((word>>12) & 0b111111) | 0b10000000);
      *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
      *utf8_output++ = char((word & 0b111111) | 0b10000000);
      pos ++;
    }
  }
  return utf8_output - start;
}

inline result convert_with_errors(const char32_t* buf, size_t len, char* utf8_output) {
  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
  size_t pos = 0;
  char* start{utf8_output};
  while (pos < len) {
    // try to convert the next block of 2 ASCII characters
    if (pos + 2 <= len) { // if it is safe to read 8 more bytes, check that they are ascii
      uint64_t v;
      ::memcpy(&v, data + pos, sizeof(uint64_t));
      if ((v & 0xFFFFFF80FFFFFF80) == 0) {
        *utf8_output++ = char(buf[pos]);
				*utf8_output++ = char(buf[pos+1]);
        pos += 2;
        continue;
      }
    }
    uint32_t word = data[pos];
    if((word & 0xFFFFFF80)==0) {
      // will generate one UTF-8 bytes
      *utf8_output++ = char(word);
      pos++;
    } else if((word & 0xFFFFF800)==0) {
      // will generate two UTF-8 bytes
      // we have 0b110XXXXX 0b10XXXXXX
      *utf8_output++ = char((word>>6) | 0b11000000);
      *utf8_output++ = char((word & 0b111111) | 0b10000000);
      pos++;
    } else if((word & 0xFFFF0000)==0) {
      // will generate three UTF-8 bytes
      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
			if (word >= 0xD800 && word <= 0xDFFF) { return result(error_code::SURROGATE, pos); }
      *utf8_output++ = char((word>>12) | 0b11100000);
      *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
      *utf8_output++ = char((word & 0b111111) | 0b10000000);
      pos++;
    } else {
      // will generate four UTF-8 bytes
      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
			if (word > 0x10FFFF) { return result(error_code::TOO_LARGE, pos); }
      *utf8_output++ = char((word>>18) | 0b11110000);
      *utf8_output++ = char(((word>>12) & 0b111111) | 0b10000000);
      *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
      *utf8_output++ = char((word & 0b111111) | 0b10000000);
      pos ++;
    }
  }
  return result(error_code::SUCCESS, utf8_output - start);
}

} // utf32_to_utf8 namespace
} // unnamed namespace
} // namespace scalar
} // namespace simdutf

#endif
/* end file src/scalar/utf32_to_utf8/utf32_to_utf8.h */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=scalar/utf32_to_utf16/valid_utf32_to_utf16.h
/* begin file src/scalar/utf32_to_utf16/valid_utf32_to_utf16.h */
#ifndef SIMDUTF_VALID_UTF32_TO_UTF16_H
#define SIMDUTF_VALID_UTF32_TO_UTF16_H

namespace simdutf {
namespace scalar {
namespace {
namespace utf32_to_utf16 {

template <endianness big_endian>
inline size_t convert_valid(const char32_t* buf, size_t len, char16_t* utf16_output) {
  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
  size_t pos = 0;
  char16_t* start{utf16_output};
  while (pos < len) {
    uint32_t word = data[pos];
    if((word & 0xFFFF0000)==0) {
      // will not generate a surrogate pair
      *utf16_output++ = !match_system(big_endian) ? char16_t(utf16::swap_bytes(uint16_t(word))) : char16_t(word);
      pos++;
    } else {
      // will generate a surrogate pair
      word -= 0x10000;
      uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
      uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
      if (!match_system(big_endian)) {
        high_surrogate = utf16::swap_bytes(high_surrogate);
        low_surrogate = utf16::swap_bytes(low_surrogate);
      }
      *utf16_output++ = char16_t(high_surrogate);
      *utf16_output++ = char16_t(low_surrogate);
      pos++;
    }
  }
  return utf16_output - start;
}

} // utf32_to_utf16 namespace
} // unnamed namespace
} // namespace scalar
} // namespace simdutf

#endif
/* end file src/scalar/utf32_to_utf16/valid_utf32_to_utf16.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=scalar/utf32_to_utf16/utf32_to_utf16.h
/* begin file src/scalar/utf32_to_utf16/utf32_to_utf16.h */
#ifndef SIMDUTF_UTF32_TO_UTF16_H
#define SIMDUTF_UTF32_TO_UTF16_H

namespace simdutf {
namespace scalar {
namespace {
namespace utf32_to_utf16 {

template <endianness big_endian>
inline size_t convert(const char32_t* buf, size_t len, char16_t* utf16_output) {
  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
  size_t pos = 0;
  char16_t* start{utf16_output};
  while (pos < len) {
    uint32_t word = data[pos];
    if((word & 0xFFFF0000)==0) {
      if (word >= 0xD800 && word <= 0xDFFF) { return 0; }
      // will not generate a surrogate pair
      *utf16_output++ = !match_system(big_endian) ? char16_t(utf16::swap_bytes(uint16_t(word))) : char16_t(word);
    } else {
      // will generate a surrogate pair
      if (word > 0x10FFFF) { return 0; }
      word -= 0x10000;
      uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
      uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
      if (!match_system(big_endian)) {
        high_surrogate = utf16::swap_bytes(high_surrogate);
        low_surrogate = utf16::swap_bytes(low_surrogate);
      }
      *utf16_output++ = char16_t(high_surrogate);
      *utf16_output++ = char16_t(low_surrogate);
    }
    pos++;
  }
  return utf16_output - start;
}

template <endianness big_endian>
inline result convert_with_errors(const char32_t* buf, size_t len, char16_t* utf16_output) {
  const uint32_t *data = reinterpret_cast<const uint32_t *>(buf);
  size_t pos = 0;
  char16_t* start{utf16_output};
  while (pos < len) {
    uint32_t word = data[pos];
    if((word & 0xFFFF0000)==0) {
      if (word >= 0xD800 && word <= 0xDFFF) { return result(error_code::SURROGATE, pos); }
      // will not generate a surrogate pair
      *utf16_output++ = !match_system(big_endian) ? char16_t(utf16::swap_bytes(uint16_t(word))) : char16_t(word);
    } else {
      // will generate a surrogate pair
      if (word > 0x10FFFF) { return result(error_code::TOO_LARGE, pos); }
      word -= 0x10000;
      uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
      uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
      if (!match_system(big_endian)) {
        high_surrogate = utf16::swap_bytes(high_surrogate);
        low_surrogate = utf16::swap_bytes(low_surrogate);
      }
      *utf16_output++ = char16_t(high_surrogate);
      *utf16_output++ = char16_t(low_surrogate);
    }
    pos++;
  }
  return result(error_code::SUCCESS, utf16_output - start);
}

} // utf32_to_utf16 namespace
} // unnamed namespace
} // namespace scalar
} // namespace simdutf

#endif
/* end file src/scalar/utf32_to_utf16/utf32_to_utf16.h */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=scalar/utf16_to_utf8/valid_utf16_to_utf8.h
/* begin file src/scalar/utf16_to_utf8/valid_utf16_to_utf8.h */
#ifndef SIMDUTF_VALID_UTF16_TO_UTF8_H
#define SIMDUTF_VALID_UTF16_TO_UTF8_H

namespace simdutf {
namespace scalar {
namespace {
namespace utf16_to_utf8 {

template <endianness big_endian>
inline size_t convert_valid(const char16_t* buf, size_t len, char* utf8_output) {
 const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
  size_t pos = 0;
  char* start{utf8_output};
  while (pos < len) {
    // try to convert the next block of 4 ASCII characters
    if (pos + 4 <= len) { // if it is safe to read 8 more bytes, check that they are ascii
      uint64_t v;
      ::memcpy(&v, data + pos, sizeof(uint64_t));
      if (!match_system(big_endian)) v = (v >> 8) | (v << (64 - 8));
      if ((v & 0xFF80FF80FF80FF80) == 0) {
        size_t final_pos = pos + 4;
        while(pos < final_pos) {
          *utf8_output++ = !match_system(big_endian) ? char(utf16::swap_bytes(buf[pos])) : char(buf[pos]);
          pos++;
        }
        continue;
      }
    }

    uint16_t word = !match_system(big_endian) ? utf16::swap_bytes(data[pos]) : data[pos];
    if((word & 0xFF80)==0) {
      // will generate one UTF-8 bytes
      *utf8_output++ = char(word);
      pos++;
    } else if((word & 0xF800)==0) {
      // will generate two UTF-8 bytes
      // we have 0b110XXXXX 0b10XXXXXX
      *utf8_output++ = char((word>>6) | 0b11000000);
      *utf8_output++ = char((word & 0b111111) | 0b10000000);
      pos++;
    } else if((word &0xF800 ) != 0xD800) {
      // will generate three UTF-8 bytes
      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
      *utf8_output++ = char((word>>12) | 0b11100000);
      *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
      *utf8_output++ = char((word & 0b111111) | 0b10000000);
      pos++;
    } else {
      // must be a surrogate pair
      uint16_t diff = uint16_t(word - 0xD800);
      if(pos + 1 >= len) { return 0; } // minimal bound checking
      uint16_t next_word = !match_system(big_endian) ? utf16::swap_bytes(data[pos + 1]) : data[pos + 1];
      uint16_t diff2 = uint16_t(next_word - 0xDC00);
      uint32_t value = (diff << 10) + diff2 + 0x10000;
      // will generate four UTF-8 bytes
      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
      *utf8_output++ = char((value>>18) | 0b11110000);
      *utf8_output++ = char(((value>>12) & 0b111111) | 0b10000000);
      *utf8_output++ = char(((value>>6) & 0b111111) | 0b10000000);
      *utf8_output++ = char((value & 0b111111) | 0b10000000);
      pos += 2;
    }
  }
  return utf8_output - start;
}

} // utf16_to_utf8 namespace
} // unnamed namespace
} // namespace scalar
} // namespace simdutf

#endif
/* end file src/scalar/utf16_to_utf8/valid_utf16_to_utf8.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=scalar/utf16_to_utf8/utf16_to_utf8.h
/* begin file src/scalar/utf16_to_utf8/utf16_to_utf8.h */
#ifndef SIMDUTF_UTF16_TO_UTF8_H
#define SIMDUTF_UTF16_TO_UTF8_H

namespace simdutf {
namespace scalar {
namespace {
namespace utf16_to_utf8 {

template <endianness big_endian>
inline size_t convert(const char16_t* buf, size_t len, char* utf8_output) {
 const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
  size_t pos = 0;
  char* start{utf8_output};
  while (pos < len) {
    // try to convert the next block of 8 ASCII characters
    if (pos + 4 <= len) { // if it is safe to read 8 more bytes, check that they are ascii
      uint64_t v;
      ::memcpy(&v, data + pos, sizeof(uint64_t));
      if (!match_system(big_endian)) v = (v >> 8) | (v << (64 - 8));
      if ((v & 0xFF80FF80FF80FF80) == 0) {
        size_t final_pos = pos + 4;
        while(pos < final_pos) {
          *utf8_output++ = !match_system(big_endian) ? char(utf16::swap_bytes(buf[pos])) : char(buf[pos]);
          pos++;
        }
        continue;
      }
    }
    uint16_t word = !match_system(big_endian) ? utf16::swap_bytes(data[pos]) : data[pos];
    if((word & 0xFF80)==0) {
      // will generate one UTF-8 bytes
      *utf8_output++ = char(word);
      pos++;
    } else if((word & 0xF800)==0) {
      // will generate two UTF-8 bytes
      // we have 0b110XXXXX 0b10XXXXXX
      *utf8_output++ = char((word>>6) | 0b11000000);
      *utf8_output++ = char((word & 0b111111) | 0b10000000);
      pos++;
    } else if((word &0xF800 ) != 0xD800) {
      // will generate three UTF-8 bytes
      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
      *utf8_output++ = char((word>>12) | 0b11100000);
      *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
      *utf8_output++ = char((word & 0b111111) | 0b10000000);
      pos++;
    } else {
      // must be a surrogate pair
      if(pos + 1 >= len) { return 0; }
      uint16_t diff = uint16_t(word - 0xD800);
      if(diff > 0x3FF) { return 0; }
      uint16_t next_word = !match_system(big_endian) ? utf16::swap_bytes(data[pos + 1]) : data[pos + 1];
      uint16_t diff2 = uint16_t(next_word - 0xDC00);
      if(diff2 > 0x3FF) { return 0; }
      uint32_t value = (diff << 10) + diff2 + 0x10000;
      // will generate four UTF-8 bytes
      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
      *utf8_output++ = char((value>>18) | 0b11110000);
      *utf8_output++ = char(((value>>12) & 0b111111) | 0b10000000);
      *utf8_output++ = char(((value>>6) & 0b111111) | 0b10000000);
      *utf8_output++ = char((value & 0b111111) | 0b10000000);
      pos += 2;
    }
  }
  return utf8_output - start;
}

template <endianness big_endian>
inline result convert_with_errors(const char16_t* buf, size_t len, char* utf8_output) {
 const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
  size_t pos = 0;
  char* start{utf8_output};
  while (pos < len) {
    // try to convert the next block of 8 ASCII characters
    if (pos + 4 <= len) { // if it is safe to read 8 more bytes, check that they are ascii
      uint64_t v;
      ::memcpy(&v, data + pos, sizeof(uint64_t));
      if (!match_system(big_endian)) v = (v >> 8) | (v << (64 - 8));
      if ((v & 0xFF80FF80FF80FF80) == 0) {
        size_t final_pos = pos + 4;
        while(pos < final_pos) {
          *utf8_output++ = !match_system(big_endian) ? char(utf16::swap_bytes(buf[pos])) : char(buf[pos]);
          pos++;
        }
        continue;
      }
    }
    uint16_t word = !match_system(big_endian) ? utf16::swap_bytes(data[pos]) : data[pos];
    if((word & 0xFF80)==0) {
      // will generate one UTF-8 bytes
      *utf8_output++ = char(word);
      pos++;
    } else if((word & 0xF800)==0) {
      // will generate two UTF-8 bytes
      // we have 0b110XXXXX 0b10XXXXXX
      *utf8_output++ = char((word>>6) | 0b11000000);
      *utf8_output++ = char((word & 0b111111) | 0b10000000);
      pos++;
    } else if((word &0xF800 ) != 0xD800) {
      // will generate three UTF-8 bytes
      // we have 0b1110XXXX 0b10XXXXXX 0b10XXXXXX
      *utf8_output++ = char((word>>12) | 0b11100000);
      *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
      *utf8_output++ = char((word & 0b111111) | 0b10000000);
      pos++;
    } else {
      // must be a surrogate pair
      if(pos + 1 >= len) { return result(error_code::SURROGATE, pos); }
      uint16_t diff = uint16_t(word - 0xD800);
      if(diff > 0x3FF) { return result(error_code::SURROGATE, pos); }
      uint16_t next_word = !match_system(big_endian) ? utf16::swap_bytes(data[pos + 1]) : data[pos + 1];
      uint16_t diff2 = uint16_t(next_word - 0xDC00);
      if(diff2 > 0x3FF) { return result(error_code::SURROGATE, pos); }
      uint32_t value = (diff << 10) + diff2 + 0x10000;
      // will generate four UTF-8 bytes
      // we have 0b11110XXX 0b10XXXXXX 0b10XXXXXX 0b10XXXXXX
      *utf8_output++ = char((value>>18) | 0b11110000);
      *utf8_output++ = char(((value>>12) & 0b111111) | 0b10000000);
      *utf8_output++ = char(((value>>6) & 0b111111) | 0b10000000);
      *utf8_output++ = char((value & 0b111111) | 0b10000000);
      pos += 2;
    }
  }
  return result(error_code::SUCCESS, utf8_output - start);
}

} // utf16_to_utf8 namespace
} // unnamed namespace
} // namespace scalar
} // namespace simdutf

#endif
/* end file src/scalar/utf16_to_utf8/utf16_to_utf8.h */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=scalar/utf16_to_utf32/valid_utf16_to_utf32.h
/* begin file src/scalar/utf16_to_utf32/valid_utf16_to_utf32.h */
#ifndef SIMDUTF_VALID_UTF16_TO_UTF32_H
#define SIMDUTF_VALID_UTF16_TO_UTF32_H

namespace simdutf {
namespace scalar {
namespace {
namespace utf16_to_utf32 {

template <endianness big_endian>
inline size_t convert_valid(const char16_t* buf, size_t len, char32_t* utf32_output) {
 const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
  size_t pos = 0;
  char32_t* start{utf32_output};
  while (pos < len) {
    uint16_t word = !match_system(big_endian) ? utf16::swap_bytes(data[pos]) : data[pos];
    if((word &0xF800 ) != 0xD800) {
      // No surrogate pair, extend 16-bit word to 32-bit word
      *utf32_output++ = char32_t(word);
      pos++;
    } else {
      // must be a surrogate pair
      uint16_t diff = uint16_t(word - 0xD800);
      if(pos + 1 >= len) { return 0; } // minimal bound checking
      uint16_t next_word = !match_system(big_endian) ? utf16::swap_bytes(data[pos + 1]) : data[pos + 1];
      uint16_t diff2 = uint16_t(next_word - 0xDC00);
      uint32_t value = (diff << 10) + diff2 + 0x10000;
      *utf32_output++ = char32_t(value);
      pos += 2;
    }
  }
  return utf32_output - start;
}

} // utf16_to_utf32 namespace
} // unnamed namespace
} // namespace scalar
} // namespace simdutf

#endif
/* end file src/scalar/utf16_to_utf32/valid_utf16_to_utf32.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=scalar/utf16_to_utf32/utf16_to_utf32.h
/* begin file src/scalar/utf16_to_utf32/utf16_to_utf32.h */
#ifndef SIMDUTF_UTF16_TO_UTF32_H
#define SIMDUTF_UTF16_TO_UTF32_H

namespace simdutf {
namespace scalar {
namespace {
namespace utf16_to_utf32 {

template <endianness big_endian>
inline size_t convert(const char16_t* buf, size_t len, char32_t* utf32_output) {
 const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
  size_t pos = 0;
  char32_t* start{utf32_output};
  while (pos < len) {
    uint16_t word = !match_system(big_endian) ? utf16::swap_bytes(data[pos]) : data[pos];
    if((word &0xF800 ) != 0xD800) {
      // No surrogate pair, extend 16-bit word to 32-bit word
      *utf32_output++ = char32_t(word);
      pos++;
    } else {
      // must be a surrogate pair
      uint16_t diff = uint16_t(word - 0xD800);
      if(diff > 0x3FF) { return 0; }
      if(pos + 1 >= len) { return 0; } // minimal bound checking
      uint16_t next_word = !match_system(big_endian) ? utf16::swap_bytes(data[pos + 1]) : data[pos + 1];
      uint16_t diff2 = uint16_t(next_word - 0xDC00);
      if(diff2 > 0x3FF) { return 0; }
      uint32_t value = (diff << 10) + diff2 + 0x10000;
      *utf32_output++ = char32_t(value);
      pos += 2;
    }
  }
  return utf32_output - start;
}

template <endianness big_endian>
inline result convert_with_errors(const char16_t* buf, size_t len, char32_t* utf32_output) {
 const uint16_t *data = reinterpret_cast<const uint16_t *>(buf);
  size_t pos = 0;
  char32_t* start{utf32_output};
  while (pos < len) {
    uint16_t word = !match_system(big_endian) ? utf16::swap_bytes(data[pos]) : data[pos];
    if((word &0xF800 ) != 0xD800) {
      // No surrogate pair, extend 16-bit word to 32-bit word
      *utf32_output++ = char32_t(word);
      pos++;
    } else {
      // must be a surrogate pair
      uint16_t diff = uint16_t(word - 0xD800);
      if(diff > 0x3FF) { return result(error_code::SURROGATE, pos); }
      if(pos + 1 >= len) { return result(error_code::SURROGATE, pos); } // minimal bound checking
      uint16_t next_word = !match_system(big_endian) ? utf16::swap_bytes(data[pos + 1]) : data[pos + 1];
      uint16_t diff2 = uint16_t(next_word - 0xDC00);
      if(diff2 > 0x3FF) { return result(error_code::SURROGATE, pos); }
      uint32_t value = (diff << 10) + diff2 + 0x10000;
      *utf32_output++ = char32_t(value);
      pos += 2;
    }
  }
  return result(error_code::SUCCESS, utf32_output - start);
}

} // utf16_to_utf32 namespace
} // unnamed namespace
} // namespace scalar
} // namespace simdutf

#endif
/* end file src/scalar/utf16_to_utf32/utf16_to_utf32.h */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=scalar/utf8_to_utf16/valid_utf8_to_utf16.h
/* begin file src/scalar/utf8_to_utf16/valid_utf8_to_utf16.h */
#ifndef SIMDUTF_VALID_UTF8_TO_UTF16_H
#define SIMDUTF_VALID_UTF8_TO_UTF16_H

namespace simdutf {
namespace scalar {
namespace {
namespace utf8_to_utf16 {

template <endianness big_endian>
inline size_t convert_valid(const char* buf, size_t len, char16_t* utf16_output) {
 const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
  size_t pos = 0;
  char16_t* start{utf16_output};
  while (pos < len) {
    // try to convert the next block of 8 ASCII bytes
    if (pos + 8 <= len) { // if it is safe to read 8 more bytes, check that they are ascii
      uint64_t v;
      ::memcpy(&v, data + pos, sizeof(uint64_t));
      if ((v & 0x8080808080808080) == 0) {
        size_t final_pos = pos + 8;
        while(pos < final_pos) {
          *utf16_output++ = !match_system(big_endian) ? char16_t(utf16::swap_bytes(buf[pos])) : char16_t(buf[pos]);
          pos++;
        }
        continue;
      }
    }
    uint8_t leading_byte = data[pos]; // leading byte
    if (leading_byte < 0b10000000) {
      // converting one ASCII byte !!!
      *utf16_output++ = !match_system(big_endian) ? char16_t(utf16::swap_bytes(leading_byte)) : char16_t(leading_byte);
      pos++;
    } else if ((leading_byte & 0b11100000) == 0b11000000) {
      // We have a two-byte UTF-8, it should become
      // a single UTF-16 word.
      if(pos + 1 >= len) { break; } // minimal bound checking
      uint16_t code_point = uint16_t(((leading_byte &0b00011111) << 6) | (data[pos + 1] &0b00111111));
      if (!match_system(big_endian)) {
        code_point = utf16::swap_bytes(uint16_t(code_point));
      }
      *utf16_output++ = char16_t(code_point);
      pos += 2;
    } else if ((leading_byte & 0b11110000) == 0b11100000) {
      // We have a three-byte UTF-8, it should become
      // a single UTF-16 word.
      if(pos + 2 >= len) { break; } // minimal bound checking
      uint16_t code_point = uint16_t(((leading_byte &0b00001111) << 12) | ((data[pos + 1] &0b00111111) << 6) | (data[pos + 2] &0b00111111));
      if (!match_system(big_endian)) {
        code_point = utf16::swap_bytes(uint16_t(code_point));
      }
      *utf16_output++ = char16_t(code_point);
      pos += 3;
    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
      // we have a 4-byte UTF-8 word.
      if(pos + 3 >= len) { break; } // minimal bound checking
      uint32_t code_point = ((leading_byte & 0b00000111) << 18 )| ((data[pos + 1] &0b00111111) << 12)
                           | ((data[pos + 2] &0b00111111) << 6) | (data[pos + 3] &0b00111111);
      code_point -= 0x10000;
      uint16_t high_surrogate = uint16_t(0xD800 + (code_point >> 10));
      uint16_t low_surrogate = uint16_t(0xDC00 + (code_point & 0x3FF));
      if (!match_system(big_endian)) {
        high_surrogate = utf16::swap_bytes(high_surrogate);
        low_surrogate = utf16::swap_bytes(low_surrogate);
      }
      *utf16_output++ = char16_t(high_surrogate);
      *utf16_output++ = char16_t(low_surrogate);
      pos += 4;
    } else {
      // we may have a continuation but we do not do error checking
      return 0;
    }
  }
  return utf16_output - start;
}


} // namespace utf8_to_utf16
} // unnamed namespace
} // namespace scalar
} // namespace simdutf

#endif
/* end file src/scalar/utf8_to_utf16/valid_utf8_to_utf16.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=scalar/utf8_to_utf16/utf8_to_utf16.h
/* begin file src/scalar/utf8_to_utf16/utf8_to_utf16.h */
#ifndef SIMDUTF_UTF8_TO_UTF16_H
#define SIMDUTF_UTF8_TO_UTF16_H

namespace simdutf {
namespace scalar {
namespace {
namespace utf8_to_utf16 {

template <endianness big_endian>
inline size_t convert(const char* buf, size_t len, char16_t* utf16_output) {
 const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
  size_t pos = 0;
  char16_t* start{utf16_output};
  while (pos < len) {
    // try to convert the next block of 16 ASCII bytes
    if (pos + 16 <= len) { // if it is safe to read 16 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) {
        size_t final_pos = pos + 16;
        while(pos < final_pos) {
          *utf16_output++ = !match_system(big_endian) ? char16_t(utf16::swap_bytes(buf[pos])) : char16_t(buf[pos]);
          pos++;
        }
        continue;
      }
    }

    uint8_t leading_byte = data[pos]; // leading byte
    if (leading_byte < 0b10000000) {
      // converting one ASCII byte !!!
      *utf16_output++ = !match_system(big_endian) ? char16_t(utf16::swap_bytes(leading_byte)): char16_t(leading_byte);
      pos++;
    } else if ((leading_byte & 0b11100000) == 0b11000000) {
      // We have a two-byte UTF-8, it should become
      // a single UTF-16 word.
      if(pos + 1 >= len) { return 0; } // minimal bound checking
      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return 0; }
      // range check
      uint32_t code_point = (leading_byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
      if (code_point < 0x80 || 0x7ff < code_point) { return 0; }
      if (!match_system(big_endian)) {
        code_point = uint32_t(utf16::swap_bytes(uint16_t(code_point)));
      }
      *utf16_output++ = char16_t(code_point);
      pos += 2;
    } else if ((leading_byte & 0b11110000) == 0b11100000) {
      // We have a three-byte UTF-8, it should become
      // a single UTF-16 word.
      if(pos + 2 >= len) { return 0; } // minimal bound checking

      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return 0; }
      if ((data[pos + 2] & 0b11000000) != 0b10000000) { return 0; }
      // range check
      uint32_t code_point = (leading_byte & 0b00001111) << 12 |
                   (data[pos + 1] & 0b00111111) << 6 |
                   (data[pos + 2] & 0b00111111);
      if (code_point < 0x800 || 0xffff < code_point ||
          (0xd7ff < code_point && code_point < 0xe000)) {
        return 0;
      }
      if (!match_system(big_endian)) {
        code_point = uint32_t(utf16::swap_bytes(uint16_t(code_point)));
      }
      *utf16_output++ = char16_t(code_point);
      pos += 3;
    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
      // we have a 4-byte UTF-8 word.
      if(pos + 3 >= len) { return 0; } // minimal bound checking
      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return 0; }
      if ((data[pos + 2] & 0b11000000) != 0b10000000) { return 0; }
      if ((data[pos + 3] & 0b11000000) != 0b10000000) { return 0; }

      // range check
      uint32_t code_point =
          (leading_byte & 0b00000111) << 18 | (data[pos + 1] & 0b00111111) << 12 |
          (data[pos + 2] & 0b00111111) << 6 | (data[pos + 3] & 0b00111111);
      if (code_point <= 0xffff || 0x10ffff < code_point) { return 0; }
      code_point -= 0x10000;
      uint16_t high_surrogate = uint16_t(0xD800 + (code_point >> 10));
      uint16_t low_surrogate = uint16_t(0xDC00 + (code_point & 0x3FF));
      if (!match_system(big_endian)) {
        high_surrogate = utf16::swap_bytes(high_surrogate);
        low_surrogate = utf16::swap_bytes(low_surrogate);
      }
      *utf16_output++ = char16_t(high_surrogate);
      *utf16_output++ = char16_t(low_surrogate);
      pos += 4;
    } else {
      return 0;
    }
  }
  return utf16_output - start;
}

template <endianness big_endian>
inline result convert_with_errors(const char* buf, size_t len, char16_t* utf16_output) {
 const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
  size_t pos = 0;
  char16_t* start{utf16_output};
  while (pos < len) {
    // try to convert the next block of 16 ASCII bytes
    if (pos + 16 <= len) { // if it is safe to read 16 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) {
        size_t final_pos = pos + 16;
        while(pos < final_pos) {
          *utf16_output++ = !match_system(big_endian) ? char16_t(utf16::swap_bytes(buf[pos])) : char16_t(buf[pos]);
          pos++;
        }
        continue;
      }
    }
    uint8_t leading_byte = data[pos]; // leading byte
    if (leading_byte < 0b10000000) {
      // converting one ASCII byte !!!
      *utf16_output++ = !match_system(big_endian) ? char16_t(utf16::swap_bytes(leading_byte)): char16_t(leading_byte);
      pos++;
    } else if ((leading_byte & 0b11100000) == 0b11000000) {
      // We have a two-byte UTF-8, it should become
      // a single UTF-16 word.
      if(pos + 1 >= len) { return result(error_code::TOO_SHORT, pos); } // minimal bound checking
      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }
      // range check
      uint32_t code_point = (leading_byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
      if (code_point < 0x80 || 0x7ff < code_point) { return result(error_code::OVERLONG, pos); }
      if (!match_system(big_endian)) {
        code_point = uint32_t(utf16::swap_bytes(uint16_t(code_point)));
      }
      *utf16_output++ = char16_t(code_point);
      pos += 2;
    } else if ((leading_byte & 0b11110000) == 0b11100000) {
      // We have a three-byte UTF-8, it should become
      // a single UTF-16 word.
      if(pos + 2 >= len) { return result(error_code::TOO_SHORT, pos); } // minimal bound checking

      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }
      if ((data[pos + 2] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }
      // range check
      uint32_t code_point = (leading_byte & 0b00001111) << 12 |
                   (data[pos + 1] & 0b00111111) << 6 |
                   (data[pos + 2] & 0b00111111);
      if ((code_point < 0x800) || (0xffff < code_point)) { return result(error_code::OVERLONG, pos);}
      if (0xd7ff < code_point && code_point < 0xe000) { return result(error_code::SURROGATE, pos); }
      if (!match_system(big_endian)) {
        code_point = uint32_t(utf16::swap_bytes(uint16_t(code_point)));
      }
      *utf16_output++ = char16_t(code_point);
      pos += 3;
    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
      // we have a 4-byte UTF-8 word.
      if(pos + 3 >= len) { return result(error_code::TOO_SHORT, pos); } // minimal bound checking
      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }
      if ((data[pos + 2] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }
      if ((data[pos + 3] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }

      // range check
      uint32_t code_point =
          (leading_byte & 0b00000111) << 18 | (data[pos + 1] & 0b00111111) << 12 |
          (data[pos + 2] & 0b00111111) << 6 | (data[pos + 3] & 0b00111111);
      if (code_point <= 0xffff) { return result(error_code::OVERLONG, pos); }
      if (0x10ffff < code_point) { return result(error_code::TOO_LARGE, pos); }
      code_point -= 0x10000;
      uint16_t high_surrogate = uint16_t(0xD800 + (code_point >> 10));
      uint16_t low_surrogate = uint16_t(0xDC00 + (code_point & 0x3FF));
      if (!match_system(big_endian)) {
        high_surrogate = utf16::swap_bytes(high_surrogate);
        low_surrogate = utf16::swap_bytes(low_surrogate);
      }
      *utf16_output++ = char16_t(high_surrogate);
      *utf16_output++ = char16_t(low_surrogate);
      pos += 4;
    } else {
      // we either have too many continuation bytes or an invalid leading byte
      if ((leading_byte & 0b11000000) == 0b10000000) { return result(error_code::TOO_LONG, pos); }
      else { return result(error_code::HEADER_BITS, pos); }
    }
  }
  return result(error_code::SUCCESS, utf16_output - start);
}

/**
 * When rewind_and_convert_with_errors is called, we are pointing at 'buf' and we have
 * up to len input bytes left, and we encountered some error. It is possible that
 * the error is at 'buf' exactly, but it could also be in the previous bytes  (up to 3 bytes back).
 *
 * prior_bytes indicates how many bytes, prior to 'buf' may belong to the current memory section
 * and can be safely accessed. We prior_bytes to access safely up to three bytes before 'buf'.
 *
 * The caller is responsible to ensure that len > 0.
 *
 * If the error is believed to have occured prior to 'buf', the count value contain in the result
 * will be SIZE_T - 1, SIZE_T - 2, or SIZE_T - 3.
 */
template <endianness endian>
inline result rewind_and_convert_with_errors(size_t prior_bytes, const char* buf, size_t len, char16_t* utf16_output) {
  size_t extra_len{0};
  // We potentially need to go back in time and find a leading byte.
  // In theory '3' would be sufficient, but sometimes the error can go back quite far.
  size_t how_far_back = prior_bytes;
  // size_t how_far_back = 3; // 3 bytes in the past + current position
  // if(how_far_back >= prior_bytes) { how_far_back = prior_bytes; }
  bool found_leading_bytes{false};
  // important: it is i <= how_far_back and not 'i < how_far_back'.
  for(size_t i = 0; i <= how_far_back; i++) {
    unsigned char byte = buf[0-i];
    found_leading_bytes = ((byte & 0b11000000) != 0b10000000);
    if(found_leading_bytes) {
      buf -= i;
      extra_len = i;
      break;
    }
  }
  //
  // It is possible for this function to return a negative count in its result.
  // C++ Standard Section 18.1 defines size_t is in <cstddef> which is described in C Standard as <stddef.h>.
  // C Standard Section 4.1.5 defines size_t as an unsigned integral type of the result of the sizeof operator
  //
  // An unsigned type will simply wrap round arithmetically (well defined).
  //
  if(!found_leading_bytes) {
    // If how_far_back == 3, we may have four consecutive continuation bytes!!!
    // [....] [continuation] [continuation] [continuation] | [buf is continuation]
    // Or we possibly have a stream that does not start with a leading byte.
    return result(error_code::TOO_LONG, 0-how_far_back);
  }
  result res = convert_with_errors<endian>(buf, len + extra_len, utf16_output);
  if (res.error) {
    res.count -= extra_len;
  }
  return res;
}

} // utf8_to_utf16 namespace
} // unnamed namespace
} // namespace scalar
} // namespace simdutf

#endif
/* end file src/scalar/utf8_to_utf16/utf8_to_utf16.h */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=scalar/utf8_to_utf32/valid_utf8_to_utf32.h
/* begin file src/scalar/utf8_to_utf32/valid_utf8_to_utf32.h */
#ifndef SIMDUTF_VALID_UTF8_TO_UTF32_H
#define SIMDUTF_VALID_UTF8_TO_UTF32_H

namespace simdutf {
namespace scalar {
namespace {
namespace utf8_to_utf32 {

inline size_t convert_valid(const char* buf, size_t len, char32_t* utf32_output) {
 const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
  size_t pos = 0;
  char32_t* start{utf32_output};
  while (pos < len) {
    // try to convert the next block of 8 ASCII bytes
    if (pos + 8 <= len) { // if it is safe to read 8 more bytes, check that they are ascii
      uint64_t v;
      ::memcpy(&v, data + pos, sizeof(uint64_t));
      if ((v & 0x8080808080808080) == 0) {
        size_t final_pos = pos + 8;
        while(pos < final_pos) {
          *utf32_output++ = char32_t(buf[pos]);
          pos++;
        }
        continue;
      }
    }
    uint8_t leading_byte = data[pos]; // leading byte
    if (leading_byte < 0b10000000) {
      // converting one ASCII byte !!!
      *utf32_output++ = char32_t(leading_byte);
      pos++;
    } else if ((leading_byte & 0b11100000) == 0b11000000) {
      // We have a two-byte UTF-8
      if(pos + 1 >= len) { break; } // minimal bound checking
      *utf32_output++ = char32_t(((leading_byte &0b00011111) << 6) | (data[pos + 1] &0b00111111));
      pos += 2;
    } else if ((leading_byte & 0b11110000) == 0b11100000) {
      // We have a three-byte UTF-8
      if(pos + 2 >= len) { break; } // minimal bound checking
      *utf32_output++ = char32_t(((leading_byte &0b00001111) << 12) | ((data[pos + 1] &0b00111111) << 6) | (data[pos + 2] &0b00111111));
      pos += 3;
    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
      // we have a 4-byte UTF-8 word.
      if(pos + 3 >= len) { break; } // minimal bound checking
      uint32_t code_word = ((leading_byte & 0b00000111) << 18 )| ((data[pos + 1] &0b00111111) << 12)
                           | ((data[pos + 2] &0b00111111) << 6) | (data[pos + 3] &0b00111111);
      *utf32_output++ = char32_t(code_word);
      pos += 4;
    } else {
      // we may have a continuation but we do not do error checking
      return 0;
    }
  }
  return utf32_output - start;
}


} // namespace utf8_to_utf32
} // unnamed namespace
} // namespace scalar
} // namespace simdutf

#endif
/* end file src/scalar/utf8_to_utf32/valid_utf8_to_utf32.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=scalar/utf8_to_utf32/utf8_to_utf32.h
/* begin file src/scalar/utf8_to_utf32/utf8_to_utf32.h */
#ifndef SIMDUTF_UTF8_TO_UTF32_H
#define SIMDUTF_UTF8_TO_UTF32_H

namespace simdutf {
namespace scalar {
namespace {
namespace utf8_to_utf32 {

inline size_t convert(const char* buf, size_t len, char32_t* utf32_output) {
 const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
  size_t pos = 0;
  char32_t* start{utf32_output};
  while (pos < len) {
    // try to convert the next block of 16 ASCII bytes
    if (pos + 16 <= len) { // if it is safe to read 16 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) {
        size_t final_pos = pos + 16;
        while(pos < final_pos) {
          *utf32_output++ = char32_t(buf[pos]);
          pos++;
        }
        continue;
      }
    }
    uint8_t leading_byte = data[pos]; // leading byte
    if (leading_byte < 0b10000000) {
      // converting one ASCII byte !!!
      *utf32_output++ = char32_t(leading_byte);
      pos++;
    } else if ((leading_byte & 0b11100000) == 0b11000000) {
      // We have a two-byte UTF-8
      if(pos + 1 >= len) { return 0; } // minimal bound checking
      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return 0; }
      // range check
      uint32_t code_point = (leading_byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
      if (code_point < 0x80 || 0x7ff < code_point) { return 0; }
      *utf32_output++ = char32_t(code_point);
      pos += 2;
    } else if ((leading_byte & 0b11110000) == 0b11100000) {
      // We have a three-byte UTF-8
      if(pos + 2 >= len) { return 0; } // minimal bound checking

      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return 0; }
      if ((data[pos + 2] & 0b11000000) != 0b10000000) { return 0; }
      // range check
      uint32_t code_point = (leading_byte & 0b00001111) << 12 |
                   (data[pos + 1] & 0b00111111) << 6 |
                   (data[pos + 2] & 0b00111111);
      if (code_point < 0x800 || 0xffff < code_point ||
          (0xd7ff < code_point && code_point < 0xe000)) {
        return 0;
      }
      *utf32_output++ = char32_t(code_point);
      pos += 3;
    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
      // we have a 4-byte UTF-8 word.
      if(pos + 3 >= len) { return 0; } // minimal bound checking
      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return 0; }
      if ((data[pos + 2] & 0b11000000) != 0b10000000) { return 0; }
      if ((data[pos + 3] & 0b11000000) != 0b10000000) { return 0; }

      // range check
      uint32_t code_point =
          (leading_byte & 0b00000111) << 18 | (data[pos + 1] & 0b00111111) << 12 |
          (data[pos + 2] & 0b00111111) << 6 | (data[pos + 3] & 0b00111111);
      if (code_point <= 0xffff || 0x10ffff < code_point) { return 0; }
      *utf32_output++ = char32_t(code_point);
      pos += 4;
    } else {
      return 0;
    }
  }
  return utf32_output - start;
}

inline result convert_with_errors(const char* buf, size_t len, char32_t* utf32_output) {
 const uint8_t *data = reinterpret_cast<const uint8_t *>(buf);
  size_t pos = 0;
  char32_t* start{utf32_output};
  while (pos < len) {
    // try to convert the next block of 16 ASCII bytes
    if (pos + 16 <= len) { // if it is safe to read 16 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) {
        size_t final_pos = pos + 16;
        while(pos < final_pos) {
          *utf32_output++ = char32_t(buf[pos]);
          pos++;
        }
        continue;
      }
    }
    uint8_t leading_byte = data[pos]; // leading byte
    if (leading_byte < 0b10000000) {
      // converting one ASCII byte !!!
      *utf32_output++ = char32_t(leading_byte);
      pos++;
    } else if ((leading_byte & 0b11100000) == 0b11000000) {
      // We have a two-byte UTF-8
      if(pos + 1 >= len) { return result(error_code::TOO_SHORT, pos); } // minimal bound checking
      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }
      // range check
      uint32_t code_point = (leading_byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
      if (code_point < 0x80 || 0x7ff < code_point) { return result(error_code::OVERLONG, pos); }
      *utf32_output++ = char32_t(code_point);
      pos += 2;
    } else if ((leading_byte & 0b11110000) == 0b11100000) {
      // We have a three-byte UTF-8
      if(pos + 2 >= len) { return result(error_code::TOO_SHORT, pos); } // minimal bound checking

      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }
      if ((data[pos + 2] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }
      // range check
      uint32_t code_point = (leading_byte & 0b00001111) << 12 |
                   (data[pos + 1] & 0b00111111) << 6 |
                   (data[pos + 2] & 0b00111111);
      if (code_point < 0x800 || 0xffff < code_point) { return result(error_code::OVERLONG, pos); }
      if (0xd7ff < code_point && code_point < 0xe000) { return result(error_code::SURROGATE, pos); }
      *utf32_output++ = char32_t(code_point);
      pos += 3;
    } else if ((leading_byte & 0b11111000) == 0b11110000) { // 0b11110000
      // we have a 4-byte UTF-8 word.
      if(pos + 3 >= len) { return result(error_code::TOO_SHORT, pos); } // minimal bound checking
      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos);}
      if ((data[pos + 2] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }
      if ((data[pos + 3] & 0b11000000) != 0b10000000) { return result(error_code::TOO_SHORT, pos); }

      // range check
      uint32_t code_point =
          (leading_byte & 0b00000111) << 18 | (data[pos + 1] & 0b00111111) << 12 |
          (data[pos + 2] & 0b00111111) << 6 | (data[pos + 3] & 0b00111111);
      if (code_point <= 0xffff) { return result(error_code::OVERLONG, pos); }
      if (0x10ffff < code_point) { return result(error_code::TOO_LARGE, pos); }
      *utf32_output++ = char32_t(code_point);
      pos += 4;
    } else {
      // we either have too many continuation bytes or an invalid leading byte
      if ((leading_byte & 0b11000000) == 0b10000000) { return result(error_code::TOO_LONG, pos); }
      else { return result(error_code::HEADER_BITS, pos); }
    }
  }
  return result(error_code::SUCCESS, utf32_output - start);
}

/**
 * When rewind_and_convert_with_errors is called, we are pointing at 'buf' and we have
 * up to len input bytes left, and we encountered some error. It is possible that
 * the error is at 'buf' exactly, but it could also be in the previous bytes location (up to 3 bytes back).
 *
 * prior_bytes indicates how many bytes, prior to 'buf' may belong to the current memory section
 * and can be safely accessed. We prior_bytes to access safely up to three bytes before 'buf'.
 *
 * The caller is responsible to ensure that len > 0.
 *
 * If the error is believed to have occured prior to 'buf', the count value contain in the result
 * will be SIZE_T - 1, SIZE_T - 2, or SIZE_T - 3.
 */
inline result rewind_and_convert_with_errors(size_t prior_bytes, const char* buf, size_t len, char32_t* utf32_output) {
  size_t extra_len{0};
  // We potentially need to go back in time and find a leading byte.
  size_t how_far_back = 3; // 3 bytes in the past + current position
  if(how_far_back > prior_bytes) { how_far_back = prior_bytes; }
  bool found_leading_bytes{false};
  // important: it is i <= how_far_back and not 'i < how_far_back'.
  for(size_t i = 0; i <= how_far_back; i++) {
    unsigned char byte = buf[0-i];
    found_leading_bytes = ((byte & 0b11000000) != 0b10000000);
    if(found_leading_bytes) {
      buf -= i;
      extra_len = i;
      break;
    }
  }
  //
  // It is possible for this function to return a negative count in its result.
  // C++ Standard Section 18.1 defines size_t is in <cstddef> which is described in C Standard as <stddef.h>.
  // C Standard Section 4.1.5 defines size_t as an unsigned integral type of the result of the sizeof operator
  //
  // An unsigned type will simply wrap round arithmetically (well defined).
  //
  if(!found_leading_bytes) {
    // If how_far_back == 3, we may have four consecutive continuation bytes!!!
    // [....] [continuation] [continuation] [continuation] | [buf is continuation]
    // Or we possibly have a stream that does not start with a leading byte.
    return result(error_code::TOO_LONG, 0-how_far_back);
  }

  result res = convert_with_errors(buf, len + extra_len, utf32_output);
  if (res.error) {
    res.count -= extra_len;
  }
  return res;
}

} // utf8_to_utf32 namespace
} // unnamed namespace
} // namespace scalar
} // namespace simdutf

#endif
/* end file src/scalar/utf8_to_utf32/utf8_to_utf32.h */
//


SIMDUTF_PUSH_DISABLE_WARNINGS
SIMDUTF_DISABLE_UNDESIRED_WARNINGS


#if SIMDUTF_IMPLEMENTATION_ARM64
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=arm64/implementation.cpp
/* begin file src/arm64/implementation.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/arm64/begin.h
/* begin file src/simdutf/arm64/begin.h */
// redefining SIMDUTF_IMPLEMENTATION to "arm64"
// #define SIMDUTF_IMPLEMENTATION arm64
/* end file src/simdutf/arm64/begin.h */
namespace simdutf {
namespace arm64 {
namespace {
#ifndef SIMDUTF_ARM64_H
#error "arm64.h must be included"
#endif
using namespace simd;

simdutf_really_inline bool is_ascii(const simd8x64<uint8_t>& input) {
    simd8<uint8_t> bits = input.reduce_or();
    return bits.max_val() < 0b10000000u;
}

simdutf_unused simdutf_really_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(0b11000000u);
    simd8<bool> is_third_byte  = prev2 >= uint8_t(0b11100000u);
    simd8<bool> is_fourth_byte = prev3 >= uint8_t(0b11110000u);
    // 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;
}

simdutf_really_inline simd8<bool> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
    simd8<bool> is_third_byte  = prev2 >= uint8_t(0b11100000u);
    simd8<bool> is_fourth_byte = prev3 >= uint8_t(0b11110000u);
    return is_third_byte ^ is_fourth_byte;
}

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=arm64/arm_detect_encodings.cpp
/* begin file src/arm64/arm_detect_encodings.cpp */
template<class checker>
// len is known to be a multiple of 2 when this is called
int arm_detect_encodings(const char * buf, size_t len) {
    const char* start = buf;
    const char* end = buf + len;

    bool is_utf8 = true;
    bool is_utf16 = true;
    bool is_utf32 = true;

    int out = 0;

    const auto v_d8 = simd8<uint8_t>::splat(0xd8);
    const auto v_f8 = simd8<uint8_t>::splat(0xf8);

    uint32x4_t currentmax = vmovq_n_u32(0x0);

    checker check{};

    while(buf + 64 <= end) {
        uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t*>(buf));
        uint16x8_t secondin = vld1q_u16(reinterpret_cast<const uint16_t*>(buf) + simd16<uint16_t>::SIZE / sizeof(char16_t));
        uint16x8_t thirdin = vld1q_u16(reinterpret_cast<const uint16_t*>(buf) + 2*simd16<uint16_t>::SIZE / sizeof(char16_t));
        uint16x8_t fourthin = vld1q_u16(reinterpret_cast<const uint16_t*>(buf) + 3*simd16<uint16_t>::SIZE / sizeof(char16_t));

        const auto u0 = simd16<uint16_t>(in);
        const auto u1 = simd16<uint16_t>(secondin);
        const auto u2 = simd16<uint16_t>(thirdin);
        const auto u3 = simd16<uint16_t>(fourthin);

        const auto v0 = u0.shr<8>();
        const auto v1 = u1.shr<8>();
        const auto v2 = u2.shr<8>();
        const auto v3 = u3.shr<8>();

        const auto in16 = simd16<uint16_t>::pack(v0, v1);
        const auto nextin16 = simd16<uint16_t>::pack(v2, v3);

        const uint64_t surrogates_wordmask0 = ((in16 & v_f8) == v_d8).to_bitmask64();
        const uint64_t surrogates_wordmask1 = ((nextin16 & v_f8) == v_d8).to_bitmask64();

        // Check for surrogates
        if (surrogates_wordmask0 != 0 || surrogates_wordmask1 != 0) {
            // Cannot be UTF8
            is_utf8 = false;
            // Can still be either UTF-16LE or UTF-32 depending on the positions of the surrogates
            // To be valid UTF-32, a surrogate cannot be in the two most significant bytes of any 32-bit word.
            // On the other hand, to be valid UTF-16LE, at least one surrogate must be in the two most significant
            // bytes of a 32-bit word since they always come in pairs in UTF-16LE.
            // Note that we always proceed in multiple of 4 before this point so there is no offset in 32-bit words.

            if (((surrogates_wordmask0 | surrogates_wordmask1) & 0xf0f0f0f0f0f0f0f0) != 0) {
                is_utf32 = false;
                // Code from arm_validate_utf16le.cpp
                // Not efficient, we do not process surrogates_wordmask1
                const char16_t * input = reinterpret_cast<const char16_t*>(buf);
                const char16_t* end16 = reinterpret_cast<const char16_t*>(start) + len/2;

                const auto v_fc = simd8<uint8_t>::splat(0xfc);
                const auto v_dc = simd8<uint8_t>::splat(0xdc);

                const uint64_t V0 = ~surrogates_wordmask0;

                const auto vH0 = ((in16 & v_fc) ==  v_dc);
                const uint64_t H0 = vH0.to_bitmask64();

                const uint64_t L0 = ~H0 & surrogates_wordmask0;

                const uint64_t a0 = L0 & (H0 >> 4);

                const uint64_t b0 = a0 << 4;

                const uint64_t c0 = V0 | a0 | b0;
                if (c0 == ~0ull) {
                    input += 16;
                } else if (c0 == 0xfffffffffffffffull) {
                    input += 15;
                } else {
                    is_utf16 = false;
                    break;
                }

                while (input + 16 < end16) {
                    const auto in0 = simd16<uint16_t>(input);
                    const auto in1 = simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));
                    const auto t0 = in0.shr<8>();
                    const auto t1 = in1.shr<8>();
                    const simd8<uint8_t> in_16 = simd16<uint16_t>::pack(t0, t1);

                    const uint64_t surrogates_wordmask = ((in_16 & v_f8) == v_d8).to_bitmask64();
                    if(surrogates_wordmask == 0) {
                        input += 16;
                    } else {
                        const uint64_t V = ~surrogates_wordmask;

                        const auto vH = ((in_16 & v_fc) ==  v_dc);
                        const uint64_t H = vH.to_bitmask64();

                        const uint64_t L = ~H & surrogates_wordmask;

                        const uint64_t a = L & (H >> 4);

                        const uint64_t b = a << 4;

                        const uint64_t c = V | a | b;
                        if (c == ~0ull) {
                            input += 16;
                        } else if (c == 0xfffffffffffffffull) {
                            input += 15;
                        } else {
                            is_utf16 = false;
                            break;
                        }
                    }
                }
            } else {
                is_utf16 = false;
                // Check for UTF-32
                if (len % 4 == 0) {
                    const char32_t * input = reinterpret_cast<const char32_t*>(buf);
                    const char32_t* end32 = reinterpret_cast<const char32_t*>(start) + len/4;

                    // Must start checking for surrogates
                    uint32x4_t currentoffsetmax = vmovq_n_u32(0x0);
                    const uint32x4_t offset = vmovq_n_u32(0xffff2000);
                    const uint32x4_t standardoffsetmax = vmovq_n_u32(0xfffff7ff);

                    const uint32x4_t in32 =  vreinterpretq_u32_u16(in);
                    const uint32x4_t secondin32 =  vreinterpretq_u32_u16(secondin);
                    const uint32x4_t thirdin32 =  vreinterpretq_u32_u16(thirdin);
                    const uint32x4_t fourthin32 =  vreinterpretq_u32_u16(fourthin);

                    currentmax = vmaxq_u32(in32,currentmax);
                    currentmax = vmaxq_u32(secondin32,currentmax);
                    currentmax = vmaxq_u32(thirdin32,currentmax);
                    currentmax = vmaxq_u32(fourthin32,currentmax);

                    currentoffsetmax = vmaxq_u32(vaddq_u32(in32, offset), currentoffsetmax);
                    currentoffsetmax = vmaxq_u32(vaddq_u32(secondin32, offset), currentoffsetmax);
                    currentoffsetmax = vmaxq_u32(vaddq_u32(thirdin32, offset), currentoffsetmax);
                    currentoffsetmax = vmaxq_u32(vaddq_u32(fourthin32, offset), currentoffsetmax);

                    while (input + 4 < end32) {
                        const uint32x4_t in_32 = vld1q_u32(reinterpret_cast<const uint32_t*>(input));
                        currentmax = vmaxq_u32(in_32,currentmax);
                        currentoffsetmax = vmaxq_u32(vaddq_u32(in_32, offset), currentoffsetmax);
                        input += 4;
                    }

                    uint32x4_t forbidden_words = veorq_u32(vmaxq_u32(currentoffsetmax, standardoffsetmax), standardoffsetmax);
                    if(vmaxvq_u32(forbidden_words) != 0) {
                        is_utf32 = false;
                    }
                } else {
                    is_utf32 = false;
                }
            }
            break;
        }
        // If no surrogate, validate under other encodings as well

        // UTF-32 validation
        currentmax = vmaxq_u32(vreinterpretq_u32_u16(in),currentmax);
        currentmax = vmaxq_u32(vreinterpretq_u32_u16(secondin),currentmax);
        currentmax = vmaxq_u32(vreinterpretq_u32_u16(thirdin),currentmax);
        currentmax = vmaxq_u32(vreinterpretq_u32_u16(fourthin),currentmax);

        // UTF-8 validation
        // Relies on ../generic/utf8_validation/utf8_lookup4_algorithm.h
        simd::simd8x64<uint8_t> in8(vreinterpretq_u8_u16(in), vreinterpretq_u8_u16(secondin), vreinterpretq_u8_u16(thirdin), vreinterpretq_u8_u16(fourthin));
        check.check_next_input(in8);

        buf += 64;
    }

    // Check which encodings are possible

    if (is_utf8) {
        if (static_cast<size_t>(buf - start) != len) {
            uint8_t block[64]{};
            std::memset(block, 0x20, 64);
            std::memcpy(block, buf, len - (buf - start));
            simd::simd8x64<uint8_t> in(block);
            check.check_next_input(in);
        }
        if (!check.errors()) {
            out |= simdutf::encoding_type::UTF8;
        }
    }

    if (is_utf16 && scalar::utf16::validate<endianness::LITTLE>(reinterpret_cast<const char16_t*>(buf), (len - (buf - start))/2)) {
        out |= simdutf::encoding_type::UTF16_LE;
    }

    if (is_utf32 && (len % 4 == 0)) {
        const uint32x4_t standardmax = vmovq_n_u32(0x10ffff);
        uint32x4_t is_zero = veorq_u32(vmaxq_u32(currentmax, standardmax), standardmax);
        if (vmaxvq_u32(is_zero) == 0 && scalar::utf32::validate(reinterpret_cast<const char32_t*>(buf), (len - (buf - start))/4)) {
            out |= simdutf::encoding_type::UTF32_LE;
        }
    }

    return out;
}
/* end file src/arm64/arm_detect_encodings.cpp */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=arm64/arm_validate_utf16.cpp
/* begin file src/arm64/arm_validate_utf16.cpp */
template <endianness big_endian>
const char16_t* arm_validate_utf16(const char16_t* input, size_t size) {
    const char16_t* end = input + size;
    const auto v_d8 = simd8<uint8_t>::splat(0xd8);
    const auto v_f8 = simd8<uint8_t>::splat(0xf8);
    const auto v_fc = simd8<uint8_t>::splat(0xfc);
    const auto v_dc = simd8<uint8_t>::splat(0xdc);
    while (input + 16 < end) {
        // 0. Load data: since the validation takes into account only higher
        //    byte of each word, we compress the two vectors into one which
        //    consists only the higher bytes.
        auto in0 = simd16<uint16_t>(input);
        auto in1 = simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));
        if (!match_system(big_endian)) {
            #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
            const uint8x16_t swap = make_uint8x16_t(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
            #else
            const uint8x16_t swap = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
            #endif
            in0 = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(in0), swap));
            in1 = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(in1), swap));
        }
        const auto t0 = in0.shr<8>();
        const auto t1 = in1.shr<8>();
        const simd8<uint8_t> in = simd16<uint16_t>::pack(t0, t1);
        // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
        const uint64_t surrogates_wordmask = ((in & v_f8) == v_d8).to_bitmask64();
        if(surrogates_wordmask == 0) {
            input += 16;
        } else {
            // 2. We have some surrogates that have to be distinguished:
            //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
            //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
            //
            //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)

            // V - non-surrogate words
            //     V = not surrogates_wordmask
            const uint64_t V = ~surrogates_wordmask;

            // H - word-mask for high surrogates: the six highest bits are 0b1101'11
            const auto vH = ((in & v_fc) ==  v_dc);
            const uint64_t H = vH.to_bitmask64();

            // L - word mask for low surrogates
            //     L = not H and surrogates_wordmask
            const uint64_t L = ~H & surrogates_wordmask;

            const uint64_t a = L & (H >> 4); // A low surrogate must be followed by high one.
                              // (A low surrogate placed in the 7th register's word
                              // is an exception we handle.)
            const uint64_t b = a << 4; // Just mark that the opposite fact is hold,
                          // thanks to that we have only two masks for valid case.
            const uint64_t c = V | a | b;      // Combine all the masks into the final one.
            if (c == ~0ull) {
                // The whole input register contains valid UTF-16, i.e.,
                // either single words or proper surrogate pairs.
                input += 16;
            } else if (c == 0xfffffffffffffffull) {
                // The 15 lower words of the input register contains valid UTF-16.
                // The 15th word may be either a low or high surrogate. It the next
                // iteration we 1) check if the low surrogate is followed by a high
                // one, 2) reject sole high surrogate.
                input += 15;
            } else {
                return nullptr;
            }
        }
    }
    return input;
}


template <endianness big_endian>
const result arm_validate_utf16_with_errors(const char16_t* input, size_t size) {
    const char16_t* start = input;
    const char16_t* end = input + size;

    const auto v_d8 = simd8<uint8_t>::splat(0xd8);
    const auto v_f8 = simd8<uint8_t>::splat(0xf8);
    const auto v_fc = simd8<uint8_t>::splat(0xfc);
    const auto v_dc = simd8<uint8_t>::splat(0xdc);
    while (input + 16 < end) {
        // 0. Load data: since the validation takes into account only higher
        //    byte of each word, we compress the two vectors into one which
        //    consists only the higher bytes.
        auto in0 = simd16<uint16_t>(input);
        auto in1 = simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));

        if (!match_system(big_endian)) {
            #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
            const uint8x16_t swap = make_uint8x16_t(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
            #else
            const uint8x16_t swap = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
            #endif
            in0 = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(in0), swap));
            in1 = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(in1), swap));
        }
        const auto t0 = in0.shr<8>();
        const auto t1 = in1.shr<8>();
        const simd8<uint8_t> in = simd16<uint16_t>::pack(t0, t1);
        // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
        const uint64_t surrogates_wordmask = ((in & v_f8) == v_d8).to_bitmask64();
        if(surrogates_wordmask == 0) {
            input += 16;
        } else {
            // 2. We have some surrogates that have to be distinguished:
            //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
            //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
            //
            //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)

            // V - non-surrogate words
            //     V = not surrogates_wordmask
            const uint64_t V = ~surrogates_wordmask;

            // H - word-mask for high surrogates: the six highest bits are 0b1101'11
            const auto vH = ((in & v_fc) ==  v_dc);
            const uint64_t H = vH.to_bitmask64();

            // L - word mask for low surrogates
            //     L = not H and surrogates_wordmask
            const uint64_t L = ~H & surrogates_wordmask;

            const uint64_t a = L & (H >> 4); // A low surrogate must be followed by high one.
                              // (A low surrogate placed in the 7th register's word
                              // is an exception we handle.)
            const uint64_t b = a << 4; // Just mark that the opposite fact is hold,
                          // thanks to that we have only two masks for valid case.
            const uint64_t c = V | a | b;      // Combine all the masks into the final one.
            if (c == ~0ull) {
                // The whole input register contains valid UTF-16, i.e.,
                // either single words or proper surrogate pairs.
                input += 16;
            } else if (c == 0xfffffffffffffffull) {
                // The 15 lower words of the input register contains valid UTF-16.
                // The 15th word may be either a low or high surrogate. It the next
                // iteration we 1) check if the low surrogate is followed by a high
                // one, 2) reject sole high surrogate.
                input += 15;
            } else {
                return result(error_code::SURROGATE, input - start);
            }
        }
    }
    return result(error_code::SUCCESS, input - start);
}
/* end file src/arm64/arm_validate_utf16.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=arm64/arm_validate_utf32le.cpp
/* begin file src/arm64/arm_validate_utf32le.cpp */

const char32_t* arm_validate_utf32le(const char32_t* input, size_t size) {
    const char32_t* end = input + size;

    const uint32x4_t standardmax = vmovq_n_u32(0x10ffff);
    const uint32x4_t offset = vmovq_n_u32(0xffff2000);
    const uint32x4_t standardoffsetmax = vmovq_n_u32(0xfffff7ff);
    uint32x4_t currentmax = vmovq_n_u32(0x0);
    uint32x4_t currentoffsetmax = vmovq_n_u32(0x0);

    while (input + 4 < end) {
        const uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t*>(input));
        currentmax = vmaxq_u32(in,currentmax);
        currentoffsetmax = vmaxq_u32(vaddq_u32(in, offset), currentoffsetmax);
        input += 4;
    }

    uint32x4_t is_zero = veorq_u32(vmaxq_u32(currentmax, standardmax), standardmax);
    if(vmaxvq_u32(is_zero) != 0) {
        return nullptr;
    }

    is_zero = veorq_u32(vmaxq_u32(currentoffsetmax, standardoffsetmax), standardoffsetmax);
    if(vmaxvq_u32(is_zero) != 0) {
        return nullptr;
    }

    return input;
}


const result arm_validate_utf32le_with_errors(const char32_t* input, size_t size) {
    const char32_t* start = input;
    const char32_t* end = input + size;

    const uint32x4_t standardmax = vmovq_n_u32(0x10ffff);
    const uint32x4_t offset = vmovq_n_u32(0xffff2000);
    const uint32x4_t standardoffsetmax = vmovq_n_u32(0xfffff7ff);
    uint32x4_t currentmax = vmovq_n_u32(0x0);
    uint32x4_t currentoffsetmax = vmovq_n_u32(0x0);

    while (input + 4 < end) {
        const uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t*>(input));
        currentmax = vmaxq_u32(in,currentmax);
        currentoffsetmax = vmaxq_u32(vaddq_u32(in, offset), currentoffsetmax);

        uint32x4_t is_zero = veorq_u32(vmaxq_u32(currentmax, standardmax), standardmax);
        if(vmaxvq_u32(is_zero) != 0) {
            return result(error_code::TOO_LARGE, input - start);
        }

        is_zero = veorq_u32(vmaxq_u32(currentoffsetmax, standardoffsetmax), standardoffsetmax);
        if(vmaxvq_u32(is_zero) != 0) {
            return result(error_code::SURROGATE, input - start);
        }

        input += 4;
    }

    return result(error_code::SUCCESS, input - start);
}
/* end file src/arm64/arm_validate_utf32le.cpp */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=arm64/arm_convert_utf8_to_utf16.cpp
/* begin file src/arm64/arm_convert_utf8_to_utf16.cpp */
// Convert up to 12 bytes from utf8 to utf16 using a mask indicating the
// end of the code points. Only the least significant 12 bits of the mask
// are accessed.
// It returns how many bytes were consumed (up to 12).
template <endianness big_endian>
size_t convert_masked_utf8_to_utf16(const char *input,
                           uint64_t utf8_end_of_code_point_mask,
                           char16_t *&utf16_output) {
  // we use an approach where we try to process up to 12 input bytes.
  // Why 12 input bytes and not 16? Because we are concerned with the size of
  // the lookup tables. Also 12 is nicely divisible by two and three.
  //
  #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
  const uint8x16_t swap = make_uint8x16_t(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
  #else
  const uint8x16_t swap = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
  #endif
  uint8x16_t in = vld1q_u8(reinterpret_cast<const uint8_t*>(input));
  const uint16_t input_utf8_end_of_code_point_mask =
      utf8_end_of_code_point_mask & 0xfff;
  //
  // Optimization note: our main path below is load-latency dependent. Thus it is maybe
  // beneficial to have fast paths that depend on branch prediction but have less latency.
  // This results in more instructions but, potentially, also higher speeds.
  //
  // We first try a few fast paths.
  if((utf8_end_of_code_point_mask & 0xffff) == 0xffff) {
    // We process in chunks of 16 bytes
    uint16x8_t ascii_first = vmovl_u8(vget_low_u8 (in));
    uint16x8_t ascii_second = vmovl_high_u8(in);
    if (!match_system(big_endian)) {
      ascii_first = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(ascii_first), swap));
      ascii_second = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(ascii_second), swap));
    }
    vst1q_u16(reinterpret_cast<uint16_t*>(utf16_output), ascii_first);
    vst1q_u16(reinterpret_cast<uint16_t*>(utf16_output) + 8, ascii_second);
    utf16_output += 16; // We wrote 16 16-bit characters.
    return 16; // We consumed 16 bytes.
  }
  if((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa) {
    // We want to take 8 2-byte UTF-8 words and turn them into 8 2-byte UTF-16 words.
    // There is probably a more efficient sequence, but the following might do.
    uint8x16_t perm = vqtbl1q_u8(in, swap);
    uint8x16_t ascii = vandq_u8(perm, vreinterpretq_u8_u16(vmovq_n_u16(0x7f)));
    uint8x16_t highbyte = vandq_u8(perm, vreinterpretq_u8_u16(vmovq_n_u16(0x1f00)));
    uint8x16_t composed = vorrq_u8(ascii, vreinterpretq_u8_u16(vshrq_n_u16(vreinterpretq_u16_u8(highbyte), 2)));
    if (!match_system(big_endian)) composed = vqtbl1q_u8(composed, swap);
    vst1q_u8(reinterpret_cast<uint8_t*>(utf16_output), composed);
    utf16_output += 8; // We wrote 16 bytes, 8 code points.
    return 16;
  }
  if(input_utf8_end_of_code_point_mask == 0x924) {
    // We want to take 4 3-byte UTF-8 words and turn them into 4 2-byte UTF-16 words.
    // There is probably a more efficient sequence, but the following might do.
#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
    const uint8x16_t sh = make_uint8x16_t(2, 1, 0, 255, 5, 4, 3, 255, 8, 7, 6, 255, 11, 10, 9, 255);
#else
    const uint8x16_t sh = {2, 1, 0, 255, 5, 4, 3, 255, 8, 7, 6, 255, 11, 10, 9, 255};
#endif
    uint8x16_t perm = vqtbl1q_u8(in, sh);
    uint8x16_t ascii =
        vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x7f))); // 7 or 6 bits
    uint8x16_t middlebyte =
        vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x3f00))); // 5 or 6 bits
    uint8x16_t middlebyte_shifted = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(middlebyte), 2));
    uint32x4_t highbyte =
        vreinterpretq_u32_u8(vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x0f0000)))); // 4 bits
    uint32x4_t highbyte_shifted = vshrq_n_u32(highbyte, 4);
    uint32x4_t composed =
        vorrq_u32(vorrq_u32(vreinterpretq_u32_u8(ascii), vreinterpretq_u32_u8(middlebyte_shifted)), highbyte_shifted);
    uint16x8_t composed_repacked = vmovn_high_u32(vmovn_u32(composed), composed);
    if (!match_system(big_endian)) composed_repacked = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(composed_repacked), swap));
    vst1q_u16(reinterpret_cast<uint16_t*>(utf16_output), composed_repacked);
    utf16_output += 4;
    return 12;
  }
  /// We do not have a fast path available, so we fallback.

  const uint8_t idx =
      simdutf::tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0];
  const uint8_t consumed =
      simdutf::tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1];


  if (idx < 64) {
    // SIX (6) input code-words
    // this is a relatively easy scenario
    // we process SIX (6) input code-words. The max length in bytes of six code
    // words spanning between 1 and 2 bytes each is 12 bytes.
    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t*>(simdutf::tables::utf8_to_utf16::shufutf8[idx]));
    uint8x16_t perm = vqtbl1q_u8(in, sh);
    uint8x16_t ascii = vandq_u8(perm, vreinterpretq_u8_u16(vmovq_n_u16(0x7f)));
    uint8x16_t highbyte = vandq_u8(perm, vreinterpretq_u8_u16(vmovq_n_u16(0x1f00)));
    uint8x16_t composed = vorrq_u8(ascii, vreinterpretq_u8_u16(vshrq_n_u16(vreinterpretq_u16_u8(highbyte), 2)));
    if (!match_system(big_endian)) composed = vqtbl1q_u8(composed, swap);
    vst1q_u8(reinterpret_cast<uint8_t*>(utf16_output), composed);
    utf16_output += 6; // We wrote 12 bytes, 6 code points.
  } else if (idx < 145) {
    // FOUR (4) input code-words
    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t*>(simdutf::tables::utf8_to_utf16::shufutf8[idx]));
    uint8x16_t perm = vqtbl1q_u8(in, sh);
    uint8x16_t ascii =
        vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x7f))); // 7 or 6 bits
    uint8x16_t middlebyte =
        vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x3f00))); // 5 or 6 bits
    uint8x16_t middlebyte_shifted = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(middlebyte), 2));
    uint32x4_t highbyte =
        vreinterpretq_u32_u8(vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x0f0000)))); // 4 bits
    uint32x4_t highbyte_shifted = vshrq_n_u32(highbyte, 4);
    uint32x4_t composed =
        vorrq_u32(vorrq_u32(vreinterpretq_u32_u8(ascii), vreinterpretq_u32_u8(middlebyte_shifted)), highbyte_shifted);
    uint16x8_t composed_repacked = vmovn_high_u32(vmovn_u32(composed), composed);
    if (!match_system(big_endian)) composed_repacked = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(composed_repacked), swap));
    vst1q_u16(reinterpret_cast<uint16_t*>(utf16_output), composed_repacked);
    utf16_output += 4;
  } else if (idx < 209) {
    // TWO (2) input code-words
    //////////////
    // There might be garbage inputs where a leading byte mascarades as a four-byte
    // leading byte (by being followed by 3 continuation byte), but is not greater than
    // 0xf0. This could trigger a buffer overflow if we only counted leading
    // bytes of the form 0xf0 as generating surrogate pairs, without further UTF-8 validation.
    // Thus we must be careful to ensure that only leading bytes at least as large as 0xf0 generate surrogate pairs.
    // We do as at the cost of an extra mask.
    /////////////
    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t*>(simdutf::tables::utf8_to_utf16::shufutf8[idx]));
    uint8x16_t perm = vqtbl1q_u8(in, sh);
    uint8x16_t ascii = vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x7f)));
    uint8x16_t middlebyte = vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x3f00)));
    uint8x16_t middlebyte_shifted = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(middlebyte), 2));
    uint8x16_t middlehighbyte = vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x3f0000)));
    // correct for spurious high bit
    uint8x16_t correct =
        vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x400000)))), 1));
    middlehighbyte = veorq_u8(correct, middlehighbyte);
    uint8x16_t middlehighbyte_shifted = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(middlehighbyte), 4));
    // We deliberately carry the leading four bits if they are present, we remove
    // them later when computing hightenbits.
    uint8x16_t highbyte = vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0xff000000)));
    uint8x16_t highbyte_shifted = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(highbyte), 6));
    // When we need to generate a surrogate pair (leading byte > 0xF0), then
    // the corresponding 32-bit value in 'composed'  will be greater than
    // > (0xff00000>>6) or > 0x3c00000. This can be used later to identify the
    // location of the surrogate pairs.
    uint8x16_t composed =
        vorrq_u8(vorrq_u8(ascii, middlebyte_shifted),
                     vorrq_u8(highbyte_shifted, middlehighbyte_shifted));
    uint32x4_t composedminus =
        vsubq_u32(vreinterpretq_u32_u8(composed), vmovq_n_u32(0x10000));
    uint32x4_t lowtenbits =
        vandq_u32(composedminus, vmovq_n_u32(0x3ff));
    // Notice the 0x3ff mask:
    uint32x4_t hightenbits = vandq_u32(vshrq_n_u32(composedminus, 10), vmovq_n_u32(0x3ff));
    uint32x4_t lowtenbitsadd =
        vaddq_u32(lowtenbits, vmovq_n_u32(0xDC00));
    uint32x4_t hightenbitsadd =
        vaddq_u32(hightenbits, vmovq_n_u32(0xD800));
    uint32x4_t lowtenbitsaddshifted = vshlq_n_u32(lowtenbitsadd, 16);
    uint32x4_t surrogates =
        vorrq_u32(hightenbitsadd, lowtenbitsaddshifted);
    uint32_t basic_buffer[4];
    uint32_t basic_buffer_swap[4];
    if (!match_system(big_endian)) {
      vst1q_u32(basic_buffer_swap, vreinterpretq_u32_u8(vqtbl1q_u8(composed, swap)));
      surrogates = vreinterpretq_u32_u8(vqtbl1q_u8(vreinterpretq_u8_u32(surrogates), swap));
    }
    vst1q_u32(basic_buffer, vreinterpretq_u32_u8(composed));
    uint32_t surrogate_buffer[4];
    vst1q_u32(surrogate_buffer, surrogates);
    for (size_t i = 0; i < 3; i++) {
      if(basic_buffer[i] > 0x3c00000) {
        utf16_output[0] = uint16_t(surrogate_buffer[i] & 0xffff);
        utf16_output[1] = uint16_t(surrogate_buffer[i] >> 16);
        utf16_output += 2;
      } else {
        utf16_output[0] = !match_system(big_endian) ? uint16_t(basic_buffer_swap[i]) : uint16_t(basic_buffer[i]);
        utf16_output++;
      }
    }
  } else {
    // here we know that there is an error but we do not handle errors
  }
  return consumed;
}
/* end file src/arm64/arm_convert_utf8_to_utf16.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=arm64/arm_convert_utf8_to_utf32.cpp
/* begin file src/arm64/arm_convert_utf8_to_utf32.cpp */
// Convert up to 12 bytes from utf8 to utf32 using a mask indicating the
// end of the code points. Only the least significant 12 bits of the mask
// are accessed.
// It returns how many bytes were consumed (up to 12).
size_t convert_masked_utf8_to_utf32(const char *input,
                           uint64_t utf8_end_of_code_point_mask,
                           char32_t *&utf32_out) {
  // we use an approach where we try to process up to 12 input bytes.
  // Why 12 input bytes and not 16? Because we are concerned with the size of
  // the lookup tables. Also 12 is nicely divisible by two and three.
  //
  uint32_t*& utf32_output = reinterpret_cast<uint32_t*&>(utf32_out);
  uint8x16_t in = vld1q_u8(reinterpret_cast<const uint8_t*>(input));
  const uint16_t input_utf8_end_of_code_point_mask =
      utf8_end_of_code_point_mask & 0xFFF;
  //
  // Optimization note: our main path below is load-latency dependent. Thus it is maybe
  // beneficial to have fast paths that depend on branch prediction but have less latency.
  // This results in more instructions but, potentially, also higher speeds.
  //
  // We first try a few fast paths.
  if((utf8_end_of_code_point_mask & 0xffff) == 0xffff) {
    // We process in chunks of 16 bytes
    vst1q_u32(utf32_output, vmovl_u16(vget_low_u16(vmovl_u8(vget_low_u8 (in)))));
    vst1q_u32(utf32_output + 4, vmovl_high_u16(vmovl_u8(vget_low_u8 (in))));
    vst1q_u32(utf32_output + 8, vmovl_u16(vget_low_u16(vmovl_high_u8(in))));
    vst1q_u32(utf32_output + 12, vmovl_high_u16(vmovl_high_u8(in)));
    utf32_output += 16; // We wrote 16 16-bit characters.
    return 16; // We consumed 16 bytes.
  }
  if((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa) {
    // We want to take 8 2-byte UTF-8 words and turn them into 8 4-byte UTF-32 words.
    // There is probably a more efficient sequence, but the following might do.
#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
    const uint8x16_t sh = make_uint8x16_t(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
#else
    //const uint8x16_t sh = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
    const uint8x16_t sh = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
#endif
    uint8x16_t perm = vqtbl1q_u8(in, sh);
    uint8x16_t ascii = vandq_u8(perm, vreinterpretq_u8_u16(vmovq_n_u16(0x7f)));
    uint8x16_t highbyte = vandq_u8(perm, vreinterpretq_u8_u16(vmovq_n_u16(0x1f00)));
    uint8x16_t composed = vorrq_u8(ascii, vreinterpretq_u8_u16(vshrq_n_u16(vreinterpretq_u16_u8(highbyte), 2)));
    vst1q_u32(utf32_output,  vmovl_u16(vget_low_u16(vreinterpretq_u16_u8(composed))));
    vst1q_u32(utf32_output+4,  vmovl_high_u16(vreinterpretq_u16_u8(composed)));
    utf32_output += 8; // We wrote 32 bytes, 8 code points.
    return 16;
  }
  if(input_utf8_end_of_code_point_mask == 0x924) {
    // We want to take 4 3-byte UTF-8 words and turn them into 4 4-byte UTF-32 words.
    // There is probably a more efficient sequence, but the following might do.
#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
    const uint8x16_t sh = make_uint8x16_t(2, 1, 0, 255, 5, 4, 3, 255, 8, 7, 6, 255, 11, 10, 9, 255);
#else
    const uint8x16_t sh = {2, 1, 0, 255, 5, 4, 3, 255, 8, 7, 6, 255, 11, 10, 9, 255};
#endif
    uint8x16_t perm = vqtbl1q_u8(in, sh);
    uint8x16_t ascii =
        vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x7f))); // 7 or 6 bits
    uint8x16_t middlebyte =
        vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x3f00))); // 5 or 6 bits
    uint8x16_t middlebyte_shifted = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(middlebyte), 2));
    uint32x4_t highbyte =
        vreinterpretq_u32_u8(vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x0f0000)))); // 4 bits
    uint32x4_t highbyte_shifted = vshrq_n_u32(highbyte, 4);
    uint32x4_t composed =
        vorrq_u32(vorrq_u32(vreinterpretq_u32_u8(ascii), vreinterpretq_u32_u8(middlebyte_shifted)), highbyte_shifted);
    vst1q_u32(utf32_output, composed);
    utf32_output += 4;
    return 12;
  }
  /// We do not have a fast path available, so we fallback.

  const uint8_t idx =
      simdutf::tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0];
  const uint8_t consumed =
      simdutf::tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1];


  if (idx < 64) {
    // SIX (6) input code-words
    // this is a relatively easy scenario
    // we process SIX (6) input code-words. The max length in bytes of six code
    // words spanning between 1 and 2 bytes each is 12 bytes.
    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t*>(simdutf::tables::utf8_to_utf16::shufutf8[idx]));
    uint8x16_t perm = vqtbl1q_u8(in, sh);
    uint8x16_t ascii = vandq_u8(perm, vreinterpretq_u8_u16(vmovq_n_u16(0x7f)));
    uint8x16_t highbyte = vandq_u8(perm, vreinterpretq_u8_u16(vmovq_n_u16(0x1f00)));
    uint8x16_t composed = vorrq_u8(ascii, vreinterpretq_u8_u16(vshrq_n_u16(vreinterpretq_u16_u8(highbyte), 2)));
    vst1q_u32(utf32_output,  vmovl_u16(vget_low_u16(vreinterpretq_u16_u8(composed))));
    vst1q_u32(utf32_output+4,  vmovl_high_u16(vreinterpretq_u16_u8(composed)));
    utf32_output += 6; // We wrote 12 bytes, 6 code points.
  } else if (idx < 145) {
    // FOUR (4) input code-words
    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t*>(simdutf::tables::utf8_to_utf16::shufutf8[idx]));
    uint8x16_t perm = vqtbl1q_u8(in, sh);
    uint8x16_t ascii =
        vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x7f))); // 7 or 6 bits
    uint8x16_t middlebyte =
        vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x3f00))); // 5 or 6 bits
    uint8x16_t middlebyte_shifted = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(middlebyte), 2));
    uint32x4_t highbyte =
        vreinterpretq_u32_u8(vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x0f0000)))); // 4 bits
    uint32x4_t highbyte_shifted = vshrq_n_u32(highbyte, 4);
    uint32x4_t composed =
        vorrq_u32(vorrq_u32(vreinterpretq_u32_u8(ascii), vreinterpretq_u32_u8(middlebyte_shifted)), highbyte_shifted);
    vst1q_u32(utf32_output, composed);
    utf32_output += 4;
  } else if (idx < 209) {
    // TWO (2) input code-words
    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t*>(simdutf::tables::utf8_to_utf16::shufutf8[idx]));
    uint8x16_t perm = vqtbl1q_u8(in, sh);
    uint8x16_t ascii = vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x7f)));
    uint8x16_t middlebyte = vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x3f00)));
    uint8x16_t middlebyte_shifted = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(middlebyte), 2));
    uint8x16_t middlehighbyte = vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x3f0000)));
    // correct for spurious high bit
    uint8x16_t correct =
        vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x400000)))), 1));
    middlehighbyte = veorq_u8(correct, middlehighbyte);
    uint8x16_t middlehighbyte_shifted = vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(middlehighbyte), 4));
    uint8x16_t highbyte = vandq_u8(perm, vreinterpretq_u8_u32(vmovq_n_u32(0x07000000)));
    uint8x16_t highbyte_shifted =vreinterpretq_u8_u32(vshrq_n_u32(vreinterpretq_u32_u8(highbyte), 6));
    uint8x16_t composed =
        vorrq_u8(vorrq_u8(ascii, middlebyte_shifted),
                     vorrq_u8(highbyte_shifted, middlehighbyte_shifted));
    vst1q_u32(utf32_output, vreinterpretq_u32_u8(composed));
    utf32_output += 3;
  } else {
    // here we know that there is an error but we do not handle errors
  }
  return consumed;
}
/* end file src/arm64/arm_convert_utf8_to_utf32.cpp */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=arm64/arm_convert_utf16_to_utf8.cpp
/* begin file src/arm64/arm_convert_utf16_to_utf8.cpp */
/*
    The vectorized algorithm works on single SSE register i.e., it
    loads eight 16-bit words.

