@ohos-ports/react-native-quick-crypto 0.0.0-stage → 1.1.8-beta.0

package/deps/simdutf/src/icelake/icelake_convert_latin1_to_utf16.inl.cpp

@@ -0,0 +1,36 @@
+// file included directly
+template <endianness big_endian>
+size_t icelake_convert_latin1_to_utf16(const char *latin1_input, size_t len,
+                                       char16_t *utf16_output) {
+  size_t rounded_len = len & ~0x1F; // Round down to nearest multiple of 32
+
+  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
+  for (size_t i = 0; i < rounded_len; i += 32) {
+    // Load 32 Latin1 characters into a 256-bit register
+    __m256i in = _mm256_loadu_si256((__m256i *)&latin1_input[i]);
+    // Zero extend each set of 8 Latin1 characters to 32 16-bit integers
+    __m512i out = _mm512_cvtepu8_epi16(in);
+    if (big_endian) {
+      out = _mm512_shuffle_epi8(out, byteflip);
+    }
+    // Store the results back to memory
+    _mm512_storeu_si512((__m512i *)&utf16_output[i], out);
+  }
+  if (rounded_len != len) {
+    uint32_t mask = uint32_t(1 << (len - rounded_len)) - 1;
+    __m256i in = _mm256_maskz_loadu_epi8(mask, latin1_input + rounded_len);
+
+    // Zero extend each set of 8 Latin1 characters to 32 16-bit integers
+    __m512i out = _mm512_cvtepu8_epi16(in);
+    if (big_endian) {
+      out = _mm512_shuffle_epi8(out, byteflip);
+    }
+    // Store the results back to memory
+    _mm512_mask_storeu_epi16(utf16_output + rounded_len, mask, out);
+  }
+
+  return len;
+}

package/deps/simdutf/src/icelake/icelake_convert_latin1_to_utf32.inl.cpp

@@ -0,0 +1,23 @@
+void avx512_convert_latin1_to_utf32(const char *buf, size_t len,
+                                    char32_t *utf32_output) {
+  while (len >= 16) {
+    // Load 16 Latin1 characters into a 128-bit register
+    __m128i in = _mm_loadu_si128((__m128i *)buf);
+
+    // Zero extend each set of 8 Latin1 characters to 16 32-bit integers using
+    // vpmovzxbd
+    __m512i out = _mm512_cvtepu8_epi32(in);
+
+    // Store the results back to memory
+    _mm512_storeu_si512((__m512i *)utf32_output, out);
+
+    len -= 16;
+    buf += 16;
+    utf32_output += 16;
+  }
+
+  __mmask16 mask = __mmask16((1 << len) - 1);
+  __m128i in = _mm_maskz_loadu_epi8(mask, buf);
+  __m512i out = _mm512_cvtepu8_epi32(in);
+  _mm512_mask_storeu_epi32((__m512i *)utf32_output, mask, out);
+}

