react-native-quick-crypto 1.0.18 → 1.1.0

package/deps/simdutf/src/simdutf/haswell/simd.h

@@ -0,0 +1,363 @@
+#ifndef SIMDUTF_HASWELL_SIMD_H
+#define SIMDUTF_HASWELL_SIMD_H
+
+namespace simdutf {
+namespace SIMDUTF_IMPLEMENTATION {
+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) {}
+
+  simdutf_really_inline operator const __m256i &() const { 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 &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>> {
+  simdutf_really_inline base8() : base<simd8<T>>() {}
+
+  simdutf_really_inline base8(const __m256i _value) : base<simd8<T>>(_value) {}
+
+  friend simdutf_always_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(const __m256i _value) : base8<bool>(_value) {}
+
+  simdutf_really_inline simd8(bool _value) : base8<bool>(splat(_value)) {}
+
+  simdutf_really_inline uint32_t to_bitmask() const {
+    return uint32_t(_mm256_movemask_epi8(value));
+  }
+};
+
+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_sub_epi8(*this, other);
+  }
+  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;
+
+  simdutf_really_inline bool is_ascii() const {
+    return _mm256_movemask_epi8(*this) == 0;
+  }
+  // Order-sensitive comparisons
+  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)) {}
+
+  // Saturated math
+  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>
+  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);
+  }
+  simdutf_really_inline simd8<bool>
+  operator>=(const simd8<uint8_t> other) const {
+    return other.min_val(*this) == other;
+  }
+
+  // Bit-specific operations
+  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();
+  }
+
+  template <int N> simdutf_really_inline simd8<uint8_t> shr() const {
+    return simd8<uint8_t>(_mm256_srli_epi16(*this, N)) & uint8_t(0xFFu >> N);
+  }
+
+  simdutf_really_inline uint64_t sum_bytes() const {
+    const auto tmp = _mm256_sad_epu8(value, _mm256_setzero_si256());
+
+    return _mm256_extract_epi64(tmp, 0) + _mm256_extract_epi64(tmp, 1) +
+           _mm256_extract_epi64(tmp, 2) + _mm256_extract_epi64(tmp, 3);
+  }
+};
+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 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))
+        .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 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 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>
+
+#include "simdutf/haswell/simd16-inl.h"
+#include "simdutf/haswell/simd32-inl.h"
+#include "simdutf/haswell/simd64-inl.h"
+
+simdutf_really_inline simd64<uint64_t> sum_8bytes(const simd8<uint8_t> v) {
+  return _mm256_sad_epu8(v.value, simd8<uint8_t>::zero());
+}
+
+} // namespace simd
+
+} // unnamed namespace
+} // namespace SIMDUTF_IMPLEMENTATION
+} // namespace simdutf
+
+#endif // SIMDUTF_HASWELL_SIMD_H

package/deps/simdutf/src/simdutf/haswell/simd16-inl.h

@@ -0,0 +1,261 @@
+#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_always_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);
+};
+
+// 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() : base16() {}
+
+  simdutf_really_inline simd16(const __m256i _value) : base16<bool>(_value) {}
+
+  // Splat constructor
+  simdutf_really_inline simd16(bool _value) : base16<bool>(splat(_value)) {}
+
+  simdutf_really_inline bitmask_type to_bitmask() const {
+    return _mm256_movemask_epi8(*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) {
+    *this = *this + other;
+    return *static_cast<simd16<T> *>(this);
+  }
+};
+
+// Unsigned code units
+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))) {}
+
+  // 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<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;
+  }
+
+  // Bit-specific operations
+  simdutf_really_inline simd16<bool> bits_not_set() const {
+    return *this == uint16_t(0);
+  }
+
+  simdutf_really_inline simd16<bool> any_bits_set() const {
+    return ~this->bits_not_set();
+  }
+
+  template <int N> simdutf_really_inline simd16<uint16_t> shr() const {
+    return simd16<uint16_t>(_mm256_srli_epi16(*this, 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 of two uint16_t code units 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 code units in linear order from v0 and v1
+    return _mm256_packus_epi16(t0, t1);
+  }
+
+  simdutf_really_inline uint64_t sum() const {
+    const auto lo_u16 = _mm256_and_si256(value, _mm256_set1_epi32(0x0000ffff));
+    const auto hi_u16 = _mm256_srli_epi32(value, 16);
+    const auto sum_u32 = _mm256_add_epi32(lo_u16, hi_u16);
+
+    const auto lo_u32 =
+        _mm256_and_si256(sum_u32, _mm256_set1_epi64x(0xffffffff));
+    const auto hi_u32 = _mm256_srli_epi64(sum_u32, 32);
+    const auto sum_u64 = _mm256_add_epi64(lo_u32, hi_u32);
+
+    return uint64_t(_mm256_extract_epi64(sum_u64, 0)) +
+           uint64_t(_mm256_extract_epi64(sum_u64, 1)) +
+           uint64_t(_mm256_extract_epi64(sum_u64, 2)) +
+           uint64_t(_mm256_extract_epi64(sum_u64, 3));
+  }
+};
+
+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 void swap_bytes() {
+    this->chunks[0] = this->chunks[0].swap_bytes();
+    this->chunks[1] = this->chunks[1].swap_bytes();
+  }
+  simdutf_really_inline uint64_t gt(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 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 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 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();
+  }
+}; // struct simd16x32<T>
+
+simd16<uint16_t> min(const simd16<uint16_t> a, simd16<uint16_t> b) {
+  return _mm256_min_epu16(a.value, b.value);
+}