    We consider three cases:
    1. an input register contains no surrogates and each value
       is in range 0x0000 .. 0x07ff.
    2. an input register contains no surrogates and values are
       is in range 0x0000 .. 0xffff.
    3. an input register contains surrogates --- i.e. codepoints
       can have 16 or 32 bits.

    Ad 1.

    When values are less than 0x0800, it means that a 16-bit words
    can be converted into: 1) single UTF8 byte (when it's an ASCII
    char) or 2) two UTF8 bytes.

    For this case we do only some shuffle to obtain these 2-byte
    codes and finally compress the whole SSE register with a single
    shuffle.

    We need 256-entry lookup table to get a compression pattern
    and the number of output bytes in the compressed vector register.
    Each entry occupies 17 bytes.

    Ad 2.

    When values fit in 16-bit words, but are above 0x07ff, then
    a single word may produce one, two or three UTF8 bytes.

    We prepare data for all these three cases in two registers.
    The first register contains lower two UTF8 bytes (used in all
    cases), while the second one contains just the third byte for
    the three-UTF8-bytes case.

    Finally these two registers are interleaved forming eight-element
    array of 32-bit values. The array spans two SSE registers.
    The bytes from the registers are compressed using two shuffles.

    We need 256-entry lookup table to get a compression pattern
    and the number of output bytes in the compressed vector register.
    Each entry occupies 17 bytes.


    To summarize:
    - We need two 256-entry tables that have 8704 bytes in total.
*/
/*
  Returns a pair: the first unprocessed byte from buf and utf8_output
  A scalar routing should carry on the conversion of the tail.
*/
template <endianness big_endian>
std::pair<const char16_t*, char*> arm_convert_utf16_to_utf8(const char16_t* buf, size_t len, char* utf8_out) {
  uint8_t * utf8_output = reinterpret_cast<uint8_t*>(utf8_out);
  const char16_t* end = buf + len;

  const uint16x8_t v_f800 = vmovq_n_u16((uint16_t)0xf800);
  const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);
  const uint16x8_t v_c080 = vmovq_n_u16((uint16_t)0xc080);

  while (buf + 16 <= end) {
    uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t *>(buf));
    if (!match_system(big_endian)) {
      #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
      const uint8x16_t swap = make_uint8x16_t(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
      #else
      const uint8x16_t swap = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
      #endif
      in = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(in), swap));
    }
    if(vmaxvq_u16(in) <= 0x7F) { // ASCII fast path!!!!
        // It is common enough that we have sequences of 16 consecutive ASCII characters.
        uint16x8_t nextin = vld1q_u16(reinterpret_cast<const uint16_t *>(buf) + 8);
        if (!match_system(big_endian)) {
          #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
          const uint8x16_t swap = make_uint8x16_t(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
          #else
          const uint8x16_t swap = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
          #endif
          nextin = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(nextin), swap));
        }
        if(vmaxvq_u16(nextin) > 0x7F) {
          // 1. pack the bytes
          // obviously suboptimal.
          uint8x8_t utf8_packed = vmovn_u16(in);
          // 2. store (8 bytes)
          vst1_u8(utf8_output, utf8_packed);
          // 3. adjust pointers
          buf += 8;
          utf8_output += 8;
          in = nextin;
        } else {
          // 1. pack the bytes
          // obviously suboptimal.
          uint8x16_t utf8_packed = vmovn_high_u16(vmovn_u16(in), nextin);
          // 2. store (16 bytes)
          vst1q_u8(utf8_output, utf8_packed);
          // 3. adjust pointers
          buf += 16;
          utf8_output += 16;
          continue; // we are done for this round!
        }
    }

    if (vmaxvq_u16(in) <= 0x7FF) {
          // 1. prepare 2-byte values
          // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
          // expected output   : [110a|aaaa|10bb|bbbb] x 8
          const uint16x8_t v_1f00 = vmovq_n_u16((int16_t)0x1f00);
          const uint16x8_t v_003f = vmovq_n_u16((int16_t)0x003f);

          // t0 = [000a|aaaa|bbbb|bb00]
          const uint16x8_t t0 = vshlq_n_u16(in, 2);
          // t1 = [000a|aaaa|0000|0000]
          const uint16x8_t t1 = vandq_u16(t0, v_1f00);
          // t2 = [0000|0000|00bb|bbbb]
          const uint16x8_t t2 = vandq_u16(in, v_003f);
          // t3 = [000a|aaaa|00bb|bbbb]
          const uint16x8_t t3 = vorrq_u16(t1, t2);
          // t4 = [110a|aaaa|10bb|bbbb]
          const uint16x8_t t4 = vorrq_u16(t3, v_c080);
          // 2. merge ASCII and 2-byte codewords
          const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
          const uint16x8_t one_byte_bytemask = vcleq_u16(in, v_007f);
          const uint8x16_t utf8_unpacked = vreinterpretq_u8_u16(vbslq_u16(one_byte_bytemask, in, t4));
          // 3. prepare bitmask for 8-bit lookup
#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
          const uint16x8_t mask = make_uint16x8_t(0x0001, 0x0004,
                                    0x0010, 0x0040,
                                    0x0002, 0x0008,
                                    0x0020, 0x0080);
#else
          const uint16x8_t mask = { 0x0001, 0x0004,
                                    0x0010, 0x0040,
                                    0x0002, 0x0008,
                                    0x0020, 0x0080 };
#endif
          uint16_t m2 = vaddvq_u16(vandq_u16(one_byte_bytemask, mask));
          // 4. pack the bytes
          const uint8_t* row = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
          const uint8x16_t shuffle = vld1q_u8(row + 1);
          const uint8x16_t utf8_packed = vqtbl1q_u8(utf8_unpacked, shuffle);

          // 5. store bytes
          vst1q_u8(utf8_output, utf8_packed);

          // 6. adjust pointers
          buf += 8;
          utf8_output += row[0];
          continue;

    }
    const uint16x8_t surrogates_bytemask = vceqq_u16(vandq_u16(in, v_f800), v_d800);
    // It might seem like checking for surrogates_bitmask == 0xc000 could help. However,
    // it is likely an uncommon occurrence.
      if (vmaxvq_u16(surrogates_bytemask) == 0) {
      // case: words from register produce either 1, 2 or 3 UTF-8 bytes
#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
        const uint16x8_t dup_even = make_uint16x8_t(0x0000, 0x0202, 0x0404, 0x0606,
                                     0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
#else
        const uint16x8_t dup_even = {0x0000, 0x0202, 0x0404, 0x0606,
                                     0x0808, 0x0a0a, 0x0c0c, 0x0e0e};
#endif
        /* In this branch we handle three cases:
           1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           - single UFT-8 byte
           2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two UTF-8 bytes
           3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - three UTF-8 bytes

          We expand the input word (16-bit) into two words (32-bit), thus
          we have room for four bytes. However, we need five distinct bit
          layouts. Note that the last byte in cases #2 and #3 is the same.

          We precompute byte 1 for case #1 and the common byte for cases #2 & #3
          in register t2.

          We precompute byte 1 for case #3 and -- **conditionally** -- precompute
          either byte 1 for case #2 or byte 2 for case #3. Note that they
          differ by exactly one bit.

          Finally from these two words we build proper UTF-8 sequence, taking
          into account the case (i.e, the number of bytes to write).
        */
        /**
         * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
         * t2 => [0ccc|cccc] [10cc|cccc]
         * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
         */
#define simdutf_vec(x) vmovq_n_u16(static_cast<uint16_t>(x))
        // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
        const uint16x8_t t0 = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(in), vreinterpretq_u8_u16(dup_even)));
        // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
        const uint16x8_t t1 = vandq_u16(t0, simdutf_vec(0b0011111101111111));
        // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
        const uint16x8_t t2 = vorrq_u16 (t1, simdutf_vec(0b1000000000000000));

        // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
        const uint16x8_t s0 = vshrq_n_u16(in, 12);
        // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
        const uint16x8_t s1 = vandq_u16(in, simdutf_vec(0b0000111111000000));
        // [0000|bbbb|bb00|0000] => [00bb|bbbb|0000|0000]
        const uint16x8_t s1s = vshlq_n_u16(s1, 2);
        // [00bb|bbbb|0000|aaaa]
        const uint16x8_t s2 = vorrq_u16(s0, s1s);
        // s3: [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
        const uint16x8_t s3 = vorrq_u16(s2, simdutf_vec(0b1100000011100000));
        const uint16x8_t v_07ff = vmovq_n_u16((uint16_t)0x07FF);
        const uint16x8_t one_or_two_bytes_bytemask = vcleq_u16(in, v_07ff);
        const uint16x8_t m0 = vbicq_u16(simdutf_vec(0b0100000000000000), one_or_two_bytes_bytemask);
        const uint16x8_t s4 = veorq_u16(s3, m0);
#undef simdutf_vec

        // 4. expand words 16-bit => 32-bit
        const uint8x16_t out0 = vreinterpretq_u8_u16(vzip1q_u16(t2, s4));
        const uint8x16_t out1 = vreinterpretq_u8_u16(vzip2q_u16(t2, s4));

        // 5. compress 32-bit words into 1, 2 or 3 bytes -- 2 x shuffle
        const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
        const uint16x8_t one_byte_bytemask = vcleq_u16(in, v_007f);
#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
        const uint16x8_t onemask = make_uint16x8_t(0x0001, 0x0004,
                                    0x0010, 0x0040,
                                    0x0100, 0x0400,
                                    0x1000, 0x4000 );
        const uint16x8_t twomask = make_uint16x8_t(0x0002, 0x0008,
                                    0x0020, 0x0080,
                                    0x0200, 0x0800,
                                    0x2000, 0x8000 );
#else
        const uint16x8_t onemask = { 0x0001, 0x0004,
                                    0x0010, 0x0040,
                                    0x0100, 0x0400,
                                    0x1000, 0x4000 };
        const uint16x8_t twomask = { 0x0002, 0x0008,
                                    0x0020, 0x0080,
                                    0x0200, 0x0800,
                                    0x2000, 0x8000 };
#endif
        const uint16x8_t combined = vorrq_u16(vandq_u16(one_byte_bytemask, onemask), vandq_u16(one_or_two_bytes_bytemask, twomask));
        const uint16_t mask = vaddvq_u16(combined);
        // The following fast path may or may not be beneficial.
        /*if(mask == 0) {
          // We only have three-byte words. Use fast path.
          const uint8x16_t shuffle = {2,3,1,6,7,5,10,11,9,14,15,13,0,0,0,0};
          const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle);
          const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle);
          vst1q_u8(utf8_output, utf8_0);
          utf8_output += 12;
          vst1q_u8(utf8_output, utf8_1);
          utf8_output += 12;
          buf += 8;
          continue;
        }*/
        const uint8_t mask0 = uint8_t(mask);

        const uint8_t* row0 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
        const uint8x16_t shuffle0 = vld1q_u8(row0 + 1);
        const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle0);

        const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
        const uint8_t* row1 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
        const uint8x16_t shuffle1 = vld1q_u8(row1 + 1);
        const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle1);

        vst1q_u8(utf8_output, utf8_0);
        utf8_output += row0[0];
        vst1q_u8(utf8_output, utf8_1);
        utf8_output += row1[0];

        buf += 8;
    // surrogate pair(s) in a register
    } else {
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // in the presence of surrogate pairs may require non-trivial tables.
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint16_t word = !match_system(big_endian) ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
        if((word & 0xFF80)==0) {
          *utf8_output++ = char(word);
        } else if((word & 0xF800)==0) {
          *utf8_output++ = char((word>>6) | 0b11000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else if((word &0xF800 ) != 0xD800) {
          *utf8_output++ = char((word>>12) | 0b11100000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else {
          // must be a surrogate pair
          uint16_t diff = uint16_t(word - 0xD800);
          uint16_t next_word = !match_system(big_endian) ? scalar::utf16::swap_bytes(buf[k + 1]) : buf[k + 1];
          k++;
          uint16_t diff2 = uint16_t(next_word - 0xDC00);
          if((diff | diff2) > 0x3FF)  { return std::make_pair(nullptr, reinterpret_cast<char*>(utf8_output)); }
          uint32_t value = (diff << 10) + diff2 + 0x10000;
          *utf8_output++ = char((value>>18) | 0b11110000);
          *utf8_output++ = char(((value>>12) & 0b111111) | 0b10000000);
          *utf8_output++ = char(((value>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((value & 0b111111) | 0b10000000);
        }
      }
      buf += k;
    }
  } // while

  return std::make_pair(buf, reinterpret_cast<char*>(utf8_output));
}


/*
  Returns a pair: a result struct and utf8_output.
  If there is an error, the count field of the result is the position of the error.
  Otherwise, it is the position of the first unprocessed byte in buf (even if finished).
  A scalar routing should carry on the conversion of the tail if needed.
*/
template <endianness big_endian>
std::pair<result, char*> arm_convert_utf16_to_utf8_with_errors(const char16_t* buf, size_t len, char* utf8_out) {
  uint8_t * utf8_output = reinterpret_cast<uint8_t*>(utf8_out);
    const char16_t* start = buf;
  const char16_t* end = buf + len;

  const uint16x8_t v_f800 = vmovq_n_u16((uint16_t)0xf800);
  const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);
  const uint16x8_t v_c080 = vmovq_n_u16((uint16_t)0xc080);

  while (buf + 16 <= end) {
    uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t *>(buf));
    if (!match_system(big_endian)) {
      #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
      const uint8x16_t swap = make_uint8x16_t(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
      #else
      const uint8x16_t swap = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
      #endif
      in = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(in), swap));
    }
    if(vmaxvq_u16(in) <= 0x7F) { // ASCII fast path!!!!
        // It is common enough that we have sequences of 16 consecutive ASCII characters.
        uint16x8_t nextin = vld1q_u16(reinterpret_cast<const uint16_t *>(buf) + 8);
        if (!match_system(big_endian)) {
          #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
          const uint8x16_t swap = make_uint8x16_t(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
          #else
          const uint8x16_t swap = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
          #endif
          nextin = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(nextin), swap));
        }
        if(vmaxvq_u16(nextin) > 0x7F) {
          // 1. pack the bytes
          // obviously suboptimal.
          uint8x8_t utf8_packed = vmovn_u16(in);
          // 2. store (8 bytes)
          vst1_u8(utf8_output, utf8_packed);
          // 3. adjust pointers
          buf += 8;
          utf8_output += 8;
          in = nextin;
        } else {
          // 1. pack the bytes
          // obviously suboptimal.
          uint8x16_t utf8_packed = vmovn_high_u16(vmovn_u16(in), nextin);
          // 2. store (16 bytes)
          vst1q_u8(utf8_output, utf8_packed);
          // 3. adjust pointers
          buf += 16;
          utf8_output += 16;
          continue; // we are done for this round!
        }
    }

    if (vmaxvq_u16(in) <= 0x7FF) {
          // 1. prepare 2-byte values
          // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
          // expected output   : [110a|aaaa|10bb|bbbb] x 8
          const uint16x8_t v_1f00 = vmovq_n_u16((int16_t)0x1f00);
          const uint16x8_t v_003f = vmovq_n_u16((int16_t)0x003f);

          // t0 = [000a|aaaa|bbbb|bb00]
          const uint16x8_t t0 = vshlq_n_u16(in, 2);
          // t1 = [000a|aaaa|0000|0000]
          const uint16x8_t t1 = vandq_u16(t0, v_1f00);
          // t2 = [0000|0000|00bb|bbbb]
          const uint16x8_t t2 = vandq_u16(in, v_003f);
          // t3 = [000a|aaaa|00bb|bbbb]
          const uint16x8_t t3 = vorrq_u16(t1, t2);
          // t4 = [110a|aaaa|10bb|bbbb]
          const uint16x8_t t4 = vorrq_u16(t3, v_c080);
          // 2. merge ASCII and 2-byte codewords
          const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
          const uint16x8_t one_byte_bytemask = vcleq_u16(in, v_007f);
          const uint8x16_t utf8_unpacked = vreinterpretq_u8_u16(vbslq_u16(one_byte_bytemask, in, t4));
          // 3. prepare bitmask for 8-bit lookup
#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
          const uint16x8_t mask = make_uint16x8_t(0x0001, 0x0004,
                                    0x0010, 0x0040,
                                    0x0002, 0x0008,
                                    0x0020, 0x0080);
#else
          const uint16x8_t mask = { 0x0001, 0x0004,
                                    0x0010, 0x0040,
                                    0x0002, 0x0008,
                                    0x0020, 0x0080 };
#endif
          uint16_t m2 = vaddvq_u16(vandq_u16(one_byte_bytemask, mask));
          // 4. pack the bytes
          const uint8_t* row = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
          const uint8x16_t shuffle = vld1q_u8(row + 1);
          const uint8x16_t utf8_packed = vqtbl1q_u8(utf8_unpacked, shuffle);

          // 5. store bytes
          vst1q_u8(utf8_output, utf8_packed);

          // 6. adjust pointers
          buf += 8;
          utf8_output += row[0];
          continue;

    }
    const uint16x8_t surrogates_bytemask = vceqq_u16(vandq_u16(in, v_f800), v_d800);
    // It might seem like checking for surrogates_bitmask == 0xc000 could help. However,
    // it is likely an uncommon occurrence.
      if (vmaxvq_u16(surrogates_bytemask) == 0) {
      // case: words from register produce either 1, 2 or 3 UTF-8 bytes
#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
        const uint16x8_t dup_even = make_uint16x8_t(0x0000, 0x0202, 0x0404, 0x0606,
                                     0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
#else
        const uint16x8_t dup_even = {0x0000, 0x0202, 0x0404, 0x0606,
                                     0x0808, 0x0a0a, 0x0c0c, 0x0e0e};
#endif
        /* In this branch we handle three cases:
           1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           - single UFT-8 byte
           2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two UTF-8 bytes
           3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - three UTF-8 bytes

          We expand the input word (16-bit) into two words (32-bit), thus
          we have room for four bytes. However, we need five distinct bit
          layouts. Note that the last byte in cases #2 and #3 is the same.

          We precompute byte 1 for case #1 and the common byte for cases #2 & #3
          in register t2.

          We precompute byte 1 for case #3 and -- **conditionally** -- precompute
          either byte 1 for case #2 or byte 2 for case #3. Note that they
          differ by exactly one bit.

          Finally from these two words we build proper UTF-8 sequence, taking
          into account the case (i.e, the number of bytes to write).
        */
        /**
         * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
         * t2 => [0ccc|cccc] [10cc|cccc]
         * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
         */
#define simdutf_vec(x) vmovq_n_u16(static_cast<uint16_t>(x))
        // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
        const uint16x8_t t0 = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(in), vreinterpretq_u8_u16(dup_even)));
        // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
        const uint16x8_t t1 = vandq_u16(t0, simdutf_vec(0b0011111101111111));
        // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
        const uint16x8_t t2 = vorrq_u16 (t1, simdutf_vec(0b1000000000000000));

        // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
        const uint16x8_t s0 = vshrq_n_u16(in, 12);
        // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
        const uint16x8_t s1 = vandq_u16(in, simdutf_vec(0b0000111111000000));
        // [0000|bbbb|bb00|0000] => [00bb|bbbb|0000|0000]
        const uint16x8_t s1s = vshlq_n_u16(s1, 2);
        // [00bb|bbbb|0000|aaaa]
        const uint16x8_t s2 = vorrq_u16(s0, s1s);
        // s3: [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
        const uint16x8_t s3 = vorrq_u16(s2, simdutf_vec(0b1100000011100000));
        const uint16x8_t v_07ff = vmovq_n_u16((uint16_t)0x07FF);
        const uint16x8_t one_or_two_bytes_bytemask = vcleq_u16(in, v_07ff);
        const uint16x8_t m0 = vbicq_u16(simdutf_vec(0b0100000000000000), one_or_two_bytes_bytemask);
        const uint16x8_t s4 = veorq_u16(s3, m0);
#undef simdutf_vec

        // 4. expand words 16-bit => 32-bit
        const uint8x16_t out0 = vreinterpretq_u8_u16(vzip1q_u16(t2, s4));
        const uint8x16_t out1 = vreinterpretq_u8_u16(vzip2q_u16(t2, s4));

        // 5. compress 32-bit words into 1, 2 or 3 bytes -- 2 x shuffle
        const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
        const uint16x8_t one_byte_bytemask = vcleq_u16(in, v_007f);
#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
        const uint16x8_t onemask = make_uint16x8_t(0x0001, 0x0004,
                                    0x0010, 0x0040,
                                    0x0100, 0x0400,
                                    0x1000, 0x4000 );
        const uint16x8_t twomask = make_uint16x8_t(0x0002, 0x0008,
                                    0x0020, 0x0080,
                                    0x0200, 0x0800,
                                    0x2000, 0x8000 );
#else
        const uint16x8_t onemask = { 0x0001, 0x0004,
                                    0x0010, 0x0040,
                                    0x0100, 0x0400,
                                    0x1000, 0x4000 };
        const uint16x8_t twomask = { 0x0002, 0x0008,
                                    0x0020, 0x0080,
                                    0x0200, 0x0800,
                                    0x2000, 0x8000 };
#endif
        const uint16x8_t combined = vorrq_u16(vandq_u16(one_byte_bytemask, onemask), vandq_u16(one_or_two_bytes_bytemask, twomask));
        const uint16_t mask = vaddvq_u16(combined);
        // The following fast path may or may not be beneficial.
        /*if(mask == 0) {
          // We only have three-byte words. Use fast path.
          const uint8x16_t shuffle = {2,3,1,6,7,5,10,11,9,14,15,13,0,0,0,0};
          const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle);
          const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle);
          vst1q_u8(utf8_output, utf8_0);
          utf8_output += 12;
          vst1q_u8(utf8_output, utf8_1);
          utf8_output += 12;
          buf += 8;
          continue;
        }*/
        const uint8_t mask0 = uint8_t(mask);

        const uint8_t* row0 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
        const uint8x16_t shuffle0 = vld1q_u8(row0 + 1);
        const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle0);

        const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
        const uint8_t* row1 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
        const uint8x16_t shuffle1 = vld1q_u8(row1 + 1);
        const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle1);

        vst1q_u8(utf8_output, utf8_0);
        utf8_output += row0[0];
        vst1q_u8(utf8_output, utf8_1);
        utf8_output += row1[0];

        buf += 8;
    // surrogate pair(s) in a register
    } else {
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // in the presence of surrogate pairs may require non-trivial tables.
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint16_t word = !match_system(big_endian) ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
        if((word & 0xFF80)==0) {
          *utf8_output++ = char(word);
        } else if((word & 0xF800)==0) {
          *utf8_output++ = char((word>>6) | 0b11000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else if((word &0xF800 ) != 0xD800) {
          *utf8_output++ = char((word>>12) | 0b11100000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else {
          // must be a surrogate pair
          uint16_t diff = uint16_t(word - 0xD800);
          uint16_t next_word = !match_system(big_endian) ? scalar::utf16::swap_bytes(buf[k + 1]) : buf[k + 1];
          k++;
          uint16_t diff2 = uint16_t(next_word - 0xDC00);
          if((diff | diff2) > 0x3FF)  { return std::make_pair(result(error_code::SURROGATE, buf - start + k - 1), reinterpret_cast<char*>(utf8_output)); }
          uint32_t value = (diff << 10) + diff2 + 0x10000;
          *utf8_output++ = char((value>>18) | 0b11110000);
          *utf8_output++ = char(((value>>12) & 0b111111) | 0b10000000);
          *utf8_output++ = char(((value>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((value & 0b111111) | 0b10000000);
        }
      }
      buf += k;
    }
  } // while

  return std::make_pair(result(error_code::SUCCESS, buf - start), reinterpret_cast<char*>(utf8_output));
}
/* end file src/arm64/arm_convert_utf16_to_utf8.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=arm64/arm_convert_utf16_to_utf32.cpp
/* begin file src/arm64/arm_convert_utf16_to_utf32.cpp */
/*
    The vectorized algorithm works on single SSE register i.e., it
    loads eight 16-bit words.

    We consider three cases:
    1. an input register contains no surrogates and each value
       is in range 0x0000 .. 0x07ff.
    2. an input register contains no surrogates and values are
       is in range 0x0000 .. 0xffff.
    3. an input register contains surrogates --- i.e. codepoints
       can have 16 or 32 bits.

    Ad 1.

    When values are less than 0x0800, it means that a 16-bit words
    can be converted into: 1) single UTF8 byte (when it's an ASCII
    char) or 2) two UTF8 bytes.

    For this case we do only some shuffle to obtain these 2-byte
    codes and finally compress the whole SSE register with a single
    shuffle.

    We need 256-entry lookup table to get a compression pattern
    and the number of output bytes in the compressed vector register.
    Each entry occupies 17 bytes.

    Ad 2.

    When values fit in 16-bit words, but are above 0x07ff, then
    a single word may produce one, two or three UTF8 bytes.

    We prepare data for all these three cases in two registers.
    The first register contains lower two UTF8 bytes (used in all
    cases), while the second one contains just the third byte for
    the three-UTF8-bytes case.

    Finally these two registers are interleaved forming eight-element
    array of 32-bit values. The array spans two SSE registers.
    The bytes from the registers are compressed using two shuffles.

    We need 256-entry lookup table to get a compression pattern
    and the number of output bytes in the compressed vector register.
    Each entry occupies 17 bytes.


    To summarize:
    - We need two 256-entry tables that have 8704 bytes in total.
*/
/*
  Returns a pair: the first unprocessed byte from buf and utf8_output
  A scalar routing should carry on the conversion of the tail.
*/
template <endianness big_endian>
std::pair<const char16_t*, char32_t*> arm_convert_utf16_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_out) {
  uint32_t * utf32_output = reinterpret_cast<uint32_t*>(utf32_out);
  const char16_t* end = buf + len;

  const uint16x8_t v_f800 = vmovq_n_u16((uint16_t)0xf800);
  const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);

  while (buf + 16 <= end) {
    uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t *>(buf));
    if (!match_system(big_endian)) {
      #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
      const uint8x16_t swap = make_uint8x16_t(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
      #else
      const uint8x16_t swap = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
      #endif
      in = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(in), swap));
    }

    const uint16x8_t surrogates_bytemask = vceqq_u16(vandq_u16(in, v_f800), v_d800);
    // It might seem like checking for surrogates_bitmask == 0xc000 could help. However,
    // it is likely an uncommon occurrence.
      if (vmaxvq_u16(surrogates_bytemask) == 0) {
      // case: no surrogate pairs, extend all 16-bit words to 32-bit words
      vst1q_u32(utf32_output,  vmovl_u16(vget_low_u16(in)));
      vst1q_u32(utf32_output+4,  vmovl_high_u16(in));
      utf32_output += 8;
      buf += 8;
    // surrogate pair(s) in a register
    } else {
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // in the presence of surrogate pairs may require non-trivial tables.
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint16_t word = !match_system(big_endian) ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
        if((word &0xF800 ) != 0xD800) {
          *utf32_output++ = char32_t(word);
        } else {
          // must be a surrogate pair
          uint16_t diff = uint16_t(word - 0xD800);
          uint16_t next_word = !match_system(big_endian) ? scalar::utf16::swap_bytes(buf[k + 1]) : buf[k + 1];
          k++;
          uint16_t diff2 = uint16_t(next_word - 0xDC00);
          if((diff | diff2) > 0x3FF)  { return std::make_pair(nullptr, reinterpret_cast<char32_t*>(utf32_output)); }
          uint32_t value = (diff << 10) + diff2 + 0x10000;
          *utf32_output++ = char32_t(value);
        }
      }
      buf += k;
    }
  } // while
  return std::make_pair(buf, reinterpret_cast<char32_t*>(utf32_output));
}


/*
  Returns a pair: a result struct and utf8_output.
  If there is an error, the count field of the result is the position of the error.
  Otherwise, it is the position of the first unprocessed byte in buf (even if finished).
  A scalar routing should carry on the conversion of the tail if needed.
*/
template <endianness big_endian>
std::pair<result, char32_t*> arm_convert_utf16_to_utf32_with_errors(const char16_t* buf, size_t len, char32_t* utf32_out) {
  uint32_t * utf32_output = reinterpret_cast<uint32_t*>(utf32_out);
  const char16_t* start = buf;
  const char16_t* end = buf + len;

  const uint16x8_t v_f800 = vmovq_n_u16((uint16_t)0xf800);
  const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);

  while (buf + 16 <= end) {
    uint16x8_t in = vld1q_u16(reinterpret_cast<const uint16_t *>(buf));
    if (!match_system(big_endian)) {
      #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
      const uint8x16_t swap = make_uint8x16_t(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
      #else
      const uint8x16_t swap = {1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14};
      #endif
      in = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(in), swap));
    }

    const uint16x8_t surrogates_bytemask = vceqq_u16(vandq_u16(in, v_f800), v_d800);
    // It might seem like checking for surrogates_bitmask == 0xc000 could help. However,
    // it is likely an uncommon occurrence.
      if (vmaxvq_u16(surrogates_bytemask) == 0) {
      // case: no surrogate pairs, extend all 16-bit words to 32-bit words
      vst1q_u32(utf32_output,  vmovl_u16(vget_low_u16(in)));
      vst1q_u32(utf32_output+4,  vmovl_high_u16(in));
      utf32_output += 8;
      buf += 8;
    // surrogate pair(s) in a register
    } else {
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // in the presence of surrogate pairs may require non-trivial tables.
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint16_t word = !match_system(big_endian) ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
        if((word &0xF800 ) != 0xD800) {
          *utf32_output++ = char32_t(word);
        } else {
          // must be a surrogate pair
          uint16_t diff = uint16_t(word - 0xD800);
          uint16_t next_word = !match_system(big_endian) ? scalar::utf16::swap_bytes(buf[k + 1]) : buf[k + 1];
          k++;
          uint16_t diff2 = uint16_t(next_word - 0xDC00);
          if((diff | diff2) > 0x3FF)  { return std::make_pair(result(error_code::SURROGATE, buf - start + k - 1), reinterpret_cast<char32_t*>(utf32_output)); }
          uint32_t value = (diff << 10) + diff2 + 0x10000;
          *utf32_output++ = char32_t(value);
        }
      }
      buf += k;
    }
  } // while
  return std::make_pair(result(error_code::SUCCESS, buf - start), reinterpret_cast<char32_t*>(utf32_output));
}
/* end file src/arm64/arm_convert_utf16_to_utf32.cpp */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=arm64/arm_convert_utf32_to_utf8.cpp
/* begin file src/arm64/arm_convert_utf32_to_utf8.cpp */
std::pair<const char32_t*, char*> arm_convert_utf32_to_utf8(const char32_t* buf, size_t len, char* utf8_out) {
  uint8_t * utf8_output = reinterpret_cast<uint8_t*>(utf8_out);
  const char32_t* end = buf + len;

  const uint16x8_t v_c080 = vmovq_n_u16((uint16_t)0xc080);

  uint16x8_t forbidden_bytemask = vmovq_n_u16(0x0);

  while (buf + 16 <= end) {
    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(buf));
    uint32x4_t nextin = vld1q_u32(reinterpret_cast<const uint32_t *>(buf+4));

    // Check if no bits set above 16th
    if(vmaxvq_u32(vorrq_u32(in, nextin)) <= 0xFFFF) {
      // Pack UTF-32 to UTF-16 safely (without surrogate pairs)
      // Apply UTF-16 => UTF-8 routine (arm_convert_utf16_to_utf8.cpp)
      uint16x8_t utf16_packed = vcombine_u16(vmovn_u32(in), vmovn_u32(nextin));
      if(vmaxvq_u16(utf16_packed) <= 0x7F) { // ASCII fast path!!!!
          // 1. pack the bytes
          // obviously suboptimal.
          uint8x8_t utf8_packed = vmovn_u16(utf16_packed);
          // 2. store (8 bytes)
          vst1_u8(utf8_output, utf8_packed);
          // 3. adjust pointers
          buf += 8;
          utf8_output += 8;
          continue; // we are done for this round!
      }

      if (vmaxvq_u16(utf16_packed) <= 0x7FF) {
            // 1. prepare 2-byte values
            // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
            // expected output   : [110a|aaaa|10bb|bbbb] x 8
            const uint16x8_t v_1f00 = vmovq_n_u16((int16_t)0x1f00);
            const uint16x8_t v_003f = vmovq_n_u16((int16_t)0x003f);

            // t0 = [000a|aaaa|bbbb|bb00]
            const uint16x8_t t0 = vshlq_n_u16(utf16_packed, 2);
            // t1 = [000a|aaaa|0000|0000]
            const uint16x8_t t1 = vandq_u16(t0, v_1f00);
            // t2 = [0000|0000|00bb|bbbb]
            const uint16x8_t t2 = vandq_u16(utf16_packed, v_003f);
            // t3 = [000a|aaaa|00bb|bbbb]
            const uint16x8_t t3 = vorrq_u16(t1, t2);
            // t4 = [110a|aaaa|10bb|bbbb]
            const uint16x8_t t4 = vorrq_u16(t3, v_c080);
            // 2. merge ASCII and 2-byte codewords
            const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
            const uint16x8_t one_byte_bytemask = vcleq_u16(utf16_packed, v_007f);
            const uint8x16_t utf8_unpacked = vreinterpretq_u8_u16(vbslq_u16(one_byte_bytemask, utf16_packed, t4));
            // 3. prepare bitmask for 8-bit lookup
  #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
            const uint16x8_t mask = make_uint16x8_t(0x0001, 0x0004,
                                      0x0010, 0x0040,
                                      0x0002, 0x0008,
                                      0x0020, 0x0080);
  #else
            const uint16x8_t mask = { 0x0001, 0x0004,
                                      0x0010, 0x0040,
                                      0x0002, 0x0008,
                                      0x0020, 0x0080 };
  #endif
            uint16_t m2 = vaddvq_u16(vandq_u16(one_byte_bytemask, mask));
            // 4. pack the bytes
            const uint8_t* row = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
            const uint8x16_t shuffle = vld1q_u8(row + 1);
            const uint8x16_t utf8_packed = vqtbl1q_u8(utf8_unpacked, shuffle);

            // 5. store bytes
            vst1q_u8(utf8_output, utf8_packed);

            // 6. adjust pointers
            buf += 8;
            utf8_output += row[0];
            continue;

      } else {
        // case: words from register produce either 1, 2 or 3 UTF-8 bytes
        const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);
        const uint16x8_t v_dfff = vmovq_n_u16((uint16_t)0xdfff);
        forbidden_bytemask = vorrq_u16(vandq_u16(vcleq_u16(utf16_packed, v_dfff), vcgeq_u16(utf16_packed, v_d800)), forbidden_bytemask);

  #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
          const uint16x8_t dup_even = make_uint16x8_t(0x0000, 0x0202, 0x0404, 0x0606,
                                      0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
  #else
          const uint16x8_t dup_even = {0x0000, 0x0202, 0x0404, 0x0606,
                                      0x0808, 0x0a0a, 0x0c0c, 0x0e0e};
  #endif
          /* In this branch we handle three cases:
            1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           - single UFT-8 byte
            2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two UTF-8 bytes
            3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - three UTF-8 bytes

            We expand the input word (16-bit) into two words (32-bit), thus
            we have room for four bytes. However, we need five distinct bit
            layouts. Note that the last byte in cases #2 and #3 is the same.

            We precompute byte 1 for case #1 and the common byte for cases #2 & #3
            in register t2.

            We precompute byte 1 for case #3 and -- **conditionally** -- precompute
            either byte 1 for case #2 or byte 2 for case #3. Note that they
            differ by exactly one bit.

            Finally from these two words we build proper UTF-8 sequence, taking
            into account the case (i.e, the number of bytes to write).
          */
          /**
           * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
           * t2 => [0ccc|cccc] [10cc|cccc]
           * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
           */
  #define simdutf_vec(x) vmovq_n_u16(static_cast<uint16_t>(x))
          // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
          const uint16x8_t t0 = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(utf16_packed), vreinterpretq_u8_u16(dup_even)));
          // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
          const uint16x8_t t1 = vandq_u16(t0, simdutf_vec(0b0011111101111111));
          // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
          const uint16x8_t t2 = vorrq_u16 (t1, simdutf_vec(0b1000000000000000));

          // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
          const uint16x8_t s0 = vshrq_n_u16(utf16_packed, 12);
          // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
          const uint16x8_t s1 = vandq_u16(utf16_packed, simdutf_vec(0b0000111111000000));
          // [0000|bbbb|bb00|0000] => [00bb|bbbb|0000|0000]
          const uint16x8_t s1s = vshlq_n_u16(s1, 2);
          // [00bb|bbbb|0000|aaaa]
          const uint16x8_t s2 = vorrq_u16(s0, s1s);
          // s3: [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
          const uint16x8_t s3 = vorrq_u16(s2, simdutf_vec(0b1100000011100000));
          const uint16x8_t v_07ff = vmovq_n_u16((uint16_t)0x07FF);
          const uint16x8_t one_or_two_bytes_bytemask = vcleq_u16(utf16_packed, v_07ff);
          const uint16x8_t m0 = vbicq_u16(simdutf_vec(0b0100000000000000), one_or_two_bytes_bytemask);
          const uint16x8_t s4 = veorq_u16(s3, m0);
  #undef simdutf_vec

          // 4. expand words 16-bit => 32-bit
          const uint8x16_t out0 = vreinterpretq_u8_u16(vzip1q_u16(t2, s4));
          const uint8x16_t out1 = vreinterpretq_u8_u16(vzip2q_u16(t2, s4));

          // 5. compress 32-bit words into 1, 2 or 3 bytes -- 2 x shuffle
          const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
          const uint16x8_t one_byte_bytemask = vcleq_u16(utf16_packed, v_007f);
  #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
          const uint16x8_t onemask = make_uint16x8_t(0x0001, 0x0004,
                                      0x0010, 0x0040,
                                      0x0100, 0x0400,
                                      0x1000, 0x4000 );
          const uint16x8_t twomask = make_uint16x8_t(0x0002, 0x0008,
                                      0x0020, 0x0080,
                                      0x0200, 0x0800,
                                      0x2000, 0x8000 );
  #else
          const uint16x8_t onemask = { 0x0001, 0x0004,
                                      0x0010, 0x0040,
                                      0x0100, 0x0400,
                                      0x1000, 0x4000 };
          const uint16x8_t twomask = { 0x0002, 0x0008,
                                      0x0020, 0x0080,
                                      0x0200, 0x0800,
                                      0x2000, 0x8000 };
  #endif
          const uint16x8_t combined = vorrq_u16(vandq_u16(one_byte_bytemask, onemask), vandq_u16(one_or_two_bytes_bytemask, twomask));
          const uint16_t mask = vaddvq_u16(combined);
          // The following fast path may or may not be beneficial.
          /*if(mask == 0) {
            // We only have three-byte words. Use fast path.
            const uint8x16_t shuffle = {2,3,1,6,7,5,10,11,9,14,15,13,0,0,0,0};
            const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle);
            const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle);
            vst1q_u8(utf8_output, utf8_0);
            utf8_output += 12;
            vst1q_u8(utf8_output, utf8_1);
            utf8_output += 12;
            buf += 8;
            continue;
          }*/
          const uint8_t mask0 = uint8_t(mask);
          const uint8_t* row0 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
          const uint8x16_t shuffle0 = vld1q_u8(row0 + 1);
          const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle0);

          const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
          const uint8_t* row1 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
          const uint8x16_t shuffle1 = vld1q_u8(row1 + 1);
          const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle1);

          vst1q_u8(utf8_output, utf8_0);
          utf8_output += row0[0];
          vst1q_u8(utf8_output, utf8_1);
          utf8_output += row1[0];

          buf += 8;
      }
    // At least one 32-bit word will produce a surrogate pair in UTF-16 <=> will produce four UTF-8 bytes.
    } else {
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // in the presence of surrogate pairs may require non-trivial tables.
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint32_t word = buf[k];
        if((word & 0xFFFFFF80)==0) {
          *utf8_output++ = char(word);
        } else if((word & 0xFFFFF800)==0) {
          *utf8_output++ = char((word>>6) | 0b11000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else if((word & 0xFFFF0000)==0) {
          if (word >= 0xD800 && word <= 0xDFFF) { return std::make_pair(nullptr, reinterpret_cast<char*>(utf8_output)); }
          *utf8_output++ = char((word>>12) | 0b11100000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else {
          if (word > 0x10FFFF) { return std::make_pair(nullptr, reinterpret_cast<char*>(utf8_output)); }
          *utf8_output++ = char((word>>18) | 0b11110000);
          *utf8_output++ = char(((word>>12) & 0b111111) | 0b10000000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        }
      }
      buf += k;
    }
  } // while

  // check for invalid input
  if (vmaxvq_u16(forbidden_bytemask) != 0) {
    return std::make_pair(nullptr, reinterpret_cast<char*>(utf8_output));
  }
  return std::make_pair(buf, reinterpret_cast<char*>(utf8_output));
}


std::pair<result, char*> arm_convert_utf32_to_utf8_with_errors(const char32_t* buf, size_t len, char* utf8_out) {
  uint8_t * utf8_output = reinterpret_cast<uint8_t*>(utf8_out);
  const char32_t* start = buf;
  const char32_t* end = buf + len;

  const uint16x8_t v_c080 = vmovq_n_u16((uint16_t)0xc080);

  while (buf + 16 <= end) {
    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(buf));
    uint32x4_t nextin = vld1q_u32(reinterpret_cast<const uint32_t *>(buf+4));

    // Check if no bits set above 16th
    if(vmaxvq_u32(vorrq_u32(in, nextin)) <= 0xFFFF) {
      // Pack UTF-32 to UTF-16 safely (without surrogate pairs)
      // Apply UTF-16 => UTF-8 routine (arm_convert_utf16_to_utf8.cpp)
      uint16x8_t utf16_packed = vcombine_u16(vmovn_u32(in), vmovn_u32(nextin));
      if(vmaxvq_u16(utf16_packed) <= 0x7F) { // ASCII fast path!!!!
          // 1. pack the bytes
          // obviously suboptimal.
          uint8x8_t utf8_packed = vmovn_u16(utf16_packed);
          // 2. store (8 bytes)
          vst1_u8(utf8_output, utf8_packed);
          // 3. adjust pointers
          buf += 8;
          utf8_output += 8;
          continue; // we are done for this round!
      }

      if (vmaxvq_u16(utf16_packed) <= 0x7FF) {
            // 1. prepare 2-byte values
            // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
            // expected output   : [110a|aaaa|10bb|bbbb] x 8
            const uint16x8_t v_1f00 = vmovq_n_u16((int16_t)0x1f00);
            const uint16x8_t v_003f = vmovq_n_u16((int16_t)0x003f);

            // t0 = [000a|aaaa|bbbb|bb00]
            const uint16x8_t t0 = vshlq_n_u16(utf16_packed, 2);
            // t1 = [000a|aaaa|0000|0000]
            const uint16x8_t t1 = vandq_u16(t0, v_1f00);
            // t2 = [0000|0000|00bb|bbbb]
            const uint16x8_t t2 = vandq_u16(utf16_packed, v_003f);
            // t3 = [000a|aaaa|00bb|bbbb]
            const uint16x8_t t3 = vorrq_u16(t1, t2);
            // t4 = [110a|aaaa|10bb|bbbb]
            const uint16x8_t t4 = vorrq_u16(t3, v_c080);
            // 2. merge ASCII and 2-byte codewords
            const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
            const uint16x8_t one_byte_bytemask = vcleq_u16(utf16_packed, v_007f);
            const uint8x16_t utf8_unpacked = vreinterpretq_u8_u16(vbslq_u16(one_byte_bytemask, utf16_packed, t4));
            // 3. prepare bitmask for 8-bit lookup
  #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
            const uint16x8_t mask = make_uint16x8_t(0x0001, 0x0004,
                                      0x0010, 0x0040,
                                      0x0002, 0x0008,
                                      0x0020, 0x0080);
  #else
            const uint16x8_t mask = { 0x0001, 0x0004,
                                      0x0010, 0x0040,
                                      0x0002, 0x0008,
                                      0x0020, 0x0080 };
  #endif
            uint16_t m2 = vaddvq_u16(vandq_u16(one_byte_bytemask, mask));
            // 4. pack the bytes
            const uint8_t* row = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
            const uint8x16_t shuffle = vld1q_u8(row + 1);
            const uint8x16_t utf8_packed = vqtbl1q_u8(utf8_unpacked, shuffle);

            // 5. store bytes
            vst1q_u8(utf8_output, utf8_packed);

            // 6. adjust pointers
            buf += 8;
            utf8_output += row[0];
            continue;

      } else {
        // case: words from register produce either 1, 2 or 3 UTF-8 bytes

        // check for invalid input
        const uint16x8_t v_d800 = vmovq_n_u16((uint16_t)0xd800);
        const uint16x8_t v_dfff = vmovq_n_u16((uint16_t)0xdfff);
        const uint16x8_t forbidden_bytemask = vandq_u16(vcleq_u16(utf16_packed, v_dfff), vcgeq_u16(utf16_packed, v_d800));
        if (vmaxvq_u16(forbidden_bytemask) != 0) {
          return std::make_pair(result(error_code::SURROGATE, buf - start), reinterpret_cast<char*>(utf8_output));
        }

  #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
          const uint16x8_t dup_even = make_uint16x8_t(0x0000, 0x0202, 0x0404, 0x0606,
                                      0x0808, 0x0a0a, 0x0c0c, 0x0e0e);
  #else
          const uint16x8_t dup_even = {0x0000, 0x0202, 0x0404, 0x0606,
                                      0x0808, 0x0a0a, 0x0c0c, 0x0e0e};
  #endif
          /* In this branch we handle three cases:
            1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           - single UFT-8 byte
            2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two UTF-8 bytes
            3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - three UTF-8 bytes

            We expand the input word (16-bit) into two words (32-bit), thus
            we have room for four bytes. However, we need five distinct bit
            layouts. Note that the last byte in cases #2 and #3 is the same.

            We precompute byte 1 for case #1 and the common byte for cases #2 & #3
            in register t2.

            We precompute byte 1 for case #3 and -- **conditionally** -- precompute
            either byte 1 for case #2 or byte 2 for case #3. Note that they
            differ by exactly one bit.

            Finally from these two words we build proper UTF-8 sequence, taking
            into account the case (i.e, the number of bytes to write).
          */
          /**
           * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
           * t2 => [0ccc|cccc] [10cc|cccc]
           * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
           */
  #define simdutf_vec(x) vmovq_n_u16(static_cast<uint16_t>(x))
          // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
          const uint16x8_t t0 = vreinterpretq_u16_u8(vqtbl1q_u8(vreinterpretq_u8_u16(utf16_packed), vreinterpretq_u8_u16(dup_even)));
          // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
          const uint16x8_t t1 = vandq_u16(t0, simdutf_vec(0b0011111101111111));
          // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
          const uint16x8_t t2 = vorrq_u16 (t1, simdutf_vec(0b1000000000000000));

          // s0: [aaaa|bbbb|bbcc|cccc] => [0000|0000|0000|aaaa]
          const uint16x8_t s0 = vshrq_n_u16(utf16_packed, 12);
          // s1: [aaaa|bbbb|bbcc|cccc] => [0000|bbbb|bb00|0000]
          const uint16x8_t s1 = vandq_u16(utf16_packed, simdutf_vec(0b0000111111000000));
          // [0000|bbbb|bb00|0000] => [00bb|bbbb|0000|0000]
          const uint16x8_t s1s = vshlq_n_u16(s1, 2);
          // [00bb|bbbb|0000|aaaa]
          const uint16x8_t s2 = vorrq_u16(s0, s1s);
          // s3: [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
          const uint16x8_t s3 = vorrq_u16(s2, simdutf_vec(0b1100000011100000));
          const uint16x8_t v_07ff = vmovq_n_u16((uint16_t)0x07FF);
          const uint16x8_t one_or_two_bytes_bytemask = vcleq_u16(utf16_packed, v_07ff);
          const uint16x8_t m0 = vbicq_u16(simdutf_vec(0b0100000000000000), one_or_two_bytes_bytemask);
          const uint16x8_t s4 = veorq_u16(s3, m0);
  #undef simdutf_vec

          // 4. expand words 16-bit => 32-bit
          const uint8x16_t out0 = vreinterpretq_u8_u16(vzip1q_u16(t2, s4));
          const uint8x16_t out1 = vreinterpretq_u8_u16(vzip2q_u16(t2, s4));

          // 5. compress 32-bit words into 1, 2 or 3 bytes -- 2 x shuffle
          const uint16x8_t v_007f = vmovq_n_u16((uint16_t)0x007F);
          const uint16x8_t one_byte_bytemask = vcleq_u16(utf16_packed, v_007f);
  #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
          const uint16x8_t onemask = make_uint16x8_t(0x0001, 0x0004,
                                      0x0010, 0x0040,
                                      0x0100, 0x0400,
                                      0x1000, 0x4000 );
          const uint16x8_t twomask = make_uint16x8_t(0x0002, 0x0008,
                                      0x0020, 0x0080,
                                      0x0200, 0x0800,
                                      0x2000, 0x8000 );
  #else
          const uint16x8_t onemask = { 0x0001, 0x0004,
                                      0x0010, 0x0040,
                                      0x0100, 0x0400,
                                      0x1000, 0x4000 };
          const uint16x8_t twomask = { 0x0002, 0x0008,
                                      0x0020, 0x0080,
                                      0x0200, 0x0800,
                                      0x2000, 0x8000 };
  #endif
          const uint16x8_t combined = vorrq_u16(vandq_u16(one_byte_bytemask, onemask), vandq_u16(one_or_two_bytes_bytemask, twomask));
          const uint16_t mask = vaddvq_u16(combined);
          // The following fast path may or may not be beneficial.
          /*if(mask == 0) {
            // We only have three-byte words. Use fast path.
            const uint8x16_t shuffle = {2,3,1,6,7,5,10,11,9,14,15,13,0,0,0,0};
            const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle);
            const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle);
            vst1q_u8(utf8_output, utf8_0);
            utf8_output += 12;
            vst1q_u8(utf8_output, utf8_1);
            utf8_output += 12;
            buf += 8;
            continue;
          }*/
          const uint8_t mask0 = uint8_t(mask);

          const uint8_t* row0 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
          const uint8x16_t shuffle0 = vld1q_u8(row0 + 1);
          const uint8x16_t utf8_0 = vqtbl1q_u8(out0, shuffle0);

          const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
          const uint8_t* row1 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
          const uint8x16_t shuffle1 = vld1q_u8(row1 + 1);
          const uint8x16_t utf8_1 = vqtbl1q_u8(out1, shuffle1);

          vst1q_u8(utf8_output, utf8_0);
          utf8_output += row0[0];
          vst1q_u8(utf8_output, utf8_1);
          utf8_output += row1[0];

          buf += 8;
      }
    // At least one 32-bit word will produce a surrogate pair in UTF-16 <=> will produce four UTF-8 bytes.
    } else {
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // in the presence of surrogate pairs may require non-trivial tables.
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint32_t word = buf[k];
        if((word & 0xFFFFFF80)==0) {
          *utf8_output++ = char(word);
        } else if((word & 0xFFFFF800)==0) {
          *utf8_output++ = char((word>>6) | 0b11000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else if((word & 0xFFFF0000)==0) {
          if (word >= 0xD800 && word <= 0xDFFF) { return std::make_pair(result(error_code::SURROGATE, buf - start + k), reinterpret_cast<char*>(utf8_output)); }
          *utf8_output++ = char((word>>12) | 0b11100000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else {
          if (word > 0x10FFFF) { return std::make_pair(result(error_code::TOO_LARGE, buf - start + k), reinterpret_cast<char*>(utf8_output)); }
          *utf8_output++ = char((word>>18) | 0b11110000);
          *utf8_output++ = char(((word>>12) & 0b111111) | 0b10000000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        }
      }
      buf += k;
    }
  } // while

  return std::make_pair(result(error_code::SUCCESS, buf - start), reinterpret_cast<char*>(utf8_output));
}
/* end file src/arm64/arm_convert_utf32_to_utf8.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=arm64/arm_convert_utf32_to_utf16.cpp
/* begin file src/arm64/arm_convert_utf32_to_utf16.cpp */
template <endianness big_endian>
std::pair<const char32_t*, char16_t*> arm_convert_utf32_to_utf16(const char32_t* buf, size_t len, char16_t* utf16_out) {
  uint16_t * utf16_output = reinterpret_cast<uint16_t*>(utf16_out);
  const char32_t* end = buf + len;

  uint16x4_t forbidden_bytemask = vmov_n_u16(0x0);

  while(buf + 4 <= end) {
    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(buf));

    // Check if no bits set above 16th
    if(vmaxvq_u32(in) <= 0xFFFF) {
      uint16x4_t utf16_packed = vmovn_u32(in);

      const uint16x4_t v_d800 = vmov_n_u16((uint16_t)0xd800);
      const uint16x4_t v_dfff = vmov_n_u16((uint16_t)0xdfff);
      forbidden_bytemask = vorr_u16(vand_u16(vcle_u16(utf16_packed, v_dfff), vcge_u16(utf16_packed, v_d800)), forbidden_bytemask);

      if (!match_system(big_endian)) {
        #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
        const uint8x8_t swap = make_uint8x8_t(1, 0, 3, 2, 5, 4, 7, 6);
        #else
        const uint8x8_t swap = {1, 0, 3, 2, 5, 4, 7, 6};
        #endif
        utf16_packed = vreinterpret_u16_u8(vtbl1_u8(vreinterpret_u8_u16(utf16_packed), swap));
      }
      vst1_u16(utf16_output, utf16_packed);
      utf16_output += 4;
      buf += 4;
    } else {
      size_t forward = 3;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint32_t word = buf[k];
        if((word & 0xFFFF0000)==0) {
          // will not generate a surrogate pair
          if (word >= 0xD800 && word <= 0xDFFF) { return std::make_pair(nullptr, reinterpret_cast<char16_t*>(utf16_output)); }
          *utf16_output++ = !match_system(big_endian) ? char16_t(word >> 8 | word << 8) : char16_t(word);
        } else {
          // will generate a surrogate pair
          if (word > 0x10FFFF) { return std::make_pair(nullptr, reinterpret_cast<char16_t*>(utf16_output)); }
          word -= 0x10000;
          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
          if (!match_system(big_endian)) {
            high_surrogate = uint16_t(high_surrogate >> 8 | high_surrogate << 8);
            low_surrogate = uint16_t(low_surrogate << 8 | low_surrogate >> 8);
          }
          *utf16_output++ = char16_t(high_surrogate);
          *utf16_output++ = char16_t(low_surrogate);
        }
      }
      buf += k;
    }
  }

  // check for invalid input
  if (vmaxv_u16(forbidden_bytemask) != 0) {
    return std::make_pair(nullptr, reinterpret_cast<char16_t*>(utf16_output));
  }

  return std::make_pair(buf, reinterpret_cast<char16_t*>(utf16_output));
}


template <endianness big_endian>
std::pair<result, char16_t*> arm_convert_utf32_to_utf16_with_errors(const char32_t* buf, size_t len, char16_t* utf16_out) {
  uint16_t * utf16_output = reinterpret_cast<uint16_t*>(utf16_out);
  const char32_t* start = buf;
  const char32_t* end = buf + len;

  while(buf + 4 <= end) {
    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(buf));

    // Check if no bits set above 16th
    if(vmaxvq_u32(in) <= 0xFFFF) {
      uint16x4_t utf16_packed = vmovn_u32(in);

      const uint16x4_t v_d800 = vmov_n_u16((uint16_t)0xd800);
      const uint16x4_t v_dfff = vmov_n_u16((uint16_t)0xdfff);
      const uint16x4_t forbidden_bytemask = vand_u16(vcle_u16(utf16_packed, v_dfff), vcge_u16(utf16_packed, v_d800));
      if (vmaxv_u16(forbidden_bytemask) != 0) {
        return std::make_pair(result(error_code::SURROGATE, buf - start), reinterpret_cast<char16_t*>(utf16_output));
      }

      if (!match_system(big_endian)) {
        #ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
        const uint8x8_t swap = make_uint8x8_t(1, 0, 3, 2, 5, 4, 7, 6);
        #else
        const uint8x8_t swap = {1, 0, 3, 2, 5, 4, 7, 6};
        #endif
        utf16_packed = vreinterpret_u16_u8(vtbl1_u8(vreinterpret_u8_u16(utf16_packed), swap));
      }
      vst1_u16(utf16_output, utf16_packed);
      utf16_output += 4;
      buf += 4;
    } else {
      size_t forward = 3;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint32_t word = buf[k];
        if((word & 0xFFFF0000)==0) {
          // will not generate a surrogate pair
          if (word >= 0xD800 && word <= 0xDFFF) { return std::make_pair(result(error_code::SURROGATE, buf - start + k), reinterpret_cast<char16_t*>(utf16_output)); }
          *utf16_output++ = !match_system(big_endian) ? char16_t(word >> 8 | word << 8) : char16_t(word);
        } else {
          // will generate a surrogate pair
          if (word > 0x10FFFF) { return std::make_pair(result(error_code::TOO_LARGE, buf - start + k), reinterpret_cast<char16_t*>(utf16_output)); }
          word -= 0x10000;
          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
          if (!match_system(big_endian)) {
            high_surrogate = uint16_t(high_surrogate >> 8 | high_surrogate << 8);
            low_surrogate = uint16_t(low_surrogate << 8 | low_surrogate >> 8);
          }
          *utf16_output++ = char16_t(high_surrogate);
          *utf16_output++ = char16_t(low_surrogate);
        }
      }
      buf += k;
    }
  }

  return std::make_pair(result(error_code::SUCCESS, buf - start), reinterpret_cast<char16_t*>(utf16_output));
}
/* end file src/arm64/arm_convert_utf32_to_utf16.cpp */
} // unnamed namespace
} // namespace arm64
} // namespace simdutf
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/buf_block_reader.h
/* begin file src/generic/buf_block_reader.h */
namespace simdutf {
namespace arm64 {
namespace {

// Walks through a buffer in block-sized increments, loading the last part with spaces
template<size_t STEP_SIZE>
struct buf_block_reader {
public:
  simdutf_really_inline buf_block_reader(const uint8_t *_buf, size_t _len);
  simdutf_really_inline size_t block_index();
  simdutf_really_inline bool has_full_block() const;
  simdutf_really_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.
   */
  simdutf_really_inline size_t get_remainder(uint8_t *dst) const;
  simdutf_really_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
simdutf_unused static char * format_input_text_64(const uint8_t *text) {
  static char *buf = reinterpret_cast<char*>(malloc(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
simdutf_unused static char * format_input_text(const simd8x64<uint8_t>& in) {
  static char *buf = reinterpret_cast<char*>(malloc(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;
}

simdutf_unused static char * format_mask(uint64_t mask) {
  static char *buf = reinterpret_cast<char*>(malloc(64 + 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>
simdutf_really_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>
simdutf_really_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }

template<size_t STEP_SIZE>
simdutf_really_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
  return idx < lenminusstep;
}

template<size_t STEP_SIZE>
simdutf_really_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
  return &buf[idx];
}

template<size_t STEP_SIZE>
simdutf_really_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>
simdutf_really_inline void buf_block_reader<STEP_SIZE>::advance() {
  idx += STEP_SIZE;
}

} // unnamed namespace
} // namespace arm64
} // namespace simdutf
/* end file src/generic/buf_block_reader.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_validation/utf8_lookup4_algorithm.h
/* begin file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
namespace simdutf {
namespace arm64 {
namespace {
namespace utf8_validation {

using namespace simd;

  simdutf_really_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);
  }
  simdutf_really_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 = simd8<uint8_t>(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.
  //
  simdutf_really_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______
    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, 0b11110000u-1, 0b11100000u-1, 0b11000000u-1
    };
    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.
    //
    simdutf_really_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.
    simdutf_really_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;
    }

    simdutf_really_inline void check_next_input(const simd8x64<uint8_t>& input) {
      if(simdutf_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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
            "We support either two or four chunks per 64-byte block.");
        if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
          this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
        } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
          this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
        }
        this->prev_incomplete = is_incomplete(input.chunks[simd8x64<uint8_t>::NUM_CHUNKS-1]);
        this->prev_input_block = input.chunks[simd8x64<uint8_t>::NUM_CHUNKS-1];

      }
    }

    // do not forget to call check_eof!
    simdutf_really_inline bool errors() const {
      return this->error.any_bits_set_anywhere();
    }

  }; // struct utf8_checker
} // namespace utf8_validation

using utf8_validation::utf8_checker;

} // unnamed namespace
} // namespace arm64
} // namespace simdutf
/* end file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_validation/utf8_validator.h
/* begin file src/generic/utf8_validation/utf8_validator.h */
namespace simdutf {
namespace arm64 {
namespace {
namespace utf8_validation {

/**
 * 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();
}

bool generic_validate_utf8(const char * input, size_t length) {
  return generic_validate_utf8<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
}

/**
 * Validates that the string is actual UTF-8 and stops on errors.
 */
template<class checker>
result generic_validate_utf8_with_errors(const uint8_t * input, size_t length) {
    checker c{};
    buf_block_reader<64> reader(input, length);
    size_t count{0};
    while (reader.has_full_block()) {
      simd::simd8x64<uint8_t> in(reader.full_block());
      c.check_next_input(in);
      if(c.errors()) {
        if (count != 0) { count--; } // Sometimes the error is only detected in the next chunk
        result res = scalar::utf8::rewind_and_validate_with_errors(reinterpret_cast<const char*>(input + count), length - count);
        res.count += count;
        return res;
      }
      reader.advance();
      count += 64;
    }
    uint8_t block[64]{};
    reader.get_remainder(block);
    simd::simd8x64<uint8_t> in(block);
    c.check_next_input(in);
    reader.advance();
    c.check_eof();
    if (c.errors()) {
      result res = scalar::utf8::rewind_and_validate_with_errors(reinterpret_cast<const char*>(input) + count, length - count);
      res.count += count;
      return res;
    } else {
      return result(error_code::SUCCESS, length);
    }
}

result generic_validate_utf8_with_errors(const char * input, size_t length) {
  return generic_validate_utf8_with_errors<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
}

template<class checker>
bool generic_validate_ascii(const uint8_t * input, size_t length) {
    buf_block_reader<64> reader(input, length);
    uint8_t blocks[64]{};
    simd::simd8x64<uint8_t> running_or(blocks);
    while (reader.has_full_block()) {
      simd::simd8x64<uint8_t> in(reader.full_block());
      running_or |= in;
      reader.advance();
    }
    uint8_t block[64]{};
    reader.get_remainder(block);
    simd::simd8x64<uint8_t> in(block);
    running_or |= in;
    return running_or.is_ascii();
}

bool generic_validate_ascii(const char * input, size_t length) {
  return generic_validate_ascii<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
}

template<class checker>
result generic_validate_ascii_with_errors(const uint8_t * input, size_t length) {
  buf_block_reader<64> reader(input, length);
  size_t count{0};
  while (reader.has_full_block()) {
    simd::simd8x64<uint8_t> in(reader.full_block());
    if (!in.is_ascii()) {
      result res = scalar::ascii::validate_with_errors(reinterpret_cast<const char*>(input + count), length - count);
      return result(res.error, count + res.count);
    }
    reader.advance();

    count += 64;
  }
  uint8_t block[64]{};
  reader.get_remainder(block);
  simd::simd8x64<uint8_t> in(block);
  if (!in.is_ascii()) {
    result res = scalar::ascii::validate_with_errors(reinterpret_cast<const char*>(input + count), length - count);
    return result(res.error, count + res.count);
  } else {
    return result(error_code::SUCCESS, length);
  }
}

result generic_validate_ascii_with_errors(const char * input, size_t length) {
  return generic_validate_ascii_with_errors<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
}

} // namespace utf8_validation
} // unnamed namespace
} // namespace arm64
} // namespace simdutf
/* end file src/generic/utf8_validation/utf8_validator.h */
// transcoding from UTF-8 to UTF-16
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_to_utf16/valid_utf8_to_utf16.h
/* begin file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */


namespace simdutf {
namespace arm64 {
namespace {
namespace utf8_to_utf16 {

using namespace simd;

template <endianness endian>
simdutf_warn_unused size_t convert_valid(const char* input, size_t size,
    char16_t* utf16_output) noexcept {
  // The implementation is not specific to haswell and should be moved to the generic directory.
  size_t pos = 0;
  char16_t* start{utf16_output};
  const size_t safety_margin = 16; // to avoid overruns!
  while(pos + 64 + safety_margin <= size) {
    // this loop could be unrolled further. For example, we could process the mask
    // far more than 64 bytes.
    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
    if(in.is_ascii()) {
      in.store_ascii_as_utf16<endian>(utf16_output);
      utf16_output += 64;
      pos += 64;
    } else {
      // Slow path. We hope that the compiler will recognize that this is a slow path.
      // Anything that is not a continuation mask is a 'leading byte', that is, the
      // start of a new code point.
      uint64_t utf8_continuation_mask = in.lt(-65 + 1);
      // -65 is 0b10111111 in two-complement's, so largest possible continuation byte
      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
      // The *start* of code points is not so useful, rather, we want the *end* of code points.
      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
      // We process in blocks of up to 12 bytes except possibly
      // for fast paths which may process up to 16 bytes. For the
      // slow path to work, we should have at least 12 input bytes left.
      size_t max_starting_point = (pos + 64) - 12;
      // Next loop is going to run at least five times when using solely
      // the slow/regular path, and at least four times if there are fast paths.
      while(pos < max_starting_point) {
        // Performance note: our ability to compute 'consumed' and
        // then shift and recompute is critical. If there is a
        // latency of, say, 4 cycles on getting 'consumed', then
        // the inner loop might have a total latency of about 6 cycles.
        // Yet we process between 6 to 12 inputs bytes, thus we get
        // a speed limit between 1 cycle/byte and 0.5 cycle/byte
        // for this section of the code. Hence, there is a limit
        // to how much we can further increase this latency before
        // it seriously harms performance.
        //
        // Thus we may allow convert_masked_utf8_to_utf16 to process
        // more bytes at a time under a fast-path mode where 16 bytes
        // are consumed at once (e.g., when encountering ASCII).
        size_t consumed = convert_masked_utf8_to_utf16<endian>(input + pos,
                            utf8_end_of_code_point_mask, utf16_output);
        pos += consumed;
        utf8_end_of_code_point_mask >>= consumed;
      }
      // At this point there may remain between 0 and 12 bytes in the
      // 64-byte block. These bytes will be processed again. So we have an
      // 80% efficiency (in the worst case). In practice we expect an
      // 85% to 90% efficiency.
    }
  }
  utf16_output += scalar::utf8_to_utf16::convert_valid<endian>(input + pos, size - pos, utf16_output);
  return utf16_output - start;
}

} // namespace utf8_to_utf16
} // unnamed namespace
} // namespace arm64
} // namespace simdutf
/* end file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_to_utf16/utf8_to_utf16.h
/* begin file src/generic/utf8_to_utf16/utf8_to_utf16.h */


namespace simdutf {
namespace arm64 {
namespace {
namespace utf8_to_utf16 {
using namespace simd;


  simdutf_really_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);
  }
  simdutf_really_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 = simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
    simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
    return must23_80 ^ sc;
  }


  struct validating_transcoder {
    // If this is nonzero, there has been a UTF-8 error.
    simd8<uint8_t> error;

    validating_transcoder() : error(uint8_t(0)) {}
    //
    // Check whether the current bytes are valid UTF-8.
    //
    simdutf_really_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);
    }


    template <endianness endian>
    simdutf_really_inline size_t convert(const char* in, size_t size, char16_t* utf16_output) {
      size_t pos = 0;
      char16_t* start{utf16_output};
      // In the worst case, we have the haswell kernel which can cause an overflow of
      // 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the last 16 bytes,
      // and if the data is valid, then it is entirely safe because 16 UTF-8 bytes generate
      // much more than 8 bytes. However, you cannot generally assume that you have valid
      // UTF-8 input, so we are going to go back from the end counting 8 leading bytes,
      // to give us a good margin.
      size_t leading_byte = 0;
      size_t margin = size;
      for(; margin > 0 && leading_byte < 8; margin--) {
        leading_byte += (int8_t(in[margin-1]) > -65);
      }
      // If the input is long enough, then we have that margin-1 is the eight last leading byte.
      const size_t safety_margin = size - margin + 1; // to avoid overruns!
      while(pos + 64 + safety_margin <= size) {
        simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
        if(input.is_ascii()) {
          input.store_ascii_as_utf16<endian>(utf16_output);
          utf16_output += 64;
          pos += 64;
        } 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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
              "We support either two or four chunks per 64-byte block.");
          auto zero = simd8<uint8_t>{uint8_t(0)};
          if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
            this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
            this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
          }
          uint64_t utf8_continuation_mask = input.lt(-65 + 1);
          uint64_t utf8_leading_mask = ~utf8_continuation_mask;
          uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
          // We process in blocks of up to 12 bytes except possibly
          // for fast paths which may process up to 16 bytes. For the
          // slow path to work, we should have at least 12 input bytes left.
          size_t max_starting_point = (pos + 64) - 12;
          // Next loop is going to run at least five times.
          while(pos < max_starting_point) {
            // Performance note: our ability to compute 'consumed' and
            // then shift and recompute is critical. If there is a
            // latency of, say, 4 cycles on getting 'consumed', then
            // the inner loop might have a total latency of about 6 cycles.
            // Yet we process between 6 to 12 inputs bytes, thus we get
            // a speed limit between 1 cycle/byte and 0.5 cycle/byte
            // for this section of the code. Hence, there is a limit
            // to how much we can further increase this latency before
            // it seriously harms performance.
            size_t consumed = convert_masked_utf8_to_utf16<endian>(in + pos,
                            utf8_end_of_code_point_mask, utf16_output);
            pos += consumed;
            utf8_end_of_code_point_mask >>= consumed;
          }
          // At this point there may remain between 0 and 12 bytes in the
          // 64-byte block. These bytes will be processed again. So we have an
          // 80% efficiency (in the worst case). In practice we expect an
          // 85% to 90% efficiency.
        }
      }
      if(errors()) { return 0; }
      if(pos < size) {
        size_t howmany  = scalar::utf8_to_utf16::convert<endian>(in + pos, size - pos, utf16_output);
        if(howmany == 0) { return 0; }
        utf16_output += howmany;
      }
      return utf16_output - start;
    }

    template <endianness endian>
    simdutf_really_inline result convert_with_errors(const char* in, size_t size, char16_t* utf16_output) {
      size_t pos = 0;
      char16_t* start{utf16_output};
      // In the worst case, we have the haswell kernel which can cause an overflow of
      // 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the last 16 bytes,
      // and if the data is valid, then it is entirely safe because 16 UTF-8 bytes generate
      // much more than 8 bytes. However, you cannot generally assume that you have valid
      // UTF-8 input, so we are going to go back from the end counting 8 leading bytes,
      // to give us a good margin.
      size_t leading_byte = 0;
      size_t margin = size;
      for(; margin > 0 && leading_byte < 8; margin--) {
        leading_byte += (int8_t(in[margin-1]) > -65);
      }
      // If the input is long enough, then we have that margin-1 is the eight last leading byte.
      const size_t safety_margin = size - margin + 1; // to avoid overruns!
      while(pos + 64 + safety_margin <= size) {
        simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
        if(input.is_ascii()) {
          input.store_ascii_as_utf16<endian>(utf16_output);
          utf16_output += 64;
          pos += 64;
        } 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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
              "We support either two or four chunks per 64-byte block.");
          auto zero = simd8<uint8_t>{uint8_t(0)};
          if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
            this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
            this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
          }
          if (errors()) {
            // rewind_and_convert_with_errors will seek a potential error from in+pos onward,
            // with the ability to go back up to pos bytes, and read size-pos bytes forward.
            result res = scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(pos, in + pos, size - pos, utf16_output);
            res.count += pos;
            return res;
          }
          uint64_t utf8_continuation_mask = input.lt(-65 + 1);
          uint64_t utf8_leading_mask = ~utf8_continuation_mask;
          uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
          // We process in blocks of up to 12 bytes except possibly
          // for fast paths which may process up to 16 bytes. For the
          // slow path to work, we should have at least 12 input bytes left.
          size_t max_starting_point = (pos + 64) - 12;
          // Next loop is going to run at least five times.
          while(pos < max_starting_point) {
            // Performance note: our ability to compute 'consumed' and
            // then shift and recompute is critical. If there is a
            // latency of, say, 4 cycles on getting 'consumed', then
            // the inner loop might have a total latency of about 6 cycles.
            // Yet we process between 6 to 12 inputs bytes, thus we get
            // a speed limit between 1 cycle/byte and 0.5 cycle/byte
            // for this section of the code. Hence, there is a limit
            // to how much we can further increase this latency before
            // it seriously harms performance.
            size_t consumed = convert_masked_utf8_to_utf16<endian>(in + pos,
                            utf8_end_of_code_point_mask, utf16_output);
            pos += consumed;
            utf8_end_of_code_point_mask >>= consumed;
          }
          // At this point there may remain between 0 and 12 bytes in the
          // 64-byte block. These bytes will be processed again. So we have an
          // 80% efficiency (in the worst case). In practice we expect an
          // 85% to 90% efficiency.
        }
      }
      if(errors()) {
        // rewind_and_convert_with_errors will seek a potential error from in+pos onward,
        // with the ability to go back up to pos bytes, and read size-pos bytes forward.
        result res = scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(pos, in + pos, size - pos, utf16_output);
        res.count += pos;
        return res;
      }
      if(pos < size) {
        // rewind_and_convert_with_errors will seek a potential error from in+pos onward,
        // with the ability to go back up to pos bytes, and read size-pos bytes forward.
        result res = scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(pos, in + pos, size - pos, utf16_output);
        if (res.error) {    // In case of error, we want the error position
          res.count += pos;
          return res;
        } else {    // In case of success, we want the number of word written
          utf16_output += res.count;
        }
      }
      return result(error_code::SUCCESS, utf16_output - start);
    }

    simdutf_really_inline bool errors() const {
      return this->error.any_bits_set_anywhere();
    }

  }; // struct utf8_checker
} // utf8_to_utf16 namespace
} // unnamed namespace
} // namespace arm64
} // namespace simdutf
/* end file src/generic/utf8_to_utf16/utf8_to_utf16.h */
// transcoding from UTF-8 to UTF-32
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_to_utf32/valid_utf8_to_utf32.h
/* begin file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */

namespace simdutf {
namespace arm64 {
namespace {
namespace utf8_to_utf32 {

using namespace simd;


simdutf_warn_unused size_t convert_valid(const char* input, size_t size,
    char32_t* utf32_output) noexcept {
  size_t pos = 0;
  char32_t* start{utf32_output};
  const size_t safety_margin = 16; // to avoid overruns!
  while(pos + 64 + safety_margin <= size) {
    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
    if(in.is_ascii()) {
      in.store_ascii_as_utf32(utf32_output);
      utf32_output += 64;
      pos += 64;
    } else {
    // -65 is 0b10111111 in two-complement's, so largest possible continuation byte
    uint64_t utf8_continuation_mask = in.lt(-65 + 1);
    uint64_t utf8_leading_mask = ~utf8_continuation_mask;
    uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
    size_t max_starting_point = (pos + 64) - 12;
    while(pos < max_starting_point) {
      size_t consumed = convert_masked_utf8_to_utf32(input + pos,
                          utf8_end_of_code_point_mask, utf32_output);
      pos += consumed;
      utf8_end_of_code_point_mask >>= consumed;
      }
    }
  }
  utf32_output += scalar::utf8_to_utf32::convert_valid(input + pos, size - pos, utf32_output);
  return utf32_output - start;
}


} // namespace utf8_to_utf32
} // unnamed namespace
} // namespace arm64
} // namespace simdutf
/* end file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_to_utf32/utf8_to_utf32.h
/* begin file src/generic/utf8_to_utf32/utf8_to_utf32.h */


namespace simdutf {
namespace arm64 {
namespace {
namespace utf8_to_utf32 {
using namespace simd;


  simdutf_really_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);
  }
  simdutf_really_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 = simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
    simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
    return must23_80 ^ sc;
  }


  struct validating_transcoder {
    // If this is nonzero, there has been a UTF-8 error.
    simd8<uint8_t> error;

    validating_transcoder() : error(uint8_t(0)) {}
    //
    // Check whether the current bytes are valid UTF-8.
    //
    simdutf_really_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);
    }



    simdutf_really_inline size_t convert(const char* in, size_t size, char32_t* utf32_output) {
      size_t pos = 0;
      char32_t* start{utf32_output};
      // In the worst case, we have the haswell kernel which can cause an overflow of
      // 8 bytes when calling convert_masked_utf8_to_utf32. If you skip the last 16 bytes,
      // and if the data is valid, then it is entirely safe because 16 UTF-8 bytes generate
      // much more than 8 bytes. However, you cannot generally assume that you have valid
      // UTF-8 input, so we are going to go back from the end counting 4 leading bytes,
      // to give us a good margin.
      size_t leading_byte = 0;
      size_t margin = size;
      for(; margin > 0 && leading_byte < 4; margin--) {
        leading_byte += (int8_t(in[margin-1]) > -65);
      }
      // If the input is long enough, then we have that margin-1 is the fourth last leading byte.
      const size_t safety_margin = size - margin + 1; // to avoid overruns!
      while(pos + 64 + safety_margin <= size) {
        simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
        if(input.is_ascii()) {
          input.store_ascii_as_utf32(utf32_output);
          utf32_output += 64;
          pos += 64;
        } 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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
              "We support either two or four chunks per 64-byte block.");
          auto zero = simd8<uint8_t>{uint8_t(0)};
          if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
            this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
            this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
          }
          uint64_t utf8_continuation_mask = input.lt(-65 + 1);
          uint64_t utf8_leading_mask = ~utf8_continuation_mask;
          uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
          // We process in blocks of up to 12 bytes except possibly
          // for fast paths which may process up to 16 bytes. For the
          // slow path to work, we should have at least 12 input bytes left.
          size_t max_starting_point = (pos + 64) - 12;
          // Next loop is going to run at least five times.
          while(pos < max_starting_point) {
            // Performance note: our ability to compute 'consumed' and
            // then shift and recompute is critical. If there is a
            // latency of, say, 4 cycles on getting 'consumed', then
            // the inner loop might have a total latency of about 6 cycles.
            // Yet we process between 6 to 12 inputs bytes, thus we get
            // a speed limit between 1 cycle/byte and 0.5 cycle/byte
            // for this section of the code. Hence, there is a limit
            // to how much we can further increase this latency before
            // it seriously harms performance.
            size_t consumed = convert_masked_utf8_to_utf32(in + pos,
                            utf8_end_of_code_point_mask, utf32_output);
            pos += consumed;
            utf8_end_of_code_point_mask >>= consumed;
          }
          // At this point there may remain between 0 and 12 bytes in the
          // 64-byte block. These bytes will be processed again. So we have an
          // 80% efficiency (in the worst case). In practice we expect an
          // 85% to 90% efficiency.
        }
      }
      if(errors()) { return 0; }
      if(pos < size) {
        size_t howmany  = scalar::utf8_to_utf32::convert(in + pos, size - pos, utf32_output);
        if(howmany == 0) { return 0; }
        utf32_output += howmany;
      }
      return utf32_output - start;
    }

    simdutf_really_inline result convert_with_errors(const char* in, size_t size, char32_t* utf32_output) {
      size_t pos = 0;
      char32_t* start{utf32_output};
      // In the worst case, we have the haswell kernel which can cause an overflow of
      // 8 bytes when calling convert_masked_utf8_to_utf32. If you skip the last 16 bytes,
      // and if the data is valid, then it is entirely safe because 16 UTF-8 bytes generate
      // much more than 8 bytes. However, you cannot generally assume that you have valid
      // UTF-8 input, so we are going to go back from the end counting 4 leading bytes,
      // to give us a good margin.
      size_t leading_byte = 0;
      size_t margin = size;
      for(; margin > 0 && leading_byte < 4; margin--) {
        leading_byte += (int8_t(in[margin-1]) > -65);
      }
      // If the input is long enough, then we have that margin-1 is the fourth last leading byte.
      const size_t safety_margin = size - margin + 1; // to avoid overruns!
      while(pos + 64 + safety_margin <= size) {
        simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
        if(input.is_ascii()) {
          input.store_ascii_as_utf32(utf32_output);
          utf32_output += 64;
          pos += 64;
        } 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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
              "We support either two or four chunks per 64-byte block.");
          auto zero = simd8<uint8_t>{uint8_t(0)};
          if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
            this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
            this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
          }
          if (errors()) {
            result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(pos, in + pos, size - pos, utf32_output);
            res.count += pos;
            return res;
          }
          uint64_t utf8_continuation_mask = input.lt(-65 + 1);
          uint64_t utf8_leading_mask = ~utf8_continuation_mask;
          uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
          // We process in blocks of up to 12 bytes except possibly
          // for fast paths which may process up to 16 bytes. For the
          // slow path to work, we should have at least 12 input bytes left.
          size_t max_starting_point = (pos + 64) - 12;
          // Next loop is going to run at least five times.
          while(pos < max_starting_point) {
            // Performance note: our ability to compute 'consumed' and
            // then shift and recompute is critical. If there is a
            // latency of, say, 4 cycles on getting 'consumed', then
            // the inner loop might have a total latency of about 6 cycles.
            // Yet we process between 6 to 12 inputs bytes, thus we get
            // a speed limit between 1 cycle/byte and 0.5 cycle/byte
            // for this section of the code. Hence, there is a limit
            // to how much we can further increase this latency before
            // it seriously harms performance.
            size_t consumed = convert_masked_utf8_to_utf32(in + pos,
                            utf8_end_of_code_point_mask, utf32_output);
            pos += consumed;
            utf8_end_of_code_point_mask >>= consumed;
          }
          // At this point there may remain between 0 and 12 bytes in the
          // 64-byte block. These bytes will be processed again. So we have an
          // 80% efficiency (in the worst case). In practice we expect an
          // 85% to 90% efficiency.
        }
      }
      if(errors()) {
        result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(pos, in + pos, size - pos, utf32_output);
        res.count += pos;
        return res;
      }
      if(pos < size) {
        result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(pos, in + pos, size - pos, utf32_output);
        if (res.error) {    // In case of error, we want the error position
          res.count += pos;
          return res;
        } else {    // In case of success, we want the number of word written
          utf32_output += res.count;
        }
      }
      return result(error_code::SUCCESS, utf32_output - start);
    }

    simdutf_really_inline bool errors() const {
      return this->error.any_bits_set_anywhere();
    }

