react-native-quick-crypto 1.0.19 → 1.1.0

package/deps/simdutf/src/haswell/avx2_convert_latin1_to_utf16.cpp

@@ -0,0 +1,28 @@
+template <endianness big_endian>
+std::pair<const char *, char16_t *>
+avx2_convert_latin1_to_utf16(const char *latin1_input, size_t len,
+                             char16_t *utf16_output) {
+  size_t rounded_len = len & ~0xF; // Round down to nearest multiple of 16
+
+  size_t i = 0;
+  for (; i < rounded_len; i += 16) {
+    // Load 16 bytes from the address (input + i) into a xmm register
+    const __m128i latin1 =
+        _mm_loadu_si128(reinterpret_cast<const __m128i *>(latin1_input + i));
+
+    // Zero extend each byte in `in` to word
+    __m256i utf16 = _mm256_cvtepu8_epi16(latin1);
+
+    if (big_endian) {
+      const __m128i swap128 =
+          _mm_setr_epi8(1, 0, 3, 2, 5, 4, 7, 6, 9, 8, 11, 10, 13, 12, 15, 14);
+      const __m256i swap = _mm256_set_m128i(swap128, swap128);
+      utf16 = _mm256_shuffle_epi8(utf16, swap);
+    }
+
+    // Store the contents of xmm1 into the address pointed by (output + i)
+    _mm256_storeu_si256(reinterpret_cast<__m256i *>(utf16_output + i), utf16);
+  }
+
+  return std::make_pair(latin1_input + rounded_len, utf16_output + rounded_len);
+}

package/deps/simdutf/src/haswell/avx2_convert_latin1_to_utf32.cpp

@@ -0,0 +1,20 @@
+std::pair<const char *, char32_t *>
+avx2_convert_latin1_to_utf32(const char *buf, size_t len,
+                             char32_t *utf32_output) {
+  size_t rounded_len = ((len | 7) ^ 7); // Round down to nearest multiple of 8
+
+  for (size_t i = 0; i < rounded_len; i += 8) {
+    // Load 8 Latin1 characters into a 64-bit register
+    __m128i in = _mm_loadl_epi64((__m128i *)&buf[i]);
+
+    // Zero extend each set of 8 Latin1 characters to 8 32-bit integers using
+    // vpmovzxbd
+    __m256i out = _mm256_cvtepu8_epi32(in);
+
+    // Store the results back to memory
+    _mm256_storeu_si256((__m256i *)&utf32_output[i], out);
+  }
+
+  // return pointers pointing to where we left off
+  return std::make_pair(buf + rounded_len, utf32_output + rounded_len);
+}

package/deps/simdutf/src/haswell/avx2_convert_latin1_to_utf8.cpp

@@ -0,0 +1,83 @@
+std::pair<const char *, char *>
+avx2_convert_latin1_to_utf8(const char *latin1_input, size_t len,
+                            char *utf8_output) {
+  const char *end = latin1_input + len;
+  const __m256i v_0000 = _mm256_setzero_si256();
+  const __m256i v_c080 = _mm256_set1_epi16((int16_t)0xc080);
+  const __m256i v_ff80 = _mm256_set1_epi16((int16_t)0xff80);
+  const size_t safety_margin = 12;
+
+  while (end - latin1_input >= std::ptrdiff_t(16 + safety_margin)) {
+    __m128i in8 = _mm_loadu_si128((__m128i *)latin1_input);
+    // a single 16-bit UTF-16 word can yield 1, 2 or 3 UTF-8 bytes
+    const __m128i v_80 = _mm_set1_epi8((char)0x80);
+    if (_mm_testz_si128(in8, v_80)) { // ASCII fast path!!!!
+      // 1. store (16 bytes)
+      _mm_storeu_si128((__m128i *)utf8_output, in8);
+      // 2. adjust pointers
+      latin1_input += 16;
+      utf8_output += 16;
+      continue; // we are done for this round!
+    }
+    // We proceed only with the first 16 bytes.
+    const __m256i in = _mm256_cvtepu8_epi16((in8));
+
+    // 1. prepare 2-byte values
+    // input 16-bit word : [0000|0000|aabb|bbbb] x 8
+    // expected output   : [1100|00aa|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 = [0000|00aa|bbbb|bb00]
+    const __m256i t0 = _mm256_slli_epi16(in, 2);
+    // t1 = [0000|00aa|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 = [1100|00aa|10bb|bbbb]
+    const __m256i t4 = _mm256_or_si256(t3, v_c080);
+
+    // 2. merge ASCII and 2-byte codewords
+
+    // 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));
+
+    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
+    latin1_input += 16;
+    continue;
+
+  } // while
+  return std::make_pair(latin1_input, utf8_output);
+}

