react-native-quick-crypto 1.0.19 → 1.1.1

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

@@ -0,0 +1,19 @@
+// file included directly
+
+bool validate_ascii(const char *buf, size_t len) {
+  const char *end = buf + len;
+  const __m512i ascii = _mm512_set1_epi8((uint8_t)0x80);
+  __m512i running_or = _mm512_setzero_si512();
+  for (; end - buf >= 64; buf += 64) {
+    const __m512i utf8 = _mm512_loadu_si512((const __m512i *)buf);
+    running_or = _mm512_ternarylogic_epi32(running_or, utf8, ascii,
+                                           0xf8); // running_or | (utf8 & ascii)
+  }
+  if (buf < end) {
+    const __m512i utf8 = _mm512_maskz_loadu_epi8(
+        (uint64_t(1) << (end - buf)) - 1, (const __m512i *)buf);
+    running_or = _mm512_ternarylogic_epi32(running_or, utf8, ascii,
+                                           0xf8); // running_or | (utf8 & ascii)
+  }
+  return (_mm512_test_epi8_mask(running_or, running_or) == 0);
+}

package/deps/simdutf/src/icelake/icelake_base64.inl.cpp

@@ -0,0 +1,630 @@
+// file included directly
+/**
+ * References and further reading:
+ *
+ * Wojciech Muła, Daniel Lemire, Base64 encoding and decoding at almost the
+ * speed of a memory copy, Software: Practice and Experience 50 (2), 2020.
+ * https://arxiv.org/abs/1910.05109
+ *
+ * Wojciech Muła, Daniel Lemire, Faster Base64 Encoding and Decoding using AVX2
+ * Instructions, ACM Transactions on the Web 12 (3), 2018.
+ * https://arxiv.org/abs/1704.00605
+ *
+ * Simon Josefsson. 2006. The Base16, Base32, and Base64 Data Encodings.
+ * https://tools.ietf.org/html/rfc4648. (2006). Internet Engineering Task Force,
+ * Request for Comments: 4648.
+ *
+ * Alfred Klomp. 2014a. Fast Base64 encoding/decoding with SSE vectorization.
+ * http://www.alfredklomp.com/programming/sse-base64/. (2014).
+ *
+ * Alfred Klomp. 2014b. Fast Base64 stream encoder/decoder in C99, with SIMD
+ * acceleration. https://github.com/aklomp/base64. (2014).
+ *
+ * Hanson Char. 2014. A Fast and Correct Base 64 Codec. (2014).
+ * https://aws.amazon.com/blogs/developer/a-fast-and-correct-base-64-codec/
+ *
+ * Nick Kopp. 2013. Base64 Encoding on a GPU.
+ * https://www.codeproject.com/Articles/276993/Base-Encoding-on-a-GPU. (2013).
+ */
+
+struct block64 {
+  __m512i chunks[1];
+};
+
+template <bool base64_url, bool use_lines>
+size_t encode_base64_impl(char *dst, const char *src, size_t srclen,
+                          base64_options options,
+                          size_t line_length = simdutf::default_line_length) {
+  size_t offset = 0;
+  if (line_length < 4) {
+    line_length = 4; // We do not support line_length less than 4
+  }
+  // credit: Wojciech Muła
+  const uint8_t *input = (const uint8_t *)src;
+
+  uint8_t *out = (uint8_t *)dst;
+  static const char *lookup_tbl =
+      base64_url
+          ? "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789-_"
+          : "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz0123456789+/";
+  const __m512i shuffle_input = _mm512_setr_epi32(
+      0x01020001, 0x04050304, 0x07080607, 0x0a0b090a, 0x0d0e0c0d, 0x10110f10,
+      0x13141213, 0x16171516, 0x191a1819, 0x1c1d1b1c, 0x1f201e1f, 0x22232122,
+      0x25262425, 0x28292728, 0x2b2c2a2b, 0x2e2f2d2e);
+  const __m512i lookup =
+      _mm512_loadu_si512(reinterpret_cast<const __m512i *>(lookup_tbl));
+  const __m512i multi_shifts = _mm512_set1_epi64(UINT64_C(0x3036242a1016040a));
+  size_t size = srclen;
+  __mmask64 input_mask = 0xffffffffffff; // (1 << 48) - 1
+  // We want that input == end_input means that we must stop.