  }; // struct utf8_checker
} // utf8_to_utf32 namespace
} // unnamed namespace
} // namespace arm64
} // namespace simdutf
/* end file src/generic/utf8_to_utf32/utf8_to_utf32.h */
// other functions
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8.h
/* begin file src/generic/utf8.h */

namespace simdutf {
namespace arm64 {
namespace {
namespace utf8 {

using namespace simd;

simdutf_really_inline size_t count_code_points(const char* in, size_t size) {
    size_t pos = 0;
    size_t count = 0;
    for(;pos + 64 <= size; pos += 64) {
      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
      uint64_t utf8_continuation_mask = input.lt(-65 + 1);
      count += 64 - count_ones(utf8_continuation_mask);
    }
    return count + scalar::utf8::count_code_points(in + pos, size - pos);
}


simdutf_really_inline size_t utf16_length_from_utf8(const char* in, size_t size) {
    size_t pos = 0;
    size_t count = 0;
    // This algorithm could no doubt be improved!
    for(;pos + 64 <= size; pos += 64) {
      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
      uint64_t utf8_continuation_mask = input.lt(-65 + 1);
      // We count one word for anything that is not a continuation (so
      // leading bytes).
      count += 64 - count_ones(utf8_continuation_mask);
      int64_t utf8_4byte = input.gteq_unsigned(240);
      count += count_ones(utf8_4byte);
    }
    return count + scalar::utf8::utf16_length_from_utf8(in + pos, size - pos);
}


simdutf_really_inline size_t utf32_length_from_utf8(const char* in, size_t size) {
    return count_code_points(in, size);
}
} // utf8 namespace
} // unnamed namespace
} // namespace arm64
} // namespace simdutf
/* end file src/generic/utf8.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf16.h
/* begin file src/generic/utf16.h */
namespace simdutf {
namespace arm64 {
namespace {
namespace utf16 {

template <endianness big_endian>
simdutf_really_inline size_t count_code_points(const char16_t* in, size_t size) {
    size_t pos = 0;
    size_t count = 0;
    for(;pos + 32 <= size; pos += 32) {
      simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
      if (!match_system(big_endian)) input.swap_bytes();
      uint64_t not_pair = input.not_in_range(0xDC00, 0xDFFF);
      count += count_ones(not_pair) / 2;
    }
    return count + scalar::utf16::count_code_points<big_endian>(in + pos, size - pos);
}

template <endianness big_endian>
simdutf_really_inline size_t utf8_length_from_utf16(const char16_t* in, size_t size) {
    size_t pos = 0;
    size_t count = 0;
    // This algorithm could no doubt be improved!
    for(;pos + 32 <= size; pos += 32) {
      simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
      if (!match_system(big_endian)) input.swap_bytes();
      uint64_t ascii_mask = input.lteq(0x7F);
      uint64_t twobyte_mask = input.lteq(0x7FF);
      uint64_t not_pair_mask = input.not_in_range(0xD800, 0xDFFF);

      size_t ascii_count = count_ones(ascii_mask) / 2;
      size_t twobyte_count = count_ones(twobyte_mask & ~ ascii_mask) / 2;
      size_t threebyte_count = count_ones(not_pair_mask & ~ twobyte_mask) / 2;
      size_t fourbyte_count = 32 - count_ones(not_pair_mask) / 2;
      count += 2 * fourbyte_count + 3 * threebyte_count + 2 * twobyte_count + ascii_count;
    }
    return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos, size - pos);
}

template <endianness big_endian>
simdutf_really_inline size_t utf32_length_from_utf16(const char16_t* in, size_t size) {
    return count_code_points<big_endian>(in, size);
}

simdutf_really_inline void change_endianness_utf16(const char16_t* in, size_t size, char16_t* output) {
  size_t pos = 0;

  while (pos + 32 <= size) {
    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
    input.swap_bytes();
    input.store(reinterpret_cast<uint16_t *>(output));
    pos += 32;
    output += 32;
  }

  scalar::utf16::change_endianness_utf16(in + pos, size - pos, output);
}

} // utf16
} // unnamed namespace
} // namespace arm64
} // namespace simdutf
/* end file src/generic/utf16.h */
//
// Implementation-specific overrides
//
namespace simdutf {
namespace arm64 {

simdutf_warn_unused int implementation::detect_encodings(const char * input, size_t length) const noexcept {
  // If there is a BOM, then we trust it.
  auto bom_encoding = simdutf::BOM::check_bom(input, length);
  if(bom_encoding != encoding_type::unspecified) { return bom_encoding; }
  if (length % 2 == 0) {
    return arm_detect_encodings<utf8_validation::utf8_checker>(input, length);
  } else {
    if (implementation::validate_utf8(input, length)) {
      return simdutf::encoding_type::UTF8;
    } else {
      return simdutf::encoding_type::unspecified;
    }
  }
}

simdutf_warn_unused bool implementation::validate_utf8(const char *buf, size_t len) const noexcept {
  return arm64::utf8_validation::generic_validate_utf8(buf,len);
}

simdutf_warn_unused result implementation::validate_utf8_with_errors(const char *buf, size_t len) const noexcept {
  return arm64::utf8_validation::generic_validate_utf8_with_errors(buf,len);
}

simdutf_warn_unused bool implementation::validate_ascii(const char *buf, size_t len) const noexcept {
  return arm64::utf8_validation::generic_validate_ascii(buf,len);
}

simdutf_warn_unused result implementation::validate_ascii_with_errors(const char *buf, size_t len) const noexcept {
  return arm64::utf8_validation::generic_validate_ascii_with_errors(buf,len);
}

simdutf_warn_unused bool implementation::validate_utf16le(const char16_t *buf, size_t len) const noexcept {
  const char16_t* tail = arm_validate_utf16<endianness::LITTLE>(buf, len);
  if (tail) {
    return scalar::utf16::validate<endianness::LITTLE>(tail, len - (tail - buf));
  } else {
    return false;
  }
}

simdutf_warn_unused bool implementation::validate_utf16be(const char16_t *buf, size_t len) const noexcept {
  const char16_t* tail = arm_validate_utf16<endianness::BIG>(buf, len);
  if (tail) {
    return scalar::utf16::validate<endianness::BIG>(tail, len - (tail - buf));
  } else {
    return false;
  }
}

simdutf_warn_unused result implementation::validate_utf16le_with_errors(const char16_t *buf, size_t len) const noexcept {
  result res = arm_validate_utf16_with_errors<endianness::LITTLE>(buf, len);
  if (res.count != len) {
    result scalar_res = scalar::utf16::validate_with_errors<endianness::LITTLE>(buf + res.count, len - res.count);
    return result(scalar_res.error, res.count + scalar_res.count);
  } else {
    return res;
  }
}

simdutf_warn_unused result implementation::validate_utf16be_with_errors(const char16_t *buf, size_t len) const noexcept {
  result res = arm_validate_utf16_with_errors<endianness::BIG>(buf, len);
  if (res.count != len) {
    result scalar_res = scalar::utf16::validate_with_errors<endianness::BIG>(buf + res.count, len - res.count);
    return result(scalar_res.error, res.count + scalar_res.count);
  } else {
    return res;
  }
}

simdutf_warn_unused bool implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
  const char32_t* tail = arm_validate_utf32le(buf, len);
  if (tail) {
    return scalar::utf32::validate(tail, len - (tail - buf));
  } else {
    return false;
  }
}

simdutf_warn_unused result implementation::validate_utf32_with_errors(const char32_t *buf, size_t len) const noexcept {
  result res = arm_validate_utf32le_with_errors(buf, len);
  if (res.count != len) {
    result scalar_res = scalar::utf32::validate_with_errors(buf + res.count, len - res.count);
    return result(scalar_res.error, res.count + scalar_res.count);
  } else {
    return res;
  }
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
  utf8_to_utf16::validating_transcoder converter;
  return converter.convert<endianness::LITTLE>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf16be(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
  utf8_to_utf16::validating_transcoder converter;
  return converter.convert<endianness::BIG>(buf, len, utf16_output);
}

simdutf_warn_unused result implementation::convert_utf8_to_utf16le_with_errors(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
  utf8_to_utf16::validating_transcoder converter;
  return converter.convert_with_errors<endianness::LITTLE>(buf, len, utf16_output);
}

simdutf_warn_unused result implementation::convert_utf8_to_utf16be_with_errors(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
  utf8_to_utf16::validating_transcoder converter;
  return converter.convert_with_errors<endianness::BIG>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16le(const char* input, size_t size,
    char16_t* utf16_output) const noexcept {
  return utf8_to_utf16::convert_valid<endianness::LITTLE>(input, size,  utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(const char* input, size_t size,
    char16_t* utf16_output) const noexcept {
  return utf8_to_utf16::convert_valid<endianness::BIG>(input, size,  utf16_output);
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(const char* buf, size_t len, char32_t* utf32_output) const noexcept {
  utf8_to_utf32::validating_transcoder converter;
  return converter.convert(buf, len, utf32_output);
}

simdutf_warn_unused result implementation::convert_utf8_to_utf32_with_errors(const char* buf, size_t len, char32_t* utf32_output) const noexcept {
  utf8_to_utf32::validating_transcoder converter;
  return converter.convert_with_errors(buf, len, utf32_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(const char* input, size_t size,
    char32_t* utf32_output) const noexcept {
  return utf8_to_utf32::convert_valid(input, size,  utf32_output);
}

simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  std::pair<const char16_t*, char*> ret = arm_convert_utf16_to_utf8<endianness::LITTLE>(buf, len, utf8_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf8_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf16_to_utf8::convert<endianness::LITTLE>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused size_t implementation::convert_utf16be_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  std::pair<const char16_t*, char*> ret = arm_convert_utf16_to_utf8<endianness::BIG>(buf, len, utf8_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf8_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf16_to_utf8::convert<endianness::BIG>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused result implementation::convert_utf16le_to_utf8_with_errors(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char*> ret = arm_convert_utf16_to_utf8_with_errors<endianness::LITTLE>(buf, len, utf8_output);
  if (ret.first.error) { return ret.first; }  // Can return directly since scalar fallback already found correct ret.first.count
  if (ret.first.count != len) { // All good so far, but not finished
    result scalar_res = scalar::utf16_to_utf8::convert_with_errors<endianness::LITTLE>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf8_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused result implementation::convert_utf16be_to_utf8_with_errors(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char*> ret = arm_convert_utf16_to_utf8_with_errors<endianness::BIG>(buf, len, utf8_output);
  if (ret.first.error) { return ret.first; }  // Can return directly since scalar fallback already found correct ret.first.count
  if (ret.first.count != len) { // All good so far, but not finished
    result scalar_res = scalar::utf16_to_utf8::convert_with_errors<endianness::BIG>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf8_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return convert_utf16le_to_utf8(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return convert_utf16be_to_utf8(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  std::pair<const char32_t*, char*> ret = arm_convert_utf32_to_utf8(buf, len, utf8_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf8_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf32_to_utf8::convert(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char*> ret = arm_convert_utf32_to_utf8_with_errors(buf, len, utf8_output);
  if (ret.first.count != len) {
    result scalar_res = scalar::utf32_to_utf8::convert_with_errors(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf8_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  std::pair<const char16_t*, char32_t*> ret = arm_convert_utf16_to_utf32<endianness::LITTLE>(buf, len, utf32_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf32_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf16_to_utf32::convert<endianness::LITTLE>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused size_t implementation::convert_utf16be_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  std::pair<const char16_t*, char32_t*> ret = arm_convert_utf16_to_utf32<endianness::BIG>(buf, len, utf32_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf32_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf16_to_utf32::convert<endianness::BIG>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused result implementation::convert_utf16le_to_utf32_with_errors(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char32_t*> ret = arm_convert_utf16_to_utf32_with_errors<endianness::LITTLE>(buf, len, utf32_output);
  if (ret.first.error) { return ret.first; }  // Can return directly since scalar fallback already found correct ret.first.count
  if (ret.first.count != len) { // All good so far, but not finished
    result scalar_res = scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf32_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char32_t*> ret = arm_convert_utf16_to_utf32_with_errors<endianness::BIG>(buf, len, utf32_output);
  if (ret.first.error) { return ret.first; }  // Can return directly since scalar fallback already found correct ret.first.count
  if (ret.first.count != len) { // All good so far, but not finished
    result scalar_res = scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf32_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  return convert_utf32_to_utf8(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  std::pair<const char32_t*, char16_t*> ret = arm_convert_utf32_to_utf16<endianness::LITTLE>(buf, len, utf16_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf16_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf32_to_utf16::convert<endianness::LITTLE>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf16be(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  std::pair<const char32_t*, char16_t*> ret = arm_convert_utf32_to_utf16<endianness::BIG>(buf, len, utf16_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf16_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf32_to_utf16::convert<endianness::BIG>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused result implementation::convert_utf32_to_utf16le_with_errors(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char16_t*> ret = arm_convert_utf32_to_utf16_with_errors<endianness::LITTLE>(buf, len, utf16_output);
  if (ret.first.count != len) {
    result scalar_res = scalar::utf32_to_utf16::convert_with_errors<endianness::LITTLE>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf16_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused result implementation::convert_utf32_to_utf16be_with_errors(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char16_t*> ret = arm_convert_utf32_to_utf16_with_errors<endianness::BIG>(buf, len, utf16_output);
  if (ret.first.count != len) {
    result scalar_res = scalar::utf32_to_utf16::convert_with_errors<endianness::BIG>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf16_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16le(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return convert_utf32_to_utf16le(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16be(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return convert_utf32_to_utf16be(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return convert_utf16le_to_utf32(buf, len, utf32_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return convert_utf16be_to_utf32(buf, len, utf32_output);
}

void implementation::change_endianness_utf16(const char16_t * input, size_t length, char16_t * output) const noexcept {
  utf16::change_endianness_utf16(input, length, output);
}

simdutf_warn_unused size_t implementation::count_utf16le(const char16_t * input, size_t length) const noexcept {
  return utf16::count_code_points<endianness::LITTLE>(input, length);
}

simdutf_warn_unused size_t implementation::count_utf16be(const char16_t * input, size_t length) const noexcept {
  return utf16::count_code_points<endianness::BIG>(input, length);
}

simdutf_warn_unused size_t implementation::count_utf8(const char * input, size_t length) const noexcept {
  return utf8::count_code_points(input, length);
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(const char16_t * input, size_t length) const noexcept {
  return utf16::utf8_length_from_utf16<endianness::LITTLE>(input, length);
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(const char16_t * input, size_t length) const noexcept {
  return utf16::utf8_length_from_utf16<endianness::BIG>(input, length);
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(const char16_t * input, size_t length) const noexcept {
  return utf16::utf32_length_from_utf16<endianness::LITTLE>(input, length);
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(const char16_t * input, size_t length) const noexcept {
  return utf16::utf32_length_from_utf16<endianness::BIG>(input, length);
}

simdutf_warn_unused size_t implementation::utf16_length_from_utf8(const char * input, size_t length) const noexcept {
  return utf8::utf16_length_from_utf8(input, length);
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf32(const char32_t * input, size_t length) const noexcept {
  const uint32x4_t v_7f = vmovq_n_u32((uint32_t)0x7f);
  const uint32x4_t v_7ff = vmovq_n_u32((uint32_t)0x7ff);
  const uint32x4_t v_ffff = vmovq_n_u32((uint32_t)0xffff);
  const uint32x4_t v_1 = vmovq_n_u32((uint32_t)0x1);
  size_t pos = 0;
  size_t count = 0;
  for(;pos + 4 <= length; pos += 4) {
    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(input + pos));
    const uint32x4_t ascii_bytes_bytemask = vcleq_u32(in, v_7f);
    const uint32x4_t one_two_bytes_bytemask = vcleq_u32(in, v_7ff);
    const uint32x4_t two_bytes_bytemask = veorq_u32(one_two_bytes_bytemask, ascii_bytes_bytemask);
    const uint32x4_t three_bytes_bytemask = veorq_u32(vcleq_u32(in, v_ffff), one_two_bytes_bytemask);

    const uint16x8_t reduced_ascii_bytes_bytemask = vreinterpretq_u16_u32(vandq_u32(ascii_bytes_bytemask, v_1));
    const uint16x8_t reduced_two_bytes_bytemask = vreinterpretq_u16_u32(vandq_u32(two_bytes_bytemask, v_1));
    const uint16x8_t reduced_three_bytes_bytemask = vreinterpretq_u16_u32(vandq_u32(three_bytes_bytemask, v_1));

    const uint16x8_t compressed_bytemask0 = vpaddq_u16(reduced_ascii_bytes_bytemask, reduced_two_bytes_bytemask);
    const uint16x8_t compressed_bytemask1 = vpaddq_u16(reduced_three_bytes_bytemask, reduced_three_bytes_bytemask);

    size_t ascii_count = count_ones(vgetq_lane_u64(vreinterpretq_u64_u16(compressed_bytemask0), 0));
    size_t two_bytes_count = count_ones(vgetq_lane_u64(vreinterpretq_u64_u16(compressed_bytemask0), 1));
    size_t three_bytes_count = count_ones(vgetq_lane_u64(vreinterpretq_u64_u16(compressed_bytemask1), 0));

    count += 16 - 3*ascii_count - 2*two_bytes_count - three_bytes_count;
  }
  return count + scalar::utf32::utf8_length_from_utf32(input + pos, length - pos);
}

simdutf_warn_unused size_t implementation::utf16_length_from_utf32(const char32_t * input, size_t length) const noexcept {
  const uint32x4_t v_ffff = vmovq_n_u32((uint32_t)0xffff);
  const uint32x4_t v_1 = vmovq_n_u32((uint32_t)0x1);
  size_t pos = 0;
  size_t count = 0;
  for(;pos + 4 <= length; pos += 4) {
    uint32x4_t in = vld1q_u32(reinterpret_cast<const uint32_t *>(input + pos));
    const uint32x4_t surrogate_bytemask = vcgtq_u32(in, v_ffff);
    const uint16x8_t reduced_bytemask = vreinterpretq_u16_u32(vandq_u32(surrogate_bytemask, v_1));
    const uint16x8_t compressed_bytemask = vpaddq_u16(reduced_bytemask, reduced_bytemask);
    size_t surrogate_count = count_ones(vgetq_lane_u64(vreinterpretq_u64_u16(compressed_bytemask), 0));
    count += 4 + surrogate_count;
  }
  return count + scalar::utf32::utf16_length_from_utf32(input + pos, length - pos);
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf8(const char * input, size_t length) const noexcept {
  return utf8::utf32_length_from_utf8(input, length);
}

} // namespace arm64
} // namespace simdutf

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/arm64/end.h
/* begin file src/simdutf/arm64/end.h */
/* end file src/simdutf/arm64/end.h */
/* end file src/arm64/implementation.cpp */
#endif
#if SIMDUTF_IMPLEMENTATION_FALLBACK
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=fallback/implementation.cpp
/* begin file src/fallback/implementation.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/fallback/begin.h
/* begin file src/simdutf/fallback/begin.h */
// redefining SIMDUTF_IMPLEMENTATION to "fallback"
// #define SIMDUTF_IMPLEMENTATION fallback
/* end file src/simdutf/fallback/begin.h */








namespace simdutf {
namespace fallback {

simdutf_warn_unused int implementation::detect_encodings(const char * input, size_t length) const noexcept {
  // If there is a BOM, then we trust it.
  auto bom_encoding = simdutf::BOM::check_bom(input, length);
  if(bom_encoding != encoding_type::unspecified) { return bom_encoding; }
  int out = 0;
  if(validate_utf8(input, length)) { out |= encoding_type::UTF8; }
  if((length % 2) == 0) {
    if(validate_utf16le(reinterpret_cast<const char16_t*>(input), length/2)) { out |= encoding_type::UTF16_LE; }
  }
  if((length % 4) == 0) {
    if(validate_utf32(reinterpret_cast<const char32_t*>(input), length/4)) { out |= encoding_type::UTF32_LE; }
  }

  return out;
}

simdutf_warn_unused bool implementation::validate_utf8(const char *buf, size_t len) const noexcept {
    return scalar::utf8::validate(buf, len);
}

simdutf_warn_unused result implementation::validate_utf8_with_errors(const char *buf, size_t len) const noexcept {
    return scalar::utf8::validate_with_errors(buf, len);
}

simdutf_warn_unused bool implementation::validate_ascii(const char *buf, size_t len) const noexcept {
    return scalar::ascii::validate(buf, len);
}

simdutf_warn_unused result implementation::validate_ascii_with_errors(const char *buf, size_t len) const noexcept {
    return scalar::ascii::validate_with_errors(buf, len);
}

simdutf_warn_unused bool implementation::validate_utf16le(const char16_t *buf, size_t len) const noexcept {
    return scalar::utf16::validate<endianness::LITTLE>(buf, len);
}

simdutf_warn_unused bool implementation::validate_utf16be(const char16_t *buf, size_t len) const noexcept {
    return scalar::utf16::validate<endianness::BIG>(buf, len);
}

simdutf_warn_unused result implementation::validate_utf16le_with_errors(const char16_t *buf, size_t len) const noexcept {
    return scalar::utf16::validate_with_errors<endianness::LITTLE>(buf, len);
}

simdutf_warn_unused result implementation::validate_utf16be_with_errors(const char16_t *buf, size_t len) const noexcept {
    return scalar::utf16::validate_with_errors<endianness::BIG>(buf, len);
}

simdutf_warn_unused bool implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
    return scalar::utf32::validate(buf, len);
}

simdutf_warn_unused result implementation::validate_utf32_with_errors(const char32_t *buf, size_t len) const noexcept {
    return scalar::utf32::validate_with_errors(buf, len);
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
   return scalar::utf8_to_utf16::convert<endianness::LITTLE>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf16be(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
   return scalar::utf8_to_utf16::convert<endianness::BIG>(buf, len, utf16_output);
}

simdutf_warn_unused result implementation::convert_utf8_to_utf16le_with_errors(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
   return scalar::utf8_to_utf16::convert_with_errors<endianness::LITTLE>(buf, len, utf16_output);
}

simdutf_warn_unused result implementation::convert_utf8_to_utf16be_with_errors(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
   return scalar::utf8_to_utf16::convert_with_errors<endianness::BIG>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16le(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
   return scalar::utf8_to_utf16::convert_valid<endianness::LITTLE>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
   return scalar::utf8_to_utf16::convert_valid<endianness::BIG>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(const char* buf, size_t len, char32_t* utf32_output) const noexcept {
   return scalar::utf8_to_utf32::convert(buf, len, utf32_output);
}

simdutf_warn_unused result implementation::convert_utf8_to_utf32_with_errors(const char* buf, size_t len, char32_t* utf32_output) const noexcept {
   return scalar::utf8_to_utf32::convert_with_errors(buf, len, utf32_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(const char* input, size_t size,
    char32_t* utf32_output) const noexcept {
  return scalar::utf8_to_utf32::convert_valid(input, size,  utf32_output);
}

simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf16_to_utf8::convert<endianness::LITTLE>(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_utf16be_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf16_to_utf8::convert<endianness::BIG>(buf, len, utf8_output);
}

simdutf_warn_unused result implementation::convert_utf16le_to_utf8_with_errors(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf16_to_utf8::convert_with_errors<endianness::LITTLE>(buf, len, utf8_output);
}

simdutf_warn_unused result implementation::convert_utf16be_to_utf8_with_errors(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf16_to_utf8::convert_with_errors<endianness::BIG>(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf16_to_utf8::convert_valid<endianness::LITTLE>(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf16_to_utf8::convert_valid<endianness::BIG>(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf32_to_utf8::convert(buf, len, utf8_output);
}

simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf32_to_utf8::convert_with_errors(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf32_to_utf8::convert_valid(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return scalar::utf32_to_utf16::convert<endianness::LITTLE>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf16be(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return scalar::utf32_to_utf16::convert<endianness::BIG>(buf, len, utf16_output);
}

simdutf_warn_unused result implementation::convert_utf32_to_utf16le_with_errors(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return scalar::utf32_to_utf16::convert_with_errors<endianness::LITTLE>(buf, len, utf16_output);
}

simdutf_warn_unused result implementation::convert_utf32_to_utf16be_with_errors(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return scalar::utf32_to_utf16::convert_with_errors<endianness::BIG>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16le(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return scalar::utf32_to_utf16::convert_valid<endianness::LITTLE>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16be(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return scalar::utf32_to_utf16::convert_valid<endianness::BIG>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return scalar::utf16_to_utf32::convert<endianness::LITTLE>(buf, len, utf32_output);
}

simdutf_warn_unused size_t implementation::convert_utf16be_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return scalar::utf16_to_utf32::convert<endianness::BIG>(buf, len, utf32_output);
}

simdutf_warn_unused result implementation::convert_utf16le_to_utf32_with_errors(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(buf, len, utf32_output);
}

simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(buf, len, utf32_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return scalar::utf16_to_utf32::convert_valid<endianness::LITTLE>(buf, len, utf32_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return scalar::utf16_to_utf32::convert_valid<endianness::BIG>(buf, len, utf32_output);
}

void implementation::change_endianness_utf16(const char16_t * input, size_t length, char16_t * output) const noexcept {
  scalar::utf16::change_endianness_utf16(input, length, output);
}

simdutf_warn_unused size_t implementation::count_utf16le(const char16_t * input, size_t length) const noexcept {
  return scalar::utf16::count_code_points<endianness::LITTLE>(input, length);
}

simdutf_warn_unused size_t implementation::count_utf16be(const char16_t * input, size_t length) const noexcept {
  return scalar::utf16::count_code_points<endianness::BIG>(input, length);
}

simdutf_warn_unused size_t implementation::count_utf8(const char * input, size_t length) const noexcept {
  return scalar::utf8::count_code_points(input, length);
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(const char16_t * input, size_t length) const noexcept {
  return scalar::utf16::utf8_length_from_utf16<endianness::LITTLE>(input, length);
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(const char16_t * input, size_t length) const noexcept {
  return scalar::utf16::utf8_length_from_utf16<endianness::BIG>(input, length);
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(const char16_t * input, size_t length) const noexcept {
  return scalar::utf16::utf32_length_from_utf16<endianness::LITTLE>(input, length);
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(const char16_t * input, size_t length) const noexcept {
  return scalar::utf16::utf32_length_from_utf16<endianness::BIG>(input, length);
}

simdutf_warn_unused size_t implementation::utf16_length_from_utf8(const char * input, size_t length) const noexcept {
  return scalar::utf8::utf16_length_from_utf8(input, length);
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf32(const char32_t * input, size_t length) const noexcept {
  return scalar::utf32::utf8_length_from_utf32(input, length);
}

simdutf_warn_unused size_t implementation::utf16_length_from_utf32(const char32_t * input, size_t length) const noexcept {
  return scalar::utf32::utf16_length_from_utf32(input, length);
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf8(const char * input, size_t length) const noexcept {
  return scalar::utf8::count_code_points(input, length);
}

} // namespace fallback
} // namespace simdutf

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/fallback/end.h
/* begin file src/simdutf/fallback/end.h */
/* end file src/simdutf/fallback/end.h */
/* end file src/fallback/implementation.cpp */
#endif
#if SIMDUTF_IMPLEMENTATION_ICELAKE
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=icelake/implementation.cpp
/* begin file src/icelake/implementation.cpp */


// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/icelake/begin.h
/* begin file src/simdutf/icelake/begin.h */
// redefining SIMDUTF_IMPLEMENTATION to "icelake"
// #define SIMDUTF_IMPLEMENTATION icelake

#if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
// nothing needed.
#else
SIMDUTF_TARGET_ICELAKE
#endif

#if SIMDUTF_GCC11ORMORE // workaround for https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
SIMDUTF_DISABLE_GCC_WARNING(-Wmaybe-uninitialized)
#endif // end of workaround
/* end file src/simdutf/icelake/begin.h */
namespace simdutf {
namespace icelake {
namespace {
#ifndef SIMDUTF_ICELAKE_H
#error "icelake.h must be included"
#endif
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=icelake/icelake_utf8_common.inl.cpp
/* begin file src/icelake/icelake_utf8_common.inl.cpp */
// Common procedures for both validating and non-validating conversions from UTF-8.
enum block_processing_mode { SIMDUTF_FULL, SIMDUTF_TAIL};

using utf8_to_utf16_result = std::pair<const char*, char16_t*>;
using utf8_to_utf32_result = std::pair<const char*, uint32_t*>;

/*
    process_block_utf8_to_utf16 converts up to 64 bytes from 'in' from UTF-8
    to UTF-16. When tail = SIMDUTF_FULL, then the full input buffer (64 bytes)
    might be used. When tail = SIMDUTF_TAIL, we take into account 'gap' which
    indicates how many input bytes are relevant.

    Returns true when the result is correct, otherwise it returns false.

    The provided in and out pointers are advanced according to how many input
    bytes have been processed, upon success.
*/
template <block_processing_mode tail, endianness big_endian>
simdutf_really_inline bool process_block_utf8_to_utf16(const char *&in, char16_t *&out, size_t gap) {
  // constants
  __m512i mask_identity = _mm512_set_epi8(63, 62, 61, 60, 59, 58, 57, 56, 55, 54, 53, 52, 51, 50, 49, 48, 47, 46, 45, 44, 43, 42, 41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 27, 26, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0);
  __m512i mask_c0c0c0c0 = _mm512_set1_epi32(0xc0c0c0c0);
  __m512i mask_80808080 = _mm512_set1_epi32(0x80808080);
  __m512i mask_f0f0f0f0 = _mm512_set1_epi32(0xf0f0f0f0);
  __m512i mask_dfdfdfdf_tail = _mm512_set_epi64(0xffffdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf, 0xdfdfdfdfdfdfdfdf);
  __m512i mask_c2c2c2c2 = _mm512_set1_epi32(0xc2c2c2c2);
  __m512i mask_ffffffff = _mm512_set1_epi32(0xffffffff);
  __m512i mask_d7c0d7c0 = _mm512_set1_epi32(0xd7c0d7c0);
  __m512i mask_dc00dc00 = _mm512_set1_epi32(0xdc00dc00);
  __m512i byteflip = _mm512_setr_epi64(
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809
        );
  // Note that 'tail' is a compile-time constant !
  __mmask64 b = (tail == SIMDUTF_FULL) ? 0xFFFFFFFFFFFFFFFF : (uint64_t(1) << gap) - 1;
  __m512i input = (tail == SIMDUTF_FULL) ? _mm512_loadu_si512(in) : _mm512_maskz_loadu_epi8(b, in);
  __mmask64 m1 = (tail == SIMDUTF_FULL) ? _mm512_cmplt_epu8_mask(input, mask_80808080) : _mm512_mask_cmplt_epu8_mask(b, input, mask_80808080);
  if(_ktestc_mask64_u8(m1, b)) {// NOT(m1) AND b -- if all zeroes, then all ASCII
  // alternatively, we could do 'if (m1 == b) { '
    if (tail == SIMDUTF_FULL) {
      in += 64;          // consumed 64 bytes
      // we convert a full 64-byte block, writing 128 bytes.
      __m512i input1 = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(input));
      if(big_endian) { input1 = _mm512_shuffle_epi8(input1, byteflip); }
      _mm512_storeu_si512(out, input1);
      out += 32;
      __m512i input2 = _mm512_cvtepu8_epi16(_mm512_extracti64x4_epi64(input, 1));
      if(big_endian) { input2 = _mm512_shuffle_epi8(input2, byteflip); }
      _mm512_storeu_si512(out, input2);
      out += 32;
      return true; // we are done
    } else {
      in += gap;
      if (gap <= 32) {
        __m512i input1 = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(input));
        if(big_endian) { input1 = _mm512_shuffle_epi8(input1, byteflip); }
        _mm512_mask_storeu_epi16(out, __mmask32((uint64_t(1) << (gap)) - 1), input1);
        out += gap;
      } else {
        __m512i input1 = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(input));
        if(big_endian) { input1 = _mm512_shuffle_epi8(input1, byteflip); }
        _mm512_storeu_si512(out, input1);
        out += 32;
        __m512i input2 = _mm512_cvtepu8_epi16(_mm512_extracti64x4_epi64(input, 1));
        if(big_endian) { input2 = _mm512_shuffle_epi8(input2, byteflip); }
        _mm512_mask_storeu_epi16(out, __mmask32((uint32_t(1) << (gap - 32)) - 1), input2);
        out += gap - 32;
      }
      return true; // we are done
    }
  }
  // classify characters further
  __mmask64 m234 = _mm512_cmp_epu8_mask(mask_c0c0c0c0, input,
                                        _MM_CMPINT_LE); // 0xc0 <= input, 2, 3, or 4 leading byte
  __mmask64 m34 = _mm512_cmp_epu8_mask(mask_dfdfdfdf_tail, input,
                                       _MM_CMPINT_LT); // 0xdf < input,  3 or 4 leading byte

  __mmask64 milltwobytes = _mm512_mask_cmp_epu8_mask(m234, input, mask_c2c2c2c2,
                                                     _MM_CMPINT_LT); // 0xc0 <= input < 0xc2 (illegal two byte sequence)
                                                                     // Overlong 2-byte sequence
  if (_ktestz_mask64_u8(milltwobytes, milltwobytes) == 0) {
    // Overlong 2-byte sequence
    return false;
  }
  if (_ktestz_mask64_u8(m34, m34) == 0) {
    // We have a 3-byte sequence and/or a 2-byte sequence, or possibly even a 4-byte sequence!
    __mmask64 m4 = _mm512_cmp_epu8_mask(input, mask_f0f0f0f0,
                                        _MM_CMPINT_NLT); // 0xf0 <= zmm0 (4 byte start bytes)

    __mmask64 mask_not_ascii = (tail == SIMDUTF_FULL) ? _knot_mask64(m1) : _kand_mask64(_knot_mask64(m1), b);

    __mmask64 mp1 = _kshiftli_mask64(m234, 1);
    __mmask64 mp2 = _kshiftli_mask64(m34, 2);
    // We could do it as follows...
    // if (_kortestz_mask64_u8(m4,m4)) { // compute the bitwise OR of the 64-bit masks a and b and return 1 if all zeroes
    // but GCC generates better code when we do:
    if (m4 == 0) { // compute the bitwise OR of the 64-bit masks a and b and return 1 if all zeroes
      // Fast path with 1,2,3 bytes
      __mmask64 mc = _kor_mask64(mp1, mp2); // expected continuation bytes
      __mmask64 m1234 = _kor_mask64(m1, m234);
      // mismatched continuation bytes:
      if (tail == SIMDUTF_FULL) {
        __mmask64 xnormcm1234 = _kxnor_mask64(mc, m1234); // XNOR of mc and m1234 should be all zero if they differ
        // the presence of a 1 bit indicates that they overlap.
        // _kortestz_mask64_u8: compute the bitwise OR of 64-bit masksand return 1 if all zeroes.
        if (!_kortestz_mask64_u8(xnormcm1234, xnormcm1234)) { return false; }
      } else {
        __mmask64 bxorm1234 = _kxor_mask64(b, m1234);
        if (mc != bxorm1234) { return false; }
      }
      // mend: identifying the last bytes of each sequence to be decoded
      __mmask64 mend = _kshiftri_mask64(m1234, 1);
      if (tail != SIMDUTF_FULL) {
        mend = _kor_mask64(mend, (uint64_t(1) << (gap - 1)));
      }


      __m512i last_and_third = _mm512_maskz_compress_epi8(mend, mask_identity);
      __m512i last_and_thirdu16 = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(last_and_third));

      __m512i nonasciitags = _mm512_maskz_mov_epi8(mask_not_ascii, mask_c0c0c0c0); // ASCII: 00000000  other: 11000000
      __m512i clearedbytes = _mm512_andnot_si512(nonasciitags, input);             // high two bits cleared where not ASCII
      __m512i lastbytes = _mm512_maskz_permutexvar_epi8(0x5555555555555555, last_and_thirdu16,
                                                        clearedbytes); // the last byte of each character

      __mmask64 mask_before_non_ascii = _kshiftri_mask64(mask_not_ascii, 1);               // bytes that precede non-ASCII bytes
      __m512i indexofsecondlastbytes = _mm512_add_epi16(mask_ffffffff, last_and_thirdu16); // indices of the second last bytes
      __m512i beforeasciibytes = _mm512_maskz_mov_epi8(mask_before_non_ascii, clearedbytes);
      __m512i secondlastbytes = _mm512_maskz_permutexvar_epi8(0x5555555555555555, indexofsecondlastbytes,
                                                              beforeasciibytes); // the second last bytes (of two, three byte seq,
                                                                                 // surrogates)
      secondlastbytes = _mm512_slli_epi16(secondlastbytes, 6);                   // shifted into position

      __m512i indexofthirdlastbytes = _mm512_add_epi16(mask_ffffffff,
                                                       indexofsecondlastbytes); // indices of the second last bytes
      __m512i thirdlastbyte = _mm512_maskz_mov_epi8(m34,
                                                    clearedbytes); // only those that are the third last byte of a sequece
      __m512i thirdlastbytes = _mm512_maskz_permutexvar_epi8(0x5555555555555555, indexofthirdlastbytes,
                                                             thirdlastbyte); // the third last bytes (of three byte sequences, hi
                                                                             // surrogate)
      thirdlastbytes = _mm512_slli_epi16(thirdlastbytes, 12);                // shifted into position
      __m512i Wout = _mm512_ternarylogic_epi32(lastbytes, secondlastbytes, thirdlastbytes, 254);
      // the elements of Wout excluding the last element if it happens to be a high surrogate:

      __mmask64 mprocessed = (tail == SIMDUTF_FULL) ? _pdep_u64(0xFFFFFFFF, mend) : _pdep_u64(0xFFFFFFFF, _kand_mask64(mend, b)); // we adjust mend at the end of the output.


      // Encodings out of range...
      {
        // the location of 3-byte sequence start bytes in the input
        __mmask64 m3 = m34 & (b ^ m4);
        // words in Wout corresponding to 3-byte sequences.
        __mmask32 M3 = __mmask32(_pext_u64(m3 << 2, mend));
        __m512i mask_08000800 = _mm512_set1_epi32(0x08000800);
        __mmask32 Msmall800 = _mm512_mask_cmplt_epu16_mask(M3, Wout, mask_08000800);
        __m512i mask_d800d800 = _mm512_set1_epi32(0xd800d800);
        __m512i Moutminusd800 = _mm512_sub_epi16(Wout, mask_d800d800);
        __mmask32 M3s = _mm512_mask_cmplt_epu16_mask(M3, Moutminusd800, mask_08000800);
        if (_kor_mask32(Msmall800, M3s)) { return false; }
      }
      int64_t nout = _mm_popcnt_u64(mprocessed);
      in +=  64 - _lzcnt_u64(mprocessed);
      if(big_endian) { Wout = _mm512_shuffle_epi8(Wout, byteflip); }
      _mm512_mask_storeu_epi16(out, __mmask32((uint64_t(1) << nout) - 1), Wout);
      out += nout;
      return true; // ok
    }
    //
    // We have a 4-byte sequence, this is the general case.
    // Slow!
    __mmask64 mp3 = _kshiftli_mask64(m4, 3);
    __mmask64 mc = _kor_mask64(_kor_mask64(mp1, mp2), mp3); // expected continuation bytes
    __mmask64 m1234 = _kor_mask64(m1, m234);

    // mend: identifying the last bytes of each sequence to be decoded
    __mmask64 mend = _kor_mask64(_kshiftri_mask64(_kor_mask64(mp3, m1234), 1), mp3);
    if (tail != SIMDUTF_FULL) {
      mend = _kor_mask64(mend, __mmask64(uint64_t(1) << (gap - 1)));
    }
    __m512i last_and_third = _mm512_maskz_compress_epi8(mend, mask_identity);
    __m512i last_and_thirdu16 = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(last_and_third));

    __m512i nonasciitags = _mm512_maskz_mov_epi8(mask_not_ascii, mask_c0c0c0c0); // ASCII: 00000000  other: 11000000
    __m512i clearedbytes = _mm512_andnot_si512(nonasciitags, input);             // high two bits cleared where not ASCII
    __m512i lastbytes = _mm512_maskz_permutexvar_epi8(0x5555555555555555, last_and_thirdu16,
                                                      clearedbytes); // the last byte of each character

    __mmask64 mask_before_non_ascii = _kshiftri_mask64(mask_not_ascii, 1);               // bytes that precede non-ASCII bytes
    __m512i indexofsecondlastbytes = _mm512_add_epi16(mask_ffffffff, last_and_thirdu16); // indices of the second last bytes
    __m512i beforeasciibytes = _mm512_maskz_mov_epi8(mask_before_non_ascii, clearedbytes);
    __m512i secondlastbytes = _mm512_maskz_permutexvar_epi8(0x5555555555555555, indexofsecondlastbytes,
                                                            beforeasciibytes); // the second last bytes (of two, three byte seq,
                                                                               // surrogates)
    secondlastbytes = _mm512_slli_epi16(secondlastbytes, 6);                   // shifted into position

    __m512i indexofthirdlastbytes = _mm512_add_epi16(mask_ffffffff,
                                                     indexofsecondlastbytes); // indices of the second last bytes
    __m512i thirdlastbyte = _mm512_maskz_mov_epi8(m34,
                                                  clearedbytes); // only those that are the third last byte of a sequece
    __m512i thirdlastbytes = _mm512_maskz_permutexvar_epi8(0x5555555555555555, indexofthirdlastbytes,
                                                           thirdlastbyte); // the third last bytes (of three byte sequences, hi
                                                                           // surrogate)
    thirdlastbytes = _mm512_slli_epi16(thirdlastbytes, 12);                // shifted into position
    __m512i thirdsecondandlastbytes = _mm512_ternarylogic_epi32(lastbytes, secondlastbytes, thirdlastbytes, 254);
    uint64_t Mlo_uint64 = _pext_u64(mp3, mend);
    __mmask32 Mlo = __mmask32(Mlo_uint64);
    __mmask32 Mhi = __mmask32(Mlo_uint64 >> 1);
    __m512i lo_surr_mask = _mm512_maskz_mov_epi16(Mlo,
                                                  mask_dc00dc00); // lo surr: 1101110000000000, other:  0000000000000000
    __m512i shifted4_thirdsecondandlastbytes = _mm512_srli_epi16(thirdsecondandlastbytes,
                                                                 4); // hi surr: 00000WVUTSRQPNML  vuts = WVUTS - 1
    __m512i tagged_lo_surrogates = _mm512_or_si512(thirdsecondandlastbytes,
                                                   lo_surr_mask); // lo surr: 110111KJHGFEDCBA, other:  unchanged
    __m512i Wout = _mm512_mask_add_epi16(tagged_lo_surrogates, Mhi, shifted4_thirdsecondandlastbytes,
                                         mask_d7c0d7c0); // hi sur: 110110vutsRQPNML, other:  unchanged
    // the elements of Wout excluding the last element if it happens to be a high surrogate:
    __mmask32 Mout = ~(Mhi & 0x80000000);
    __mmask64 mprocessed = (tail == SIMDUTF_FULL) ? _pdep_u64(Mout, mend) : _pdep_u64(Mout, _kand_mask64(mend, b)); // we adjust mend at the end of the output.


    // mismatched continuation bytes:
    if (tail == SIMDUTF_FULL) {
      __mmask64 xnormcm1234 = _kxnor_mask64(mc, m1234); // XNOR of mc and m1234 should be all zero if they differ
      // the presence of a 1 bit indicates that they overlap.
      // _kortestz_mask64_u8: compute the bitwise OR of 64-bit masksand return 1 if all zeroes.
      if (!_kortestz_mask64_u8(xnormcm1234, xnormcm1234)) { return false; }
    } else {
      __mmask64 bxorm1234 = _kxor_mask64(b, m1234);
      if (mc != bxorm1234) { return false; }
    }
    // Encodings out of range...
    {
      // the location of 3-byte sequence start bytes in the input
      __mmask64 m3 = m34 & (b ^ m4);
      // words in Wout corresponding to 3-byte sequences.
      __mmask32 M3 = __mmask32(_pext_u64(m3 << 2, mend));
      __m512i mask_08000800 = _mm512_set1_epi32(0x08000800);
      __mmask32 Msmall800 = _mm512_mask_cmplt_epu16_mask(M3, Wout, mask_08000800);
      __m512i mask_d800d800 = _mm512_set1_epi32(0xd800d800);
      __m512i Moutminusd800 = _mm512_sub_epi16(Wout, mask_d800d800);
      __mmask32 M3s = _mm512_mask_cmplt_epu16_mask(M3, Moutminusd800, mask_08000800);
      __m512i mask_04000400 = _mm512_set1_epi32(0x04000400);
      __mmask32 M4s = _mm512_mask_cmpge_epu16_mask(Mhi, Moutminusd800, mask_04000400);
      if (!_kortestz_mask32_u8(M4s, _kor_mask32(Msmall800, M3s))) { return false; }
    }
    in += 64 - _lzcnt_u64(mprocessed);
    int64_t nout = _mm_popcnt_u64(mprocessed);
    if(big_endian) { Wout = _mm512_shuffle_epi8(Wout, byteflip); }
    _mm512_mask_storeu_epi16(out, __mmask32((uint64_t(1) << nout) - 1), Wout);
    out += nout;
    return true; // ok
  }
  // Fast path 2: all ASCII or 2 byte
  __mmask64 continuation_or_ascii = (tail == SIMDUTF_FULL) ? _knot_mask64(m234) : _kand_mask64(_knot_mask64(m234), b);
  // on top of -0xc0 we substract -2 which we get back later of the
  // continuation byte tags
  __m512i leading2byte = _mm512_maskz_sub_epi8(m234, input, mask_c2c2c2c2);
  __mmask64 leading = tail == (tail == SIMDUTF_FULL) ? _kor_mask64(m1, m234) : _kand_mask64(_kor_mask64(m1, m234), b); // first bytes of each sequence
  if (tail == SIMDUTF_FULL) {
    __mmask64 xnor234leading = _kxnor_mask64(_kshiftli_mask64(m234, 1), leading);
    if (!_kortestz_mask64_u8(xnor234leading, xnor234leading)) { return false; }
  } else {
    __mmask64 bxorleading = _kxor_mask64(b, leading);
    if (_kshiftli_mask64(m234, 1) != bxorleading) { return false; }
  }
  //
  if (tail == SIMDUTF_FULL) {
    // In the two-byte/ASCII scenario, we are easily latency bound, so we want
    // to increment the input buffer as quickly as possible.
    // We process 32 bytes unless the byte at index 32 is a continuation byte,
    // in which case we include it as well for a total of 33 bytes.
    // Note that if x is an ASCII byte, then the following is false:
    // int8_t(x) <= int8_t(0xc0) under two's complement.
    in += 32;
    if(int8_t(*in) <= int8_t(0xc0)) in++;
    // The alternative is to do
    // in += 64 - _lzcnt_u64(_pdep_u64(0xFFFFFFFF, continuation_or_ascii));
    // but it requires loading the input, doing the mask computation, and converting
    // back the mask to a general register. It just takes too long, leaving the
    // processor likely to be idle.
  } else {
    in += 64 - _lzcnt_u64(_pdep_u64(0xFFFFFFFF, continuation_or_ascii));
  }
  __m512i lead = _mm512_maskz_compress_epi8(leading, leading2byte);          // will contain zero for ascii, and the data
  lead = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(lead));                 // ... zero extended into words
  __m512i follow = _mm512_maskz_compress_epi8(continuation_or_ascii, input); // the last bytes of each sequence
  follow = _mm512_cvtepu8_epi16(_mm512_castsi512_si256(follow));             // ... zero extended into words
  lead = _mm512_slli_epi16(lead, 6);                                         // shifted into position
  __m512i final = _mm512_add_epi16(follow, lead);                            // combining lead and follow

  if(big_endian) { final = _mm512_shuffle_epi8(final, byteflip); }
  if (tail == SIMDUTF_FULL) {
    // Next part is UTF-16 specific and can be generalized to UTF-32.
    int nout = _mm_popcnt_u32(uint32_t(leading));
    _mm512_mask_storeu_epi16(out, __mmask32((uint64_t(1) << nout) - 1), final);
    out += nout; // UTF-8 to UTF-16 is only expansionary in this case.
  } else {
    int nout = int(_mm_popcnt_u64(_pdep_u64(0xFFFFFFFF, leading)));
    _mm512_mask_storeu_epi16(out, __mmask32((uint64_t(1) << nout) - 1), final);
    out += nout; // UTF-8 to UTF-16 is only expansionary in this case.
  }

  return true; // we are fine.
}




/*
    utf32_to_utf16_masked converts `count` lower UTF-32 words
    from input `utf32` into UTF-16. It differs from utf32_to_utf16
    in that it 'masks' the writes.

    Returns how many 16-bit words were stored.

    byteflip is used for flipping 16-bit words, and it should be
        __m512i byteflip = _mm512_setr_epi64(
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809
        );
    We pass it to the (always inlined) function to encourage the compiler to
    keep the value in a (constant) register.
*/
template <endianness big_endian>
simdutf_really_inline size_t utf32_to_utf16_masked(const __m512i byteflip, __m512i utf32, unsigned int count, char16_t* output) {

    const __mmask16 valid = uint16_t((1 << count) - 1);
    // 1. check if we have any surrogate pairs
    const __m512i v_0000_ffff = _mm512_set1_epi32(0x0000ffff);
    const __mmask16 sp_mask = _mm512_mask_cmpgt_epu32_mask(valid, utf32, v_0000_ffff);

    if (sp_mask == 0) {
        if(big_endian) {
          _mm256_mask_storeu_epi16((__m256i*)output, valid, _mm256_shuffle_epi8(_mm512_cvtepi32_epi16(utf32), _mm512_castsi512_si256(byteflip)));

        } else {
          _mm256_mask_storeu_epi16((__m256i*)output, valid, _mm512_cvtepi32_epi16(utf32));
        }
        return count;
    }

    {
        // build surrogate pair words in 32-bit lanes

        //    t0 = 8 x [000000000000aaaa|aaaaaabbbbbbbbbb]
        const __m512i v_0001_0000 = _mm512_set1_epi32(0x00010000);
        const __m512i t0 = _mm512_sub_epi32(utf32, v_0001_0000);

        //    t1 = 8 x [000000aaaaaaaaaa|bbbbbbbbbb000000]
        const __m512i t1 = _mm512_slli_epi32(t0, 6);

        //    t2 = 8 x [000000aaaaaaaaaa|aaaaaabbbbbbbbbb] -- copy hi word from t1 to t0
        //         0xe4 = (t1 and v_ffff_0000) or (t0 and not v_ffff_0000)
        const __m512i v_ffff_0000 = _mm512_set1_epi32(0xffff0000);
        const __m512i t2 = _mm512_ternarylogic_epi32(t1, t0, v_ffff_0000, 0xe4);

        //    t2 = 8 x [110110aaaaaaaaaa|110111bbbbbbbbbb] -- copy hi word from t1 to t0
        //         0xba = (t2 and not v_fc00_fc000) or v_d800_dc00
        const __m512i v_fc00_fc00 = _mm512_set1_epi32(0xfc00fc00);
        const __m512i v_d800_dc00 = _mm512_set1_epi32(0xd800dc00);
        const __m512i t3 = _mm512_ternarylogic_epi32(t2, v_fc00_fc00, v_d800_dc00, 0xba);
        const __m512i t4 = _mm512_mask_blend_epi32(sp_mask, utf32, t3);
        __m512i t5 = _mm512_ror_epi32(t4, 16);
        // Here we want to trim all of the upper 16-bit words from the 2-byte
        // characters represented as 4-byte values. We can compute it from
        // sp_mask or the following... It can be more optimized!
        const  __mmask32 nonzero = _kor_mask32(0xaaaaaaaa,_mm512_cmpneq_epi16_mask(t5, _mm512_setzero_si512()));
        const  __mmask32 nonzero_masked = _kand_mask32(nonzero, __mmask32((uint64_t(1) << (2*count)) - 1));
        if(big_endian) { t5 = _mm512_shuffle_epi8(t5, byteflip); }
        // we deliberately avoid _mm512_mask_compressstoreu_epi16 for portability (zen4)
        __m512i compressed = _mm512_maskz_compress_epi16(nonzero_masked, t5);
        _mm512_mask_storeu_epi16(output, (1<<(count + static_cast<unsigned int>(count_ones(sp_mask)))) - 1, compressed);
        //_mm512_mask_compressstoreu_epi16(output, nonzero_masked, t5);
    }

    return count + static_cast<unsigned int>(count_ones(sp_mask));
}

/*
    utf32_to_utf16 converts `count` lower UTF-32 words
    from input `utf32` into UTF-16. It may overflow.

    Returns how many 16-bit words were stored.

    byteflip is used for flipping 16-bit words, and it should be
        __m512i byteflip = _mm512_setr_epi64(
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809
        );
    We pass it to the (always inlined) function to encourage the compiler to
    keep the value in a (constant) register.
*/
template <endianness big_endian>
simdutf_really_inline size_t utf32_to_utf16(const __m512i byteflip, __m512i utf32, unsigned int count, char16_t* output) {
    // check if we have any surrogate pairs
    const __m512i v_0000_ffff = _mm512_set1_epi32(0x0000ffff);
    const __mmask16 sp_mask = _mm512_cmpgt_epu32_mask(utf32, v_0000_ffff);

    if (sp_mask == 0) {
        // technically, it should be _mm256_storeu_epi16
        if(big_endian) {
          _mm256_storeu_si256((__m256i*)output, _mm256_shuffle_epi8(_mm512_cvtepi32_epi16(utf32),_mm512_castsi512_si256(byteflip)));
        } else {
          _mm256_storeu_si256((__m256i*)output, _mm512_cvtepi32_epi16(utf32));
        }
        return count;
    }

    {
        // build surrogate pair words in 32-bit lanes

        //    t0 = 8 x [000000000000aaaa|aaaaaabbbbbbbbbb]
        const __m512i v_0001_0000 = _mm512_set1_epi32(0x00010000);
        const __m512i t0 = _mm512_sub_epi32(utf32, v_0001_0000);

        //    t1 = 8 x [000000aaaaaaaaaa|bbbbbbbbbb000000]
        const __m512i t1 = _mm512_slli_epi32(t0, 6);

        //    t2 = 8 x [000000aaaaaaaaaa|aaaaaabbbbbbbbbb] -- copy hi word from t1 to t0
        //         0xe4 = (t1 and v_ffff_0000) or (t0 and not v_ffff_0000)
        const __m512i v_ffff_0000 = _mm512_set1_epi32(0xffff0000);
        const __m512i t2 = _mm512_ternarylogic_epi32(t1, t0, v_ffff_0000, 0xe4);

        //    t2 = 8 x [110110aaaaaaaaaa|110111bbbbbbbbbb] -- copy hi word from t1 to t0
        //         0xba = (t2 and not v_fc00_fc000) or v_d800_dc00
        const __m512i v_fc00_fc00 = _mm512_set1_epi32(0xfc00fc00);
        const __m512i v_d800_dc00 = _mm512_set1_epi32(0xd800dc00);
        const __m512i t3 = _mm512_ternarylogic_epi32(t2, v_fc00_fc00, v_d800_dc00, 0xba);
        const __m512i t4 = _mm512_mask_blend_epi32(sp_mask, utf32, t3);
        __m512i t5 = _mm512_ror_epi32(t4, 16);
        const  __mmask32 nonzero = _kor_mask32(0xaaaaaaaa,_mm512_cmpneq_epi16_mask(t5, _mm512_setzero_si512()));
        if(big_endian) { t5 = _mm512_shuffle_epi8(t5, byteflip); }
        // we deliberately avoid _mm512_mask_compressstoreu_epi16 for portability (zen4)
        __m512i compressed = _mm512_maskz_compress_epi16(nonzero, t5);
        _mm512_mask_storeu_epi16(output, (1<<(count + static_cast<unsigned int>(count_ones(sp_mask)))) - 1, compressed);
        //_mm512_mask_compressstoreu_epi16(output, nonzero, t5);
    }

    return count + static_cast<unsigned int>(count_ones(sp_mask));
}

/**
 * Store the last N bytes of previous followed by 512-N bytes from input.
 */
template <int N>
__m512i prev(__m512i input, __m512i previous) {
    static_assert(N<=32, "N must be no larger than 32");
    const __m512i movemask = _mm512_setr_epi32(28,29,30,31,0,1,2,3,4,5,6,7,8,9,10,11);
    const __m512i rotated = _mm512_permutex2var_epi32(input, movemask, previous);
#if SIMDUTF_GCC8 || SIMDUTF_GCC9
    constexpr int shift = 16-N; // workaround for GCC8,9
    return _mm512_alignr_epi8(input, rotated, shift);
#else
    return _mm512_alignr_epi8(input, rotated, 16-N);
#endif // SIMDUTF_GCC8 || SIMDUTF_GCC9
}

template <unsigned idx0, unsigned idx1, unsigned idx2, unsigned idx3>
__m512i shuffle_epi128(__m512i v) {
    static_assert((idx0 >= 0 && idx0 <= 3), "idx0 must be in range 0..3");
    static_assert((idx1 >= 0 && idx1 <= 3), "idx1 must be in range 0..3");
    static_assert((idx2 >= 0 && idx2 <= 3), "idx2 must be in range 0..3");
    static_assert((idx3 >= 0 && idx3 <= 3), "idx3 must be in range 0..3");

    constexpr unsigned shuffle = idx0 | (idx1 << 2) | (idx2 << 4) | (idx3 << 6);
    return _mm512_shuffle_i32x4(v, v, shuffle);
}

template <unsigned idx>
constexpr __m512i broadcast_epi128(__m512i v) {
    return shuffle_epi128<idx, idx, idx, idx>(v);
}

/**
 * Current unused.
 */
template <int N>
__m512i rotate_by_N_epi8(const __m512i input) {

    // lanes order: 1, 2, 3, 0 => 0b00_11_10_01
    const __m512i permuted = _mm512_shuffle_i32x4(input, input, 0x39);

    return _mm512_alignr_epi8(permuted, input, N);
}

/*
    expanded_utf8_to_utf32 converts expanded UTF-8 characters (`utf8`)
    stored at separate 32-bit lanes.

    For each lane we have also a character class (`char_class), given in form
    0x8080800N, where N is 4 higest bits from the leading byte; 0x80 resets
    corresponding bytes during pshufb.
*/
simdutf_really_inline __m512i expanded_utf8_to_utf32(__m512i char_class, __m512i utf8) {
    /*
        Input:
        - utf8: bytes stored at separate 32-bit words
        - valid: which words have valid UTF-8 characters

        Bit layout of single word. We show 4 cases for each possible
        UTF-8 character encoding. The `?` denotes bits we must not
        assume their value.

        |10dd.dddd|10cc.cccc|10bb.bbbb|1111.0aaa| 4-byte char
        |????.????|10cc.cccc|10bb.bbbb|1110.aaaa| 3-byte char
        |????.????|????.????|10bb.bbbb|110a.aaaa| 2-byte char
        |????.????|????.????|????.????|0aaa.aaaa| ASCII char
          byte 3    byte 2    byte 1     byte 0
    */

    /* 1. Reset control bits of continuation bytes and the MSB
          of the leading byte; this makes all bytes unsigned (and
          does not alter ASCII char).

        |00dd.dddd|00cc.cccc|00bb.bbbb|0111.0aaa| 4-byte char
        |00??.????|00cc.cccc|00bb.bbbb|0110.aaaa| 3-byte char
        |00??.????|00??.????|00bb.bbbb|010a.aaaa| 2-byte char
        |00??.????|00??.????|00??.????|0aaa.aaaa| ASCII char
         ^^        ^^        ^^        ^
    */
    __m512i values;
    const __m512i v_3f3f_3f7f = _mm512_set1_epi32(0x3f3f3f7f);
    values = _mm512_and_si512(utf8, v_3f3f_3f7f);

    /* 2. Swap and join fields A-B and C-D

        |0000.cccc|ccdd.dddd|0001.110a|aabb.bbbb| 4-byte char
        |0000.cccc|cc??.????|0001.10aa|aabb.bbbb| 3-byte char
        |0000.????|????.????|0001.0aaa|aabb.bbbb| 2-byte char
        |0000.????|????.????|000a.aaaa|aa??.????| ASCII char */
    const __m512i v_0140_0140 = _mm512_set1_epi32(0x01400140);
    values = _mm512_maddubs_epi16(values, v_0140_0140);

    /* 3. Swap and join fields AB & CD

        |0000.0001|110a.aabb|bbbb.cccc|ccdd.dddd| 4-byte char
        |0000.0001|10aa.aabb|bbbb.cccc|cc??.????| 3-byte char
        |0000.0001|0aaa.aabb|bbbb.????|????.????| 2-byte char
        |0000.000a|aaaa.aa??|????.????|????.????| ASCII char */
    const __m512i v_0001_1000 = _mm512_set1_epi32(0x00011000);
    values = _mm512_madd_epi16(values, v_0001_1000);

    /* 4. Shift left the values by variable amounts to reset highest UTF-8 bits
        |aaab.bbbb|bccc.cccd|dddd.d000|0000.0000| 4-byte char -- by 11
        |aaaa.bbbb|bbcc.cccc|????.??00|0000.0000| 3-byte char -- by 10
        |aaaa.abbb|bbb?.????|????.???0|0000.0000| 2-byte char -- by 9
        |aaaa.aaa?|????.????|????.????|?000.0000| ASCII char -- by 7 */
    {
        /** pshufb

        continuation = 0
        ascii    = 7
        _2_bytes = 9
        _3_bytes = 10
        _4_bytes = 11

        shift_left_v3 = 4 * [
            ascii, # 0000
            ascii, # 0001
            ascii, # 0010
            ascii, # 0011
            ascii, # 0100
            ascii, # 0101
            ascii, # 0110
            ascii, # 0111
            continuation, # 1000
            continuation, # 1001
            continuation, # 1010
            continuation, # 1011
            _2_bytes, # 1100
            _2_bytes, # 1101
            _3_bytes, # 1110
            _4_bytes, # 1111
        ] */
        const __m512i shift_left_v3 = _mm512_setr_epi64(
            0x0707070707070707,
            0x0b0a090900000000,
            0x0707070707070707,
            0x0b0a090900000000,
            0x0707070707070707,
            0x0b0a090900000000,
            0x0707070707070707,
            0x0b0a090900000000
        );

        const __m512i shift = _mm512_shuffle_epi8(shift_left_v3, char_class);
        values = _mm512_sllv_epi32(values, shift);
    }

    /* 5. Shift right the values by variable amounts to reset lowest bits
        |0000.0000|000a.aabb|bbbb.cccc|ccdd.dddd| 4-byte char -- by 11
        |0000.0000|0000.0000|aaaa.bbbb|bbcc.cccc| 3-byte char -- by 16
        |0000.0000|0000.0000|0000.0aaa|aabb.bbbb| 2-byte char -- by 21
        |0000.0000|0000.0000|0000.0000|0aaa.aaaa| ASCII char -- by 25 */
    {
        // 4 * [25, 25, 25, 25, 25, 25, 25, 25, 0, 0, 0, 0, 21, 21, 16, 11]
        const __m512i shift_right = _mm512_setr_epi64(
            0x1919191919191919,
            0x0b10151500000000,
            0x1919191919191919,
            0x0b10151500000000,
            0x1919191919191919,
            0x0b10151500000000,
            0x1919191919191919,
            0x0b10151500000000
        );

        const __m512i shift = _mm512_shuffle_epi8(shift_right, char_class);
        values = _mm512_srlv_epi32(values, shift);
    }

    return values;
}


simdutf_really_inline __m512i expand_and_identify(__m512i lane0, __m512i lane1, int &count) {
    const __m512i merged = _mm512_mask_mov_epi32(lane0, 0x1000, lane1);
    const __m512i expand_ver2 = _mm512_setr_epi64(
                0x0403020103020100,
                0x0605040305040302,
                0x0807060507060504,
                0x0a09080709080706,
                0x0c0b0a090b0a0908,
                0x0e0d0c0b0d0c0b0a,
                0x000f0e0d0f0e0d0c,
                0x0201000f01000f0e
    );
    const __m512i input = _mm512_shuffle_epi8(merged, expand_ver2);
    const __m512i v_0000_00c0 = _mm512_set1_epi32(0xc0);
    const __m512i t0 = _mm512_and_si512(input, v_0000_00c0);
    const __m512i v_0000_0080 = _mm512_set1_epi32(0x80);
    const __mmask16 leading_bytes = _mm512_cmpneq_epu32_mask(t0, v_0000_0080);
    count = static_cast<int>(count_ones(leading_bytes));
    return  _mm512_mask_compress_epi32(_mm512_setzero_si512(), leading_bytes, input);
}

simdutf_really_inline __m512i expand_utf8_to_utf32(__m512i input) {
    __m512i char_class = _mm512_srli_epi32(input, 4);
    /*  char_class = ((input >> 4) & 0x0f) | 0x80808000 */
    const __m512i v_0000_000f = _mm512_set1_epi32(0x0f);
    const __m512i v_8080_8000 = _mm512_set1_epi32(0x80808000);
    char_class = _mm512_ternarylogic_epi32(char_class, v_0000_000f, v_8080_8000, 0xea);
    return expanded_utf8_to_utf32(char_class, input);
}
/* end file src/icelake/icelake_utf8_common.inl.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=icelake/icelake_macros.inl.cpp
/* begin file src/icelake/icelake_macros.inl.cpp */

/*
    This upcoming macro (SIMDUTF_ICELAKE_TRANSCODE16) takes 16 + 4 bytes (of a UTF-8 string)
    and loads all possible 4-byte substring into an AVX512 register.

    For example if we have bytes abcdefgh... we create following 32-bit lanes

    [abcd|bcde|cdef|defg|efgh|...]
     ^                          ^
     byte 0 of reg              byte 63 of reg
*/
/** pshufb
        # lane{0,1,2} have got bytes: [  0,  1,  2,  3,  4,  5,  6,  8,  9, 10, 11, 12, 13, 14, 15]
        # lane3 has got bytes:        [ 16, 17, 18, 19,  4,  5,  6,  8,  9, 10, 11, 12, 13, 14, 15]

        expand_ver2 = [
            # lane 0:
            0, 1, 2, 3,
            1, 2, 3, 4,
            2, 3, 4, 5,
            3, 4, 5, 6,

            # lane 1:
            4, 5, 6, 7,
            5, 6, 7, 8,
            6, 7, 8, 9,
            7, 8, 9, 10,

            # lane 2:
             8,  9, 10, 11,
             9, 10, 11, 12,
            10, 11, 12, 13,
            11, 12, 13, 14,

            # lane 3 order: 13, 14, 15, 16 14, 15, 16, 17, 15, 16, 17, 18, 16, 17, 18, 19
            12, 13, 14, 15,
            13, 14, 15,  0,
            14, 15,  0,  1,
            15,  0,  1,  2,
        ]
*/

#define SIMDUTF_ICELAKE_TRANSCODE16(LANE0, LANE1, MASKED)                                                    \
        {                                                                                                    \
            const __m512i merged = _mm512_mask_mov_epi32(LANE0, 0x1000, LANE1);                              \
            const __m512i expand_ver2 = _mm512_setr_epi64(                                                   \
                0x0403020103020100,                                                                          \
                0x0605040305040302,                                                                          \
                0x0807060507060504,                                                                          \
                0x0a09080709080706,                                                                          \
                0x0c0b0a090b0a0908,                                                                          \
                0x0e0d0c0b0d0c0b0a,                                                                          \
                0x000f0e0d0f0e0d0c,                                                                          \
                0x0201000f01000f0e                                                                           \
            );                                                                                               \
            const __m512i input = _mm512_shuffle_epi8(merged, expand_ver2);                                  \
                                                                                                             \
            __mmask16 leading_bytes;                                                                         \
            const __m512i v_0000_00c0 = _mm512_set1_epi32(0xc0);                                             \
            const __m512i t0 = _mm512_and_si512(input, v_0000_00c0);                                         \
            const __m512i v_0000_0080 = _mm512_set1_epi32(0x80);                                             \
            leading_bytes = _mm512_cmpneq_epu32_mask(t0, v_0000_0080);                                       \
                                                                                                             \
            __m512i char_class;                                                                              \
            char_class = _mm512_srli_epi32(input, 4);                                                        \
            /*  char_class = ((input >> 4) & 0x0f) | 0x80808000 */                                           \
            const __m512i v_0000_000f = _mm512_set1_epi32(0x0f);                                             \
            const __m512i v_8080_8000 = _mm512_set1_epi32(0x80808000);                                       \
            char_class = _mm512_ternarylogic_epi32(char_class, v_0000_000f, v_8080_8000, 0xea);              \
                                                                                                             \
            const int valid_count = static_cast<int>(count_ones(leading_bytes));                             \
            const __m512i utf32 = expanded_utf8_to_utf32(char_class, input);                                 \
                                                                                                             \
            const __m512i out = _mm512_mask_compress_epi32(_mm512_setzero_si512(), leading_bytes, utf32);    \
                                                                                                             \
            if (UTF32) {                                                                                     \
                if(MASKED) {                                                                                 \
                    const __mmask16 valid = uint16_t((1 << valid_count) - 1);                                \
                    _mm512_mask_storeu_epi32((__m512i*)output, valid, out);                                  \
                } else {                                                                                     \
                    _mm512_storeu_si512((__m512i*)output, out);                                              \
                }                                                                                            \
                output += valid_count;                                                                       \
            } else {                                                                                         \
                if(MASKED) {                                                                                 \
                    output += utf32_to_utf16_masked<big_endian>(byteflip, out, valid_count, reinterpret_cast<char16_t *>(output)); \
                } else {                                                                                     \
                    output += utf32_to_utf16<big_endian>(byteflip, out, valid_count, reinterpret_cast<char16_t *>(output));        \
                }                                                                                            \
            }                                                                                                \
        }

#define SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(INPUT, VALID_COUNT, MASKED)                                    \
{                                                                                                           \
    if (UTF32) {                                                                                            \
        if(MASKED) {                                                                                        \
            const __mmask16 valid_mask = uint16_t((1 << VALID_COUNT) - 1);                                  \
            _mm512_mask_storeu_epi32((__m512i*)output, valid_mask, INPUT);                                  \
        } else {                                                                                            \
            _mm512_storeu_si512((__m512i*)output, INPUT);                                              \
        }                                                                                                   \
        output += VALID_COUNT;                                                                              \
    } else {                                                                                                \
        if(MASKED) {                                                                                        \
            output += utf32_to_utf16_masked<big_endian>(byteflip, INPUT, VALID_COUNT, reinterpret_cast<char16_t *>(output));      \
        } else {                                                                                            \
            output += utf32_to_utf16<big_endian>(byteflip, INPUT, VALID_COUNT, reinterpret_cast<char16_t *>(output));             \
        }                                                                                                   \
    }                                                                                                       \
}


#define SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)                                  \
        if (UTF32) {                                                                      \
                const __m128i t0 = _mm512_castsi512_si128(utf8);                          \
                const __m128i t1 = _mm512_extracti32x4_epi32(utf8, 1);                    \
                const __m128i t2 = _mm512_extracti32x4_epi32(utf8, 2);                    \
                const __m128i t3 = _mm512_extracti32x4_epi32(utf8, 3);                    \
                _mm512_storeu_si512((__m512i*)(output + 0*16), _mm512_cvtepu8_epi32(t0)); \
                _mm512_storeu_si512((__m512i*)(output + 1*16), _mm512_cvtepu8_epi32(t1)); \
                _mm512_storeu_si512((__m512i*)(output + 2*16), _mm512_cvtepu8_epi32(t2)); \
                _mm512_storeu_si512((__m512i*)(output + 3*16), _mm512_cvtepu8_epi32(t3)); \
        } else {                                                                          \
                const __m256i h0 = _mm512_castsi512_si256(utf8);                          \
                const __m256i h1 = _mm512_extracti64x4_epi64(utf8, 1);                    \
                if(big_endian) {                                                          \
                _mm512_storeu_si512((__m512i*)(output + 0*16), _mm512_shuffle_epi8(_mm512_cvtepu8_epi16(h0), byteflip)); \
                _mm512_storeu_si512((__m512i*)(output + 2*16), _mm512_shuffle_epi8(_mm512_cvtepu8_epi16(h1), byteflip)); \
                } else {                                                                  \
                _mm512_storeu_si512((__m512i*)(output + 0*16), _mm512_cvtepu8_epi16(h0)); \
                _mm512_storeu_si512((__m512i*)(output + 2*16), _mm512_cvtepu8_epi16(h1)); \
                }                                                                         \
        }
/* end file src/icelake/icelake_macros.inl.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=icelake/icelake_from_valid_utf8.inl.cpp
/* begin file src/icelake/icelake_from_valid_utf8.inl.cpp */
// file included directly

// File contains conversion procedure from VALID UTF-8 strings.

/*
    valid_utf8_to_fixed_length converts a valid UTF-8 string into UTF-32.

    The `OUTPUT` template type decides what to do with UTF-32: store
    it directly or convert into UTF-16 (with AVX512).

    Input:
    - str           - valid UTF-8 string
    - len           - string length
    - out_buffer    - output buffer

    Result:
    - pair.first    - the first unprocessed input byte
    - pair.second   - the first unprocessed output word
*/
template <endianness big_endian, typename OUTPUT>
std::pair<const char*, OUTPUT*> valid_utf8_to_fixed_length(const char* str, size_t len, OUTPUT* dwords) {
    constexpr bool UTF32 = std::is_same<OUTPUT, uint32_t>::value;
    constexpr bool UTF16 = std::is_same<OUTPUT, char16_t>::value;
    static_assert(UTF32 or UTF16, "output type has to be uint32_t (for UTF-32) or char16_t (for UTF-16)");
    static_assert(!(UTF32 and big_endian), "we do not currently support big-endian UTF-32");

    __m512i byteflip = _mm512_setr_epi64(
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809
        );
    const char* ptr = str;
    const char* end = ptr + len;

    OUTPUT* output = dwords;
    /**
     * In the main loop, we consume 64 bytes per iteration,
     * but we access 64 + 4 bytes.
     * We check for ptr + 64 + 64 <= end because
     * we want to be do maskless writes without overruns.
     */
    while (ptr + 64 + 64 <= end) {
        const __m512i utf8 = _mm512_loadu_si512((const __m512i*)ptr);
        const __m512i v_80 = _mm512_set1_epi8(char(0x80));
        const __mmask64 ascii = _mm512_test_epi8_mask(utf8, v_80);
        if(ascii == 0) {
            SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
            output += 64;
            ptr += 64;
            continue;
        }

        const __m512i lane0 = broadcast_epi128<0>(utf8);
        const __m512i lane1 = broadcast_epi128<1>(utf8);
        int valid_count0;
        __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
        const __m512i lane2 = broadcast_epi128<2>(utf8);
        int valid_count1;
        __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
        if(valid_count0 + valid_count1 <= 16) {
            vec0 = _mm512_mask_expand_epi32(vec0, __mmask16(((1<<valid_count1)-1)<<valid_count0), vec1);
            valid_count0 += valid_count1;
            vec0 = expand_utf8_to_utf32(vec0);
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, false)
        } else {
            vec0 = expand_utf8_to_utf32(vec0);
            vec1 = expand_utf8_to_utf32(vec1);
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, false)
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, false)
        }
        const __m512i lane3 = broadcast_epi128<3>(utf8);
        int valid_count2;
        __m512i vec2 = expand_and_identify(lane2, lane3, valid_count2);
        uint32_t tmp1;
        ::memcpy(&tmp1, ptr + 64, sizeof(tmp1));
        const __m512i lane4 = _mm512_set1_epi32(tmp1);
        int valid_count3;
        __m512i vec3 = expand_and_identify(lane3, lane4, valid_count3);
        if(valid_count2 + valid_count3 <= 16) {
            vec2 = _mm512_mask_expand_epi32(vec2, __mmask16(((1<<valid_count3)-1)<<valid_count2), vec3);
            valid_count2 += valid_count3;
            vec2 = expand_utf8_to_utf32(vec2);
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, false)
        } else {
            vec2 = expand_utf8_to_utf32(vec2);
            vec3 = expand_utf8_to_utf32(vec3);
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, false)
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec3, valid_count3, false)
        }
        ptr += 4*16;
    }

    if (ptr + 64 <= end) {
        const __m512i utf8 = _mm512_loadu_si512((const __m512i*)ptr);
        const __m512i v_80 = _mm512_set1_epi8(char(0x80));
        const __mmask64 ascii = _mm512_test_epi8_mask(utf8, v_80);
        if(ascii == 0) {
            SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
            output += 64;
            ptr += 64;
        } else {
            const __m512i lane0 = broadcast_epi128<0>(utf8);
            const __m512i lane1 = broadcast_epi128<1>(utf8);
            int valid_count0;
            __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
            const __m512i lane2 = broadcast_epi128<2>(utf8);
            int valid_count1;
            __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
            if(valid_count0 + valid_count1 <= 16) {
                vec0 = _mm512_mask_expand_epi32(vec0, __mmask16(((1<<valid_count1)-1)<<valid_count0), vec1);
                valid_count0 += valid_count1;
                vec0 = expand_utf8_to_utf32(vec0);
                SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
            } else {
                vec0 = expand_utf8_to_utf32(vec0);
                vec1 = expand_utf8_to_utf32(vec1);
                SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
                SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, true)
            }

            const __m512i lane3 = broadcast_epi128<3>(utf8);
            SIMDUTF_ICELAKE_TRANSCODE16(lane2, lane3, true)

            ptr += 3*16;
        }
    }
    return {ptr, output};
}


using utf8_to_utf16_result = std::pair<const char*, char16_t*>;
/* end file src/icelake/icelake_from_valid_utf8.inl.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=icelake/icelake_utf8_validation.inl.cpp
/* begin file src/icelake/icelake_utf8_validation.inl.cpp */
// file included directly


simdutf_really_inline __m512i check_special_cases(__m512i input, const __m512i prev1) {
  __m512i mask1 = _mm512_setr_epi64(
        0x0202020202020202,
        0x4915012180808080,
        0x0202020202020202,
        0x4915012180808080,
        0x0202020202020202,
        0x4915012180808080,
        0x0202020202020202,
        0x4915012180808080);
    const __m512i v_0f = _mm512_set1_epi8(0x0f);
    __m512i index1 = _mm512_and_si512(_mm512_srli_epi16(prev1, 4), v_0f);

    __m512i byte_1_high = _mm512_shuffle_epi8(mask1, index1);
    __m512i mask2 = _mm512_setr_epi64(
        0xcbcbcb8b8383a3e7,
        0xcbcbdbcbcbcbcbcb,
        0xcbcbcb8b8383a3e7,
        0xcbcbdbcbcbcbcbcb,
        0xcbcbcb8b8383a3e7,
        0xcbcbdbcbcbcbcbcb,
        0xcbcbcb8b8383a3e7,
        0xcbcbdbcbcbcbcbcb);
     __m512i index2 = _mm512_and_si512(prev1, v_0f);

    __m512i byte_1_low = _mm512_shuffle_epi8(mask2, index2);
    __m512i mask3 = _mm512_setr_epi64(
        0x101010101010101,
        0x1010101babaaee6,
        0x101010101010101,
        0x1010101babaaee6,
        0x101010101010101,
        0x1010101babaaee6,
        0x101010101010101,
        0x1010101babaaee6
    );
    __m512i index3 = _mm512_and_si512(_mm512_srli_epi16(input, 4), v_0f);
    __m512i byte_2_high = _mm512_shuffle_epi8(mask3, index3);
    return _mm512_ternarylogic_epi64(byte_1_high, byte_1_low, byte_2_high, 128);
  }

  simdutf_really_inline __m512i check_multibyte_lengths(const __m512i input,
      const __m512i prev_input, const __m512i sc) {
    __m512i prev2 = prev<2>(input, prev_input);
    __m512i prev3 = prev<3>(input, prev_input);
    __m512i is_third_byte  = _mm512_subs_epu8(prev2, _mm512_set1_epi8(0b11100000u-1)); // Only 111_____ will be > 0
    __m512i is_fourth_byte  = _mm512_subs_epu8(prev3, _mm512_set1_epi8(0b11110000u-1)); // Only 1111____ will be > 0
    __m512i is_third_or_fourth_byte = _mm512_or_si512(is_third_byte, is_fourth_byte);
    const __m512i v_7f = _mm512_set1_epi8(char(0x7f));
    is_third_or_fourth_byte = _mm512_adds_epu8(v_7f, is_third_or_fourth_byte);
    // We want to compute (is_third_or_fourth_byte AND v80) XOR sc.
    const __m512i v_80 = _mm512_set1_epi8(char(0x80));
    return _mm512_ternarylogic_epi32(is_third_or_fourth_byte, v_80, sc, 0b1101010);
    //__m512i is_third_or_fourth_byte_mask = _mm512_and_si512(is_third_or_fourth_byte, v_80);
    //return _mm512_xor_si512(is_third_or_fourth_byte_mask, 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.
  //
  simdutf_really_inline __m512i is_incomplete(const __m512i input) {
    // If the previous input's last 3 bytes match this, they're too short (they ended at EOF):
    // ... 1111____ 111_____ 11______
    __m512i max_value = _mm512_setr_epi64(
        0xffffffffffffffff,
        0xffffffffffffffff,
        0xffffffffffffffff,
        0xffffffffffffffff,
        0xffffffffffffffff,
        0xffffffffffffffff,
        0xffffffffffffffff,
        0xbfdfefffffffffff);
    return _mm512_subs_epu8(input, max_value);
  }

  struct avx512_utf8_checker {
    // If this is nonzero, there has been a UTF-8 error.
    __m512i error{};

    // The last input we received
    __m512i prev_input_block{};
    // Whether the last input we received was incomplete (used for ASCII fast path)
    __m512i prev_incomplete{};

    //
    // Check whether the current bytes are valid UTF-8.
    //
    simdutf_really_inline void check_utf8_bytes(const __m512i input, const __m512i 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)
      __m512i prev1 = prev<1>(input, prev_input);
      __m512i sc = check_special_cases(input, prev1);
      this->error = _mm512_or_si512(check_multibyte_lengths(input, prev_input, sc), this->error);
    }

    // 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.
    simdutf_really_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 = _mm512_or_si512(this->error, this->prev_incomplete);
    }

    // returns true if ASCII.
    simdutf_really_inline bool check_next_input(const __m512i input) {
      const __m512i v_80 = _mm512_set1_epi8(char(0x80));
      const __mmask64 ascii = _mm512_test_epi8_mask(input, v_80);
      if(ascii == 0) {
        this->error = _mm512_or_si512(this->error, this->prev_incomplete);
        return true;
      } else {
        this->check_utf8_bytes(input, this->prev_input_block);
        this->prev_incomplete = is_incomplete(input);
        this->prev_input_block = input;
        return false;
      }
    }
    // do not forget to call check_eof!
    simdutf_really_inline bool errors() const {
        return _mm512_test_epi8_mask(this->error, this->error) != 0;
    }

  }; // struct avx512_utf8_checker
/* end file src/icelake/icelake_utf8_validation.inl.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=icelake/icelake_from_utf8.inl.cpp
/* begin file src/icelake/icelake_from_utf8.inl.cpp */
// file included directly

// File contains conversion procedure from possibly invalid UTF-8 strings.