package/deps/simdutf/src/icelake/icelake_convert_latin1_to_utf8.inl.cpp

@@ -0,0 +1,107 @@
+// file included directly
+
+static inline size_t latin1_to_utf8_avx512_vec(__m512i input, size_t input_len,
+                                               char *utf8_output,
+                                               int mask_output) {
+  __mmask64 nonascii = _mm512_movepi8_mask(input);
+  size_t output_size = input_len + (size_t)count_ones(nonascii);
+
+  // Mask to denote whether the byte is a leading byte that is not ascii
+  __mmask64 sixth = _mm512_cmpge_epu8_mask(
+      input, _mm512_set1_epi8(-64)); // binary representation of -64: 1100 0000
+
+  const uint64_t alternate_bits = UINT64_C(0x5555555555555555);
+  uint64_t ascii = ~nonascii;
+  // the bits in ascii are inverted and zeros are interspersed in between them
+  uint64_t maskA = ~_pdep_u64(ascii, alternate_bits);
+  uint64_t maskB = ~_pdep_u64(ascii >> 32, alternate_bits);
+
+  // interleave bytes from top and bottom halves (abcd...ABCD -> aAbBcCdD)
+  __m512i input_interleaved = _mm512_permutexvar_epi8(
+      _mm512_set_epi32(0x3f1f3e1e, 0x3d1d3c1c, 0x3b1b3a1a, 0x39193818,
+                       0x37173616, 0x35153414, 0x33133212, 0x31113010,
+                       0x2f0f2e0e, 0x2d0d2c0c, 0x2b0b2a0a, 0x29092808,
+                       0x27072606, 0x25052404, 0x23032202, 0x21012000),
+      input);
+
+  // double size of each byte, and insert the leading byte 1100 0010
+
+  /*
+  upscale the bytes to 16-bit value, adding the 0b11000000 leading byte in the
+  process. We adjust for the bytes that have their two most significant bits.
+  This takes care of the first 32 bytes, assuming we interleaved the bytes. */
+  __m512i outputA =
+      _mm512_shldi_epi16(input_interleaved, _mm512_set1_epi8(-62), 8);
+  outputA = _mm512_mask_add_epi16(
+      outputA, (__mmask32)sixth, outputA,
+      _mm512_set1_epi16(1 - 0x4000)); // 1- 0x4000 = 1100 0000 0000 0001????
+
+  // in the second 32-bit half, set first or second option based on whether
+  // original input is leading byte (second case) or not (first case)
+  __m512i leadingB =
+      _mm512_mask_blend_epi16((__mmask32)(sixth >> 32),
+                              _mm512_set1_epi16(0x00c2),  // 0000 0000 1101 0010
+                              _mm512_set1_epi16(0x40c3)); // 0100 0000 1100 0011
+  __m512i outputB = _mm512_ternarylogic_epi32(
+      input_interleaved, leadingB, _mm512_set1_epi16((short)0xff00),
+      (240 & 170) ^ 204); // (input_interleaved & 0xff00) ^ leadingB
+
+  // prune redundant bytes
+  outputA = _mm512_maskz_compress_epi8(maskA, outputA);
+  outputB = _mm512_maskz_compress_epi8(maskB, outputB);
+
+  size_t output_sizeA = (size_t)count_ones((uint32_t)nonascii) + 32;
+
+  if (mask_output) {
+    if (input_len > 32) { // is the second half of the input vector used?
+      __mmask64 write_mask = _bzhi_u64(~0ULL, (unsigned int)output_sizeA);
+      _mm512_mask_storeu_epi8(utf8_output, write_mask, outputA);
+      utf8_output += output_sizeA;
+      write_mask = _bzhi_u64(~0ULL, (unsigned int)(output_size - output_sizeA));
+      _mm512_mask_storeu_epi8(utf8_output, write_mask, outputB);
+    } else {
+      __mmask64 write_mask = _bzhi_u64(~0ULL, (unsigned int)output_size);
+      _mm512_mask_storeu_epi8(utf8_output, write_mask, outputA);
+    }
+  } else {
+    _mm512_storeu_si512(utf8_output, outputA);
+    utf8_output += output_sizeA;
+    _mm512_storeu_si512(utf8_output, outputB);
+  }
+  return output_size;
+}
+
+static inline size_t latin1_to_utf8_avx512_branch(__m512i input,
+                                                  char *utf8_output) {
+  __mmask64 nonascii = _mm512_movepi8_mask(input);
+  if (nonascii) {
+    return latin1_to_utf8_avx512_vec(input, 64, utf8_output, 0);
+  } else {
+    _mm512_storeu_si512(utf8_output, input);
+    return 64;
+  }
+}
+
+size_t latin1_to_utf8_avx512_start(const char *buf, size_t len,
+                                   char *utf8_output) {
+  char *start = utf8_output;
+  size_t pos = 0;
+  // if there's at least 128 bytes remaining, we don't need to mask the output
+  for (; pos + 128 <= len; pos += 64) {
+    __m512i input = _mm512_loadu_si512((__m512i *)(buf + pos));
+    utf8_output += latin1_to_utf8_avx512_branch(input, utf8_output);
+  }
+  // in the last 128 bytes, the first 64 may require masking the output
+  if (pos + 64 <= len) {
+    __m512i input = _mm512_loadu_si512((__m512i *)(buf + pos));
+    utf8_output += latin1_to_utf8_avx512_vec(input, 64, utf8_output, 1);
+    pos += 64;
+  }
+  // with the last 64 bytes, the input also needs to be masked
+  if (pos < len) {
+    __mmask64 load_mask = _bzhi_u64(~0ULL, (unsigned int)(len - pos));
+    __m512i input = _mm512_maskz_loadu_epi8(load_mask, (__m512i *)(buf + pos));
+    utf8_output += latin1_to_utf8_avx512_vec(input, len - pos, utf8_output, 1);
+  }
+  return (size_t)(utf8_output - start);
+}