package/deps/simdutf/src/haswell/avx2_convert_utf16_to_latin1.cpp

@@ -0,0 +1,83 @@
+template <endianness big_endian>
+std::pair<const char16_t *, char *>
+avx2_convert_utf16_to_latin1(const char16_t *buf, size_t len,
+                             char *latin1_output) {
+  const char16_t *end = buf + len;
+  while (end - buf >= 32) {
+    // Load 16 UTF-16 characters into 256-bit AVX2 register
+    __m256i in0 = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buf));
+    __m256i in1 =
+        _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buf + 16));
+
+    if simdutf_constexpr (!match_system(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);
+      in0 = _mm256_shuffle_epi8(in0, swap);
+      in1 = _mm256_shuffle_epi8(in1, swap);
+    }
+
+    __m256i high_byte_mask = _mm256_set1_epi16((int16_t)0xFF00);
+    if (_mm256_testz_si256(_mm256_or_si256(in0, in1), high_byte_mask)) {
+      // Pack 16-bit characters into 8-bit and store in latin1_output
+      const __m256i packed = _mm256_packus_epi16(in0, in1);
+
+      const __m256i result = _mm256_permute4x64_epi64(packed, 0b11011000);
+
+      _mm256_storeu_si256(reinterpret_cast<__m256i *>(latin1_output), result);
+      // Adjust pointers for the next iteration
+      buf += 32;
+      latin1_output += 32;
+    } else {
+      return std::make_pair(nullptr, reinterpret_cast<char *>(latin1_output));
+    }
+  } // while
+  return std::make_pair(buf, latin1_output);
+}
+
+template <endianness big_endian>
+std::pair<result, char *>
+avx2_convert_utf16_to_latin1_with_errors(const char16_t *buf, size_t len,
+                                         char *latin1_output) {
+  const char16_t *start = buf;
+  const char16_t *end = buf + len;
+  while (end - buf >= 16) {
+    __m256i in = _mm256_loadu_si256(reinterpret_cast<const __m256i *>(buf));
+
+    if simdutf_constexpr (!match_system(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);
+    }
+
+    __m256i high_byte_mask = _mm256_set1_epi16((int16_t)0xFF00);
+    if (_mm256_testz_si256(in, high_byte_mask)) {
+      __m128i lo = _mm256_extractf128_si256(in, 0);
+      __m128i hi = _mm256_extractf128_si256(in, 1);
+      __m128i latin1_packed_lo = _mm_packus_epi16(lo, lo);
+      __m128i latin1_packed_hi = _mm_packus_epi16(hi, hi);
+      _mm_storel_epi64(reinterpret_cast<__m128i *>(latin1_output),
+                       latin1_packed_lo);
+      _mm_storel_epi64(reinterpret_cast<__m128i *>(latin1_output + 8),
+                       latin1_packed_hi);
+      buf += 16;
+      latin1_output += 16;
+    } else {
+      // Fallback to scalar code for handling errors
+      for (int k = 0; k < 16; k++) {
+        uint16_t word = scalar::utf16::swap_if_needed<big_endian>(buf[k]);
+        if (word <= 0xff) {
+          *latin1_output++ = char(word);
+        } else {
+          return std::make_pair(
+              result{error_code::TOO_LARGE, (size_t)(buf - start + k)},
+              latin1_output);
+        }
+      }
+      buf += 16;
+    }
+  } // while
+  return std::make_pair(result{error_code::SUCCESS, (size_t)(buf - start)},
+                        latin1_output);
+}

package/deps/simdutf/src/haswell/avx2_convert_utf16_to_utf32.cpp

@@ -0,0 +1,210 @@
+/*
+    The vectorized algorithm works on single SSE register i.e., it
+    loads eight 16-bit code units.
+
+    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
+       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 code unit
+    can be converted into: 1) single UTF8 byte (when it is 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 code units, 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 (end - buf >= 16) {
+    __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 code units to sixteen 32-bit code
+      // units
+      _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 = scalar::utf16::swap_if_needed<big_endian>(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 =
+              scalar::utf16::swap_if_needed<big_endian>(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 (end - buf >= 16) {
+    __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 code units to sixteen 32-bit code
+      // units
+      _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 = scalar::utf16::swap_if_needed<big_endian>(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 =
+              scalar::utf16::swap_if_needed<big_endian>(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);
+}