@hxa-rn/react-native-quick-crypto 1.1.0-beta.1

package/harmony/quick_crypto/deps/simdutf/src/arm64/arm_convert_utf8_to_utf16.cpp

@@ -0,0 +1,313 @@
+// Convert up to 16 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 16, usually 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.
+  //
+  uint8x16_t in = vld1q_u8(reinterpret_cast<const uint8_t *>(input));
+  const uint16_t input_utf8_end_of_code_point_mask =
+      utf8_end_of_code_point_mask & 0xfff;
+  //
+  // 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.
+  // The obvious first test is ASCII, which actually consumes the full 16.
+  if ((utf8_end_of_code_point_mask & 0xFFFF) == 0xffff) {
+    // We process in chunks of 16 bytes
+    // The routine in simd.h is reused.
+    simd8<int8_t> temp{vreinterpretq_s8_u8(in)};
+    temp.store_ascii_as_utf16<big_endian>(utf16_output);
+    utf16_output += 16; // We wrote 16 16-bit characters.
+    return 16;          // We consumed 16 bytes.
+  }
+
+  // 3 byte sequences are the next most common, as seen in CJK, which has long
+  // sequences of these.
+  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.
+    uint16x4_t composed = convert_utf8_3_byte_to_utf16(in);
+    // Byte swap if necessary
+    if simdutf_constexpr (!match_system(big_endian)) {
+      composed = vreinterpret_u16_u8(vrev16_u8(vreinterpret_u8_u16(composed)));
+    }
+    vst1_u16(reinterpret_cast<uint16_t *>(utf16_output), composed);
+    utf16_output += 4; // We wrote 4 16-bit characters.
+    return 12;         // We consumed 12 bytes.
+  }
+
+  // 2 byte sequences occur in short bursts in languages like Greek and Russian.
+  if ((utf8_end_of_code_point_mask & 0xFFF) == 0xaaa) {
+    // We want to take 6 2-byte UTF-8 code units and turn them into 6 2-byte
+    // UTF-16 code units.
+    uint16x8_t composed = convert_utf8_2_byte_to_utf16(in);
+    // Byte swap if necessary
+    if simdutf_constexpr (!match_system(big_endian)) {
+      composed =
+          vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(composed)));
+    }
+    vst1q_u16(reinterpret_cast<uint16_t *>(utf16_output), composed);
+
+    utf16_output += 6; // We wrote 6 16-bit characters.
+    return 12;         // We consumed 12 bytes.
+  }
+
+  /// We do not have a fast path available, or the fast path is unimportant, so
+  /// we fallback.
+  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
+    // Convert to UTF-16
+    uint16x8_t composed = convert_utf8_1_to_2_byte_to_utf16(in, idx);
+    // Byte swap if necessary
+    if simdutf_constexpr (!match_system(big_endian)) {
+      composed =
+          vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(composed)));
+    }
+    // Store
+    vst1q_u16(reinterpret_cast<uint16_t *>(utf16_output), composed);
+    utf16_output += 6; // We wrote 6 16-bit characters.
+    return consumed;
+  } else if (idx < 145) {
+    // FOUR (4) input code-code units
+    // UTF-16 and UTF-32 use similar algorithms, but UTF-32 skips the narrowing.
+    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t *>(
+        simdutf::tables::utf8_to_utf16::shufutf8[idx]));
+    // XXX: depending on the system scalar instructions might be faster.
+    // 1 byte: 00000000 00000000 0ccccccc
+    // 2 byte: 00000000 110bbbbb 10cccccc
+    // 3 byte: 1110aaaa 10bbbbbb 10cccccc
+    uint32x4_t perm = vreinterpretq_u32_u8(vqtbl1q_u8(in, sh));
+    // 1 byte: 00000000 0ccccccc
+    // 2 byte: xx0bbbbb x0cccccc
+    // 3 byte: xxbbbbbb x0cccccc
+    uint16x4_t lowperm = vmovn_u32(perm);
+    // Partially mask with bic (doesn't require a temporary register unlike and)
+    // The shift left insert below will clear the top bits.
