react-native-quick-crypto 1.0.18 → 1.1.0

package/deps/simdutf/src/icelake/icelake_utf8_length_from_utf16.inl.cpp

@@ -0,0 +1,269 @@
+template <endianness big_endian>
+simdutf_really_inline size_t icelake_utf8_length_from_utf16(const char16_t *in,
+                                                            size_t size) {
+
+  using vector_u16 = simd16<uint16_t>;
+  constexpr size_t N = vector_u16::ELEMENTS; // 32 on AVX-512
+  if (N + 1 > size) {
+    result scalar_result =
+        scalar::utf16::utf8_length_from_utf16_with_replacement<big_endian>(
+            in, size);
+    return scalar_result.count;
+  } // special case for short inputs
+  size_t pos = 0;
+
+  const __m512i byteflip = _mm512_setr_epi64(
+      0x0607040502030001, 0x0e0f0c0d0a0b0809, 0x0607040502030001,
+      0x0e0f0c0d0a0b0809, 0x0607040502030001, 0x0e0f0c0d0a0b0809,
+      0x0607040502030001, 0x0e0f0c0d0a0b0809);
+
+  size_t count = 0;
+
+  for (; pos < size / (2 * N) * (2 * N); pos += 2 * N) {
+
+    __m512i input1 =
+        _mm512_loadu_si512(reinterpret_cast<const __m512i *>(in + pos));
+    __m512i input2 =
+        _mm512_loadu_si512(reinterpret_cast<const __m512i *>(in + pos + N));
+
+    if simdutf_constexpr (!match_system(big_endian)) {
+      input1 = _mm512_shuffle_epi8(input1, byteflip);
+      input2 = _mm512_shuffle_epi8(input2, byteflip);
+    }
+    // 0xd800 .. 0xdbff - low surrogate
+    // 0xdc00 .. 0xdfff - high surrogate
+    __mmask32 is_surrogate1 = _mm512_cmpeq_epi16_mask(
+        _mm512_and_si512(input1, _mm512_set1_epi16(uint16_t(0xf800))),
+        _mm512_set1_epi16(uint16_t(0xd800)));
+    __mmask32 is_surrogate2 = _mm512_cmpeq_epi16_mask(
+        _mm512_and_si512(input2, _mm512_set1_epi16(uint16_t(0xf800))),
+        _mm512_set1_epi16(uint16_t(0xd800)));
+    // c0 - chars that yield 2- or 3-byte UTF-8 codes
+    __mmask32 c01 =
+        _mm512_test_epi16_mask(input1, _mm512_set1_epi16(uint16_t(0xff80)));
+    __mmask32 c02 =
+        _mm512_test_epi16_mask(input2, _mm512_set1_epi16(uint16_t(0xff80)));
+
+    // c1 - chars that yield 3-byte UTF-8 codes (including surrogates)
+    __mmask32 c11 =
+        _mm512_test_epi16_mask(input1, _mm512_set1_epi16(uint16_t(0xf800)));
+    __mmask32 c12 =
+        _mm512_test_epi16_mask(input2, _mm512_set1_epi16(uint16_t(0xf800)));
+    count += count_ones32(c01);
+    count += count_ones32(c11);
+    count -= count_ones32(is_surrogate1);
+    count += count_ones32(c02);
+    count += count_ones32(c12);
+    count -= count_ones32(is_surrogate2);
+  }
+  if (pos + N <= size) {
+    __m512i input =
+        _mm512_loadu_si512(reinterpret_cast<const __m512i *>(in + pos));
+    if simdutf_constexpr (!match_system(big_endian)) {
+      input = _mm512_shuffle_epi8(input, byteflip);
+    }
+    // 0xd800 .. 0xdbff - low surrogate
+    // 0xdc00 .. 0xdfff - high surrogate
+    __mmask32 is_surrogate = _mm512_cmpeq_epi16_mask(
+        _mm512_and_si512(input, _mm512_set1_epi16(uint16_t(0xf800))),
+        _mm512_set1_epi16(uint16_t(0xd800)));
+
+    // c0 - chars that yield 2- or 3-byte UTF-8 codes
+    __mmask32 c0 =
+        _mm512_test_epi16_mask(input, _mm512_set1_epi16(uint16_t(0xff80)));
+
+    // c1 - chars that yield 3-byte UTF-8 codes (including surrogates)
+    __mmask32 c1 =
+        _mm512_test_epi16_mask(input, _mm512_set1_epi16(uint16_t(0xf800)));
+    count += count_ones32(c0);
+    count += count_ones32(c1);
+    count -= count_ones32(is_surrogate);
+    pos += N;
+  }
+  // At this point, we have processed 'pos' char16 values and we have less than
+  // N remaining.