+  const uint8_t *end_input = input + (size - (size % 48));
+  while (input != end_input) {
+    const __m512i v = _mm512_maskz_loadu_epi8(
+        input_mask, reinterpret_cast<const __m512i *>(input));
+    const __m512i in = _mm512_permutexvar_epi8(shuffle_input, v);
+    const __m512i indices = _mm512_multishift_epi64_epi8(multi_shifts, in);
+    const __m512i result = _mm512_permutexvar_epi8(indices, lookup);
+    if (use_lines) {
+      if (offset + 64 > line_length) {
+        if (line_length >= 64) {
+          __m512i expanded = _mm512_mask_expand_epi8(
+              _mm512_set1_epi8('\n'), ~(1ULL << ((line_length - offset))),
+              result);
+          _mm512_storeu_si512(reinterpret_cast<__m512i *>(out), expanded);
+          __m128i last_lane =
+              _mm512_extracti32x4_epi32(result, 3); // Lane 3 (bytes 48-63)
+          uint8_t last_byte =
+              static_cast<uint8_t>(_mm_extract_epi8(last_lane, 15));
+          out[64] = last_byte;
+          out += 65;
+          offset = 64 - (line_length - offset);
+        } else { // slow path
+          alignas(64) uint8_t local_buffer[64];
+          _mm512_storeu_si512(reinterpret_cast<__m512i *>(local_buffer),
+                              result);
+          size_t out_pos = 0;
+          size_t local_offset = offset;
+          for (size_t j = 0; j < 64;) {
+            if (local_offset == line_length) {
+              out[out_pos++] = '\n';
+              local_offset = 0;
+            }
+            out[out_pos++] = local_buffer[j++];
+            local_offset++;
+          }
+          offset = local_offset;
+          out += out_pos;
+        }
+      } else {
+        _mm512_storeu_si512(reinterpret_cast<__m512i *>(out), result);
+        offset += 64;
+        out += 64;
+      }
+    } else {
+      _mm512_storeu_si512(reinterpret_cast<__m512i *>(out), result);
+      out += 64;
+    }
+    input += 48;
+  }
+  size = size % 48;
+
+  input_mask = ((__mmask64)1 << size) - 1;
+  const __m512i v = _mm512_maskz_loadu_epi8(
+      input_mask, reinterpret_cast<const __m512i *>(input));
+  const __m512i in = _mm512_permutexvar_epi8(shuffle_input, v);
+  const __m512i indices = _mm512_multishift_epi64_epi8(multi_shifts, in);
+  bool padding_needed =
+      (((options & base64_url) == 0) ^
+       ((options & base64_reverse_padding) == base64_reverse_padding));
+  size_t padding_amount = ((size % 3) > 0) ? (3 - (size % 3)) : 0;
+  size_t output_len = ((size + 2) / 3) * 4;
+  size_t non_padded_output_len = output_len - padding_amount;
+  if (!padding_needed) {
+    output_len = non_padded_output_len;
+  }
+  // If no output, we are done.