package/deps/simdutf/src/icelake/icelake_convert_utf16_to_latin1.inl.cpp

@@ -0,0 +1,103 @@
+// file included directly
+template <endianness big_endian>
+size_t icelake_convert_utf16_to_latin1(const char16_t *buf, size_t len,
+                                       char *latin1_output) {
+  const char16_t *end = buf + len;
+  __m512i v_0xFF = _mm512_set1_epi16(0xff);
+  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
+  __m512i shufmask = _mm512_set_epi8(
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 62, 60, 58, 56, 54, 52, 50, 48, 46, 44, 42, 40, 38,
+      36, 34, 32, 30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
+  while (end - buf >= 32) {
+    __m512i in = _mm512_loadu_si512((__m512i *)buf);
+    if (big_endian) {
+      in = _mm512_shuffle_epi8(in, byteflip);
+    }
+    if (_mm512_cmpgt_epu16_mask(in, v_0xFF)) {
+      return 0;
+    }
+    _mm256_storeu_si256(
+        (__m256i *)latin1_output,
+        _mm512_castsi512_si256(_mm512_permutexvar_epi8(shufmask, in)));
+    latin1_output += 32;
+    buf += 32;
+  }
+  if (buf < end) {
+    uint32_t mask(uint32_t(1 << (end - buf)) - 1);
+    __m512i in = _mm512_maskz_loadu_epi16(mask, buf);
+    if (big_endian) {
+      in = _mm512_shuffle_epi8(in, byteflip);
+    }
+    if (_mm512_cmpgt_epu16_mask(in, v_0xFF)) {
+      return 0;
+    }
+    _mm256_mask_storeu_epi8(
+        latin1_output, mask,
+        _mm512_castsi512_si256(_mm512_permutexvar_epi8(shufmask, in)));
+  }
+  return len;
+}
+
+template <endianness big_endian>
+std::pair<result, char *>
+icelake_convert_utf16_to_latin1_with_errors(const char16_t *buf, size_t len,
+                                            char *latin1_output) {
+  const char16_t *end = buf + len;
+  const char16_t *start = buf;
+  __m512i byteflip = _mm512_setr_epi64(0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809,
+                                       0x0607040502030001, 0x0e0f0c0d0a0b0809);
+  __m512i v_0xFF = _mm512_set1_epi16(0xff);
+  __m512i shufmask = _mm512_set_epi8(
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 62, 60, 58, 56, 54, 52, 50, 48, 46, 44, 42, 40, 38,
+      36, 34, 32, 30, 28, 26, 24, 22, 20, 18, 16, 14, 12, 10, 8, 6, 4, 2, 0);
+  while (end - buf >= 32) {
+    __m512i in = _mm512_loadu_si512((__m512i *)buf);
+    if (big_endian) {
+      in = _mm512_shuffle_epi8(in, byteflip);
+    }
+    if (_mm512_cmpgt_epu16_mask(in, v_0xFF)) {
+      uint16_t word;
+      while ((word = scalar::utf16::swap_if_needed<big_endian>(
+                  uint16_t(*buf))) <= 0xff) {
+        *latin1_output++ = uint8_t(word);
+        buf++;
+      }
+      return std::make_pair(result(error_code::TOO_LARGE, buf - start),
+                            latin1_output);
+    }
+    _mm256_storeu_si256(
+        (__m256i *)latin1_output,
+        _mm512_castsi512_si256(_mm512_permutexvar_epi8(shufmask, in)));
+    latin1_output += 32;
+    buf += 32;
+  }
+  if (buf < end) {
+    uint32_t mask(uint32_t(1 << (end - buf)) - 1);
+    __m512i in = _mm512_maskz_loadu_epi16(mask, buf);
+    if (big_endian) {
+      in = _mm512_shuffle_epi8(in, byteflip);
+    }
+    if (_mm512_cmpgt_epu16_mask(in, v_0xFF)) {
+
+      uint16_t word;
+      while ((word = scalar::utf16::swap_if_needed<big_endian>(
+                  uint16_t(*buf))) <= 0xff) {
+        *latin1_output++ = uint8_t(word);
+        buf++;
+      }
+      return std::make_pair(result(error_code::TOO_LARGE, buf - start),
+                            latin1_output);
+    }
+    _mm256_mask_storeu_epi8(
+        latin1_output, mask,
+        _mm512_castsi512_si256(_mm512_permutexvar_epi8(shufmask, in)));
+  }
+  return std::make_pair(result(error_code::SUCCESS, len), latin1_output);
+}