+  __mmask32 remaining_mask =
+      0xFFFFFFFFULL >>
+      (32 - (size - pos)); // mask for the remaining char16 values
+  __m512i input = _mm512_maskz_loadu_epi16(remaining_mask, in + pos);
+  if simdutf_constexpr (!match_system(big_endian)) {
+    input = _mm512_shuffle_epi8(input, byteflip);
+  }
+  // 0xd800 .. 0xdbff - low surrogate
+  // 0xdc00 .. 0xdfff - high surrogate
+  __mmask32 is_surrogate = _mm512_cmpeq_epi16_mask(
+      _mm512_and_si512(input, _mm512_set1_epi16(uint16_t(0xf800))),
+      _mm512_set1_epi16(uint16_t(0xd800)));
+
+  // c0 - chars that yield 2- or 3-byte UTF-8 codes
+  __mmask32 c0 =
+      _mm512_test_epi16_mask(input, _mm512_set1_epi16(uint16_t(0xff80)));
+
+  // c1 - chars that yield 3-byte UTF-8 codes (including surrogates)
+  __mmask32 c1 =
+      _mm512_test_epi16_mask(input, _mm512_set1_epi16(uint16_t(0xf800)));
+  count += count_ones32(c0);
+  count += count_ones32(c1);
+  count -= count_ones32(is_surrogate);
+  pos = size;
+
+  count += pos;
+  return count;
+}
+
+template <endianness big_endian>
+simdutf_really_inline result icelake_utf8_length_from_utf16_with_replacement(
+    const char16_t *in, size_t size) {
+  ///////
+  // We repeat 3 times the same algorithm.
+  // First, we proceed with an unrolled loop of 2*N char16 values (for speed).
+  // Second, we process N char16 values.
+  // Finally, we process the remaining char16 values (less than N).
+  ///////
+  using vector_u16 = simd16<uint16_t>;
+  constexpr size_t N = vector_u16::ELEMENTS; // 32 on AVX-512
+  if (N + 1 > size) {
+    return scalar::utf16::utf8_length_from_utf16_with_replacement<big_endian>(
+        in, size);
+  } // special case for short inputs
+  size_t pos = 0;
+
+  const __m512i byteflip = _mm512_setr_epi64(
+      0x0607040502030001, 0x0e0f0c0d0a0b0809, 0x0607040502030001,
+      0x0e0f0c0d0a0b0809, 0x0607040502030001, 0x0e0f0c0d0a0b0809,
+      0x0607040502030001, 0x0e0f0c0d0a0b0809);
+
+  constexpr uint32_t straddle_mask =
+      match_system(big_endian) ? 0xfc00fc00 : 0x00fc00fc;
+  constexpr uint32_t straddle_pair =
+      match_system(big_endian) ? 0xdc00d800 : 0x00dc00d8;
+
+  size_t count = 0;
+  bool any_surrogates = false;
+  // We assume all surrogates are mismatched and count here the matched
+  // ones.
+  size_t matches = 0;
+
+  for (; pos < (size - 1) / (2 * N) * (2 * N); pos += 2 * N) {
+    __m512i current1 =
+        _mm512_loadu_si512(reinterpret_cast<const __m512i *>(in + pos));
+    if simdutf_constexpr (!match_system(big_endian)) {
+      current1 = _mm512_shuffle_epi8(current1, byteflip);
+    }
+    __m512i current2 =
+        _mm512_loadu_si512(reinterpret_cast<const __m512i *>(in + pos + N));
+    if simdutf_constexpr (!match_system(big_endian)) {
+      current2 = _mm512_shuffle_epi8(current2, byteflip);
+    }
+
+    __mmask32 is_surrogate1 = _mm512_cmpeq_epi16_mask(
+        _mm512_and_si512(current1, _mm512_set1_epi16(uint16_t(0xf800))),
+        _mm512_set1_epi16(uint16_t(0xd800)));
+    __mmask32 is_surrogate2 = _mm512_cmpeq_epi16_mask(
+        _mm512_and_si512(current2, _mm512_set1_epi16(uint16_t(0xf800))),
+        _mm512_set1_epi16(uint16_t(0xd800)));
+    __mmask32 c01 =
+        _mm512_test_epi16_mask(current1, _mm512_set1_epi16(uint16_t(0xff80)));
+    __mmask32 c11 =
+        _mm512_test_epi16_mask(current1, _mm512_set1_epi16(uint16_t(0xf800)));
+    __mmask32 c02 =
+        _mm512_test_epi16_mask(current2, _mm512_set1_epi16(uint16_t(0xff80)));
+    __mmask32 c12 =
+        _mm512_test_epi16_mask(current2, _mm512_set1_epi16(uint16_t(0xf800)));
+    count += count_ones32(c01);
+    count += count_ones32(c11);
+    count += count_ones32(c02);
+    count += count_ones32(c12);
+    if (_kor_mask32(is_surrogate1, is_surrogate2)) {
+      any_surrogates = true;
+      __m512i lb_masked1 =
+          _mm512_and_si512(current1, _mm512_set1_epi16(uint16_t(0xfc00)));
+      __mmask32 hi_surrogates1 = _mm512_cmpeq_epi16_mask(