+    // 1 byte: 00000000 00000000
+    // 2 byte: xx0bbbbb 00000000
+    // 3 byte: xxbbbbbb 00000000
+    uint16x4_t middlebyte = vbic_u16(lowperm, vmov_n_u16(uint16_t(~0xFF00)));
+    // ASCII
+    // 1 byte: 00000000 0ccccccc
+    // 2+byte: 00000000 00cccccc
+    uint16x4_t ascii = vand_u16(lowperm, vmov_n_u16(0x7F));
+    // Split into narrow vectors.
+    // 2 byte: 00000000 00000000
+    // 3 byte: 00000000 xxxxaaaa
+    uint16x4_t highperm = vshrn_n_u32(perm, 16);
+    // Shift right accumulate the middle byte
+    // 1 byte: 00000000 0ccccccc
+    // 2 byte: 00xx0bbb bbcccccc
+    // 3 byte: 00xxbbbb bbcccccc
+    uint16x4_t middlelow = vsra_n_u16(ascii, middlebyte, 2);
+    // Shift left and insert the top 4 bits, overwriting the garbage
+    // 1 byte: 00000000 0ccccccc
+    // 2 byte: 00000bbb bbcccccc
+    // 3 byte: aaaabbbb bbcccccc
+    uint16x4_t composed = vsli_n_u16(middlelow, highperm, 12);
+    // Byte swap if necessary
+    if simdutf_constexpr (!match_system(big_endian)) {
+      composed = vreinterpret_u16_u8(vrev16_u8(vreinterpret_u8_u16(composed)));
+    }
+    vst1_u16(reinterpret_cast<uint16_t *>(utf16_output), composed);
+
+    utf16_output += 4; // We wrote 4 16-bit codepoints
+    return consumed;
+  } else if (idx < 209) {
+    // THREE (3) input code-code units
+    if (input_utf8_end_of_code_point_mask == 0x888) {
+      // We want to take 3 4-byte UTF-8 code units and turn them into 3 4-byte
+      // UTF-16 pairs. Generating surrogate pairs is a little tricky though, but
+      // it is easier when we can assume they are all pairs. This version does
+      // not use the LUT, but 4 byte sequences are less common and the overhead
+      // of the extra memory access is less important than the early branch
+      // overhead in shorter sequences.
+
+      // Swap byte pairs
+      // 10dddddd 10cccccc|10bbbbbb 11110aaa
+      // 10cccccc 10dddddd|11110aaa 10bbbbbb
+      uint8x16_t swap = vrev16q_u8(in);
+      // Shift left 2 bits
+      // cccccc00 dddddd00 xxxxxxxx bbbbbb00
+      uint32x4_t shift = vreinterpretq_u32_u8(vshlq_n_u8(swap, 2));
+      // Create a magic number containing the low 2 bits of the trail surrogate
+      // and all the corrections needed to create the pair. UTF-8 4b prefix   =
+      // -0x0000|0xF000 surrogate offset  = -0x0000|0x0040 (0x10000 << 6)
+      // surrogate high    = +0x0000|0xD800
+      // surrogate low     = +0xDC00|0x0000
+      // -------------------------------
+      //                   = +0xDC00|0xE7C0
+      uint32x4_t magic = vmovq_n_u32(0xDC00E7C0);
+      // Generate unadjusted trail surrogate minus lowest 2 bits
+      // xxxxxxxx xxxxxxxx|11110aaa bbbbbb00
+      uint32x4_t trail =