/**
 * Attempts to convert up to len 1-byte words from in (in UTF-8 format) to
 * out.
 * Returns the position of the input and output after the processing is
 * completed. Upon error, the output is set to null.
 */

template <endianness big_endian>
utf8_to_utf16_result fast_avx512_convert_utf8_to_utf16(const char *in, size_t len, char16_t *out) {
  const char *const final_in = in + len;
  bool result = true;
  while (result) {
    if (in + 64 <= final_in) {
        result = process_block_utf8_to_utf16<SIMDUTF_FULL, big_endian>(in, out, final_in - in);
    } else if(in < final_in) {
        result = process_block_utf8_to_utf16<SIMDUTF_TAIL, big_endian>(in, out, final_in - in);
    } else { break; }
  }
  if(!result) { out = nullptr; }
  return std::make_pair(in, out);
}

template <endianness big_endian>
simdutf::result fast_avx512_convert_utf8_to_utf16_with_errors(const char *in, size_t len, char16_t *out) {
  const char *const init_in = in;
  const char16_t *const init_out = out;
  const char *const final_in = in + len;
  bool  result = true;
  while (result) {
    if (in + 64 <= final_in) {
        result = process_block_utf8_to_utf16<SIMDUTF_FULL, big_endian>(in, out, final_in - in);
    } else if(in < final_in) {
        result = process_block_utf8_to_utf16<SIMDUTF_TAIL, big_endian>(in, out, final_in - in);
    } else { break; }
  }
  if(!result) {
    // rewind_and_convert_with_errors will seek a potential error from in onward,
    // with the ability to go back up to in - init_in bytes, and read final_in - in bytes forward.
    simdutf::result res = scalar::utf8_to_utf16::rewind_and_convert_with_errors<big_endian>(in - init_in, in, final_in - in, out);
    res.count += (in - init_in);
    return res;
  } else {
    return simdutf::result(error_code::SUCCESS,out - init_out);
  }
}


template <endianness big_endian, typename OUTPUT>
std::pair<const char*, OUTPUT*> validating_utf8_to_fixed_length(const char* str, size_t len, OUTPUT* dwords) {
    constexpr bool UTF32 = std::is_same<OUTPUT, uint32_t>::value;
    constexpr bool UTF16 = std::is_same<OUTPUT, char16_t>::value;
    static_assert(UTF32 or UTF16, "output type has to be uint32_t (for UTF-32) or char16_t (for UTF-16)");
    static_assert(!(UTF32 and big_endian), "we do not currently support big-endian UTF-32");

    const char* ptr = str;
    const char* end = ptr + len;
    __m512i byteflip = _mm512_setr_epi64(
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809
        );
    OUTPUT* output = dwords;
    avx512_utf8_checker checker{};
    /**
     * In the main loop, we consume 64 bytes per iteration,
     * but we access 64 + 4 bytes.
     * We check for ptr + 64 + 64 <= end because
     * we want to be do maskless writes without overruns.
     */
    while (ptr + 64 + 64 <= end) {
        const __m512i utf8 = _mm512_loadu_si512((const __m512i*)ptr);
        if(checker.check_next_input(utf8)) {
            SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
            output += 64;
            ptr += 64;
            continue;
        }
        const __m512i lane0 = broadcast_epi128<0>(utf8);
        const __m512i lane1 = broadcast_epi128<1>(utf8);
        int valid_count0;
        __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
        const __m512i lane2 = broadcast_epi128<2>(utf8);
        int valid_count1;
        __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
        if(valid_count0 + valid_count1 <= 16) {
            vec0 = _mm512_mask_expand_epi32(vec0, __mmask16(((1<<valid_count1)-1)<<valid_count0), vec1);
            valid_count0 += valid_count1;
            vec0 = expand_utf8_to_utf32(vec0);
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, false)
        } else {
            vec0 = expand_utf8_to_utf32(vec0);
            vec1 = expand_utf8_to_utf32(vec1);
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, false)
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, false)
        }
        const __m512i lane3 = broadcast_epi128<3>(utf8);
        int valid_count2;
        __m512i vec2 = expand_and_identify(lane2, lane3, valid_count2);
        uint32_t tmp1;
        ::memcpy(&tmp1, ptr + 64, sizeof(tmp1));
        const __m512i lane4 = _mm512_set1_epi32(tmp1);
        int valid_count3;
        __m512i vec3 = expand_and_identify(lane3, lane4, valid_count3);
        if(valid_count2 + valid_count3 <= 16) {
            vec2 = _mm512_mask_expand_epi32(vec2, __mmask16(((1<<valid_count3)-1)<<valid_count2), vec3);
            valid_count2 += valid_count3;
            vec2 = expand_utf8_to_utf32(vec2);
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, false)
        } else {
            vec2 = expand_utf8_to_utf32(vec2);
            vec3 = expand_utf8_to_utf32(vec3);
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, false)
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec3, valid_count3, false)
        }
        ptr += 4*16;
    }
    const char* validatedptr = ptr; // validated up to ptr

    // For the final pass, we validate 64 bytes, but we only transcode
    // 3*16 bytes, so we may end up double-validating 16 bytes.
    if (ptr + 64 <= end) {
        const __m512i utf8 = _mm512_loadu_si512((const __m512i*)ptr);
        if(checker.check_next_input(utf8)) {
            SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
            output += 64;
            ptr += 64;
        } else {
            const __m512i lane0 = broadcast_epi128<0>(utf8);
            const __m512i lane1 = broadcast_epi128<1>(utf8);
            int valid_count0;
            __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
            const __m512i lane2 = broadcast_epi128<2>(utf8);
            int valid_count1;
            __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
            if(valid_count0 + valid_count1 <= 16) {
                vec0 = _mm512_mask_expand_epi32(vec0, __mmask16(((1<<valid_count1)-1)<<valid_count0), vec1);
                valid_count0 += valid_count1;
                vec0 = expand_utf8_to_utf32(vec0);
                SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
            } else {
                vec0 = expand_utf8_to_utf32(vec0);
                vec1 = expand_utf8_to_utf32(vec1);
                SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
                SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, true)
            }

            const __m512i lane3 = broadcast_epi128<3>(utf8);
            SIMDUTF_ICELAKE_TRANSCODE16(lane2, lane3, true)

            ptr += 3*16;
        }
        validatedptr += 4*16;
    }
    {
       const __m512i utf8 = _mm512_maskz_loadu_epi8((1ULL<<(end - validatedptr))-1, (const __m512i*)validatedptr);
       checker.check_next_input(utf8);
    }
    checker.check_eof();
    if(checker.errors()) {
        return {ptr, nullptr}; // We found an error.
    }
    return {ptr, output};
}

// Like validating_utf8_to_fixed_length but returns as soon as an error is identified
template <endianness big_endian, typename OUTPUT>
std::tuple<const char*, OUTPUT*, bool> validating_utf8_to_fixed_length_with_constant_checks(const char* str, size_t len, OUTPUT* dwords) {
    constexpr bool UTF32 = std::is_same<OUTPUT, uint32_t>::value;
    constexpr bool UTF16 = std::is_same<OUTPUT, char16_t>::value;
    static_assert(UTF32 or UTF16, "output type has to be uint32_t (for UTF-32) or char16_t (for UTF-16)");
    static_assert(!(UTF32 and big_endian), "we do not currently support big-endian UTF-32");

    const char* ptr = str;
    const char* end = ptr + len;
    __m512i byteflip = _mm512_setr_epi64(
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809
        );
    OUTPUT* output = dwords;
    avx512_utf8_checker checker{};
    /**
     * In the main loop, we consume 64 bytes per iteration,
     * but we access 64 + 4 bytes.
     * We check for ptr + 64 + 64 <= end because
     * we want to be do maskless writes without overruns.
     */
    while (ptr + 64 + 64 <= end) {
        const __m512i utf8 = _mm512_loadu_si512((const __m512i*)ptr);
        if(checker.check_next_input(utf8)) {
            SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
            output += 64;
            ptr += 64;
            continue;
        }
        if(checker.errors()) {
            return {ptr, output, false}; // We found an error.
        }
        const __m512i lane0 = broadcast_epi128<0>(utf8);
        const __m512i lane1 = broadcast_epi128<1>(utf8);
        int valid_count0;
        __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
        const __m512i lane2 = broadcast_epi128<2>(utf8);
        int valid_count1;
        __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
        if(valid_count0 + valid_count1 <= 16) {
            vec0 = _mm512_mask_expand_epi32(vec0, __mmask16(((1<<valid_count1)-1)<<valid_count0), vec1);
            valid_count0 += valid_count1;
            vec0 = expand_utf8_to_utf32(vec0);
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, false)
        } else {
            vec0 = expand_utf8_to_utf32(vec0);
            vec1 = expand_utf8_to_utf32(vec1);
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, false)
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, false)
        }
        const __m512i lane3 = broadcast_epi128<3>(utf8);
        int valid_count2;
        __m512i vec2 = expand_and_identify(lane2, lane3, valid_count2);
        uint32_t tmp1;
        ::memcpy(&tmp1, ptr + 64, sizeof(tmp1));
        const __m512i lane4 = _mm512_set1_epi32(tmp1);
        int valid_count3;
        __m512i vec3 = expand_and_identify(lane3, lane4, valid_count3);
        if(valid_count2 + valid_count3 <= 16) {
            vec2 = _mm512_mask_expand_epi32(vec2, __mmask16(((1<<valid_count3)-1)<<valid_count2), vec3);
            valid_count2 += valid_count3;
            vec2 = expand_utf8_to_utf32(vec2);
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, false)
        } else {
            vec2 = expand_utf8_to_utf32(vec2);
            vec3 = expand_utf8_to_utf32(vec3);
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec2, valid_count2, false)
            SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec3, valid_count3, false)
        }
        ptr += 4*16;
    }
    const char* validatedptr = ptr; // validated up to ptr

    // For the final pass, we validate 64 bytes, but we only transcode
    // 3*16 bytes, so we may end up double-validating 16 bytes.
    if (ptr + 64 <= end) {
        const __m512i utf8 = _mm512_loadu_si512((const __m512i*)ptr);
        if(checker.check_next_input(utf8)) {
            SIMDUTF_ICELAKE_STORE_ASCII(UTF32, utf8, output)
            output += 64;
            ptr += 64;
        } else if(checker.errors()) {
            return {ptr, output, false}; // We found an error.
        } else {
            const __m512i lane0 = broadcast_epi128<0>(utf8);
            const __m512i lane1 = broadcast_epi128<1>(utf8);
            int valid_count0;
            __m512i vec0 = expand_and_identify(lane0, lane1, valid_count0);
            const __m512i lane2 = broadcast_epi128<2>(utf8);
            int valid_count1;
            __m512i vec1 = expand_and_identify(lane1, lane2, valid_count1);
            if(valid_count0 + valid_count1 <= 16) {
                vec0 = _mm512_mask_expand_epi32(vec0, __mmask16(((1<<valid_count1)-1)<<valid_count0), vec1);
                valid_count0 += valid_count1;
                vec0 = expand_utf8_to_utf32(vec0);
                SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
            } else {
                vec0 = expand_utf8_to_utf32(vec0);
                vec1 = expand_utf8_to_utf32(vec1);
                SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec0, valid_count0, true)
                SIMDUTF_ICELAKE_WRITE_UTF16_OR_UTF32(vec1, valid_count1, true)
            }

            const __m512i lane3 = broadcast_epi128<3>(utf8);
            SIMDUTF_ICELAKE_TRANSCODE16(lane2, lane3, true)

            ptr += 3*16;
        }
        validatedptr += 4*16;
    }
    {
       const __m512i utf8 = _mm512_maskz_loadu_epi8((1ULL<<(end - validatedptr))-1, (const __m512i*)validatedptr);
       checker.check_next_input(utf8);
    }
    checker.check_eof();
    if(checker.errors()) {
        return {ptr, output, false}; // We found an error.
    }
    return {ptr, output, true};
}
/* end file src/icelake/icelake_from_utf8.inl.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=icelake/icelake_convert_utf16_to_utf32.inl.cpp
/* begin file src/icelake/icelake_convert_utf16_to_utf32.inl.cpp */
// file included directly

/*
  Returns a pair: the first unprocessed byte from buf and utf32_output
  A scalar routing should carry on the conversion of the tail.
*/
template <endianness big_endian>
std::tuple<const char16_t*, char32_t*, bool> convert_utf16_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) {
  const char16_t* end = buf + len;
  const __m512i v_fc00 = _mm512_set1_epi16((uint16_t)0xfc00);
  const __m512i v_d800 = _mm512_set1_epi16((uint16_t)0xd800);
  const __m512i v_dc00 = _mm512_set1_epi16((uint16_t)0xdc00);
  __mmask32 carry{0};
  const __m512i byteflip = _mm512_setr_epi64(
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809
        );
  while (buf + 32 <= end) {
    // Always safe because buf + 32 <= end so that end - buf >= 32 bytes:
    __m512i in = _mm512_loadu_si512((__m512i*)buf);
    if(big_endian) { in = _mm512_shuffle_epi8(in, byteflip); }

    // H - bitmask for high surrogates
    const __mmask32 H = _mm512_cmpeq_epi16_mask(_mm512_and_si512(in, v_fc00), v_d800);
    // H - bitmask for low surrogates
    const __mmask32 L = _mm512_cmpeq_epi16_mask(_mm512_and_si512(in, v_fc00), v_dc00);

    if ((H|L)) {
      // surrogate pair(s) in a register
      const __mmask32 V = (L ^ (carry | (H << 1)));   // A high surrogate must be followed by low one and a low one must be preceded by a high one.
                                                      // If valid, V should be equal to 0

      if(V == 0) {
        // valid case
        /*
            Input surrogate pair:
            |1101.11aa.aaaa.aaaa|1101.10bb.bbbb.bbbb|
                low surrogate      high surrogate
        */
        /*  1. Expand all words to 32-bit words
            in  |0000.0000.0000.0000.1101.11aa.aaaa.aaaa|0000.0000.0000.0000.1101.10bb.bbbb.bbbb|
        */
        const __m512i first = _mm512_cvtepu16_epi32(_mm512_castsi512_si256(in));
        const __m512i second = _mm512_cvtepu16_epi32(_mm512_extracti32x8_epi32(in,1));

        /*  2. Shift by one 16-bit word to align low surrogates with high surrogates
            in      |0000.0000.0000.0000.1101.11aa.aaaa.aaaa|0000.0000.0000.0000.1101.10bb.bbbb.bbbb|
            shifted |????.????.????.????.????.????.????.????|0000.0000.0000.0000.1101.11aa.aaaa.aaaa|
        */
        const __m512i shifted_first = _mm512_alignr_epi32(second, first, 1);
        const __m512i shifted_second = _mm512_alignr_epi32(_mm512_setzero_si512(), second, 1);

        /*  3. Align all high surrogates in first and second by shifting to the left by 10 bits
            |0000.0000.0000.0000.1101.11aa.aaaa.aaaa|0000.0011.0110.bbbb.bbbb.bb00.0000.0000|
        */
        const __m512i aligned_first = _mm512_mask_slli_epi32(first, (__mmask16)H, first, 10);
        const __m512i aligned_second = _mm512_mask_slli_epi32(second, (__mmask16)(H>>16), second, 10);

        /*  4. Remove surrogate prefixes and add offset 0x10000 by adding in, shifted and constant
            in      |0000.0000.0000.0000.1101.11aa.aaaa.aaaa|0000.0011.0110.bbbb.bbbb.bb00.0000.0000|
            shifted |????.????.????.????.????.????.????.????|0000.0000.0000.0000.1101.11aa.aaaa.aaaa|
            constant|1111.1100.1010.0000.0010.0100.0000.0000|1111.1100.1010.0000.0010.0100.0000.0000|
        */
        const __m512i constant = _mm512_set1_epi32((uint32_t)0xfca02400);
        const __m512i added_first = _mm512_mask_add_epi32(aligned_first, (__mmask16)H, aligned_first, shifted_first);
        const __m512i utf32_first = _mm512_mask_add_epi32(added_first, (__mmask16)H, added_first, constant);

        const __m512i added_second = _mm512_mask_add_epi32(aligned_second, (__mmask16)(H>>16), aligned_second, shifted_second);
        const __m512i utf32_second = _mm512_mask_add_epi32(added_second, (__mmask16)(H>>16), added_second, constant);

        //  5. Store all valid UTF-32 words (low surrogate positions and 32nd word are invalid)
        const __mmask32 valid = ~L & 0x7fffffff;
        // We deliberately do a _mm512_maskz_compress_epi32 followed by storeu_epi32
        // to ease performance portability to Zen 4.
        const __m512i compressed_first = _mm512_maskz_compress_epi32((__mmask16)(valid), utf32_first);
        const size_t howmany1 = count_ones((uint16_t)(valid));
        _mm512_storeu_si512((__m512i *) utf32_output,  compressed_first);
        utf32_output += howmany1;
        const __m512i compressed_second = _mm512_maskz_compress_epi32((__mmask16)(valid >> 16), utf32_second);
        const size_t howmany2 = count_ones((uint16_t)(valid >> 16));
        // The following could be unsafe in some cases?
        //_mm512_storeu_epi32((__m512i *) utf32_output, compressed_second);
        _mm512_mask_storeu_epi32((__m512i *) utf32_output, __mmask16((1<<howmany2)-1), compressed_second);
        utf32_output += howmany2;
        // Only process 31 words, but keep track if the 31st word is a high surrogate as a carry
        buf += 31;
        carry = (H >> 30) & 0x1;
      } else {
        // invalid case
        return std::make_tuple(buf+carry, utf32_output, false);
      }
    } else {
      // no surrogates
      // extend all thirty-two 16-bit words to thirty-two 32-bit words
      _mm512_storeu_si512((__m512i *)(utf32_output), _mm512_cvtepu16_epi32(_mm512_castsi512_si256(in)));
      _mm512_storeu_si512((__m512i *)(utf32_output) + 1, _mm512_cvtepu16_epi32(_mm512_extracti32x8_epi32(in,1)));
      utf32_output += 32;
      buf += 32;
      carry = 0;
    }
  } // while
  return std::make_tuple(buf+carry, utf32_output, true);
}
/* end file src/icelake/icelake_convert_utf16_to_utf32.inl.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=icelake/icelake_convert_utf32_to_utf8.inl.cpp
/* begin file src/icelake/icelake_convert_utf32_to_utf8.inl.cpp */
// file included directly

// Todo: currently, this is just the haswell code, optimize for icelake kernel.
std::pair<const char32_t*, char*> avx512_convert_utf32_to_utf8(const char32_t* buf, size_t len, char* utf8_output) {
  const char32_t* end = buf + len;
  const __m256i v_0000 = _mm256_setzero_si256();
  const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000);
  const __m256i v_ff80 = _mm256_set1_epi16((uint16_t)0xff80);
  const __m256i v_f800 = _mm256_set1_epi16((uint16_t)0xf800);
  const __m256i v_c080 = _mm256_set1_epi16((uint16_t)0xc080);
  const __m256i v_7fffffff = _mm256_set1_epi32((uint32_t)0x7fffffff);
  __m256i running_max = _mm256_setzero_si256();
  __m256i forbidden_bytemask = _mm256_setzero_si256();

  const size_t safety_margin = 12; // to avoid overruns, see issue https://github.com/simdutf/simdutf/issues/92

  while (buf + 16 + safety_margin <= end) {
    __m256i in = _mm256_loadu_si256((__m256i*)buf);
    __m256i nextin = _mm256_loadu_si256((__m256i*)buf+1);
    running_max = _mm256_max_epu32(_mm256_max_epu32(in, running_max), nextin);

    // Pack 32-bit UTF-32 words to 16-bit UTF-16 words with unsigned saturation
    __m256i in_16 = _mm256_packus_epi32(_mm256_and_si256(in, v_7fffffff), _mm256_and_si256(nextin, v_7fffffff));
    in_16 = _mm256_permute4x64_epi64(in_16, 0b11011000);

    // Try to apply UTF-16 => UTF-8 routine on 256 bits (haswell/avx2_convert_utf16_to_utf8.cpp)

    if(_mm256_testz_si256(in_16, v_ff80)) { // ASCII fast path!!!!
      // 1. pack the bytes
      const __m128i utf8_packed = _mm_packus_epi16(_mm256_castsi256_si128(in_16),_mm256_extractf128_si256(in_16,1));
      // 2. store (16 bytes)
      _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);
      // 3. adjust pointers
      buf += 16;
      utf8_output += 16;
      continue; // we are done for this round!
    }
    // no bits set above 7th bit
    const __m256i one_byte_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_ff80), v_0000);
    const uint32_t one_byte_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));

    // no bits set above 11th bit
    const __m256i one_or_two_bytes_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_0000);
    const uint32_t one_or_two_bytes_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(one_or_two_bytes_bytemask));
    if (one_or_two_bytes_bitmask == 0xffffffff) {
      // 1. prepare 2-byte values
      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
      // expected output   : [110a|aaaa|10bb|bbbb] x 8
      const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
      const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);

      // t0 = [000a|aaaa|bbbb|bb00]
      const __m256i t0 = _mm256_slli_epi16(in_16, 2);
      // t1 = [000a|aaaa|0000|0000]
      const __m256i t1 = _mm256_and_si256(t0, v_1f00);
      // t2 = [0000|0000|00bb|bbbb]
      const __m256i t2 = _mm256_and_si256(in_16, v_003f);
      // t3 = [000a|aaaa|00bb|bbbb]
      const __m256i t3 = _mm256_or_si256(t1, t2);
      // t4 = [110a|aaaa|10bb|bbbb]
      const __m256i t4 = _mm256_or_si256(t3, v_c080);

      // 2. merge ASCII and 2-byte codewords
      const __m256i utf8_unpacked = _mm256_blendv_epi8(t4, in_16, one_byte_bytemask);

      // 3. prepare bitmask for 8-bit lookup
      const uint32_t M0 = one_byte_bitmask & 0x55555555;
      const uint32_t M1 = M0 >> 7;
      const uint32_t M2 = (M1 | M0)  & 0x00ff00ff;
      // 4. pack the bytes

      const uint8_t* row = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
      const uint8_t* row_2 = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2>>16)][0];

      const __m128i shuffle = _mm_loadu_si128((__m128i*)(row + 1));
      const __m128i shuffle_2 = _mm_loadu_si128((__m128i*)(row_2 + 1));

      const __m256i utf8_packed = _mm256_shuffle_epi8(utf8_unpacked, _mm256_setr_m128i(shuffle,shuffle_2));
      // 5. store bytes
      _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_packed));
      utf8_output += row[0];
      _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_packed,1));
      utf8_output += row_2[0];

      // 6. adjust pointers
      buf += 16;
      continue;
    }
    // Must check for overflow in packing
    const __m256i saturation_bytemask = _mm256_cmpeq_epi32(_mm256_and_si256(_mm256_or_si256(in, nextin), v_ffff0000), v_0000);
    const uint32_t saturation_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));
    if (saturation_bitmask == 0xffffffff) {
      // case: words from register produce either 1, 2 or 3 UTF-8 bytes
      const __m256i v_d800 = _mm256_set1_epi16((uint16_t)0xd800);
      forbidden_bytemask = _mm256_or_si256(forbidden_bytemask, _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_d800));

      const __m256i dup_even = _mm256_setr_epi16(0x0000, 0x0202, 0x0404, 0x0606,
                                              0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
                                              0x0000, 0x0202, 0x0404, 0x0606,
                                              0x0808, 0x0a0a, 0x0c0c, 0x0e0e);

      /* In this branch we handle three cases:
        1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           - single UFT-8 byte
        2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two UTF-8 bytes
        3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - three UTF-8 bytes

        We expand the input word (16-bit) into two words (32-bit), thus
        we have room for four bytes. However, we need five distinct bit
        layouts. Note that the last byte in cases #2 and #3 is the same.

        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
        in register t2.

        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
        either byte 1 for case #2 or byte 2 for case #3. Note that they
        differ by exactly one bit.

        Finally from these two words we build proper UTF-8 sequence, taking
        into account the case (i.e, the number of bytes to write).
      */
      /**
       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
       * t2 => [0ccc|cccc] [10cc|cccc]
       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
       */
#define simdutf_vec(x) _mm256_set1_epi16(static_cast<uint16_t>(x))
      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
      const __m256i t0 = _mm256_shuffle_epi8(in_16, dup_even);
      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
      const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111));
      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
      const __m256i t2 = _mm256_or_si256 (t1, simdutf_vec(0b1000000000000000));

      // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
      const __m256i s0 = _mm256_srli_epi16(in_16, 4);
      // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
      const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100));
      // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
      const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140));
      // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
      const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000));
      const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask, simdutf_vec(0b0100000000000000));
      const __m256i s4 = _mm256_xor_si256(s3, m0);
#undef simdutf_vec

      // 4. expand words 16-bit => 32-bit
      const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
      const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);

      // 5. compress 32-bit words into 1, 2 or 3 bytes -- 2 x shuffle
      const uint32_t mask = (one_byte_bitmask & 0x55555555) |
                            (one_or_two_bytes_bitmask & 0xaaaaaaaa);
      // Due to the wider registers, the following path is less likely to be useful.
      /*if(mask == 0) {
        // We only have three-byte words. Use fast path.
        const __m256i shuffle = _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1, 2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1);
        const __m256i utf8_0 = _mm256_shuffle_epi8(out0, shuffle);
        const __m256i utf8_1 = _mm256_shuffle_epi8(out1, shuffle);
        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
        utf8_output += 12;
        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
        utf8_output += 12;
        _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_0,1));
        utf8_output += 12;
        _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_1,1));
        utf8_output += 12;
        buf += 16;
        continue;
      }*/
      const uint8_t mask0 = uint8_t(mask);
      const uint8_t* row0 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
      const __m128i shuffle0 = _mm_loadu_si128((__m128i*)(row0 + 1));
      const __m128i utf8_0 = _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);

      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
      const uint8_t* row1 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
      const __m128i shuffle1 = _mm_loadu_si128((__m128i*)(row1 + 1));
      const __m128i utf8_1 = _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);

      const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
      const uint8_t* row2 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
      const __m128i shuffle2 = _mm_loadu_si128((__m128i*)(row2 + 1));
      const __m128i utf8_2 = _mm_shuffle_epi8(_mm256_extractf128_si256(out0,1), shuffle2);


      const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
      const uint8_t* row3 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
      const __m128i shuffle3 = _mm_loadu_si128((__m128i*)(row3 + 1));
      const __m128i utf8_3 = _mm_shuffle_epi8(_mm256_extractf128_si256(out1,1), shuffle3);

      _mm_storeu_si128((__m128i*)utf8_output, utf8_0);
      utf8_output += row0[0];
      _mm_storeu_si128((__m128i*)utf8_output, utf8_1);
      utf8_output += row1[0];
      _mm_storeu_si128((__m128i*)utf8_output, utf8_2);
      utf8_output += row2[0];
      _mm_storeu_si128((__m128i*)utf8_output, utf8_3);
      utf8_output += row3[0];
      buf += 16;
    } else {
      // case: at least one 32-bit word is larger than 0xFFFF <=> it will produce four UTF-8 bytes.
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // may require large, non-trivial tables?
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint32_t word = buf[k];
        if((word & 0xFFFFFF80)==0) {  // 1-byte (ASCII)
          *utf8_output++ = char(word);
        } else if((word & 0xFFFFF800)==0) { // 2-byte
          *utf8_output++ = char((word>>6) | 0b11000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else if((word & 0xFFFF0000 )==0) {  // 3-byte
          if (word >= 0xD800 && word <= 0xDFFF) { return std::make_pair(nullptr, utf8_output); }
          *utf8_output++ = char((word>>12) | 0b11100000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else {  // 4-byte
          if (word > 0x10FFFF) { return std::make_pair(nullptr, utf8_output); }
          *utf8_output++ = char((word>>18) | 0b11110000);
          *utf8_output++ = char(((word>>12) & 0b111111) | 0b10000000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        }
      }
      buf += k;
    }
  } // while

  // check for invalid input
  const __m256i v_10ffff = _mm256_set1_epi32((uint32_t)0x10ffff);
  if(static_cast<uint32_t>(_mm256_movemask_epi8(_mm256_cmpeq_epi32(_mm256_max_epu32(running_max, v_10ffff), v_10ffff))) != 0xffffffff) {
    return std::make_pair(nullptr, utf8_output);
  }

  if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) != 0) { return std::make_pair(nullptr, utf8_output); }

  return std::make_pair(buf, utf8_output);
}

// Todo: currently, this is just the haswell code, optimize for icelake kernel.
std::pair<result, char*> avx512_convert_utf32_to_utf8_with_errors(const char32_t* buf, size_t len, char* utf8_output) {
  const char32_t* end = buf + len;
  const char32_t* start = buf;

  const __m256i v_0000 = _mm256_setzero_si256();
  const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000);
  const __m256i v_ff80 = _mm256_set1_epi16((uint16_t)0xff80);
  const __m256i v_f800 = _mm256_set1_epi16((uint16_t)0xf800);
  const __m256i v_c080 = _mm256_set1_epi16((uint16_t)0xc080);
  const __m256i v_7fffffff = _mm256_set1_epi32((uint32_t)0x7fffffff);
  const __m256i v_10ffff = _mm256_set1_epi32((uint32_t)0x10ffff);

  const size_t safety_margin = 12; // to avoid overruns, see issue https://github.com/simdutf/simdutf/issues/92

  while (buf + 16 + safety_margin <= end) {
    __m256i in = _mm256_loadu_si256((__m256i*)buf);
    __m256i nextin = _mm256_loadu_si256((__m256i*)buf+1);
    // Check for too large input
    const __m256i max_input = _mm256_max_epu32(_mm256_max_epu32(in, nextin), v_10ffff);
    if(static_cast<uint32_t>(_mm256_movemask_epi8(_mm256_cmpeq_epi32(max_input, v_10ffff))) != 0xffffffff) {
      return std::make_pair(result(error_code::TOO_LARGE, buf - start), utf8_output);
    }

    // Pack 32-bit UTF-32 words to 16-bit UTF-16 words with unsigned saturation
    __m256i in_16 = _mm256_packus_epi32(_mm256_and_si256(in, v_7fffffff), _mm256_and_si256(nextin, v_7fffffff));
    in_16 = _mm256_permute4x64_epi64(in_16, 0b11011000);

    // Try to apply UTF-16 => UTF-8 routine on 256 bits (haswell/avx2_convert_utf16_to_utf8.cpp)

    if(_mm256_testz_si256(in_16, v_ff80)) { // ASCII fast path!!!!
      // 1. pack the bytes
      const __m128i utf8_packed = _mm_packus_epi16(_mm256_castsi256_si128(in_16),_mm256_extractf128_si256(in_16,1));
      // 2. store (16 bytes)
      _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);
      // 3. adjust pointers
      buf += 16;
      utf8_output += 16;
      continue; // we are done for this round!
    }
    // no bits set above 7th bit
    const __m256i one_byte_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_ff80), v_0000);
    const uint32_t one_byte_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));

    // no bits set above 11th bit
    const __m256i one_or_two_bytes_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_0000);
    const uint32_t one_or_two_bytes_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(one_or_two_bytes_bytemask));
    if (one_or_two_bytes_bitmask == 0xffffffff) {
      // 1. prepare 2-byte values
      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
      // expected output   : [110a|aaaa|10bb|bbbb] x 8
      const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
      const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);

      // t0 = [000a|aaaa|bbbb|bb00]
      const __m256i t0 = _mm256_slli_epi16(in_16, 2);
      // t1 = [000a|aaaa|0000|0000]
      const __m256i t1 = _mm256_and_si256(t0, v_1f00);
      // t2 = [0000|0000|00bb|bbbb]
      const __m256i t2 = _mm256_and_si256(in_16, v_003f);
      // t3 = [000a|aaaa|00bb|bbbb]
      const __m256i t3 = _mm256_or_si256(t1, t2);
      // t4 = [110a|aaaa|10bb|bbbb]
      const __m256i t4 = _mm256_or_si256(t3, v_c080);

      // 2. merge ASCII and 2-byte codewords
      const __m256i utf8_unpacked = _mm256_blendv_epi8(t4, in_16, one_byte_bytemask);

      // 3. prepare bitmask for 8-bit lookup
      const uint32_t M0 = one_byte_bitmask & 0x55555555;
      const uint32_t M1 = M0 >> 7;
      const uint32_t M2 = (M1 | M0)  & 0x00ff00ff;
      // 4. pack the bytes

      const uint8_t* row = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
      const uint8_t* row_2 = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2>>16)][0];

      const __m128i shuffle = _mm_loadu_si128((__m128i*)(row + 1));
      const __m128i shuffle_2 = _mm_loadu_si128((__m128i*)(row_2 + 1));

      const __m256i utf8_packed = _mm256_shuffle_epi8(utf8_unpacked, _mm256_setr_m128i(shuffle,shuffle_2));
      // 5. store bytes
      _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_packed));
      utf8_output += row[0];
      _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_packed,1));
      utf8_output += row_2[0];

      // 6. adjust pointers
      buf += 16;
      continue;
    }
    // Must check for overflow in packing
    const __m256i saturation_bytemask = _mm256_cmpeq_epi32(_mm256_and_si256(_mm256_or_si256(in, nextin), v_ffff0000), v_0000);
    const uint32_t saturation_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));
    if (saturation_bitmask == 0xffffffff) {
      // case: words from register produce either 1, 2 or 3 UTF-8 bytes

      // Check for illegal surrogate words
      const __m256i v_d800 = _mm256_set1_epi16((uint16_t)0xd800);
      const __m256i forbidden_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_d800);
      if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) != 0x0) {
        return std::make_pair(result(error_code::SURROGATE, buf - start), utf8_output);
      }

      const __m256i dup_even = _mm256_setr_epi16(0x0000, 0x0202, 0x0404, 0x0606,
                                              0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
                                              0x0000, 0x0202, 0x0404, 0x0606,
                                              0x0808, 0x0a0a, 0x0c0c, 0x0e0e);

      /* In this branch we handle three cases:
        1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           - single UFT-8 byte
        2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two UTF-8 bytes
        3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - three UTF-8 bytes

        We expand the input word (16-bit) into two words (32-bit), thus
        we have room for four bytes. However, we need five distinct bit
        layouts. Note that the last byte in cases #2 and #3 is the same.

        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
        in register t2.

        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
        either byte 1 for case #2 or byte 2 for case #3. Note that they
        differ by exactly one bit.

        Finally from these two words we build proper UTF-8 sequence, taking
        into account the case (i.e, the number of bytes to write).
      */
      /**
       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
       * t2 => [0ccc|cccc] [10cc|cccc]
       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
       */
#define simdutf_vec(x) _mm256_set1_epi16(static_cast<uint16_t>(x))
      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
      const __m256i t0 = _mm256_shuffle_epi8(in_16, dup_even);
      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
      const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111));
      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
      const __m256i t2 = _mm256_or_si256 (t1, simdutf_vec(0b1000000000000000));

      // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
      const __m256i s0 = _mm256_srli_epi16(in_16, 4);
      // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
      const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100));
      // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
      const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140));
      // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
      const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000));
      const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask, simdutf_vec(0b0100000000000000));
      const __m256i s4 = _mm256_xor_si256(s3, m0);
#undef simdutf_vec

      // 4. expand words 16-bit => 32-bit
      const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
      const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);

      // 5. compress 32-bit words into 1, 2 or 3 bytes -- 2 x shuffle
      const uint32_t mask = (one_byte_bitmask & 0x55555555) |
                            (one_or_two_bytes_bitmask & 0xaaaaaaaa);
      // Due to the wider registers, the following path is less likely to be useful.
      /*if(mask == 0) {
        // We only have three-byte words. Use fast path.
        const __m256i shuffle = _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1, 2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1);
        const __m256i utf8_0 = _mm256_shuffle_epi8(out0, shuffle);
        const __m256i utf8_1 = _mm256_shuffle_epi8(out1, shuffle);
        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
        utf8_output += 12;
        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
        utf8_output += 12;
        _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_0,1));
        utf8_output += 12;
        _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_1,1));
        utf8_output += 12;
        buf += 16;
        continue;
      }*/
      const uint8_t mask0 = uint8_t(mask);
      const uint8_t* row0 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
      const __m128i shuffle0 = _mm_loadu_si128((__m128i*)(row0 + 1));
      const __m128i utf8_0 = _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);

      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
      const uint8_t* row1 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
      const __m128i shuffle1 = _mm_loadu_si128((__m128i*)(row1 + 1));
      const __m128i utf8_1 = _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);

      const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
      const uint8_t* row2 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
      const __m128i shuffle2 = _mm_loadu_si128((__m128i*)(row2 + 1));
      const __m128i utf8_2 = _mm_shuffle_epi8(_mm256_extractf128_si256(out0,1), shuffle2);


      const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
      const uint8_t* row3 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
      const __m128i shuffle3 = _mm_loadu_si128((__m128i*)(row3 + 1));
      const __m128i utf8_3 = _mm_shuffle_epi8(_mm256_extractf128_si256(out1,1), shuffle3);

      _mm_storeu_si128((__m128i*)utf8_output, utf8_0);
      utf8_output += row0[0];
      _mm_storeu_si128((__m128i*)utf8_output, utf8_1);
      utf8_output += row1[0];
      _mm_storeu_si128((__m128i*)utf8_output, utf8_2);
      utf8_output += row2[0];
      _mm_storeu_si128((__m128i*)utf8_output, utf8_3);
      utf8_output += row3[0];
      buf += 16;
    } else {
      // case: at least one 32-bit word is larger than 0xFFFF <=> it will produce four UTF-8 bytes.
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // may require large, non-trivial tables?
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint32_t word = buf[k];
        if((word & 0xFFFFFF80)==0) {  // 1-byte (ASCII)
          *utf8_output++ = char(word);
        } else if((word & 0xFFFFF800)==0) { // 2-byte
          *utf8_output++ = char((word>>6) | 0b11000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else if((word & 0xFFFF0000 )==0) {  // 3-byte
          if (word >= 0xD800 && word <= 0xDFFF) { return std::make_pair(result(error_code::SURROGATE, buf - start + k), utf8_output); }
          *utf8_output++ = char((word>>12) | 0b11100000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else {  // 4-byte
          if (word > 0x10FFFF) { return std::make_pair(result(error_code::TOO_LARGE, buf - start + k), utf8_output); }
          *utf8_output++ = char((word>>18) | 0b11110000);
          *utf8_output++ = char(((word>>12) & 0b111111) | 0b10000000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        }
      }
      buf += k;
    }
  } // while

  return std::make_pair(result(error_code::SUCCESS, buf - start), utf8_output);
}
/* end file src/icelake/icelake_convert_utf32_to_utf8.inl.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=icelake/icelake_convert_utf32_to_utf16.inl.cpp
/* begin file src/icelake/icelake_convert_utf32_to_utf16.inl.cpp */
// file included directly

// Todo: currently, this is just the haswell code, optimize for icelake kernel.
template <endianness big_endian>
std::pair<const char32_t*, char16_t*> avx512_convert_utf32_to_utf16(const char32_t* buf, size_t len, char16_t* utf16_output) {
  const char32_t* end = buf + len;

  const size_t safety_margin = 12; // to avoid overruns, see issue https://github.com/simdutf/simdutf/issues/92
  __m256i forbidden_bytemask = _mm256_setzero_si256();


  while (buf + 8 + safety_margin <= end) {
    __m256i in = _mm256_loadu_si256((__m256i*)buf);

    const __m256i v_00000000 = _mm256_setzero_si256();
    const __m256i v_ffff0000 = _mm256_set1_epi32((int32_t)0xffff0000);

    // no bits set above 16th bit <=> can pack to UTF16 without surrogate pairs
    const __m256i saturation_bytemask = _mm256_cmpeq_epi32(_mm256_and_si256(in, v_ffff0000), v_00000000);
    const uint32_t saturation_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));

    if (saturation_bitmask == 0xffffffff) {
      const __m256i v_f800 = _mm256_set1_epi32((uint32_t)0xf800);
      const __m256i v_d800 = _mm256_set1_epi32((uint32_t)0xd800);
      forbidden_bytemask = _mm256_or_si256(forbidden_bytemask, _mm256_cmpeq_epi32(_mm256_and_si256(in, v_f800), v_d800));

      __m128i utf16_packed = _mm_packus_epi32(_mm256_castsi256_si128(in),_mm256_extractf128_si256(in,1));
      if (big_endian) {
        const __m128i swap = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
        utf16_packed = _mm_shuffle_epi8(utf16_packed, swap);
      }
      _mm_storeu_si128((__m128i*)utf16_output, utf16_packed);
      utf16_output += 8;
      buf += 8;
    } else {
      size_t forward = 7;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint32_t word = buf[k];
        if((word & 0xFFFF0000)==0) {
          // will not generate a surrogate pair
          if (word >= 0xD800 && word <= 0xDFFF) { return std::make_pair(nullptr, utf16_output); }
          *utf16_output++ = big_endian ? char16_t((uint16_t(word) >> 8) | (uint16_t(word) << 8)) : char16_t(word);
        } else {
          // will generate a surrogate pair
          if (word > 0x10FFFF) { return std::make_pair(nullptr, utf16_output); }
          word -= 0x10000;
          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
          if (big_endian) {
            high_surrogate = uint16_t((high_surrogate >> 8) | (high_surrogate << 8));
            low_surrogate = uint16_t((low_surrogate >> 8) | (low_surrogate << 8));
          }
          *utf16_output++ = char16_t(high_surrogate);
          *utf16_output++ = char16_t(low_surrogate);
        }
      }
      buf += k;
    }
  }

  // check for invalid input
  if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) != 0) { return std::make_pair(nullptr, utf16_output); }

  return std::make_pair(buf, utf16_output);
}

// Todo: currently, this is just the haswell code, optimize for icelake kernel.
template <endianness big_endian>
std::pair<result, char16_t*> avx512_convert_utf32_to_utf16_with_errors(const char32_t* buf, size_t len, char16_t* utf16_output) {
  const char32_t* start = buf;
  const char32_t* end = buf + len;

  const size_t safety_margin = 12; // to avoid overruns, see issue https://github.com/simdutf/simdutf/issues/92

  while (buf + 8 + safety_margin <= end) {
    __m256i in = _mm256_loadu_si256((__m256i*)buf);

    const __m256i v_00000000 = _mm256_setzero_si256();
    const __m256i v_ffff0000 = _mm256_set1_epi32((int32_t)0xffff0000);

    // no bits set above 16th bit <=> can pack to UTF16 without surrogate pairs
    const __m256i saturation_bytemask = _mm256_cmpeq_epi32(_mm256_and_si256(in, v_ffff0000), v_00000000);
    const uint32_t saturation_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));

    if (saturation_bitmask == 0xffffffff) {
      const __m256i v_f800 = _mm256_set1_epi32((uint32_t)0xf800);
      const __m256i v_d800 = _mm256_set1_epi32((uint32_t)0xd800);
      const __m256i forbidden_bytemask = _mm256_cmpeq_epi32(_mm256_and_si256(in, v_f800), v_d800);
      if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) != 0x0) {
        return std::make_pair(result(error_code::SURROGATE, buf - start), utf16_output);
      }

      __m128i utf16_packed = _mm_packus_epi32(_mm256_castsi256_si128(in),_mm256_extractf128_si256(in,1));
      if (big_endian) {
        const __m128i swap = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
        utf16_packed = _mm_shuffle_epi8(utf16_packed, swap);
      }
      _mm_storeu_si128((__m128i*)utf16_output, utf16_packed);
      utf16_output += 8;
      buf += 8;
    } else {
      size_t forward = 7;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint32_t word = buf[k];
        if((word & 0xFFFF0000)==0) {
          // will not generate a surrogate pair
          if (word >= 0xD800 && word <= 0xDFFF) { return std::make_pair(result(error_code::SURROGATE, buf - start + k), utf16_output); }
          *utf16_output++ = big_endian ? char16_t((uint16_t(word) >> 8) | (uint16_t(word) << 8)) : char16_t(word);
        } else {
          // will generate a surrogate pair
          if (word > 0x10FFFF) { return std::make_pair(result(error_code::TOO_LARGE, buf - start + k), utf16_output); }
          word -= 0x10000;
          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
          if (big_endian) {
            high_surrogate = uint16_t((high_surrogate >> 8) | (high_surrogate << 8));
            low_surrogate = uint16_t((low_surrogate >> 8) | (low_surrogate << 8));
          }
          *utf16_output++ = char16_t(high_surrogate);
          *utf16_output++ = char16_t(low_surrogate);
        }
      }
      buf += k;
    }
  }

  return std::make_pair(result(error_code::SUCCESS, buf - start), utf16_output);
}
/* end file src/icelake/icelake_convert_utf32_to_utf16.inl.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=icelake/icelake_ascii_validation.inl.cpp
/* begin file src/icelake/icelake_ascii_validation.inl.cpp */
// file included directly

bool validate_ascii(const char* buf, size_t len) {
  const char* end = buf + len;
  const __m512i ascii = _mm512_set1_epi8((uint8_t)0x80);
  __m512i running_or = _mm512_setzero_si512();
  for (; buf + 64 <= end; buf += 64) {
    const __m512i utf8 = _mm512_loadu_si512((const __m512i*)buf);
    running_or = _mm512_ternarylogic_epi32(running_or, utf8, ascii, 0xf8); // running_or | (utf8 & ascii)
  }
  if(buf < end) {
     const __m512i utf8 = _mm512_maskz_loadu_epi8((uint64_t(1) << (end-buf)) - 1,(const __m512i*)buf);
    running_or = _mm512_ternarylogic_epi32(running_or, utf8, ascii, 0xf8); // running_or | (utf8 & ascii)
  }
  return (_mm512_test_epi8_mask(running_or, running_or) == 0);
}
/* end file src/icelake/icelake_ascii_validation.inl.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=icelake/icelake_utf32_validation.inl.cpp
/* begin file src/icelake/icelake_utf32_validation.inl.cpp */
// file included directly

const char32_t* validate_utf32(const char32_t* buf, size_t len) {
    const char32_t* end = len >= 16 ? buf + len - 16 : nullptr;

    const __m512i offset = _mm512_set1_epi32((uint32_t)0xffff2000);
    __m512i currentmax = _mm512_setzero_si512();
    __m512i currentoffsetmax = _mm512_setzero_si512();

    while (buf <= end) {
      __m512i utf32 = _mm512_loadu_si512((const __m512i*)buf);
      buf += 16;
      currentoffsetmax = _mm512_max_epu32(_mm512_add_epi32(utf32, offset), currentoffsetmax);
      currentmax = _mm512_max_epu32(utf32, currentmax);
    }

    const __m512i standardmax = _mm512_set1_epi32((uint32_t)0x10ffff);
    const __m512i standardoffsetmax = _mm512_set1_epi32((uint32_t)0xfffff7ff);
    __m512i is_zero = _mm512_xor_si512(_mm512_max_epu32(currentmax, standardmax), standardmax);
    if (_mm512_test_epi8_mask(is_zero, is_zero) != 0) {
      return nullptr;
    }
    is_zero = _mm512_xor_si512(_mm512_max_epu32(currentoffsetmax, standardoffsetmax), standardoffsetmax);
    if (_mm512_test_epi8_mask(is_zero, is_zero) != 0) {
      return nullptr;
    }

    return buf;
}
/* end file src/icelake/icelake_utf32_validation.inl.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=icelake/icelake_convert_utf16_to_utf8.inl.cpp
/* begin file src/icelake/icelake_convert_utf16_to_utf8.inl.cpp */
// file included directly

/**
 * This function converts the input (inbuf, inlen), assumed to be valid
 * UTF16 (little endian) into UTF-8 (to outbuf). The number of words written
 * is written to 'outlen' and the function reports the number of input word
 * consumed.
 */
template <endianness big_endian>
size_t utf16_to_utf8_avx512i(const char16_t *inbuf, size_t inlen,
                               unsigned char *outbuf, size_t *outlen) {
  __m512i in;
  __mmask32 inmask = _cvtu32_mask32(0x7fffffff);
  __m512i byteflip = _mm512_setr_epi64(
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809
        );
  const char16_t * const inbuf_orig = inbuf;
  const unsigned char * const outbuf_orig = outbuf;
  size_t adjust = 0;
  int carry = 0;

  while (inlen >= 32) {
    in = _mm512_loadu_si512(inbuf);
    if(big_endian) { in = _mm512_shuffle_epi8(in, byteflip); }
    inlen -= 31;
  lastiteration:
    inbuf += 31;

  failiteration:
    const __mmask32 is234byte = _mm512_mask_cmp_epu16_mask(
      inmask, in, _mm512_set1_epi16(0x0080), _MM_CMPINT_NLT);

    if (_ktestz_mask32_u8(inmask, is234byte)) {
      // fast path for ASCII only
      _mm512_mask_cvtepi16_storeu_epi8(outbuf, inmask, in);
      outbuf += 31;
      carry = 0;

      if (inlen < 32) {
        goto tail;
      } else {
        continue;
      }
    }

    const __mmask32 is12byte =
        _mm512_cmp_epu16_mask(in, _mm512_set1_epi16(0x0800), _MM_CMPINT_LT);

    if (_ktestc_mask32_u8(is12byte, inmask)) {
      // fast path for 1 and 2 byte only

      const __m512i twobytes = _mm512_ternarylogic_epi32(
          _mm512_slli_epi16(in, 8), _mm512_srli_epi16(in, 6),
          _mm512_set1_epi16(0x3f3f), 0xa8); // (A|B)&C
      in = _mm512_mask_add_epi16(in, is234byte, twobytes,
                                 _mm512_set1_epi16(int16_t(0x80c0)));
      const __m512i cmpmask =
          _mm512_mask_blend_epi16(inmask, _mm512_set1_epi16(int16_t(0xffff)),
                                  _mm512_set1_epi16(0x0800));
      const __mmask64 smoosh = _mm512_cmp_epu8_mask(in, cmpmask, _MM_CMPINT_NLT);
      const __m512i out = _mm512_maskz_compress_epi8(smoosh, in);
      _mm512_mask_storeu_epi8(outbuf, _cvtu64_mask64(_pext_u64(_cvtmask64_u64(smoosh), _cvtmask64_u64(smoosh))),
                              out);
      outbuf += 31 + _mm_popcnt_u32(_cvtmask32_u32(is234byte));
      carry = 0;

      if (inlen < 32) {
        goto tail;
      } else {
        continue;
      }
    }
    __m512i lo = _mm512_cvtepu16_epi32(_mm512_castsi512_si256(in));
    __m512i hi = _mm512_cvtepu16_epi32(_mm512_extracti32x8_epi32(in, 1));


    __m512i taglo = _mm512_set1_epi32(0x8080e000);
    __m512i taghi = taglo;

    const __m512i fc00masked = _mm512_and_epi32(in, _mm512_set1_epi16(int16_t(0xfc00)));
    const __mmask32 hisurr = _mm512_mask_cmp_epu16_mask(
        inmask, fc00masked, _mm512_set1_epi16(int16_t(0xd800)), _MM_CMPINT_EQ);
    const __mmask32 losurr = _mm512_cmp_epu16_mask(
        fc00masked, _mm512_set1_epi16(int16_t(0xdc00)), _MM_CMPINT_EQ);

    int carryout = 0;
    if (!_kortestz_mask32_u8(hisurr, losurr)) {
      // handle surrogates

      __m512i los = _mm512_alignr_epi32(hi, lo, 1);
      __m512i his = _mm512_alignr_epi32(lo, hi, 1);

      const __mmask32 hisurrhi = _kshiftri_mask32(hisurr, 16);
      taglo =
          _mm512_mask_mov_epi32(taglo,__mmask16(hisurr), _mm512_set1_epi32(0x808080f0));
      taghi =
          _mm512_mask_mov_epi32(taghi, __mmask16(hisurrhi), _mm512_set1_epi32(0x808080f0));

      lo = _mm512_mask_slli_epi32(lo, __mmask16(hisurr), lo, 10);
      hi = _mm512_mask_slli_epi32(hi, __mmask16(hisurrhi), hi, 10);
      los = _mm512_add_epi32(los, _mm512_set1_epi32(0xfca02400));
      his = _mm512_add_epi32(his, _mm512_set1_epi32(0xfca02400));
      lo = _mm512_mask_add_epi32(lo, __mmask16(hisurr), lo, los);
      hi = _mm512_mask_add_epi32(hi, __mmask16(hisurrhi), hi, his);

      carryout = _cvtu32_mask32(_kshiftri_mask32(hisurr, 30));

      const uint32_t  h = _cvtmask32_u32(hisurr);
      const uint32_t  l = _cvtmask32_u32(losurr);
      // check for mismatched surrogates
      if ((h + h + carry) ^ l) {
        const uint32_t lonohi = l & ~(h + h + carry);
        const uint32_t hinolo = h & ~(l >> 1);
        inlen = _tzcnt_u32(hinolo | lonohi);
        inmask = __mmask32(0x7fffffff & ((1 << inlen) - 1));
        in = _mm512_maskz_mov_epi16(inmask, in);
        adjust = (int)inlen - 31;
        inlen = 0;
        goto failiteration;
      }
    }

    hi = _mm512_maskz_mov_epi32(_cvtu32_mask16(0x7fff),hi);
    carry = carryout;

    __m512i mslo =
        _mm512_multishift_epi64_epi8(_mm512_set1_epi64(0x20262c3200060c12), lo);

    __m512i mshi =
        _mm512_multishift_epi64_epi8(_mm512_set1_epi64(0x20262c3200060c12), hi);

    const __mmask32 outmask = __mmask32(_kandn_mask64(losurr, inmask));
    const __mmask64 outmhi = _kshiftri_mask64(outmask, 16);

    const __mmask32 is1byte = __mmask32(_knot_mask64(is234byte));
    const __mmask64 is1bhi = _kshiftri_mask64(is1byte, 16);
    const __mmask64 is12bhi = _kshiftri_mask64(is12byte, 16);

    taglo =
        _mm512_mask_mov_epi32(taglo, __mmask16(is12byte), _mm512_set1_epi32(0x80c00000));
    taghi =
        _mm512_mask_mov_epi32(taghi, __mmask16(is12bhi), _mm512_set1_epi32(0x80c00000));
    __m512i magiclo = _mm512_mask_blend_epi32(__mmask16(outmask), _mm512_set1_epi32(0xffffffff),
                                      _mm512_set1_epi32(0x00010101));
    __m512i magichi = _mm512_mask_blend_epi32(__mmask16(outmhi), _mm512_set1_epi32(0xffffffff),
                                      _mm512_set1_epi32(0x00010101));


    magiclo = _mm512_mask_blend_epi32(__mmask16(outmask), _mm512_set1_epi32(0xffffffff),
                                      _mm512_set1_epi32(0x00010101));
    magichi = _mm512_mask_blend_epi32(__mmask16(outmhi), _mm512_set1_epi32(0xffffffff),
                                      _mm512_set1_epi32(0x00010101));

    mslo = _mm512_ternarylogic_epi32(mslo, _mm512_set1_epi32(0x3f3f3f3f), taglo,
                                     0xea); // A&B|C
    mshi = _mm512_ternarylogic_epi32(mshi, _mm512_set1_epi32(0x3f3f3f3f), taghi,
                                     0xea);
    mslo = _mm512_mask_slli_epi32(mslo, __mmask16(is1byte), lo, 24);

    mshi = _mm512_mask_slli_epi32(mshi, __mmask16(is1bhi), hi, 24);

    const __mmask64 wantlo = _mm512_cmp_epu8_mask(mslo, magiclo, _MM_CMPINT_NLT);
    const __mmask64 wanthi = _mm512_cmp_epu8_mask(mshi, magichi, _MM_CMPINT_NLT);
    const __m512i outlo = _mm512_maskz_compress_epi8(wantlo, mslo);
    const __m512i outhi = _mm512_maskz_compress_epi8(wanthi, mshi);
    const uint64_t wantlo_uint64 = _cvtmask64_u64(wantlo);
    const uint64_t wanthi_uint64 = _cvtmask64_u64(wanthi);

    uint64_t advlo = _mm_popcnt_u64(wantlo_uint64);
    uint64_t advhi = _mm_popcnt_u64(wanthi_uint64);

    _mm512_mask_storeu_epi8(outbuf, _cvtu64_mask64(_pext_u64(wantlo_uint64, wantlo_uint64)), outlo);
    _mm512_mask_storeu_epi8(outbuf + advlo, _cvtu64_mask64(_pext_u64(wanthi_uint64, wanthi_uint64)), outhi);
    outbuf += advlo + advhi;
  }
  outbuf -= adjust;

tail:
  if (inlen != 0) {
    // We must have inlen < 31.
    inmask = _cvtu32_mask32((1 << inlen) - 1);
    in = _mm512_maskz_loadu_epi16(inmask, inbuf);
    if(big_endian) { in = _mm512_shuffle_epi8(in, byteflip); }
    adjust = inlen - 31;
    inlen = 0;
    goto lastiteration;
  }
  *outlen = (outbuf - outbuf_orig) + adjust;
  return ((inbuf - inbuf_orig) + adjust);
}
/* end file src/icelake/icelake_convert_utf16_to_utf8.inl.cpp */

} // namespace
} // namespace icelake
} // namespace simdutf

namespace simdutf {
namespace icelake {


simdutf_warn_unused int
implementation::detect_encodings(const char *input,
                                 size_t length) const noexcept {
  // If there is a BOM, then we trust it.
  auto bom_encoding = simdutf::BOM::check_bom(input, length);
  if(bom_encoding != encoding_type::unspecified) { return bom_encoding; }
  if (length % 2 == 0) {
    const char *buf = input;

    const char *start = buf;
    const char *end = input + length;

    bool is_utf8 = true;
    bool is_utf16 = true;
    bool is_utf32 = true;

    int out = 0;

    avx512_utf8_checker checker{};
    __m512i currentmax = _mm512_setzero_si512();
    while (buf + 64 <= end) {
      __m512i in = _mm512_loadu_si512((__m512i *)buf);
      __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
      __mmask32 surrogates =
          _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
      if (surrogates) {
        is_utf8 = false;

        // Can still be either UTF-16LE or UTF-32 depending on the positions
        // of the surrogates To be valid UTF-32, a surrogate cannot be in the
        // two most significant bytes of any 32-bit word. On the other hand, to
        // be valid UTF-16LE, at least one surrogate must be in the two most
        // significant bytes of a 32-bit word since they always come in pairs in
        // UTF-16LE. Note that we always proceed in multiple of 4 before this
        // point so there is no offset in 32-bit words.

        if ((surrogates & 0xaaaaaaaa) != 0) {
          is_utf32 = false;
          __mmask32 highsurrogates = _mm512_cmplt_epu16_mask(
              diff, _mm512_set1_epi16(uint16_t(0x0400)));
          __mmask32 lowsurrogates = surrogates ^ highsurrogates;
          // high must be followed by low
          if ((highsurrogates << 1) != lowsurrogates) {
            return simdutf::encoding_type::unspecified;
          }

          bool ends_with_high = ((highsurrogates & 0x80000000) != 0);
          if (ends_with_high) {
            buf +=
                31 *
                sizeof(char16_t); // advance only by 31 words so that we start
                                  // with the high surrogate on the next round.
          } else {
            buf += 32 * sizeof(char16_t);
          }
          is_utf16 = validate_utf16le(reinterpret_cast<const char16_t *>(buf),
                                      (end - buf) / sizeof(char16_t));
          if (!is_utf16) {
            return simdutf::encoding_type::unspecified;

          } else {
            return simdutf::encoding_type::UTF16_LE;
          }

        } else {
          is_utf16 = false;
          // Check for UTF-32
          if (length % 4 == 0) {
            const char32_t *input32 = reinterpret_cast<const char32_t *>(buf);
            const char32_t *end32 =
                reinterpret_cast<const char32_t *>(start) + length / 4;
            if (validate_utf32(input32, end32 - input32)) {
              return simdutf::encoding_type::UTF32_LE;
            }
          }
          return simdutf::encoding_type::unspecified;
        }
        break;
      }
      // If no surrogate, validate under other encodings as well

      // UTF-32 validation
      currentmax = _mm512_max_epu32(in, currentmax);

      // UTF-8 validation
      checker.check_next_input(in);

      buf += 64;
    }

    // Check which encodings are possible

    if (is_utf8) {
      size_t current_length = static_cast<size_t>(buf - start);
      if (current_length != length) {
        const __m512i utf8 = _mm512_maskz_loadu_epi8(
            (1ULL << (length - current_length)) - 1, (const __m512i *)buf);
        checker.check_next_input(utf8);
      }
      checker.check_eof();
      if (!checker.errors()) {
        out |= simdutf::encoding_type::UTF8;
      }
    }

    if (is_utf16 && scalar::utf16::validate<endianness::LITTLE>(
                        reinterpret_cast<const char16_t *>(buf),
                        (length - (buf - start)) / 2)) {
      out |= simdutf::encoding_type::UTF16_LE;
    }

    if (is_utf32 && (length % 4 == 0)) {
      currentmax = _mm512_max_epu32(
          _mm512_maskz_loadu_epi8(
              (1ULL << (length - static_cast<size_t>(buf - start))) - 1,
              (const __m512i *)buf),
          currentmax);
      __mmask16 outside_range = _mm512_cmp_epu32_mask(currentmax, _mm512_set1_epi32(0x10ffff),
                                _MM_CMPINT_GT);
      if (outside_range == 0) {
        out |= simdutf::encoding_type::UTF32_LE;
      }
    }

    return out;
  } else if (implementation::validate_utf8(input, length)) {
    return simdutf::encoding_type::UTF8;
  } else {
    return simdutf::encoding_type::unspecified;
  }
}

simdutf_warn_unused bool implementation::validate_utf8(const char *buf, size_t len) const noexcept {
    avx512_utf8_checker checker{};
    const char* ptr = buf;
    const char* end = ptr + len;
    for (; ptr + 64 <= end; ptr += 64) {
        const __m512i utf8 = _mm512_loadu_si512((const __m512i*)ptr);
        checker.check_next_input(utf8);
    }
    {
       const __m512i utf8 = _mm512_maskz_loadu_epi8((1ULL<<(end - ptr))-1, (const __m512i*)ptr);
       checker.check_next_input(utf8);
    }
    checker.check_eof();
    return ! checker.errors();
}

simdutf_warn_unused result implementation::validate_utf8_with_errors(const char *buf, size_t len) const noexcept {
    avx512_utf8_checker checker{};
    const char* ptr = buf;
    const char* end = ptr + len;
    size_t count{0};
    for (; ptr + 64 <= end; ptr += 64) {
      const __m512i utf8 = _mm512_loadu_si512((const __m512i*)ptr);
      checker.check_next_input(utf8);
      if(checker.errors()) {
        if (count != 0) { count--; } // Sometimes the error is only detected in the next chunk
        result res = scalar::utf8::rewind_and_validate_with_errors(reinterpret_cast<const char*>(buf + count), len - count);
        res.count += count;
        return res;
      }
      count += 64;
    }
    {
      const __m512i utf8 = _mm512_maskz_loadu_epi8((1ULL<<(end - ptr))-1, (const __m512i*)ptr);
      checker.check_next_input(utf8);
      if(checker.errors()) {
        if (count != 0) { count--; } // Sometimes the error is only detected in the next chunk
        result res = scalar::utf8::rewind_and_validate_with_errors(reinterpret_cast<const char*>(buf + count), len - count);
        res.count += count;
        return res;
      } else {
        return result(error_code::SUCCESS, len);
      }
    }
}

simdutf_warn_unused bool implementation::validate_ascii(const char *buf, size_t len) const noexcept {
  return icelake::validate_ascii(buf, len);
}

simdutf_warn_unused result implementation::validate_ascii_with_errors(const char *buf, size_t len) const noexcept {
  const char* buf_orig = buf;
  const char* end = buf + len;
  const __m512i ascii = _mm512_set1_epi8((uint8_t)0x80);
  for (; buf + 64 <= end; buf += 64) {
    const __m512i input = _mm512_loadu_si512((const __m512i*)buf);
    __mmask64 notascii = _mm512_cmp_epu8_mask(input, ascii, _MM_CMPINT_NLT);
    if(notascii) {
      return result(error_code::TOO_LARGE, buf - buf_orig + _tzcnt_u64(notascii));
    }
  }
  {
    const __m512i input = _mm512_maskz_loadu_epi8((1ULL<<(end - buf))-1, (const __m512i*)buf);
    __mmask64 notascii = _mm512_cmp_epu8_mask(input, ascii, _MM_CMPINT_NLT);
    if(notascii) {
      return result(error_code::TOO_LARGE, buf - buf_orig + _tzcnt_u64(notascii));
    }
  }
  return result(error_code::SUCCESS, len);
}

simdutf_warn_unused bool implementation::validate_utf16le(const char16_t *buf, size_t len) const noexcept {
    const char16_t *end = buf + len;

    for(;buf + 32 <= end; ) {
      __m512i in = _mm512_loadu_si512((__m512i*)buf);
      __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
      __mmask32 surrogates = _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
      if(surrogates) {
        __mmask32 highsurrogates = _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
        __mmask32 lowsurrogates = surrogates ^ highsurrogates;
        // high must be followed by low
        if ((highsurrogates << 1) != lowsurrogates) {
           return false;
        }
        bool ends_with_high = ((highsurrogates & 0x80000000) != 0);
        if(ends_with_high) {
          buf += 31; // advance only by 31 words so that we start with the high surrogate on the next round.
        } else {
          buf += 32;
        }
      } else {
        buf += 32;
      }
    }
    if(buf < end) {
      __m512i in = _mm512_maskz_loadu_epi16((1<<(end-buf))-1,(__m512i*)buf);
      __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
      __mmask32 surrogates = _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
      if(surrogates) {
        __mmask32 highsurrogates = _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
        __mmask32 lowsurrogates = surrogates ^ highsurrogates;
        // high must be followed by low
        if ((highsurrogates << 1) != lowsurrogates) {
           return false;
        }
      }
    }
    return true;
}

simdutf_warn_unused bool implementation::validate_utf16be(const char16_t *buf, size_t len) const noexcept {
   const char16_t *end = buf + len;
   const __m512i byteflip = _mm512_setr_epi64(
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809
        );
    for(;buf + 32 <= end; ) {
      __m512i in = _mm512_shuffle_epi8(_mm512_loadu_si512((__m512i*)buf), byteflip);
      __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
      __mmask32 surrogates = _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
      if(surrogates) {
        __mmask32 highsurrogates = _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
        __mmask32 lowsurrogates = surrogates ^ highsurrogates;
        // high must be followed by low
        if ((highsurrogates << 1) != lowsurrogates) {
           return false;
        }
        bool ends_with_high = ((highsurrogates & 0x80000000) != 0);
        if(ends_with_high) {
          buf += 31; // advance only by 31 words so that we start with the high surrogate on the next round.
        } else {
          buf += 32;
        }
      } else {
        buf += 32;
      }
    }
    if(buf < end) {
      __m512i in = _mm512_shuffle_epi8(_mm512_maskz_loadu_epi16((1<<(end-buf))-1,(__m512i*)buf), byteflip);
      __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
      __mmask32 surrogates = _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
      if(surrogates) {
        __mmask32 highsurrogates = _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
        __mmask32 lowsurrogates = surrogates ^ highsurrogates;
        // high must be followed by low
        if ((highsurrogates << 1) != lowsurrogates) {
           return false;
        }
      }
    }
    return true;
}

simdutf_warn_unused result implementation::validate_utf16le_with_errors(const char16_t *buf, size_t len) const noexcept {
    const char16_t *start_buf = buf;
    const char16_t *end = buf + len;
    for(;buf + 32 <= end; ) {
      __m512i in = _mm512_loadu_si512((__m512i*)buf);
      __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
      __mmask32 surrogates = _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
      if(surrogates) {
        __mmask32 highsurrogates = _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
        __mmask32 lowsurrogates = surrogates ^ highsurrogates;
        // high must be followed by low
        if ((highsurrogates << 1) != lowsurrogates) {
          uint32_t extra_low = _tzcnt_u32(lowsurrogates &~(highsurrogates << 1));
          uint32_t extra_high = _tzcnt_u32(highsurrogates &~(lowsurrogates >> 1));
          return result(error_code::SURROGATE, (buf - start_buf) + (extra_low < extra_high ? extra_low : extra_high));
        }
        bool ends_with_high = ((highsurrogates & 0x80000000) != 0);
        if(ends_with_high) {
          buf += 31; // advance only by 31 words so that we start with the high surrogate on the next round.
        } else {
          buf += 32;
        }
      } else {
        buf += 32;
      }
    }
    if(buf < end) {
      __m512i in = _mm512_maskz_loadu_epi16((1<<(end-buf))-1,(__m512i*)buf);
      __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
      __mmask32 surrogates = _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
      if(surrogates) {
        __mmask32 highsurrogates = _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
        __mmask32 lowsurrogates = surrogates ^ highsurrogates;
        // high must be followed by low
        if ((highsurrogates << 1) != lowsurrogates) {
          uint32_t extra_low = _tzcnt_u32(lowsurrogates &~(highsurrogates << 1));
          uint32_t extra_high = _tzcnt_u32(highsurrogates &~(lowsurrogates >> 1));
          return result(error_code::SURROGATE, (buf - start_buf) + (extra_low < extra_high ? extra_low : extra_high));
        }
      }
    }
    return result(error_code::SUCCESS, len);
}

simdutf_warn_unused result implementation::validate_utf16be_with_errors(const char16_t *buf, size_t len) const noexcept {
    const char16_t *start_buf = buf;
    const char16_t *end = buf + len;
    const __m512i byteflip = _mm512_setr_epi64(
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809
        );
    for(;buf + 32 <= end; ) {
      __m512i in = _mm512_shuffle_epi8(_mm512_loadu_si512((__m512i*)buf), byteflip);
      __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
      __mmask32 surrogates = _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
      if(surrogates) {
        __mmask32 highsurrogates = _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
        __mmask32 lowsurrogates = surrogates ^ highsurrogates;
        // high must be followed by low
        if ((highsurrogates << 1) != lowsurrogates) {
          uint32_t extra_low = _tzcnt_u32(lowsurrogates &~(highsurrogates << 1));
          uint32_t extra_high = _tzcnt_u32(highsurrogates &~(lowsurrogates >> 1));
          return result(error_code::SURROGATE, (buf - start_buf) + (extra_low < extra_high ? extra_low : extra_high));
        }
        bool ends_with_high = ((highsurrogates & 0x80000000) != 0);
        if(ends_with_high) {
          buf += 31; // advance only by 31 words so that we start with the high surrogate on the next round.
        } else {
          buf += 32;
        }
      } else {
        buf += 32;
      }
    }
    if(buf < end) {
      __m512i in = _mm512_shuffle_epi8(_mm512_maskz_loadu_epi16((1<<(end-buf))-1,(__m512i*)buf), byteflip);
      __m512i diff = _mm512_sub_epi16(in, _mm512_set1_epi16(uint16_t(0xD800)));
      __mmask32 surrogates = _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0800)));
      if(surrogates) {
        __mmask32 highsurrogates = _mm512_cmplt_epu16_mask(diff, _mm512_set1_epi16(uint16_t(0x0400)));
        __mmask32 lowsurrogates = surrogates ^ highsurrogates;
        // high must be followed by low
        if ((highsurrogates << 1) != lowsurrogates) {
          uint32_t extra_low = _tzcnt_u32(lowsurrogates &~(highsurrogates << 1));
          uint32_t extra_high = _tzcnt_u32(highsurrogates &~(lowsurrogates >> 1));
          return result(error_code::SURROGATE, (buf - start_buf) + (extra_low < extra_high ? extra_low : extra_high));
        }
      }
    }
    return result(error_code::SUCCESS, len);
}

simdutf_warn_unused bool implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
  const char32_t * tail = icelake::validate_utf32(buf, len);
  if (tail) {
    return scalar::utf32::validate(tail, len - (tail - buf));
  } else {
    return false;
  }
}

simdutf_warn_unused result implementation::validate_utf32_with_errors(const char32_t *buf, size_t len) const noexcept {

    const char32_t* end = len >= 16 ? buf + len - 16 : nullptr;
    const char32_t* buf_orig = buf;
    while (buf <= end) {
      __m512i utf32 = _mm512_loadu_si512((const __m512i*)buf);
      __mmask16 outside_range = _mm512_cmp_epu32_mask(utf32, _mm512_set1_epi32(0x10ffff),
                                _MM_CMPINT_GT);
      if (outside_range) {
        return result(error_code::TOO_LARGE, buf - buf_orig + _tzcnt_u32(outside_range));
      }

      __m512i utf32_off = _mm512_add_epi32(utf32, _mm512_set1_epi32(0xffff2000));

      __mmask16 surrogate_range = _mm512_cmp_epu32_mask(utf32_off, _mm512_set1_epi32(0xfffff7ff),
                                _MM_CMPINT_GT);
      if (surrogate_range) {
        return result(error_code::SURROGATE, buf - buf_orig + _tzcnt_u32(surrogate_range));
      }
      buf += 16;
    }
    if(buf < buf_orig + len) {
      __m512i utf32 = _mm512_maskz_loadu_epi32(__mmask16((1<<(buf_orig + len - buf))-1),(const __m512i*)buf);
      __mmask16 outside_range = _mm512_cmp_epu32_mask(utf32, _mm512_set1_epi32(0x10ffff),
                                _MM_CMPINT_GT);
      if (outside_range) {
        return result(error_code::TOO_LARGE, buf - buf_orig + _tzcnt_u32(outside_range));
      }
      __m512i utf32_off = _mm512_add_epi32(utf32, _mm512_set1_epi32(0xffff2000));

      __mmask16 surrogate_range = _mm512_cmp_epu32_mask(utf32_off, _mm512_set1_epi32(0xfffff7ff),
                                _MM_CMPINT_GT);
      if (surrogate_range) {
        return result(error_code::SURROGATE, buf - buf_orig + _tzcnt_u32(surrogate_range));
      }
    }

    return result(error_code::SUCCESS, len);
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
  utf8_to_utf16_result ret = fast_avx512_convert_utf8_to_utf16<endianness::LITTLE>(buf, len, utf16_output);
  if (ret.second == nullptr) {
    return 0;
  }
  return ret.second - utf16_output;
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf16be(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
  utf8_to_utf16_result ret = fast_avx512_convert_utf8_to_utf16<endianness::BIG>(buf, len, utf16_output);
  if (ret.second == nullptr) {
    return 0;
  }
  return ret.second - utf16_output;
}

simdutf_warn_unused result implementation::convert_utf8_to_utf16le_with_errors(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
   return fast_avx512_convert_utf8_to_utf16_with_errors<endianness::LITTLE>(buf, len, utf16_output);
}

simdutf_warn_unused result implementation::convert_utf8_to_utf16be_with_errors(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
   return fast_avx512_convert_utf8_to_utf16_with_errors<endianness::BIG>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16le(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
  utf8_to_utf16_result ret = icelake::valid_utf8_to_fixed_length<endianness::LITTLE, char16_t>(buf, len, utf16_output);
  size_t saved_bytes = ret.second - utf16_output;
  const char* end = buf + len;
  if (ret.first == end) {
    return saved_bytes;
  }

  // Note: AVX512 procedure looks up 4 bytes forward, and
  //       correctly converts multi-byte chars even if their
  //       continuation bytes lie outsiede 16-byte window.
  //       It meas, we have to skip continuation bytes from
  //       the beginning ret.first, as they were already consumed.
  while (ret.first != end && ((uint8_t(*ret.first) & 0xc0) == 0x80)) {
      ret.first += 1;
  }

  if (ret.first != end) {
    const size_t scalar_saved_bytes = scalar::utf8_to_utf16::convert_valid<endianness::LITTLE>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }

  return saved_bytes;
}

simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
  utf8_to_utf16_result ret = icelake::valid_utf8_to_fixed_length<endianness::BIG, char16_t>(buf, len, utf16_output);
  size_t saved_bytes = ret.second - utf16_output;
  const char* end = buf + len;
  if (ret.first == end) {
    return saved_bytes;
  }

  // Note: AVX512 procedure looks up 4 bytes forward, and
  //       correctly converts multi-byte chars even if their
  //       continuation bytes lie outsiede 16-byte window.
  //       It meas, we have to skip continuation bytes from
  //       the beginning ret.first, as they were already consumed.
  while (ret.first != end && ((uint8_t(*ret.first) & 0xc0) == 0x80)) {
      ret.first += 1;
  }

  if (ret.first != end) {
    const size_t scalar_saved_bytes = scalar::utf8_to_utf16::convert_valid<endianness::BIG>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }

  return saved_bytes;
}


simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(const char* buf, size_t len, char32_t* utf32_out) const noexcept {
  uint32_t * utf32_output = reinterpret_cast<uint32_t *>(utf32_out);
  utf8_to_utf32_result ret = icelake::validating_utf8_to_fixed_length<endianness::LITTLE, uint32_t>(buf, len, utf32_output);
  if (ret.second == nullptr)
    return 0;

  size_t saved_bytes = ret.second - utf32_output;
  const char* end = buf + len;
  if (ret.first == end) {
    return saved_bytes;
  }

  // Note: the AVX512 procedure looks up 4 bytes forward, and
  //       correctly converts multi-byte chars even if their
  //       continuation bytes lie outside 16-byte window.
  //       It means, we have to skip continuation bytes from
  //       the beginning ret.first, as they were already consumed.
  while (ret.first != end and ((uint8_t(*ret.first) & 0xc0) == 0x80)) {
      ret.first += 1;
  }

  if (ret.first != end) {
    const size_t scalar_saved_bytes = scalar::utf8_to_utf32::convert(
                                        ret.first, len - (ret.first - buf), utf32_out + saved_bytes);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }

  return saved_bytes;
}

simdutf_warn_unused result implementation::convert_utf8_to_utf32_with_errors(const char* buf, size_t len, char32_t* utf32) const noexcept {
  uint32_t * utf32_output = reinterpret_cast<uint32_t *>(utf32);
  auto ret = icelake::validating_utf8_to_fixed_length_with_constant_checks<endianness::LITTLE, uint32_t>(buf, len, utf32_output);
  if (!std::get<2>(ret)) {
    auto new_buf = std::get<0>(ret);
    // rewind_and_convert_with_errors will seek a potential error from new_buf onward,
    // with the ability to go back up to new_buf - buf bytes, and read len - (new_buf - buf) bytes forward.
    result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(new_buf - buf, new_buf, len - (new_buf - buf), reinterpret_cast<char32_t *>(std::get<1>(ret)));
    res.count += (std::get<0>(ret) - buf);
    return res;
  }
  size_t saved_bytes = std::get<1>(ret) - utf32_output;
  const char* end = buf + len;
  if (std::get<0>(ret) == end) {
    return {simdutf::SUCCESS, saved_bytes};
  }

  // Note: the AVX512 procedure looks up 4 bytes forward, and
  //       correctly converts multi-byte chars even if their
  //       continuation bytes lie outside 16-byte window.
  //       It means, we have to skip continuation bytes from
  //       the beginning ret.first, as they were already consumed.
  while (std::get<0>(ret) != end and ((uint8_t(*std::get<0>(ret)) & 0xc0) == 0x80)) {
      std::get<0>(ret) += 1;
  }

  if (std::get<0>(ret) != end) {
    auto scalar_result = scalar::utf8_to_utf32::convert_with_errors(
                                        std::get<0>(ret), len - (std::get<0>(ret) - buf), reinterpret_cast<char32_t *>(utf32_output) + saved_bytes);
    if (scalar_result.error != simdutf::SUCCESS) {
      scalar_result.count +=  (std::get<0>(ret) - buf);
    } else {
      scalar_result.count += saved_bytes;
    }
    return scalar_result;
  }

  return {simdutf::SUCCESS, size_t(std::get<1>(ret) - utf32_output)};
}


simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(const char* buf, size_t len, char32_t* utf32_out) const noexcept {
  uint32_t * utf32_output = reinterpret_cast<uint32_t *>(utf32_out);
  utf8_to_utf32_result ret = icelake::valid_utf8_to_fixed_length<endianness::LITTLE, uint32_t>(buf, len, utf32_output);
  size_t saved_bytes = ret.second - utf32_output;
  const char* end = buf + len;
  if (ret.first == end) {
    return saved_bytes;
  }

