multi_compress 0.3.1 → 0.3.3

data/ext/multi_compress/vendor/brotli/c/enc/hash.h

@@ -10,11 +10,6 @@
 #ifndef BROTLI_ENC_HASH_H_
 #define BROTLI_ENC_HASH_H_
 
-#include <stdlib.h>  /* exit */
-#include <string.h>  /* memcmp, memset */
-
-#include <brotli/types.h>
-
 #include "../common/constants.h"
 #include "../common/dictionary.h"
 #include "../common/platform.h"
@@ -22,7 +17,10 @@
 #include "encoder_dict.h"
 #include "fast_log.h"
 #include "find_match_length.h"
+#include "hash_base.h"
+#include "matching_tag_mask.h"
 #include "memory.h"
+#include "params.h"
 #include "quality.h"
 #include "static_dict.h"
 
@@ -38,7 +36,7 @@ typedef struct {
   void* extra[4];
 
   /**
-   * False before the fisrt invocation of HasherSetup (where "extra" memory)
+   * False before the first invocation of HasherSetup (where "extra" memory)
    * is allocated.
    */
   BROTLI_BOOL is_setup_;
@@ -71,23 +69,6 @@ typedef struct HasherSearchResult {
   int len_code_delta; /* == len_code - len */
 } HasherSearchResult;
 
-/* kHashMul32 multiplier has these properties:
-   * The multiplier must be odd. Otherwise we may lose the highest bit.
-   * No long streaks of ones or zeros.
-   * There is no effort to ensure that it is a prime, the oddity is enough
-     for this use.
-   * The number has been tuned heuristically against compression benchmarks. */
-static const uint32_t kHashMul32 = 0x1E35A7BD;
-static const uint64_t kHashMul64 =
-    BROTLI_MAKE_UINT64_T(0x1FE35A7Bu, 0xD3579BD3u);
-
-static BROTLI_INLINE uint32_t Hash14(const uint8_t* data) {
-  uint32_t h = BROTLI_UNALIGNED_LOAD32LE(data) * kHashMul32;
-  /* The higher bits contain more mixture from the multiplication,
-     so we take our results from there. */
-  return h >> (32 - 14);
-}
-
 static BROTLI_INLINE void PrepareDistanceCache(
     int* BROTLI_RESTRICT distance_cache, const int num_distances) {
   if (num_distances > 4) {
@@ -297,6 +278,16 @@ static BROTLI_INLINE size_t BackwardMatchLengthCode(const BackwardMatch* self) {
 #include "hash_longest_match64_inc.h"  /* NOLINT(build/include) */
 #undef HASHER
 
+#if defined(BROTLI_MAX_SIMD_QUALITY)
+#define HASHER() H58
+#include "hash_longest_match_simd_inc.h" /* NOLINT(build/include) */
+#undef HASHER
+
+#define HASHER() H68
+#include "hash_longest_match64_simd_inc.h" /* NOLINT(build/include) */
+#undef HASHER
+#endif
+
 #define BUCKET_BITS 15
 
 #define NUM_LAST_DISTANCES_TO_CHECK 4
@@ -388,7 +379,13 @@ static BROTLI_INLINE size_t BackwardMatchLengthCode(const BackwardMatch* self) {
 #undef CAT
 #undef EXPAND_CAT
 
-#define FOR_SIMPLE_HASHERS(H) H(2) H(3) H(4) H(5) H(6) H(40) H(41) H(42) H(54)
+#if defined(BROTLI_MAX_SIMD_QUALITY)
+#define FOR_SIMPLE_HASHERS(H) \
+  H(2) H(3) H(4) H(5) H(6) H(40) H(41) H(42) H(54) H(58) H(68)
+#else
+#define FOR_SIMPLE_HASHERS(H) \
+  H(2) H(3) H(4) H(5) H(6) H(40) H(41) H(42) H(54)
+#endif
 #define FOR_COMPOSITE_HASHERS(H) H(35) H(55) H(65)
 #define FOR_GENERIC_HASHERS(H) FOR_SIMPLE_HASHERS(H) FOR_COMPOSITE_HASHERS(H)
 #define FOR_ALL_HASHERS(H) FOR_GENERIC_HASHERS(H) H(10)
@@ -523,8 +520,8 @@ static BROTLI_INLINE void FindCompoundDictionaryMatch(
   const uint64_t hash_mask = (~((uint64_t)0U)) >> (64 - hash_bits);
 
