react-native-quick-crypto 1.0.18 → 1.1.0

package/deps/simdutf/src/haswell/avx2_convert_utf8_to_utf16.cpp

@@ -0,0 +1,195 @@
+// 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 == 0xfff) {
+    // We process the data in chunks of 12 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 += 12; // We wrote 12 16-bit characters.
+    return 12;          // We consumed 12 bytes.
+  }
+  if (((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) {
+    // We want to take 8 2-byte UTF-8 code units and turn them into 8 2-byte
+    // UTF-16 code units. 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 code units and turn them into 4 2-byte
+    // UTF-16 code units. 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-code units
+    // this is a relatively easy scenario
+    // we process SIX (6) input code-code units. The max length in bytes of six
+    // code code units 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-code units
+    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-code units
+    //////////////
+    // 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;
+}

package/deps/simdutf/src/haswell/avx2_convert_utf8_to_utf32.cpp

@@ -0,0 +1,135 @@
+// 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 == 0xfff) {
+    // We process the data in chunks of 12 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 += 12; // We wrote 12 32-bit characters.
+    return 12;          // We consumed 12 bytes.
+  }
+  if (((utf8_end_of_code_point_mask & 0xffff) == 0xaaaa)) {
+    // We want to take 8 2-byte UTF-8 code units and turn them into 8 4-byte
+    // UTF-32 code units. 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 code units and turn them into 4 4-byte
+    // UTF-32 code units. 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-code units
+    // this is a relatively easy scenario
+    // we process SIX (6) input code-code units. The max length in bytes of six
+    // code code units 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-code units
+    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-code units
+    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;
+}