  // Note: AVX512 procedure looks up 4 bytes forward, and
  //       correctly converts multi-byte chars even if their
  //       continuation bytes lie outsiede 16-byte window.
  //       It meas, we have to skip continuation bytes from
  //       the beginning ret.first, as they were already consumed.
  while (ret.first != end && ((uint8_t(*ret.first) & 0xc0) == 0x80)) {
      ret.first += 1;
  }

  if (ret.first != end) {
    const size_t scalar_saved_bytes = scalar::utf8_to_utf32::convert_valid(
                                        ret.first, len - (ret.first - buf), utf32_out + saved_bytes);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }

  return saved_bytes;
}

simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  size_t outlen;
  size_t inlen = utf16_to_utf8_avx512i<endianness::LITTLE>(buf, len, (unsigned char*)utf8_output, &outlen);
  if(inlen != len) { return 0; }
  return outlen;
}

simdutf_warn_unused size_t implementation::convert_utf16be_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  size_t outlen;
  size_t inlen = utf16_to_utf8_avx512i<endianness::BIG>(buf, len, (unsigned char*)utf8_output, &outlen);
  if(inlen != len) { return 0; }
  return outlen;
}

simdutf_warn_unused result implementation::convert_utf16le_to_utf8_with_errors(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  size_t outlen;
  size_t inlen = utf16_to_utf8_avx512i<endianness::LITTLE>(buf, len, (unsigned char*)utf8_output, &outlen);
  if(inlen != len) {
    result res = scalar::utf16_to_utf8::convert_with_errors<endianness::LITTLE>(buf + inlen, len - outlen, utf8_output + outlen);
    res.count += inlen;
    return res;
  }
  return {simdutf::SUCCESS, outlen};
}

simdutf_warn_unused result implementation::convert_utf16be_to_utf8_with_errors(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  size_t outlen;
  size_t inlen = utf16_to_utf8_avx512i<endianness::BIG>(buf, len, (unsigned char*)utf8_output, &outlen);
  if(inlen != len) {
    result res = scalar::utf16_to_utf8::convert_with_errors<endianness::BIG>(buf + inlen, len - outlen, utf8_output + outlen);
    res.count += inlen;
    return res;
  }
  return {simdutf::SUCCESS, outlen};
}

simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return convert_utf16le_to_utf8(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return convert_utf16be_to_utf8(buf, len, utf8_output);
}


simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  std::pair<const char32_t*, char*> ret = avx512_convert_utf32_to_utf8(buf, len, utf8_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf8_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf32_to_utf8::convert(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char*> ret = icelake::avx512_convert_utf32_to_utf8_with_errors(buf, len, utf8_output);
  if (ret.first.count != len) {
    result scalar_res = scalar::utf32_to_utf8::convert_with_errors(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf8_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  return convert_utf32_to_utf8(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  std::pair<const char32_t*, char16_t*> ret = avx512_convert_utf32_to_utf16<endianness::LITTLE>(buf, len, utf16_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf16_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf32_to_utf16::convert<endianness::LITTLE>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf16be(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  std::pair<const char32_t*, char16_t*> ret = avx512_convert_utf32_to_utf16<endianness::BIG>(buf, len, utf16_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf16_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf32_to_utf16::convert<endianness::BIG>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused result implementation::convert_utf32_to_utf16le_with_errors(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char16_t*> ret = avx512_convert_utf32_to_utf16_with_errors<endianness::LITTLE>(buf, len, utf16_output);
  if (ret.first.count != len) {
    result scalar_res = scalar::utf32_to_utf16::convert_with_errors<endianness::LITTLE>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf16_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused result implementation::convert_utf32_to_utf16be_with_errors(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char16_t*> ret = avx512_convert_utf32_to_utf16_with_errors<endianness::BIG>(buf, len, utf16_output);
  if (ret.first.count != len) {
    result scalar_res = scalar::utf32_to_utf16::convert_with_errors<endianness::BIG>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf16_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16le(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return convert_utf32_to_utf16le(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16be(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return convert_utf32_to_utf16be(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  std::tuple<const char16_t*, char32_t*, bool> ret = icelake::convert_utf16_to_utf32<endianness::LITTLE>(buf, len, utf32_output);
  if (!std::get<2>(ret)) { return 0; }
  size_t saved_bytes = std::get<1>(ret) - utf32_output;
  if (std::get<0>(ret) != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf16_to_utf32::convert<endianness::LITTLE>(
                                        std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused size_t implementation::convert_utf16be_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  std::tuple<const char16_t*, char32_t*, bool> ret = icelake::convert_utf16_to_utf32<endianness::BIG>(buf, len, utf32_output);
  if (!std::get<2>(ret)) { return 0; }
  size_t saved_bytes = std::get<1>(ret) - utf32_output;
  if (std::get<0>(ret) != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf16_to_utf32::convert<endianness::BIG>(
                                        std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused result implementation::convert_utf16le_to_utf32_with_errors(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  std::tuple<const char16_t*, char32_t*, bool> ret = icelake::convert_utf16_to_utf32<endianness::LITTLE>(buf, len, utf32_output);
  if (!std::get<2>(ret)) {
    result scalar_res = scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(
                                        std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
    scalar_res.count += (std::get<0>(ret) - buf);
    return scalar_res;
  }
  size_t saved_bytes = std::get<1>(ret) - utf32_output;
  if (std::get<0>(ret) != buf + len) {
    result scalar_res = scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(
                                        std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
    if (scalar_res.error) {
      scalar_res.count += (std::get<0>(ret) - buf);
      return scalar_res;
    } else {
      scalar_res.count += saved_bytes;
      return scalar_res;
    }
  }
  return simdutf::result(simdutf::SUCCESS, saved_bytes);
}

simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  std::tuple<const char16_t*, char32_t*, bool> ret = icelake::convert_utf16_to_utf32<endianness::BIG>(buf, len, utf32_output);
  if (!std::get<2>(ret)) {
    result scalar_res = scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(
                                        std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
    scalar_res.count += (std::get<0>(ret) - buf);
    return scalar_res;
  }
  size_t saved_bytes = std::get<1>(ret) - utf32_output;
  if (std::get<0>(ret) != buf + len) {
    result scalar_res = scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(
                                        std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
    if (scalar_res.error) {
      scalar_res.count += (std::get<0>(ret) - buf);
      return scalar_res;
    } else {
      scalar_res.count += saved_bytes;
      return scalar_res;
    }
  }
  return simdutf::result(simdutf::SUCCESS, saved_bytes);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  std::tuple<const char16_t*, char32_t*, bool> ret = icelake::convert_utf16_to_utf32<endianness::LITTLE>(buf, len, utf32_output);
  if (!std::get<2>(ret)) { return 0; }
  size_t saved_bytes = std::get<1>(ret) - utf32_output;
  if (std::get<0>(ret) != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf16_to_utf32::convert<endianness::LITTLE>(
                                        std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  std::tuple<const char16_t*, char32_t*, bool> ret = icelake::convert_utf16_to_utf32<endianness::BIG>(buf, len, utf32_output);
  if (!std::get<2>(ret)) { return 0; }
  size_t saved_bytes = std::get<1>(ret) - utf32_output;
  if (std::get<0>(ret) != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf16_to_utf32::convert<endianness::BIG>(
                                        std::get<0>(ret), len - (std::get<0>(ret) - buf), std::get<1>(ret));
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

void implementation::change_endianness_utf16(const char16_t * input, size_t length, char16_t * output) const noexcept {
  size_t pos = 0;
  const __m512i byteflip = _mm512_setr_epi64(
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809
        );
  while (pos + 32 <= length) {
    __m512i utf16 = _mm512_loadu_si512((const __m512i*)(input + pos));
    utf16 = _mm512_shuffle_epi8(utf16, byteflip);
    _mm512_storeu_si512(output + pos, utf16);
    pos += 32;
  }
  if(pos < length) {
    __mmask32 m((1<< (length - pos))-1);
    __m512i utf16 = _mm512_maskz_loadu_epi16(m, (const __m512i*)(input + pos));
    utf16 = _mm512_shuffle_epi8(utf16, byteflip);
    _mm512_mask_storeu_epi16(output + pos, m, utf16);
  }
}


simdutf_warn_unused size_t implementation::count_utf16le(const char16_t * input, size_t length) const noexcept {
  const char16_t* end = length >= 32 ? input + length - 32 : nullptr;
  const char16_t* ptr = input;

  const __m512i low = _mm512_set1_epi16((uint16_t)0xdc00);
  const __m512i high = _mm512_set1_epi16((uint16_t)0xdfff);

  size_t count{0};

  while (ptr <= end) {
    __m512i utf16 = _mm512_loadu_si512((const __m512i*)ptr);
    ptr += 32;
    uint64_t not_high_surrogate = static_cast<uint64_t>(_mm512_cmpgt_epu16_mask(utf16, high) | _mm512_cmplt_epu16_mask(utf16, low));
    count += count_ones(not_high_surrogate);
  }

  return count + scalar::utf16::count_code_points<endianness::LITTLE>(ptr, length - (ptr - input));
}

simdutf_warn_unused size_t implementation::count_utf16be(const char16_t * input, size_t length) const noexcept {
  const char16_t* end = length >= 32 ? input + length - 32 : nullptr;
  const char16_t* ptr = input;

  const __m512i low = _mm512_set1_epi16((uint16_t)0xdc00);
  const __m512i high = _mm512_set1_epi16((uint16_t)0xdfff);

  size_t count{0};
  const __m512i byteflip = _mm512_setr_epi64(
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809
        );
  while (ptr <= end) {
    __m512i utf16 = _mm512_shuffle_epi8(_mm512_loadu_si512((__m512i*)ptr), byteflip);
    ptr += 32;
    uint64_t not_high_surrogate = static_cast<uint64_t>(_mm512_cmpgt_epu16_mask(utf16, high) | _mm512_cmplt_epu16_mask(utf16, low));
    count += count_ones(not_high_surrogate);
  }

  return count + scalar::utf16::count_code_points<endianness::BIG>(ptr, length - (ptr - input));
}


simdutf_warn_unused size_t implementation::count_utf8(const char * input, size_t length) const noexcept {
  const char* end = length >= 64 ? input + length - 64 : nullptr;
  const char* ptr = input;

  const __m512i continuation = _mm512_set1_epi8(char(0b10111111));

  size_t count{0};

  while (ptr <= end) {
    __m512i utf8 = _mm512_loadu_si512((const __m512i*)ptr);
    ptr += 64;
    uint64_t continuation_bitmask = static_cast<uint64_t>(_mm512_cmple_epi8_mask(utf8, continuation));
    count += 64 - count_ones(continuation_bitmask);
  }

  return count + scalar::utf8::count_code_points(ptr, length - (ptr - input));
}


simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(const char16_t * input, size_t length) const noexcept {
  const char16_t* end = length >= 32 ? input + length - 32 : nullptr;
  const char16_t* ptr = input;

  const __m512i v_007f = _mm512_set1_epi16((uint16_t)0x007f);
  const __m512i v_07ff = _mm512_set1_epi16((uint16_t)0x07ff);
  const __m512i v_dfff = _mm512_set1_epi16((uint16_t)0xdfff);
  const __m512i v_d800 = _mm512_set1_epi16((uint16_t)0xd800);

  size_t count{0};

  while (ptr <= end) {
    __m512i utf16 = _mm512_loadu_si512((const __m512i*)ptr);
    ptr += 32;
    __mmask32 ascii_bitmask = _mm512_cmple_epu16_mask(utf16, v_007f);
    __mmask32 two_bytes_bitmask = _mm512_mask_cmple_epu16_mask(~ascii_bitmask, utf16, v_07ff);
    __mmask32 not_one_two_bytes = ~(ascii_bitmask | two_bytes_bitmask);
    __mmask32 surrogates_bitmask = _mm512_mask_cmple_epu16_mask(not_one_two_bytes, utf16, v_dfff) & _mm512_mask_cmpge_epu16_mask(not_one_two_bytes, utf16, v_d800);

    size_t ascii_count = count_ones(ascii_bitmask);
    size_t two_bytes_count = count_ones(two_bytes_bitmask);
    size_t surrogate_bytes_count = count_ones(surrogates_bitmask);
    size_t three_bytes_count = 32 - ascii_count - two_bytes_count - surrogate_bytes_count;

    count += ascii_count + 2*two_bytes_count + 3*three_bytes_count + 2*surrogate_bytes_count;
  }

  return count + scalar::utf16::utf8_length_from_utf16<endianness::LITTLE>(ptr, length - (ptr - input));
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(const char16_t * input, size_t length) const noexcept {
  const char16_t* end = length >= 32 ? input + length - 32 : nullptr;
  const char16_t* ptr = input;

  const __m512i v_007f = _mm512_set1_epi16((uint16_t)0x007f);
  const __m512i v_07ff = _mm512_set1_epi16((uint16_t)0x07ff);
  const __m512i v_dfff = _mm512_set1_epi16((uint16_t)0xdfff);
  const __m512i v_d800 = _mm512_set1_epi16((uint16_t)0xd800);

  size_t count{0};
  const __m512i byteflip = _mm512_setr_epi64(
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809,
            0x0607040502030001,
            0x0e0f0c0d0a0b0809
        );
  while (ptr <= end) {
    __m512i utf16 = _mm512_loadu_si512((const __m512i*)ptr);
    utf16 = _mm512_shuffle_epi8(utf16, byteflip);
    ptr += 32;
    __mmask32 ascii_bitmask = _mm512_cmple_epu16_mask(utf16, v_007f);
    __mmask32 two_bytes_bitmask = _mm512_mask_cmple_epu16_mask(~ascii_bitmask, utf16, v_07ff);
    __mmask32 not_one_two_bytes = ~(ascii_bitmask | two_bytes_bitmask);
    __mmask32 surrogates_bitmask = _mm512_mask_cmple_epu16_mask(not_one_two_bytes, utf16, v_dfff) & _mm512_mask_cmpge_epu16_mask(not_one_two_bytes, utf16, v_d800);

    size_t ascii_count = count_ones(ascii_bitmask);
    size_t two_bytes_count = count_ones(two_bytes_bitmask);
    size_t surrogate_bytes_count = count_ones(surrogates_bitmask);
    size_t three_bytes_count = 32 - ascii_count - two_bytes_count - surrogate_bytes_count;
    count += ascii_count + 2*two_bytes_count + 3*three_bytes_count + 2*surrogate_bytes_count;
  }

  return count + scalar::utf16::utf8_length_from_utf16<endianness::BIG>(ptr, length - (ptr - input));
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(const char16_t * input, size_t length) const noexcept {
  return implementation::count_utf16le(input, length);
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(const char16_t * input, size_t length) const noexcept {
  return implementation::count_utf16be(input, length);
}

simdutf_warn_unused size_t implementation::utf16_length_from_utf8(const char * input, size_t length) const noexcept {
    size_t pos = 0;
    size_t count = 0;
    // This algorithm could no doubt be improved!
    for(;pos + 64 <= length; pos += 64) {
      __m512i utf8 = _mm512_loadu_si512((const __m512i*)(input+pos));
      uint64_t utf8_continuation_mask = _mm512_cmple_epi8_mask(utf8, _mm512_set1_epi8(-65+1));
      // We count one word for anything that is not a continuation (so
      // leading bytes).
      count += 64 - count_ones(utf8_continuation_mask);
      uint64_t utf8_4byte = _mm512_cmpge_epu8_mask(utf8, _mm512_set1_epi8(int8_t(240)));
      count += count_ones(utf8_4byte);
    }
    return count + scalar::utf8::utf16_length_from_utf8(input + pos, length - pos);
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf32(const char32_t * input, size_t length) const noexcept {
  const char32_t* end = length >= 16 ? input + length - 16 : nullptr;
  const char32_t* ptr = input;

  const __m512i v_0000_007f = _mm512_set1_epi32((uint32_t)0x7f);
  const __m512i v_0000_07ff = _mm512_set1_epi32((uint32_t)0x7ff);
  const __m512i v_0000_ffff = _mm512_set1_epi32((uint32_t)0x0000ffff);

  size_t count{0};

  while (ptr <= end) {
    __m512i utf32 = _mm512_loadu_si512((const __m512i*)ptr);
    ptr += 16;
    __mmask16 ascii_bitmask = _mm512_cmple_epu32_mask(utf32, v_0000_007f);
    __mmask16 two_bytes_bitmask = _mm512_mask_cmple_epu32_mask(_knot_mask16(ascii_bitmask), utf32, v_0000_07ff);
    __mmask16 three_bytes_bitmask = _mm512_mask_cmple_epu32_mask(_knot_mask16(_mm512_kor(ascii_bitmask, two_bytes_bitmask)), utf32, v_0000_ffff);

    size_t ascii_count = count_ones(ascii_bitmask);
    size_t two_bytes_count = count_ones(two_bytes_bitmask);
    size_t three_bytes_count = count_ones(three_bytes_bitmask);
    size_t four_bytes_count = 16 - ascii_count - two_bytes_count - three_bytes_count;
    count += ascii_count + 2*two_bytes_count + 3*three_bytes_count + 4*four_bytes_count;
  }

  return count + scalar::utf32::utf8_length_from_utf32(ptr, length - (ptr - input));
}

simdutf_warn_unused size_t implementation::utf16_length_from_utf32(const char32_t * input, size_t length) const noexcept {
  const char32_t* end = length >= 16 ? input + length - 16 : nullptr;
  const char32_t* ptr = input;

  const __m512i v_0000_ffff = _mm512_set1_epi32((uint32_t)0x0000ffff);

  size_t count{0};

  while (ptr <= end) {
    __m512i utf32 = _mm512_loadu_si512((const __m512i*)ptr);
    ptr += 16;
    __mmask16 surrogates_bitmask = _mm512_cmpgt_epu32_mask(utf32, v_0000_ffff);

    count += 16 + count_ones(surrogates_bitmask);
  }

  return count + scalar::utf32::utf16_length_from_utf32(ptr, length - (ptr - input));
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf8(const char * input, size_t length) const noexcept {
  return implementation::count_utf8(input, length);
}

} // namespace icelake
} // namespace simdutf

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/icelake/end.h
/* begin file src/simdutf/icelake/end.h */
#if SIMDUTF_CAN_ALWAYS_RUN_ICELAKE
// nothing needed.
#else
SIMDUTF_UNTARGET_REGION
#endif


#if SIMDUTF_GCC11ORMORE // workaround for https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
SIMDUTF_POP_DISABLE_WARNINGS
#endif // end of workaround
/* end file src/simdutf/icelake/end.h */
/* end file src/icelake/implementation.cpp */
#endif
#if SIMDUTF_IMPLEMENTATION_HASWELL
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=haswell/implementation.cpp
/* begin file src/haswell/implementation.cpp */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/haswell/begin.h
/* begin file src/simdutf/haswell/begin.h */
// redefining SIMDUTF_IMPLEMENTATION to "haswell"
// #define SIMDUTF_IMPLEMENTATION haswell

#if SIMDUTF_CAN_ALWAYS_RUN_HASWELL
// nothing needed.
#else
SIMDUTF_TARGET_HASWELL
#endif

#if SIMDUTF_GCC11ORMORE // workaround for https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
SIMDUTF_DISABLE_GCC_WARNING(-Wmaybe-uninitialized)
#endif // end of workaround
/* end file src/simdutf/haswell/begin.h */
namespace simdutf {
namespace haswell {
namespace {
#ifndef SIMDUTF_HASWELL_H
#error "haswell.h must be included"
#endif
using namespace simd;


simdutf_really_inline bool is_ascii(const simd8x64<uint8_t>& input) {
  return input.reduce_or().is_ascii();
}

simdutf_unused simdutf_really_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(0b11000000u-1); // Only 11______ will be > 0
  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0b11100000u-1); // Only 111_____ will be > 0
  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0b11110000u-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);
}

simdutf_really_inline simd8<bool> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0b11100000u-1); // Only 111_____ will be > 0
  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0b11110000u-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_third_byte | is_fourth_byte) > int8_t(0);
}

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=haswell/avx2_detect_encodings.cpp
/* begin file src/haswell/avx2_detect_encodings.cpp */
template<class checker>
// len is known to be a multiple of 2 when this is called
int avx2_detect_encodings(const char * buf, size_t len) {
    const char* start = buf;
    const char* end = buf + len;

    bool is_utf8 = true;
    bool is_utf16 = true;
    bool is_utf32 = true;

    int out = 0;

    const auto v_d8 = simd8<uint8_t>::splat(0xd8);
    const auto v_f8 = simd8<uint8_t>::splat(0xf8);

    __m256i currentmax = _mm256_setzero_si256();

    checker check{};

    while(buf + 64 <= end) {
        __m256i in = _mm256_loadu_si256((__m256i*)buf);
        __m256i nextin = _mm256_loadu_si256((__m256i*)buf+1);

        const auto u0 = simd16<uint16_t>(in);
        const auto u1 = simd16<uint16_t>(nextin);

        const auto v0 = u0.shr<8>();
        const auto v1 = u1.shr<8>();

        const auto in16 = simd16<uint16_t>::pack(v0, v1);

        const auto surrogates_wordmask0 = (in16 & v_f8) == v_d8;
        uint32_t surrogates_bitmask0 = surrogates_wordmask0.to_bitmask();

        // Check for surrogates
        if (surrogates_bitmask0 != 0x0) {
            // Cannot be UTF8
            is_utf8 = false;
            // Can still be either UTF-16LE or UTF-32 depending on the positions of the surrogates
            // To be valid UTF-32, a surrogate cannot be in the two most significant bytes of any 32-bit word.
            // On the other hand, to be valid UTF-16LE, at least one surrogate must be in the two most significant
            // bytes of a 32-bit word since they always come in pairs in UTF-16LE.
            // Note that we always proceed in multiple of 4 before this point so there is no offset in 32-bit words.

            if ((surrogates_bitmask0 & 0xaaaaaaaa) != 0) {
                is_utf32 = false;
                // Code from avx2_validate_utf16le.cpp
                const char16_t * input = reinterpret_cast<const char16_t*>(buf);
                const char16_t* end16 = reinterpret_cast<const char16_t*>(start) + len/2;

                const auto v_fc = simd8<uint8_t>::splat(0xfc);
                const auto v_dc = simd8<uint8_t>::splat(0xdc);

                const uint32_t V0 = ~surrogates_bitmask0;

                const auto    vH0 = (in16 & v_fc) == v_dc;
                const uint32_t H0 = vH0.to_bitmask();

                const uint32_t L0 = ~H0 & surrogates_bitmask0;

                const uint32_t a0 = L0 & (H0 >> 1);
                const uint32_t b0 = a0 << 1;
                const uint32_t c0 = V0 | a0 | b0;

                if (c0 == 0xffffffff) {
                    input += simd16<uint16_t>::ELEMENTS * 2;
                } else if (c0 == 0x7fffffff) {
                    input += simd16<uint16_t>::ELEMENTS * 2 - 1;
                } else {
                    return simdutf::encoding_type::unspecified;
                }

                while (input + simd16<uint16_t>::ELEMENTS * 2 < end16) {
                    const auto in0 = simd16<uint16_t>(input);
                    const auto in1 = simd16<uint16_t>(input + simd16<uint16_t>::ELEMENTS);

                    const auto t0 = in0.shr<8>();
                    const auto t1 = in1.shr<8>();

                    const auto in_16 = simd16<uint16_t>::pack(t0, t1);

                    const auto surrogates_wordmask = (in_16 & v_f8) == v_d8;
                    const uint32_t surrogates_bitmask = surrogates_wordmask.to_bitmask();
                    if (surrogates_bitmask == 0x0) {
                        input += simd16<uint16_t>::ELEMENTS * 2;
                    } else {
                        const uint32_t V = ~surrogates_bitmask;

                        const auto    vH = (in_16 & v_fc) == v_dc;
                        const uint32_t H = vH.to_bitmask();

                        const uint32_t L = ~H & surrogates_bitmask;

                        const uint32_t a = L & (H >> 1);

                        const uint32_t b = a << 1;

                        const uint32_t c = V | a | b;

                        if (c == 0xffffffff) {
                            input += simd16<uint16_t>::ELEMENTS * 2;
                        } else if (c == 0x7fffffff) {
                            input += simd16<uint16_t>::ELEMENTS * 2 - 1;
                        } else {
                            return simdutf::encoding_type::unspecified;
                        }
                    }
                }
            } else {
                is_utf16 = false;
                // Check for UTF-32
                if (len % 4 == 0) {
                    const char32_t * input = reinterpret_cast<const char32_t*>(buf);
                    const char32_t* end32 = reinterpret_cast<const char32_t*>(start) + len/4;

                    // Must start checking for surrogates
                    __m256i currentoffsetmax = _mm256_setzero_si256();
                    const __m256i offset = _mm256_set1_epi32(0xffff2000);
                    const __m256i standardoffsetmax = _mm256_set1_epi32(0xfffff7ff);

                    currentmax = _mm256_max_epu32(in, currentmax);
                    currentmax = _mm256_max_epu32(nextin, currentmax);

                    currentoffsetmax = _mm256_max_epu32(_mm256_add_epi32(in, offset), currentoffsetmax);
                    currentoffsetmax = _mm256_max_epu32(_mm256_add_epi32(nextin, offset), currentoffsetmax);

                    while (input + 8 < end32) {
                        const __m256i in32 = _mm256_loadu_si256((__m256i *)input);
                        currentmax = _mm256_max_epu32(in32,currentmax);
                        currentoffsetmax = _mm256_max_epu32(_mm256_add_epi32(in32, offset), currentoffsetmax);
                        input += 8;
                    }

                    __m256i forbidden_words = _mm256_xor_si256(_mm256_max_epu32(currentoffsetmax, standardoffsetmax), standardoffsetmax);
                    if(_mm256_testz_si256(forbidden_words, forbidden_words) == 0) {
                        return simdutf::encoding_type::unspecified;
                    }
                } else {
                    return simdutf::encoding_type::unspecified;
                }
            }
            break;
        }
        // If no surrogate, validate under other encodings as well

        // UTF-32 validation
        currentmax = _mm256_max_epu32(in, currentmax);
        currentmax = _mm256_max_epu32(nextin, currentmax);

        // UTF-8 validation
        // Relies on ../generic/utf8_validation/utf8_lookup4_algorithm.h
        simd::simd8x64<uint8_t> in8(in, nextin);
        check.check_next_input(in8);

        buf += 64;
    }

    // Check which encodings are possible

    if (is_utf8) {
        if (static_cast<size_t>(buf - start) != len) {
            uint8_t block[64]{};
            std::memset(block, 0x20, 64);
            std::memcpy(block, buf, len - (buf - start));
            simd::simd8x64<uint8_t> in(block);
            check.check_next_input(in);
        }
        if (!check.errors()) {
            out |= simdutf::encoding_type::UTF8;
        }
    }

    if (is_utf16 && scalar::utf16::validate<endianness::LITTLE>(reinterpret_cast<const char16_t*>(buf), (len - (buf - start))/2)) {
        out |= simdutf::encoding_type::UTF16_LE;
    }

    if (is_utf32 && (len % 4 == 0)) {
        const __m256i standardmax = _mm256_set1_epi32(0x10ffff);
        __m256i is_zero = _mm256_xor_si256(_mm256_max_epu32(currentmax, standardmax), standardmax);
        if (_mm256_testz_si256(is_zero, is_zero) == 1 && scalar::utf32::validate(reinterpret_cast<const char32_t*>(buf), (len - (buf - start))/4)) {
            out |= simdutf::encoding_type::UTF32_LE;
        }
    }

    return out;
}
/* end file src/haswell/avx2_detect_encodings.cpp */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=haswell/avx2_validate_utf16.cpp
/* begin file src/haswell/avx2_validate_utf16.cpp */
/*
    In UTF-16 words in range 0xD800 to 0xDFFF have special meaning.

    In a vectorized algorithm we want to examine the most significant
    nibble in order to select a fast path. If none of highest nibbles
    are 0xD (13), than we are sure that UTF-16 chunk in a vector
    register is valid.

    Let us analyze what we need to check if the nibble is 0xD. The
    value of the preceding nibble determines what we have:

    0xd000 .. 0xd7ff - a valid word
    0xd800 .. 0xdbff - low surrogate
    0xdc00 .. 0xdfff - high surrogate

    Other constraints we have to consider:
    - there must not be two consecutive low surrogates (0xd800 .. 0xdbff)
    - there must not be two consecutive high surrogates (0xdc00 .. 0xdfff)
    - there must not be sole low surrogate nor high surrogate

    We're going to build three bitmasks based on the 3rd nibble:
    - V = valid word,
    - L = low surrogate (0xd800 .. 0xdbff)
    - H = high surrogate (0xdc00 .. 0xdfff)

      0   1   2   3   4   5   6   7    <--- word index
    [ V | L | H | L | H | V | V | L ]
      1   0   0   0   0   1   1   0     - V = valid masks
      0   1   0   1   0   0   0   1     - L = low surrogate
      0   0   1   0   1   0   0   0     - H high surrogate


      1   0   0   0   0   1   1   0   V = valid masks
      0   1   0   1   0   0   0   0   a = L & (H >> 1)
      0   0   1   0   1   0   0   0   b = a << 1
      1   1   1   1   1   1   1   0   c = V | a | b
                                  ^
                                  the last bit can be zero, we just consume 7 words
                                  and recheck this word in the next iteration
*/

/* Returns:
   - pointer to the last unprocessed character (a scalar fallback should check the rest);
   - nullptr if an error was detected.
*/
template <endianness big_endian>
const char16_t* avx2_validate_utf16(const char16_t* input, size_t size) {
    const char16_t* end = input + size;

    const auto v_d8 = simd8<uint8_t>::splat(0xd8);
    const auto v_f8 = simd8<uint8_t>::splat(0xf8);
    const auto v_fc = simd8<uint8_t>::splat(0xfc);
    const auto v_dc = simd8<uint8_t>::splat(0xdc);

    while (input + simd16<uint16_t>::ELEMENTS * 2 < end) {
        // 0. Load data: since the validation takes into account only higher
        //    byte of each word, we compress the two vectors into one which
        //    consists only the higher bytes.
        auto in0 = simd16<uint16_t>(input);
        auto in1 = simd16<uint16_t>(input + simd16<uint16_t>::ELEMENTS);

        if (big_endian) {
            in0 = in0.swap_bytes();
            in1 = in1.swap_bytes();
        }

        const auto t0 = in0.shr<8>();
        const auto t1 = in1.shr<8>();

        const auto in = simd16<uint16_t>::pack(t0, t1);

        // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
        const auto surrogates_wordmask = (in & v_f8) == v_d8;
        const uint32_t surrogates_bitmask = surrogates_wordmask.to_bitmask();
        if (surrogates_bitmask == 0x0) {
            input += simd16<uint16_t>::ELEMENTS * 2;
        } else {
            // 2. We have some surrogates that have to be distinguished:
            //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
            //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
            //
            //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)

            // V - non-surrogate words
            //     V = not surrogates_wordmask
            const uint32_t V = ~surrogates_bitmask;

            // H - word-mask for high surrogates: the six highest bits are 0b1101'11
            const auto    vH = (in & v_fc) == v_dc;
            const uint32_t H = vH.to_bitmask();

            // L - word mask for low surrogates
            //     L = not H and surrogates_wordmask
            const uint32_t L = ~H & surrogates_bitmask;

            const uint32_t a = L & (H >> 1);  // A low surrogate must be followed by high one.
                                              // (A low surrogate placed in the 7th register's word
                                              // is an exception we handle.)
            const uint32_t b = a << 1;        // Just mark that the opposite fact is hold,
                                              // thanks to that we have only two masks for valid case.
            const uint32_t c = V | a | b;     // Combine all the masks into the final one.

            if (c == 0xffffffff) {
                // The whole input register contains valid UTF-16, i.e.,
                // either single words or proper surrogate pairs.
                input += simd16<uint16_t>::ELEMENTS * 2;
            } else if (c == 0x7fffffff) {
                // The 31 lower words of the input register contains valid UTF-16.
                // The 31 word may be either a low or high surrogate. It the next
                // iteration we 1) check if the low surrogate is followed by a high
                // one, 2) reject sole high surrogate.
                input += simd16<uint16_t>::ELEMENTS * 2 - 1;
            } else {
                return nullptr;
            }
        }
    }

    return input;
}


template <endianness big_endian>
const result avx2_validate_utf16_with_errors(const char16_t* input, size_t size) {
    const char16_t* start = input;
    const char16_t* end = input + size;

    const auto v_d8 = simd8<uint8_t>::splat(0xd8);
    const auto v_f8 = simd8<uint8_t>::splat(0xf8);
    const auto v_fc = simd8<uint8_t>::splat(0xfc);
    const auto v_dc = simd8<uint8_t>::splat(0xdc);

    while (input + simd16<uint16_t>::ELEMENTS * 2 < end) {
        // 0. Load data: since the validation takes into account only higher
        //    byte of each word, we compress the two vectors into one which
        //    consists only the higher bytes.
        auto in0 = simd16<uint16_t>(input);
        auto in1 = simd16<uint16_t>(input + simd16<uint16_t>::ELEMENTS);

        if (big_endian) {
            in0 = in0.swap_bytes();
            in1 = in1.swap_bytes();
        }

        const auto t0 = in0.shr<8>();
        const auto t1 = in1.shr<8>();

        const auto in = simd16<uint16_t>::pack(t0, t1);

        // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
        const auto surrogates_wordmask = (in & v_f8) == v_d8;
        const uint32_t surrogates_bitmask = surrogates_wordmask.to_bitmask();
        if (surrogates_bitmask == 0x0) {
            input += simd16<uint16_t>::ELEMENTS * 2;
        } else {
            // 2. We have some surrogates that have to be distinguished:
            //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
            //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
            //
            //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)

            // V - non-surrogate words
            //     V = not surrogates_wordmask
            const uint32_t V = ~surrogates_bitmask;

            // H - word-mask for high surrogates: the six highest bits are 0b1101'11
            const auto    vH = (in & v_fc) == v_dc;
            const uint32_t H = vH.to_bitmask();

            // L - word mask for low surrogates
            //     L = not H and surrogates_wordmask
            const uint32_t L = ~H & surrogates_bitmask;

            const uint32_t a = L & (H >> 1);  // A low surrogate must be followed by high one.
                                              // (A low surrogate placed in the 7th register's word
                                              // is an exception we handle.)
            const uint32_t b = a << 1;        // Just mark that the opposite fact is hold,
                                              // thanks to that we have only two masks for valid case.
            const uint32_t c = V | a | b;     // Combine all the masks into the final one.

            if (c == 0xffffffff) {
                // The whole input register contains valid UTF-16, i.e.,
                // either single words or proper surrogate pairs.
                input += simd16<uint16_t>::ELEMENTS * 2;
            } else if (c == 0x7fffffff) {
                // The 31 lower words of the input register contains valid UTF-16.
                // The 31 word may be either a low or high surrogate. It the next
                // iteration we 1) check if the low surrogate is followed by a high
                // one, 2) reject sole high surrogate.
                input += simd16<uint16_t>::ELEMENTS * 2 - 1;
            } else {
                return result(error_code::SURROGATE, input - start);
            }
        }
    }

    return result(error_code::SUCCESS, input - start);
}
/* end file src/haswell/avx2_validate_utf16.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=haswell/avx2_validate_utf32le.cpp
/* begin file src/haswell/avx2_validate_utf32le.cpp */
/* Returns:
   - pointer to the last unprocessed character (a scalar fallback should check the rest);
   - nullptr if an error was detected.
*/
const char32_t* avx2_validate_utf32le(const char32_t* input, size_t size) {
    const char32_t* end = input + size;

    const __m256i standardmax = _mm256_set1_epi32(0x10ffff);
    const __m256i offset = _mm256_set1_epi32(0xffff2000);
    const __m256i standardoffsetmax = _mm256_set1_epi32(0xfffff7ff);
    __m256i currentmax = _mm256_setzero_si256();
    __m256i currentoffsetmax = _mm256_setzero_si256();

    while (input + 8 < end) {
        const __m256i in = _mm256_loadu_si256((__m256i *)input);
        currentmax = _mm256_max_epu32(in,currentmax);
        currentoffsetmax = _mm256_max_epu32(_mm256_add_epi32(in, offset), currentoffsetmax);
        input += 8;
    }
    __m256i is_zero = _mm256_xor_si256(_mm256_max_epu32(currentmax, standardmax), standardmax);
    if(_mm256_testz_si256(is_zero, is_zero) == 0) {
        return nullptr;
    }

    is_zero = _mm256_xor_si256(_mm256_max_epu32(currentoffsetmax, standardoffsetmax), standardoffsetmax);
    if(_mm256_testz_si256(is_zero, is_zero) == 0) {
        return nullptr;
    }

    return input;
}


const result avx2_validate_utf32le_with_errors(const char32_t* input, size_t size) {
    const char32_t* start = input;
    const char32_t* end = input + size;

    const __m256i standardmax = _mm256_set1_epi32(0x10ffff);
    const __m256i offset = _mm256_set1_epi32(0xffff2000);
    const __m256i standardoffsetmax = _mm256_set1_epi32(0xfffff7ff);
    __m256i currentmax = _mm256_setzero_si256();
    __m256i currentoffsetmax = _mm256_setzero_si256();

    while (input + 8 < end) {
        const __m256i in = _mm256_loadu_si256((__m256i *)input);
        currentmax = _mm256_max_epu32(in,currentmax);
        currentoffsetmax = _mm256_max_epu32(_mm256_add_epi32(in, offset), currentoffsetmax);

        __m256i is_zero = _mm256_xor_si256(_mm256_max_epu32(currentmax, standardmax), standardmax);
        if(_mm256_testz_si256(is_zero, is_zero) == 0) {
            return result(error_code::TOO_LARGE, input - start);
        }

        is_zero = _mm256_xor_si256(_mm256_max_epu32(currentoffsetmax, standardoffsetmax), standardoffsetmax);
        if(_mm256_testz_si256(is_zero, is_zero) == 0) {
            return result(error_code::SURROGATE, input - start);
        }
        input += 8;
    }

    return result(error_code::SUCCESS, input - start);
}
/* end file src/haswell/avx2_validate_utf32le.cpp */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=haswell/avx2_convert_utf8_to_utf16.cpp
/* begin file src/haswell/avx2_convert_utf8_to_utf16.cpp */
// depends on "tables/utf8_to_utf16_tables.h"


// Convert up to 12 bytes from utf8 to utf16 using a mask indicating the
// end of the code points. Only the least significant 12 bits of the mask
// are accessed.
// It returns how many bytes were consumed (up to 12).
template <endianness big_endian>
size_t convert_masked_utf8_to_utf16(const char *input,
                           uint64_t utf8_end_of_code_point_mask,
                           char16_t *&utf16_output) {
  // we use an approach where we try to process up to 12 input bytes.
  // Why 12 input bytes and not 16? Because we are concerned with the size of
  // the lookup tables. Also 12 is nicely divisible by two and three.
  //
  //
  // Optimization note: our main path below is load-latency dependent. Thus it is maybe
  // beneficial to have fast paths that depend on branch prediction but have less latency.
  // This results in more instructions but, potentially, also higher speeds.
  //
  // We first try a few fast paths.
  const __m128i swap = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
  const __m128i in = _mm_loadu_si128((__m128i *)input);
  const uint16_t input_utf8_end_of_code_point_mask =
      utf8_end_of_code_point_mask & 0xfff;
  if(((utf8_end_of_code_point_mask & 0xffff) == 0xffff)) {
    // We process the data in chunks of 16 bytes.
    __m256i ascii = _mm256_cvtepu8_epi16(in);
    if (big_endian) {
      const __m256i swap256 = _mm256_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14,
                                  17, 16, 19, 18, 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
      ascii = _mm256_shuffle_epi8(ascii, swap256);
    }
    _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf16_output), ascii);
    utf16_output += 16; // We wrote 16 16-bit characters.
    return 16; // We consumed 16 bytes.
  }
  if(((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) {
    // We want to take 8 2-byte UTF-8 words and turn them into 8 2-byte UTF-16 words.
    // There is probably a more efficient sequence, but the following might do.
    const __m128i sh = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
    __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
    if (big_endian) composed = _mm_shuffle_epi8(composed, swap);
    _mm_storeu_si128((__m128i *)utf16_output, composed);
    utf16_output += 8; // We wrote 16 bytes, 8 code points.
    return 16;
  }
  if(input_utf8_end_of_code_point_mask == 0x924) {
    // We want to take 4 3-byte UTF-8 words and turn them into 4 2-byte UTF-16 words.
    // There is probably a more efficient sequence, but the following might do.
    const __m128i sh = _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii =
        _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits
    const __m128i middlebyte =
        _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits
    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
    const __m128i highbyte =
        _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits
    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4);
    const __m128i composed =
        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted);
    __m128i composed_repacked = _mm_packus_epi32(composed, composed);
    if (big_endian) composed_repacked = _mm_shuffle_epi8(composed_repacked, swap);
    _mm_storeu_si128((__m128i *)utf16_output, composed_repacked);
    utf16_output += 4;
    return 12;
  }

  const uint8_t idx =
      simdutf::tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0];
  const uint8_t consumed =
      simdutf::tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1];
  if (idx < 64) {
    // SIX (6) input code-words
    // this is a relatively easy scenario
    // we process SIX (6) input code-words. The max length in bytes of six code
    // words spanning between 1 and 2 bytes each is 12 bytes. On processors
    // where pdep/pext is fast, we might be able to use a small lookup table.
    const __m128i sh =
        _mm_loadu_si128((const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
    __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
    if (big_endian) composed = _mm_shuffle_epi8(composed, swap);
    _mm_storeu_si128((__m128i *)utf16_output, composed);
    utf16_output += 6; // We wrote 12 bytes, 6 code points. There is a potential overflow of 4 bytes.
  } else if (idx < 145) {
    // FOUR (4) input code-words
    const __m128i sh =
        _mm_loadu_si128((const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii =
        _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits
    const __m128i middlebyte =
        _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits
    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
    const __m128i highbyte =
        _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits
    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4);
    const __m128i composed =
        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted);
    __m128i composed_repacked = _mm_packus_epi32(composed, composed);
    if (big_endian) composed_repacked = _mm_shuffle_epi8(composed_repacked, swap);
    _mm_storeu_si128((__m128i *)utf16_output, composed_repacked);
    utf16_output += 4; // Here we overflow by 8 bytes.
  } else if (idx < 209) {
    // TWO (2) input code-words
    //////////////
    // There might be garbage inputs where a leading byte mascarades as a four-byte
    // leading byte (by being followed by 3 continuation byte), but is not greater than
    // 0xf0. This could trigger a buffer overflow if we only counted leading
    // bytes of the form 0xf0 as generating surrogate pairs, without further UTF-8 validation.
    // Thus we must be careful to ensure that only leading bytes at least as large as 0xf0 generate surrogate pairs.
    // We do as at the cost of an extra mask.
    /////////////
    const __m128i sh =
        _mm_loadu_si128((const __m128i *)simdutf::tables::utf8_to_utf16::shufutf8[idx]);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f));
    const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00));
    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
    __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000));
    // correct for spurious high bit
    const __m128i correct =
        _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1);
    middlehighbyte = _mm_xor_si128(correct, middlehighbyte);
    const __m128i middlehighbyte_shifted = _mm_srli_epi32(middlehighbyte, 4);
    // We deliberately carry the leading four bits in highbyte if they are present,
    // we remove them later when computing hightenbits.
    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0xff000000));
    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 6);
    // When we need to generate a surrogate pair (leading byte > 0xF0), then
    // the corresponding 32-bit value in 'composed'  will be greater than
    // > (0xff00000>>6) or > 0x3c00000. This can be used later to identify the
    // location of the surrogate pairs.
    const __m128i composed =
        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted),
                     _mm_or_si128(highbyte_shifted, middlehighbyte_shifted));
    const __m128i composedminus =
        _mm_sub_epi32(composed, _mm_set1_epi32(0x10000));
    const __m128i lowtenbits =
        _mm_and_si128(composedminus, _mm_set1_epi32(0x3ff));
    // Notice the 0x3ff mask:
    const __m128i hightenbits = _mm_and_si128(_mm_srli_epi32(composedminus, 10), _mm_set1_epi32(0x3ff));
    const __m128i lowtenbitsadd =
        _mm_add_epi32(lowtenbits, _mm_set1_epi32(0xDC00));
    const __m128i hightenbitsadd =
        _mm_add_epi32(hightenbits, _mm_set1_epi32(0xD800));
    const __m128i lowtenbitsaddshifted = _mm_slli_epi32(lowtenbitsadd, 16);
    __m128i surrogates =
        _mm_or_si128(hightenbitsadd, lowtenbitsaddshifted);
    uint32_t basic_buffer[4];
    uint32_t basic_buffer_swap[4];
    if (big_endian) {
      _mm_storeu_si128((__m128i *)basic_buffer_swap, _mm_shuffle_epi8(composed, swap));
      surrogates = _mm_shuffle_epi8(surrogates, swap);
    }
    _mm_storeu_si128((__m128i *)basic_buffer, composed);
    uint32_t surrogate_buffer[4];
    _mm_storeu_si128((__m128i *)surrogate_buffer, surrogates);
    for (size_t i = 0; i < 3; i++) {
      if(basic_buffer[i] > 0x3c00000) {
        utf16_output[0] = uint16_t(surrogate_buffer[i] & 0xffff);
        utf16_output[1] = uint16_t(surrogate_buffer[i] >> 16);
        utf16_output += 2;
      } else  {
        utf16_output[0] = big_endian ? uint16_t(basic_buffer_swap[i]) : uint16_t(basic_buffer[i]);
        utf16_output++;
      }
    }
  } else {
    // here we know that there is an error but we do not handle errors
  }
  return consumed;
}
/* end file src/haswell/avx2_convert_utf8_to_utf16.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=haswell/avx2_convert_utf8_to_utf32.cpp
/* begin file src/haswell/avx2_convert_utf8_to_utf32.cpp */
// depends on "tables/utf8_to_utf16_tables.h"


// Convert up to 12 bytes from utf8 to utf32 using a mask indicating the
// end of the code points. Only the least significant 12 bits of the mask
// are accessed.
// It returns how many bytes were consumed (up to 12).
size_t convert_masked_utf8_to_utf32(const char *input,
                           uint64_t utf8_end_of_code_point_mask,
                           char32_t *&utf32_output) {
  // we use an approach where we try to process up to 12 input bytes.
  // Why 12 input bytes and not 16? Because we are concerned with the size of
  // the lookup tables. Also 12 is nicely divisible by two and three.
  //
  //
  // Optimization note: our main path below is load-latency dependent. Thus it is maybe
  // beneficial to have fast paths that depend on branch prediction but have less latency.
  // This results in more instructions but, potentially, also higher speeds.
  //
  // We first try a few fast paths.
  const __m128i in = _mm_loadu_si128((__m128i *)input);
  const uint16_t input_utf8_end_of_code_point_mask =
      utf8_end_of_code_point_mask & 0xfff;
  if(((utf8_end_of_code_point_mask & 0xffff) == 0xffff)) {
    // We process the data in chunks of 16 bytes.
    _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output), _mm256_cvtepu8_epi32(in));
    _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output+8), _mm256_cvtepu8_epi32(_mm_srli_si128(in,8)));
    utf32_output += 16; // We wrote 16 32-bit characters.
    return 16; // We consumed 16 bytes.
  }
  if(((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) {
    // We want to take 8 2-byte UTF-8 words and turn them into 8 4-byte UTF-32 words.
    // There is probably a more efficient sequence, but the following might do.
    const __m128i sh = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
    const __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
    _mm256_storeu_si256((__m256i *)utf32_output, _mm256_cvtepu16_epi32(composed));
    utf32_output += 8; // We wrote 16 bytes, 8 code points.
    return 16;
  }
  if(input_utf8_end_of_code_point_mask == 0x924) {
    // We want to take 4 3-byte UTF-8 words and turn them into 4 4-byte UTF-32 words.
    // There is probably a more efficient sequence, but the following might do.
    const __m128i sh = _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii =
        _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits
    const __m128i middlebyte =
        _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits
    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
    const __m128i highbyte =
        _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits
    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4);
    const __m128i composed =
        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted);
    _mm_storeu_si128((__m128i *)utf32_output, composed);
    utf32_output += 4;
    return 12;
  }
  /// We do not have a fast path available, so we fallback.

  const uint8_t idx =
      tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0];
  const uint8_t consumed =
      tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1];
  if (idx < 64) {
    // SIX (6) input code-words
    // this is a relatively easy scenario
    // we process SIX (6) input code-words. The max length in bytes of six code
    // words spanning between 1 and 2 bytes each is 12 bytes. On processors
    // where pdep/pext is fast, we might be able to use a small lookup table.
    const __m128i sh =
        _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
    const __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
    _mm256_storeu_si256((__m256i *)utf32_output, _mm256_cvtepu16_epi32(composed));
    utf32_output += 6; // We wrote 24 bytes, 6 code points. There is a potential
    // overflow of 32 - 24 = 8 bytes.
  } else if (idx < 145) {
    // FOUR (4) input code-words
    const __m128i sh =
        _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii =
        _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits
    const __m128i middlebyte =
        _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits
    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
    const __m128i highbyte =
        _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits
    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4);
    const __m128i composed =
        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted);
    _mm_storeu_si128((__m128i *)utf32_output, composed);
    utf32_output += 4;
  } else if (idx < 209) {
    // TWO (2) input code-words
    const __m128i sh =
        _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f));
    const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00));
    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
    __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000));
    // correct for spurious high bit
    const __m128i correct =
        _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1);
    middlehighbyte = _mm_xor_si128(correct, middlehighbyte);
    const __m128i middlehighbyte_shifted = _mm_srli_epi32(middlehighbyte, 4);
    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0x07000000));
    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 6);
    const __m128i composed =
        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted),
                     _mm_or_si128(highbyte_shifted, middlehighbyte_shifted));
    _mm_storeu_si128((__m128i *)utf32_output, composed);
    utf32_output += 3; // We wrote 3 * 4 bytes, there is a potential overflow of 4 bytes.
  } else {
    // here we know that there is an error but we do not handle errors
  }
  return consumed;
}
/* end file src/haswell/avx2_convert_utf8_to_utf32.cpp */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=haswell/avx2_convert_utf16_to_utf8.cpp
/* begin file src/haswell/avx2_convert_utf16_to_utf8.cpp */
/*
    The vectorized algorithm works on single SSE register i.e., it
    loads eight 16-bit words.

    We consider three cases:
    1. an input register contains no surrogates and each value
       is in range 0x0000 .. 0x07ff.
    2. an input register contains no surrogates and values are
       is in range 0x0000 .. 0xffff.
    3. an input register contains surrogates --- i.e. codepoints
       can have 16 or 32 bits.

    Ad 1.

    When values are less than 0x0800, it means that a 16-bit words
    can be converted into: 1) single UTF8 byte (when it's an ASCII
    char) or 2) two UTF8 bytes.

    For this case we do only some shuffle to obtain these 2-byte
    codes and finally compress the whole SSE register with a single
    shuffle.

    We need 256-entry lookup table to get a compression pattern
    and the number of output bytes in the compressed vector register.
    Each entry occupies 17 bytes.

    Ad 2.

    When values fit in 16-bit words, but are above 0x07ff, then
    a single word may produce one, two or three UTF8 bytes.

    We prepare data for all these three cases in two registers.
    The first register contains lower two UTF8 bytes (used in all
    cases), while the second one contains just the third byte for
    the three-UTF8-bytes case.

    Finally these two registers are interleaved forming eight-element
    array of 32-bit values. The array spans two SSE registers.
    The bytes from the registers are compressed using two shuffles.

    We need 256-entry lookup table to get a compression pattern
    and the number of output bytes in the compressed vector register.
    Each entry occupies 17 bytes.


    To summarize:
    - We need two 256-entry tables that have 8704 bytes in total.
*/


/*
  Returns a pair: the first unprocessed byte from buf and utf8_output
  A scalar routing should carry on the conversion of the tail.
*/
template <endianness big_endian>
std::pair<const char16_t*, char*> avx2_convert_utf16_to_utf8(const char16_t* buf, size_t len, char* utf8_output) {
  const char16_t* end = buf + len;
  const __m256i v_0000 = _mm256_setzero_si256();
  const __m256i v_f800 = _mm256_set1_epi16((int16_t)0xf800);
  const __m256i v_d800 = _mm256_set1_epi16((int16_t)0xd800);
  const __m256i v_c080 = _mm256_set1_epi16((int16_t)0xc080);
  const size_t safety_margin = 12; // to avoid overruns, see issue https://github.com/simdutf/simdutf/issues/92

  while (buf + 16 + safety_margin <= end) {
    __m256i in = _mm256_loadu_si256((__m256i*)buf);
    if (big_endian) {
      const __m256i swap = _mm256_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14,
                                  17, 16, 19, 18, 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
      in = _mm256_shuffle_epi8(in, swap);
    }
    // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes
    const __m256i v_ff80 = _mm256_set1_epi16((int16_t)0xff80);
    if(_mm256_testz_si256(in, v_ff80)) { // ASCII fast path!!!!
        // 1. pack the bytes
        const __m128i utf8_packed = _mm_packus_epi16(_mm256_castsi256_si128(in),_mm256_extractf128_si256(in,1));
        // 2. store (16 bytes)
        _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);
        // 3. adjust pointers
        buf += 16;
        utf8_output += 16;
        continue; // we are done for this round!
    }
    // no bits set above 7th bit
    const __m256i one_byte_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in, v_ff80), v_0000);
    const uint32_t one_byte_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));

    // no bits set above 11th bit
    const __m256i one_or_two_bytes_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_0000);
    const uint32_t one_or_two_bytes_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(one_or_two_bytes_bytemask));
    if (one_or_two_bytes_bitmask == 0xffffffff) {

          // 1. prepare 2-byte values
          // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
          // expected output   : [110a|aaaa|10bb|bbbb] x 8
          const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
          const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);

          // t0 = [000a|aaaa|bbbb|bb00]
          const __m256i t0 = _mm256_slli_epi16(in, 2);
          // t1 = [000a|aaaa|0000|0000]
          const __m256i t1 = _mm256_and_si256(t0, v_1f00);
          // t2 = [0000|0000|00bb|bbbb]
          const __m256i t2 = _mm256_and_si256(in, v_003f);
          // t3 = [000a|aaaa|00bb|bbbb]
          const __m256i t3 = _mm256_or_si256(t1, t2);
          // t4 = [110a|aaaa|10bb|bbbb]
          const __m256i t4 = _mm256_or_si256(t3, v_c080);

          // 2. merge ASCII and 2-byte codewords
          const __m256i utf8_unpacked = _mm256_blendv_epi8(t4, in, one_byte_bytemask);

          // 3. prepare bitmask for 8-bit lookup
          const uint32_t M0 = one_byte_bitmask & 0x55555555;
          const uint32_t M1 = M0 >> 7;
          const uint32_t M2 = (M1 | M0)  & 0x00ff00ff;
          // 4. pack the bytes

          const uint8_t* row = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
          const uint8_t* row_2 = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2>>16)][0];

          const __m128i shuffle = _mm_loadu_si128((__m128i*)(row + 1));
          const __m128i shuffle_2 = _mm_loadu_si128((__m128i*)(row_2 + 1));

          const __m256i utf8_packed = _mm256_shuffle_epi8(utf8_unpacked, _mm256_setr_m128i(shuffle,shuffle_2));
          // 5. store bytes
          _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_packed));
          utf8_output += row[0];
          _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_packed,1));
          utf8_output += row_2[0];

          // 6. adjust pointers
          buf += 16;
          continue;
    }
    // 1. Check if there are any surrogate word in the input chunk.
    //    We have also deal with situation when there is a surrogate word
    //    at the end of a chunk.
    const __m256i surrogates_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_d800);

    // bitmask = 0x0000 if there are no surrogates
    //         = 0xc000 if the last word is a surrogate
    const uint32_t surrogates_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(surrogates_bytemask));
    // It might seem like checking for surrogates_bitmask == 0xc000 could help. However,
    // it is likely an uncommon occurrence.
    if (surrogates_bitmask == 0x00000000) {
      // case: words from register produce either 1, 2 or 3 UTF-8 bytes
        const __m256i dup_even = _mm256_setr_epi16(0x0000, 0x0202, 0x0404, 0x0606,
                                                0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
                                                0x0000, 0x0202, 0x0404, 0x0606,
                                                0x0808, 0x0a0a, 0x0c0c, 0x0e0e);

        /* In this branch we handle three cases:
           1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           - single UFT-8 byte
           2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two UTF-8 bytes
           3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - three UTF-8 bytes

          We expand the input word (16-bit) into two words (32-bit), thus
          we have room for four bytes. However, we need five distinct bit
          layouts. Note that the last byte in cases #2 and #3 is the same.

          We precompute byte 1 for case #1 and the common byte for cases #2 & #3
          in register t2.

          We precompute byte 1 for case #3 and -- **conditionally** -- precompute
          either byte 1 for case #2 or byte 2 for case #3. Note that they
          differ by exactly one bit.

          Finally from these two words we build proper UTF-8 sequence, taking
          into account the case (i.e, the number of bytes to write).
        */
        /**
         * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
         * t2 => [0ccc|cccc] [10cc|cccc]
         * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
         */
#define simdutf_vec(x) _mm256_set1_epi16(static_cast<uint16_t>(x))
        // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
        const __m256i t0 = _mm256_shuffle_epi8(in, dup_even);
        // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
        const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111));
        // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
        const __m256i t2 = _mm256_or_si256 (t1, simdutf_vec(0b1000000000000000));

        // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
        const __m256i s0 = _mm256_srli_epi16(in, 4);
        // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
        const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100));
        // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
        const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140));
        // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
        const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000));
        const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask, simdutf_vec(0b0100000000000000));
        const __m256i s4 = _mm256_xor_si256(s3, m0);
#undef simdutf_vec

        // 4. expand words 16-bit => 32-bit
        const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
        const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);

        // 5. compress 32-bit words into 1, 2 or 3 bytes -- 2 x shuffle
        const uint32_t mask = (one_byte_bitmask & 0x55555555) |
                              (one_or_two_bytes_bitmask & 0xaaaaaaaa);
        // Due to the wider registers, the following path is less likely to be useful.
        /*if(mask == 0) {
          // We only have three-byte words. Use fast path.
          const __m256i shuffle = _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1, 2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1);
          const __m256i utf8_0 = _mm256_shuffle_epi8(out0, shuffle);
          const __m256i utf8_1 = _mm256_shuffle_epi8(out1, shuffle);
          _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
          utf8_output += 12;
          _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
          utf8_output += 12;
          _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_0,1));
          utf8_output += 12;
          _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_1,1));
          utf8_output += 12;
          buf += 16;
          continue;
        }*/
        const uint8_t mask0 = uint8_t(mask);
        const uint8_t* row0 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
        const __m128i shuffle0 = _mm_loadu_si128((__m128i*)(row0 + 1));
        const __m128i utf8_0 = _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);

        const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
        const uint8_t* row1 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
        const __m128i shuffle1 = _mm_loadu_si128((__m128i*)(row1 + 1));
        const __m128i utf8_1 = _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);

        const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
        const uint8_t* row2 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
        const __m128i shuffle2 = _mm_loadu_si128((__m128i*)(row2 + 1));
        const __m128i utf8_2 = _mm_shuffle_epi8(_mm256_extractf128_si256(out0,1), shuffle2);


        const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
        const uint8_t* row3 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
        const __m128i shuffle3 = _mm_loadu_si128((__m128i*)(row3 + 1));
        const __m128i utf8_3 = _mm_shuffle_epi8(_mm256_extractf128_si256(out1,1), shuffle3);

        _mm_storeu_si128((__m128i*)utf8_output, utf8_0);
        utf8_output += row0[0];
        _mm_storeu_si128((__m128i*)utf8_output, utf8_1);
        utf8_output += row1[0];
        _mm_storeu_si128((__m128i*)utf8_output, utf8_2);
        utf8_output += row2[0];
        _mm_storeu_si128((__m128i*)utf8_output, utf8_3);
        utf8_output += row3[0];
        buf += 16;
    // surrogate pair(s) in a register
    } else {
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // in the presence of surrogate pairs may require non-trivial tables.
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint16_t word = big_endian ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
        if((word & 0xFF80)==0) {
          *utf8_output++ = char(word);
        } else if((word & 0xF800)==0) {
          *utf8_output++ = char((word>>6) | 0b11000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else if((word &0xF800 ) != 0xD800) {
          *utf8_output++ = char((word>>12) | 0b11100000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else {
          // must be a surrogate pair
          uint16_t diff = uint16_t(word - 0xD800);
          uint16_t next_word = big_endian ? scalar::utf16::swap_bytes(buf[k+1]) : buf[k+1];
          k++;
          uint16_t diff2 = uint16_t(next_word - 0xDC00);
          if((diff | diff2) > 0x3FF)  { return std::make_pair(nullptr, utf8_output); }
          uint32_t value = (diff << 10) + diff2 + 0x10000;
          *utf8_output++ = char((value>>18) | 0b11110000);
          *utf8_output++ = char(((value>>12) & 0b111111) | 0b10000000);
          *utf8_output++ = char(((value>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((value & 0b111111) | 0b10000000);
        }
      }
      buf += k;
    }
  } // while
  return std::make_pair(buf, utf8_output);
}


/*
  Returns a pair: a result struct and utf8_output.
  If there is an error, the count field of the result is the position of the error.
  Otherwise, it is the position of the first unprocessed byte in buf (even if finished).
  A scalar routing should carry on the conversion of the tail if needed.
*/
template <endianness big_endian>
std::pair<result, char*> avx2_convert_utf16_to_utf8_with_errors(const char16_t* buf, size_t len, char* utf8_output) {
  const char16_t* start = buf;
  const char16_t* end = buf + len;

  const __m256i v_0000 = _mm256_setzero_si256();
  const __m256i v_f800 = _mm256_set1_epi16((int16_t)0xf800);
  const __m256i v_d800 = _mm256_set1_epi16((int16_t)0xd800);
  const __m256i v_c080 = _mm256_set1_epi16((int16_t)0xc080);
  const size_t safety_margin = 12; // to avoid overruns, see issue https://github.com/simdutf/simdutf/issues/92

  while (buf + 16 + safety_margin <= end) {
    __m256i in = _mm256_loadu_si256((__m256i*)buf);
    if (big_endian) {
      const __m256i swap = _mm256_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14,
                                  17, 16, 19, 18, 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
      in = _mm256_shuffle_epi8(in, swap);
    }
    // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes
    const __m256i v_ff80 = _mm256_set1_epi16((int16_t)0xff80);
    if(_mm256_testz_si256(in, v_ff80)) { // ASCII fast path!!!!
        // 1. pack the bytes
        const __m128i utf8_packed = _mm_packus_epi16(_mm256_castsi256_si128(in),_mm256_extractf128_si256(in,1));
        // 2. store (16 bytes)
        _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);
        // 3. adjust pointers
        buf += 16;
        utf8_output += 16;
        continue; // we are done for this round!
    }
    // no bits set above 7th bit
    const __m256i one_byte_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in, v_ff80), v_0000);
    const uint32_t one_byte_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));

    // no bits set above 11th bit
    const __m256i one_or_two_bytes_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_0000);
    const uint32_t one_or_two_bytes_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(one_or_two_bytes_bytemask));
    if (one_or_two_bytes_bitmask == 0xffffffff) {

          // 1. prepare 2-byte values
          // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
          // expected output   : [110a|aaaa|10bb|bbbb] x 8
          const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
          const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);

          // t0 = [000a|aaaa|bbbb|bb00]
          const __m256i t0 = _mm256_slli_epi16(in, 2);
          // t1 = [000a|aaaa|0000|0000]
          const __m256i t1 = _mm256_and_si256(t0, v_1f00);
          // t2 = [0000|0000|00bb|bbbb]
          const __m256i t2 = _mm256_and_si256(in, v_003f);
          // t3 = [000a|aaaa|00bb|bbbb]
          const __m256i t3 = _mm256_or_si256(t1, t2);
          // t4 = [110a|aaaa|10bb|bbbb]
          const __m256i t4 = _mm256_or_si256(t3, v_c080);

          // 2. merge ASCII and 2-byte codewords
          const __m256i utf8_unpacked = _mm256_blendv_epi8(t4, in, one_byte_bytemask);

          // 3. prepare bitmask for 8-bit lookup
          const uint32_t M0 = one_byte_bitmask & 0x55555555;
          const uint32_t M1 = M0 >> 7;
          const uint32_t M2 = (M1 | M0)  & 0x00ff00ff;
          // 4. pack the bytes

          const uint8_t* row = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
          const uint8_t* row_2 = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2>>16)][0];

          const __m128i shuffle = _mm_loadu_si128((__m128i*)(row + 1));
          const __m128i shuffle_2 = _mm_loadu_si128((__m128i*)(row_2 + 1));

          const __m256i utf8_packed = _mm256_shuffle_epi8(utf8_unpacked, _mm256_setr_m128i(shuffle,shuffle_2));
          // 5. store bytes
          _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_packed));
          utf8_output += row[0];
          _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_packed,1));
          utf8_output += row_2[0];

          // 6. adjust pointers
          buf += 16;
          continue;
    }
    // 1. Check if there are any surrogate word in the input chunk.
    //    We have also deal with situation when there is a surrogate word
    //    at the end of a chunk.
    const __m256i surrogates_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_d800);

    // bitmask = 0x0000 if there are no surrogates
    //         = 0xc000 if the last word is a surrogate
    const uint32_t surrogates_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(surrogates_bytemask));
    // It might seem like checking for surrogates_bitmask == 0xc000 could help. However,
    // it is likely an uncommon occurrence.
    if (surrogates_bitmask == 0x00000000) {
      // case: words from register produce either 1, 2 or 3 UTF-8 bytes
        const __m256i dup_even = _mm256_setr_epi16(0x0000, 0x0202, 0x0404, 0x0606,
                                                0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
                                                0x0000, 0x0202, 0x0404, 0x0606,
                                                0x0808, 0x0a0a, 0x0c0c, 0x0e0e);

        /* In this branch we handle three cases:
           1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           - single UFT-8 byte
           2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two UTF-8 bytes
           3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - three UTF-8 bytes

          We expand the input word (16-bit) into two words (32-bit), thus
          we have room for four bytes. However, we need five distinct bit
          layouts. Note that the last byte in cases #2 and #3 is the same.

          We precompute byte 1 for case #1 and the common byte for cases #2 & #3
          in register t2.

          We precompute byte 1 for case #3 and -- **conditionally** -- precompute
          either byte 1 for case #2 or byte 2 for case #3. Note that they
          differ by exactly one bit.

          Finally from these two words we build proper UTF-8 sequence, taking
          into account the case (i.e, the number of bytes to write).
        */
        /**
         * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
         * t2 => [0ccc|cccc] [10cc|cccc]
         * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
         */
#define simdutf_vec(x) _mm256_set1_epi16(static_cast<uint16_t>(x))
        // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
        const __m256i t0 = _mm256_shuffle_epi8(in, dup_even);
        // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
        const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111));
        // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
        const __m256i t2 = _mm256_or_si256 (t1, simdutf_vec(0b1000000000000000));

        // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
        const __m256i s0 = _mm256_srli_epi16(in, 4);
        // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
        const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100));
        // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
        const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140));
        // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
        const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000));
        const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask, simdutf_vec(0b0100000000000000));
        const __m256i s4 = _mm256_xor_si256(s3, m0);
#undef simdutf_vec

        // 4. expand words 16-bit => 32-bit
        const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
        const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);

        // 5. compress 32-bit words into 1, 2 or 3 bytes -- 2 x shuffle
        const uint32_t mask = (one_byte_bitmask & 0x55555555) |
                              (one_or_two_bytes_bitmask & 0xaaaaaaaa);
        // Due to the wider registers, the following path is less likely to be useful.
        /*if(mask == 0) {
          // We only have three-byte words. Use fast path.
          const __m256i shuffle = _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1, 2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1);
          const __m256i utf8_0 = _mm256_shuffle_epi8(out0, shuffle);
          const __m256i utf8_1 = _mm256_shuffle_epi8(out1, shuffle);
          _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
          utf8_output += 12;
          _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
          utf8_output += 12;
          _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_0,1));
          utf8_output += 12;
          _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_1,1));
          utf8_output += 12;
          buf += 16;
          continue;
        }*/
        const uint8_t mask0 = uint8_t(mask);
        const uint8_t* row0 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
        const __m128i shuffle0 = _mm_loadu_si128((__m128i*)(row0 + 1));
        const __m128i utf8_0 = _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);

        const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
        const uint8_t* row1 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
        const __m128i shuffle1 = _mm_loadu_si128((__m128i*)(row1 + 1));
        const __m128i utf8_1 = _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);

        const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
        const uint8_t* row2 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
        const __m128i shuffle2 = _mm_loadu_si128((__m128i*)(row2 + 1));
        const __m128i utf8_2 = _mm_shuffle_epi8(_mm256_extractf128_si256(out0,1), shuffle2);


        const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
        const uint8_t* row3 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
        const __m128i shuffle3 = _mm_loadu_si128((__m128i*)(row3 + 1));
        const __m128i utf8_3 = _mm_shuffle_epi8(_mm256_extractf128_si256(out1,1), shuffle3);

        _mm_storeu_si128((__m128i*)utf8_output, utf8_0);
        utf8_output += row0[0];
        _mm_storeu_si128((__m128i*)utf8_output, utf8_1);
        utf8_output += row1[0];
        _mm_storeu_si128((__m128i*)utf8_output, utf8_2);
        utf8_output += row2[0];
        _mm_storeu_si128((__m128i*)utf8_output, utf8_3);
        utf8_output += row3[0];
        buf += 16;
    // surrogate pair(s) in a register
    } else {
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // in the presence of surrogate pairs may require non-trivial tables.
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint16_t word = big_endian ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
        if((word & 0xFF80)==0) {
          *utf8_output++ = char(word);
        } else if((word & 0xF800)==0) {
          *utf8_output++ = char((word>>6) | 0b11000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else if((word &0xF800 ) != 0xD800) {
          *utf8_output++ = char((word>>12) | 0b11100000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else {
          // must be a surrogate pair
          uint16_t diff = uint16_t(word - 0xD800);
          uint16_t next_word = big_endian ? scalar::utf16::swap_bytes(buf[k+1]) : buf[k+1];
          k++;
          uint16_t diff2 = uint16_t(next_word - 0xDC00);
          if((diff | diff2) > 0x3FF)  { return std::make_pair(result(error_code::SURROGATE, buf - start + k - 1), utf8_output); }
          uint32_t value = (diff << 10) + diff2 + 0x10000;
          *utf8_output++ = char((value>>18) | 0b11110000);
          *utf8_output++ = char(((value>>12) & 0b111111) | 0b10000000);
          *utf8_output++ = char(((value>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((value & 0b111111) | 0b10000000);
        }
      }
      buf += k;
    }
  } // while
  return std::make_pair(result(error_code::SUCCESS, buf - start), utf8_output);
}
/* end file src/haswell/avx2_convert_utf16_to_utf8.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=haswell/avx2_convert_utf16_to_utf32.cpp
/* begin file src/haswell/avx2_convert_utf16_to_utf32.cpp */
/*
    The vectorized algorithm works on single SSE register i.e., it
    loads eight 16-bit words.

    We consider three cases:
    1. an input register contains no surrogates and each value
       is in range 0x0000 .. 0x07ff.
    2. an input register contains no surrogates and values are
       is in range 0x0000 .. 0xffff.
    3. an input register contains surrogates --- i.e. codepoints
       can have 16 or 32 bits.

    Ad 1.

    When values are less than 0x0800, it means that a 16-bit words
    can be converted into: 1) single UTF8 byte (when it's an ASCII
    char) or 2) two UTF8 bytes.

    For this case we do only some shuffle to obtain these 2-byte
    codes and finally compress the whole SSE register with a single
    shuffle.

    We need 256-entry lookup table to get a compression pattern
    and the number of output bytes in the compressed vector register.
    Each entry occupies 17 bytes.

    Ad 2.

    When values fit in 16-bit words, but are above 0x07ff, then
    a single word may produce one, two or three UTF8 bytes.

    We prepare data for all these three cases in two registers.
    The first register contains lower two UTF8 bytes (used in all
    cases), while the second one contains just the third byte for
    the three-UTF8-bytes case.

    Finally these two registers are interleaved forming eight-element
    array of 32-bit values. The array spans two SSE registers.
    The bytes from the registers are compressed using two shuffles.

    We need 256-entry lookup table to get a compression pattern
    and the number of output bytes in the compressed vector register.
    Each entry occupies 17 bytes.


    To summarize:
    - We need two 256-entry tables that have 8704 bytes in total.
*/


/*
  Returns a pair: the first unprocessed byte from buf and utf32_output
  A scalar routing should carry on the conversion of the tail.
*/
template <endianness big_endian>
std::pair<const char16_t*, char32_t*> avx2_convert_utf16_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) {
  const char16_t* end = buf + len;
  const __m256i v_f800 = _mm256_set1_epi16((int16_t)0xf800);
  const __m256i v_d800 = _mm256_set1_epi16((int16_t)0xd800);

  while (buf + 16 <= end) {
    __m256i in = _mm256_loadu_si256((__m256i*)buf);
    if (big_endian) {
      const __m256i swap = _mm256_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14,
                                  17, 16, 19, 18, 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
      in = _mm256_shuffle_epi8(in, swap);
    }

    // 1. Check if there are any surrogate word in the input chunk.
    //    We have also deal with situation when there is a surrogate word
    //    at the end of a chunk.
    const __m256i surrogates_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_d800);

    // bitmask = 0x0000 if there are no surrogates
    //         = 0xc000 if the last word is a surrogate
    const uint32_t surrogates_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(surrogates_bytemask));
    // It might seem like checking for surrogates_bitmask == 0xc000 could help. However,
    // it is likely an uncommon occurrence.
    if (surrogates_bitmask == 0x00000000) {
      // case: we extend all sixteen 16-bit words to sixteen 32-bit words
        _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output), _mm256_cvtepu16_epi32(_mm256_castsi256_si128(in)));
        _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output + 8), _mm256_cvtepu16_epi32(_mm256_extractf128_si256(in,1)));
        utf32_output += 16;
        buf += 16;
    // surrogate pair(s) in a register
    } else {
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // in the presence of surrogate pairs may require non-trivial tables.
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint16_t word = big_endian ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
        if((word &0xF800 ) != 0xD800) {
          // No surrogate pair
          *utf32_output++ = char32_t(word);
        } else {
          // must be a surrogate pair
          uint16_t diff = uint16_t(word - 0xD800);
          uint16_t next_word = big_endian ? scalar::utf16::swap_bytes(buf[k+1]) : buf[k+1];
          k++;
          uint16_t diff2 = uint16_t(next_word - 0xDC00);
          if((diff | diff2) > 0x3FF)  { return std::make_pair(nullptr, utf32_output); }
          uint32_t value = (diff << 10) + diff2 + 0x10000;
          *utf32_output++ = char32_t(value);
        }
      }
      buf += k;
    }
  } // while
  return std::make_pair(buf, utf32_output);
}


/*
  Returns a pair: a result struct and utf8_output.
  If there is an error, the count field of the result is the position of the error.
  Otherwise, it is the position of the first unprocessed byte in buf (even if finished).
  A scalar routing should carry on the conversion of the tail if needed.
*/
template <endianness big_endian>
std::pair<result, char32_t*> avx2_convert_utf16_to_utf32_with_errors(const char16_t* buf, size_t len, char32_t* utf32_output) {
  const char16_t* start = buf;
  const char16_t* end = buf + len;
  const __m256i v_f800 = _mm256_set1_epi16((int16_t)0xf800);
  const __m256i v_d800 = _mm256_set1_epi16((int16_t)0xd800);

  while (buf + 16 <= end) {
    __m256i in = _mm256_loadu_si256((__m256i*)buf);
    if (big_endian) {
      const __m256i swap = _mm256_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14,
                                  17, 16, 19, 18, 21, 20, 23, 22, 25, 24, 27, 26, 29, 28, 31, 30);
      in = _mm256_shuffle_epi8(in, swap);
    }

    // 1. Check if there are any surrogate word in the input chunk.
    //    We have also deal with situation when there is a surrogate word
    //    at the end of a chunk.
    const __m256i surrogates_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in, v_f800), v_d800);

    // bitmask = 0x0000 if there are no surrogates
    //         = 0xc000 if the last word is a surrogate
    const uint32_t surrogates_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(surrogates_bytemask));
    // It might seem like checking for surrogates_bitmask == 0xc000 could help. However,
    // it is likely an uncommon occurrence.
    if (surrogates_bitmask == 0x00000000) {
      // case: we extend all sixteen 16-bit words to sixteen 32-bit words
        _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output), _mm256_cvtepu16_epi32(_mm256_castsi256_si128(in)));
        _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf32_output + 8), _mm256_cvtepu16_epi32(_mm256_extractf128_si256(in,1)));
        utf32_output += 16;
        buf += 16;
    // surrogate pair(s) in a register
    } else {
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // in the presence of surrogate pairs may require non-trivial tables.
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint16_t word = big_endian ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
        if((word &0xF800 ) != 0xD800) {
          // No surrogate pair
          *utf32_output++ = char32_t(word);
        } else {
          // must be a surrogate pair
          uint16_t diff = uint16_t(word - 0xD800);
          uint16_t next_word = big_endian ? scalar::utf16::swap_bytes(buf[k+1]) : buf[k+1];
          k++;
          uint16_t diff2 = uint16_t(next_word - 0xDC00);
          if((diff | diff2) > 0x3FF)  { return std::make_pair(result(error_code::SURROGATE, buf - start + k - 1), utf32_output); }
          uint32_t value = (diff << 10) + diff2 + 0x10000;
          *utf32_output++ = char32_t(value);
        }
      }
      buf += k;
    }
  } // while
  return std::make_pair(result(error_code::SUCCESS, buf - start), utf32_output);
}
/* end file src/haswell/avx2_convert_utf16_to_utf32.cpp */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=haswell/avx2_convert_utf32_to_utf8.cpp
/* begin file src/haswell/avx2_convert_utf32_to_utf8.cpp */
std::pair<const char32_t*, char*> avx2_convert_utf32_to_utf8(const char32_t* buf, size_t len, char* utf8_output) {
  const char32_t* end = buf + len;
  const __m256i v_0000 = _mm256_setzero_si256();
  const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000);
  const __m256i v_ff80 = _mm256_set1_epi16((uint16_t)0xff80);
  const __m256i v_f800 = _mm256_set1_epi16((uint16_t)0xf800);
  const __m256i v_c080 = _mm256_set1_epi16((uint16_t)0xc080);
  const __m256i v_7fffffff = _mm256_set1_epi32((uint32_t)0x7fffffff);
  __m256i running_max = _mm256_setzero_si256();
  __m256i forbidden_bytemask = _mm256_setzero_si256();

  const size_t safety_margin = 12; // to avoid overruns, see issue https://github.com/simdutf/simdutf/issues/92

  while (buf + 16 + safety_margin <= end) {
    __m256i in = _mm256_loadu_si256((__m256i*)buf);
    __m256i nextin = _mm256_loadu_si256((__m256i*)buf+1);
    running_max = _mm256_max_epu32(_mm256_max_epu32(in, running_max), nextin);

    // Pack 32-bit UTF-32 words to 16-bit UTF-16 words with unsigned saturation
    __m256i in_16 = _mm256_packus_epi32(_mm256_and_si256(in, v_7fffffff), _mm256_and_si256(nextin, v_7fffffff));
    in_16 = _mm256_permute4x64_epi64(in_16, 0b11011000);

    // Try to apply UTF-16 => UTF-8 routine on 256 bits (haswell/avx2_convert_utf16_to_utf8.cpp)

    if(_mm256_testz_si256(in_16, v_ff80)) { // ASCII fast path!!!!
      // 1. pack the bytes
      const __m128i utf8_packed = _mm_packus_epi16(_mm256_castsi256_si128(in_16),_mm256_extractf128_si256(in_16,1));
      // 2. store (16 bytes)
      _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);
      // 3. adjust pointers
      buf += 16;
      utf8_output += 16;
      continue; // we are done for this round!
    }
    // no bits set above 7th bit
    const __m256i one_byte_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_ff80), v_0000);
    const uint32_t one_byte_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));

    // no bits set above 11th bit
    const __m256i one_or_two_bytes_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_0000);
    const uint32_t one_or_two_bytes_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(one_or_two_bytes_bytemask));
    if (one_or_two_bytes_bitmask == 0xffffffff) {
      // 1. prepare 2-byte values
      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
      // expected output   : [110a|aaaa|10bb|bbbb] x 8
      const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
      const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);

      // t0 = [000a|aaaa|bbbb|bb00]
      const __m256i t0 = _mm256_slli_epi16(in_16, 2);
      // t1 = [000a|aaaa|0000|0000]
      const __m256i t1 = _mm256_and_si256(t0, v_1f00);
      // t2 = [0000|0000|00bb|bbbb]
      const __m256i t2 = _mm256_and_si256(in_16, v_003f);
      // t3 = [000a|aaaa|00bb|bbbb]
      const __m256i t3 = _mm256_or_si256(t1, t2);
      // t4 = [110a|aaaa|10bb|bbbb]
      const __m256i t4 = _mm256_or_si256(t3, v_c080);

      // 2. merge ASCII and 2-byte codewords
      const __m256i utf8_unpacked = _mm256_blendv_epi8(t4, in_16, one_byte_bytemask);

      // 3. prepare bitmask for 8-bit lookup
      const uint32_t M0 = one_byte_bitmask & 0x55555555;
      const uint32_t M1 = M0 >> 7;
      const uint32_t M2 = (M1 | M0)  & 0x00ff00ff;
      // 4. pack the bytes

      const uint8_t* row = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
      const uint8_t* row_2 = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2>>16)][0];

      const __m128i shuffle = _mm_loadu_si128((__m128i*)(row + 1));
      const __m128i shuffle_2 = _mm_loadu_si128((__m128i*)(row_2 + 1));

      const __m256i utf8_packed = _mm256_shuffle_epi8(utf8_unpacked, _mm256_setr_m128i(shuffle,shuffle_2));
      // 5. store bytes
      _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_packed));
      utf8_output += row[0];
      _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_packed,1));
      utf8_output += row_2[0];

      // 6. adjust pointers
      buf += 16;
      continue;
    }
    // Must check for overflow in packing
    const __m256i saturation_bytemask = _mm256_cmpeq_epi32(_mm256_and_si256(_mm256_or_si256(in, nextin), v_ffff0000), v_0000);
    const uint32_t saturation_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));
    if (saturation_bitmask == 0xffffffff) {
      // case: words from register produce either 1, 2 or 3 UTF-8 bytes
      const __m256i v_d800 = _mm256_set1_epi16((uint16_t)0xd800);
      forbidden_bytemask = _mm256_or_si256(forbidden_bytemask, _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_d800));

      const __m256i dup_even = _mm256_setr_epi16(0x0000, 0x0202, 0x0404, 0x0606,
                                              0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
                                              0x0000, 0x0202, 0x0404, 0x0606,
                                              0x0808, 0x0a0a, 0x0c0c, 0x0e0e);

      /* In this branch we handle three cases:
        1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           - single UFT-8 byte
        2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two UTF-8 bytes
        3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - three UTF-8 bytes

        We expand the input word (16-bit) into two words (32-bit), thus
        we have room for four bytes. However, we need five distinct bit
        layouts. Note that the last byte in cases #2 and #3 is the same.