   const uint32_t* slot_offsets = (uint32_t*)(&self[1]);
-  const uint16_t* heads = (uint16_t*)(&slot_offsets[1u << slot_bits]);
-  const uint32_t* items = (uint32_t*)(&heads[1u << bucket_bits]);
+  const uint16_t* heads = (uint16_t*)(&slot_offsets[(size_t)1u << slot_bits]);
+  const uint32_t* items = (uint32_t*)(&heads[(size_t)1u << bucket_bits]);
   const uint8_t* source = NULL;
 
   const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
@@ -548,6 +545,8 @@ static BROTLI_INLINE void FindCompoundDictionaryMatch(
     source = (const uint8_t*)BROTLI_UNALIGNED_LOAD_PTR((const uint8_t**)tail);
   }
 
+  BROTLI_DCHECK(cur_ix_masked + max_length <= ring_buffer_mask);
+
   for (i = 0; i < 4; ++i) {
     const size_t distance = (size_t)distance_cache[i];
     size_t offset;
@@ -574,6 +573,11 @@ static BROTLI_INLINE void FindCompoundDictionaryMatch(
       }
     }
   }
+  /* we require matches of len >4, so increase best_len to 3, so we can compare
+   * 4 bytes all the time. */
+  if (best_len < 3) {
+    best_len = 3;
+  }
   while (item == 0) {
     size_t offset;
     size_t distance;
@@ -586,9 +590,10 @@ static BROTLI_INLINE void FindCompoundDictionaryMatch(
     limit = source_size - offset;
     limit = (limit > max_length) ? max_length : limit;
     if (distance > max_distance) continue;
-    if (cur_ix_masked + best_len > ring_buffer_mask ||
-        best_len >= limit ||
-        data[cur_ix_masked + best_len] != source[offset + best_len]) {
+    if (cur_ix_masked + best_len > ring_buffer_mask || best_len >= limit ||
+        /* compare 4 bytes ending at best_len + 1 */
+        BrotliUnalignedRead32(&data[cur_ix_masked + best_len - 3]) !=
+            BrotliUnalignedRead32(&source[offset + best_len - 3])) {
       continue;
     }
     {
@@ -627,8 +632,8 @@ static BROTLI_INLINE size_t FindAllCompoundDictionaryMatches(
   const uint64_t hash_mask = (~((uint64_t)0U)) >> (64 - hash_bits);
 
   const uint32_t* slot_offsets = (uint32_t*)(&self[1]);
-  const uint16_t* heads = (uint16_t*)(&slot_offsets[1u << slot_bits]);
-  const uint32_t* items = (uint32_t*)(&heads[1u << bucket_bits]);
+  const uint16_t* heads = (uint16_t*)(&slot_offsets[(size_t)1u << slot_bits]);
+  const uint32_t* items = (uint32_t*)(&heads[(size_t)1u << bucket_bits]);
   const uint8_t* source = NULL;
 
   const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
@@ -651,6 +656,8 @@ static BROTLI_INLINE size_t FindAllCompoundDictionaryMatches(
     source = (const uint8_t*)BROTLI_UNALIGNED_LOAD_PTR((const uint8_t**)tail);
   }
 