+  if (output_len == 0) {
+    return (size_t)(out - (uint8_t *)dst);
+  }
+  __mmask64 output_mask = 0xFFFFFFFFFFFFFFFF >> (64 - output_len);
+  __m512i result = _mm512_mask_permutexvar_epi8(
+      _mm512_set1_epi8('='), ((__mmask64)1 << non_padded_output_len) - 1,
+      indices, lookup);
+  if (use_lines) {
+    if (offset + output_len > line_length) {
+      if (line_length >= 64) {
+        __m512i expanded = _mm512_mask_expand_epi8(
+            _mm512_set1_epi8('\n'), ~(1ULL << ((line_length - offset))),
+            result);
+        if (output_len == 64) {
+          _mm512_storeu_si512(reinterpret_cast<__m512i *>(out), expanded);
+          out += 64;
+          _mm512_mask_storeu_epi8(reinterpret_cast<__m512i *>(out - 63),
+                                  1ULL << 63, result);
+          out++;
+        } else {
+          output_mask = 0xFFFFFFFFFFFFFFFF >> (64 - output_len - 1);
+          _mm512_mask_storeu_epi8(reinterpret_cast<__m512i *>(out), output_mask,
+                                  expanded);
+          out += output_len + 1;
+        }
+      } else {
+        alignas(64) uint8_t local_buffer[64];
+        _mm512_storeu_si512(reinterpret_cast<__m512i *>(local_buffer), result);
+        size_t out_pos = 0;
+        size_t local_offset = offset;
+        for (size_t j = 0; j < output_len;) {
+          if (local_offset == line_length) {
+            out[out_pos++] = '\n';
+            local_offset = 0;
+          }
+          out[out_pos++] = local_buffer[j++];
+          local_offset++;
+        }
+        offset = local_offset;
+        out += out_pos;
+      }
+    } else {
+      _mm512_mask_storeu_epi8(reinterpret_cast<__m512i *>(out), output_mask,
+                              result);
+      out += output_len;
+    }
+  } else {
+    _mm512_mask_storeu_epi8(reinterpret_cast<__m512i *>(out), output_mask,
+                            result);
+    out += output_len;
+  }
+  return (size_t)(out - (uint8_t *)dst);
+}
+
+template <bool base64_url>
+size_t encode_base64(char *dst, const char *src, size_t srclen,
+                     base64_options options) {
+  return encode_base64_impl<base64_url, false>(dst, src, srclen, options);
+}
+
+template <bool base64_url, bool ignore_garbage, bool default_or_url>
+static inline uint64_t to_base64_mask(block64 *b, uint64_t *error,
+                                      uint64_t input_mask = UINT64_MAX) {
+  __m512i input = b->chunks[0];
+  const __m512i ascii_space_tbl = _mm512_set_epi8(
+      0, 0, 13, 12, 0, 10, 9, 0, 0, 0, 0, 0, 0, 0, 0, 32, 0, 0, 13, 12, 0, 10,
+      9, 0, 0, 0, 0, 0, 0, 0, 0, 32, 0, 0, 13, 12, 0, 10, 9, 0, 0, 0, 0, 0, 0,
+      0, 0, 32, 0, 0, 13, 12, 0, 10, 9, 0, 0, 0, 0, 0, 0, 0, 0, 32);
+  __m512i lookup0;
+  if (default_or_url) {
+    lookup0 = _mm512_set_epi8(
+        -128, -128, -128, -128, -128, -128, 61, 60, 59, 58, 57, 56, 55, 54, 53,
+        52, 63, -128, 62, -128, 62, -128, -128, -128, -128, -128, -128, -128,
+        -128, -128, -128, -1, -128, -128, -128, -128, -128, -128, -128, -128,
+        -128, -128, -128, -128, -128, -128, -128, -128, -128, -128, -1, -128,
+        -128, -1, -1, -128, -128, -128, -128, -128, -128, -128, -128, -1);
+  } else if (base64_url) {
+    lookup0 = _mm512_set_epi8(
+        -128, -128, -128, -128, -128, -128, 61, 60, 59, 58, 57, 56, 55, 54, 53,
+        52, -128, -128, 62, -128, -128, -128, -128, -128, -128, -128, -128,
+        -128, -128, -128, -128, -1, -128, -128, -128, -128, -128, -128, -128,
+        -128, -128, -128, -128, -128, -128, -128, -128, -128, -128, -128, -1,
+        -128, -128, -1, -1, -128, -128, -128, -128, -128, -128, -128, -128, -1);
+  } else {
+    lookup0 = _mm512_set_epi8(
+        -128, -128, -128, -128, -128, -128, 61, 60, 59, 58, 57, 56, 55, 54, 53,
+        52, 63, -128, -128, -128, 62, -128, -128, -128, -128, -128, -128, -128,
+        -128, -128, -128, -1, -128, -128, -128, -128, -128, -128, -128, -128,
+        -128, -128, -128, -128, -128, -128, -128, -128, -128, -128, -1, -128,
+        -128, -1, -1, -128, -128, -128, -128, -128, -128, -128, -128, -128);
+  }
+  __m512i lookup1;
+  if (default_or_url) {
+    lookup1 = _mm512_set_epi8(
+        -128, -128, -128, -128, -128, 51, 50, 49, 48, 47, 46, 45, 44, 43, 42,
+        41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 27, 26, -128,