+          vbslq_u32(vmovq_n_u32(0x0000FF00), vreinterpretq_u32_u8(swap), shift);
+      // Insert low 2 bits of trail surrogate to magic number for later
+      // 11011100 00000000 11100111 110000cc
+      uint16x8_t magic_with_low_2 =
+          vreinterpretq_u16_u32(vsraq_n_u32(magic, shift, 30));
+      // Generate lead surrogate
+      // xxxxcccc ccdddddd|xxxxxxxx xxxxxxxx
+      uint32x4_t lead = vreinterpretq_u32_u16(
+          vsliq_n_u16(vreinterpretq_u16_u8(swap), vreinterpretq_u16_u8(in), 6));
+      // Mask out lead
+      // 000000cc ccdddddd|xxxxxxxx xxxxxxxx
+      lead = vbicq_u32(lead, vmovq_n_u32(uint32_t(~0x03FFFFFF)));
+      // Blend pairs
+      // 000000cc ccdddddd|11110aaa bbbbbb00
+      uint16x8_t blend = vreinterpretq_u16_u32(
+          vbslq_u32(vmovq_n_u32(0x0000FFFF), trail, lead));
+      // Add magic number to finish the result
+      // 110111CC CCDDDDDD|110110AA BBBBBBCC
+      uint16x8_t composed = vaddq_u16(blend, magic_with_low_2);
+      // Byte swap if necessary
+      if simdutf_constexpr (!match_system(big_endian)) {
+        composed =
+            vreinterpretq_u16_u8(vrev16q_u8(vreinterpretq_u8_u16(composed)));
+      }
+      uint16_t buffer[8];
+      vst1q_u16(reinterpret_cast<uint16_t *>(buffer), composed);
+      for (int k = 0; k < 6; k++) {
+        utf16_output[k] = buffer[k];
+      } // the loop might compiler to a couple of instructions.
+      // We need some validation. See
+      // https://github.com/simdutf/simdutf/pull/631
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+      uint8x16_t expected_mask = simdutf_make_uint8x16_t(
+          0xf8, 0xc0, 0xc0, 0xc0, 0xf8, 0xc0, 0xc0, 0xc0, 0xf8, 0xc0, 0xc0,
+          0xc0, 0x0, 0x0, 0x0, 0x0);
+#else
+      uint8x16_t expected_mask = {0xf8, 0xc0, 0xc0, 0xc0, 0xf8, 0xc0,
+                                  0xc0, 0xc0, 0xf8, 0xc0, 0xc0, 0xc0,
+                                  0x0,  0x0,  0x0,  0x0};
+#endif
+#ifdef SIMDUTF_REGULAR_VISUAL_STUDIO
+      uint8x16_t expected = simdutf_make_uint8x16_t(
+          0xf0, 0x80, 0x80, 0x80, 0xf0, 0x80, 0x80, 0x80, 0xf0, 0x80, 0x80,
+          0x80, 0x0, 0x0, 0x0, 0x0);
+#else
+      uint8x16_t expected = {0xf0, 0x80, 0x80, 0x80, 0xf0, 0x80, 0x80, 0x80,
+                             0xf0, 0x80, 0x80, 0x80, 0x0,  0x0,  0x0,  0x0};
+#endif
+      uint8x16_t check = vceqq_u8(vandq_u8(in, expected_mask), expected);
+      bool correct = (vminvq_u32(vreinterpretq_u32_u8(check)) == 0xFFFFFFFF);
+      // The validation is just three instructions and it is not on a critical
+      // path.
+      if (correct) {
+        utf16_output += 6; // We wrote 3 32-bit surrogate pairs.
+      }
+      return 12; // We consumed 12 bytes.