+  BROTLI_DCHECK(cur_ix_masked + max_length <= ring_buffer_mask);
+
   while (item == 0) {
     size_t offset;
     size_t distance;

data/ext/multi_compress/vendor/brotli/c/enc/hash_base.h

@@ -0,0 +1,38 @@
+/* Copyright 2025 Google Inc. All Rights Reserved.
+
+   Distributed under MIT license.
+   See file LICENSE for detail or copy at https://opensource.org/licenses/MIT
+*/
+
+/* Basic common hash functions / constants. */
+
+#ifndef THIRD_PARTY_BROTLI_ENC_HASH_BASE_H_
+#define THIRD_PARTY_BROTLI_ENC_HASH_BASE_H_
+
+#include "../common/platform.h"
+
+/* kHashMul32 multiplier has these properties:
+   * The multiplier must be odd. Otherwise we may lose the highest bit.
+   * No long streaks of ones or zeros.
+   * There is no effort to ensure that it is a prime, the oddity is enough
+     for this use.
+   * The number has been tuned heuristically against compression benchmarks. */
+static const uint32_t kHashMul32 = 0x1E35A7BD;
+static const uint64_t kHashMul64 =
+    BROTLI_MAKE_UINT64_T(0x1FE35A7Bu, 0xD3579BD3u);
+
+static BROTLI_INLINE uint32_t Hash14(const uint8_t* data) {
+  uint32_t h = BROTLI_UNALIGNED_LOAD32LE(data) * kHashMul32;
+  /* The higher bits contain more mixture from the multiplication,
+     so we take our results from there. */
+  return h >> (32 - 14);
+}
+
+static BROTLI_INLINE uint32_t Hash15(const uint8_t* data) {
+  uint32_t h = BROTLI_UNALIGNED_LOAD32LE(data) * kHashMul32;
+  /* The higher bits contain more mixture from the multiplication,
+     so we take our results from there. */
+  return h >> (32 - 15);
+}
+
+#endif  // THIRD_PARTY_BROTLI_ENC_HASH_BASE_H_

data/ext/multi_compress/vendor/brotli/c/enc/hash_forgetful_chain_inc.h

@@ -212,6 +212,9 @@ static BROTLI_INLINE void FN(FindLongestMatch)(
   const uint8_t tiny_hash = (uint8_t)(key);
   out->len = 0;
   out->len_code_delta = 0;
+
+  BROTLI_DCHECK(cur_ix_masked + max_length <= ring_buffer_mask);
+
   /* Try last distance first. */
   for (i = 0; i < NUM_LAST_DISTANCES_TO_CHECK; ++i) {
     const size_t backward = (size_t)distance_cache[i];
@@ -241,6 +244,11 @@ static BROTLI_INLINE void FN(FindLongestMatch)(
       }
     }
   }
+  /* we require matches of len >4, so increase best_len to 3, so we can compare
+   * 4 bytes all the time. */
+  if (best_len < 3) {
+    best_len = 3;
+  }
   {
     const size_t bank = key & (NUM_BANKS - 1);
     size_t backward = 0;
@@ -257,7 +265,9 @@ static BROTLI_INLINE void FN(FindLongestMatch)(
       delta = banks[bank].slots[last].delta;
       if (cur_ix_masked + best_len > ring_buffer_mask ||
           prev_ix + best_len > ring_buffer_mask ||
-          data[cur_ix_masked + best_len] != data[prev_ix + best_len]) {
+          /* compare 4 bytes ending at best_len + 1 */
+          BrotliUnalignedRead32(&data[cur_ix_masked + best_len - 3]) !=
+              BrotliUnalignedRead32(&data[prev_ix + best_len - 3])) {
         continue;
       }
       {