+        63, -128, -128, -128, -128, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15,
+        14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, -128);
+  } else if (base64_url) {
+    lookup1 = _mm512_set_epi8(
+        -128, -128, -128, -128, -128, 51, 50, 49, 48, 47, 46, 45, 44, 43, 42,
+        41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 27, 26, -128,
+        63, -128, -128, -128, -128, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16, 15,
+        14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, -128);
+  } else {
+    lookup1 = _mm512_set_epi8(
+        -128, -128, -128, -128, -128, 51, 50, 49, 48, 47, 46, 45, 44, 43, 42,
+        41, 40, 39, 38, 37, 36, 35, 34, 33, 32, 31, 30, 29, 28, 27, 26, -128,
+        -128, -128, -128, -128, -128, 25, 24, 23, 22, 21, 20, 19, 18, 17, 16,
+        15, 14, 13, 12, 11, 10, 9, 8, 7, 6, 5, 4, 3, 2, 1, 0, -128);
+  }
+
+  const __m512i translated = _mm512_permutex2var_epi8(lookup0, input, lookup1);
+  const __m512i combined = _mm512_or_si512(translated, input);
+  const __mmask64 mask = _mm512_movepi8_mask(combined) & input_mask;
+  if (!ignore_garbage && mask) {
+    const __mmask64 spaces =
+        _mm512_cmpeq_epi8_mask(_mm512_shuffle_epi8(ascii_space_tbl, input),
+                               input) &
+        input_mask;
+    *error = (mask ^ spaces);
+  }
+  b->chunks[0] = translated;
+
+  return mask | (~input_mask);
+}
+
+static inline void copy_block(block64 *b, char *output) {
+  _mm512_storeu_si512(reinterpret_cast<__m512i *>(output), b->chunks[0]);
+}
+
+static inline uint64_t compress_block(block64 *b, uint64_t mask, char *output) {
+  uint64_t nmask = ~mask;
+  __m512i c = _mm512_maskz_compress_epi8(nmask, b->chunks[0]);
+  _mm512_storeu_si512(reinterpret_cast<__m512i *>(output), c);
+  return _mm_popcnt_u64(nmask);
+}
+
+// The caller of this function is responsible to ensure that there are 64 bytes
+// available from reading at src. The data is read into a block64 structure.
+static inline void load_block(block64 *b, const char *src) {
+  b->chunks[0] = _mm512_loadu_si512(reinterpret_cast<const __m512i *>(src));
+}
+
+static inline void load_block_partial(block64 *b, const char *src,
+                                      __mmask64 input_mask) {
+  b->chunks[0] = _mm512_maskz_loadu_epi8(
+      input_mask, reinterpret_cast<const __m512i *>(src));
+}
+
+// The caller of this function is responsible to ensure that there are 128 bytes
+// available from reading at src. The data is read into a block64 structure.
+static inline void load_block(block64 *b, const char16_t *src) {
+  __m512i m1 = _mm512_loadu_si512(reinterpret_cast<const __m512i *>(src));
+  __m512i m2 = _mm512_loadu_si512(reinterpret_cast<const __m512i *>(src + 32));
+  __m512i p = _mm512_packus_epi16(m1, m2);
+  b->chunks[0] =
+      _mm512_permutexvar_epi64(_mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7), p);
+}
+
+static inline void load_block_partial(block64 *b, const char16_t *src,
+                                      __mmask64 input_mask) {
+  __m512i m1 = _mm512_maskz_loadu_epi16((__mmask32)input_mask,
+                                        reinterpret_cast<const __m512i *>(src));
+  __m512i m2 =
+      _mm512_maskz_loadu_epi16((__mmask32)(input_mask >> 32),
+                               reinterpret_cast<const __m512i *>(src + 32));
+  __m512i p = _mm512_packus_epi16(m1, m2);
+  b->chunks[0] =
+      _mm512_permutexvar_epi64(_mm512_setr_epi64(0, 2, 4, 6, 1, 3, 5, 7), p);
+}
+
+static inline void base64_decode(char *out, __m512i str) {
+  const __m512i merge_ab_and_bc =
+      _mm512_maddubs_epi16(str, _mm512_set1_epi32(0x01400140));
+  const __m512i merged =
+      _mm512_madd_epi16(merge_ab_and_bc, _mm512_set1_epi32(0x00011000));
+  const __m512i pack = _mm512_set_epi8(
+      0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 60, 61, 62, 56, 57, 58,
+      52, 53, 54, 48, 49, 50, 44, 45, 46, 40, 41, 42, 36, 37, 38, 32, 33, 34,
+      28, 29, 30, 24, 25, 26, 20, 21, 22, 16, 17, 18, 12, 13, 14, 8, 9, 10, 4,
+      5, 6, 0, 1, 2);
+  const __m512i shuffled = _mm512_permutexvar_epi8(pack, merged);
+  _mm512_mask_storeu_epi8(
+      (__m512i *)out, 0xffffffffffff,
+      shuffled); // mask would be 0xffffffffffff since we write 48 bytes.