+    }
+    // 3 1-4 byte sequences
+    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t *>(
+        simdutf::tables::utf8_to_utf16::shufutf8[idx]));
+
+    // 1 byte: 00000000 00000000 00000000 0ddddddd
+    // 3 byte: 00000000 00000000 110ccccc 10dddddd
+    // 3 byte: 00000000 1110bbbb 10cccccc 10dddddd
+    // 4 byte: 11110aaa 10bbbbbb 10cccccc 10dddddd
+    uint32x4_t perm = vreinterpretq_u32_u8(vqtbl1q_u8(in, sh));
+    // added to fix issue https://github.com/simdutf/simdutf/issues/514
+    // We only want to write 2 * 16-bit code units when that is actually what we
+    // have. Unfortunately, we cannot trust the input. So it is possible to get
+    // 0xff as an input byte and it should not result in a surrogate pair. We
+    // need to check for that.
+    uint32_t permbuffer[4];
+    vst1q_u32(permbuffer, perm);
+    // Mask the low and middle bytes
+    // 00000000 00000000 00000000 0ddddddd
+    uint32x4_t ascii = vandq_u32(perm, vmovq_n_u32(0x7f));
+    // Because the surrogates need more work, the high surrogate is computed
+    // first.
+    uint32x4_t middlehigh = vshlq_n_u32(perm, 2);
+    // 00000000 00000000 00cccccc 00000000
+    uint32x4_t middlebyte = vandq_u32(perm, vmovq_n_u32(0x3F00));
+    // Start assembling the sequence. Since the 4th byte is in the same position
+    // as it would be in a surrogate and there is no dependency, shift left
+    // instead of right. 3 byte: 00000000 10bbbbxx xxxxxxxx xxxxxxxx 4 byte:
+    // 11110aaa bbbbbbxx xxxxxxxx xxxxxxxx
+    uint32x4_t ab = vbslq_u32(vmovq_n_u32(0xFF000000), perm, middlehigh);
+    // Top 16 bits contains the high ten bits of the surrogate pair before
+    // correction 3 byte: 00000000 10bbbbcc|cccc0000 00000000 4 byte: 11110aaa
+    // bbbbbbcc|cccc0000 00000000 - high 10 bits correct w/o correction
+    uint32x4_t abc =
+        vbslq_u32(vmovq_n_u32(0xFFFC0000), ab, vshlq_n_u32(middlebyte, 4));
+    // Combine the low 6 or 7 bits by a shift right accumulate
+    // 3 byte: 00000000 00000010|bbbbcccc ccdddddd - low 16 bits correct
+    // 4 byte: 00000011 110aaabb|bbbbcccc ccdddddd - low 10 bits correct w/o
+    // correction
+    uint32x4_t composed = vsraq_n_u32(ascii, abc, 6);
+    // After this is for surrogates
+    // Blend the low and high surrogates
+    // 4 byte: 11110aaa bbbbbbcc|bbbbcccc ccdddddd
+    uint32x4_t mixed = vbslq_u32(vmovq_n_u32(0xFFFF0000), abc, composed);
+    // Clear the upper 6 bits of the low surrogate. Don't clear the upper bits
+    // yet as 0x10000 was not subtracted from the codepoint yet. 4 byte:
+    // 11110aaa bbbbbbcc|000000cc ccdddddd
+    uint16x8_t masked_pair = vreinterpretq_u16_u32(
+        vbicq_u32(mixed, vmovq_n_u32(uint32_t(~0xFFFF03FF))));
+    // Correct the remaining UTF-8 prefix, surrogate offset, and add the
+    // surrogate prefixes in one magic 16-bit addition. similar magic number but
+    // without the continue byte adjust and halfword swapped UTF-8 4b prefix   =
+    // -0xF000|0x0000 surrogate offset  = -0x0040|0x0000 (0x10000 << 6)
+    // surrogate high    = +0xD800|0x0000
+    // surrogate low     = +0x0000|0xDC00
+    // -----------------------------------
+    //                   = +0xE7C0|0xDC00
+    uint16x8_t magic = vreinterpretq_u16_u32(vmovq_n_u32(0xE7C0DC00));
+    // 4 byte: 110110AA BBBBBBCC|110111CC CCDDDDDD - surrogate pair complete
+    uint32x4_t surrogates =
+        vreinterpretq_u32_u16(vaddq_u16(masked_pair, magic));
+    // If the high bit is 1 (s32 less than zero), this needs a surrogate pair
+    uint32x4_t is_pair = vcltzq_s32(vreinterpretq_s32_u32(perm));
+
+    // Select either the 4 byte surrogate pair or the 2 byte solo codepoint
+    // 3 byte: 0xxxxxxx xxxxxxxx|bbbbcccc ccdddddd
+    // 4 byte: 110110AA BBBBBBCC|110111CC CCDDDDDD
+    uint32x4_t selected = vbslq_u32(is_pair, surrogates, composed);
+    // Byte swap if necessary
+    if simdutf_constexpr (!match_system(big_endian)) {
+      selected =
+          vreinterpretq_u32_u8(vrev16q_u8(vreinterpretq_u8_u32(selected)));
+    }
+    // Attempting to shuffle and store would be complex, just scalarize.