data/ext/multi_compress/vendor/brotli/c/enc/hash_longest_match64_inc.h

@@ -170,8 +170,16 @@ static BROTLI_INLINE void FN(FindLongestMatch)(
   score_t best_score = out->score;
   size_t best_len = out->len;
   size_t i;
+  /* Precalculate the hash key and prefetch the bucket. */
+  const size_t key = FN(HashBytes)(&data[cur_ix_masked], self->hash_mul_);
+  uint32_t* BROTLI_RESTRICT bucket = &buckets[key << self->block_bits_];
+  PREFETCH_L1(bucket);
+  if (self->block_bits_ > 4) PREFETCH_L1(bucket + 16);
   out->len = 0;
   out->len_code_delta = 0;
+
+  BROTLI_DCHECK(cur_ix_masked + max_length <= ring_buffer_mask);
+
   /* Try last distance first. */
   for (i = 0; i < (size_t)self->num_last_distances_to_check_; ++i) {
     const size_t backward = (size_t)distance_cache[i];
@@ -184,8 +192,10 @@ static BROTLI_INLINE void FN(FindLongestMatch)(
     }
     prev_ix &= ring_buffer_mask;
 
-    if (cur_ix_masked + best_len > ring_buffer_mask ||
-        prev_ix + best_len > ring_buffer_mask ||
+    if (cur_ix_masked + best_len > ring_buffer_mask) {
+      break;
+    }
+    if (prev_ix + best_len > ring_buffer_mask ||
         data[cur_ix_masked + best_len] != data[prev_ix + best_len]) {
       continue;
     }
@@ -211,9 +221,12 @@ static BROTLI_INLINE void FN(FindLongestMatch)(
       }
     }
   }
+  /* we require matches of len >4, so increase best_len to 3, so we can compare
+   * 4 bytes all the time. */
+  if (best_len < 3) {
+    best_len = 3;
+  }
   {
-    const size_t key = FN(HashBytes)(&data[cur_ix_masked], self->hash_mul_);
-    uint32_t* BROTLI_RESTRICT bucket = &buckets[key << self->block_bits_];
     const size_t down =
         (num[key] > self->block_size_) ?
         (num[key] - self->block_size_) : 0u;
@@ -228,9 +241,13 @@ static BROTLI_INLINE void FN(FindLongestMatch)(
         break;
       }
       prev_ix &= ring_buffer_mask;
-      if (cur_ix_masked + best_len > ring_buffer_mask ||
-          prev_ix + best_len > ring_buffer_mask ||
-          data[cur_ix_masked + best_len] != data[prev_ix + best_len]) {
+      if (cur_ix_masked + best_len > ring_buffer_mask) {
+        break;
+      }
+      if (prev_ix + best_len > ring_buffer_mask ||
+          /* compare 4 bytes ending at best_len + 1 */
+          BrotliUnalignedRead32(&data[cur_ix_masked + best_len - 3]) !=
+              BrotliUnalignedRead32(&data[prev_ix + best_len - 3])) {
         continue;
       }
       current4 = BrotliUnalignedRead32(data + prev_ix);