+}
+// decode 64 bytes and output 48 bytes
+static inline void base64_decode_block(char *out, const char *src) {
+  base64_decode(out,
+                _mm512_loadu_si512(reinterpret_cast<const __m512i *>(src)));
+}
+static inline void base64_decode_block(char *out, block64 *b) {
+  base64_decode(out, b->chunks[0]);
+}
+
+template <bool base64_url, bool ignore_garbage, bool default_or_url,
+          typename chartype>
+full_result
+compress_decode_base64(char *dst, const chartype *src, size_t srclen,
+                       base64_options options,
+                       last_chunk_handling_options last_chunk_options) {
+  (void)options;
+  const uint8_t *to_base64 =
+      default_or_url ? tables::base64::to_base64_default_or_url_value
+                     : (base64_url ? tables::base64::to_base64_url_value
+                                   : tables::base64::to_base64_value);
+  auto ri = simdutf::scalar::base64::find_end(src, srclen, options);
+  size_t equallocation = ri.equallocation;
+  size_t padding_characters = ri.equalsigns;
+  srclen = ri.srclen;
+  size_t full_input_length = ri.full_input_length;
+  if (srclen == 0) {
+    if (!ignore_garbage && padding_characters > 0) {
+      return {INVALID_BASE64_CHARACTER, equallocation, 0};
+    }
+    return {SUCCESS, full_input_length, 0};
+  }
+  const chartype *const srcinit = src;
+  const char *const dstinit = dst;
+  const chartype *const srcend = src + srclen;
+
+  // figure out why block_size == 2 is sometimes best???
+  constexpr size_t block_size = 6;
+  char buffer[block_size * 64];
+  char *bufferptr = buffer;
+  if (srclen >= 64) {
+    const chartype *const srcend64 = src + srclen - 64;
+    while (src <= srcend64) {
+      block64 b;
+      load_block(&b, src);
+      src += 64;
+      uint64_t error = 0;
+      uint64_t badcharmask =
+          to_base64_mask<base64_url, ignore_garbage, default_or_url>(&b,
+                                                                     &error);
+      if (!ignore_garbage && error) {
+        src -= 64;
+        size_t error_offset = _tzcnt_u64(error);
+        return {error_code::INVALID_BASE64_CHARACTER,
+                size_t(src - srcinit + error_offset), size_t(dst - dstinit)};
+      }
+      if (badcharmask != 0) {
+        // optimization opportunity: check for simple masks like those made of
+        // continuous 1s followed by continuous 0s. And masks containing a
+        // single bad character.