package/deps/simdutf/src/icelake/icelake_convert_utf16_to_utf32.inl.cpp

@@ -0,0 +1,136 @@
+// 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 (std::distance(buf, end) >= 32) {
+    // 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 code units to 32-bit code units
+            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 code units (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 code units, 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 code units to thirty-two 32-bit code units
+      _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);
+}

package/deps/simdutf/src/icelake/icelake_convert_utf16_to_utf8.inl.cpp

@@ -0,0 +1,206 @@
+// 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 code units
+ * 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;
+  int 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 & ((1U << 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((1U << inlen) - 1);
+    in = _mm512_maskz_loadu_epi16(inmask, inbuf);
+    if (big_endian) {
+      in = _mm512_shuffle_epi8(in, byteflip);
+    }
+    adjust = (int)inlen - 31;
+    inlen = 0;
+    goto lastiteration;
+  }
+  *outlen = (outbuf - outbuf_orig) + adjust;
+  return ((inbuf - inbuf_orig) + adjust);
+}

package/deps/simdutf/src/icelake/icelake_convert_utf32_to_latin1.inl.cpp

@@ -0,0 +1,74 @@
+// file included directly
+size_t icelake_convert_utf32_to_latin1(const char32_t *buf, size_t len,
+                                       char *latin1_output) {
+  const char32_t *end = buf + len;
+  __m512i v_0xFF = _mm512_set1_epi32(0xff);
+  __m512i shufmask = _mm512_set_epi8(
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 60,
+      56, 52, 48, 44, 40, 36, 32, 28, 24, 20, 16, 12, 8, 4, 0);
+  while (end - buf >= 16) {
+    __m512i in = _mm512_loadu_si512((__m512i *)buf);
+    if (_mm512_cmpgt_epu32_mask(in, v_0xFF)) {
+      return 0;
+    }
+    _mm_storeu_si128(
+        (__m128i *)latin1_output,
+        _mm512_castsi512_si128(_mm512_permutexvar_epi8(shufmask, in)));
+    latin1_output += 16;
+    buf += 16;
+  }
+  if (buf < end) {
+    uint16_t mask = uint16_t((1 << (end - buf)) - 1);
+    __m512i in = _mm512_maskz_loadu_epi32(mask, buf);
+    if (_mm512_cmpgt_epu32_mask(in, v_0xFF)) {
+      return 0;
+    }
+    _mm_mask_storeu_epi8(
+        latin1_output, mask,
+        _mm512_castsi512_si128(_mm512_permutexvar_epi8(shufmask, in)));
+  }
+  return len;
+}
+
+std::pair<result, char *>
+icelake_convert_utf32_to_latin1_with_errors(const char32_t *buf, size_t len,
+                                            char *latin1_output) {
+  const char32_t *end = buf + len;
+  const char32_t *start = buf;
+  __m512i v_0xFF = _mm512_set1_epi32(0xff);
+  __m512i shufmask = _mm512_set_epi8(
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 60,
+      56, 52, 48, 44, 40, 36, 32, 28, 24, 20, 16, 12, 8, 4, 0);
+  while (end - buf >= 16) {
+    __m512i in = _mm512_loadu_si512((__m512i *)buf);
+    if (_mm512_cmpgt_epu32_mask(in, v_0xFF)) {
+      while (uint32_t(*buf) <= 0xff) {
+        *latin1_output++ = uint8_t(*buf++);
+      }
+      return std::make_pair(result(error_code::TOO_LARGE, buf - start),
+                            latin1_output);
+    }
+    _mm_storeu_si128(
+        (__m128i *)latin1_output,
+        _mm512_castsi512_si128(_mm512_permutexvar_epi8(shufmask, in)));
+    latin1_output += 16;
+    buf += 16;
+  }
+  if (buf < end) {
+    uint16_t mask = uint16_t((1 << (end - buf)) - 1);
+    __m512i in = _mm512_maskz_loadu_epi32(mask, buf);
+    if (_mm512_cmpgt_epu32_mask(in, v_0xFF)) {
+      while (uint32_t(*buf) <= 0xff) {
+        *latin1_output++ = uint8_t(*buf++);
+      }
+      return std::make_pair(result(error_code::TOO_LARGE, buf - start),
+                            latin1_output);
+    }
+    _mm_mask_storeu_epi8(
+        latin1_output, mask,
+        _mm512_castsi512_si128(_mm512_permutexvar_epi8(shufmask, in)));
+  }
+  return std::make_pair(result(error_code::SUCCESS, len), latin1_output);
+}