+          lb_masked1, _mm512_set1_epi16(uint16_t(0xd800)));
+      __mmask32 lo_surrogates1 = _mm512_cmpeq_epi16_mask(
+          lb_masked1, _mm512_set1_epi16(uint16_t(0xdc00)));
+      __m512i lb_masked2 =
+          _mm512_and_si512(current2, _mm512_set1_epi16(uint16_t(0xfc00)));
+      __mmask32 hi_surrogates2 = _mm512_cmpeq_epi16_mask(
+          lb_masked2, _mm512_set1_epi16(uint16_t(0xd800)));
+      __mmask32 lo_surrogates2 = _mm512_cmpeq_epi16_mask(
+          lb_masked2, _mm512_set1_epi16(uint16_t(0xdc00)));
+      matches += count_ones32(
+          _kand_mask32(_kshiftli_mask32(hi_surrogates1, 1), lo_surrogates1));
+      matches += count_ones32(
+          _kand_mask32(_kshiftli_mask32(hi_surrogates2, 1), lo_surrogates2));
+      uint32_t straddle1, straddle2;
+      memcpy(&straddle1, in + pos + 1 * N - 1, sizeof(uint32_t));
+      memcpy(&straddle2, in + pos + 2 * N - 1, sizeof(uint32_t));
+      matches += ((straddle1 & straddle_mask) == straddle_pair) +
+                 ((straddle2 & straddle_mask) == straddle_pair);
+    }
+  }
+  if (pos + N + 1 <= size) {
+    __m512i input =
+        _mm512_loadu_si512(reinterpret_cast<const __m512i *>(in + pos));
+    if simdutf_constexpr (!match_system(big_endian)) {
+      input = _mm512_shuffle_epi8(input, byteflip);
+    }
+
+    __mmask32 is_surrogate = _mm512_cmpeq_epi16_mask(
+        _mm512_and_si512(input, _mm512_set1_epi16(uint16_t(0xf800))),
+        _mm512_set1_epi16(uint16_t(0xd800)));
+    __mmask32 c0 =
+        _mm512_test_epi16_mask(input, _mm512_set1_epi16(uint16_t(0xff80)));
+    __mmask32 c1 =
+        _mm512_test_epi16_mask(input, _mm512_set1_epi16(uint16_t(0xf800)));
+    count += count_ones32(c0);
+    count += count_ones32(c1);
+    if (is_surrogate) {
+      any_surrogates = true;
+      __m512i lb_masked =
+          _mm512_and_si512(input, _mm512_set1_epi16(uint16_t(0xfc00)));
+      __mmask32 hi_surrogates = _mm512_cmpeq_epi16_mask(
+          lb_masked, _mm512_set1_epi16(uint16_t(0xd800)));
+      __mmask32 lo_surrogates = _mm512_cmpeq_epi16_mask(
+          lb_masked, _mm512_set1_epi16(uint16_t(0xdc00)));
+      matches += count_ones32(
+          _kand_mask32(_kshiftli_mask32(hi_surrogates, 1), lo_surrogates));
+      uint32_t straddle;
+      memcpy(&straddle, in + pos + N - 1, sizeof(uint32_t));
+      matches += (straddle & straddle_mask) == straddle_pair;
+    }
+    pos += N;
+  }
+
+  size_t overshoot = 32 - (size - pos);
+  __mmask32 remaining_mask(uint32_t(0xFFFFFFFFULL << overshoot));
+  __m512i input =
+      _mm512_maskz_loadu_epi16(remaining_mask, in + pos - overshoot);
+  if simdutf_constexpr (!match_system(big_endian)) {
+    input = _mm512_shuffle_epi8(input, byteflip);
+  }
+
+  __mmask32 is_surrogate = _mm512_cmpeq_epi16_mask(
+      _mm512_and_si512(input, _mm512_set1_epi16(uint16_t(0xf800))),
+      _mm512_set1_epi16(uint16_t(0xd800)));
+  __mmask32 c0 =
+      _mm512_test_epi16_mask(input, _mm512_set1_epi16(uint16_t(0xff80)));
+  __mmask32 c1 =
+      _mm512_test_epi16_mask(input, _mm512_set1_epi16(uint16_t(0xf800)));
+
+  count += count_ones32(c0);
+  count += count_ones32(c1);
+  if (is_surrogate) {
+    any_surrogates = true;
+    __m512i lb_masked =
+        _mm512_and_si512(input, _mm512_set1_epi16(uint16_t(0xfc00)));
+    __mmask32 hi_surrogates =
+        _mm512_cmpeq_epi16_mask(lb_masked, _mm512_set1_epi16(uint16_t(0xd800)));
+    __mmask32 lo_surrogates =
+        _mm512_cmpeq_epi16_mask(lb_masked, _mm512_set1_epi16(uint16_t(0xdc00)));
+    matches += count_ones32(
+        _kand_mask32(_kshiftli_mask32(hi_surrogates, 1), lo_surrogates));
+  }
+  pos = size;
+  count += pos;
+
+  count -= 2 * matches;
+  return {any_surrogates ? SURROGATE : SUCCESS, count};
+}

package/deps/simdutf/src/icelake/icelake_utf8_validation.inl.cpp

@@ -0,0 +1,116 @@
+// 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 is 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