+        bufferptr += compress_block(&b, badcharmask, bufferptr);
+      } else if (bufferptr != buffer) {
+        copy_block(&b, bufferptr);
+        bufferptr += 64;
+      } else {
+        base64_decode_block(dst, &b);
+        dst += 48;
+      }
+      if (bufferptr >= (block_size - 1) * 64 + buffer) {
+        for (size_t i = 0; i < (block_size - 1); i++) {
+          base64_decode_block(dst, buffer + i * 64);
+          dst += 48;
+        }
+        std::memcpy(buffer, buffer + (block_size - 1) * 64,
+                    64); // 64 might be too much
+        bufferptr -= (block_size - 1) * 64;
+      }
+    }
+  }
+
+  int last_block_len = (int)(srcend - src);
+  if (last_block_len != 0) {
+    __mmask64 input_mask = ((__mmask64)1 << last_block_len) - 1;
+    block64 b;
+    load_block_partial(&b, src, input_mask);
+    uint64_t error = 0;
+    uint64_t badcharmask =
+        to_base64_mask<base64_url, ignore_garbage, default_or_url>(&b, &error,
+                                                                   input_mask);
+    if (!ignore_garbage && error) {
+      size_t error_offset = _tzcnt_u64(error);
+      return {error_code::INVALID_BASE64_CHARACTER,
+              size_t(src - srcinit + error_offset), size_t(dst - dstinit)};
+    }
+    src += last_block_len;
+    bufferptr += compress_block(&b, badcharmask, bufferptr);
+  }
+
+  char *buffer_start = buffer;
+  for (; buffer_start + 64 <= bufferptr; buffer_start += 64) {
+    base64_decode_block(dst, buffer_start);
+    dst += 48;
+  }
+  if ((bufferptr - buffer_start) != 0) {
+    // For efficiency reasons, we end up reproducing much of the code
+    // in base64_tail_decode_impl. Better engineering would be to
+    // refactor the code so that we can call it without a performance hit.
+    size_t rem = (bufferptr - buffer_start);
+    int idx = rem % 4;
+    __mmask64 mask = ((__mmask64)1 << rem) - 1;
+    __m512i input = _mm512_maskz_loadu_epi8(mask, buffer_start);
+    size_t output_len = (rem / 4) * 3;
+    __mmask64 output_mask = mask >> (rem - output_len);
+    const __m512i merge_ab_and_bc =
+        _mm512_maddubs_epi16(input, _mm512_set1_epi32(0x01400140));
+    const __m512i merged =
+        _mm512_madd_epi16(merge_ab_and_bc, _mm512_set1_epi32(0x00011000));
+    const __m512i pack = _mm512_set_epi8(
+        0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 60, 61, 62, 56, 57, 58,
+        52, 53, 54, 48, 49, 50, 44, 45, 46, 40, 41, 42, 36, 37, 38, 32, 33, 34,
+        28, 29, 30, 24, 25, 26, 20, 21, 22, 16, 17, 18, 12, 13, 14, 8, 9, 10, 4,
+        5, 6, 0, 1, 2);
+    const __m512i shuffled = _mm512_permutexvar_epi8(pack, merged);
+    // We never should have that the number of base64 characters + the
+    // number of padding characters is more than 4.
+    if (!ignore_garbage && (idx + padding_characters > 4)) {
+      return {INVALID_BASE64_CHARACTER, size_t(src - srcinit),
+              size_t(dst - dstinit), true};
+    }
+    // The idea here is that in loose mode,
+    // if there is padding at all, it must be used
+    // to form 4-wise chunk. However, in loose mode,
+    // we do accept no padding at all.
+    if (!ignore_garbage &&
+        last_chunk_options == last_chunk_handling_options::loose &&
+        (idx >= 2) && padding_characters > 0 &&
+        ((idx + padding_characters) & 3) != 0) {
+      return {INVALID_BASE64_CHARACTER, size_t(src - srcinit),
+              size_t(dst - dstinit), true};
+    } else
+      // The idea here is that in strict mode, we do not want to accept
+      // incomplete base64 chunks. So if the chunk was otherwise valid, we
+      // return BASE64_INPUT_REMAINDER.
+      if (!ignore_garbage &&
+          last_chunk_options == last_chunk_handling_options::strict &&
+          (idx >= 2) && ((idx + padding_characters) & 3) != 0) {
+        // The partial chunk was at src - idx
+        _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+        dst += output_len;
+        return {BASE64_INPUT_REMAINDER, equallocation, size_t(dst - dstinit)};
+      } else
+        // If there is a partial chunk with insufficient padding, with
+        // stop_before_partial, we need to just ignore it. In "only full" mode,
+        // skip the minute there are padding characters.