data/ext/multi_compress/vendor/brotli/c/enc/hash_longest_match64_simd_inc.h

@@ -0,0 +1,304 @@
+/* NOLINT(build/header_guard) */
+/* Copyright 2010 Google Inc. All Rights Reserved.
+
+   Distributed under MIT license.
+   See file LICENSE for detail or copy at https://opensource.org/licenses/MIT
+*/
+
+/* template parameters: FN */
+
+/* A (forgetful) hash table to the data seen by the compressor, to
+   help create backward references to previous data.
+
+   This is a hash map of fixed size (bucket_size_) to a ring buffer of
+   fixed size (block_size_). The ring buffer contains the last block_size_
+   index positions of the given hash key in the compressed data. */
+
+#define HashLongestMatch HASHER()
+
+#define TAG_HASH_BITS 8
+#define TAG_HASH_MASK ((1 << TAG_HASH_BITS) - 1)
+
+static BROTLI_INLINE size_t FN(HashTypeLength)(void) { return 8; }
+static BROTLI_INLINE size_t FN(StoreLookahead)(void) { return 8; }
+
+/* HashBytes is the function that chooses the bucket to place the address in. */
+static BROTLI_INLINE size_t FN(HashBytes)(const uint8_t* BROTLI_RESTRICT data,
+                                          uint64_t hash_mul) {
+  const uint64_t h = BROTLI_UNALIGNED_LOAD64LE(data) * hash_mul;
+  /* The higher bits contain more mixture from the multiplication,
+     so we take our results from there. */
+  return (size_t)(h >> (64 - 15 - TAG_HASH_BITS));
+}
+
+typedef struct HashLongestMatch {
+  /* Number of hash buckets. */
+  size_t bucket_size_;
+  /* Only block_size_ newest backward references are kept,
+     and the older are forgotten. */
+  size_t block_size_;
+  /* Hash multiplier tuned to match length. */
+  uint64_t hash_mul_;
+  /* Mask for accessing entries in a block (in a ring-buffer manner). */
+  uint32_t block_mask_;
+
+  int block_bits_;
+  int num_last_distances_to_check_;
+
+  /* Shortcuts. */
+  HasherCommon* common_;
+
+  /* --- Dynamic size members --- */
+
+  /* Number of entries in a particular bucket. */
+  uint16_t* num_;  /* uint16_t[bucket_size]; */
+
+  uint8_t* tags_;
+
+  /* Buckets containing block_size_ of backward references. */
+  uint32_t* buckets_;  /* uint32_t[bucket_size * block_size]; */
+} HashLongestMatch;
+
+static void FN(Initialize)(
+    HasherCommon* common, HashLongestMatch* BROTLI_RESTRICT self,
+    const BrotliEncoderParams* params) {
+  self->common_ = common;
+
+  BROTLI_UNUSED(params);
+  self->hash_mul_ = kHashMul64 << (64 - 5 * 8);
+  BROTLI_DCHECK(common->params.bucket_bits == 15);
+  self->bucket_size_ = (size_t)1 << common->params.bucket_bits;
+  self->block_bits_ = common->params.block_bits;
+  self->block_size_ = (size_t)1 << common->params.block_bits;
+  self->block_mask_ = (uint32_t)(self->block_size_ - 1);
+  self->num_last_distances_to_check_ =
+      common->params.num_last_distances_to_check;
+  self->num_ = (uint16_t*)common->extra[0];
+  self->tags_ = (uint8_t*)common->extra[1];
+  self->buckets_ = (uint32_t*)common->extra[2];
+}
+
+static void FN(Prepare)(
+    HashLongestMatch* BROTLI_RESTRICT self, BROTLI_BOOL one_shot,
+    size_t input_size, const uint8_t* BROTLI_RESTRICT data) {
+  uint16_t* BROTLI_RESTRICT num = self->num_;
+  /* Partial preparation is 100 times slower (per socket). */
+  size_t partial_prepare_threshold = self->bucket_size_ >> 6;
+  if (one_shot && input_size <= partial_prepare_threshold) {
+    size_t i;
+    for (i = 0; i < input_size; ++i) {
+      const size_t hash = FN(HashBytes)(&data[i], self->hash_mul_);
+      const size_t key = hash >> TAG_HASH_BITS;
+      num[key] = 65535;
+    }
+  } else {
+    /* Set all the bytes of num to 255, which makes each uint16_t 65535. */
+    memset(num, 255, self->bucket_size_ * sizeof(num[0]));
+  }
+}
+
+static BROTLI_INLINE void FN(HashMemAllocInBytes)(
+    const BrotliEncoderParams* params, BROTLI_BOOL one_shot,
+    size_t input_size, size_t* alloc_size) {
+  size_t bucket_size = (size_t)1 << params->hasher.bucket_bits;
+  size_t block_size = (size_t)1 << params->hasher.block_bits;
+  BROTLI_UNUSED(one_shot);
+  BROTLI_UNUSED(input_size);