        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
        in register t2.

        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
        either byte 1 for case #2 or byte 2 for case #3. Note that they
        differ by exactly one bit.

        Finally from these two words we build proper UTF-8 sequence, taking
        into account the case (i.e, the number of bytes to write).
      */
      /**
       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
       * t2 => [0ccc|cccc] [10cc|cccc]
       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
       */
#define simdutf_vec(x) _mm256_set1_epi16(static_cast<uint16_t>(x))
      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
      const __m256i t0 = _mm256_shuffle_epi8(in_16, dup_even);
      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
      const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111));
      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
      const __m256i t2 = _mm256_or_si256 (t1, simdutf_vec(0b1000000000000000));

      // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
      const __m256i s0 = _mm256_srli_epi16(in_16, 4);
      // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
      const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100));
      // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
      const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140));
      // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
      const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000));
      const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask, simdutf_vec(0b0100000000000000));
      const __m256i s4 = _mm256_xor_si256(s3, m0);
#undef simdutf_vec

      // 4. expand words 16-bit => 32-bit
      const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
      const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);

      // 5. compress 32-bit words into 1, 2 or 3 bytes -- 2 x shuffle
      const uint32_t mask = (one_byte_bitmask & 0x55555555) |
                            (one_or_two_bytes_bitmask & 0xaaaaaaaa);
      // Due to the wider registers, the following path is less likely to be useful.
      /*if(mask == 0) {
        // We only have three-byte words. Use fast path.
        const __m256i shuffle = _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1, 2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1);
        const __m256i utf8_0 = _mm256_shuffle_epi8(out0, shuffle);
        const __m256i utf8_1 = _mm256_shuffle_epi8(out1, shuffle);
        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
        utf8_output += 12;
        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
        utf8_output += 12;
        _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_0,1));
        utf8_output += 12;
        _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_1,1));
        utf8_output += 12;
        buf += 16;
        continue;
      }*/
      const uint8_t mask0 = uint8_t(mask);
      const uint8_t* row0 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
      const __m128i shuffle0 = _mm_loadu_si128((__m128i*)(row0 + 1));
      const __m128i utf8_0 = _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);

      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
      const uint8_t* row1 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
      const __m128i shuffle1 = _mm_loadu_si128((__m128i*)(row1 + 1));
      const __m128i utf8_1 = _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);

      const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
      const uint8_t* row2 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
      const __m128i shuffle2 = _mm_loadu_si128((__m128i*)(row2 + 1));
      const __m128i utf8_2 = _mm_shuffle_epi8(_mm256_extractf128_si256(out0,1), shuffle2);


      const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
      const uint8_t* row3 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
      const __m128i shuffle3 = _mm_loadu_si128((__m128i*)(row3 + 1));
      const __m128i utf8_3 = _mm_shuffle_epi8(_mm256_extractf128_si256(out1,1), shuffle3);

      _mm_storeu_si128((__m128i*)utf8_output, utf8_0);
      utf8_output += row0[0];
      _mm_storeu_si128((__m128i*)utf8_output, utf8_1);
      utf8_output += row1[0];
      _mm_storeu_si128((__m128i*)utf8_output, utf8_2);
      utf8_output += row2[0];
      _mm_storeu_si128((__m128i*)utf8_output, utf8_3);
      utf8_output += row3[0];
      buf += 16;
    } else {
      // case: at least one 32-bit word is larger than 0xFFFF <=> it will produce four UTF-8 bytes.
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // may require large, non-trivial tables?
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint32_t word = buf[k];
        if((word & 0xFFFFFF80)==0) {  // 1-byte (ASCII)
          *utf8_output++ = char(word);
        } else if((word & 0xFFFFF800)==0) { // 2-byte
          *utf8_output++ = char((word>>6) | 0b11000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else if((word & 0xFFFF0000 )==0) {  // 3-byte
          if (word >= 0xD800 && word <= 0xDFFF) { return std::make_pair(nullptr, utf8_output); }
          *utf8_output++ = char((word>>12) | 0b11100000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else {  // 4-byte
          if (word > 0x10FFFF) { return std::make_pair(nullptr, utf8_output); }
          *utf8_output++ = char((word>>18) | 0b11110000);
          *utf8_output++ = char(((word>>12) & 0b111111) | 0b10000000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        }
      }
      buf += k;
    }
  } // while

  // check for invalid input
  const __m256i v_10ffff = _mm256_set1_epi32((uint32_t)0x10ffff);
  if(static_cast<uint32_t>(_mm256_movemask_epi8(_mm256_cmpeq_epi32(_mm256_max_epu32(running_max, v_10ffff), v_10ffff))) != 0xffffffff) {
    return std::make_pair(nullptr, utf8_output);
  }

  if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) != 0) { return std::make_pair(nullptr, utf8_output); }

  return std::make_pair(buf, utf8_output);
}


std::pair<result, char*> avx2_convert_utf32_to_utf8_with_errors(const char32_t* buf, size_t len, char* utf8_output) {
  const char32_t* end = buf + len;
  const char32_t* start = buf;

  const __m256i v_0000 = _mm256_setzero_si256();
  const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000);
  const __m256i v_ff80 = _mm256_set1_epi16((uint16_t)0xff80);
  const __m256i v_f800 = _mm256_set1_epi16((uint16_t)0xf800);
  const __m256i v_c080 = _mm256_set1_epi16((uint16_t)0xc080);
  const __m256i v_7fffffff = _mm256_set1_epi32((uint32_t)0x7fffffff);
  const __m256i v_10ffff = _mm256_set1_epi32((uint32_t)0x10ffff);

  const size_t safety_margin = 12; // to avoid overruns, see issue https://github.com/simdutf/simdutf/issues/92

  while (buf + 16 + safety_margin <= end) {
    __m256i in = _mm256_loadu_si256((__m256i*)buf);
    __m256i nextin = _mm256_loadu_si256((__m256i*)buf+1);
    // Check for too large input
    const __m256i max_input = _mm256_max_epu32(_mm256_max_epu32(in, nextin), v_10ffff);
    if(static_cast<uint32_t>(_mm256_movemask_epi8(_mm256_cmpeq_epi32(max_input, v_10ffff))) != 0xffffffff) {
      return std::make_pair(result(error_code::TOO_LARGE, buf - start), utf8_output);
    }

    // Pack 32-bit UTF-32 words to 16-bit UTF-16 words with unsigned saturation
    __m256i in_16 = _mm256_packus_epi32(_mm256_and_si256(in, v_7fffffff), _mm256_and_si256(nextin, v_7fffffff));
    in_16 = _mm256_permute4x64_epi64(in_16, 0b11011000);

    // Try to apply UTF-16 => UTF-8 routine on 256 bits (haswell/avx2_convert_utf16_to_utf8.cpp)

    if(_mm256_testz_si256(in_16, v_ff80)) { // ASCII fast path!!!!
      // 1. pack the bytes
      const __m128i utf8_packed = _mm_packus_epi16(_mm256_castsi256_si128(in_16),_mm256_extractf128_si256(in_16,1));
      // 2. store (16 bytes)
      _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);
      // 3. adjust pointers
      buf += 16;
      utf8_output += 16;
      continue; // we are done for this round!
    }
    // no bits set above 7th bit
    const __m256i one_byte_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_ff80), v_0000);
    const uint32_t one_byte_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(one_byte_bytemask));

    // no bits set above 11th bit
    const __m256i one_or_two_bytes_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_0000);
    const uint32_t one_or_two_bytes_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(one_or_two_bytes_bytemask));
    if (one_or_two_bytes_bitmask == 0xffffffff) {
      // 1. prepare 2-byte values
      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
      // expected output   : [110a|aaaa|10bb|bbbb] x 8
      const __m256i v_1f00 = _mm256_set1_epi16((int16_t)0x1f00);
      const __m256i v_003f = _mm256_set1_epi16((int16_t)0x003f);

      // t0 = [000a|aaaa|bbbb|bb00]
      const __m256i t0 = _mm256_slli_epi16(in_16, 2);
      // t1 = [000a|aaaa|0000|0000]
      const __m256i t1 = _mm256_and_si256(t0, v_1f00);
      // t2 = [0000|0000|00bb|bbbb]
      const __m256i t2 = _mm256_and_si256(in_16, v_003f);
      // t3 = [000a|aaaa|00bb|bbbb]
      const __m256i t3 = _mm256_or_si256(t1, t2);
      // t4 = [110a|aaaa|10bb|bbbb]
      const __m256i t4 = _mm256_or_si256(t3, v_c080);

      // 2. merge ASCII and 2-byte codewords
      const __m256i utf8_unpacked = _mm256_blendv_epi8(t4, in_16, one_byte_bytemask);

      // 3. prepare bitmask for 8-bit lookup
      const uint32_t M0 = one_byte_bitmask & 0x55555555;
      const uint32_t M1 = M0 >> 7;
      const uint32_t M2 = (M1 | M0)  & 0x00ff00ff;
      // 4. pack the bytes

      const uint8_t* row = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2)][0];
      const uint8_t* row_2 = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[uint8_t(M2>>16)][0];

      const __m128i shuffle = _mm_loadu_si128((__m128i*)(row + 1));
      const __m128i shuffle_2 = _mm_loadu_si128((__m128i*)(row_2 + 1));

      const __m256i utf8_packed = _mm256_shuffle_epi8(utf8_unpacked, _mm256_setr_m128i(shuffle,shuffle_2));
      // 5. store bytes
      _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_packed));
      utf8_output += row[0];
      _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_packed,1));
      utf8_output += row_2[0];

      // 6. adjust pointers
      buf += 16;
      continue;
    }
    // Must check for overflow in packing
    const __m256i saturation_bytemask = _mm256_cmpeq_epi32(_mm256_and_si256(_mm256_or_si256(in, nextin), v_ffff0000), v_0000);
    const uint32_t saturation_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));
    if (saturation_bitmask == 0xffffffff) {
      // case: words from register produce either 1, 2 or 3 UTF-8 bytes

      // Check for illegal surrogate words
      const __m256i v_d800 = _mm256_set1_epi16((uint16_t)0xd800);
      const __m256i forbidden_bytemask = _mm256_cmpeq_epi16(_mm256_and_si256(in_16, v_f800), v_d800);
      if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) != 0x0) {
        return std::make_pair(result(error_code::SURROGATE, buf - start), utf8_output);
      }

      const __m256i dup_even = _mm256_setr_epi16(0x0000, 0x0202, 0x0404, 0x0606,
                                              0x0808, 0x0a0a, 0x0c0c, 0x0e0e,
                                              0x0000, 0x0202, 0x0404, 0x0606,
                                              0x0808, 0x0a0a, 0x0c0c, 0x0e0e);

      /* In this branch we handle three cases:
        1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           - single UFT-8 byte
        2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two UTF-8 bytes
        3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - three UTF-8 bytes

        We expand the input word (16-bit) into two words (32-bit), thus
        we have room for four bytes. However, we need five distinct bit
        layouts. Note that the last byte in cases #2 and #3 is the same.

        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
        in register t2.

        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
        either byte 1 for case #2 or byte 2 for case #3. Note that they
        differ by exactly one bit.

        Finally from these two words we build proper UTF-8 sequence, taking
        into account the case (i.e, the number of bytes to write).
      */
      /**
       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
       * t2 => [0ccc|cccc] [10cc|cccc]
       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
       */
#define simdutf_vec(x) _mm256_set1_epi16(static_cast<uint16_t>(x))
      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
      const __m256i t0 = _mm256_shuffle_epi8(in_16, dup_even);
      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
      const __m256i t1 = _mm256_and_si256(t0, simdutf_vec(0b0011111101111111));
      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
      const __m256i t2 = _mm256_or_si256 (t1, simdutf_vec(0b1000000000000000));

      // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
      const __m256i s0 = _mm256_srli_epi16(in_16, 4);
      // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
      const __m256i s1 = _mm256_and_si256(s0, simdutf_vec(0b0000111111111100));
      // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
      const __m256i s2 = _mm256_maddubs_epi16(s1, simdutf_vec(0x0140));
      // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
      const __m256i s3 = _mm256_or_si256(s2, simdutf_vec(0b1100000011100000));
      const __m256i m0 = _mm256_andnot_si256(one_or_two_bytes_bytemask, simdutf_vec(0b0100000000000000));
      const __m256i s4 = _mm256_xor_si256(s3, m0);
#undef simdutf_vec

      // 4. expand words 16-bit => 32-bit
      const __m256i out0 = _mm256_unpacklo_epi16(t2, s4);
      const __m256i out1 = _mm256_unpackhi_epi16(t2, s4);

      // 5. compress 32-bit words into 1, 2 or 3 bytes -- 2 x shuffle
      const uint32_t mask = (one_byte_bitmask & 0x55555555) |
                            (one_or_two_bytes_bitmask & 0xaaaaaaaa);
      // Due to the wider registers, the following path is less likely to be useful.
      /*if(mask == 0) {
        // We only have three-byte words. Use fast path.
        const __m256i shuffle = _mm256_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1, 2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1);
        const __m256i utf8_0 = _mm256_shuffle_epi8(out0, shuffle);
        const __m256i utf8_1 = _mm256_shuffle_epi8(out1, shuffle);
        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_0));
        utf8_output += 12;
        _mm_storeu_si128((__m128i*)utf8_output, _mm256_castsi256_si128(utf8_1));
        utf8_output += 12;
        _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_0,1));
        utf8_output += 12;
        _mm_storeu_si128((__m128i*)utf8_output, _mm256_extractf128_si256(utf8_1,1));
        utf8_output += 12;
        buf += 16;
        continue;
      }*/
      const uint8_t mask0 = uint8_t(mask);
      const uint8_t* row0 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
      const __m128i shuffle0 = _mm_loadu_si128((__m128i*)(row0 + 1));
      const __m128i utf8_0 = _mm_shuffle_epi8(_mm256_castsi256_si128(out0), shuffle0);

      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);
      const uint8_t* row1 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
      const __m128i shuffle1 = _mm_loadu_si128((__m128i*)(row1 + 1));
      const __m128i utf8_1 = _mm_shuffle_epi8(_mm256_castsi256_si128(out1), shuffle1);

      const uint8_t mask2 = static_cast<uint8_t>(mask >> 16);
      const uint8_t* row2 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask2][0];
      const __m128i shuffle2 = _mm_loadu_si128((__m128i*)(row2 + 1));
      const __m128i utf8_2 = _mm_shuffle_epi8(_mm256_extractf128_si256(out0,1), shuffle2);


      const uint8_t mask3 = static_cast<uint8_t>(mask >> 24);
      const uint8_t* row3 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask3][0];
      const __m128i shuffle3 = _mm_loadu_si128((__m128i*)(row3 + 1));
      const __m128i utf8_3 = _mm_shuffle_epi8(_mm256_extractf128_si256(out1,1), shuffle3);

      _mm_storeu_si128((__m128i*)utf8_output, utf8_0);
      utf8_output += row0[0];
      _mm_storeu_si128((__m128i*)utf8_output, utf8_1);
      utf8_output += row1[0];
      _mm_storeu_si128((__m128i*)utf8_output, utf8_2);
      utf8_output += row2[0];
      _mm_storeu_si128((__m128i*)utf8_output, utf8_3);
      utf8_output += row3[0];
      buf += 16;
    } else {
      // case: at least one 32-bit word is larger than 0xFFFF <=> it will produce four UTF-8 bytes.
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // may require large, non-trivial tables?
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint32_t word = buf[k];
        if((word & 0xFFFFFF80)==0) {  // 1-byte (ASCII)
          *utf8_output++ = char(word);
        } else if((word & 0xFFFFF800)==0) { // 2-byte
          *utf8_output++ = char((word>>6) | 0b11000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else if((word & 0xFFFF0000 )==0) {  // 3-byte
          if (word >= 0xD800 && word <= 0xDFFF) { return std::make_pair(result(error_code::SURROGATE, buf - start + k), utf8_output); }
          *utf8_output++ = char((word>>12) | 0b11100000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else {  // 4-byte
          if (word > 0x10FFFF) { return std::make_pair(result(error_code::TOO_LARGE, buf - start + k), utf8_output); }
          *utf8_output++ = char((word>>18) | 0b11110000);
          *utf8_output++ = char(((word>>12) & 0b111111) | 0b10000000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        }
      }
      buf += k;
    }
  } // while

  return std::make_pair(result(error_code::SUCCESS, buf - start), utf8_output);
}
/* end file src/haswell/avx2_convert_utf32_to_utf8.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=haswell/avx2_convert_utf32_to_utf16.cpp
/* begin file src/haswell/avx2_convert_utf32_to_utf16.cpp */
template <endianness big_endian>
std::pair<const char32_t*, char16_t*> avx2_convert_utf32_to_utf16(const char32_t* buf, size_t len, char16_t* utf16_output) {
  const char32_t* end = buf + len;

  const size_t safety_margin = 12; // to avoid overruns, see issue https://github.com/simdutf/simdutf/issues/92
  __m256i forbidden_bytemask = _mm256_setzero_si256();


  while (buf + 8 + safety_margin <= end) {
    __m256i in = _mm256_loadu_si256((__m256i*)buf);

    const __m256i v_00000000 = _mm256_setzero_si256();
    const __m256i v_ffff0000 = _mm256_set1_epi32((int32_t)0xffff0000);

    // no bits set above 16th bit <=> can pack to UTF16 without surrogate pairs
    const __m256i saturation_bytemask = _mm256_cmpeq_epi32(_mm256_and_si256(in, v_ffff0000), v_00000000);
    const uint32_t saturation_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));

    if (saturation_bitmask == 0xffffffff) {
      const __m256i v_f800 = _mm256_set1_epi32((uint32_t)0xf800);
      const __m256i v_d800 = _mm256_set1_epi32((uint32_t)0xd800);
      forbidden_bytemask = _mm256_or_si256(forbidden_bytemask, _mm256_cmpeq_epi32(_mm256_and_si256(in, v_f800), v_d800));

      __m128i utf16_packed = _mm_packus_epi32(_mm256_castsi256_si128(in),_mm256_extractf128_si256(in,1));
      if (big_endian) {
        const __m128i swap = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
        utf16_packed = _mm_shuffle_epi8(utf16_packed, swap);
      }
      _mm_storeu_si128((__m128i*)utf16_output, utf16_packed);
      utf16_output += 8;
      buf += 8;
    } else {
      size_t forward = 7;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint32_t word = buf[k];
        if((word & 0xFFFF0000)==0) {
          // will not generate a surrogate pair
          if (word >= 0xD800 && word <= 0xDFFF) { return std::make_pair(nullptr, utf16_output); }
          *utf16_output++ = big_endian ? char16_t((uint16_t(word) >> 8) | (uint16_t(word) << 8)) : char16_t(word);
        } else {
          // will generate a surrogate pair
          if (word > 0x10FFFF) { return std::make_pair(nullptr, utf16_output); }
          word -= 0x10000;
          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
          if (big_endian) {
            high_surrogate = uint16_t((high_surrogate >> 8) | (high_surrogate << 8));
            low_surrogate = uint16_t((low_surrogate >> 8) | (low_surrogate << 8));
          }
          *utf16_output++ = char16_t(high_surrogate);
          *utf16_output++ = char16_t(low_surrogate);
        }
      }
      buf += k;
    }
  }

  // check for invalid input
  if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) != 0) { return std::make_pair(nullptr, utf16_output); }

  return std::make_pair(buf, utf16_output);
}


template <endianness big_endian>
std::pair<result, char16_t*> avx2_convert_utf32_to_utf16_with_errors(const char32_t* buf, size_t len, char16_t* utf16_output) {
  const char32_t* start = buf;
  const char32_t* end = buf + len;

  const size_t safety_margin = 12; // to avoid overruns, see issue https://github.com/simdutf/simdutf/issues/92

  while (buf + 8 + safety_margin <= end) {
    __m256i in = _mm256_loadu_si256((__m256i*)buf);

    const __m256i v_00000000 = _mm256_setzero_si256();
    const __m256i v_ffff0000 = _mm256_set1_epi32((int32_t)0xffff0000);

    // no bits set above 16th bit <=> can pack to UTF16 without surrogate pairs
    const __m256i saturation_bytemask = _mm256_cmpeq_epi32(_mm256_and_si256(in, v_ffff0000), v_00000000);
    const uint32_t saturation_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(saturation_bytemask));

    if (saturation_bitmask == 0xffffffff) {
      const __m256i v_f800 = _mm256_set1_epi32((uint32_t)0xf800);
      const __m256i v_d800 = _mm256_set1_epi32((uint32_t)0xd800);
      const __m256i forbidden_bytemask = _mm256_cmpeq_epi32(_mm256_and_si256(in, v_f800), v_d800);
      if (static_cast<uint32_t>(_mm256_movemask_epi8(forbidden_bytemask)) != 0x0) {
        return std::make_pair(result(error_code::SURROGATE, buf - start), utf16_output);
      }

      __m128i utf16_packed = _mm_packus_epi32(_mm256_castsi256_si128(in),_mm256_extractf128_si256(in,1));
      if (big_endian) {
        const __m128i swap = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
        utf16_packed = _mm_shuffle_epi8(utf16_packed, swap);
      }
      _mm_storeu_si128((__m128i*)utf16_output, utf16_packed);
      utf16_output += 8;
      buf += 8;
    } else {
      size_t forward = 7;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint32_t word = buf[k];
        if((word & 0xFFFF0000)==0) {
          // will not generate a surrogate pair
          if (word >= 0xD800 && word <= 0xDFFF) { return std::make_pair(result(error_code::SURROGATE, buf - start + k), utf16_output); }
          *utf16_output++ = big_endian ? char16_t((uint16_t(word) >> 8) | (uint16_t(word) << 8)) : char16_t(word);
        } else {
          // will generate a surrogate pair
          if (word > 0x10FFFF) { return std::make_pair(result(error_code::TOO_LARGE, buf - start + k), utf16_output); }
          word -= 0x10000;
          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
          if (big_endian) {
            high_surrogate = uint16_t((high_surrogate >> 8) | (high_surrogate << 8));
            low_surrogate = uint16_t((low_surrogate >> 8) | (low_surrogate << 8));
          }
          *utf16_output++ = char16_t(high_surrogate);
          *utf16_output++ = char16_t(low_surrogate);
        }
      }
      buf += k;
    }
  }

  return std::make_pair(result(error_code::SUCCESS, buf - start), utf16_output);
}
/* end file src/haswell/avx2_convert_utf32_to_utf16.cpp */
} // unnamed namespace
} // namespace haswell
} // namespace simdutf

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/buf_block_reader.h
/* begin file src/generic/buf_block_reader.h */
namespace simdutf {
namespace haswell {
namespace {

// Walks through a buffer in block-sized increments, loading the last part with spaces
template<size_t STEP_SIZE>
struct buf_block_reader {
public:
  simdutf_really_inline buf_block_reader(const uint8_t *_buf, size_t _len);
  simdutf_really_inline size_t block_index();
  simdutf_really_inline bool has_full_block() const;
  simdutf_really_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.
   */
  simdutf_really_inline size_t get_remainder(uint8_t *dst) const;
  simdutf_really_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
simdutf_unused static char * format_input_text_64(const uint8_t *text) {
  static char *buf = reinterpret_cast<char*>(malloc(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
simdutf_unused static char * format_input_text(const simd8x64<uint8_t>& in) {
  static char *buf = reinterpret_cast<char*>(malloc(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;
}

simdutf_unused static char * format_mask(uint64_t mask) {
  static char *buf = reinterpret_cast<char*>(malloc(64 + 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>
simdutf_really_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>
simdutf_really_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }

template<size_t STEP_SIZE>
simdutf_really_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
  return idx < lenminusstep;
}

template<size_t STEP_SIZE>
simdutf_really_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
  return &buf[idx];
}

template<size_t STEP_SIZE>
simdutf_really_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>
simdutf_really_inline void buf_block_reader<STEP_SIZE>::advance() {
  idx += STEP_SIZE;
}

} // unnamed namespace
} // namespace haswell
} // namespace simdutf
/* end file src/generic/buf_block_reader.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_validation/utf8_lookup4_algorithm.h
/* begin file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
namespace simdutf {
namespace haswell {
namespace {
namespace utf8_validation {

using namespace simd;

  simdutf_really_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);
  }
  simdutf_really_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 = simd8<uint8_t>(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.
  //
  simdutf_really_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______
    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, 0b11110000u-1, 0b11100000u-1, 0b11000000u-1
    };
    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.
    //
    simdutf_really_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.
    simdutf_really_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;
    }

    simdutf_really_inline void check_next_input(const simd8x64<uint8_t>& input) {
      if(simdutf_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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
            "We support either two or four chunks per 64-byte block.");
        if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
          this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
        } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
          this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
        }
        this->prev_incomplete = is_incomplete(input.chunks[simd8x64<uint8_t>::NUM_CHUNKS-1]);
        this->prev_input_block = input.chunks[simd8x64<uint8_t>::NUM_CHUNKS-1];

      }
    }

    // do not forget to call check_eof!
    simdutf_really_inline bool errors() const {
      return this->error.any_bits_set_anywhere();
    }

  }; // struct utf8_checker
} // namespace utf8_validation

using utf8_validation::utf8_checker;

} // unnamed namespace
} // namespace haswell
} // namespace simdutf
/* end file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_validation/utf8_validator.h
/* begin file src/generic/utf8_validation/utf8_validator.h */
namespace simdutf {
namespace haswell {
namespace {
namespace utf8_validation {

/**
 * 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();
}

bool generic_validate_utf8(const char * input, size_t length) {
  return generic_validate_utf8<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
}

/**
 * Validates that the string is actual UTF-8 and stops on errors.
 */
template<class checker>
result generic_validate_utf8_with_errors(const uint8_t * input, size_t length) {
    checker c{};
    buf_block_reader<64> reader(input, length);
    size_t count{0};
    while (reader.has_full_block()) {
      simd::simd8x64<uint8_t> in(reader.full_block());
      c.check_next_input(in);
      if(c.errors()) {
        if (count != 0) { count--; } // Sometimes the error is only detected in the next chunk
        result res = scalar::utf8::rewind_and_validate_with_errors(reinterpret_cast<const char*>(input + count), length - count);
        res.count += count;
        return res;
      }
      reader.advance();
      count += 64;
    }
    uint8_t block[64]{};
    reader.get_remainder(block);
    simd::simd8x64<uint8_t> in(block);
    c.check_next_input(in);
    reader.advance();
    c.check_eof();
    if (c.errors()) {
      result res = scalar::utf8::rewind_and_validate_with_errors(reinterpret_cast<const char*>(input) + count, length - count);
      res.count += count;
      return res;
    } else {
      return result(error_code::SUCCESS, length);
    }
}

result generic_validate_utf8_with_errors(const char * input, size_t length) {
  return generic_validate_utf8_with_errors<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
}

template<class checker>
bool generic_validate_ascii(const uint8_t * input, size_t length) {
    buf_block_reader<64> reader(input, length);
    uint8_t blocks[64]{};
    simd::simd8x64<uint8_t> running_or(blocks);
    while (reader.has_full_block()) {
      simd::simd8x64<uint8_t> in(reader.full_block());
      running_or |= in;
      reader.advance();
    }
    uint8_t block[64]{};
    reader.get_remainder(block);
    simd::simd8x64<uint8_t> in(block);
    running_or |= in;
    return running_or.is_ascii();
}

bool generic_validate_ascii(const char * input, size_t length) {
  return generic_validate_ascii<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
}

template<class checker>
result generic_validate_ascii_with_errors(const uint8_t * input, size_t length) {
  buf_block_reader<64> reader(input, length);
  size_t count{0};
  while (reader.has_full_block()) {
    simd::simd8x64<uint8_t> in(reader.full_block());
    if (!in.is_ascii()) {
      result res = scalar::ascii::validate_with_errors(reinterpret_cast<const char*>(input + count), length - count);
      return result(res.error, count + res.count);
    }
    reader.advance();

    count += 64;
  }
  uint8_t block[64]{};
  reader.get_remainder(block);
  simd::simd8x64<uint8_t> in(block);
  if (!in.is_ascii()) {
    result res = scalar::ascii::validate_with_errors(reinterpret_cast<const char*>(input + count), length - count);
    return result(res.error, count + res.count);
  } else {
    return result(error_code::SUCCESS, length);
  }
}

result generic_validate_ascii_with_errors(const char * input, size_t length) {
  return generic_validate_ascii_with_errors<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
}

} // namespace utf8_validation
} // unnamed namespace
} // namespace haswell
} // namespace simdutf
/* end file src/generic/utf8_validation/utf8_validator.h */
// transcoding from UTF-8 to UTF-16
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_to_utf16/valid_utf8_to_utf16.h
/* begin file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */


namespace simdutf {
namespace haswell {
namespace {
namespace utf8_to_utf16 {

using namespace simd;

template <endianness endian>
simdutf_warn_unused size_t convert_valid(const char* input, size_t size,
    char16_t* utf16_output) noexcept {
  // The implementation is not specific to haswell and should be moved to the generic directory.
  size_t pos = 0;
  char16_t* start{utf16_output};
  const size_t safety_margin = 16; // to avoid overruns!
  while(pos + 64 + safety_margin <= size) {
    // this loop could be unrolled further. For example, we could process the mask
    // far more than 64 bytes.
    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
    if(in.is_ascii()) {
      in.store_ascii_as_utf16<endian>(utf16_output);
      utf16_output += 64;
      pos += 64;
    } else {
      // Slow path. We hope that the compiler will recognize that this is a slow path.
      // Anything that is not a continuation mask is a 'leading byte', that is, the
      // start of a new code point.
      uint64_t utf8_continuation_mask = in.lt(-65 + 1);
      // -65 is 0b10111111 in two-complement's, so largest possible continuation byte
      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
      // The *start* of code points is not so useful, rather, we want the *end* of code points.
      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
      // We process in blocks of up to 12 bytes except possibly
      // for fast paths which may process up to 16 bytes. For the
      // slow path to work, we should have at least 12 input bytes left.
      size_t max_starting_point = (pos + 64) - 12;
      // Next loop is going to run at least five times when using solely
      // the slow/regular path, and at least four times if there are fast paths.
      while(pos < max_starting_point) {
        // Performance note: our ability to compute 'consumed' and
        // then shift and recompute is critical. If there is a
        // latency of, say, 4 cycles on getting 'consumed', then
        // the inner loop might have a total latency of about 6 cycles.
        // Yet we process between 6 to 12 inputs bytes, thus we get
        // a speed limit between 1 cycle/byte and 0.5 cycle/byte
        // for this section of the code. Hence, there is a limit
        // to how much we can further increase this latency before
        // it seriously harms performance.
        //
        // Thus we may allow convert_masked_utf8_to_utf16 to process
        // more bytes at a time under a fast-path mode where 16 bytes
        // are consumed at once (e.g., when encountering ASCII).
        size_t consumed = convert_masked_utf8_to_utf16<endian>(input + pos,
                            utf8_end_of_code_point_mask, utf16_output);
        pos += consumed;
        utf8_end_of_code_point_mask >>= consumed;
      }
      // At this point there may remain between 0 and 12 bytes in the
      // 64-byte block. These bytes will be processed again. So we have an
      // 80% efficiency (in the worst case). In practice we expect an
      // 85% to 90% efficiency.
    }
  }
  utf16_output += scalar::utf8_to_utf16::convert_valid<endian>(input + pos, size - pos, utf16_output);
  return utf16_output - start;
}

} // namespace utf8_to_utf16
} // unnamed namespace
} // namespace haswell
} // namespace simdutf
/* end file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_to_utf16/utf8_to_utf16.h
/* begin file src/generic/utf8_to_utf16/utf8_to_utf16.h */


namespace simdutf {
namespace haswell {
namespace {
namespace utf8_to_utf16 {
using namespace simd;


  simdutf_really_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);
  }
  simdutf_really_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 = simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
    simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
    return must23_80 ^ sc;
  }


  struct validating_transcoder {
    // If this is nonzero, there has been a UTF-8 error.
    simd8<uint8_t> error;

    validating_transcoder() : error(uint8_t(0)) {}
    //
    // Check whether the current bytes are valid UTF-8.
    //
    simdutf_really_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);
    }


    template <endianness endian>
    simdutf_really_inline size_t convert(const char* in, size_t size, char16_t* utf16_output) {
      size_t pos = 0;
      char16_t* start{utf16_output};
      // In the worst case, we have the haswell kernel which can cause an overflow of
      // 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the last 16 bytes,
      // and if the data is valid, then it is entirely safe because 16 UTF-8 bytes generate
      // much more than 8 bytes. However, you cannot generally assume that you have valid
      // UTF-8 input, so we are going to go back from the end counting 8 leading bytes,
      // to give us a good margin.
      size_t leading_byte = 0;
      size_t margin = size;
      for(; margin > 0 && leading_byte < 8; margin--) {
        leading_byte += (int8_t(in[margin-1]) > -65);
      }
      // If the input is long enough, then we have that margin-1 is the eight last leading byte.
      const size_t safety_margin = size - margin + 1; // to avoid overruns!
      while(pos + 64 + safety_margin <= size) {
        simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
        if(input.is_ascii()) {
          input.store_ascii_as_utf16<endian>(utf16_output);
          utf16_output += 64;
          pos += 64;
        } 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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
              "We support either two or four chunks per 64-byte block.");
          auto zero = simd8<uint8_t>{uint8_t(0)};
          if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
            this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
            this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
          }
          uint64_t utf8_continuation_mask = input.lt(-65 + 1);
          uint64_t utf8_leading_mask = ~utf8_continuation_mask;
          uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
          // We process in blocks of up to 12 bytes except possibly
          // for fast paths which may process up to 16 bytes. For the
          // slow path to work, we should have at least 12 input bytes left.
          size_t max_starting_point = (pos + 64) - 12;
          // Next loop is going to run at least five times.
          while(pos < max_starting_point) {
            // Performance note: our ability to compute 'consumed' and
            // then shift and recompute is critical. If there is a
            // latency of, say, 4 cycles on getting 'consumed', then
            // the inner loop might have a total latency of about 6 cycles.
            // Yet we process between 6 to 12 inputs bytes, thus we get
            // a speed limit between 1 cycle/byte and 0.5 cycle/byte
            // for this section of the code. Hence, there is a limit
            // to how much we can further increase this latency before
            // it seriously harms performance.
            size_t consumed = convert_masked_utf8_to_utf16<endian>(in + pos,
                            utf8_end_of_code_point_mask, utf16_output);
            pos += consumed;
            utf8_end_of_code_point_mask >>= consumed;
          }
          // At this point there may remain between 0 and 12 bytes in the
          // 64-byte block. These bytes will be processed again. So we have an
          // 80% efficiency (in the worst case). In practice we expect an
          // 85% to 90% efficiency.
        }
      }
      if(errors()) { return 0; }
      if(pos < size) {
        size_t howmany  = scalar::utf8_to_utf16::convert<endian>(in + pos, size - pos, utf16_output);
        if(howmany == 0) { return 0; }
        utf16_output += howmany;
      }
      return utf16_output - start;
    }

    template <endianness endian>
    simdutf_really_inline result convert_with_errors(const char* in, size_t size, char16_t* utf16_output) {
      size_t pos = 0;
      char16_t* start{utf16_output};
      // In the worst case, we have the haswell kernel which can cause an overflow of
      // 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the last 16 bytes,
      // and if the data is valid, then it is entirely safe because 16 UTF-8 bytes generate
      // much more than 8 bytes. However, you cannot generally assume that you have valid
      // UTF-8 input, so we are going to go back from the end counting 8 leading bytes,
      // to give us a good margin.
      size_t leading_byte = 0;
      size_t margin = size;
      for(; margin > 0 && leading_byte < 8; margin--) {
        leading_byte += (int8_t(in[margin-1]) > -65);
      }
      // If the input is long enough, then we have that margin-1 is the eight last leading byte.
      const size_t safety_margin = size - margin + 1; // to avoid overruns!
      while(pos + 64 + safety_margin <= size) {
        simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
        if(input.is_ascii()) {
          input.store_ascii_as_utf16<endian>(utf16_output);
          utf16_output += 64;
          pos += 64;
        } 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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
              "We support either two or four chunks per 64-byte block.");
          auto zero = simd8<uint8_t>{uint8_t(0)};
          if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
            this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
            this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
          }
          if (errors()) {
            // rewind_and_convert_with_errors will seek a potential error from in+pos onward,
            // with the ability to go back up to pos bytes, and read size-pos bytes forward.
            result res = scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(pos, in + pos, size - pos, utf16_output);
            res.count += pos;
            return res;
          }
          uint64_t utf8_continuation_mask = input.lt(-65 + 1);
          uint64_t utf8_leading_mask = ~utf8_continuation_mask;
          uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
          // We process in blocks of up to 12 bytes except possibly
          // for fast paths which may process up to 16 bytes. For the
          // slow path to work, we should have at least 12 input bytes left.
          size_t max_starting_point = (pos + 64) - 12;
          // Next loop is going to run at least five times.
          while(pos < max_starting_point) {
            // Performance note: our ability to compute 'consumed' and
            // then shift and recompute is critical. If there is a
            // latency of, say, 4 cycles on getting 'consumed', then
            // the inner loop might have a total latency of about 6 cycles.
            // Yet we process between 6 to 12 inputs bytes, thus we get
            // a speed limit between 1 cycle/byte and 0.5 cycle/byte
            // for this section of the code. Hence, there is a limit
            // to how much we can further increase this latency before
            // it seriously harms performance.
            size_t consumed = convert_masked_utf8_to_utf16<endian>(in + pos,
                            utf8_end_of_code_point_mask, utf16_output);
            pos += consumed;
            utf8_end_of_code_point_mask >>= consumed;
          }
          // At this point there may remain between 0 and 12 bytes in the
          // 64-byte block. These bytes will be processed again. So we have an
          // 80% efficiency (in the worst case). In practice we expect an
          // 85% to 90% efficiency.
        }
      }
      if(errors()) {
        // rewind_and_convert_with_errors will seek a potential error from in+pos onward,
        // with the ability to go back up to pos bytes, and read size-pos bytes forward.
        result res = scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(pos, in + pos, size - pos, utf16_output);
        res.count += pos;
        return res;
      }
      if(pos < size) {
        // rewind_and_convert_with_errors will seek a potential error from in+pos onward,
        // with the ability to go back up to pos bytes, and read size-pos bytes forward.
        result res = scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(pos, in + pos, size - pos, utf16_output);
        if (res.error) {    // In case of error, we want the error position
          res.count += pos;
          return res;
        } else {    // In case of success, we want the number of word written
          utf16_output += res.count;
        }
      }
      return result(error_code::SUCCESS, utf16_output - start);
    }

    simdutf_really_inline bool errors() const {
      return this->error.any_bits_set_anywhere();
    }

  }; // struct utf8_checker
} // utf8_to_utf16 namespace
} // unnamed namespace
} // namespace haswell
} // namespace simdutf
/* end file src/generic/utf8_to_utf16/utf8_to_utf16.h */
// transcoding from UTF-8 to UTF-32
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_to_utf32/valid_utf8_to_utf32.h
/* begin file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */

namespace simdutf {
namespace haswell {
namespace {
namespace utf8_to_utf32 {

using namespace simd;


simdutf_warn_unused size_t convert_valid(const char* input, size_t size,
    char32_t* utf32_output) noexcept {
  size_t pos = 0;
  char32_t* start{utf32_output};
  const size_t safety_margin = 16; // to avoid overruns!
  while(pos + 64 + safety_margin <= size) {
    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
    if(in.is_ascii()) {
      in.store_ascii_as_utf32(utf32_output);
      utf32_output += 64;
      pos += 64;
    } else {
    // -65 is 0b10111111 in two-complement's, so largest possible continuation byte
    uint64_t utf8_continuation_mask = in.lt(-65 + 1);
    uint64_t utf8_leading_mask = ~utf8_continuation_mask;
    uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
    size_t max_starting_point = (pos + 64) - 12;
    while(pos < max_starting_point) {
      size_t consumed = convert_masked_utf8_to_utf32(input + pos,
                          utf8_end_of_code_point_mask, utf32_output);
      pos += consumed;
      utf8_end_of_code_point_mask >>= consumed;
      }
    }
  }
  utf32_output += scalar::utf8_to_utf32::convert_valid(input + pos, size - pos, utf32_output);
  return utf32_output - start;
}


} // namespace utf8_to_utf32
} // unnamed namespace
} // namespace haswell
} // namespace simdutf
/* end file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_to_utf32/utf8_to_utf32.h
/* begin file src/generic/utf8_to_utf32/utf8_to_utf32.h */


namespace simdutf {
namespace haswell {
namespace {
namespace utf8_to_utf32 {
using namespace simd;


  simdutf_really_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);
  }
  simdutf_really_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 = simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
    simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
    return must23_80 ^ sc;
  }


  struct validating_transcoder {
    // If this is nonzero, there has been a UTF-8 error.
    simd8<uint8_t> error;

    validating_transcoder() : error(uint8_t(0)) {}
    //
    // Check whether the current bytes are valid UTF-8.
    //
    simdutf_really_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);
    }



    simdutf_really_inline size_t convert(const char* in, size_t size, char32_t* utf32_output) {
      size_t pos = 0;
      char32_t* start{utf32_output};
      // In the worst case, we have the haswell kernel which can cause an overflow of
      // 8 bytes when calling convert_masked_utf8_to_utf32. If you skip the last 16 bytes,
      // and if the data is valid, then it is entirely safe because 16 UTF-8 bytes generate
      // much more than 8 bytes. However, you cannot generally assume that you have valid
      // UTF-8 input, so we are going to go back from the end counting 4 leading bytes,
      // to give us a good margin.
      size_t leading_byte = 0;
      size_t margin = size;
      for(; margin > 0 && leading_byte < 4; margin--) {
        leading_byte += (int8_t(in[margin-1]) > -65);
      }
      // If the input is long enough, then we have that margin-1 is the fourth last leading byte.
      const size_t safety_margin = size - margin + 1; // to avoid overruns!
      while(pos + 64 + safety_margin <= size) {
        simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
        if(input.is_ascii()) {
          input.store_ascii_as_utf32(utf32_output);
          utf32_output += 64;
          pos += 64;
        } 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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
              "We support either two or four chunks per 64-byte block.");
          auto zero = simd8<uint8_t>{uint8_t(0)};
          if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
            this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
            this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
          }
          uint64_t utf8_continuation_mask = input.lt(-65 + 1);
          uint64_t utf8_leading_mask = ~utf8_continuation_mask;
          uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
          // We process in blocks of up to 12 bytes except possibly
          // for fast paths which may process up to 16 bytes. For the
          // slow path to work, we should have at least 12 input bytes left.
          size_t max_starting_point = (pos + 64) - 12;
          // Next loop is going to run at least five times.
          while(pos < max_starting_point) {
            // Performance note: our ability to compute 'consumed' and
            // then shift and recompute is critical. If there is a
            // latency of, say, 4 cycles on getting 'consumed', then
            // the inner loop might have a total latency of about 6 cycles.
            // Yet we process between 6 to 12 inputs bytes, thus we get
            // a speed limit between 1 cycle/byte and 0.5 cycle/byte
            // for this section of the code. Hence, there is a limit
            // to how much we can further increase this latency before
            // it seriously harms performance.
            size_t consumed = convert_masked_utf8_to_utf32(in + pos,
                            utf8_end_of_code_point_mask, utf32_output);
            pos += consumed;
            utf8_end_of_code_point_mask >>= consumed;
          }
          // At this point there may remain between 0 and 12 bytes in the
          // 64-byte block. These bytes will be processed again. So we have an
          // 80% efficiency (in the worst case). In practice we expect an
          // 85% to 90% efficiency.
        }
      }
      if(errors()) { return 0; }
      if(pos < size) {
        size_t howmany  = scalar::utf8_to_utf32::convert(in + pos, size - pos, utf32_output);
        if(howmany == 0) { return 0; }
        utf32_output += howmany;
      }
      return utf32_output - start;
    }

    simdutf_really_inline result convert_with_errors(const char* in, size_t size, char32_t* utf32_output) {
      size_t pos = 0;
      char32_t* start{utf32_output};
      // In the worst case, we have the haswell kernel which can cause an overflow of
      // 8 bytes when calling convert_masked_utf8_to_utf32. If you skip the last 16 bytes,
      // and if the data is valid, then it is entirely safe because 16 UTF-8 bytes generate
      // much more than 8 bytes. However, you cannot generally assume that you have valid
      // UTF-8 input, so we are going to go back from the end counting 4 leading bytes,
      // to give us a good margin.
      size_t leading_byte = 0;
      size_t margin = size;
      for(; margin > 0 && leading_byte < 4; margin--) {
        leading_byte += (int8_t(in[margin-1]) > -65);
      }
      // If the input is long enough, then we have that margin-1 is the fourth last leading byte.
      const size_t safety_margin = size - margin + 1; // to avoid overruns!
      while(pos + 64 + safety_margin <= size) {
        simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
        if(input.is_ascii()) {
          input.store_ascii_as_utf32(utf32_output);
          utf32_output += 64;
          pos += 64;
        } 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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
              "We support either two or four chunks per 64-byte block.");
          auto zero = simd8<uint8_t>{uint8_t(0)};
          if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
            this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
            this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
          }
          if (errors()) {
            result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(pos, in + pos, size - pos, utf32_output);
            res.count += pos;
            return res;
          }
          uint64_t utf8_continuation_mask = input.lt(-65 + 1);
          uint64_t utf8_leading_mask = ~utf8_continuation_mask;
          uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
          // We process in blocks of up to 12 bytes except possibly
          // for fast paths which may process up to 16 bytes. For the
          // slow path to work, we should have at least 12 input bytes left.
          size_t max_starting_point = (pos + 64) - 12;
          // Next loop is going to run at least five times.
          while(pos < max_starting_point) {
            // Performance note: our ability to compute 'consumed' and
            // then shift and recompute is critical. If there is a
            // latency of, say, 4 cycles on getting 'consumed', then
            // the inner loop might have a total latency of about 6 cycles.
            // Yet we process between 6 to 12 inputs bytes, thus we get
            // a speed limit between 1 cycle/byte and 0.5 cycle/byte
            // for this section of the code. Hence, there is a limit
            // to how much we can further increase this latency before
            // it seriously harms performance.
            size_t consumed = convert_masked_utf8_to_utf32(in + pos,
                            utf8_end_of_code_point_mask, utf32_output);
            pos += consumed;
            utf8_end_of_code_point_mask >>= consumed;
          }
          // At this point there may remain between 0 and 12 bytes in the
          // 64-byte block. These bytes will be processed again. So we have an
          // 80% efficiency (in the worst case). In practice we expect an
          // 85% to 90% efficiency.
        }
      }
      if(errors()) {
        result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(pos, in + pos, size - pos, utf32_output);
        res.count += pos;
        return res;
      }
      if(pos < size) {
        result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(pos, in + pos, size - pos, utf32_output);
        if (res.error) {    // In case of error, we want the error position
          res.count += pos;
          return res;
        } else {    // In case of success, we want the number of word written
          utf32_output += res.count;
        }
      }
      return result(error_code::SUCCESS, utf32_output - start);
    }

    simdutf_really_inline bool errors() const {
      return this->error.any_bits_set_anywhere();
    }

  }; // struct utf8_checker
} // utf8_to_utf32 namespace
} // unnamed namespace
} // namespace haswell
} // namespace simdutf
/* end file src/generic/utf8_to_utf32/utf8_to_utf32.h */
// other functions
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8.h
/* begin file src/generic/utf8.h */

namespace simdutf {
namespace haswell {
namespace {
namespace utf8 {

using namespace simd;

simdutf_really_inline size_t count_code_points(const char* in, size_t size) {
    size_t pos = 0;
    size_t count = 0;
    for(;pos + 64 <= size; pos += 64) {
      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
      uint64_t utf8_continuation_mask = input.lt(-65 + 1);
      count += 64 - count_ones(utf8_continuation_mask);
    }
    return count + scalar::utf8::count_code_points(in + pos, size - pos);
}


simdutf_really_inline size_t utf16_length_from_utf8(const char* in, size_t size) {
    size_t pos = 0;
    size_t count = 0;
    // This algorithm could no doubt be improved!
    for(;pos + 64 <= size; pos += 64) {
      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
      uint64_t utf8_continuation_mask = input.lt(-65 + 1);
      // We count one word for anything that is not a continuation (so
      // leading bytes).
      count += 64 - count_ones(utf8_continuation_mask);
      int64_t utf8_4byte = input.gteq_unsigned(240);
      count += count_ones(utf8_4byte);
    }
    return count + scalar::utf8::utf16_length_from_utf8(in + pos, size - pos);
}


simdutf_really_inline size_t utf32_length_from_utf8(const char* in, size_t size) {
    return count_code_points(in, size);
}
} // utf8 namespace
} // unnamed namespace
} // namespace haswell
} // namespace simdutf
/* end file src/generic/utf8.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf16.h
/* begin file src/generic/utf16.h */
namespace simdutf {
namespace haswell {
namespace {
namespace utf16 {

template <endianness big_endian>
simdutf_really_inline size_t count_code_points(const char16_t* in, size_t size) {
    size_t pos = 0;
    size_t count = 0;
    for(;pos + 32 <= size; pos += 32) {
      simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
      if (!match_system(big_endian)) input.swap_bytes();
      uint64_t not_pair = input.not_in_range(0xDC00, 0xDFFF);
      count += count_ones(not_pair) / 2;
    }
    return count + scalar::utf16::count_code_points<big_endian>(in + pos, size - pos);
}

template <endianness big_endian>
simdutf_really_inline size_t utf8_length_from_utf16(const char16_t* in, size_t size) {
    size_t pos = 0;
    size_t count = 0;
    // This algorithm could no doubt be improved!
    for(;pos + 32 <= size; pos += 32) {
      simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
      if (!match_system(big_endian)) input.swap_bytes();
      uint64_t ascii_mask = input.lteq(0x7F);
      uint64_t twobyte_mask = input.lteq(0x7FF);
      uint64_t not_pair_mask = input.not_in_range(0xD800, 0xDFFF);

      size_t ascii_count = count_ones(ascii_mask) / 2;
      size_t twobyte_count = count_ones(twobyte_mask & ~ ascii_mask) / 2;
      size_t threebyte_count = count_ones(not_pair_mask & ~ twobyte_mask) / 2;
      size_t fourbyte_count = 32 - count_ones(not_pair_mask) / 2;
      count += 2 * fourbyte_count + 3 * threebyte_count + 2 * twobyte_count + ascii_count;
    }
    return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos, size - pos);
}

template <endianness big_endian>
simdutf_really_inline size_t utf32_length_from_utf16(const char16_t* in, size_t size) {
    return count_code_points<big_endian>(in, size);
}

simdutf_really_inline void change_endianness_utf16(const char16_t* in, size_t size, char16_t* output) {
  size_t pos = 0;

  while (pos + 32 <= size) {
    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
    input.swap_bytes();
    input.store(reinterpret_cast<uint16_t *>(output));
    pos += 32;
    output += 32;
  }

  scalar::utf16::change_endianness_utf16(in + pos, size - pos, output);
}

} // utf16
} // unnamed namespace
} // namespace haswell
} // namespace simdutf
/* end file src/generic/utf16.h */

namespace simdutf {
namespace haswell {

simdutf_warn_unused int implementation::detect_encodings(const char * input, size_t length) const noexcept {
  // If there is a BOM, then we trust it.
  auto bom_encoding = simdutf::BOM::check_bom(input, length);
  if(bom_encoding != encoding_type::unspecified) { return bom_encoding; }
  if (length % 2 == 0) {
    return avx2_detect_encodings<utf8_validation::utf8_checker>(input, length);
  } else {
    if (implementation::validate_utf8(input, length)) {
      return simdutf::encoding_type::UTF8;
    } else {
      return simdutf::encoding_type::unspecified;
    }
  }
}

simdutf_warn_unused bool implementation::validate_utf8(const char *buf, size_t len) const noexcept {
  return haswell::utf8_validation::generic_validate_utf8(buf,len);
}

simdutf_warn_unused result implementation::validate_utf8_with_errors(const char *buf, size_t len) const noexcept {
  return haswell::utf8_validation::generic_validate_utf8_with_errors(buf,len);
}

simdutf_warn_unused bool implementation::validate_ascii(const char *buf, size_t len) const noexcept {
  return haswell::utf8_validation::generic_validate_ascii(buf,len);
}

simdutf_warn_unused result implementation::validate_ascii_with_errors(const char *buf, size_t len) const noexcept {
  return haswell::utf8_validation::generic_validate_ascii_with_errors(buf,len);
}

simdutf_warn_unused bool implementation::validate_utf16le(const char16_t *buf, size_t len) const noexcept {
  const char16_t* tail = avx2_validate_utf16<endianness::LITTLE>(buf, len);
  if (tail) {
    return scalar::utf16::validate<endianness::LITTLE>(tail, len - (tail - buf));
  } else {
    return false;
  }
}

simdutf_warn_unused bool implementation::validate_utf16be(const char16_t *buf, size_t len) const noexcept {
  const char16_t* tail = avx2_validate_utf16<endianness::BIG>(buf, len);
  if (tail) {
    return scalar::utf16::validate<endianness::BIG>(tail, len - (tail - buf));
  } else {
    return false;
  }
}

simdutf_warn_unused result implementation::validate_utf16le_with_errors(const char16_t *buf, size_t len) const noexcept {
  result res = avx2_validate_utf16_with_errors<endianness::LITTLE>(buf, len);
  if (res.count != len) {
    result scalar_res = scalar::utf16::validate_with_errors<endianness::LITTLE>(buf + res.count, len - res.count);
    return result(scalar_res.error, res.count + scalar_res.count);
  } else {
    return res;
  }
}

simdutf_warn_unused result implementation::validate_utf16be_with_errors(const char16_t *buf, size_t len) const noexcept {
  result res = avx2_validate_utf16_with_errors<endianness::BIG>(buf, len);
  if (res.count != len) {
    result scalar_res = scalar::utf16::validate_with_errors<endianness::BIG>(buf + res.count, len - res.count);
    return result(scalar_res.error, res.count + scalar_res.count);
  } else {
    return res;
  }
}

simdutf_warn_unused bool implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
  const char32_t* tail = avx2_validate_utf32le(buf, len);
  if (tail) {
    return scalar::utf32::validate(tail, len - (tail - buf));
  } else {
    return false;
  }
}

simdutf_warn_unused result implementation::validate_utf32_with_errors(const char32_t *buf, size_t len) const noexcept {
  result res = avx2_validate_utf32le_with_errors(buf, len);
  if (res.count != len) {
    result scalar_res = scalar::utf32::validate_with_errors(buf + res.count, len - res.count);
    return result(scalar_res.error, res.count + scalar_res.count);
  } else {
    return res;
  }
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
  utf8_to_utf16::validating_transcoder converter;
  return converter.convert<endianness::LITTLE>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf16be(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
  utf8_to_utf16::validating_transcoder converter;
  return converter.convert<endianness::BIG>(buf, len, utf16_output);
}

simdutf_warn_unused result implementation::convert_utf8_to_utf16le_with_errors(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
  utf8_to_utf16::validating_transcoder converter;
  return converter.convert_with_errors<endianness::LITTLE>(buf, len, utf16_output);
}

simdutf_warn_unused result implementation::convert_utf8_to_utf16be_with_errors(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
  utf8_to_utf16::validating_transcoder converter;
  return converter.convert_with_errors<endianness::BIG>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16le(const char* input, size_t size,
    char16_t* utf16_output) const noexcept {
   return utf8_to_utf16::convert_valid<endianness::LITTLE>(input, size,  utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(const char* input, size_t size,
    char16_t* utf16_output) const noexcept {
   return utf8_to_utf16::convert_valid<endianness::BIG>(input, size,  utf16_output);
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(const char* buf, size_t len, char32_t* utf32_output) const noexcept {
  utf8_to_utf32::validating_transcoder converter;
  return converter.convert(buf, len, utf32_output);
}

simdutf_warn_unused result implementation::convert_utf8_to_utf32_with_errors(const char* buf, size_t len, char32_t* utf32_output) const noexcept {
  utf8_to_utf32::validating_transcoder converter;
  return converter.convert_with_errors(buf, len, utf32_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(const char* input, size_t size,
    char32_t* utf32_output) const noexcept {
  return utf8_to_utf32::convert_valid(input, size,  utf32_output);
}

simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  std::pair<const char16_t*, char*> ret = haswell::avx2_convert_utf16_to_utf8<endianness::LITTLE>(buf, len, utf8_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf8_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf16_to_utf8::convert<endianness::LITTLE>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused size_t implementation::convert_utf16be_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  std::pair<const char16_t*, char*> ret = haswell::avx2_convert_utf16_to_utf8<endianness::BIG>(buf, len, utf8_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf8_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf16_to_utf8::convert<endianness::BIG>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused result implementation::convert_utf16le_to_utf8_with_errors(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char*> ret = haswell::avx2_convert_utf16_to_utf8_with_errors<endianness::LITTLE>(buf, len, utf8_output);
  if (ret.first.error) { return ret.first; }  // Can return directly since scalar fallback already found correct ret.first.count
  if (ret.first.count != len) { // All good so far, but not finished
    result scalar_res = scalar::utf16_to_utf8::convert_with_errors<endianness::LITTLE>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf8_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused result implementation::convert_utf16be_to_utf8_with_errors(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char*> ret = haswell::avx2_convert_utf16_to_utf8_with_errors<endianness::BIG>(buf, len, utf8_output);
  if (ret.first.error) { return ret.first; }  // Can return directly since scalar fallback already found correct ret.first.count
  if (ret.first.count != len) { // All good so far, but not finished
    result scalar_res = scalar::utf16_to_utf8::convert_with_errors<endianness::BIG>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf8_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return convert_utf16le_to_utf8(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return convert_utf16be_to_utf8(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  std::pair<const char32_t*, char*> ret = avx2_convert_utf32_to_utf8(buf, len, utf8_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf8_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf32_to_utf8::convert(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char*> ret = haswell::avx2_convert_utf32_to_utf8_with_errors(buf, len, utf8_output);
  if (ret.first.count != len) {
    result scalar_res = scalar::utf32_to_utf8::convert_with_errors(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf8_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  std::pair<const char16_t*, char32_t*> ret = haswell::avx2_convert_utf16_to_utf32<endianness::LITTLE>(buf, len, utf32_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf32_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf16_to_utf32::convert<endianness::LITTLE>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused size_t implementation::convert_utf16be_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  std::pair<const char16_t*, char32_t*> ret = haswell::avx2_convert_utf16_to_utf32<endianness::BIG>(buf, len, utf32_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf32_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf16_to_utf32::convert<endianness::BIG>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused result implementation::convert_utf16le_to_utf32_with_errors(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char32_t*> ret = haswell::avx2_convert_utf16_to_utf32_with_errors<endianness::LITTLE>(buf, len, utf32_output);
  if (ret.first.error) { return ret.first; }  // Can return directly since scalar fallback already found correct ret.first.count
  if (ret.first.count != len) { // All good so far, but not finished
    result scalar_res = scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf32_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char32_t*> ret = haswell::avx2_convert_utf16_to_utf32_with_errors<endianness::BIG>(buf, len, utf32_output);
  if (ret.first.error) { return ret.first; }  // Can return directly since scalar fallback already found correct ret.first.count
  if (ret.first.count != len) { // All good so far, but not finished
    result scalar_res = scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf32_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  return convert_utf32_to_utf8(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  std::pair<const char32_t*, char16_t*> ret = avx2_convert_utf32_to_utf16<endianness::LITTLE>(buf, len, utf16_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf16_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf32_to_utf16::convert<endianness::LITTLE>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf16be(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  std::pair<const char32_t*, char16_t*> ret = avx2_convert_utf32_to_utf16<endianness::BIG>(buf, len, utf16_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf16_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf32_to_utf16::convert<endianness::BIG>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused result implementation::convert_utf32_to_utf16le_with_errors(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char16_t*> ret = haswell::avx2_convert_utf32_to_utf16_with_errors<endianness::LITTLE>(buf, len, utf16_output);
  if (ret.first.count != len) {
    result scalar_res = scalar::utf32_to_utf16::convert_with_errors<endianness::LITTLE>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf16_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused result implementation::convert_utf32_to_utf16be_with_errors(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char16_t*> ret = haswell::avx2_convert_utf32_to_utf16_with_errors<endianness::BIG>(buf, len, utf16_output);
  if (ret.first.count != len) {
    result scalar_res = scalar::utf32_to_utf16::convert_with_errors<endianness::BIG>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf16_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16le(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return convert_utf32_to_utf16le(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16be(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return convert_utf32_to_utf16be(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return convert_utf16le_to_utf32(buf, len, utf32_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return convert_utf16be_to_utf32(buf, len, utf32_output);
}

void implementation::change_endianness_utf16(const char16_t * input, size_t length, char16_t * output) const noexcept {
  utf16::change_endianness_utf16(input, length, output);
}

simdutf_warn_unused size_t implementation::count_utf16le(const char16_t * input, size_t length) const noexcept {
  return utf16::count_code_points<endianness::LITTLE>(input, length);
}

simdutf_warn_unused size_t implementation::count_utf16be(const char16_t * input, size_t length) const noexcept {
  return utf16::count_code_points<endianness::BIG>(input, length);
}

simdutf_warn_unused size_t implementation::count_utf8(const char * input, size_t length) const noexcept {
  return utf8::count_code_points(input, length);
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(const char16_t * input, size_t length) const noexcept {
  return utf16::utf8_length_from_utf16<endianness::LITTLE>(input, length);
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(const char16_t * input, size_t length) const noexcept {
  return utf16::utf8_length_from_utf16<endianness::BIG>(input, length);
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(const char16_t * input, size_t length) const noexcept {
  return utf16::utf32_length_from_utf16<endianness::LITTLE>(input, length);
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(const char16_t * input, size_t length) const noexcept {
  return utf16::utf32_length_from_utf16<endianness::BIG>(input, length);
}

simdutf_warn_unused size_t implementation::utf16_length_from_utf8(const char * input, size_t length) const noexcept {
  return utf8::utf16_length_from_utf8(input, length);
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf32(const char32_t * input, size_t length) const noexcept {
  const __m256i v_00000000 = _mm256_setzero_si256();
  const __m256i v_ffffff80 = _mm256_set1_epi32((uint32_t)0xffffff80);
  const __m256i v_fffff800 = _mm256_set1_epi32((uint32_t)0xfffff800);
  const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000);
  size_t pos = 0;
  size_t count = 0;
  for(;pos + 8 <= length; pos += 8) {
    __m256i in = _mm256_loadu_si256((__m256i*)(input + pos));
    const __m256i ascii_bytes_bytemask = _mm256_cmpeq_epi32(_mm256_and_si256(in, v_ffffff80), v_00000000);
    const __m256i one_two_bytes_bytemask = _mm256_cmpeq_epi32(_mm256_and_si256(in, v_fffff800), v_00000000);
    const __m256i two_bytes_bytemask = _mm256_xor_si256(one_two_bytes_bytemask, ascii_bytes_bytemask);
    const __m256i one_two_three_bytes_bytemask = _mm256_cmpeq_epi32(_mm256_and_si256(in, v_ffff0000), v_00000000);
    const __m256i three_bytes_bytemask = _mm256_xor_si256(one_two_three_bytes_bytemask, one_two_bytes_bytemask);
    const uint32_t ascii_bytes_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(ascii_bytes_bytemask));
    const uint32_t two_bytes_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(two_bytes_bytemask));
    const uint32_t three_bytes_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(three_bytes_bytemask));

    size_t ascii_count = count_ones(ascii_bytes_bitmask) / 4;
    size_t two_bytes_count = count_ones(two_bytes_bitmask) / 4;
    size_t three_bytes_count = count_ones(three_bytes_bitmask) / 4;
    count += 32 - 3*ascii_count - 2*two_bytes_count - three_bytes_count;
  }
  return count + scalar::utf32::utf8_length_from_utf32(input + pos, length - pos);
}

simdutf_warn_unused size_t implementation::utf16_length_from_utf32(const char32_t * input, size_t length) const noexcept {
  const __m256i v_00000000 = _mm256_setzero_si256();
  const __m256i v_ffff0000 = _mm256_set1_epi32((uint32_t)0xffff0000);
  size_t pos = 0;
  size_t count = 0;
  for(;pos + 8 <= length; pos += 8) {
    __m256i in = _mm256_loadu_si256((__m256i*)(input + pos));
    const __m256i surrogate_bytemask = _mm256_cmpeq_epi32(_mm256_and_si256(in, v_ffff0000), v_00000000);
    const uint32_t surrogate_bitmask = static_cast<uint32_t>(_mm256_movemask_epi8(surrogate_bytemask));
    size_t surrogate_count = (32-count_ones(surrogate_bitmask))/4;
    count += 8 + surrogate_count;
  }
  return count + scalar::utf32::utf16_length_from_utf32(input + pos, length - pos);
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf8(const char * input, size_t length) const noexcept {
  return scalar::utf8::count_code_points(input, length);
}

} // namespace haswell
} // namespace simdutf

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/haswell/end.h
/* begin file src/simdutf/haswell/end.h */
#if SIMDUTF_CAN_ALWAYS_RUN_HASWELL
// nothing needed.
#else
SIMDUTF_UNTARGET_REGION
#endif


#if SIMDUTF_GCC11ORMORE // workaround for https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
SIMDUTF_POP_DISABLE_WARNINGS
#endif // end of workaround
/* end file src/simdutf/haswell/end.h */
/* end file src/haswell/implementation.cpp */
#endif
#if SIMDUTF_IMPLEMENTATION_PPC64
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=ppc64/implementation.cpp
/* begin file src/ppc64/implementation.cpp */





// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/ppc64/begin.h
/* begin file src/simdutf/ppc64/begin.h */
// redefining SIMDUTF_IMPLEMENTATION to "ppc64"
// #define SIMDUTF_IMPLEMENTATION ppc64
/* end file src/simdutf/ppc64/begin.h */
namespace simdutf {
namespace ppc64 {
namespace {
#ifndef SIMDUTF_PPC64_H
#error "ppc64.h must be included"
#endif
using namespace simd;


simdutf_really_inline bool is_ascii(const simd8x64<uint8_t>& input) {
  // careful: 0x80 is not ascii.
  return input.reduce_or().saturating_sub(0b01111111u).bits_not_set_anywhere();
}

simdutf_unused simdutf_really_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(0b11000000u-1); // Only 11______ will be > 0
  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0b11100000u-1); // Only 111_____ will be > 0
  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0b11110000u-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);
}

simdutf_really_inline simd8<bool> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0b11100000u-1); // Only 111_____ will be > 0
  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0b11110000u-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_third_byte | is_fourth_byte) > int8_t(0);
}

} // unnamed namespace
} // namespace ppc64
} // namespace simdutf

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/buf_block_reader.h
/* begin file src/generic/buf_block_reader.h */
namespace simdutf {
namespace ppc64 {
namespace {

// Walks through a buffer in block-sized increments, loading the last part with spaces
template<size_t STEP_SIZE>
struct buf_block_reader {
public:
  simdutf_really_inline buf_block_reader(const uint8_t *_buf, size_t _len);
  simdutf_really_inline size_t block_index();
  simdutf_really_inline bool has_full_block() const;
  simdutf_really_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.
   */
  simdutf_really_inline size_t get_remainder(uint8_t *dst) const;
  simdutf_really_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
simdutf_unused static char * format_input_text_64(const uint8_t *text) {
  static char *buf = reinterpret_cast<char*>(malloc(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
simdutf_unused static char * format_input_text(const simd8x64<uint8_t>& in) {
  static char *buf = reinterpret_cast<char*>(malloc(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;
}

simdutf_unused static char * format_mask(uint64_t mask) {
  static char *buf = reinterpret_cast<char*>(malloc(64 + 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>
simdutf_really_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>
simdutf_really_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }

template<size_t STEP_SIZE>
simdutf_really_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
  return idx < lenminusstep;
}

template<size_t STEP_SIZE>
simdutf_really_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
  return &buf[idx];
}

template<size_t STEP_SIZE>
simdutf_really_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>
simdutf_really_inline void buf_block_reader<STEP_SIZE>::advance() {
  idx += STEP_SIZE;
}

} // unnamed namespace
} // namespace ppc64
} // namespace simdutf
/* end file src/generic/buf_block_reader.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_validation/utf8_lookup4_algorithm.h
/* begin file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
namespace simdutf {
namespace ppc64 {
namespace {
namespace utf8_validation {

using namespace simd;

  simdutf_really_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);
  }
  simdutf_really_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 = simd8<uint8_t>(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.
  //
  simdutf_really_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______
    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, 0b11110000u-1, 0b11100000u-1, 0b11000000u-1
    };
    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.
    //
    simdutf_really_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.
    simdutf_really_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;
    }

    simdutf_really_inline void check_next_input(const simd8x64<uint8_t>& input) {
      if(simdutf_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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
            "We support either two or four chunks per 64-byte block.");
        if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
          this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
        } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
          this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
        }
        this->prev_incomplete = is_incomplete(input.chunks[simd8x64<uint8_t>::NUM_CHUNKS-1]);
        this->prev_input_block = input.chunks[simd8x64<uint8_t>::NUM_CHUNKS-1];

      }
    }

    // do not forget to call check_eof!
    simdutf_really_inline bool errors() const {
      return this->error.any_bits_set_anywhere();
    }

  }; // struct utf8_checker
} // namespace utf8_validation

using utf8_validation::utf8_checker;

} // unnamed namespace
} // namespace ppc64
} // namespace simdutf
/* end file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_validation/utf8_validator.h
/* begin file src/generic/utf8_validation/utf8_validator.h */
namespace simdutf {
namespace ppc64 {
namespace {
namespace utf8_validation {

/**
 * 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();
}

bool generic_validate_utf8(const char * input, size_t length) {
  return generic_validate_utf8<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
}

/**
 * Validates that the string is actual UTF-8 and stops on errors.
 */
template<class checker>
result generic_validate_utf8_with_errors(const uint8_t * input, size_t length) {
    checker c{};
    buf_block_reader<64> reader(input, length);
    size_t count{0};
    while (reader.has_full_block()) {
      simd::simd8x64<uint8_t> in(reader.full_block());
      c.check_next_input(in);
      if(c.errors()) {
        if (count != 0) { count--; } // Sometimes the error is only detected in the next chunk
        result res = scalar::utf8::rewind_and_validate_with_errors(reinterpret_cast<const char*>(input + count), length - count);
        res.count += count;
        return res;
      }
      reader.advance();
      count += 64;
    }
    uint8_t block[64]{};
    reader.get_remainder(block);
    simd::simd8x64<uint8_t> in(block);
    c.check_next_input(in);
    reader.advance();
    c.check_eof();
    if (c.errors()) {
      result res = scalar::utf8::rewind_and_validate_with_errors(reinterpret_cast<const char*>(input) + count, length - count);
      res.count += count;
      return res;
    } else {
      return result(error_code::SUCCESS, length);
    }
}

result generic_validate_utf8_with_errors(const char * input, size_t length) {
  return generic_validate_utf8_with_errors<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
}

template<class checker>
bool generic_validate_ascii(const uint8_t * input, size_t length) {
    buf_block_reader<64> reader(input, length);
    uint8_t blocks[64]{};
    simd::simd8x64<uint8_t> running_or(blocks);
    while (reader.has_full_block()) {
      simd::simd8x64<uint8_t> in(reader.full_block());
      running_or |= in;
      reader.advance();
    }
    uint8_t block[64]{};
    reader.get_remainder(block);
    simd::simd8x64<uint8_t> in(block);
    running_or |= in;
    return running_or.is_ascii();
}

bool generic_validate_ascii(const char * input, size_t length) {
  return generic_validate_ascii<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
}

template<class checker>
result generic_validate_ascii_with_errors(const uint8_t * input, size_t length) {
  buf_block_reader<64> reader(input, length);
  size_t count{0};
  while (reader.has_full_block()) {
    simd::simd8x64<uint8_t> in(reader.full_block());
    if (!in.is_ascii()) {
      result res = scalar::ascii::validate_with_errors(reinterpret_cast<const char*>(input + count), length - count);
      return result(res.error, count + res.count);
    }
    reader.advance();

    count += 64;
  }
  uint8_t block[64]{};
  reader.get_remainder(block);
  simd::simd8x64<uint8_t> in(block);
  if (!in.is_ascii()) {
    result res = scalar::ascii::validate_with_errors(reinterpret_cast<const char*>(input + count), length - count);
    return result(res.error, count + res.count);
  } else {
    return result(error_code::SUCCESS, length);
  }
}

result generic_validate_ascii_with_errors(const char * input, size_t length) {
  return generic_validate_ascii_with_errors<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
}

} // namespace utf8_validation
} // unnamed namespace
} // namespace ppc64
} // namespace simdutf
/* end file src/generic/utf8_validation/utf8_validator.h */
// transcoding from UTF-8 to UTF-16
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_to_utf16/valid_utf8_to_utf16.h
/* begin file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */


namespace simdutf {
namespace ppc64 {
namespace {
namespace utf8_to_utf16 {

using namespace simd;

template <endianness endian>
simdutf_warn_unused size_t convert_valid(const char* input, size_t size,
    char16_t* utf16_output) noexcept {
  // The implementation is not specific to haswell and should be moved to the generic directory.
  size_t pos = 0;
  char16_t* start{utf16_output};
  const size_t safety_margin = 16; // to avoid overruns!
  while(pos + 64 + safety_margin <= size) {
    // this loop could be unrolled further. For example, we could process the mask
    // far more than 64 bytes.
    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
    if(in.is_ascii()) {
      in.store_ascii_as_utf16<endian>(utf16_output);
      utf16_output += 64;
      pos += 64;
    } else {
      // Slow path. We hope that the compiler will recognize that this is a slow path.
      // Anything that is not a continuation mask is a 'leading byte', that is, the
      // start of a new code point.
      uint64_t utf8_continuation_mask = in.lt(-65 + 1);
      // -65 is 0b10111111 in two-complement's, so largest possible continuation byte
      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
      // The *start* of code points is not so useful, rather, we want the *end* of code points.
      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
      // We process in blocks of up to 12 bytes except possibly
      // for fast paths which may process up to 16 bytes. For the
      // slow path to work, we should have at least 12 input bytes left.
      size_t max_starting_point = (pos + 64) - 12;
      // Next loop is going to run at least five times when using solely
      // the slow/regular path, and at least four times if there are fast paths.
      while(pos < max_starting_point) {
        // Performance note: our ability to compute 'consumed' and
        // then shift and recompute is critical. If there is a
        // latency of, say, 4 cycles on getting 'consumed', then
        // the inner loop might have a total latency of about 6 cycles.
        // Yet we process between 6 to 12 inputs bytes, thus we get
        // a speed limit between 1 cycle/byte and 0.5 cycle/byte
        // for this section of the code. Hence, there is a limit
        // to how much we can further increase this latency before
        // it seriously harms performance.
        //
        // Thus we may allow convert_masked_utf8_to_utf16 to process
        // more bytes at a time under a fast-path mode where 16 bytes
        // are consumed at once (e.g., when encountering ASCII).
        size_t consumed = convert_masked_utf8_to_utf16<endian>(input + pos,
                            utf8_end_of_code_point_mask, utf16_output);
        pos += consumed;
        utf8_end_of_code_point_mask >>= consumed;
      }
      // At this point there may remain between 0 and 12 bytes in the
      // 64-byte block. These bytes will be processed again. So we have an
      // 80% efficiency (in the worst case). In practice we expect an
      // 85% to 90% efficiency.
    }
  }
  utf16_output += scalar::utf8_to_utf16::convert_valid<endian>(input + pos, size - pos, utf16_output);
  return utf16_output - start;
}

} // namespace utf8_to_utf16
} // unnamed namespace
} // namespace ppc64
} // namespace simdutf
/* end file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_to_utf16/utf8_to_utf16.h
/* begin file src/generic/utf8_to_utf16/utf8_to_utf16.h */


namespace simdutf {
namespace ppc64 {
namespace {
namespace utf8_to_utf16 {
using namespace simd;


  simdutf_really_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);
  }
  simdutf_really_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 = simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
    simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
    return must23_80 ^ sc;
  }


  struct validating_transcoder {
    // If this is nonzero, there has been a UTF-8 error.
    simd8<uint8_t> error;

    validating_transcoder() : error(uint8_t(0)) {}
    //
    // Check whether the current bytes are valid UTF-8.
    //
    simdutf_really_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);
    }


    template <endianness endian>
    simdutf_really_inline size_t convert(const char* in, size_t size, char16_t* utf16_output) {
      size_t pos = 0;
      char16_t* start{utf16_output};
      // In the worst case, we have the haswell kernel which can cause an overflow of
      // 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the last 16 bytes,
      // and if the data is valid, then it is entirely safe because 16 UTF-8 bytes generate
      // much more than 8 bytes. However, you cannot generally assume that you have valid
      // UTF-8 input, so we are going to go back from the end counting 8 leading bytes,
      // to give us a good margin.
      size_t leading_byte = 0;
      size_t margin = size;
      for(; margin > 0 && leading_byte < 8; margin--) {
        leading_byte += (int8_t(in[margin-1]) > -65);
      }
      // If the input is long enough, then we have that margin-1 is the eight last leading byte.
      const size_t safety_margin = size - margin + 1; // to avoid overruns!
      while(pos + 64 + safety_margin <= size) {
        simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
        if(input.is_ascii()) {
          input.store_ascii_as_utf16<endian>(utf16_output);
          utf16_output += 64;
          pos += 64;
        } 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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
              "We support either two or four chunks per 64-byte block.");
          auto zero = simd8<uint8_t>{uint8_t(0)};
          if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
            this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
            this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
          }
          uint64_t utf8_continuation_mask = input.lt(-65 + 1);
          uint64_t utf8_leading_mask = ~utf8_continuation_mask;
          uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
          // We process in blocks of up to 12 bytes except possibly
          // for fast paths which may process up to 16 bytes. For the
          // slow path to work, we should have at least 12 input bytes left.
          size_t max_starting_point = (pos + 64) - 12;
          // Next loop is going to run at least five times.
          while(pos < max_starting_point) {
            // Performance note: our ability to compute 'consumed' and
            // then shift and recompute is critical. If there is a
            // latency of, say, 4 cycles on getting 'consumed', then
            // the inner loop might have a total latency of about 6 cycles.
            // Yet we process between 6 to 12 inputs bytes, thus we get
            // a speed limit between 1 cycle/byte and 0.5 cycle/byte
            // for this section of the code. Hence, there is a limit
            // to how much we can further increase this latency before
            // it seriously harms performance.
            size_t consumed = convert_masked_utf8_to_utf16<endian>(in + pos,
                            utf8_end_of_code_point_mask, utf16_output);
            pos += consumed;
            utf8_end_of_code_point_mask >>= consumed;
          }
          // At this point there may remain between 0 and 12 bytes in the
          // 64-byte block. These bytes will be processed again. So we have an
          // 80% efficiency (in the worst case). In practice we expect an
          // 85% to 90% efficiency.
        }
      }
      if(errors()) { return 0; }
      if(pos < size) {
        size_t howmany  = scalar::utf8_to_utf16::convert<endian>(in + pos, size - pos, utf16_output);
        if(howmany == 0) { return 0; }
        utf16_output += howmany;
      }
      return utf16_output - start;
    }

    template <endianness endian>
    simdutf_really_inline result convert_with_errors(const char* in, size_t size, char16_t* utf16_output) {
      size_t pos = 0;
      char16_t* start{utf16_output};
      // In the worst case, we have the haswell kernel which can cause an overflow of
      // 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the last 16 bytes,
      // and if the data is valid, then it is entirely safe because 16 UTF-8 bytes generate
      // much more than 8 bytes. However, you cannot generally assume that you have valid
      // UTF-8 input, so we are going to go back from the end counting 8 leading bytes,
      // to give us a good margin.
      size_t leading_byte = 0;
      size_t margin = size;
      for(; margin > 0 && leading_byte < 8; margin--) {
        leading_byte += (int8_t(in[margin-1]) > -65);
      }
      // If the input is long enough, then we have that margin-1 is the eight last leading byte.
      const size_t safety_margin = size - margin + 1; // to avoid overruns!
      while(pos + 64 + safety_margin <= size) {
        simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
        if(input.is_ascii()) {
          input.store_ascii_as_utf16<endian>(utf16_output);
          utf16_output += 64;
          pos += 64;
        } 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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
              "We support either two or four chunks per 64-byte block.");
          auto zero = simd8<uint8_t>{uint8_t(0)};
          if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
            this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
            this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
          }
          if (errors()) {
            // rewind_and_convert_with_errors will seek a potential error from in+pos onward,
            // with the ability to go back up to pos bytes, and read size-pos bytes forward.
            result res = scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(pos, in + pos, size - pos, utf16_output);
            res.count += pos;
            return res;
          }
          uint64_t utf8_continuation_mask = input.lt(-65 + 1);
          uint64_t utf8_leading_mask = ~utf8_continuation_mask;
          uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
          // We process in blocks of up to 12 bytes except possibly
          // for fast paths which may process up to 16 bytes. For the
          // slow path to work, we should have at least 12 input bytes left.
          size_t max_starting_point = (pos + 64) - 12;
          // Next loop is going to run at least five times.
          while(pos < max_starting_point) {
            // Performance note: our ability to compute 'consumed' and
            // then shift and recompute is critical. If there is a
            // latency of, say, 4 cycles on getting 'consumed', then
            // the inner loop might have a total latency of about 6 cycles.
            // Yet we process between 6 to 12 inputs bytes, thus we get
            // a speed limit between 1 cycle/byte and 0.5 cycle/byte
            // for this section of the code. Hence, there is a limit
            // to how much we can further increase this latency before
            // it seriously harms performance.
            size_t consumed = convert_masked_utf8_to_utf16<endian>(in + pos,
                            utf8_end_of_code_point_mask, utf16_output);
            pos += consumed;
            utf8_end_of_code_point_mask >>= consumed;
          }
          // At this point there may remain between 0 and 12 bytes in the
          // 64-byte block. These bytes will be processed again. So we have an
          // 80% efficiency (in the worst case). In practice we expect an
          // 85% to 90% efficiency.
        }
      }
      if(errors()) {
        // rewind_and_convert_with_errors will seek a potential error from in+pos onward,
        // with the ability to go back up to pos bytes, and read size-pos bytes forward.
        result res = scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(pos, in + pos, size - pos, utf16_output);
        res.count += pos;
        return res;
      }
      if(pos < size) {
        // rewind_and_convert_with_errors will seek a potential error from in+pos onward,
        // with the ability to go back up to pos bytes, and read size-pos bytes forward.
        result res = scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(pos, in + pos, size - pos, utf16_output);
        if (res.error) {    // In case of error, we want the error position
          res.count += pos;
          return res;
        } else {    // In case of success, we want the number of word written
          utf16_output += res.count;
        }
      }
      return result(error_code::SUCCESS, utf16_output - start);
    }

    simdutf_really_inline bool errors() const {
      return this->error.any_bits_set_anywhere();
    }