+        if ((last_chunk_options ==
+                 last_chunk_handling_options::stop_before_partial &&
+             (padding_characters + idx < 4) && (idx != 0) &&
+             (idx >= 2 || padding_characters == 0)) ||
+            (last_chunk_options ==
+                 last_chunk_handling_options::only_full_chunks &&
+             (idx >= 2 || padding_characters == 0))) {
+          _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+          dst += output_len;
+          // we need to rewind src to before the partial chunk
+          size_t characters_to_skip = idx;
+          while (characters_to_skip > 0) {
+            src--;
+            auto c = *src;
+            uint8_t code = to_base64[uint8_t(c)];
+            if (simdutf::scalar::base64::is_eight_byte(c) && code <= 63) {
+              characters_to_skip--;
+            }
+          }
+          // And then we need to skip ignored characters
+          // See https://github.com/simdutf/simdutf/issues/824
+          while (src > srcinit) {
+            auto c = *(src - 1);
+            uint8_t code = to_base64[uint8_t(c)];
+            if (simdutf::scalar::base64::is_eight_byte(c) && code <= 63) {
+              break;
+            }
+            src--;
+          }
+          return {SUCCESS, size_t(src - srcinit), size_t(dst - dstinit)};
+        } else {
+          if (idx == 2) {
+            if (!ignore_garbage &&
+                last_chunk_options == last_chunk_handling_options::strict) {
+              uint32_t triple = (uint32_t(bufferptr[-2]) << 3 * 6) +
+                                (uint32_t(bufferptr[-1]) << 2 * 6);
+              if (triple & 0xffff) {
+                _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+                dst += output_len;
+                return {BASE64_EXTRA_BITS, size_t(src - srcinit),
+                        size_t(dst - dstinit)};
+              }
+            }
+            output_mask = (output_mask << 1) | 1;
+            output_len += 1;
+            _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+            dst += output_len;
+          } else if (idx == 3) {
+            if (!ignore_garbage &&
+                last_chunk_options == last_chunk_handling_options::strict) {
+              uint32_t triple = (uint32_t(bufferptr[-3]) << 3 * 6) +
+                                (uint32_t(bufferptr[-2]) << 2 * 6) +
+                                (uint32_t(bufferptr[-1]) << 1 * 6);
+              if (triple & 0xff) {
+                _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+                dst += output_len;
+                return {BASE64_EXTRA_BITS, size_t(src - srcinit),
+                        size_t(dst - dstinit)};
+              }
+            }
+            output_mask = (output_mask << 2) | 3;
+            output_len += 2;
+            _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+            dst += output_len;
+          } else if (!ignore_garbage && idx == 1 &&
+                     (!is_partial(last_chunk_options) ||
+                      (is_partial(last_chunk_options) &&
+                       padding_characters > 0))) {
+            _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+            dst += output_len;
+            return {BASE64_INPUT_REMAINDER, size_t(src - srcinit),
+                    size_t(dst - dstinit)};
+          } else if (!ignore_garbage && idx == 0 && padding_characters > 0) {
+            _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+            dst += output_len;
+            return {INVALID_BASE64_CHARACTER, equallocation,
+                    size_t(dst - dstinit)};
+          } else {
+            _mm512_mask_storeu_epi8((__m512i *)dst, output_mask, shuffled);
+            dst += output_len;
+          }
+        }
+    if (!ignore_garbage && !is_partial(last_chunk_options) &&
+        padding_characters > 0) {