+    uint32_t buffer[4];
+    vst1q_u32(buffer, selected);
+    // Test for the top bit of the surrogate mask. Remove due to issue 514
+    // const uint32_t SURROGATE_MASK = match_system(big_endian) ? 0x80000000 :
+    // 0x00800000;
+    for (size_t i = 0; i < 3; i++) {
+      // Surrogate
+      // Used to be if (buffer[i] & SURROGATE_MASK) {
+      // See discussion above.
+      // patch for issue https://github.com/simdutf/simdutf/issues/514
+      if ((permbuffer[i] & 0xf8000000) == 0xf0000000) {
+        utf16_output[0] = uint16_t(buffer[i] >> 16);
+        utf16_output[1] = uint16_t(buffer[i] & 0xFFFF);
+        utf16_output += 2;
+      } else {
+        utf16_output[0] = uint16_t(buffer[i] & 0xFFFF);
+        utf16_output++;
+      }
+    }
+    return consumed;
+  } else {
+    // here we know that there is an error but we do not handle errors
+    return 12;
+  }
+}

package/harmony/quick_crypto/deps/simdutf/src/arm64/arm_convert_utf8_to_utf32.cpp

@@ -0,0 +1,179 @@
+// 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_out) {
+  // 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.
+  //
+  uint32_t *&utf32_output = reinterpret_cast<uint32_t *&>(utf32_out);
+  uint8x16_t in = vld1q_u8(reinterpret_cast<const uint8_t *>(input));
+  const uint16_t input_utf8_end_of_code_point_mask =
+      utf8_end_of_code_point_mask & 0xFFF;
+  //
+  // 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.
+  if (utf8_end_of_code_point_mask == 0xfff) {
+    // We process in chunks of 12 bytes.
+    // use fast implementation in src/simdutf/arm64/simd.h
+    // Ideally the compiler can keep the tables in registers.
+    simd8<int8_t> temp{vreinterpretq_s8_u8(in)};
+    temp.store_ascii_as_utf32_tbl(utf32_out);
+    utf32_output += 12; // We wrote 12 32-bit characters.
+    return 12;          // We consumed 12 bytes.
+  }
+  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. Convert to UTF-16
+    uint16x4_t composed_utf16 = convert_utf8_3_byte_to_utf16(in);
+    // Zero extend and store via ST2 with a zero.
+    uint16x4x2_t interleaver = {{composed_utf16, vmov_n_u16(0)}};
+    vst2_u16(reinterpret_cast<uint16_t *>(utf32_output), interleaver);
+    utf32_output += 4; // We wrote 4 32-bit characters.
+    return 12;         // We consumed 12 bytes.
+  }
+
+  // 2 byte sequences occur in short bursts in languages like Greek and Russian.
+  if (input_utf8_end_of_code_point_mask == 0xaaa) {
+    // We want to take 6 2-byte UTF-8 code units and turn them into 6 4-byte
+    // UTF-32 code units. Convert to UTF-16
+    uint16x8_t composed_utf16 = convert_utf8_2_byte_to_utf16(in);
+    // Zero extend and store via ST2 with a zero.