  }; // struct utf8_checker
} // utf8_to_utf16 namespace
} // unnamed namespace
} // namespace ppc64
} // namespace simdutf
/* end file src/generic/utf8_to_utf16/utf8_to_utf16.h */
// transcoding from UTF-8 to UTF-32
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_to_utf32/valid_utf8_to_utf32.h
/* begin file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */

namespace simdutf {
namespace ppc64 {
namespace {
namespace utf8_to_utf32 {

using namespace simd;


simdutf_warn_unused size_t convert_valid(const char* input, size_t size,
    char32_t* utf32_output) noexcept {
  size_t pos = 0;
  char32_t* start{utf32_output};
  const size_t safety_margin = 16; // to avoid overruns!
  while(pos + 64 + safety_margin <= size) {
    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
    if(in.is_ascii()) {
      in.store_ascii_as_utf32(utf32_output);
      utf32_output += 64;
      pos += 64;
    } else {
    // -65 is 0b10111111 in two-complement's, so largest possible continuation byte
    uint64_t utf8_continuation_mask = in.lt(-65 + 1);
    uint64_t utf8_leading_mask = ~utf8_continuation_mask;
    uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
    size_t max_starting_point = (pos + 64) - 12;
    while(pos < max_starting_point) {
      size_t consumed = convert_masked_utf8_to_utf32(input + pos,
                          utf8_end_of_code_point_mask, utf32_output);
      pos += consumed;
      utf8_end_of_code_point_mask >>= consumed;
      }
    }
  }
  utf32_output += scalar::utf8_to_utf32::convert_valid(input + pos, size - pos, utf32_output);
  return utf32_output - start;
}


} // namespace utf8_to_utf32
} // unnamed namespace
} // namespace ppc64
} // namespace simdutf
/* end file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_to_utf32/utf8_to_utf32.h
/* begin file src/generic/utf8_to_utf32/utf8_to_utf32.h */


namespace simdutf {
namespace ppc64 {
namespace {
namespace utf8_to_utf32 {
using namespace simd;


  simdutf_really_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);
  }
  simdutf_really_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 = simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
    simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
    return must23_80 ^ sc;
  }


  struct validating_transcoder {
    // If this is nonzero, there has been a UTF-8 error.
    simd8<uint8_t> error;

    validating_transcoder() : error(uint8_t(0)) {}
    //
    // Check whether the current bytes are valid UTF-8.
    //
    simdutf_really_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);
    }



    simdutf_really_inline size_t convert(const char* in, size_t size, char32_t* utf32_output) {
      size_t pos = 0;
      char32_t* start{utf32_output};
      // In the worst case, we have the haswell kernel which can cause an overflow of
      // 8 bytes when calling convert_masked_utf8_to_utf32. If you skip the last 16 bytes,
      // and if the data is valid, then it is entirely safe because 16 UTF-8 bytes generate
      // much more than 8 bytes. However, you cannot generally assume that you have valid
      // UTF-8 input, so we are going to go back from the end counting 4 leading bytes,
      // to give us a good margin.
      size_t leading_byte = 0;
      size_t margin = size;
      for(; margin > 0 && leading_byte < 4; margin--) {
        leading_byte += (int8_t(in[margin-1]) > -65);
      }
      // If the input is long enough, then we have that margin-1 is the fourth last leading byte.
      const size_t safety_margin = size - margin + 1; // to avoid overruns!
      while(pos + 64 + safety_margin <= size) {
        simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
        if(input.is_ascii()) {
          input.store_ascii_as_utf32(utf32_output);
          utf32_output += 64;
          pos += 64;
        } 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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
              "We support either two or four chunks per 64-byte block.");
          auto zero = simd8<uint8_t>{uint8_t(0)};
          if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
            this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
            this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
          }
          uint64_t utf8_continuation_mask = input.lt(-65 + 1);
          uint64_t utf8_leading_mask = ~utf8_continuation_mask;
          uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
          // We process in blocks of up to 12 bytes except possibly
          // for fast paths which may process up to 16 bytes. For the
          // slow path to work, we should have at least 12 input bytes left.
          size_t max_starting_point = (pos + 64) - 12;
          // Next loop is going to run at least five times.
          while(pos < max_starting_point) {
            // Performance note: our ability to compute 'consumed' and
            // then shift and recompute is critical. If there is a
            // latency of, say, 4 cycles on getting 'consumed', then
            // the inner loop might have a total latency of about 6 cycles.
            // Yet we process between 6 to 12 inputs bytes, thus we get
            // a speed limit between 1 cycle/byte and 0.5 cycle/byte
            // for this section of the code. Hence, there is a limit
            // to how much we can further increase this latency before
            // it seriously harms performance.
            size_t consumed = convert_masked_utf8_to_utf32(in + pos,
                            utf8_end_of_code_point_mask, utf32_output);
            pos += consumed;
            utf8_end_of_code_point_mask >>= consumed;
          }
          // At this point there may remain between 0 and 12 bytes in the
          // 64-byte block. These bytes will be processed again. So we have an
          // 80% efficiency (in the worst case). In practice we expect an
          // 85% to 90% efficiency.
        }
      }
      if(errors()) { return 0; }
      if(pos < size) {
        size_t howmany  = scalar::utf8_to_utf32::convert(in + pos, size - pos, utf32_output);
        if(howmany == 0) { return 0; }
        utf32_output += howmany;
      }
      return utf32_output - start;
    }

    simdutf_really_inline result convert_with_errors(const char* in, size_t size, char32_t* utf32_output) {
      size_t pos = 0;
      char32_t* start{utf32_output};
      // In the worst case, we have the haswell kernel which can cause an overflow of
      // 8 bytes when calling convert_masked_utf8_to_utf32. If you skip the last 16 bytes,
      // and if the data is valid, then it is entirely safe because 16 UTF-8 bytes generate
      // much more than 8 bytes. However, you cannot generally assume that you have valid
      // UTF-8 input, so we are going to go back from the end counting 4 leading bytes,
      // to give us a good margin.
      size_t leading_byte = 0;
      size_t margin = size;
      for(; margin > 0 && leading_byte < 4; margin--) {
        leading_byte += (int8_t(in[margin-1]) > -65);
      }
      // If the input is long enough, then we have that margin-1 is the fourth last leading byte.
      const size_t safety_margin = size - margin + 1; // to avoid overruns!
      while(pos + 64 + safety_margin <= size) {
        simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
        if(input.is_ascii()) {
          input.store_ascii_as_utf32(utf32_output);
          utf32_output += 64;
          pos += 64;
        } 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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
              "We support either two or four chunks per 64-byte block.");
          auto zero = simd8<uint8_t>{uint8_t(0)};
          if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
            this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
            this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
          }
          if (errors()) {
            result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(pos, in + pos, size - pos, utf32_output);
            res.count += pos;
            return res;
          }
          uint64_t utf8_continuation_mask = input.lt(-65 + 1);
          uint64_t utf8_leading_mask = ~utf8_continuation_mask;
          uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
          // We process in blocks of up to 12 bytes except possibly
          // for fast paths which may process up to 16 bytes. For the
          // slow path to work, we should have at least 12 input bytes left.
          size_t max_starting_point = (pos + 64) - 12;
          // Next loop is going to run at least five times.
          while(pos < max_starting_point) {
            // Performance note: our ability to compute 'consumed' and
            // then shift and recompute is critical. If there is a
            // latency of, say, 4 cycles on getting 'consumed', then
            // the inner loop might have a total latency of about 6 cycles.
            // Yet we process between 6 to 12 inputs bytes, thus we get
            // a speed limit between 1 cycle/byte and 0.5 cycle/byte
            // for this section of the code. Hence, there is a limit
            // to how much we can further increase this latency before
            // it seriously harms performance.
            size_t consumed = convert_masked_utf8_to_utf32(in + pos,
                            utf8_end_of_code_point_mask, utf32_output);
            pos += consumed;
            utf8_end_of_code_point_mask >>= consumed;
          }
          // At this point there may remain between 0 and 12 bytes in the
          // 64-byte block. These bytes will be processed again. So we have an
          // 80% efficiency (in the worst case). In practice we expect an
          // 85% to 90% efficiency.
        }
      }
      if(errors()) {
        result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(pos, in + pos, size - pos, utf32_output);
        res.count += pos;
        return res;
      }
      if(pos < size) {
        result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(pos, in + pos, size - pos, utf32_output);
        if (res.error) {    // In case of error, we want the error position
          res.count += pos;
          return res;
        } else {    // In case of success, we want the number of word written
          utf32_output += res.count;
        }
      }
      return result(error_code::SUCCESS, utf32_output - start);
    }

    simdutf_really_inline bool errors() const {
      return this->error.any_bits_set_anywhere();
    }

  }; // struct utf8_checker
} // utf8_to_utf32 namespace
} // unnamed namespace
} // namespace ppc64
} // namespace simdutf
/* end file src/generic/utf8_to_utf32/utf8_to_utf32.h */
// other functions
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8.h
/* begin file src/generic/utf8.h */

namespace simdutf {
namespace ppc64 {
namespace {
namespace utf8 {

using namespace simd;

simdutf_really_inline size_t count_code_points(const char* in, size_t size) {
    size_t pos = 0;
    size_t count = 0;
    for(;pos + 64 <= size; pos += 64) {
      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
      uint64_t utf8_continuation_mask = input.lt(-65 + 1);
      count += 64 - count_ones(utf8_continuation_mask);
    }
    return count + scalar::utf8::count_code_points(in + pos, size - pos);
}


simdutf_really_inline size_t utf16_length_from_utf8(const char* in, size_t size) {
    size_t pos = 0;
    size_t count = 0;
    // This algorithm could no doubt be improved!
    for(;pos + 64 <= size; pos += 64) {
      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
      uint64_t utf8_continuation_mask = input.lt(-65 + 1);
      // We count one word for anything that is not a continuation (so
      // leading bytes).
      count += 64 - count_ones(utf8_continuation_mask);
      int64_t utf8_4byte = input.gteq_unsigned(240);
      count += count_ones(utf8_4byte);
    }
    return count + scalar::utf8::utf16_length_from_utf8(in + pos, size - pos);
}


simdutf_really_inline size_t utf32_length_from_utf8(const char* in, size_t size) {
    return count_code_points(in, size);
}
} // utf8 namespace
} // unnamed namespace
} // namespace ppc64
} // namespace simdutf
/* end file src/generic/utf8.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf16.h
/* begin file src/generic/utf16.h */
namespace simdutf {
namespace ppc64 {
namespace {
namespace utf16 {

template <endianness big_endian>
simdutf_really_inline size_t count_code_points(const char16_t* in, size_t size) {
    size_t pos = 0;
    size_t count = 0;
    for(;pos + 32 <= size; pos += 32) {
      simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
      if (!match_system(big_endian)) input.swap_bytes();
      uint64_t not_pair = input.not_in_range(0xDC00, 0xDFFF);
      count += count_ones(not_pair) / 2;
    }
    return count + scalar::utf16::count_code_points<big_endian>(in + pos, size - pos);
}

template <endianness big_endian>
simdutf_really_inline size_t utf8_length_from_utf16(const char16_t* in, size_t size) {
    size_t pos = 0;
    size_t count = 0;
    // This algorithm could no doubt be improved!
    for(;pos + 32 <= size; pos += 32) {
      simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
      if (!match_system(big_endian)) input.swap_bytes();
      uint64_t ascii_mask = input.lteq(0x7F);
      uint64_t twobyte_mask = input.lteq(0x7FF);
      uint64_t not_pair_mask = input.not_in_range(0xD800, 0xDFFF);

      size_t ascii_count = count_ones(ascii_mask) / 2;
      size_t twobyte_count = count_ones(twobyte_mask & ~ ascii_mask) / 2;
      size_t threebyte_count = count_ones(not_pair_mask & ~ twobyte_mask) / 2;
      size_t fourbyte_count = 32 - count_ones(not_pair_mask) / 2;
      count += 2 * fourbyte_count + 3 * threebyte_count + 2 * twobyte_count + ascii_count;
    }
    return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos, size - pos);
}

template <endianness big_endian>
simdutf_really_inline size_t utf32_length_from_utf16(const char16_t* in, size_t size) {
    return count_code_points<big_endian>(in, size);
}

simdutf_really_inline void change_endianness_utf16(const char16_t* in, size_t size, char16_t* output) {
  size_t pos = 0;

  while (pos + 32 <= size) {
    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
    input.swap_bytes();
    input.store(reinterpret_cast<uint16_t *>(output));
    pos += 32;
    output += 32;
  }

  scalar::utf16::change_endianness_utf16(in + pos, size - pos, output);
}

} // utf16
} // unnamed namespace
} // namespace ppc64
} // namespace simdutf
/* end file src/generic/utf16.h */

//
// Implementation-specific overrides
//
namespace simdutf {
namespace ppc64 {

simdutf_warn_unused int implementation::detect_encodings(const char * input, size_t length) const noexcept {
  // If there is a BOM, then we trust it.
  auto bom_encoding = simdutf::BOM::check_bom(input, length);
  if(bom_encoding != encoding_type::unspecified) { return bom_encoding; }
  int out = 0;
  if(validate_utf8(input, length)) { out |= encoding_type::UTF8; }
  if((length % 2) == 0) {
    if(validate_utf16(reinterpret_cast<const char16_t*>(input), length/2)) { out |= encoding_type::UTF16_LE; }
  }
  if((length % 4) == 0) {
    if(validate_utf32(reinterpret_cast<const char32_t*>(input), length/4)) { out |= encoding_type::UTF32_LE; }
  }

  return out;
}

simdutf_warn_unused bool implementation::validate_utf8(const char *buf, size_t len) const noexcept {
  return ppc64::utf8_validation::generic_validate_utf8(buf,len);
}

simdutf_warn_unused result implementation::validate_utf8_with_errors(const char *buf, size_t len) const noexcept {
  return ppc64::utf8_validation::generic_validate_utf8_with_errors(buf,len);
}

simdutf_warn_unused bool implementation::validate_ascii(const char *buf, size_t len) const noexcept {
  return ppc64::utf8_validation::generic_validate_ascii(buf,len);
}

simdutf_warn_unused result implementation::validate_ascii_with_errors(const char *buf, size_t len) const noexcept {
  return ppc64::utf8_validation::generic_validate_ascii_with_errors(buf,len);
}

simdutf_warn_unused bool implementation::validate_utf16le(const char16_t *buf, size_t len) const noexcept {
  return scalar::utf16::validate<endianness::LITTLE>(buf, len);
}

simdutf_warn_unused bool implementation::validate_utf16be(const char16_t *buf, size_t len) const noexcept {
  return scalar::utf16::validate<endianness::BIG>(buf, len);
}

simdutf_warn_unused result implementation::validate_utf16le_with_errors(const char16_t *buf, size_t len) const noexcept {
  return scalar::utf16::validate_with_errors<endianness::LITTLE>(buf, len);
}

simdutf_warn_unused result implementation::validate_utf16be_with_errors(const char16_t *buf, size_t len) const noexcept {
  return scalar::utf16::validate_with_errors<endianness::BIG>(buf, len);
}

simdutf_warn_unused result implementation::validate_utf32_with_errors(const char32_t *buf, size_t len) const noexcept {
  return scalar::utf32::validate_with_errors(buf, len);
}

simdutf_warn_unused bool implementation::validate_utf32(const char16_t *buf, size_t len) const noexcept {
  return scalar::utf32::validate(buf, len);
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(const char* /*buf*/, size_t /*len*/, char16_t* /*utf16_output*/) const noexcept {
  return 0; // stub
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf16be(const char* /*buf*/, size_t /*len*/, char16_t* /*utf16_output*/) const noexcept {
  return 0; // stub
}

simdutf_warn_unused result implementation::convert_utf8_to_utf16le_with_errors(const char* /*buf*/, size_t /*len*/, char16_t* /*utf16_output*/) const noexcept {
  return result(error_code::OTHER, 0); // stub
}

simdutf_warn_unused result implementation::convert_utf8_to_utf16be_with_errors(const char* /*buf*/, size_t /*len*/, char16_t* /*utf16_output*/) const noexcept {
  return result(error_code::OTHER, 0); // stub
}

simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16le(const char* /*buf*/, size_t /*len*/, char16_t* /*utf16_output*/) const noexcept {
  return 0; // stub
}

simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(const char* /*buf*/, size_t /*len*/, char16_t* /*utf16_output*/) const noexcept {
  return 0; // stub
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(const char* /*buf*/, size_t /*len*/, char32_t* /*utf16_output*/) const noexcept {
  return 0; // stub
}

simdutf_warn_unused result implementation::convert_utf8_to_utf32_with_errors(const char* /*buf*/, size_t /*len*/, char32_t* /*utf16_output*/) const noexcept {
  return result(error_code::OTHER, 0); // stub
}

simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(const char* /*buf*/, size_t /*len*/, char32_t* /*utf16_output*/) const noexcept {
  return 0; // stub
}

simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf16_to_utf8::convert<endianness::LITTLE>(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_utf16be_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf16_to_utf8::convert<endianness::BIG>(buf, len, utf8_output);
}

simdutf_warn_unused result implementation::convert_utf16le_to_utf8_with_errors(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf16_to_utf8::convert_with_errors<endianness::LITTLE>(buf, len, utf8_output);
}

simdutf_warn_unused result implementation::convert_utf16be_to_utf8_with_errors(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf16_to_utf8::convert_with_errors<endianness::BIG>(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf16_to_utf8::convert_valid<endianness::LITTLE>(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf16_to_utf8::convert_valid<endianness::BIG>(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf32_to_utf8::convert(buf, len, utf8_output);
}

simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf32_to_utf8::convert_with_errors(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  return scalar::utf32_to_utf8::convert_valid(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return scalar::utf32_to_utf16::convert<endianness::LITTLE>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf16be(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return scalar::utf32_to_utf16::convert<endianness::BIG>(buf, len, utf16_output);
}

simdutf_warn_unused result implementation::convert_utf32_to_utf16le_with_errors(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return scalar::utf32_to_utf16::convert_with_errors<endianness::LITTLE>(buf, len, utf16_output);
}

simdutf_warn_unused result implementation::convert_utf32_to_utf16be_with_errors(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return scalar::utf32_to_utf16::convert_with_errors<endianness::BIG>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16le(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return scalar::utf32_to_utf16::convert_valid<endianness::LITTLE>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16be(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return scalar::utf32_to_utf16::convert_valid<endianness::BIG>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return scalar::utf16_to_utf32::convert<endianness::LITTLE>(buf, len, utf32_output);
}

simdutf_warn_unused size_t implementation::convert_utf16be_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return scalar::utf16_to_utf32::convert<endianness::BIG>(buf, len, utf32_output);
}

simdutf_warn_unused result implementation::convert_utf16le_to_utf32_with_errors(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(buf, len, utf32_output);
}

simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(buf, len, utf32_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return scalar::utf16_to_utf32::convert_valid<endianness::LITTLE>(buf, len, utf32_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return scalar::utf16_to_utf32::convert_valid<endianness::BIG>(buf, len, utf32_output);
}

void implementation::change_endianness_utf16(const char16_t * input, size_t length, char16_t * output) const noexcept {
  scalar::utf16::change_endianness_utf16(input, length, output);
}

simdutf_warn_unused size_t implementation::count_utf16le(const char16_t * input, size_t length) const noexcept {
  return scalar::utf16::count_code_points<endianness::LITTLE>(input, length);
}

simdutf_warn_unused size_t implementation::count_utf16be(const char16_t * input, size_t length) const noexcept {
  return scalar::utf16::count_code_points<endianness::BIG>(input, length);
}

simdutf_warn_unused size_t implementation::count_utf8(const char * input, size_t length) const noexcept {
  return utf8::count_code_points(input, length);
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(const char16_t * input, size_t length) const noexcept {
  return scalar::utf16::utf8_length_from_utf16<endianness::LITTLE>(input, length);
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(const char16_t * input, size_t length) const noexcept {
  return scalar::utf16::utf8_length_from_utf16<endianness::BIG>(input, length);
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(const char16_t * input, size_t length) const noexcept {
  return scalar::utf16::utf32_length_from_utf16<endianness::LITTLE>(input, length);
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(const char16_t * input, size_t length) const noexcept {
  return scalar::utf16::utf32_length_from_utf16<endianness::BIG>(input, length);
}

simdutf_warn_unused size_t implementation::utf16_length_from_utf8(const char * input, size_t length) const noexcept {
  return scalar::utf8::utf16_length_from_utf8(input, length);
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf32(const char32_t * input, size_t length) const noexcept {
  return scalar::utf32::utf8_length_from_utf32(input, length);
}

simdutf_warn_unused size_t implementation::utf16_length_from_utf32(const char32_t * input, size_t length) const noexcept {
  return scalar::utf32::utf16_length_from_utf32(input, length);
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf8(const char * input, size_t length) const noexcept {
  return scalar::utf8::count_code_points(input, length);
}

} // namespace ppc64
} // namespace simdutf

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/ppc64/end.h
/* begin file src/simdutf/ppc64/end.h */
/* end file src/simdutf/ppc64/end.h */
/* end file src/ppc64/implementation.cpp */
#endif
#if SIMDUTF_IMPLEMENTATION_WESTMERE
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=westmere/implementation.cpp
/* begin file src/westmere/implementation.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/westmere/begin.h
/* begin file src/simdutf/westmere/begin.h */
// redefining SIMDUTF_IMPLEMENTATION to "westmere"
// #define SIMDUTF_IMPLEMENTATION westmere

#if SIMDUTF_CAN_ALWAYS_RUN_WESTMERE
// nothing needed.
#else
SIMDUTF_TARGET_WESTMERE
#endif
/* end file src/simdutf/westmere/begin.h */
namespace simdutf {
namespace westmere {
namespace {
#ifndef SIMDUTF_WESTMERE_H
#error "westmere.h must be included"
#endif
using namespace simd;

simdutf_really_inline bool is_ascii(const simd8x64<uint8_t>& input) {
  return input.reduce_or().is_ascii();
}

simdutf_unused simdutf_really_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(0b11000000u-1); // Only 11______ will be > 0
  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0b11100000u-1); // Only 111_____ will be > 0
  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0b11110000u-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);
}

simdutf_really_inline simd8<bool> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0b11100000u-1); // Only 111_____ will be > 0
  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0b11110000u-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_third_byte | is_fourth_byte) > int8_t(0);
}

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=westmere/sse_detect_encodings.cpp
/* begin file src/westmere/sse_detect_encodings.cpp */
template<class checker>
// len is known to be a multiple of 2 when this is called
int sse_detect_encodings(const char * buf, size_t len) {
    const char* start = buf;
    const char* end = buf + len;

    bool is_utf8 = true;
    bool is_utf16 = true;
    bool is_utf32 = true;

    int out = 0;

    const auto v_d8 = simd8<uint8_t>::splat(0xd8);
    const auto v_f8 = simd8<uint8_t>::splat(0xf8);

    __m128i currentmax = _mm_setzero_si128();

    checker check{};

    while(buf + 64 <= end) {
        __m128i in = _mm_loadu_si128((__m128i*)buf);
        __m128i secondin = _mm_loadu_si128((__m128i*)buf+1);
        __m128i thirdin = _mm_loadu_si128((__m128i*)buf+2);
        __m128i fourthin = _mm_loadu_si128((__m128i*)buf+3);

        const auto u0 = simd16<uint16_t>(in);
        const auto u1 = simd16<uint16_t>(secondin);
        const auto u2 = simd16<uint16_t>(thirdin);
        const auto u3 = simd16<uint16_t>(fourthin);

        const auto v0 = u0.shr<8>();
        const auto v1 = u1.shr<8>();
        const auto v2 = u2.shr<8>();
        const auto v3 = u3.shr<8>();

        const auto in16 = simd16<uint16_t>::pack(v0, v1);
        const auto nextin16 = simd16<uint16_t>::pack(v2, v3);

        const auto surrogates_wordmask0 = (in16 & v_f8) == v_d8;
        const auto surrogates_wordmask1 = (nextin16 & v_f8) == v_d8;
        uint16_t surrogates_bitmask0 = static_cast<uint16_t>(surrogates_wordmask0.to_bitmask());
        uint16_t surrogates_bitmask1 = static_cast<uint16_t>(surrogates_wordmask1.to_bitmask());

        // Check for surrogates
        if (surrogates_bitmask0 != 0x0 || surrogates_bitmask1 != 0x0) {
            // Cannot be UTF8
            is_utf8 = false;
            // Can still be either UTF-16LE or UTF-32 depending on the positions of the surrogates
            // To be valid UTF-32, a surrogate cannot be in the two most significant bytes of any 32-bit word.
            // On the other hand, to be valid UTF-16LE, at least one surrogate must be in the two most significant
            // bytes of a 32-bit word since they always come in pairs in UTF-16LE.
            // Note that we always proceed in multiple of 4 before this point so there is no offset in 32-bit words.

            if (((surrogates_bitmask0 | surrogates_bitmask1) & 0xaaaa) != 0) {
                is_utf32 = false;
                // Code from sse_validate_utf16le.cpp
                // Not efficient, we do not process surrogates_bitmask1
                const char16_t * input = reinterpret_cast<const char16_t*>(buf);
                const char16_t* end16 = reinterpret_cast<const char16_t*>(start) + len/2;

                const auto v_fc = simd8<uint8_t>::splat(0xfc);
                const auto v_dc = simd8<uint8_t>::splat(0xdc);

                const uint16_t V0 = static_cast<uint16_t>(~surrogates_bitmask0);

                const auto    vH0 = (in16 & v_fc) == v_dc;
                const uint16_t H0 = static_cast<uint16_t>(vH0.to_bitmask());

                const uint16_t L0 = static_cast<uint16_t>(~H0 & surrogates_bitmask0);

                const uint16_t a0 = static_cast<uint16_t>(L0 & (H0 >> 1));

                const uint16_t b0 = static_cast<uint16_t>(a0 << 1);

                const uint16_t c0 = static_cast<uint16_t>(V0 | a0 | b0);

                if (c0 == 0xffff) {
                    input += 16;
                } else if (c0 == 0x7fff) {
                    input += 15;
                } else {
                    is_utf16 = false;
                    break;
                }

                while (input + simd16<uint16_t>::SIZE * 2 < end16) {
                    const auto in0 = simd16<uint16_t>(input);
                    const auto in1 = simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));

                    const auto t0 = in0.shr<8>();
                    const auto t1 = in1.shr<8>();

                    const auto in_16 = simd16<uint16_t>::pack(t0, t1);

                    const auto surrogates_wordmask = (in_16 & v_f8) == v_d8;
                    const uint16_t surrogates_bitmask = static_cast<uint16_t>(surrogates_wordmask.to_bitmask());
                    if (surrogates_bitmask == 0x0) {
                        input += 16;
                    } else {
                        const uint16_t V = static_cast<uint16_t>(~surrogates_bitmask);

                        const auto    vH = (in_16 & v_fc) == v_dc;
                        const uint16_t H = static_cast<uint16_t>(vH.to_bitmask());

                        const uint16_t L = static_cast<uint16_t>(~H & surrogates_bitmask);

                        const uint16_t a = static_cast<uint16_t>(L & (H >> 1));

                        const uint16_t b = static_cast<uint16_t>(a << 1);

                        const uint16_t c = static_cast<uint16_t>(V | a | b);

                        if (c == 0xffff) {
                            input += 16;
                        } else if (c == 0x7fff) {
                            input += 15;
                        } else {
                            is_utf16 = false;
                            break;
                        }
                    }
                }
            } else {
                is_utf16 = false;
                // Check for UTF-32
                if (len % 4 == 0) {
                    const char32_t * input = reinterpret_cast<const char32_t*>(buf);
                    const char32_t* end32 = reinterpret_cast<const char32_t*>(start) + len/4;

                    // Must start checking for surrogates
                    __m128i currentoffsetmax = _mm_setzero_si128();
                    const __m128i offset = _mm_set1_epi32(0xffff2000);
                    const __m128i standardoffsetmax = _mm_set1_epi32(0xfffff7ff);

                    currentmax = _mm_max_epu32(in, currentmax);
                    currentmax = _mm_max_epu32(secondin, currentmax);
                    currentmax = _mm_max_epu32(thirdin, currentmax);
                    currentmax = _mm_max_epu32(fourthin, currentmax);

                    currentoffsetmax = _mm_max_epu32(_mm_add_epi32(in, offset), currentoffsetmax);
                    currentoffsetmax = _mm_max_epu32(_mm_add_epi32(secondin, offset), currentoffsetmax);
                    currentoffsetmax = _mm_max_epu32(_mm_add_epi32(thirdin, offset), currentoffsetmax);
                    currentoffsetmax = _mm_max_epu32(_mm_add_epi32(fourthin, offset), currentoffsetmax);

                    while (input + 4 < end32) {
                        const __m128i in32 = _mm_loadu_si128((__m128i *)input);
                        currentmax = _mm_max_epu32(in32,currentmax);
                        currentoffsetmax = _mm_max_epu32(_mm_add_epi32(in32, offset), currentoffsetmax);
                        input += 4;
                    }

                    __m128i forbidden_words = _mm_xor_si128(_mm_max_epu32(currentoffsetmax, standardoffsetmax), standardoffsetmax);
                    if(_mm_testz_si128(forbidden_words, forbidden_words) == 0) {
                        is_utf32 = false;
                    }
                } else {
                    is_utf32 = false;
                }
            }
            break;
        }
        // If no surrogate, validate under other encodings as well

        // UTF-32 validation
        currentmax = _mm_max_epu32(in, currentmax);
        currentmax = _mm_max_epu32(secondin, currentmax);
        currentmax = _mm_max_epu32(thirdin, currentmax);
        currentmax = _mm_max_epu32(fourthin, currentmax);

        // UTF-8 validation
        // Relies on ../generic/utf8_validation/utf8_lookup4_algorithm.h
        simd::simd8x64<uint8_t> in8(in, secondin, thirdin, fourthin);
        check.check_next_input(in8);

        buf += 64;
    }

    // Check which encodings are possible

    if (is_utf8) {
        if (static_cast<size_t>(buf - start) != len) {
            uint8_t block[64]{};
            std::memset(block, 0x20, 64);
            std::memcpy(block, buf, len - (buf - start));
            simd::simd8x64<uint8_t> in(block);
            check.check_next_input(in);
        }
        if (!check.errors()) {
            out |= simdutf::encoding_type::UTF8;
        }
    }

    if (is_utf16 && scalar::utf16::validate<endianness::LITTLE>(reinterpret_cast<const char16_t*>(buf), (len - (buf - start))/2)) {
        out |= simdutf::encoding_type::UTF16_LE;
    }

    if (is_utf32 && (len % 4 == 0)) {
        const __m128i standardmax = _mm_set1_epi32(0x10ffff);
        __m128i is_zero = _mm_xor_si128(_mm_max_epu32(currentmax, standardmax), standardmax);
        if (_mm_testz_si128(is_zero, is_zero) == 1 && scalar::utf32::validate(reinterpret_cast<const char32_t*>(buf), (len - (buf - start))/4)) {
            out |= simdutf::encoding_type::UTF32_LE;
        }
    }

    return out;
}
/* end file src/westmere/sse_detect_encodings.cpp */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=westmere/sse_validate_utf16.cpp
/* begin file src/westmere/sse_validate_utf16.cpp */
/*
    In UTF-16 words in range 0xD800 to 0xDFFF have special meaning.

    In a vectorized algorithm we want to examine the most significant
    nibble in order to select a fast path. If none of highest nibbles
    are 0xD (13), than we are sure that UTF-16 chunk in a vector
    register is valid.

    Let us analyze what we need to check if the nibble is 0xD. The
    value of the preceding nibble determines what we have:

    0xd000 .. 0xd7ff - a valid word
    0xd800 .. 0xdbff - low surrogate
    0xdc00 .. 0xdfff - high surrogate

    Other constraints we have to consider:
    - there must not be two consecutive low surrogates (0xd800 .. 0xdbff)
    - there must not be two consecutive high surrogates (0xdc00 .. 0xdfff)
    - there must not be sole low surrogate nor high surrogate

    We're going to build three bitmasks based on the 3rd nibble:
    - V = valid word,
    - L = low surrogate (0xd800 .. 0xdbff)
    - H = high surrogate (0xdc00 .. 0xdfff)

      0   1   2   3   4   5   6   7    <--- word index
    [ V | L | H | L | H | V | V | L ]
      1   0   0   0   0   1   1   0     - V = valid masks
      0   1   0   1   0   0   0   1     - L = low surrogate
      0   0   1   0   1   0   0   0     - H high surrogate


      1   0   0   0   0   1   1   0   V = valid masks
      0   1   0   1   0   0   0   0   a = L & (H >> 1)
      0   0   1   0   1   0   0   0   b = a << 1
      1   1   1   1   1   1   1   0   c = V | a | b
                                  ^
                                  the last bit can be zero, we just consume 7 words
                                  and recheck this word in the next iteration
*/

/* Returns:
   - pointer to the last unprocessed character (a scalar fallback should check the rest);
   - nullptr if an error was detected.
*/
template <endianness big_endian>
const char16_t* sse_validate_utf16(const char16_t* input, size_t size) {
    const char16_t* end = input + size;

    const auto v_d8 = simd8<uint8_t>::splat(0xd8);
    const auto v_f8 = simd8<uint8_t>::splat(0xf8);
    const auto v_fc = simd8<uint8_t>::splat(0xfc);
    const auto v_dc = simd8<uint8_t>::splat(0xdc);

    while (input + simd16<uint16_t>::SIZE * 2 < end) {
        // 0. Load data: since the validation takes into account only higher
        //    byte of each word, we compress the two vectors into one which
        //    consists only the higher bytes.
        auto in0 = simd16<uint16_t>(input);
        auto in1 = simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));
        if (big_endian) {
            in0 = in0.swap_bytes();
            in1 = in1.swap_bytes();
        }

        const auto t0 = in0.shr<8>();
        const auto t1 = in1.shr<8>();

        const auto in = simd16<uint16_t>::pack(t0, t1);

        // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
        const auto surrogates_wordmask = (in & v_f8) == v_d8;
        const uint16_t surrogates_bitmask = static_cast<uint16_t>(surrogates_wordmask.to_bitmask());
        if (surrogates_bitmask == 0x0000) {
            input += 16;
        } else {
            // 2. We have some surrogates that have to be distinguished:
            //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
            //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
            //
            //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)

            // V - non-surrogate words
            //     V = not surrogates_wordmask
            const uint16_t V = static_cast<uint16_t>(~surrogates_bitmask);

            // H - word-mask for high surrogates: the six highest bits are 0b1101'11
            const auto    vH = (in & v_fc) == v_dc;
            const uint16_t H = static_cast<uint16_t>(vH.to_bitmask());

            // L - word mask for low surrogates
            //     L = not H and surrogates_wordmask
            const uint16_t L = static_cast<uint16_t>(~H & surrogates_bitmask);

            const uint16_t a = static_cast<uint16_t>(L & (H >> 1));  // A low surrogate must be followed by high one.
                                              // (A low surrogate placed in the 7th register's word
                                              // is an exception we handle.)
            const uint16_t b = static_cast<uint16_t>(a << 1);        // Just mark that the opinput - startite fact is hold,
                                              // thanks to that we have only two masks for valid case.
            const uint16_t c = static_cast<uint16_t>(V | a | b);     // Combine all the masks into the final one.

            if (c == 0xffff) {
                // The whole input register contains valid UTF-16, i.e.,
                // either single words or proper surrogate pairs.
                input += 16;
            } else if (c == 0x7fff) {
                // The 15 lower words of the input register contains valid UTF-16.
                // The 15th word may be either a low or high surrogate. It the next
                // iteration we 1) check if the low surrogate is followed by a high
                // one, 2) reject sole high surrogate.
                input += 15;
            } else {
                return nullptr;
            }
        }
    }

    return input;
}


template <endianness big_endian>
const result sse_validate_utf16_with_errors(const char16_t* input, size_t size) {
    const char16_t* start = input;
    const char16_t* end = input + size;

    const auto v_d8 = simd8<uint8_t>::splat(0xd8);
    const auto v_f8 = simd8<uint8_t>::splat(0xf8);
    const auto v_fc = simd8<uint8_t>::splat(0xfc);
    const auto v_dc = simd8<uint8_t>::splat(0xdc);

    while (input + simd16<uint16_t>::SIZE * 2 < end) {
        // 0. Load data: since the validation takes into account only higher
        //    byte of each word, we compress the two vectors into one which
        //    consists only the higher bytes.
        auto in0 = simd16<uint16_t>(input);
        auto in1 = simd16<uint16_t>(input + simd16<uint16_t>::SIZE / sizeof(char16_t));

        if (big_endian) {
            in0 = in0.swap_bytes();
            in1 = in1.swap_bytes();
        }

        const auto t0 = in0.shr<8>();
        const auto t1 = in1.shr<8>();

        const auto in = simd16<uint16_t>::pack(t0, t1);

        // 1. Check whether we have any 0xD800..DFFF word (0b1101'1xxx'yyyy'yyyy).
        const auto surrogates_wordmask = (in & v_f8) == v_d8;
        const uint16_t surrogates_bitmask = static_cast<uint16_t>(surrogates_wordmask.to_bitmask());
        if (surrogates_bitmask == 0x0000) {
            input += 16;
        } else {
            // 2. We have some surrogates that have to be distinguished:
            //    - low  surrogates: 0b1101'10xx'yyyy'yyyy (0xD800..0xDBFF)
            //    - high surrogates: 0b1101'11xx'yyyy'yyyy (0xDC00..0xDFFF)
            //
            //    Fact: high surrogate has 11th bit set (3rd bit in the higher word)

            // V - non-surrogate words
            //     V = not surrogates_wordmask
            const uint16_t V = static_cast<uint16_t>(~surrogates_bitmask);

            // H - word-mask for high surrogates: the six highest bits are 0b1101'11
            const auto    vH = (in & v_fc) == v_dc;
            const uint16_t H = static_cast<uint16_t>(vH.to_bitmask());

            // L - word mask for low surrogates
            //     L = not H and surrogates_wordmask
            const uint16_t L = static_cast<uint16_t>(~H & surrogates_bitmask);

            const uint16_t a = static_cast<uint16_t>(L & (H >> 1));  // A low surrogate must be followed by high one.
                                              // (A low surrogate placed in the 7th register's word
                                              // is an exception we handle.)
            const uint16_t b = static_cast<uint16_t>(a << 1);        // Just mark that the opinput - startite fact is hold,
                                              // thanks to that we have only two masks for valid case.
            const uint16_t c = static_cast<uint16_t>(V | a | b);     // Combine all the masks into the final one.

            if (c == 0xffff) {
                // The whole input register contains valid UTF-16, i.e.,
                // either single words or proper surrogate pairs.
                input += 16;
            } else if (c == 0x7fff) {
                // The 15 lower words of the input register contains valid UTF-16.
                // The 15th word may be either a low or high surrogate. It the next
                // iteration we 1) check if the low surrogate is followed by a high
                // one, 2) reject sole high surrogate.
                input += 15;
            } else {
                return result(error_code::SURROGATE, input - start);
            }
        }
    }

    return result(error_code::SUCCESS, input - start);
}
/* end file src/westmere/sse_validate_utf16.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=westmere/sse_validate_utf32le.cpp
/* begin file src/westmere/sse_validate_utf32le.cpp */
/* Returns:
   - pointer to the last unprocessed character (a scalar fallback should check the rest);
   - nullptr if an error was detected.
*/
const char32_t* sse_validate_utf32le(const char32_t* input, size_t size) {
    const char32_t* end = input + size;

    const __m128i standardmax = _mm_set1_epi32(0x10ffff);
    const __m128i offset = _mm_set1_epi32(0xffff2000);
    const __m128i standardoffsetmax = _mm_set1_epi32(0xfffff7ff);
    __m128i currentmax = _mm_setzero_si128();
    __m128i currentoffsetmax = _mm_setzero_si128();

    while (input + 4 < end) {
        const __m128i in = _mm_loadu_si128((__m128i *)input);
        currentmax = _mm_max_epu32(in,currentmax);
        currentoffsetmax = _mm_max_epu32(_mm_add_epi32(in, offset), currentoffsetmax);
        input += 4;
    }
    __m128i is_zero = _mm_xor_si128(_mm_max_epu32(currentmax, standardmax), standardmax);
    if(_mm_test_all_zeros(is_zero, is_zero) == 0) {
        return nullptr;
    }

    is_zero = _mm_xor_si128(_mm_max_epu32(currentoffsetmax, standardoffsetmax), standardoffsetmax);
    if(_mm_test_all_zeros(is_zero, is_zero) == 0) {
        return nullptr;
    }

    return input;
}


const result sse_validate_utf32le_with_errors(const char32_t* input, size_t size) {
    const char32_t* start = input;
    const char32_t* end = input + size;

    const __m128i standardmax = _mm_set1_epi32(0x10ffff);
    const __m128i offset = _mm_set1_epi32(0xffff2000);
    const __m128i standardoffsetmax = _mm_set1_epi32(0xfffff7ff);
    __m128i currentmax = _mm_setzero_si128();
    __m128i currentoffsetmax = _mm_setzero_si128();

    while (input + 4 < end) {
        const __m128i in = _mm_loadu_si128((__m128i *)input);
        currentmax = _mm_max_epu32(in,currentmax);
        currentoffsetmax = _mm_max_epu32(_mm_add_epi32(in, offset), currentoffsetmax);

        __m128i is_zero = _mm_xor_si128(_mm_max_epu32(currentmax, standardmax), standardmax);
        if(_mm_test_all_zeros(is_zero, is_zero) == 0) {
            return result(error_code::TOO_LARGE, input - start);
        }

        is_zero = _mm_xor_si128(_mm_max_epu32(currentoffsetmax, standardoffsetmax), standardoffsetmax);
        if(_mm_test_all_zeros(is_zero, is_zero) == 0) {
            return result(error_code::SURROGATE, input - start);
        }
        input += 4;
    }

    return result(error_code::SUCCESS, input - start);
}
/* end file src/westmere/sse_validate_utf32le.cpp */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=westmere/sse_convert_utf8_to_utf16.cpp
/* begin file src/westmere/sse_convert_utf8_to_utf16.cpp */
// depends on "tables/utf8_to_utf16_tables.h"


// Convert up to 12 bytes from utf8 to utf16 using a mask indicating the
// end of the code points. Only the least significant 12 bits of the mask
// are accessed.
// It returns how many bytes were consumed (up to 12).
template <endianness big_endian>
size_t convert_masked_utf8_to_utf16(const char *input,
                           uint64_t utf8_end_of_code_point_mask,
                           char16_t *&utf16_output) {
  // we use an approach where we try to process up to 12 input bytes.
  // Why 12 input bytes and not 16? Because we are concerned with the size of
  // the lookup tables. Also 12 is nicely divisible by two and three.
  //
  //
  // Optimization note: our main path below is load-latency dependent. Thus it is maybe
  // beneficial to have fast paths that depend on branch prediction but have less latency.
  // This results in more instructions but, potentially, also higher speeds.
  //
  // We first try a few fast paths.
  const __m128i swap = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
  const __m128i in = _mm_loadu_si128((__m128i *)input);
  const uint16_t input_utf8_end_of_code_point_mask =
      utf8_end_of_code_point_mask & 0xfff;
  if(((utf8_end_of_code_point_mask & 0xffff) == 0xffff)) {
    // We process the data in chunks of 16 bytes.
    __m128i ascii_first = _mm_cvtepu8_epi16(in);
    __m128i ascii_second = _mm_cvtepu8_epi16(_mm_srli_si128(in,8));
    if (big_endian) {
      ascii_first = _mm_shuffle_epi8(ascii_first, swap);
      ascii_second = _mm_shuffle_epi8(ascii_second, swap);
    }
    _mm_storeu_si128(reinterpret_cast<__m128i *>(utf16_output), ascii_first);
    _mm_storeu_si128(reinterpret_cast<__m128i *>(utf16_output + 8), ascii_second);
    utf16_output += 16; // We wrote 16 16-bit characters.
    return 16; // We consumed 16 bytes.
  }
  if(((utf8_end_of_code_point_mask & 0xFFFF) == 0xaaaa)) {
    // We want to take 8 2-byte UTF-8 words and turn them into 8 2-byte UTF-16 words.
    // There is probably a more efficient sequence, but the following might do.
    const __m128i sh = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
    __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
    if (big_endian) composed = _mm_shuffle_epi8(composed, swap);
    _mm_storeu_si128((__m128i *)utf16_output, composed);
    utf16_output += 8; // We wrote 16 bytes, 8 code points.
    return 16;
  }
  if(input_utf8_end_of_code_point_mask == 0x924) {
    // We want to take 4 3-byte UTF-8 words and turn them into 4 2-byte UTF-16 words.
    // There is probably a more efficient sequence, but the following might do.
    const __m128i sh = _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii =
        _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits
    const __m128i middlebyte =
        _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits
    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
    const __m128i highbyte =
        _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits
    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4);
    const __m128i composed =
        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted);
    __m128i composed_repacked = _mm_packus_epi32(composed, composed);
    if (big_endian) composed_repacked = _mm_shuffle_epi8(composed_repacked, swap);
    _mm_storeu_si128((__m128i *)utf16_output, composed_repacked);
    utf16_output += 4;
    return 12;
  }
  /// We do not have a fast path available, so we fallback.

  const uint8_t idx =
      tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0];
  const uint8_t consumed =
      tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1];
  if (idx < 64) {
    // SIX (6) input code-words
    // this is a relatively easy scenario
    // we process SIX (6) input code-words. The max length in bytes of six code
    // words spanning between 1 and 2 bytes each is 12 bytes. On processors
    // where pdep/pext is fast, we might be able to use a small lookup table.
    const __m128i sh =
        _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
    __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
    if (big_endian) composed = _mm_shuffle_epi8(composed, swap);
    _mm_storeu_si128((__m128i *)utf16_output, composed);
    utf16_output += 6; // We wrote 12 bytes, 6 code points.
  } else if (idx < 145) {
    // FOUR (4) input code-words
    const __m128i sh =
        _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii =
        _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits
    const __m128i middlebyte =
        _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits
    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
    const __m128i highbyte =
        _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits
    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4);
    const __m128i composed =
        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted);
     __m128i composed_repacked = _mm_packus_epi32(composed, composed);
    if (big_endian) composed_repacked = _mm_shuffle_epi8(composed_repacked, swap);
    _mm_storeu_si128((__m128i *)utf16_output, composed_repacked);
    utf16_output += 4;
  } else if (idx < 209) {
    // TWO (2) input code-words
    //////////////
    // There might be garbage inputs where a leading byte mascarades as a four-byte
    // leading byte (by being followed by 3 continuation byte), but is not greater than
    // 0xf0. This could trigger a buffer overflow if we only counted leading
    // bytes of the form 0xf0 as generating surrogate pairs, without further UTF-8 validation.
    // Thus we must be careful to ensure that only leading bytes at least as large as 0xf0 generate surrogate pairs.
    // We do as at the cost of an extra mask.
    /////////////
    const __m128i sh =
        _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f));
    const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00));
    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
    __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000));
    // correct for spurious high bit
    const __m128i correct =
        _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1);
    middlehighbyte = _mm_xor_si128(correct, middlehighbyte);
    const __m128i middlehighbyte_shifted = _mm_srli_epi32(middlehighbyte, 4);
    // We deliberately carry the leading four bits in highbyte if they are present,
    // we remove them later when computing hightenbits.
    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0xff000000));
    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 6);
    // When we need to generate a surrogate pair (leading byte > 0xF0), then
    // the corresponding 32-bit value in 'composed'  will be greater than
    // > (0xff00000>>6) or > 0x3c00000. This can be used later to identify the
    // location of the surrogate pairs.
    const __m128i composed =
        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted),
                     _mm_or_si128(highbyte_shifted, middlehighbyte_shifted));
    const __m128i composedminus =
        _mm_sub_epi32(composed, _mm_set1_epi32(0x10000));
    const __m128i lowtenbits =
        _mm_and_si128(composedminus, _mm_set1_epi32(0x3ff));
    // Notice the 0x3ff mask:
    const __m128i hightenbits = _mm_and_si128(_mm_srli_epi32(composedminus, 10), _mm_set1_epi32(0x3ff));
    const __m128i lowtenbitsadd =
        _mm_add_epi32(lowtenbits, _mm_set1_epi32(0xDC00));
    const __m128i hightenbitsadd =
        _mm_add_epi32(hightenbits, _mm_set1_epi32(0xD800));
    const __m128i lowtenbitsaddshifted = _mm_slli_epi32(lowtenbitsadd, 16);
    __m128i surrogates =
        _mm_or_si128(hightenbitsadd, lowtenbitsaddshifted);
    uint32_t basic_buffer[4];
    uint32_t basic_buffer_swap[4];
    if (big_endian) {
      _mm_storeu_si128((__m128i *)basic_buffer_swap, _mm_shuffle_epi8(composed, swap));
      surrogates = _mm_shuffle_epi8(surrogates, swap);
    }
    _mm_storeu_si128((__m128i *)basic_buffer, composed);
    uint32_t surrogate_buffer[4];
    _mm_storeu_si128((__m128i *)surrogate_buffer, surrogates);
    for (size_t i = 0; i < 3; i++) {
      if(basic_buffer[i] > 0x3c00000) {
        utf16_output[0] = uint16_t(surrogate_buffer[i] & 0xffff);
        utf16_output[1] = uint16_t(surrogate_buffer[i] >> 16);
        utf16_output += 2;
      } else {
        utf16_output[0] = big_endian ? uint16_t(basic_buffer_swap[i]) : uint16_t(basic_buffer[i]);
        utf16_output++;
      }
    }
  } else {
    // here we know that there is an error but we do not handle errors
  }
  return consumed;
}
/* end file src/westmere/sse_convert_utf8_to_utf16.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=westmere/sse_convert_utf8_to_utf32.cpp
/* begin file src/westmere/sse_convert_utf8_to_utf32.cpp */
// depends on "tables/utf8_to_utf16_tables.h"


// Convert up to 12 bytes from utf8 to utf32 using a mask indicating the
// end of the code points. Only the least significant 12 bits of the mask
// are accessed.
// It returns how many bytes were consumed (up to 12).
size_t convert_masked_utf8_to_utf32(const char *input,
                           uint64_t utf8_end_of_code_point_mask,
                           char32_t *&utf32_output) {
  // we use an approach where we try to process up to 12 input bytes.
  // Why 12 input bytes and not 16? Because we are concerned with the size of
  // the lookup tables. Also 12 is nicely divisible by two and three.
  //
  //
  // Optimization note: our main path below is load-latency dependent. Thus it is maybe
  // beneficial to have fast paths that depend on branch prediction but have less latency.
  // This results in more instructions but, potentially, also higher speeds.
  //
  // We first try a few fast paths.
  const __m128i in = _mm_loadu_si128((__m128i *)input);
  const uint16_t input_utf8_end_of_code_point_mask =
      utf8_end_of_code_point_mask & 0xfff;
  if(((utf8_end_of_code_point_mask & 0xffff) == 0xffff)) {
    // We process the data in chunks of 16 bytes.
    _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output), _mm_cvtepu8_epi32(in));
    _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output+4), _mm_cvtepu8_epi32(_mm_srli_si128(in,4)));
    _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output+8), _mm_cvtepu8_epi32(_mm_srli_si128(in,8)));
    _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output+12), _mm_cvtepu8_epi32(_mm_srli_si128(in,12)));
    utf32_output += 16; // We wrote 16 32-bit characters.
    return 16; // We consumed 16 bytes.
  }
  if(((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) {
    // We want to take 8 2-byte UTF-8 words and turn them into 8 4-byte UTF-32 words.
    // There is probably a more efficient sequence, but the following might do.
    const __m128i sh = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
    const __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
    _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output), _mm_cvtepu16_epi32(composed));
    _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output+4), _mm_cvtepu16_epi32(_mm_srli_si128(composed,8)));
    utf32_output += 8; // We wrote 32 bytes, 8 code points.
    return 16;
  }
  if(input_utf8_end_of_code_point_mask == 0x924) {
    // We want to take 4 3-byte UTF-8 words and turn them into 4 4-byte UTF-32 words.
    // There is probably a more efficient sequence, but the following might do.
    const __m128i sh = _mm_setr_epi8(2, 1, 0, -1, 5, 4, 3, -1, 8, 7, 6, -1, 11, 10, 9, -1);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii =
        _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits
    const __m128i middlebyte =
        _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits
    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
    const __m128i highbyte =
        _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits
    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4);
    const __m128i composed =
        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted);
    _mm_storeu_si128((__m128i *)utf32_output, composed);
    utf32_output += 4;
    return 12;
  }
  /// We do not have a fast path available, so we fallback.

  const uint8_t idx =
      tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][0];
  const uint8_t consumed =
      tables::utf8_to_utf16::utf8bigindex[input_utf8_end_of_code_point_mask][1];
  if (idx < 64) {
    // SIX (6) input code-words
    // this is a relatively easy scenario
    // we process SIX (6) input code-words. The max length in bytes of six code
    // words spanning between 1 and 2 bytes each is 12 bytes. On processors
    // where pdep/pext is fast, we might be able to use a small lookup table.
    const __m128i sh =
        _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi16(0x7f));
    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi16(0x1f00));
    const __m128i composed = _mm_or_si128(ascii, _mm_srli_epi16(highbyte, 2));
    _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output), _mm_cvtepu16_epi32(composed));
    _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output+4), _mm_cvtepu16_epi32(_mm_srli_si128(composed,8)));
    utf32_output += 6; // We wrote 12 bytes, 6 code points.
  } else if (idx < 145) {
    // FOUR (4) input code-words
    const __m128i sh =
        _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii =
        _mm_and_si128(perm, _mm_set1_epi32(0x7f)); // 7 or 6 bits
    const __m128i middlebyte =
        _mm_and_si128(perm, _mm_set1_epi32(0x3f00)); // 5 or 6 bits
    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
    const __m128i highbyte =
        _mm_and_si128(perm, _mm_set1_epi32(0x0f0000)); // 4 bits
    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 4);
    const __m128i composed =
        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted), highbyte_shifted);
    _mm_storeu_si128((__m128i *)utf32_output, composed);
    utf32_output += 4;
  } else if (idx < 209) {
    // TWO (2) input code-words
    const __m128i sh =
        _mm_loadu_si128((const __m128i *)tables::utf8_to_utf16::shufutf8[idx]);
    const __m128i perm = _mm_shuffle_epi8(in, sh);
    const __m128i ascii = _mm_and_si128(perm, _mm_set1_epi32(0x7f));
    const __m128i middlebyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f00));
    const __m128i middlebyte_shifted = _mm_srli_epi32(middlebyte, 2);
    __m128i middlehighbyte = _mm_and_si128(perm, _mm_set1_epi32(0x3f0000));
    // correct for spurious high bit
    const __m128i correct =
        _mm_srli_epi32(_mm_and_si128(perm, _mm_set1_epi32(0x400000)), 1);
    middlehighbyte = _mm_xor_si128(correct, middlehighbyte);
    const __m128i middlehighbyte_shifted = _mm_srli_epi32(middlehighbyte, 4);
    const __m128i highbyte = _mm_and_si128(perm, _mm_set1_epi32(0x07000000));
    const __m128i highbyte_shifted = _mm_srli_epi32(highbyte, 6);
    const __m128i composed =
        _mm_or_si128(_mm_or_si128(ascii, middlebyte_shifted),
                     _mm_or_si128(highbyte_shifted, middlehighbyte_shifted));
    _mm_storeu_si128((__m128i *)utf32_output, composed);
    utf32_output += 3;
  } else {
    // here we know that there is an error but we do not handle errors
  }
  return consumed;
}
/* end file src/westmere/sse_convert_utf8_to_utf32.cpp */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=westmere/sse_convert_utf16_to_utf8.cpp
/* begin file src/westmere/sse_convert_utf16_to_utf8.cpp */
/*
    The vectorized algorithm works on single SSE register i.e., it
    loads eight 16-bit words.

    We consider three cases:
    1. an input register contains no surrogates and each value
       is in range 0x0000 .. 0x07ff.
    2. an input register contains no surrogates and values are
       is in range 0x0000 .. 0xffff.
    3. an input register contains surrogates --- i.e. codepoints
       can have 16 or 32 bits.

    Ad 1.

    When values are less than 0x0800, it means that a 16-bit words
    can be converted into: 1) single UTF8 byte (when it's an ASCII
    char) or 2) two UTF8 bytes.

    For this case we do only some shuffle to obtain these 2-byte
    codes and finally compress the whole SSE register with a single
    shuffle.

    We need 256-entry lookup table to get a compression pattern
    and the number of output bytes in the compressed vector register.
    Each entry occupies 17 bytes.

    Ad 2.

    When values fit in 16-bit words, but are above 0x07ff, then
    a single word may produce one, two or three UTF8 bytes.

    We prepare data for all these three cases in two registers.
    The first register contains lower two UTF8 bytes (used in all
    cases), while the second one contains just the third byte for
    the three-UTF8-bytes case.

    Finally these two registers are interleaved forming eight-element
    array of 32-bit values. The array spans two SSE registers.
    The bytes from the registers are compressed using two shuffles.

    We need 256-entry lookup table to get a compression pattern
    and the number of output bytes in the compressed vector register.
    Each entry occupies 17 bytes.


    To summarize:
    - We need two 256-entry tables that have 8704 bytes in total.
*/

/*
  Returns a pair: the first unprocessed byte from buf and utf8_output
  A scalar routing should carry on the conversion of the tail.
*/
template <endianness big_endian>
std::pair<const char16_t*, char*> sse_convert_utf16_to_utf8(const char16_t* buf, size_t len, char* utf8_output) {

  const char16_t* end = buf + len;

  const __m128i v_0000 = _mm_setzero_si128();
  const __m128i v_f800 = _mm_set1_epi16((int16_t)0xf800);
  const __m128i v_d800 = _mm_set1_epi16((int16_t)0xd800);
  const __m128i v_c080 = _mm_set1_epi16((int16_t)0xc080);
  const size_t safety_margin = 12; // to avoid overruns, see issue https://github.com/simdutf/simdutf/issues/92

  while (buf + 16 + safety_margin <= end) {
    __m128i in = _mm_loadu_si128((__m128i*)buf);
    if (big_endian) {
      const __m128i swap = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
      in = _mm_shuffle_epi8(in, swap);
    }
    // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes
    const __m128i v_ff80 = _mm_set1_epi16((int16_t)0xff80);
    if(_mm_testz_si128(in, v_ff80)) { // ASCII fast path!!!!
        __m128i nextin = _mm_loadu_si128((__m128i*)buf+1);
        if (big_endian) {
          const __m128i swap = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
          nextin = _mm_shuffle_epi8(nextin, swap);
        }
        if(!_mm_testz_si128(nextin, v_ff80)) {
          // 1. pack the bytes
          // obviously suboptimal.
          const __m128i utf8_packed = _mm_packus_epi16(in,in);
          // 2. store (16 bytes)
          _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);
          // 3. adjust pointers
          buf += 8;
          utf8_output += 8;
          in = nextin;
        } else {
          // 1. pack the bytes
          // obviously suboptimal.
          const __m128i utf8_packed = _mm_packus_epi16(in,nextin);
          // 2. store (16 bytes)
          _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);
          // 3. adjust pointers
          buf += 16;
          utf8_output += 16;
          continue; // we are done for this round!
        }
    }

    // no bits set above 7th bit
    const __m128i one_byte_bytemask = _mm_cmpeq_epi16(_mm_and_si128(in, v_ff80), v_0000);
    const uint16_t one_byte_bitmask = static_cast<uint16_t>(_mm_movemask_epi8(one_byte_bytemask));

    // no bits set above 11th bit
    const __m128i one_or_two_bytes_bytemask = _mm_cmpeq_epi16(_mm_and_si128(in, v_f800), v_0000);
    const uint16_t one_or_two_bytes_bitmask = static_cast<uint16_t>(_mm_movemask_epi8(one_or_two_bytes_bytemask));

    if (one_or_two_bytes_bitmask == 0xffff) {
          // 1. prepare 2-byte values
          // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
          // expected output   : [110a|aaaa|10bb|bbbb] x 8
          const __m128i v_1f00 = _mm_set1_epi16((int16_t)0x1f00);
          const __m128i v_003f = _mm_set1_epi16((int16_t)0x003f);

          // t0 = [000a|aaaa|bbbb|bb00]
          const __m128i t0 = _mm_slli_epi16(in, 2);
          // t1 = [000a|aaaa|0000|0000]
          const __m128i t1 = _mm_and_si128(t0, v_1f00);
          // t2 = [0000|0000|00bb|bbbb]
          const __m128i t2 = _mm_and_si128(in, v_003f);
          // t3 = [000a|aaaa|00bb|bbbb]
          const __m128i t3 = _mm_or_si128(t1, t2);
          // t4 = [110a|aaaa|10bb|bbbb]
          const __m128i t4 = _mm_or_si128(t3, v_c080);

          // 2. merge ASCII and 2-byte codewords
          const __m128i utf8_unpacked = _mm_blendv_epi8(t4, in, one_byte_bytemask);

          // 3. prepare bitmask for 8-bit lookup
          //    one_byte_bitmask = hhggffeeddccbbaa -- the bits are doubled (h - MSB, a - LSB)
          const uint16_t m0 = one_byte_bitmask & 0x5555;  // m0 = 0h0g0f0e0d0c0b0a
          const uint16_t m1 = static_cast<uint16_t>(m0 >> 7);                    // m1 = 00000000h0g0f0e0
          const uint8_t  m2 = static_cast<uint8_t>((m0 | m1) & 0xff);           // m2 =         hdgcfbea
          // 4. pack the bytes
          const uint8_t* row = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
          const __m128i shuffle = _mm_loadu_si128((__m128i*)(row + 1));
          const __m128i utf8_packed = _mm_shuffle_epi8(utf8_unpacked, shuffle);

          // 5. store bytes
          _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);

          // 6. adjust pointers
          buf += 8;
          utf8_output += row[0];
          continue;

    }

    // 1. Check if there are any surrogate word in the input chunk.
    //    We have also deal with situation when there is a surrogate word
    //    at the end of a chunk.
    const __m128i surrogates_bytemask = _mm_cmpeq_epi16(_mm_and_si128(in, v_f800), v_d800);

    // bitmask = 0x0000 if there are no surrogates
    //         = 0xc000 if the last word is a surrogate
    const uint16_t surrogates_bitmask = static_cast<uint16_t>(_mm_movemask_epi8(surrogates_bytemask));
    // It might seem like checking for surrogates_bitmask == 0xc000 could help. However,
    // it is likely an uncommon occurrence.
    if (surrogates_bitmask == 0x0000) {
      // case: words from register produce either 1, 2 or 3 UTF-8 bytes
        const __m128i dup_even = _mm_setr_epi16(0x0000, 0x0202, 0x0404, 0x0606,
                                                0x0808, 0x0a0a, 0x0c0c, 0x0e0e);

        /* In this branch we handle three cases:
           1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           - single UFT-8 byte
           2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two UTF-8 bytes
           3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - three UTF-8 bytes

          We expand the input word (16-bit) into two words (32-bit), thus
          we have room for four bytes. However, we need five distinct bit
          layouts. Note that the last byte in cases #2 and #3 is the same.

          We precompute byte 1 for case #1 and the common byte for cases #2 & #3
          in register t2.

          We precompute byte 1 for case #3 and -- **conditionally** -- precompute
          either byte 1 for case #2 or byte 2 for case #3. Note that they
          differ by exactly one bit.

          Finally from these two words we build proper UTF-8 sequence, taking
          into account the case (i.e, the number of bytes to write).
        */
        /**
         * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
         * t2 => [0ccc|cccc] [10cc|cccc]
         * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
         */
#define simdutf_vec(x) _mm_set1_epi16(static_cast<uint16_t>(x))
        // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
        const __m128i t0 = _mm_shuffle_epi8(in, dup_even);
        // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
        const __m128i t1 = _mm_and_si128(t0, simdutf_vec(0b0011111101111111));
        // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
        const __m128i t2 = _mm_or_si128 (t1, simdutf_vec(0b1000000000000000));

        // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
        const __m128i s0 = _mm_srli_epi16(in, 4);
        // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
        const __m128i s1 = _mm_and_si128(s0, simdutf_vec(0b0000111111111100));
        // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
        const __m128i s2 = _mm_maddubs_epi16(s1, simdutf_vec(0x0140));
        // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
        const __m128i s3 = _mm_or_si128(s2, simdutf_vec(0b1100000011100000));
        const __m128i m0 = _mm_andnot_si128(one_or_two_bytes_bytemask, simdutf_vec(0b0100000000000000));
        const __m128i s4 = _mm_xor_si128(s3, m0);
#undef simdutf_vec

        // 4. expand words 16-bit => 32-bit
        const __m128i out0 = _mm_unpacklo_epi16(t2, s4);
        const __m128i out1 = _mm_unpackhi_epi16(t2, s4);

        // 5. compress 32-bit words into 1, 2 or 3 bytes -- 2 x shuffle
        const uint16_t mask = (one_byte_bitmask & 0x5555) |
                              (one_or_two_bytes_bitmask & 0xaaaa);
        if(mask == 0) {
          // We only have three-byte words. Use fast path.
          const __m128i shuffle = _mm_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1);
          const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle);
          const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle);
          _mm_storeu_si128((__m128i*)utf8_output, utf8_0);
          utf8_output += 12;
          _mm_storeu_si128((__m128i*)utf8_output, utf8_1);
          utf8_output += 12;
          buf += 8;
          continue;
        }
        const uint8_t mask0 = uint8_t(mask);

        const uint8_t* row0 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
        const __m128i shuffle0 = _mm_loadu_si128((__m128i*)(row0 + 1));
        const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle0);

        const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);

        const uint8_t* row1 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
        const __m128i shuffle1 = _mm_loadu_si128((__m128i*)(row1 + 1));
        const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle1);

        _mm_storeu_si128((__m128i*)utf8_output, utf8_0);
        utf8_output += row0[0];
        _mm_storeu_si128((__m128i*)utf8_output, utf8_1);
        utf8_output += row1[0];

        buf += 8;
    // surrogate pair(s) in a register
    } else {
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // in the presence of surrogate pairs may require non-trivial tables.
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint16_t word = big_endian ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
        if((word & 0xFF80)==0) {
          *utf8_output++ = char(word);
        } else if((word & 0xF800)==0) {
          *utf8_output++ = char((word>>6) | 0b11000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else if((word &0xF800 ) != 0xD800) {
          *utf8_output++ = char((word>>12) | 0b11100000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else {
          // must be a surrogate pair
          uint16_t diff = uint16_t(word - 0xD800);
          uint16_t next_word = big_endian ? scalar::utf16::swap_bytes(buf[k+1]) : buf[k+1];
          k++;
          uint16_t diff2 = uint16_t(next_word - 0xDC00);
          if((diff | diff2) > 0x3FF)  { return std::make_pair(nullptr, utf8_output); }
          uint32_t value = (diff << 10) + diff2 + 0x10000;
          *utf8_output++ = char((value>>18) | 0b11110000);
          *utf8_output++ = char(((value>>12) & 0b111111) | 0b10000000);
          *utf8_output++ = char(((value>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((value & 0b111111) | 0b10000000);
        }
      }
      buf += k;
    }
  } // while

  return std::make_pair(buf, utf8_output);
}


/*
  Returns a pair: a result struct and utf8_output.
  If there is an error, the count field of the result is the position of the error.
  Otherwise, it is the position of the first unprocessed byte in buf (even if finished).
  A scalar routing should carry on the conversion of the tail if needed.
*/
template <endianness big_endian>
std::pair<result, char*> sse_convert_utf16_to_utf8_with_errors(const char16_t* buf, size_t len, char* utf8_output) {
  const char16_t* start = buf;
  const char16_t* end = buf + len;

  const __m128i v_0000 = _mm_setzero_si128();
  const __m128i v_f800 = _mm_set1_epi16((int16_t)0xf800);
  const __m128i v_d800 = _mm_set1_epi16((int16_t)0xd800);
  const __m128i v_c080 = _mm_set1_epi16((int16_t)0xc080);
  const size_t safety_margin = 12; // to avoid overruns, see issue https://github.com/simdutf/simdutf/issues/92

  while (buf + 16 + safety_margin <= end) {
    __m128i in = _mm_loadu_si128((__m128i*)buf);
    if (big_endian) {
      const __m128i swap = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
      in = _mm_shuffle_epi8(in, swap);
    }
    // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes
    const __m128i v_ff80 = _mm_set1_epi16((int16_t)0xff80);
    if(_mm_testz_si128(in, v_ff80)) { // ASCII fast path!!!!
        __m128i nextin = _mm_loadu_si128((__m128i*)buf+1);
        if (big_endian) {
          const __m128i swap = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
          nextin = _mm_shuffle_epi8(nextin, swap);
        }
        if(!_mm_testz_si128(nextin, v_ff80)) {
          // 1. pack the bytes
          // obviously suboptimal.
          const __m128i utf8_packed = _mm_packus_epi16(in,in);
          // 2. store (16 bytes)
          _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);
          // 3. adjust pointers
          buf += 8;
          utf8_output += 8;
          in = nextin;
        } else {
          // 1. pack the bytes
          // obviously suboptimal.
          const __m128i utf8_packed = _mm_packus_epi16(in,nextin);
          // 2. store (16 bytes)
          _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);
          // 3. adjust pointers
          buf += 16;
          utf8_output += 16;
          continue; // we are done for this round!
        }
    }

    // no bits set above 7th bit
    const __m128i one_byte_bytemask = _mm_cmpeq_epi16(_mm_and_si128(in, v_ff80), v_0000);
    const uint16_t one_byte_bitmask = static_cast<uint16_t>(_mm_movemask_epi8(one_byte_bytemask));

    // no bits set above 11th bit
    const __m128i one_or_two_bytes_bytemask = _mm_cmpeq_epi16(_mm_and_si128(in, v_f800), v_0000);
    const uint16_t one_or_two_bytes_bitmask = static_cast<uint16_t>(_mm_movemask_epi8(one_or_two_bytes_bytemask));

    if (one_or_two_bytes_bitmask == 0xffff) {
          // 1. prepare 2-byte values
          // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
          // expected output   : [110a|aaaa|10bb|bbbb] x 8
          const __m128i v_1f00 = _mm_set1_epi16((int16_t)0x1f00);
          const __m128i v_003f = _mm_set1_epi16((int16_t)0x003f);

          // t0 = [000a|aaaa|bbbb|bb00]
          const __m128i t0 = _mm_slli_epi16(in, 2);
          // t1 = [000a|aaaa|0000|0000]
          const __m128i t1 = _mm_and_si128(t0, v_1f00);
          // t2 = [0000|0000|00bb|bbbb]
          const __m128i t2 = _mm_and_si128(in, v_003f);
          // t3 = [000a|aaaa|00bb|bbbb]
          const __m128i t3 = _mm_or_si128(t1, t2);
          // t4 = [110a|aaaa|10bb|bbbb]
          const __m128i t4 = _mm_or_si128(t3, v_c080);

          // 2. merge ASCII and 2-byte codewords
          const __m128i utf8_unpacked = _mm_blendv_epi8(t4, in, one_byte_bytemask);

          // 3. prepare bitmask for 8-bit lookup
          //    one_byte_bitmask = hhggffeeddccbbaa -- the bits are doubled (h - MSB, a - LSB)
          const uint16_t m0 = one_byte_bitmask & 0x5555;  // m0 = 0h0g0f0e0d0c0b0a
          const uint16_t m1 = static_cast<uint16_t>(m0 >> 7);                    // m1 = 00000000h0g0f0e0
          const uint8_t  m2 = static_cast<uint8_t>((m0 | m1) & 0xff);           // m2 =         hdgcfbea
          // 4. pack the bytes
          const uint8_t* row = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
          const __m128i shuffle = _mm_loadu_si128((__m128i*)(row + 1));
          const __m128i utf8_packed = _mm_shuffle_epi8(utf8_unpacked, shuffle);

          // 5. store bytes
          _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);

          // 6. adjust pointers
          buf += 8;
          utf8_output += row[0];
          continue;

    }

    // 1. Check if there are any surrogate word in the input chunk.
    //    We have also deal with situation when there is a surrogate word
    //    at the end of a chunk.
    const __m128i surrogates_bytemask = _mm_cmpeq_epi16(_mm_and_si128(in, v_f800), v_d800);

    // bitmask = 0x0000 if there are no surrogates
    //         = 0xc000 if the last word is a surrogate
    const uint16_t surrogates_bitmask = static_cast<uint16_t>(_mm_movemask_epi8(surrogates_bytemask));
    // It might seem like checking for surrogates_bitmask == 0xc000 could help. However,
    // it is likely an uncommon occurrence.
    if (surrogates_bitmask == 0x0000) {
      // case: words from register produce either 1, 2 or 3 UTF-8 bytes
        const __m128i dup_even = _mm_setr_epi16(0x0000, 0x0202, 0x0404, 0x0606,
                                                0x0808, 0x0a0a, 0x0c0c, 0x0e0e);

        /* In this branch we handle three cases:
           1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           - single UFT-8 byte
           2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two UTF-8 bytes
           3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - three UTF-8 bytes

          We expand the input word (16-bit) into two words (32-bit), thus
          we have room for four bytes. However, we need five distinct bit
          layouts. Note that the last byte in cases #2 and #3 is the same.

          We precompute byte 1 for case #1 and the common byte for cases #2 & #3
          in register t2.

          We precompute byte 1 for case #3 and -- **conditionally** -- precompute
          either byte 1 for case #2 or byte 2 for case #3. Note that they
          differ by exactly one bit.

          Finally from these two words we build proper UTF-8 sequence, taking
          into account the case (i.e, the number of bytes to write).
        */
        /**
         * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
         * t2 => [0ccc|cccc] [10cc|cccc]
         * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
         */
#define simdutf_vec(x) _mm_set1_epi16(static_cast<uint16_t>(x))
        // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
        const __m128i t0 = _mm_shuffle_epi8(in, dup_even);
        // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
        const __m128i t1 = _mm_and_si128(t0, simdutf_vec(0b0011111101111111));
        // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
        const __m128i t2 = _mm_or_si128 (t1, simdutf_vec(0b1000000000000000));

        // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
        const __m128i s0 = _mm_srli_epi16(in, 4);
        // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
        const __m128i s1 = _mm_and_si128(s0, simdutf_vec(0b0000111111111100));
        // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
        const __m128i s2 = _mm_maddubs_epi16(s1, simdutf_vec(0x0140));
        // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
        const __m128i s3 = _mm_or_si128(s2, simdutf_vec(0b1100000011100000));
        const __m128i m0 = _mm_andnot_si128(one_or_two_bytes_bytemask, simdutf_vec(0b0100000000000000));
        const __m128i s4 = _mm_xor_si128(s3, m0);
#undef simdutf_vec

        // 4. expand words 16-bit => 32-bit
        const __m128i out0 = _mm_unpacklo_epi16(t2, s4);
        const __m128i out1 = _mm_unpackhi_epi16(t2, s4);

        // 5. compress 32-bit words into 1, 2 or 3 bytes -- 2 x shuffle
        const uint16_t mask = (one_byte_bitmask & 0x5555) |
                              (one_or_two_bytes_bitmask & 0xaaaa);
        if(mask == 0) {
          // We only have three-byte words. Use fast path.
          const __m128i shuffle = _mm_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1);
          const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle);
          const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle);
          _mm_storeu_si128((__m128i*)utf8_output, utf8_0);
          utf8_output += 12;
          _mm_storeu_si128((__m128i*)utf8_output, utf8_1);
          utf8_output += 12;
          buf += 8;
          continue;
        }
        const uint8_t mask0 = uint8_t(mask);

        const uint8_t* row0 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
        const __m128i shuffle0 = _mm_loadu_si128((__m128i*)(row0 + 1));
        const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle0);

        const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);

        const uint8_t* row1 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
        const __m128i shuffle1 = _mm_loadu_si128((__m128i*)(row1 + 1));
        const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle1);

        _mm_storeu_si128((__m128i*)utf8_output, utf8_0);
        utf8_output += row0[0];
        _mm_storeu_si128((__m128i*)utf8_output, utf8_1);
        utf8_output += row1[0];

        buf += 8;
    // surrogate pair(s) in a register
    } else {
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // in the presence of surrogate pairs may require non-trivial tables.
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint16_t word = big_endian ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
        if((word & 0xFF80)==0) {
          *utf8_output++ = char(word);
        } else if((word & 0xF800)==0) {
          *utf8_output++ = char((word>>6) | 0b11000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else if((word &0xF800 ) != 0xD800) {
          *utf8_output++ = char((word>>12) | 0b11100000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else {
          // must be a surrogate pair
          uint16_t diff = uint16_t(word - 0xD800);
          uint16_t next_word = big_endian ? scalar::utf16::swap_bytes(buf[k+1]) : buf[k+1];
          k++;
          uint16_t diff2 = uint16_t(next_word - 0xDC00);
          if((diff | diff2) > 0x3FF)  { return std::make_pair(result(error_code::SURROGATE, buf - start + k - 1), utf8_output); }
          uint32_t value = (diff << 10) + diff2 + 0x10000;
          *utf8_output++ = char((value>>18) | 0b11110000);
          *utf8_output++ = char(((value>>12) & 0b111111) | 0b10000000);
          *utf8_output++ = char(((value>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((value & 0b111111) | 0b10000000);
        }
      }
      buf += k;
    }
  } // while

  return std::make_pair(result(error_code::SUCCESS, buf - start), utf8_output);
}
/* end file src/westmere/sse_convert_utf16_to_utf8.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=westmere/sse_convert_utf16_to_utf32.cpp
/* begin file src/westmere/sse_convert_utf16_to_utf32.cpp */
/*
    The vectorized algorithm works on single SSE register i.e., it
    loads eight 16-bit words.

    We consider three cases:
    1. an input register contains no surrogates and each value
       is in range 0x0000 .. 0x07ff.
    2. an input register contains no surrogates and values are
       is in range 0x0000 .. 0xffff.
    3. an input register contains surrogates --- i.e. codepoints
       can have 16 or 32 bits.

    Ad 1.

    When values are less than 0x0800, it means that a 16-bit words
    can be converted into: 1) single UTF8 byte (when it's an ASCII
    char) or 2) two UTF8 bytes.

    For this case we do only some shuffle to obtain these 2-byte
    codes and finally compress the whole SSE register with a single
    shuffle.

    We need 256-entry lookup table to get a compression pattern
    and the number of output bytes in the compressed vector register.
    Each entry occupies 17 bytes.

    Ad 2.

    When values fit in 16-bit words, but are above 0x07ff, then
    a single word may produce one, two or three UTF8 bytes.

    We prepare data for all these three cases in two registers.
    The first register contains lower two UTF8 bytes (used in all
    cases), while the second one contains just the third byte for
    the three-UTF8-bytes case.

    Finally these two registers are interleaved forming eight-element
    array of 32-bit values. The array spans two SSE registers.
    The bytes from the registers are compressed using two shuffles.

    We need 256-entry lookup table to get a compression pattern
    and the number of output bytes in the compressed vector register.
    Each entry occupies 17 bytes.