+      size_t output_count = size_t(dst - dstinit);
+      if ((output_count % 3 == 0) ||
+          ((output_count % 3) + 1 + padding_characters != 4)) {
+        return {INVALID_BASE64_CHARACTER, equallocation, output_count};
+      }
+    }
+    return {SUCCESS, full_input_length, size_t(dst - dstinit)};
+  }
+
+  if (!ignore_garbage && padding_characters > 0) {
+    if ((size_t(dst - dstinit) % 3 == 0) ||
+        ((size_t(dst - dstinit) % 3) + 1 + padding_characters != 4)) {
+      return {INVALID_BASE64_CHARACTER, equallocation, size_t(dst - dstinit)};
+    }
+  }
+  return {SUCCESS, srclen, size_t(dst - dstinit)};
+}
+
+simdutf_warn_unused size_t icelake_binary_length_from_base64(const char *input,
+                                                             size_t length) {
+  size_t count = 0;
+  const char *ptr = input;
+  const char *end = input + length;
+
+  __m512i spaces = _mm512_set1_epi8(0x20);
+  while (ptr + 64 <= end) {
+    __m512i data = _mm512_loadu_si512(reinterpret_cast<const __m512i *>(ptr));
+    uint64_t mask = _mm512_cmpgt_epi8_mask(data, spaces);
+    count += count_ones(mask);
+    ptr += 64;
+  }
+
+  while (ptr < end) {
+    count += (*ptr > 0x20) ? 1 : 0;
+    ptr++;
+  }
+
+  size_t padding = 0;
+  size_t pos = length;
+  while (pos > 0 && padding < 2) {
+    char c = input[--pos];
+    if (c == '=') {
+      padding++;
+    } else if (c > ' ') {
+      break;
+    }
+  }
+  return ((count - padding) * 3) / 4;
+}
+
+simdutf_warn_unused size_t
+icelake_binary_length_from_base64(const char16_t *input, size_t length) {
+  size_t count = 0;
+  const char16_t *ptr = input;
+  const char16_t *end = input + length;
+
+  __m512i spaces = _mm512_set1_epi16(0x20);
+  while (ptr + 32 <= end) {
+    __m512i data = _mm512_loadu_si512(reinterpret_cast<const __m512i *>(ptr));
+    __mmask32 mask = _mm512_cmpgt_epi16_mask(data, spaces);
+    count += _mm_popcnt_u32(mask);
+    ptr += 32;
+  }
+
+  while (ptr < end) {
+    count += (*ptr > 0x20) ? 1 : 0;
+    ptr++;
+  }
+
+  size_t padding = 0;
+  size_t pos = length;
+  while (pos > 0 && padding < 2) {
+    char16_t c = input[--pos];
+    if (c == '=') {
+      padding++;
+    } else if (c > ' ') {
+      break;
+    }
+  }
+  return ((count - padding) * 3) / 4;
+}

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

@@ -0,0 +1,37 @@
+// file included directly
+/**
+ * Store the last N bytes of previous followed by 512-N bytes from input.
+ */
+template <int N> __m512i prev(__m512i input, __m512i previous) {
+  static_assert(N <= 32, "N must be no larger than 32");
+  const __m512i movemask =
+      _mm512_setr_epi32(28, 29, 30, 31, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11);
+  const __m512i rotated = _mm512_permutex2var_epi32(input, movemask, previous);
+#if SIMDUTF_GCC8 || SIMDUTF_GCC9
+  constexpr int shift = 16 - N; // workaround for GCC8,9
+  return _mm512_alignr_epi8(input, rotated, shift);
+#else
+  return _mm512_alignr_epi8(input, rotated, 16 - N);
+#endif // SIMDUTF_GCC8 || SIMDUTF_GCC9
+}
+
+template <unsigned idx0, unsigned idx1, unsigned idx2, unsigned idx3>
+__m512i shuffle_epi128(__m512i v) {
+  static_assert((idx0 >= 0 && idx0 <= 3), "idx0 must be in range 0..3");
+  static_assert((idx1 >= 0 && idx1 <= 3), "idx1 must be in range 0..3");
+  static_assert((idx2 >= 0 && idx2 <= 3), "idx2 must be in range 0..3");
+  static_assert((idx3 >= 0 && idx3 <= 3), "idx3 must be in range 0..3");
+
+  constexpr unsigned shuffle = idx0 | (idx1 << 2) | (idx2 << 4) | (idx3 << 6);
+  return _mm512_shuffle_i32x4(v, v, shuffle);
+}
+
+template <unsigned idx> constexpr __m512i broadcast_epi128(__m512i v) {
+  return shuffle_epi128<idx, idx, idx, idx>(v);
+}
+
+simdutf_really_inline __m512i broadcast_128bit_lane(__m128i lane) {
+  const __m512i tmp = _mm512_castsi128_si512(lane);
+
+  return broadcast_epi128<0>(tmp);
+}