+  alloc_size[0] = sizeof(uint16_t) * bucket_size;
+  alloc_size[1] = sizeof(uint8_t) * bucket_size * block_size;
+  alloc_size[2] = sizeof(uint32_t) * bucket_size * block_size;
+}
+
+/* Look at 4 bytes at &data[ix & mask].
+   Compute a hash from these, and store the value of ix at that position. */
+static BROTLI_INLINE void FN(Store)(
+    HashLongestMatch* BROTLI_RESTRICT self, const uint8_t* BROTLI_RESTRICT data,
+    const size_t mask, const size_t ix) {
+  uint16_t* BROTLI_RESTRICT num = self->num_;
+  uint8_t* BROTLI_RESTRICT tags = self->tags_;
+  uint32_t* BROTLI_RESTRICT buckets = self->buckets_;
+  const size_t hash = FN(HashBytes)(&data[ix & mask], self->hash_mul_);
+  const size_t key = hash >> TAG_HASH_BITS;
+  const uint8_t tag = hash & TAG_HASH_MASK;
+  const size_t minor_ix = num[key] & self->block_mask_;
+  const size_t offset = minor_ix + (key << self->block_bits_);
+  --num[key];
+  buckets[offset] = (uint32_t)ix;
+  tags[offset] = tag;
+}
+
+static BROTLI_INLINE void FN(StoreRange)(HashLongestMatch* BROTLI_RESTRICT self,
+    const uint8_t* BROTLI_RESTRICT data, const size_t mask,
+    const size_t ix_start, const size_t ix_end) {
+  size_t i;
+  for (i = ix_start; i < ix_end; ++i) {
+    FN(Store)(self, data, mask, i);
+  }
+}
+
+static BROTLI_INLINE void FN(StitchToPreviousBlock)(
+    HashLongestMatch* BROTLI_RESTRICT self,
+    size_t num_bytes, size_t position, const uint8_t* ringbuffer,
+    size_t ringbuffer_mask) {
+  if (num_bytes >= FN(HashTypeLength)() - 1 && position >= 3) {
+    /* Prepare the hashes for three last bytes of the last write.
+       These could not be calculated before, since they require knowledge
+       of both the previous and the current block. */
+    FN(Store)(self, ringbuffer, ringbuffer_mask, position - 3);
+    FN(Store)(self, ringbuffer, ringbuffer_mask, position - 2);
+    FN(Store)(self, ringbuffer, ringbuffer_mask, position - 1);
+  }
+}
+
+static BROTLI_INLINE void FN(PrepareDistanceCache)(
+    HashLongestMatch* BROTLI_RESTRICT self,
+    int* BROTLI_RESTRICT distance_cache) {
+  PrepareDistanceCache(distance_cache, self->num_last_distances_to_check_);
+}
+
+/* Find a longest backward match of &data[cur_ix] up to the length of
+   max_length and stores the position cur_ix in the hash table.
+
+   REQUIRES: FN(PrepareDistanceCache) must be invoked for current distance cache
+             values; if this method is invoked repeatedly with the same distance
+             cache values, it is enough to invoke FN(PrepareDistanceCache) once.
+
+   Does not look for matches longer than max_length.
+   Does not look for matches further away than max_backward.
+   Writes the best match into |out|.
+   |out|->score is updated only if a better match is found. */
+static BROTLI_INLINE void FN(FindLongestMatch)(
+    HashLongestMatch* BROTLI_RESTRICT self,
+    const BrotliEncoderDictionary* dictionary,
+    const uint8_t* BROTLI_RESTRICT data, const size_t ring_buffer_mask,
+    const int* BROTLI_RESTRICT distance_cache, const size_t cur_ix,
+    const size_t max_length, const size_t max_backward,
+    const size_t dictionary_distance, const size_t max_distance,
+    HasherSearchResult* BROTLI_RESTRICT out) {
+  uint16_t* BROTLI_RESTRICT num = self->num_;
+  uint32_t* BROTLI_RESTRICT buckets = self->buckets_;
+  uint8_t* BROTLI_RESTRICT tags = self->tags_;
+  const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
+  /* Don't accept a short copy from far away. */
+  score_t min_score = out->score;
+  score_t best_score = out->score;
+  size_t best_len = out->len;
+  size_t i;
+  /* Precalculate the hash key and prefetch the bucket. */
+  const size_t hash = FN(HashBytes)(&data[cur_ix_masked], self->hash_mul_);
+  const size_t key = hash >> TAG_HASH_BITS;
+  uint32_t* BROTLI_RESTRICT bucket = &buckets[key << self->block_bits_];
+  uint8_t* BROTLI_RESTRICT tag_bucket = &tags[key << self->block_bits_];
+  PREFETCH_L1(bucket);
+  PREFETCH_L1(tag_bucket);
+  if (self->block_bits_ > 4) PREFETCH_L1(bucket + 16);
+  out->len = 0;
+  out->len_code_delta = 0;
+