package/deps/simdutf/src/haswell/avx2_utf16fix.cpp

@@ -0,0 +1,173 @@
+/*
+ * Process one block of 16 characters.  If in_place is false,
+ * copy the block from in to out.  If there is a sequencing
+ * error in the block, overwrite the illsequenced characters
+ * with the replacement character.  This function reads one
+ * character before the beginning of the buffer as a lookback.
+ * If that character is illsequenced, it too is overwritten.
+ */
+template <endianness big_endian, bool in_place>
+void utf16fix_block(char16_t *out, const char16_t *in) {
+  auto swap_if_needed = [](uint16_t x) simdutf_constexpr -> uint16_t {
+    return scalar::utf16::swap_if_needed<big_endian>(x);
+  };
+  const char16_t replacement = scalar::utf16::replacement<big_endian>();
+  __m256i lookback, block, lb_masked, block_masked, lb_is_high, block_is_low;
+  __m256i illseq, lb_illseq, block_illseq, lb_illseq_shifted;
+
+  lookback = _mm256_loadu_si256((const __m256i *)(in - 1));
+  block = _mm256_loadu_si256((const __m256i *)in);
+  lb_masked =
+      _mm256_and_si256(lookback, _mm256_set1_epi16(swap_if_needed(0xfc00u)));
+  block_masked =
+      _mm256_and_si256(block, _mm256_set1_epi16(swap_if_needed(0xfc00u)));
+  lb_is_high =
+      _mm256_cmpeq_epi16(lb_masked, _mm256_set1_epi16(swap_if_needed(0xd800u)));
+  block_is_low = _mm256_cmpeq_epi16(block_masked,
+                                    _mm256_set1_epi16(swap_if_needed(0xdc00u)));
+
+  illseq = _mm256_xor_si256(lb_is_high, block_is_low);
+  if (!_mm256_testz_si256(illseq, illseq)) {
+    int lb;
+
+    /* compute the cause of the illegal sequencing */
+    lb_illseq = _mm256_andnot_si256(block_is_low, lb_is_high);
+#if SIMDUTF_GCC9OROLDER
+    // Old GCC versions are missing _mm256_zextsi128_si256, so we emulate it.
+    __m128i tmp_legacygcc =
+        _mm_bslli_si128(_mm256_extracti128_si256(lb_illseq, 1), 14);
+    __m256i tmp_legacygcc256 =
+        _mm256_set_m128i(_mm_setzero_si128(), tmp_legacygcc);
+    lb_illseq_shifted =
+        _mm256_or_si256(_mm256_bsrli_epi128(lb_illseq, 2), tmp_legacygcc256);
+#else
+    lb_illseq_shifted =
+        _mm256_or_si256(_mm256_bsrli_epi128(lb_illseq, 2),
+                        _mm256_zextsi128_si256(_mm_bslli_si128(
+                            _mm256_extracti128_si256(lb_illseq, 1), 14)));
+#endif // SIMDUTF_GCC9OROLDER
+    block_illseq = _mm256_or_si256(
+        _mm256_andnot_si256(lb_is_high, block_is_low), lb_illseq_shifted);
+
+    /* fix illegal sequencing in the lookback */
+#if SIMDUTF_GCC10 || SIMDUTF_GCC9OROLDER
+    // GCC 10 is missing important intrinsics.
+    lb = _mm_cvtsi128_si32(_mm256_extractf128_si256(lb_illseq, 0));
+#else
+    lb = _mm256_cvtsi256_si32(lb_illseq);
+#endif
+    lb = (lb & replacement) | (~lb & out[-1]);
+    out[-1] = char16_t(lb);
+
+    /* fix illegal sequencing in the main block */
+    block =
+        _mm256_blendv_epi8(block, _mm256_set1_epi16(replacement), block_illseq);
+    _mm256_storeu_si256((__m256i *)out, block);
+  } else if (!in_place) {
+    _mm256_storeu_si256((__m256i *)out, block);
+  }
+}
+
+template <endianness big_endian, bool in_place>
+void utf16fix_block_sse(char16_t *out, const char16_t *in) {
+  auto swap_if_needed = [](uint16_t x) simdutf_constexpr -> uint16_t {
+    return scalar::utf16::swap_if_needed<big_endian>(x);
+  };
+  const char16_t replacement = scalar::utf16::replacement<big_endian>();
+  __m128i lookback, block, lb_masked, block_masked, lb_is_high, block_is_low;
+  __m128i illseq, lb_illseq, block_illseq;
+
+  lookback = _mm_loadu_si128((const __m128i *)(in - 1));
+  block = _mm_loadu_si128((const __m128i *)in);
+  lb_masked = _mm_and_si128(lookback, _mm_set1_epi16(swap_if_needed(0xfc00U)));
+  block_masked = _mm_and_si128(block, _mm_set1_epi16(swap_if_needed(0xfc00U)));
+  lb_is_high =
+      _mm_cmpeq_epi16(lb_masked, _mm_set1_epi16(swap_if_needed(0xd800U)));
+  block_is_low =
+      _mm_cmpeq_epi16(block_masked, _mm_set1_epi16(swap_if_needed(0xdc00U)));
+
+  illseq = _mm_xor_si128(lb_is_high, block_is_low);
+  if (_mm_movemask_epi8(illseq) != 0) {
+    /* compute the cause of the illegal sequencing */
+    lb_illseq = _mm_andnot_si128(block_is_low, lb_is_high);
+    block_illseq = _mm_or_si128(_mm_andnot_si128(lb_is_high, block_is_low),
+                                _mm_bsrli_si128(lb_illseq, 2));
+    /* fix illegal sequencing in the lookback */
+    int lb = _mm_cvtsi128_si32(lb_illseq);
+    lb = (lb & replacement) | (~lb & out[-1]);
+    out[-1] = char16_t(lb);
+    /* fix illegal sequencing in the main block */
+    block =
+        _mm_or_si128(_mm_andnot_si128(block_illseq, block),
+                     _mm_and_si128(block_illseq, _mm_set1_epi16(replacement)));
+    _mm_storeu_si128((__m128i *)out, block);
+  } else if (!in_place) {
+    _mm_storeu_si128((__m128i *)out, block);
+  }
+}
+
+template <endianness big_endian>
+void utf16fix_sse(const char16_t *in, size_t n, char16_t *out) {
+  const char16_t replacement = scalar::utf16::replacement<big_endian>();
+  size_t i;
+
+  if (n < 9) {
+    scalar::utf16::to_well_formed_utf16<big_endian>(in, n, out);
+    return;
+  }
+
+  out[0] =
+      scalar::utf16::is_low_surrogate<big_endian>(in[0]) ? replacement : in[0];
+
+  /* duplicate code to have the compiler specialise utf16fix_block() */
+  if (in == out) {
+    for (i = 1; i + 8 < n; i += 8) {
+      utf16fix_block_sse<big_endian, true>(out + i, in + i);
+    }
+
+    utf16fix_block_sse<big_endian, true>(out + n - 8, in + n - 8);
+  } else {
+    for (i = 1; i + 8 < n; i += 8) {
+      utf16fix_block_sse<big_endian, false>(out + i, in + i);
+    }
+
+    utf16fix_block_sse<big_endian, false>(out + n - 8, in + n - 8);
+  }
+
+  out[n - 1] = scalar::utf16::is_high_surrogate<big_endian>(out[n - 1])
+                   ? replacement
+                   : out[n - 1];
+}
+
+template <endianness big_endian>
+void utf16fix_avx(const char16_t *in, size_t n, char16_t *out) {
+  const char16_t replacement = scalar::utf16::replacement<big_endian>();
+  size_t i;
+
+  if (n < 17) {
+    utf16fix_sse<big_endian>(in, n, out);
+    return;
+  }
+
+  out[0] =
+      scalar::utf16::is_low_surrogate<big_endian>(in[0]) ? replacement : in[0];
+
+  /* duplicate code to have the compiler specialise utf16fix_block() */
+  if (in == out) {
+    for (i = 1; i + 16 < n; i += 16) {
+      utf16fix_block<big_endian, true>(out + i, in + i);
+    }
+
+    utf16fix_block<big_endian, true>(out + n - 16, in + n - 16);
+  } else {
+    for (i = 1; i + 16 < n; i += 16) {
+      utf16fix_block<big_endian, false>(out + i, in + i);
+    }
+
+    utf16fix_block<big_endian, false>(out + n - 16, in + n - 16);
+  }
+
+  out[n - 1] = scalar::utf16::is_high_surrogate<big_endian>(out[n - 1])
+                   ? replacement
+                   : out[n - 1];
+}

package/deps/simdutf/src/haswell/avx2_validate_utf16.cpp

@@ -0,0 +1,17 @@
+template <endianness big_endian>
+simd8<uint8_t> utf16_gather_high_bytes(const simd16<uint16_t> &in0,
+                                       const simd16<uint16_t> &in1) {
+  if (big_endian) {
+    // we want lower bytes
+    const auto mask = simd16<uint16_t>(0x00ff);
+    const auto t0 = in0 & mask;
+    const auto t1 = in1 & mask;
+
+    return simd16<uint16_t>::pack(t0, t1);
+  } else {
+    const auto t0 = in0.shr<8>();
+    const auto t1 = in1.shr<8>();
+
+    return simd16<uint16_t>::pack(t0, t1);
+  }
+}