+    uint16x8x2_t interleaver = {{composed_utf16, vmovq_n_u16(0)}};
+    vst2q_u16(reinterpret_cast<uint16_t *>(utf32_output), interleaver);
+    utf32_output += 6; // We wrote 6 32-bit characters.
+    return 12;         // We consumed 12 bytes.
+  }
+  /// Either no fast path or an unimportant fast path.
+
+  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
+    // Convert to UTF-16
+    uint16x8_t composed_utf16 = convert_utf8_1_to_2_byte_to_utf16(in, idx);
+    // Zero extend and store with ST2 and zero
+    uint16x8x2_t interleaver = {{composed_utf16, vmovq_n_u16(0)}};
+    vst2q_u16(reinterpret_cast<uint16_t *>(utf32_output), interleaver);
+    utf32_output += 6; // We wrote 6 32-bit characters.
+    return consumed;
+  } else if (idx < 145) {
+    // FOUR (4) input code-code units
+    // UTF-16 and UTF-32 use similar algorithms, but UTF-32 skips the narrowing.
+    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t *>(
+        simdutf::tables::utf8_to_utf16::shufutf8[idx]));
+    // Shuffle
+    // 1 byte: 00000000 00000000 0ccccccc
+    // 2 byte: 00000000 110bbbbb 10cccccc
+    // 3 byte: 1110aaaa 10bbbbbb 10cccccc
+    uint32x4_t perm = vreinterpretq_u32_u8(vqtbl1q_u8(in, sh));
+    // Split
+    // 00000000 00000000 0ccccccc
+    uint32x4_t ascii = vandq_u32(perm, vmovq_n_u32(0x7F)); // 6 or 7 bits
+    // Note: unmasked
+    // xxxxxxxx aaaaxxxx xxxxxxxx
+    uint32x4_t high = vshrq_n_u32(perm, 4); // 4 bits
+    // Use 16 bit bic instead of and.
+    // The top bits will be corrected later in the bsl
+    // 00000000 10bbbbbb 00000000
+    uint32x4_t middle = vreinterpretq_u32_u16(
+        vbicq_u16(vreinterpretq_u16_u32(perm),
+                  vmovq_n_u16(uint16_t(~0xff00)))); // 5 or 6 bits
+    // Combine low and middle with shift right accumulate
+    // 00000000 00xxbbbb bbcccccc
+    uint32x4_t lowmid = vsraq_n_u32(ascii, middle, 2);
+    // Insert top 4 bits from high byte with bitwise select
+    // 00000000 aaaabbbb bbcccccc
+    uint32x4_t composed = vbslq_u32(vmovq_n_u32(0x0000F000), high, lowmid);
+    vst1q_u32(utf32_output, composed);
+    utf32_output += 4; // We wrote 4 32-bit characters.
+    return consumed;
+  } else if (idx < 209) {
+    // THREE (3) input code-code units
+    if (input_utf8_end_of_code_point_mask == 0x888) {
+      // We want to take 3 4-byte UTF-8 code units and turn them into 3 4-byte
+      // UTF-32 code units. This uses the same method as the fixed 3 byte
+      // version, reversing and shift left insert. However, there is no need for
+      // a shuffle mask now, just rev16 and rev32.
+      //
+      // This version does not use the LUT, but 4 byte sequences are less common
+      // and the overhead of the extra memory access is less important than the
+      // early branch overhead in shorter sequences, so it comes last.