+  BROTLI_DCHECK(cur_ix_masked + max_length <= ring_buffer_mask);
+
+  /* Try last distance first. */
+  for (i = 0; i < (size_t)self->num_last_distances_to_check_; ++i) {
+    const size_t backward = (size_t)distance_cache[i];
+    size_t prev_ix = (size_t)(cur_ix - backward);
+    if (prev_ix >= cur_ix) {
+      continue;
+    }
+    if (BROTLI_PREDICT_FALSE(backward > max_backward)) {
+      continue;
+    }
+    prev_ix &= ring_buffer_mask;
+
+    if (cur_ix_masked + best_len > ring_buffer_mask) {
+      break;
+    }
+    if (prev_ix + best_len > ring_buffer_mask ||
+        data[cur_ix_masked + best_len] != data[prev_ix + best_len]) {
+      continue;
+    }
+    {
+      const size_t len = FindMatchLengthWithLimit(&data[prev_ix],
+                                                  &data[cur_ix_masked],
+                                                  max_length);
+      if (len >= 3 || (len == 2 && i < 2)) {
+        /* Comparing for >= 2 does not change the semantics, but just saves for
+           a few unnecessary binary logarithms in backward reference score,
+           since we are not interested in such short matches. */
+        score_t score = BackwardReferenceScoreUsingLastDistance(len);
+        if (best_score < score) {
+          if (i != 0) score -= BackwardReferencePenaltyUsingLastDistance(i);
+          if (best_score < score) {
+            best_score = score;
+            best_len = len;
+            out->len = best_len;
+            out->distance = backward;
+            out->score = best_score;
+          }
+        }
+      }
+    }
+  }
+  /* we require matches of len >4, so increase best_len to 3, so we can compare
+   * 4 bytes all the time. */
+  if (best_len < 3) {
+    best_len = 3;
+  }
+  {
+    const uint8_t tag = hash & TAG_HASH_MASK;
+    const uint32_t first4 = BrotliUnalignedRead32(data + cur_ix_masked);
+    const size_t max_length_m4 = max_length - 4;
+    const size_t head = (num[key] + 1) & self->block_mask_;
+    uint64_t matches =
+        GetMatchingTagMask(self->block_size_ / 16, tag, tag_bucket, head);
+    /* Mask off any matches from uninitialized tags. */
+    uint16_t n = 65535 - num[key];
+    uint64_t block_has_unused_slots = self->block_size_ > n;
+    uint64_t mask = (block_has_unused_slots << (n & (64 - 1))) - 1;
+    matches &= mask;
+    for (; matches > 0; matches &= (matches - 1)) {
+      const size_t rb_index =
+          (head + (size_t)BROTLI_TZCNT64(matches)) & self->block_mask_;
+      size_t prev_ix = bucket[rb_index];
+      uint32_t current4;
+      const size_t backward = cur_ix - prev_ix;
+      if (BROTLI_PREDICT_FALSE(backward > max_backward)) {
+        break;
+      }
+      prev_ix &= ring_buffer_mask;
+      if (cur_ix_masked + best_len > ring_buffer_mask) {
+        break;
+      }
+      if (prev_ix + best_len > ring_buffer_mask ||
+          /* compare 4 bytes ending at best_len + 1 */
+          BrotliUnalignedRead32(&data[cur_ix_masked + best_len - 3]) !=
+              BrotliUnalignedRead32(&data[prev_ix + best_len - 3])) {
+        continue;
+      }
+      current4 = BrotliUnalignedRead32(data + prev_ix);
+      if (first4 != current4) continue;
+      {
+        const size_t len = FindMatchLengthWithLimit(&data[prev_ix + 4],
+                                                    &data[cur_ix_masked + 4],
+                                                    max_length_m4) + 4;
+        const score_t score = BackwardReferenceScore(len, backward);
+        if (best_score < score) {
+          best_score = score;
+          best_len = len;
+          out->len = best_len;
+          out->distance = backward;
+          out->score = best_score;
+        }
+      }
+    }
+    bucket[num[key] & self->block_mask_] = (uint32_t)cur_ix;
+    tag_bucket[num[key] & self->block_mask_] = tag;
+    --num[key];
+  }
+  if (min_score == out->score) {
+    SearchInStaticDictionary(dictionary,
+        self->common_, &data[cur_ix_masked], max_length, dictionary_distance,
+        max_distance, out, BROTLI_FALSE);
+  }
+}
+
+#undef HashLongestMatch