    To summarize:
    - We need two 256-entry tables that have 8704 bytes in total.
*/

/*
  Returns a pair: the first unprocessed byte from buf and utf8_output
  A scalar routing should carry on the conversion of the tail.
*/
template <endianness big_endian>
std::pair<const char16_t*, char32_t*> sse_convert_utf16_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) {
  const char16_t* end = buf + len;

  const __m128i v_f800 = _mm_set1_epi16((int16_t)0xf800);
  const __m128i v_d800 = _mm_set1_epi16((int16_t)0xd800);

  while (buf + 16 <= end) {
    __m128i in = _mm_loadu_si128((__m128i*)buf);

    if (big_endian) {
      const __m128i swap = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
      in = _mm_shuffle_epi8(in, swap);
    }

    // 1. Check if there are any surrogate word in the input chunk.
    //    We have also deal with situation when there is a surrogate word
    //    at the end of a chunk.
    const __m128i surrogates_bytemask = _mm_cmpeq_epi16(_mm_and_si128(in, v_f800), v_d800);

    // bitmask = 0x0000 if there are no surrogates
    //         = 0xc000 if the last word is a surrogate
    const uint16_t surrogates_bitmask = static_cast<uint16_t>(_mm_movemask_epi8(surrogates_bytemask));
    // It might seem like checking for surrogates_bitmask == 0xc000 could help. However,
    // it is likely an uncommon occurrence.
    if (surrogates_bitmask == 0x0000) {
      // case: no surrogate pair, extend 16-bit words to 32-bit words
        _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output), _mm_cvtepu16_epi32(in));
        _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output+4), _mm_cvtepu16_epi32(_mm_srli_si128(in,8)));
        utf32_output += 8;
        buf += 8;
    // surrogate pair(s) in a register
    } else {
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // in the presence of surrogate pairs may require non-trivial tables.
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint16_t word = big_endian ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
        if((word &0xF800 ) != 0xD800) {
          *utf32_output++ = char32_t(word);
        } else {
          // must be a surrogate pair
          uint16_t diff = uint16_t(word - 0xD800);
          uint16_t next_word = big_endian ? scalar::utf16::swap_bytes(buf[k+1]) : buf[k+1];
          k++;
          uint16_t diff2 = uint16_t(next_word - 0xDC00);
          if((diff | diff2) > 0x3FF)  { return std::make_pair(nullptr, utf32_output); }
          uint32_t value = (diff << 10) + diff2 + 0x10000;
          *utf32_output++ = char32_t(value);
        }
      }
      buf += k;
    }
  } // while
  return std::make_pair(buf, utf32_output);
}


/*
  Returns a pair: a result struct and utf8_output.
  If there is an error, the count field of the result is the position of the error.
  Otherwise, it is the position of the first unprocessed byte in buf (even if finished).
  A scalar routing should carry on the conversion of the tail if needed.
*/
template <endianness big_endian>
std::pair<result, char32_t*> sse_convert_utf16_to_utf32_with_errors(const char16_t* buf, size_t len, char32_t* utf32_output) {
  const char16_t* start = buf;
  const char16_t* end = buf + len;

  const __m128i v_f800 = _mm_set1_epi16((int16_t)0xf800);
  const __m128i v_d800 = _mm_set1_epi16((int16_t)0xd800);

  while (buf + 16 <= end) {
    __m128i in = _mm_loadu_si128((__m128i*)buf);

    if (big_endian) {
      const __m128i swap = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
      in = _mm_shuffle_epi8(in, swap);
    }

    // 1. Check if there are any surrogate word in the input chunk.
    //    We have also deal with situation when there is a surrogate word
    //    at the end of a chunk.
    const __m128i surrogates_bytemask = _mm_cmpeq_epi16(_mm_and_si128(in, v_f800), v_d800);

    // bitmask = 0x0000 if there are no surrogates
    //         = 0xc000 if the last word is a surrogate
    const uint16_t surrogates_bitmask = static_cast<uint16_t>(_mm_movemask_epi8(surrogates_bytemask));
    // It might seem like checking for surrogates_bitmask == 0xc000 could help. However,
    // it is likely an uncommon occurrence.
    if (surrogates_bitmask == 0x0000) {
      // case: no surrogate pair, extend 16-bit words to 32-bit words
        _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output), _mm_cvtepu16_epi32(in));
        _mm_storeu_si128(reinterpret_cast<__m128i *>(utf32_output+4), _mm_cvtepu16_epi32(_mm_srli_si128(in,8)));
        utf32_output += 8;
        buf += 8;
    // surrogate pair(s) in a register
    } else {
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // in the presence of surrogate pairs may require non-trivial tables.
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint16_t word = big_endian ? scalar::utf16::swap_bytes(buf[k]) : buf[k];
        if((word &0xF800 ) != 0xD800) {
          *utf32_output++ = char32_t(word);
        } else {
          // must be a surrogate pair
          uint16_t diff = uint16_t(word - 0xD800);
          uint16_t next_word = big_endian ? scalar::utf16::swap_bytes(buf[k+1]) : buf[k+1];
          k++;
          uint16_t diff2 = uint16_t(next_word - 0xDC00);
          if((diff | diff2) > 0x3FF)  { return std::make_pair(result(error_code::SURROGATE, buf - start + k - 1), utf32_output); }
          uint32_t value = (diff << 10) + diff2 + 0x10000;
          *utf32_output++ = char32_t(value);
        }
      }
      buf += k;
    }
  } // while
  return std::make_pair(result(error_code::SUCCESS, buf - start), utf32_output);
}
/* end file src/westmere/sse_convert_utf16_to_utf32.cpp */

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=westmere/sse_convert_utf32_to_utf8.cpp
/* begin file src/westmere/sse_convert_utf32_to_utf8.cpp */
std::pair<const char32_t*, char*> sse_convert_utf32_to_utf8(const char32_t* buf, size_t len, char* utf8_output) {
  const char32_t* end = buf + len;

  const __m128i v_0000 = _mm_setzero_si128();
  const __m128i v_f800 = _mm_set1_epi16((uint16_t)0xf800);
  const __m128i v_c080 = _mm_set1_epi16((uint16_t)0xc080);
  const __m128i v_ff80 = _mm_set1_epi16((uint16_t)0xff80);
  const __m128i v_ffff0000 = _mm_set1_epi32((uint32_t)0xffff0000);
  const __m128i v_7fffffff = _mm_set1_epi32((uint32_t)0x7fffffff);
  __m128i running_max = _mm_setzero_si128();
  __m128i forbidden_bytemask = _mm_setzero_si128();
  const size_t safety_margin = 12; // to avoid overruns, see issue https://github.com/simdutf/simdutf/issues/92

  while (buf + 16 + safety_margin <= end) {
    // We load two 16 bytes registers for a total of 32 bytes or 16 characters.
    __m128i in = _mm_loadu_si128((__m128i*)buf);
    __m128i nextin = _mm_loadu_si128((__m128i*)buf+1);
    running_max = _mm_max_epu32(_mm_max_epu32(in, running_max), nextin);

    // Pack 32-bit UTF-32 words to 16-bit UTF-16 words with unsigned saturation
    __m128i in_16 = _mm_packus_epi32(_mm_and_si128(in, v_7fffffff), _mm_and_si128(nextin, v_7fffffff));

    // Try to apply UTF-16 => UTF-8 from ./sse_convert_utf16_to_utf8.cpp

    // Check for ASCII fast path
    if(_mm_testz_si128(in_16, v_ff80)) { // ASCII fast path!!!!
      // We eagerly load another 32 bytes, hoping that they will be ASCII too.
      // The intuition is that we try to collect 16 ASCII characters which requires
      // a total of 64 bytes of input. If we fail, we just pass thirdin and fourthin
      // as our new inputs.
      __m128i thirdin = _mm_loadu_si128((__m128i*)buf+2);
      __m128i fourthin = _mm_loadu_si128((__m128i*)buf+3);
      running_max = _mm_max_epu32(_mm_max_epu32(thirdin, running_max), fourthin);
      __m128i nextin_16 = _mm_packus_epi32(_mm_and_si128(thirdin, v_7fffffff), _mm_and_si128(fourthin, v_7fffffff));
      if(!_mm_testz_si128(nextin_16, v_ff80)) {
        // 1. pack the bytes
        // obviously suboptimal.
        const __m128i utf8_packed = _mm_packus_epi16(in_16,in_16);
        // 2. store (16 bytes)
        _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);
        // 3. adjust pointers
        buf += 8;
        utf8_output += 8;
        // Proceed with next input
        in_16 = nextin_16;
        // We need to update in and nextin because they are used later.
        in = thirdin;
        nextin = fourthin;
      } else {
        // 1. pack the bytes
        const __m128i utf8_packed = _mm_packus_epi16(in_16, nextin_16);
        // 2. store (16 bytes)
        _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);
        // 3. adjust pointers
        buf += 16;
        utf8_output += 16;
        continue; // we are done for this round!
      }
    }

    // no bits set above 7th bit
    const __m128i one_byte_bytemask = _mm_cmpeq_epi16(_mm_and_si128(in_16, v_ff80), v_0000);
    const uint16_t one_byte_bitmask = static_cast<uint16_t>(_mm_movemask_epi8(one_byte_bytemask));

    // no bits set above 11th bit
    const __m128i one_or_two_bytes_bytemask = _mm_cmpeq_epi16(_mm_and_si128(in_16, v_f800), v_0000);
    const uint16_t one_or_two_bytes_bitmask = static_cast<uint16_t>(_mm_movemask_epi8(one_or_two_bytes_bytemask));

    if (one_or_two_bytes_bitmask == 0xffff) {
      // case: all words either produce 1 or 2 UTF-8 bytes (at least one produces 2 bytes)
      // 1. prepare 2-byte values
      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
      // expected output   : [110a|aaaa|10bb|bbbb] x 8
      const __m128i v_1f00 = _mm_set1_epi16((int16_t)0x1f00);
      const __m128i v_003f = _mm_set1_epi16((int16_t)0x003f);

      // t0 = [000a|aaaa|bbbb|bb00]
      const __m128i t0 = _mm_slli_epi16(in_16, 2);
      // t1 = [000a|aaaa|0000|0000]
      const __m128i t1 = _mm_and_si128(t0, v_1f00);
      // t2 = [0000|0000|00bb|bbbb]
      const __m128i t2 = _mm_and_si128(in_16, v_003f);
      // t3 = [000a|aaaa|00bb|bbbb]
      const __m128i t3 = _mm_or_si128(t1, t2);
      // t4 = [110a|aaaa|10bb|bbbb]
      const __m128i t4 = _mm_or_si128(t3, v_c080);

      // 2. merge ASCII and 2-byte codewords
      const __m128i utf8_unpacked = _mm_blendv_epi8(t4, in_16, one_byte_bytemask);

      // 3. prepare bitmask for 8-bit lookup
      //    one_byte_bitmask = hhggffeeddccbbaa -- the bits are doubled (h - MSB, a - LSB)
      const uint16_t m0 = one_byte_bitmask & 0x5555;  // m0 = 0h0g0f0e0d0c0b0a
      const uint16_t m1 = static_cast<uint16_t>(m0 >> 7);                    // m1 = 00000000h0g0f0e0
      const uint8_t  m2 = static_cast<uint8_t>((m0 | m1) & 0xff);           // m2 =         hdgcfbea
      // 4. pack the bytes
      const uint8_t* row = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
      const __m128i shuffle = _mm_loadu_si128((__m128i*)(row + 1));
      const __m128i utf8_packed = _mm_shuffle_epi8(utf8_unpacked, shuffle);

      // 5. store bytes
      _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);

      // 6. adjust pointers
      buf += 8;
      utf8_output += row[0];
      continue;
    }

    // Check for overflow in packing

    const __m128i saturation_bytemask = _mm_cmpeq_epi32(_mm_and_si128(_mm_or_si128(in, nextin), v_ffff0000), v_0000);
    const uint32_t saturation_bitmask = static_cast<uint32_t>(_mm_movemask_epi8(saturation_bytemask));
    if (saturation_bitmask == 0xffff) {
      // case: words from register produce either 1, 2 or 3 UTF-8 bytes
      const __m128i v_d800 = _mm_set1_epi16((uint16_t)0xd800);
      forbidden_bytemask = _mm_or_si128(forbidden_bytemask, _mm_cmpeq_epi16(_mm_and_si128(in_16, v_f800), v_d800));

      const __m128i dup_even = _mm_setr_epi16(0x0000, 0x0202, 0x0404, 0x0606,
                                              0x0808, 0x0a0a, 0x0c0c, 0x0e0e);

      /* In this branch we handle three cases:
          1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           - single UFT-8 byte
          2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two UTF-8 bytes
          3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - three UTF-8 bytes

        We expand the input word (16-bit) into two words (32-bit), thus
        we have room for four bytes. However, we need five distinct bit
        layouts. Note that the last byte in cases #2 and #3 is the same.

        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
        in register t2.

        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
        either byte 1 for case #2 or byte 2 for case #3. Note that they
        differ by exactly one bit.

        Finally from these two words we build proper UTF-8 sequence, taking
        into account the case (i.e, the number of bytes to write).
      */
      /**
       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
       * t2 => [0ccc|cccc] [10cc|cccc]
       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
       */
#define simdutf_vec(x) _mm_set1_epi16(static_cast<uint16_t>(x))
      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
      const __m128i t0 = _mm_shuffle_epi8(in_16, dup_even);
      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
      const __m128i t1 = _mm_and_si128(t0, simdutf_vec(0b0011111101111111));
      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
      const __m128i t2 = _mm_or_si128 (t1, simdutf_vec(0b1000000000000000));

      // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
      const __m128i s0 = _mm_srli_epi16(in_16, 4);
      // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
      const __m128i s1 = _mm_and_si128(s0, simdutf_vec(0b0000111111111100));
      // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
      const __m128i s2 = _mm_maddubs_epi16(s1, simdutf_vec(0x0140));
      // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
      const __m128i s3 = _mm_or_si128(s2, simdutf_vec(0b1100000011100000));
      const __m128i m0 = _mm_andnot_si128(one_or_two_bytes_bytemask, simdutf_vec(0b0100000000000000));
      const __m128i s4 = _mm_xor_si128(s3, m0);
#undef simdutf_vec

      // 4. expand words 16-bit => 32-bit
      const __m128i out0 = _mm_unpacklo_epi16(t2, s4);
      const __m128i out1 = _mm_unpackhi_epi16(t2, s4);

      // 5. compress 32-bit words into 1, 2 or 3 bytes -- 2 x shuffle
      const uint16_t mask = (one_byte_bitmask & 0x5555) |
                            (one_or_two_bytes_bitmask & 0xaaaa);
      if(mask == 0) {
        // We only have three-byte words. Use fast path.
        const __m128i shuffle = _mm_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1);
        const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle);
        const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle);
        _mm_storeu_si128((__m128i*)utf8_output, utf8_0);
        utf8_output += 12;
        _mm_storeu_si128((__m128i*)utf8_output, utf8_1);
        utf8_output += 12;
        buf += 8;
        continue;
      }
      const uint8_t mask0 = uint8_t(mask);

      const uint8_t* row0 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
      const __m128i shuffle0 = _mm_loadu_si128((__m128i*)(row0 + 1));
      const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle0);

      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);

      const uint8_t* row1 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
      const __m128i shuffle1 = _mm_loadu_si128((__m128i*)(row1 + 1));
      const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle1);

      _mm_storeu_si128((__m128i*)utf8_output, utf8_0);
      utf8_output += row0[0];
      _mm_storeu_si128((__m128i*)utf8_output, utf8_1);
      utf8_output += row1[0];

      buf += 8;
    } else {
      // case: at least one 32-bit word produce a surrogate pair in UTF-16 <=> will produce four UTF-8 bytes
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // in the presence of surrogate pairs may require non-trivial tables.
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint32_t word = buf[k];
        if((word & 0xFFFFFF80)==0) {
          *utf8_output++ = char(word);
        } else if((word & 0xFFFFF800)==0) {
          *utf8_output++ = char((word>>6) | 0b11000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else if((word &0xFFFF0000 )==0) {
          if (word >= 0xD800 && word <= 0xDFFF) { return std::make_pair(nullptr, utf8_output); }
          *utf8_output++ = char((word>>12) | 0b11100000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else {
          if (word > 0x10FFFF) { return std::make_pair(nullptr, utf8_output); }
          *utf8_output++ = char((word>>18) | 0b11110000);
          *utf8_output++ = char(((word>>12) & 0b111111) | 0b10000000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        }
      }
      buf += k;
    }
  } // while

  // check for invalid input
  const __m128i v_10ffff = _mm_set1_epi32((uint32_t)0x10ffff);
  if(static_cast<uint16_t>(_mm_movemask_epi8(_mm_cmpeq_epi32(_mm_max_epu32(running_max, v_10ffff), v_10ffff))) != 0xffff) {
    return std::make_pair(nullptr, utf8_output);
  }

  if (static_cast<uint32_t>(_mm_movemask_epi8(forbidden_bytemask)) != 0) { return std::make_pair(nullptr, utf8_output); }

  return std::make_pair(buf, utf8_output);
}


std::pair<result, char*> sse_convert_utf32_to_utf8_with_errors(const char32_t* buf, size_t len, char* utf8_output) {

  const char32_t* end = buf + len;
  const char32_t* start = buf;

  const __m128i v_0000 = _mm_setzero_si128();
  const __m128i v_f800 = _mm_set1_epi16((uint16_t)0xf800);
  const __m128i v_c080 = _mm_set1_epi16((uint16_t)0xc080);
  const __m128i v_ff80 = _mm_set1_epi16((uint16_t)0xff80);
  const __m128i v_ffff0000 = _mm_set1_epi32((uint32_t)0xffff0000);
  const __m128i v_7fffffff = _mm_set1_epi32((uint32_t)0x7fffffff);
  const __m128i v_10ffff = _mm_set1_epi32((uint32_t)0x10ffff);

  const size_t safety_margin = 12; // to avoid overruns, see issue https://github.com/simdutf/simdutf/issues/92

  while (buf + 16 + safety_margin <= end) {
    // We load two 16 bytes registers for a total of 32 bytes or 16 characters.
    __m128i in = _mm_loadu_si128((__m128i*)buf);
    __m128i nextin = _mm_loadu_si128((__m128i*)buf+1);

    // Check for too large input
    __m128i max_input = _mm_max_epu32(_mm_max_epu32(in, nextin), v_10ffff);
    if(static_cast<uint16_t>(_mm_movemask_epi8(_mm_cmpeq_epi32(max_input, v_10ffff))) != 0xffff) {
      return std::make_pair(result(error_code::TOO_LARGE, buf - start), utf8_output);
    }

    // Pack 32-bit UTF-32 words to 16-bit UTF-16 words with unsigned saturation
    __m128i in_16 = _mm_packus_epi32(_mm_and_si128(in, v_7fffffff), _mm_and_si128(nextin, v_7fffffff));

    // Try to apply UTF-16 => UTF-8 from ./sse_convert_utf16_to_utf8.cpp

    // Check for ASCII fast path
    if(_mm_testz_si128(in_16, v_ff80)) { // ASCII fast path!!!!
      // We eagerly load another 32 bytes, hoping that they will be ASCII too.
      // The intuition is that we try to collect 16 ASCII characters which requires
      // a total of 64 bytes of input. If we fail, we just pass thirdin and fourthin
      // as our new inputs.
      __m128i thirdin = _mm_loadu_si128((__m128i*)buf+2);
      __m128i fourthin = _mm_loadu_si128((__m128i*)buf+3);
      __m128i nextin_16 = _mm_packus_epi32(_mm_and_si128(thirdin, v_7fffffff), _mm_and_si128(fourthin, v_7fffffff));
      if(!_mm_testz_si128(nextin_16, v_ff80)) {
        // 1. pack the bytes
        // obviously suboptimal.
        const __m128i utf8_packed = _mm_packus_epi16(in_16,in_16);
        // 2. store (16 bytes)
        _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);
        // 3. adjust pointers
        buf += 8;
        utf8_output += 8;
        // Proceed with next input
        in_16 = nextin_16;
        __m128i next_max_input = _mm_max_epu32(_mm_max_epu32(thirdin, fourthin), v_10ffff);
        if(static_cast<uint16_t>(_mm_movemask_epi8(_mm_cmpeq_epi32(next_max_input, v_10ffff))) != 0xffff) {
          return std::make_pair(result(error_code::TOO_LARGE, buf - start), utf8_output);
        }
        // We need to update in and nextin because they are used later.
        in = thirdin;
        nextin = fourthin;
      } else {
        // 1. pack the bytes
        const __m128i utf8_packed = _mm_packus_epi16(in_16, nextin_16);
        // 2. store (16 bytes)
        _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);
        // 3. adjust pointers
        buf += 16;
        utf8_output += 16;
        continue; // we are done for this round!
      }
    }

    // no bits set above 7th bit
    const __m128i one_byte_bytemask = _mm_cmpeq_epi16(_mm_and_si128(in_16, v_ff80), v_0000);
    const uint16_t one_byte_bitmask = static_cast<uint16_t>(_mm_movemask_epi8(one_byte_bytemask));

    // no bits set above 11th bit
    const __m128i one_or_two_bytes_bytemask = _mm_cmpeq_epi16(_mm_and_si128(in_16, v_f800), v_0000);
    const uint16_t one_or_two_bytes_bitmask = static_cast<uint16_t>(_mm_movemask_epi8(one_or_two_bytes_bytemask));

    if (one_or_two_bytes_bitmask == 0xffff) {
      // case: all words either produce 1 or 2 UTF-8 bytes (at least one produces 2 bytes)
      // 1. prepare 2-byte values
      // input 16-bit word : [0000|0aaa|aabb|bbbb] x 8
      // expected output   : [110a|aaaa|10bb|bbbb] x 8
      const __m128i v_1f00 = _mm_set1_epi16((int16_t)0x1f00);
      const __m128i v_003f = _mm_set1_epi16((int16_t)0x003f);

      // t0 = [000a|aaaa|bbbb|bb00]
      const __m128i t0 = _mm_slli_epi16(in_16, 2);
      // t1 = [000a|aaaa|0000|0000]
      const __m128i t1 = _mm_and_si128(t0, v_1f00);
      // t2 = [0000|0000|00bb|bbbb]
      const __m128i t2 = _mm_and_si128(in_16, v_003f);
      // t3 = [000a|aaaa|00bb|bbbb]
      const __m128i t3 = _mm_or_si128(t1, t2);
      // t4 = [110a|aaaa|10bb|bbbb]
      const __m128i t4 = _mm_or_si128(t3, v_c080);

      // 2. merge ASCII and 2-byte codewords
      const __m128i utf8_unpacked = _mm_blendv_epi8(t4, in_16, one_byte_bytemask);

      // 3. prepare bitmask for 8-bit lookup
      //    one_byte_bitmask = hhggffeeddccbbaa -- the bits are doubled (h - MSB, a - LSB)
      const uint16_t m0 = one_byte_bitmask & 0x5555;  // m0 = 0h0g0f0e0d0c0b0a
      const uint16_t m1 = static_cast<uint16_t>(m0 >> 7);                    // m1 = 00000000h0g0f0e0
      const uint8_t  m2 = static_cast<uint8_t>((m0 | m1) & 0xff);           // m2 =         hdgcfbea
      // 4. pack the bytes
      const uint8_t* row = &simdutf::tables::utf16_to_utf8::pack_1_2_utf8_bytes[m2][0];
      const __m128i shuffle = _mm_loadu_si128((__m128i*)(row + 1));
      const __m128i utf8_packed = _mm_shuffle_epi8(utf8_unpacked, shuffle);

      // 5. store bytes
      _mm_storeu_si128((__m128i*)utf8_output, utf8_packed);

      // 6. adjust pointers
      buf += 8;
      utf8_output += row[0];
      continue;
    }


    // Check for overflow in packing
    const __m128i saturation_bytemask = _mm_cmpeq_epi32(_mm_and_si128(_mm_or_si128(in, nextin), v_ffff0000), v_0000);
    const uint32_t saturation_bitmask = static_cast<uint32_t>(_mm_movemask_epi8(saturation_bytemask));

    if (saturation_bitmask == 0xffff) {
      // case: words from register produce either 1, 2 or 3 UTF-8 bytes

      // Check for illegal surrogate words
      const __m128i v_d800 = _mm_set1_epi16((uint16_t)0xd800);
      const __m128i forbidden_bytemask = _mm_cmpeq_epi16(_mm_and_si128(in_16, v_f800), v_d800);
      if (static_cast<uint32_t>(_mm_movemask_epi8(forbidden_bytemask)) != 0) {
        return std::make_pair(result(error_code::SURROGATE, buf - start), utf8_output);
      }

      const __m128i dup_even = _mm_setr_epi16(0x0000, 0x0202, 0x0404, 0x0606,
                                              0x0808, 0x0a0a, 0x0c0c, 0x0e0e);

      /* In this branch we handle three cases:
          1. [0000|0000|0ccc|cccc] => [0ccc|cccc]                           - single UFT-8 byte
          2. [0000|0bbb|bbcc|cccc] => [110b|bbbb], [10cc|cccc]              - two UTF-8 bytes
          3. [aaaa|bbbb|bbcc|cccc] => [1110|aaaa], [10bb|bbbb], [10cc|cccc] - three UTF-8 bytes

        We expand the input word (16-bit) into two words (32-bit), thus
        we have room for four bytes. However, we need five distinct bit
        layouts. Note that the last byte in cases #2 and #3 is the same.

        We precompute byte 1 for case #1 and the common byte for cases #2 & #3
        in register t2.

        We precompute byte 1 for case #3 and -- **conditionally** -- precompute
        either byte 1 for case #2 or byte 2 for case #3. Note that they
        differ by exactly one bit.

        Finally from these two words we build proper UTF-8 sequence, taking
        into account the case (i.e, the number of bytes to write).
      */
      /**
       * Given [aaaa|bbbb|bbcc|cccc] our goal is to produce:
       * t2 => [0ccc|cccc] [10cc|cccc]
       * s4 => [1110|aaaa] ([110b|bbbb] OR [10bb|bbbb])
       */
#define simdutf_vec(x) _mm_set1_epi16(static_cast<uint16_t>(x))
      // [aaaa|bbbb|bbcc|cccc] => [bbcc|cccc|bbcc|cccc]
      const __m128i t0 = _mm_shuffle_epi8(in_16, dup_even);
      // [bbcc|cccc|bbcc|cccc] => [00cc|cccc|0bcc|cccc]
      const __m128i t1 = _mm_and_si128(t0, simdutf_vec(0b0011111101111111));
      // [00cc|cccc|0bcc|cccc] => [10cc|cccc|0bcc|cccc]
      const __m128i t2 = _mm_or_si128 (t1, simdutf_vec(0b1000000000000000));

      // [aaaa|bbbb|bbcc|cccc] =>  [0000|aaaa|bbbb|bbcc]
      const __m128i s0 = _mm_srli_epi16(in_16, 4);
      // [0000|aaaa|bbbb|bbcc] => [0000|aaaa|bbbb|bb00]
      const __m128i s1 = _mm_and_si128(s0, simdutf_vec(0b0000111111111100));
      // [0000|aaaa|bbbb|bb00] => [00bb|bbbb|0000|aaaa]
      const __m128i s2 = _mm_maddubs_epi16(s1, simdutf_vec(0x0140));
      // [00bb|bbbb|0000|aaaa] => [11bb|bbbb|1110|aaaa]
      const __m128i s3 = _mm_or_si128(s2, simdutf_vec(0b1100000011100000));
      const __m128i m0 = _mm_andnot_si128(one_or_two_bytes_bytemask, simdutf_vec(0b0100000000000000));
      const __m128i s4 = _mm_xor_si128(s3, m0);
#undef simdutf_vec

      // 4. expand words 16-bit => 32-bit
      const __m128i out0 = _mm_unpacklo_epi16(t2, s4);
      const __m128i out1 = _mm_unpackhi_epi16(t2, s4);

      // 5. compress 32-bit words into 1, 2 or 3 bytes -- 2 x shuffle
      const uint16_t mask = (one_byte_bitmask & 0x5555) |
                            (one_or_two_bytes_bitmask & 0xaaaa);
      if(mask == 0) {
        // We only have three-byte words. Use fast path.
        const __m128i shuffle = _mm_setr_epi8(2,3,1,6,7,5,10,11,9,14,15,13,-1,-1,-1,-1);
        const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle);
        const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle);
        _mm_storeu_si128((__m128i*)utf8_output, utf8_0);
        utf8_output += 12;
        _mm_storeu_si128((__m128i*)utf8_output, utf8_1);
        utf8_output += 12;
        buf += 8;
        continue;
      }
      const uint8_t mask0 = uint8_t(mask);

      const uint8_t* row0 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask0][0];
      const __m128i shuffle0 = _mm_loadu_si128((__m128i*)(row0 + 1));
      const __m128i utf8_0 = _mm_shuffle_epi8(out0, shuffle0);

      const uint8_t mask1 = static_cast<uint8_t>(mask >> 8);

      const uint8_t* row1 = &simdutf::tables::utf16_to_utf8::pack_1_2_3_utf8_bytes[mask1][0];
      const __m128i shuffle1 = _mm_loadu_si128((__m128i*)(row1 + 1));
      const __m128i utf8_1 = _mm_shuffle_epi8(out1, shuffle1);

      _mm_storeu_si128((__m128i*)utf8_output, utf8_0);
      utf8_output += row0[0];
      _mm_storeu_si128((__m128i*)utf8_output, utf8_1);
      utf8_output += row1[0];

      buf += 8;
    } else {
      // case: at least one 32-bit word produce a surrogate pair in UTF-16 <=> will produce four UTF-8 bytes
      // Let us do a scalar fallback.
      // It may seem wasteful to use scalar code, but being efficient with SIMD
      // in the presence of surrogate pairs may require non-trivial tables.
      size_t forward = 15;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint32_t word = buf[k];
        if((word & 0xFFFFFF80)==0) {
          *utf8_output++ = char(word);
        } else if((word & 0xFFFFF800)==0) {
          *utf8_output++ = char((word>>6) | 0b11000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else if((word &0xFFFF0000 )==0) {
          if (word >= 0xD800 && word <= 0xDFFF) { return std::make_pair(result(error_code::SURROGATE, buf - start + k), utf8_output); }
          *utf8_output++ = char((word>>12) | 0b11100000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        } else {
          if (word > 0x10FFFF) { return std::make_pair(result(error_code::TOO_LARGE, buf- start + k), utf8_output); }
          *utf8_output++ = char((word>>18) | 0b11110000);
          *utf8_output++ = char(((word>>12) & 0b111111) | 0b10000000);
          *utf8_output++ = char(((word>>6) & 0b111111) | 0b10000000);
          *utf8_output++ = char((word & 0b111111) | 0b10000000);
        }
      }
      buf += k;
    }
  } // while

  return std::make_pair(result(error_code::SUCCESS, buf - start), utf8_output);
}
/* end file src/westmere/sse_convert_utf32_to_utf8.cpp */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=westmere/sse_convert_utf32_to_utf16.cpp
/* begin file src/westmere/sse_convert_utf32_to_utf16.cpp */
template <endianness big_endian>
std::pair<const char32_t*, char16_t*> sse_convert_utf32_to_utf16(const char32_t* buf, size_t len, char16_t* utf16_output) {

  const char32_t* end = buf + len;

  const __m128i v_0000 = _mm_setzero_si128();
  const __m128i v_ffff0000 = _mm_set1_epi32((int32_t)0xffff0000);
  __m128i forbidden_bytemask = _mm_setzero_si128();

  while (buf + 8 <= end) {
    __m128i in = _mm_loadu_si128((__m128i*)buf);
    __m128i nextin = _mm_loadu_si128((__m128i*)buf+1);
    const __m128i saturation_bytemask = _mm_cmpeq_epi32(_mm_and_si128(_mm_or_si128(in, nextin), v_ffff0000), v_0000);
    const uint32_t saturation_bitmask = static_cast<uint32_t>(_mm_movemask_epi8(saturation_bytemask));

    // Check if no bits set above 16th
    if (saturation_bitmask == 0xffff) {
      // Pack UTF-32 to UTF-16
      __m128i utf16_packed = _mm_packus_epi32(in, nextin);

      const __m128i v_f800 = _mm_set1_epi16((uint16_t)0xf800);
      const __m128i v_d800 = _mm_set1_epi16((uint16_t)0xd800);
      forbidden_bytemask = _mm_or_si128(forbidden_bytemask, _mm_cmpeq_epi16(_mm_and_si128(utf16_packed, v_f800), v_d800));

      if (big_endian) {
        const __m128i swap = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
        utf16_packed = _mm_shuffle_epi8(utf16_packed, swap);
      }

      _mm_storeu_si128((__m128i*)utf16_output, utf16_packed);
      utf16_output += 8;
      buf += 8;
    } else {
      size_t forward = 7;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint32_t word = buf[k];
        if((word & 0xFFFF0000)==0) {
          // will not generate a surrogate pair
          if (word >= 0xD800 && word <= 0xDFFF) { return std::make_pair(nullptr, utf16_output); }
          *utf16_output++ = big_endian ? char16_t((uint16_t(word) >> 8) | (uint16_t(word) << 8)) : char16_t(word);
        } else {
          // will generate a surrogate pair
          if (word > 0x10FFFF) { return std::make_pair(nullptr, utf16_output); }
          word -= 0x10000;
          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
          if (big_endian) {
            high_surrogate = uint16_t((high_surrogate >> 8) | (high_surrogate << 8));
            low_surrogate = uint16_t((low_surrogate >> 8) | (low_surrogate << 8));
          }
          *utf16_output++ = char16_t(high_surrogate);
          *utf16_output++ = char16_t(low_surrogate);
        }
      }
      buf += k;
    }
  }

  // check for invalid input
  if (static_cast<uint32_t>(_mm_movemask_epi8(forbidden_bytemask)) != 0) { return std::make_pair(nullptr, utf16_output); }

  return std::make_pair(buf, utf16_output);
}


template <endianness big_endian>
std::pair<result, char16_t*> sse_convert_utf32_to_utf16_with_errors(const char32_t* buf, size_t len, char16_t* utf16_output) {
  const char32_t* start = buf;
  const char32_t* end = buf + len;

  const __m128i v_0000 = _mm_setzero_si128();
  const __m128i v_ffff0000 = _mm_set1_epi32((int32_t)0xffff0000);

  while (buf + 8 <= end) {
    __m128i in = _mm_loadu_si128((__m128i*)buf);
    __m128i nextin = _mm_loadu_si128((__m128i*)buf+1);
    const __m128i saturation_bytemask = _mm_cmpeq_epi32(_mm_and_si128(_mm_or_si128(in, nextin), v_ffff0000), v_0000);
    const uint32_t saturation_bitmask = static_cast<uint32_t>(_mm_movemask_epi8(saturation_bytemask));

    // Check if no bits set above 16th
    if (saturation_bitmask == 0xffff) {
      // Pack UTF-32 to UTF-16
      __m128i utf16_packed = _mm_packus_epi32(in, nextin);

      const __m128i v_f800 = _mm_set1_epi16((uint16_t)0xf800);
      const __m128i v_d800 = _mm_set1_epi16((uint16_t)0xd800);
      const __m128i forbidden_bytemask = _mm_cmpeq_epi16(_mm_and_si128(utf16_packed, v_f800), v_d800);
      if (static_cast<uint32_t>(_mm_movemask_epi8(forbidden_bytemask)) != 0) {
        return std::make_pair(result(error_code::SURROGATE, buf - start), utf16_output);
      }

      if (big_endian) {
        const __m128i swap = _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
        utf16_packed = _mm_shuffle_epi8(utf16_packed, swap);
      }

      _mm_storeu_si128((__m128i*)utf16_output, utf16_packed);
      utf16_output += 8;
      buf += 8;
    } else {
      size_t forward = 7;
      size_t k = 0;
      if(size_t(end - buf) < forward + 1) { forward = size_t(end - buf - 1);}
      for(; k < forward; k++) {
        uint32_t word = buf[k];
        if((word & 0xFFFF0000)==0) {
          // will not generate a surrogate pair
          if (word >= 0xD800 && word <= 0xDFFF) { return std::make_pair(result(error_code::SURROGATE, buf - start + k), utf16_output); }
          *utf16_output++ = big_endian ? char16_t((uint16_t(word) >> 8) | (uint16_t(word) << 8)) : char16_t(word);
        } else {
          // will generate a surrogate pair
          if (word > 0x10FFFF) { return std::make_pair(result(error_code::TOO_LARGE, buf - start + k), utf16_output); }
          word -= 0x10000;
          uint16_t high_surrogate = uint16_t(0xD800 + (word >> 10));
          uint16_t low_surrogate = uint16_t(0xDC00 + (word & 0x3FF));
          if (big_endian) {
            high_surrogate = uint16_t((high_surrogate >> 8) | (high_surrogate << 8));
            low_surrogate = uint16_t((low_surrogate >> 8) | (low_surrogate << 8));
          }
          *utf16_output++ = char16_t(high_surrogate);
          *utf16_output++ = char16_t(low_surrogate);
        }
      }
      buf += k;
    }
  }

  return std::make_pair(result(error_code::SUCCESS, buf - start), utf16_output);
}
/* end file src/westmere/sse_convert_utf32_to_utf16.cpp */

} // unnamed namespace
} // namespace westmere
} // namespace simdutf

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/buf_block_reader.h
/* begin file src/generic/buf_block_reader.h */
namespace simdutf {
namespace westmere {
namespace {

// Walks through a buffer in block-sized increments, loading the last part with spaces
template<size_t STEP_SIZE>
struct buf_block_reader {
public:
  simdutf_really_inline buf_block_reader(const uint8_t *_buf, size_t _len);
  simdutf_really_inline size_t block_index();
  simdutf_really_inline bool has_full_block() const;
  simdutf_really_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.
   */
  simdutf_really_inline size_t get_remainder(uint8_t *dst) const;
  simdutf_really_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
simdutf_unused static char * format_input_text_64(const uint8_t *text) {
  static char *buf = reinterpret_cast<char*>(malloc(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
simdutf_unused static char * format_input_text(const simd8x64<uint8_t>& in) {
  static char *buf = reinterpret_cast<char*>(malloc(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;
}

simdutf_unused static char * format_mask(uint64_t mask) {
  static char *buf = reinterpret_cast<char*>(malloc(64 + 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>
simdutf_really_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>
simdutf_really_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }

template<size_t STEP_SIZE>
simdutf_really_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
  return idx < lenminusstep;
}

template<size_t STEP_SIZE>
simdutf_really_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
  return &buf[idx];
}

template<size_t STEP_SIZE>
simdutf_really_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>
simdutf_really_inline void buf_block_reader<STEP_SIZE>::advance() {
  idx += STEP_SIZE;
}

} // unnamed namespace
} // namespace westmere
} // namespace simdutf
/* end file src/generic/buf_block_reader.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_validation/utf8_lookup4_algorithm.h
/* begin file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
namespace simdutf {
namespace westmere {
namespace {
namespace utf8_validation {

using namespace simd;

  simdutf_really_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);
  }
  simdutf_really_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 = simd8<uint8_t>(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.
  //
  simdutf_really_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______
    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, 0b11110000u-1, 0b11100000u-1, 0b11000000u-1
    };
    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.
    //
    simdutf_really_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.
    simdutf_really_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;
    }

    simdutf_really_inline void check_next_input(const simd8x64<uint8_t>& input) {
      if(simdutf_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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
            "We support either two or four chunks per 64-byte block.");
        if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
          this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
        } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
          this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
          this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
          this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
        }
        this->prev_incomplete = is_incomplete(input.chunks[simd8x64<uint8_t>::NUM_CHUNKS-1]);
        this->prev_input_block = input.chunks[simd8x64<uint8_t>::NUM_CHUNKS-1];

      }
    }

    // do not forget to call check_eof!
    simdutf_really_inline bool errors() const {
      return this->error.any_bits_set_anywhere();
    }

  }; // struct utf8_checker
} // namespace utf8_validation

using utf8_validation::utf8_checker;

} // unnamed namespace
} // namespace westmere
} // namespace simdutf
/* end file src/generic/utf8_validation/utf8_lookup4_algorithm.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_validation/utf8_validator.h
/* begin file src/generic/utf8_validation/utf8_validator.h */
namespace simdutf {
namespace westmere {
namespace {
namespace utf8_validation {

/**
 * 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();
}

bool generic_validate_utf8(const char * input, size_t length) {
  return generic_validate_utf8<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
}

/**
 * Validates that the string is actual UTF-8 and stops on errors.
 */
template<class checker>
result generic_validate_utf8_with_errors(const uint8_t * input, size_t length) {
    checker c{};
    buf_block_reader<64> reader(input, length);
    size_t count{0};
    while (reader.has_full_block()) {
      simd::simd8x64<uint8_t> in(reader.full_block());
      c.check_next_input(in);
      if(c.errors()) {
        if (count != 0) { count--; } // Sometimes the error is only detected in the next chunk
        result res = scalar::utf8::rewind_and_validate_with_errors(reinterpret_cast<const char*>(input + count), length - count);
        res.count += count;
        return res;
      }
      reader.advance();
      count += 64;
    }
    uint8_t block[64]{};
    reader.get_remainder(block);
    simd::simd8x64<uint8_t> in(block);
    c.check_next_input(in);
    reader.advance();
    c.check_eof();
    if (c.errors()) {
      result res = scalar::utf8::rewind_and_validate_with_errors(reinterpret_cast<const char*>(input) + count, length - count);
      res.count += count;
      return res;
    } else {
      return result(error_code::SUCCESS, length);
    }
}

result generic_validate_utf8_with_errors(const char * input, size_t length) {
  return generic_validate_utf8_with_errors<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
}

template<class checker>
bool generic_validate_ascii(const uint8_t * input, size_t length) {
    buf_block_reader<64> reader(input, length);
    uint8_t blocks[64]{};
    simd::simd8x64<uint8_t> running_or(blocks);
    while (reader.has_full_block()) {
      simd::simd8x64<uint8_t> in(reader.full_block());
      running_or |= in;
      reader.advance();
    }
    uint8_t block[64]{};
    reader.get_remainder(block);
    simd::simd8x64<uint8_t> in(block);
    running_or |= in;
    return running_or.is_ascii();
}

bool generic_validate_ascii(const char * input, size_t length) {
  return generic_validate_ascii<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
}

template<class checker>
result generic_validate_ascii_with_errors(const uint8_t * input, size_t length) {
  buf_block_reader<64> reader(input, length);
  size_t count{0};
  while (reader.has_full_block()) {
    simd::simd8x64<uint8_t> in(reader.full_block());
    if (!in.is_ascii()) {
      result res = scalar::ascii::validate_with_errors(reinterpret_cast<const char*>(input + count), length - count);
      return result(res.error, count + res.count);
    }
    reader.advance();

    count += 64;
  }
  uint8_t block[64]{};
  reader.get_remainder(block);
  simd::simd8x64<uint8_t> in(block);
  if (!in.is_ascii()) {
    result res = scalar::ascii::validate_with_errors(reinterpret_cast<const char*>(input + count), length - count);
    return result(res.error, count + res.count);
  } else {
    return result(error_code::SUCCESS, length);
  }
}

result generic_validate_ascii_with_errors(const char * input, size_t length) {
  return generic_validate_ascii_with_errors<utf8_checker>(reinterpret_cast<const uint8_t *>(input),length);
}

} // namespace utf8_validation
} // unnamed namespace
} // namespace westmere
} // namespace simdutf
/* end file src/generic/utf8_validation/utf8_validator.h */
// transcoding from UTF-8 to UTF-16
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_to_utf16/valid_utf8_to_utf16.h
/* begin file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */


namespace simdutf {
namespace westmere {
namespace {
namespace utf8_to_utf16 {

using namespace simd;

template <endianness endian>
simdutf_warn_unused size_t convert_valid(const char* input, size_t size,
    char16_t* utf16_output) noexcept {
  // The implementation is not specific to haswell and should be moved to the generic directory.
  size_t pos = 0;
  char16_t* start{utf16_output};
  const size_t safety_margin = 16; // to avoid overruns!
  while(pos + 64 + safety_margin <= size) {
    // this loop could be unrolled further. For example, we could process the mask
    // far more than 64 bytes.
    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
    if(in.is_ascii()) {
      in.store_ascii_as_utf16<endian>(utf16_output);
      utf16_output += 64;
      pos += 64;
    } else {
      // Slow path. We hope that the compiler will recognize that this is a slow path.
      // Anything that is not a continuation mask is a 'leading byte', that is, the
      // start of a new code point.
      uint64_t utf8_continuation_mask = in.lt(-65 + 1);
      // -65 is 0b10111111 in two-complement's, so largest possible continuation byte
      uint64_t utf8_leading_mask = ~utf8_continuation_mask;
      // The *start* of code points is not so useful, rather, we want the *end* of code points.
      uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
      // We process in blocks of up to 12 bytes except possibly
      // for fast paths which may process up to 16 bytes. For the
      // slow path to work, we should have at least 12 input bytes left.
      size_t max_starting_point = (pos + 64) - 12;
      // Next loop is going to run at least five times when using solely
      // the slow/regular path, and at least four times if there are fast paths.
      while(pos < max_starting_point) {
        // Performance note: our ability to compute 'consumed' and
        // then shift and recompute is critical. If there is a
        // latency of, say, 4 cycles on getting 'consumed', then
        // the inner loop might have a total latency of about 6 cycles.
        // Yet we process between 6 to 12 inputs bytes, thus we get
        // a speed limit between 1 cycle/byte and 0.5 cycle/byte
        // for this section of the code. Hence, there is a limit
        // to how much we can further increase this latency before
        // it seriously harms performance.
        //
        // Thus we may allow convert_masked_utf8_to_utf16 to process
        // more bytes at a time under a fast-path mode where 16 bytes
        // are consumed at once (e.g., when encountering ASCII).
        size_t consumed = convert_masked_utf8_to_utf16<endian>(input + pos,
                            utf8_end_of_code_point_mask, utf16_output);
        pos += consumed;
        utf8_end_of_code_point_mask >>= consumed;
      }
      // At this point there may remain between 0 and 12 bytes in the
      // 64-byte block. These bytes will be processed again. So we have an
      // 80% efficiency (in the worst case). In practice we expect an
      // 85% to 90% efficiency.
    }
  }
  utf16_output += scalar::utf8_to_utf16::convert_valid<endian>(input + pos, size - pos, utf16_output);
  return utf16_output - start;
}

} // namespace utf8_to_utf16
} // unnamed namespace
} // namespace westmere
} // namespace simdutf
/* end file src/generic/utf8_to_utf16/valid_utf8_to_utf16.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_to_utf16/utf8_to_utf16.h
/* begin file src/generic/utf8_to_utf16/utf8_to_utf16.h */


namespace simdutf {
namespace westmere {
namespace {
namespace utf8_to_utf16 {
using namespace simd;


  simdutf_really_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);
  }
  simdutf_really_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 = simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
    simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
    return must23_80 ^ sc;
  }


  struct validating_transcoder {
    // If this is nonzero, there has been a UTF-8 error.
    simd8<uint8_t> error;

    validating_transcoder() : error(uint8_t(0)) {}
    //
    // Check whether the current bytes are valid UTF-8.
    //
    simdutf_really_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);
    }


    template <endianness endian>
    simdutf_really_inline size_t convert(const char* in, size_t size, char16_t* utf16_output) {
      size_t pos = 0;
      char16_t* start{utf16_output};
      // In the worst case, we have the haswell kernel which can cause an overflow of
      // 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the last 16 bytes,
      // and if the data is valid, then it is entirely safe because 16 UTF-8 bytes generate
      // much more than 8 bytes. However, you cannot generally assume that you have valid
      // UTF-8 input, so we are going to go back from the end counting 8 leading bytes,
      // to give us a good margin.
      size_t leading_byte = 0;
      size_t margin = size;
      for(; margin > 0 && leading_byte < 8; margin--) {
        leading_byte += (int8_t(in[margin-1]) > -65);
      }
      // If the input is long enough, then we have that margin-1 is the eight last leading byte.
      const size_t safety_margin = size - margin + 1; // to avoid overruns!
      while(pos + 64 + safety_margin <= size) {
        simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
        if(input.is_ascii()) {
          input.store_ascii_as_utf16<endian>(utf16_output);
          utf16_output += 64;
          pos += 64;
        } 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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
              "We support either two or four chunks per 64-byte block.");
          auto zero = simd8<uint8_t>{uint8_t(0)};
          if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
            this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
            this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
          }
          uint64_t utf8_continuation_mask = input.lt(-65 + 1);
          uint64_t utf8_leading_mask = ~utf8_continuation_mask;
          uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
          // We process in blocks of up to 12 bytes except possibly
          // for fast paths which may process up to 16 bytes. For the
          // slow path to work, we should have at least 12 input bytes left.
          size_t max_starting_point = (pos + 64) - 12;
          // Next loop is going to run at least five times.
          while(pos < max_starting_point) {
            // Performance note: our ability to compute 'consumed' and
            // then shift and recompute is critical. If there is a
            // latency of, say, 4 cycles on getting 'consumed', then
            // the inner loop might have a total latency of about 6 cycles.
            // Yet we process between 6 to 12 inputs bytes, thus we get
            // a speed limit between 1 cycle/byte and 0.5 cycle/byte
            // for this section of the code. Hence, there is a limit
            // to how much we can further increase this latency before
            // it seriously harms performance.
            size_t consumed = convert_masked_utf8_to_utf16<endian>(in + pos,
                            utf8_end_of_code_point_mask, utf16_output);
            pos += consumed;
            utf8_end_of_code_point_mask >>= consumed;
          }
          // At this point there may remain between 0 and 12 bytes in the
          // 64-byte block. These bytes will be processed again. So we have an
          // 80% efficiency (in the worst case). In practice we expect an
          // 85% to 90% efficiency.
        }
      }
      if(errors()) { return 0; }
      if(pos < size) {
        size_t howmany  = scalar::utf8_to_utf16::convert<endian>(in + pos, size - pos, utf16_output);
        if(howmany == 0) { return 0; }
        utf16_output += howmany;
      }
      return utf16_output - start;
    }

    template <endianness endian>
    simdutf_really_inline result convert_with_errors(const char* in, size_t size, char16_t* utf16_output) {
      size_t pos = 0;
      char16_t* start{utf16_output};
      // In the worst case, we have the haswell kernel which can cause an overflow of
      // 8 bytes when calling convert_masked_utf8_to_utf16. If you skip the last 16 bytes,
      // and if the data is valid, then it is entirely safe because 16 UTF-8 bytes generate
      // much more than 8 bytes. However, you cannot generally assume that you have valid
      // UTF-8 input, so we are going to go back from the end counting 8 leading bytes,
      // to give us a good margin.
      size_t leading_byte = 0;
      size_t margin = size;
      for(; margin > 0 && leading_byte < 8; margin--) {
        leading_byte += (int8_t(in[margin-1]) > -65);
      }
      // If the input is long enough, then we have that margin-1 is the eight last leading byte.
      const size_t safety_margin = size - margin + 1; // to avoid overruns!
      while(pos + 64 + safety_margin <= size) {
        simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
        if(input.is_ascii()) {
          input.store_ascii_as_utf16<endian>(utf16_output);
          utf16_output += 64;
          pos += 64;
        } 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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
              "We support either two or four chunks per 64-byte block.");
          auto zero = simd8<uint8_t>{uint8_t(0)};
          if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
            this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
            this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
          }
          if (errors()) {
            // rewind_and_convert_with_errors will seek a potential error from in+pos onward,
            // with the ability to go back up to pos bytes, and read size-pos bytes forward.
            result res = scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(pos, in + pos, size - pos, utf16_output);
            res.count += pos;
            return res;
          }
          uint64_t utf8_continuation_mask = input.lt(-65 + 1);
          uint64_t utf8_leading_mask = ~utf8_continuation_mask;
          uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
          // We process in blocks of up to 12 bytes except possibly
          // for fast paths which may process up to 16 bytes. For the
          // slow path to work, we should have at least 12 input bytes left.
          size_t max_starting_point = (pos + 64) - 12;
          // Next loop is going to run at least five times.
          while(pos < max_starting_point) {
            // Performance note: our ability to compute 'consumed' and
            // then shift and recompute is critical. If there is a
            // latency of, say, 4 cycles on getting 'consumed', then
            // the inner loop might have a total latency of about 6 cycles.
            // Yet we process between 6 to 12 inputs bytes, thus we get
            // a speed limit between 1 cycle/byte and 0.5 cycle/byte
            // for this section of the code. Hence, there is a limit
            // to how much we can further increase this latency before
            // it seriously harms performance.
            size_t consumed = convert_masked_utf8_to_utf16<endian>(in + pos,
                            utf8_end_of_code_point_mask, utf16_output);
            pos += consumed;
            utf8_end_of_code_point_mask >>= consumed;
          }
          // At this point there may remain between 0 and 12 bytes in the
          // 64-byte block. These bytes will be processed again. So we have an
          // 80% efficiency (in the worst case). In practice we expect an
          // 85% to 90% efficiency.
        }
      }
      if(errors()) {
        // rewind_and_convert_with_errors will seek a potential error from in+pos onward,
        // with the ability to go back up to pos bytes, and read size-pos bytes forward.
        result res = scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(pos, in + pos, size - pos, utf16_output);
        res.count += pos;
        return res;
      }
      if(pos < size) {
        // rewind_and_convert_with_errors will seek a potential error from in+pos onward,
        // with the ability to go back up to pos bytes, and read size-pos bytes forward.
        result res = scalar::utf8_to_utf16::rewind_and_convert_with_errors<endian>(pos, in + pos, size - pos, utf16_output);
        if (res.error) {    // In case of error, we want the error position
          res.count += pos;
          return res;
        } else {    // In case of success, we want the number of word written
          utf16_output += res.count;
        }
      }
      return result(error_code::SUCCESS, utf16_output - start);
    }

    simdutf_really_inline bool errors() const {
      return this->error.any_bits_set_anywhere();
    }

  }; // struct utf8_checker
} // utf8_to_utf16 namespace
} // unnamed namespace
} // namespace westmere
} // namespace simdutf
/* end file src/generic/utf8_to_utf16/utf8_to_utf16.h */
// transcoding from UTF-8 to UTF-32
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_to_utf32/valid_utf8_to_utf32.h
/* begin file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */

namespace simdutf {
namespace westmere {
namespace {
namespace utf8_to_utf32 {

using namespace simd;


simdutf_warn_unused size_t convert_valid(const char* input, size_t size,
    char32_t* utf32_output) noexcept {
  size_t pos = 0;
  char32_t* start{utf32_output};
  const size_t safety_margin = 16; // to avoid overruns!
  while(pos + 64 + safety_margin <= size) {
    simd8x64<int8_t> in(reinterpret_cast<const int8_t *>(input + pos));
    if(in.is_ascii()) {
      in.store_ascii_as_utf32(utf32_output);
      utf32_output += 64;
      pos += 64;
    } else {
    // -65 is 0b10111111 in two-complement's, so largest possible continuation byte
    uint64_t utf8_continuation_mask = in.lt(-65 + 1);
    uint64_t utf8_leading_mask = ~utf8_continuation_mask;
    uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
    size_t max_starting_point = (pos + 64) - 12;
    while(pos < max_starting_point) {
      size_t consumed = convert_masked_utf8_to_utf32(input + pos,
                          utf8_end_of_code_point_mask, utf32_output);
      pos += consumed;
      utf8_end_of_code_point_mask >>= consumed;
      }
    }
  }
  utf32_output += scalar::utf8_to_utf32::convert_valid(input + pos, size - pos, utf32_output);
  return utf32_output - start;
}


} // namespace utf8_to_utf32
} // unnamed namespace
} // namespace westmere
} // namespace simdutf
/* end file src/generic/utf8_to_utf32/valid_utf8_to_utf32.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8_to_utf32/utf8_to_utf32.h
/* begin file src/generic/utf8_to_utf32/utf8_to_utf32.h */


namespace simdutf {
namespace westmere {
namespace {
namespace utf8_to_utf32 {
using namespace simd;


  simdutf_really_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);
  }
  simdutf_really_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 = simd8<uint8_t>(must_be_2_3_continuation(prev2, prev3));
    simd8<uint8_t> must23_80 = must23 & uint8_t(0x80);
    return must23_80 ^ sc;
  }


  struct validating_transcoder {
    // If this is nonzero, there has been a UTF-8 error.
    simd8<uint8_t> error;

    validating_transcoder() : error(uint8_t(0)) {}
    //
    // Check whether the current bytes are valid UTF-8.
    //
    simdutf_really_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);
    }



    simdutf_really_inline size_t convert(const char* in, size_t size, char32_t* utf32_output) {
      size_t pos = 0;
      char32_t* start{utf32_output};
      // In the worst case, we have the haswell kernel which can cause an overflow of
      // 8 bytes when calling convert_masked_utf8_to_utf32. If you skip the last 16 bytes,
      // and if the data is valid, then it is entirely safe because 16 UTF-8 bytes generate
      // much more than 8 bytes. However, you cannot generally assume that you have valid
      // UTF-8 input, so we are going to go back from the end counting 4 leading bytes,
      // to give us a good margin.
      size_t leading_byte = 0;
      size_t margin = size;
      for(; margin > 0 && leading_byte < 4; margin--) {
        leading_byte += (int8_t(in[margin-1]) > -65);
      }
      // If the input is long enough, then we have that margin-1 is the fourth last leading byte.
      const size_t safety_margin = size - margin + 1; // to avoid overruns!
      while(pos + 64 + safety_margin <= size) {
        simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
        if(input.is_ascii()) {
          input.store_ascii_as_utf32(utf32_output);
          utf32_output += 64;
          pos += 64;
        } 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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
              "We support either two or four chunks per 64-byte block.");
          auto zero = simd8<uint8_t>{uint8_t(0)};
          if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
            this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
            this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
          }
          uint64_t utf8_continuation_mask = input.lt(-65 + 1);
          uint64_t utf8_leading_mask = ~utf8_continuation_mask;
          uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
          // We process in blocks of up to 12 bytes except possibly
          // for fast paths which may process up to 16 bytes. For the
          // slow path to work, we should have at least 12 input bytes left.
          size_t max_starting_point = (pos + 64) - 12;
          // Next loop is going to run at least five times.
          while(pos < max_starting_point) {
            // Performance note: our ability to compute 'consumed' and
            // then shift and recompute is critical. If there is a
            // latency of, say, 4 cycles on getting 'consumed', then
            // the inner loop might have a total latency of about 6 cycles.
            // Yet we process between 6 to 12 inputs bytes, thus we get
            // a speed limit between 1 cycle/byte and 0.5 cycle/byte
            // for this section of the code. Hence, there is a limit
            // to how much we can further increase this latency before
            // it seriously harms performance.
            size_t consumed = convert_masked_utf8_to_utf32(in + pos,
                            utf8_end_of_code_point_mask, utf32_output);
            pos += consumed;
            utf8_end_of_code_point_mask >>= consumed;
          }
          // At this point there may remain between 0 and 12 bytes in the
          // 64-byte block. These bytes will be processed again. So we have an
          // 80% efficiency (in the worst case). In practice we expect an
          // 85% to 90% efficiency.
        }
      }
      if(errors()) { return 0; }
      if(pos < size) {
        size_t howmany  = scalar::utf8_to_utf32::convert(in + pos, size - pos, utf32_output);
        if(howmany == 0) { return 0; }
        utf32_output += howmany;
      }
      return utf32_output - start;
    }

    simdutf_really_inline result convert_with_errors(const char* in, size_t size, char32_t* utf32_output) {
      size_t pos = 0;
      char32_t* start{utf32_output};
      // In the worst case, we have the haswell kernel which can cause an overflow of
      // 8 bytes when calling convert_masked_utf8_to_utf32. If you skip the last 16 bytes,
      // and if the data is valid, then it is entirely safe because 16 UTF-8 bytes generate
      // much more than 8 bytes. However, you cannot generally assume that you have valid
      // UTF-8 input, so we are going to go back from the end counting 4 leading bytes,
      // to give us a good margin.
      size_t leading_byte = 0;
      size_t margin = size;
      for(; margin > 0 && leading_byte < 4; margin--) {
        leading_byte += (int8_t(in[margin-1]) > -65);
      }
      // If the input is long enough, then we have that margin-1 is the fourth last leading byte.
      const size_t safety_margin = size - margin + 1; // to avoid overruns!
      while(pos + 64 + safety_margin <= size) {
        simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
        if(input.is_ascii()) {
          input.store_ascii_as_utf32(utf32_output);
          utf32_output += 64;
          pos += 64;
        } 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 == 2) || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
              "We support either two or four chunks per 64-byte block.");
          auto zero = simd8<uint8_t>{uint8_t(0)};
          if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
          } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
            this->check_utf8_bytes(input.chunks[0], zero);
            this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
            this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
            this->check_utf8_bytes(input.chunks[3], input.chunks[2]);
          }
          if (errors()) {
            result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(pos, in + pos, size - pos, utf32_output);
            res.count += pos;
            return res;
          }
          uint64_t utf8_continuation_mask = input.lt(-65 + 1);
          uint64_t utf8_leading_mask = ~utf8_continuation_mask;
          uint64_t utf8_end_of_code_point_mask = utf8_leading_mask>>1;
          // We process in blocks of up to 12 bytes except possibly
          // for fast paths which may process up to 16 bytes. For the
          // slow path to work, we should have at least 12 input bytes left.
          size_t max_starting_point = (pos + 64) - 12;
          // Next loop is going to run at least five times.
          while(pos < max_starting_point) {
            // Performance note: our ability to compute 'consumed' and
            // then shift and recompute is critical. If there is a
            // latency of, say, 4 cycles on getting 'consumed', then
            // the inner loop might have a total latency of about 6 cycles.
            // Yet we process between 6 to 12 inputs bytes, thus we get
            // a speed limit between 1 cycle/byte and 0.5 cycle/byte
            // for this section of the code. Hence, there is a limit
            // to how much we can further increase this latency before
            // it seriously harms performance.
            size_t consumed = convert_masked_utf8_to_utf32(in + pos,
                            utf8_end_of_code_point_mask, utf32_output);
            pos += consumed;
            utf8_end_of_code_point_mask >>= consumed;
          }
          // At this point there may remain between 0 and 12 bytes in the
          // 64-byte block. These bytes will be processed again. So we have an
          // 80% efficiency (in the worst case). In practice we expect an
          // 85% to 90% efficiency.
        }
      }
      if(errors()) {
        result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(pos, in + pos, size - pos, utf32_output);
        res.count += pos;
        return res;
      }
      if(pos < size) {
        result res = scalar::utf8_to_utf32::rewind_and_convert_with_errors(pos, in + pos, size - pos, utf32_output);
        if (res.error) {    // In case of error, we want the error position
          res.count += pos;
          return res;
        } else {    // In case of success, we want the number of word written
          utf32_output += res.count;
        }
      }
      return result(error_code::SUCCESS, utf32_output - start);
    }

    simdutf_really_inline bool errors() const {
      return this->error.any_bits_set_anywhere();
    }

  }; // struct utf8_checker
} // utf8_to_utf32 namespace
} // unnamed namespace
} // namespace westmere
} // namespace simdutf
/* end file src/generic/utf8_to_utf32/utf8_to_utf32.h */
// other functions
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf8.h
/* begin file src/generic/utf8.h */

namespace simdutf {
namespace westmere {
namespace {
namespace utf8 {

using namespace simd;

simdutf_really_inline size_t count_code_points(const char* in, size_t size) {
    size_t pos = 0;
    size_t count = 0;
    for(;pos + 64 <= size; pos += 64) {
      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
      uint64_t utf8_continuation_mask = input.lt(-65 + 1);
      count += 64 - count_ones(utf8_continuation_mask);
    }
    return count + scalar::utf8::count_code_points(in + pos, size - pos);
}


simdutf_really_inline size_t utf16_length_from_utf8(const char* in, size_t size) {
    size_t pos = 0;
    size_t count = 0;
    // This algorithm could no doubt be improved!
    for(;pos + 64 <= size; pos += 64) {
      simd8x64<int8_t> input(reinterpret_cast<const int8_t *>(in + pos));
      uint64_t utf8_continuation_mask = input.lt(-65 + 1);
      // We count one word for anything that is not a continuation (so
      // leading bytes).
      count += 64 - count_ones(utf8_continuation_mask);
      int64_t utf8_4byte = input.gteq_unsigned(240);
      count += count_ones(utf8_4byte);
    }
    return count + scalar::utf8::utf16_length_from_utf8(in + pos, size - pos);
}


simdutf_really_inline size_t utf32_length_from_utf8(const char* in, size_t size) {
    return count_code_points(in, size);
}
} // utf8 namespace
} // unnamed namespace
} // namespace westmere
} // namespace simdutf
/* end file src/generic/utf8.h */
// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=generic/utf16.h
/* begin file src/generic/utf16.h */
namespace simdutf {
namespace westmere {
namespace {
namespace utf16 {

template <endianness big_endian>
simdutf_really_inline size_t count_code_points(const char16_t* in, size_t size) {
    size_t pos = 0;
    size_t count = 0;
    for(;pos + 32 <= size; pos += 32) {
      simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
      if (!match_system(big_endian)) input.swap_bytes();
      uint64_t not_pair = input.not_in_range(0xDC00, 0xDFFF);
      count += count_ones(not_pair) / 2;
    }
    return count + scalar::utf16::count_code_points<big_endian>(in + pos, size - pos);
}

template <endianness big_endian>
simdutf_really_inline size_t utf8_length_from_utf16(const char16_t* in, size_t size) {
    size_t pos = 0;
    size_t count = 0;
    // This algorithm could no doubt be improved!
    for(;pos + 32 <= size; pos += 32) {
      simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
      if (!match_system(big_endian)) input.swap_bytes();
      uint64_t ascii_mask = input.lteq(0x7F);
      uint64_t twobyte_mask = input.lteq(0x7FF);
      uint64_t not_pair_mask = input.not_in_range(0xD800, 0xDFFF);

      size_t ascii_count = count_ones(ascii_mask) / 2;
      size_t twobyte_count = count_ones(twobyte_mask & ~ ascii_mask) / 2;
      size_t threebyte_count = count_ones(not_pair_mask & ~ twobyte_mask) / 2;
      size_t fourbyte_count = 32 - count_ones(not_pair_mask) / 2;
      count += 2 * fourbyte_count + 3 * threebyte_count + 2 * twobyte_count + ascii_count;
    }
    return count + scalar::utf16::utf8_length_from_utf16<big_endian>(in + pos, size - pos);
}

template <endianness big_endian>
simdutf_really_inline size_t utf32_length_from_utf16(const char16_t* in, size_t size) {
    return count_code_points<big_endian>(in, size);
}

simdutf_really_inline void change_endianness_utf16(const char16_t* in, size_t size, char16_t* output) {
  size_t pos = 0;

  while (pos + 32 <= size) {
    simd16x32<uint16_t> input(reinterpret_cast<const uint16_t *>(in + pos));
    input.swap_bytes();
    input.store(reinterpret_cast<uint16_t *>(output));
    pos += 32;
    output += 32;
  }

  scalar::utf16::change_endianness_utf16(in + pos, size - pos, output);
}

} // utf16
} // unnamed namespace
} // namespace westmere
} // namespace simdutf
/* end file src/generic/utf16.h */
//
// Implementation-specific overrides
//

namespace simdutf {
namespace westmere {

simdutf_warn_unused int implementation::detect_encodings(const char * input, size_t length) const noexcept {
  // If there is a BOM, then we trust it.
  auto bom_encoding = simdutf::BOM::check_bom(input, length);
  if(bom_encoding != encoding_type::unspecified) { return bom_encoding; }
  if (length % 2 == 0) {
    return sse_detect_encodings<utf8_validation::utf8_checker>(input, length);
  } else {
    if (implementation::validate_utf8(input, length)) {
      return simdutf::encoding_type::UTF8;
    } else {
      return simdutf::encoding_type::unspecified;
    }
  }
}

simdutf_warn_unused bool implementation::validate_utf8(const char *buf, size_t len) const noexcept {
  return westmere::utf8_validation::generic_validate_utf8(buf, len);
}

simdutf_warn_unused result implementation::validate_utf8_with_errors(const char *buf, size_t len) const noexcept {
  return westmere::utf8_validation::generic_validate_utf8_with_errors(buf, len);
}

simdutf_warn_unused bool implementation::validate_ascii(const char *buf, size_t len) const noexcept {
  return westmere::utf8_validation::generic_validate_ascii(buf, len);
}

simdutf_warn_unused result implementation::validate_ascii_with_errors(const char *buf, size_t len) const noexcept {
  return westmere::utf8_validation::generic_validate_ascii_with_errors(buf,len);
}

simdutf_warn_unused bool implementation::validate_utf16le(const char16_t *buf, size_t len) const noexcept {
  const char16_t* tail = sse_validate_utf16<endianness::LITTLE>(buf, len);
  if (tail) {
    return scalar::utf16::validate<endianness::LITTLE>(tail, len - (tail - buf));
  } else {
    return false;
  }
}

simdutf_warn_unused bool implementation::validate_utf16be(const char16_t *buf, size_t len) const noexcept {
  const char16_t* tail = sse_validate_utf16<endianness::BIG>(buf, len);
  if (tail) {
    return scalar::utf16::validate<endianness::BIG>(tail, len - (tail - buf));
  } else {
    return false;
  }
}

simdutf_warn_unused result implementation::validate_utf16le_with_errors(const char16_t *buf, size_t len) const noexcept {
  result res = sse_validate_utf16_with_errors<endianness::LITTLE>(buf, len);
  if (res.count != len) {
    result scalar_res = scalar::utf16::validate_with_errors<endianness::LITTLE>(buf + res.count, len - res.count);
    return result(scalar_res.error, res.count + scalar_res.count);
  } else {
    return res;
  }
}

simdutf_warn_unused result implementation::validate_utf16be_with_errors(const char16_t *buf, size_t len) const noexcept {
  result res = sse_validate_utf16_with_errors<endianness::BIG>(buf, len);
  if (res.count != len) {
    result scalar_res = scalar::utf16::validate_with_errors<endianness::BIG>(buf + res.count, len - res.count);
    return result(scalar_res.error, res.count + scalar_res.count);
  } else {
    return res;
  }
}

simdutf_warn_unused bool implementation::validate_utf32(const char32_t *buf, size_t len) const noexcept {
  const char32_t* tail = sse_validate_utf32le(buf, len);
  if (tail) {
    return scalar::utf32::validate(tail, len - (tail - buf));
  } else {
    return false;
  }
}

simdutf_warn_unused result implementation::validate_utf32_with_errors(const char32_t *buf, size_t len) const noexcept {
  result res = sse_validate_utf32le_with_errors(buf, len);
  if (res.count != len) {
    result scalar_res = scalar::utf32::validate_with_errors(buf + res.count, len - res.count);
    return result(scalar_res.error, res.count + scalar_res.count);
  } else {
    return res;
  }
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf16le(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
  utf8_to_utf16::validating_transcoder converter;
  return converter.convert<endianness::LITTLE>(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf16be(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
  utf8_to_utf16::validating_transcoder converter;
  return converter.convert<endianness::BIG>(buf, len, utf16_output);
}

simdutf_warn_unused result implementation::convert_utf8_to_utf16le_with_errors(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
  utf8_to_utf16::validating_transcoder converter;
  return converter.convert_with_errors<endianness::LITTLE>(buf, len, utf16_output);
}

simdutf_warn_unused result implementation::convert_utf8_to_utf16be_with_errors(const char* buf, size_t len, char16_t* utf16_output) const noexcept {
  utf8_to_utf16::validating_transcoder converter;
  return converter.convert_with_errors<endianness::BIG>(buf, len, utf16_output);
}


simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16le(const char* input, size_t size,
    char16_t* utf16_output) const noexcept {
  return utf8_to_utf16::convert_valid<endianness::LITTLE>(input, size,  utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf16be(const char* input, size_t size,
    char16_t* utf16_output) const noexcept {
  return utf8_to_utf16::convert_valid<endianness::BIG>(input, size,  utf16_output);
}

simdutf_warn_unused size_t implementation::convert_utf8_to_utf32(const char* buf, size_t len, char32_t* utf32_output) const noexcept {
  utf8_to_utf32::validating_transcoder converter;
  return converter.convert(buf, len, utf32_output);
}

simdutf_warn_unused result implementation::convert_utf8_to_utf32_with_errors(const char* buf, size_t len, char32_t* utf32_output) const noexcept {
  utf8_to_utf32::validating_transcoder converter;
  return converter.convert_with_errors(buf, len, utf32_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf8_to_utf32(const char* input, size_t size,
    char32_t* utf32_output) const noexcept {
  return utf8_to_utf32::convert_valid(input, size,  utf32_output);
}

simdutf_warn_unused size_t implementation::convert_utf16le_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  std::pair<const char16_t*, char*> ret = sse_convert_utf16_to_utf8<endianness::LITTLE>(buf, len, utf8_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf8_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf16_to_utf8::convert<endianness::LITTLE>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused size_t implementation::convert_utf16be_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  std::pair<const char16_t*, char*> ret = sse_convert_utf16_to_utf8<endianness::BIG>(buf, len, utf8_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf8_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf16_to_utf8::convert<endianness::BIG>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused result implementation::convert_utf16le_to_utf8_with_errors(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char*> ret = westmere::sse_convert_utf16_to_utf8_with_errors<endianness::LITTLE>(buf, len, utf8_output);
  if (ret.first.error) { return ret.first; }  // Can return directly since scalar fallback already found correct ret.first.count
  if (ret.first.count != len) { // All good so far, but not finished
    result scalar_res = scalar::utf16_to_utf8::convert_with_errors<endianness::LITTLE>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf8_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused result implementation::convert_utf16be_to_utf8_with_errors(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char*> ret = westmere::sse_convert_utf16_to_utf8_with_errors<endianness::BIG>(buf, len, utf8_output);
  if (ret.first.error) { return ret.first; }  // Can return directly since scalar fallback already found correct ret.first.count
  if (ret.first.count != len) { // All good so far, but not finished
    result scalar_res = scalar::utf16_to_utf8::convert_with_errors<endianness::BIG>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf8_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return convert_utf16le_to_utf8(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf8(const char16_t* buf, size_t len, char* utf8_output) const noexcept {
  return convert_utf16be_to_utf8(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf8(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  std::pair<const char32_t*, char*> ret = sse_convert_utf32_to_utf8(buf, len, utf8_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf8_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf32_to_utf8::convert(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused result implementation::convert_utf32_to_utf8_with_errors(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char*> ret = westmere::sse_convert_utf32_to_utf8_with_errors(buf, len, utf8_output);
  if (ret.first.count != len) {
    result scalar_res = scalar::utf32_to_utf8::convert_with_errors(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf8_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused size_t implementation::convert_utf16le_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  std::pair<const char16_t*, char32_t*> ret = sse_convert_utf16_to_utf32<endianness::LITTLE>(buf, len, utf32_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf32_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf16_to_utf32::convert<endianness::LITTLE>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused size_t implementation::convert_utf16be_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  std::pair<const char16_t*, char32_t*> ret = sse_convert_utf16_to_utf32<endianness::BIG>(buf, len, utf32_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf32_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf16_to_utf32::convert<endianness::BIG>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused result implementation::convert_utf16le_to_utf32_with_errors(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char32_t*> ret = westmere::sse_convert_utf16_to_utf32_with_errors<endianness::LITTLE>(buf, len, utf32_output);
  if (ret.first.error) { return ret.first; }  // Can return directly since scalar fallback already found correct ret.first.count
  if (ret.first.count != len) { // All good so far, but not finished
    result scalar_res = scalar::utf16_to_utf32::convert_with_errors<endianness::LITTLE>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf32_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused result implementation::convert_utf16be_to_utf32_with_errors(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char32_t*> ret = westmere::sse_convert_utf16_to_utf32_with_errors<endianness::BIG>(buf, len, utf32_output);
  if (ret.first.error) { return ret.first; }  // Can return directly since scalar fallback already found correct ret.first.count
  if (ret.first.count != len) { // All good so far, but not finished
    result scalar_res = scalar::utf16_to_utf32::convert_with_errors<endianness::BIG>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf32_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf8(const char32_t* buf, size_t len, char* utf8_output) const noexcept {
  return convert_utf32_to_utf8(buf, len, utf8_output);
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf16le(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  std::pair<const char32_t*, char16_t*> ret = sse_convert_utf32_to_utf16<endianness::LITTLE>(buf, len, utf16_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf16_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf32_to_utf16::convert<endianness::LITTLE>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused size_t implementation::convert_utf32_to_utf16be(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  std::pair<const char32_t*, char16_t*> ret = sse_convert_utf32_to_utf16<endianness::BIG>(buf, len, utf16_output);
  if (ret.first == nullptr) { return 0; }
  size_t saved_bytes = ret.second - utf16_output;
  if (ret.first != buf + len) {
    const size_t scalar_saved_bytes = scalar::utf32_to_utf16::convert<endianness::BIG>(
                                        ret.first, len - (ret.first - buf), ret.second);
    if (scalar_saved_bytes == 0) { return 0; }
    saved_bytes += scalar_saved_bytes;
  }
  return saved_bytes;
}

simdutf_warn_unused result implementation::convert_utf32_to_utf16le_with_errors(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char16_t*> ret = westmere::sse_convert_utf32_to_utf16_with_errors<endianness::LITTLE>(buf, len, utf16_output);
  if (ret.first.count != len) {
    result scalar_res = scalar::utf32_to_utf16::convert_with_errors<endianness::LITTLE>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf16_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused result implementation::convert_utf32_to_utf16be_with_errors(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  // ret.first.count is always the position in the buffer, not the number of words written even if finished
  std::pair<result, char16_t*> ret = westmere::sse_convert_utf32_to_utf16_with_errors<endianness::BIG>(buf, len, utf16_output);
  if (ret.first.count != len) {
    result scalar_res = scalar::utf32_to_utf16::convert_with_errors<endianness::BIG>(
                                        buf + ret.first.count, len - ret.first.count, ret.second);
    if (scalar_res.error) {
      scalar_res.count += ret.first.count;
      return scalar_res;
    } else {
      ret.second += scalar_res.count;
    }
  }
  ret.first.count = ret.second - utf16_output;   // Set count to the number of 8-bit words written
  return ret.first;
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16le(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return convert_utf32_to_utf16le(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf32_to_utf16be(const char32_t* buf, size_t len, char16_t* utf16_output) const noexcept {
  return convert_utf32_to_utf16be(buf, len, utf16_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16le_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return convert_utf16le_to_utf32(buf, len, utf32_output);
}

simdutf_warn_unused size_t implementation::convert_valid_utf16be_to_utf32(const char16_t* buf, size_t len, char32_t* utf32_output) const noexcept {
  return convert_utf16be_to_utf32(buf, len, utf32_output);
}

void implementation::change_endianness_utf16(const char16_t * input, size_t length, char16_t * output) const noexcept {
  utf16::change_endianness_utf16(input, length, output);
}

simdutf_warn_unused size_t implementation::count_utf16le(const char16_t * input, size_t length) const noexcept {
  return utf16::count_code_points<endianness::LITTLE>(input, length);
}

simdutf_warn_unused size_t implementation::count_utf16be(const char16_t * input, size_t length) const noexcept {
  return utf16::count_code_points<endianness::BIG>(input, length);
}

simdutf_warn_unused size_t implementation::count_utf8(const char * input, size_t length) const noexcept {
  return utf8::count_code_points(input, length);
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf16le(const char16_t * input, size_t length) const noexcept {
  return utf16::utf8_length_from_utf16<endianness::LITTLE>(input, length);
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf16be(const char16_t * input, size_t length) const noexcept {
  return utf16::utf8_length_from_utf16<endianness::BIG>(input, length);
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf16le(const char16_t * input, size_t length) const noexcept {
  return utf16::utf32_length_from_utf16<endianness::LITTLE>(input, length);
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf16be(const char16_t * input, size_t length) const noexcept {
  return utf16::utf32_length_from_utf16<endianness::BIG>(input, length);
}

simdutf_warn_unused size_t implementation::utf16_length_from_utf8(const char * input, size_t length) const noexcept {
  return utf8::utf16_length_from_utf8(input, length);
}

simdutf_warn_unused size_t implementation::utf8_length_from_utf32(const char32_t * input, size_t length) const noexcept {
  const __m128i v_00000000 = _mm_setzero_si128();
  const __m128i v_ffffff80 = _mm_set1_epi32((uint32_t)0xffffff80);
  const __m128i v_fffff800 = _mm_set1_epi32((uint32_t)0xfffff800);
  const __m128i v_ffff0000 = _mm_set1_epi32((uint32_t)0xffff0000);
  size_t pos = 0;
  size_t count = 0;
  for(;pos + 4 <= length; pos += 4) {
    __m128i in = _mm_loadu_si128((__m128i*)(input + pos));
    const __m128i ascii_bytes_bytemask = _mm_cmpeq_epi32(_mm_and_si128(in, v_ffffff80), v_00000000);
    const __m128i one_two_bytes_bytemask = _mm_cmpeq_epi32(_mm_and_si128(in, v_fffff800), v_00000000);
    const __m128i two_bytes_bytemask = _mm_xor_si128(one_two_bytes_bytemask, ascii_bytes_bytemask);
    const __m128i one_two_three_bytes_bytemask = _mm_cmpeq_epi32(_mm_and_si128(in, v_ffff0000), v_00000000);
    const __m128i three_bytes_bytemask = _mm_xor_si128(one_two_three_bytes_bytemask, one_two_bytes_bytemask);
    const uint16_t ascii_bytes_bitmask = static_cast<uint16_t>(_mm_movemask_epi8(ascii_bytes_bytemask));
    const uint16_t two_bytes_bitmask = static_cast<uint16_t>(_mm_movemask_epi8(two_bytes_bytemask));
    const uint16_t three_bytes_bitmask = static_cast<uint16_t>(_mm_movemask_epi8(three_bytes_bytemask));

    size_t ascii_count = count_ones(ascii_bytes_bitmask) / 4;
    size_t two_bytes_count = count_ones(two_bytes_bitmask) / 4;
    size_t three_bytes_count = count_ones(three_bytes_bitmask) / 4;
    count += 16 - 3*ascii_count - 2*two_bytes_count - three_bytes_count;
  }
  return count + scalar::utf32::utf8_length_from_utf32(input + pos, length - pos);
}

simdutf_warn_unused size_t implementation::utf16_length_from_utf32(const char32_t * input, size_t length) const noexcept {
  const __m128i v_00000000 = _mm_setzero_si128();
  const __m128i v_ffff0000 = _mm_set1_epi32((uint32_t)0xffff0000);
  size_t pos = 0;
  size_t count = 0;
  for(;pos + 4 <= length; pos += 4) {
    __m128i in = _mm_loadu_si128((__m128i*)(input + pos));
    const __m128i surrogate_bytemask = _mm_cmpeq_epi32(_mm_and_si128(in, v_ffff0000), v_00000000);
    const uint16_t surrogate_bitmask = static_cast<uint16_t>(_mm_movemask_epi8(surrogate_bytemask));
    size_t surrogate_count = (16-count_ones(surrogate_bitmask))/4;
    count += 4 + surrogate_count;
  }
  return count + scalar::utf32::utf16_length_from_utf32(input + pos, length - pos);
}

simdutf_warn_unused size_t implementation::utf32_length_from_utf8(const char * input, size_t length) const noexcept {
  return scalar::utf8::count_code_points(input, length);
}

} // namespace westmere
} // namespace simdutf

// dofile: invoked with prepath=/Users/lemire/CVS/github/simdutf/src, filename=simdutf/westmere/end.h
/* begin file src/simdutf/westmere/end.h */
#if SIMDUTF_CAN_ALWAYS_RUN_WESTMERE
// nothing needed.
#else
SIMDUTF_UNTARGET_REGION
#endif

/* end file src/simdutf/westmere/end.h */
/* end file src/westmere/implementation.cpp */
#endif

SIMDUTF_POP_DISABLE_WARNINGS
/* end file src/simdutf.cpp */