+
+      // Swap pairs of bytes
+      // 10dddddd|10cccccc|10bbbbbb|11110aaa
+      // 10cccccc 10dddddd|11110aaa 10bbbbbb
+      uint16x8_t swap1 = vreinterpretq_u16_u8(vrev16q_u8(in));
+      // Shift left and insert
+      // xxxxcccc ccdddddd|xxxxxxxa aabbbbbb
+      uint16x8_t merge1 = vsliq_n_u16(swap1, vreinterpretq_u16_u8(in), 6);
+      // Swap 16-bit lanes
+      // xxxxcccc ccdddddd xxxxxxxa aabbbbbb
+      // xxxxxxxa aabbbbbb xxxxcccc ccdddddd
+      uint32x4_t swap2 = vreinterpretq_u32_u16(vrev32q_u16(merge1));
+      // Shift insert again
+      // xxxxxxxx xxxaaabb bbbbcccc ccdddddd
+      uint32x4_t merge2 = vsliq_n_u32(swap2, vreinterpretq_u32_u16(merge1), 12);
+      // Clear the garbage
+      // 00000000 000aaabb bbbbcccc ccdddddd
+      uint32x4_t composed = vandq_u32(merge2, vmovq_n_u32(0x1FFFFF));
+      // Store
+      vst1q_u32(utf32_output, composed);
+
+      utf32_output += 3; // We wrote 3 32-bit characters.
+      return 12;         // We consumed 12 bytes.
+    }
+    // Unlike UTF-16, doing a fast codepath doesn't have nearly as much benefit
+    // due to surrogates no longer being involved.
+    uint8x16_t sh = vld1q_u8(reinterpret_cast<const uint8_t *>(
+        simdutf::tables::utf8_to_utf16::shufutf8[idx]));
+    // 1 byte: 00000000 00000000 00000000 0ddddddd
+    // 2 byte: 00000000 00000000 110ccccc 10dddddd
+    // 3 byte: 00000000 1110bbbb 10cccccc 10dddddd
+    // 4 byte: 11110aaa 10bbbbbb 10cccccc 10dddddd
+    uint32x4_t perm = vreinterpretq_u32_u8(vqtbl1q_u8(in, sh));
+    // Ascii
+    uint32x4_t ascii = vandq_u32(perm, vmovq_n_u32(0x7F));
+    uint32x4_t middle = vandq_u32(perm, vmovq_n_u32(0x3f00));
+    // When converting the way we do, the 3 byte prefix will be interpreted as
+    // the 18th bit being set, since the code would interpret the lead byte
+    // (0b1110bbbb) as a continuation byte (0b10bbbbbb). To fix this, we can
+    // either xor or do an 8 bit add of the 6th bit shifted right by 1. Since
+    // NEON has shift right accumulate, we use that.
+    //  4 byte   3 byte
+    // 10bbbbbb 1110bbbb
+    // 00000000 01000000 6th bit
+    // 00000000 00100000 shift right
+    // 10bbbbbb 0000bbbb add
+    // 00bbbbbb 0000bbbb mask
+    uint8x16_t correction =
+        vreinterpretq_u8_u32(vandq_u32(perm, vmovq_n_u32(0x00400000)));
+    uint32x4_t corrected = vreinterpretq_u32_u8(
+        vsraq_n_u8(vreinterpretq_u8_u32(perm), correction, 1));
+    // 00000000 00000000 0000cccc ccdddddd
+    uint32x4_t cd = vsraq_n_u32(ascii, middle, 2);
+    // Insert twice
+    // xxxxxxxx xxxaaabb bbbbxxxx xxxxxxxx
+    uint32x4_t ab = vbslq_u32(vmovq_n_u32(0x01C0000), vshrq_n_u32(corrected, 6),
+                              vshrq_n_u32(corrected, 4));
+    // 00000000 000aaabb bbbbcccc ccdddddd
+    uint32x4_t composed = vbslq_u32(vmovq_n_u32(0xFFE00FFF), cd, ab);
+    // Store
+    vst1q_u32(utf32_output, composed);
+    utf32_output += 3; // We wrote 3 32-bit characters.
+    return consumed;
+  } else {
+    // here we know that there is an error but we do not handle errors
+    return 12;
+  }
+}