data/ext/multi_compress/vendor/brotli/c/enc/hash_longest_match_inc.h

@@ -104,11 +104,13 @@ static BROTLI_INLINE void FN(HashMemAllocInBytes)(
 static BROTLI_INLINE void FN(Store)(
     HashLongestMatch* BROTLI_RESTRICT self, const uint8_t* BROTLI_RESTRICT data,
     const size_t mask, const size_t ix) {
+  uint16_t* BROTLI_RESTRICT num = self->num_;
+  uint32_t* BROTLI_RESTRICT buckets = self->buckets_;
   const uint32_t key = FN(HashBytes)(&data[ix & mask], self->hash_shift_);
-  const size_t minor_ix = self->num_[key] & self->block_mask_;
+  const size_t minor_ix = num[key] & self->block_mask_;
   const size_t offset = minor_ix + (key << self->block_bits_);
-  self->buckets_[offset] = (uint32_t)ix;
-  ++self->num_[key];
+  ++num[key];
+  buckets[offset] = (uint32_t)ix;
 }
 
 static BROTLI_INLINE void FN(StoreRange)(HashLongestMatch* BROTLI_RESTRICT self,
@@ -167,8 +169,17 @@ static BROTLI_INLINE void FN(FindLongestMatch)(
   score_t best_score = out->score;
   size_t best_len = out->len;
   size_t i;
+  /* Precalculate the hash key and prefetch the bucket. */
+  const uint32_t key =
+      FN(HashBytes)(&data[cur_ix_masked], self->hash_shift_);
+  uint32_t* BROTLI_RESTRICT bucket = &buckets[key << self->block_bits_];
+  PREFETCH_L1(bucket);
+  if (self->block_bits_ > 4) PREFETCH_L1(bucket + 16);
   out->len = 0;
   out->len_code_delta = 0;
+
+  BROTLI_DCHECK(cur_ix_masked + max_length <= ring_buffer_mask);
+
   /* Try last distance first. */
   for (i = 0; i < (size_t)self->num_last_distances_to_check_; ++i) {
     const size_t backward = (size_t)distance_cache[i];
@@ -181,8 +192,10 @@ static BROTLI_INLINE void FN(FindLongestMatch)(
     }
     prev_ix &= ring_buffer_mask;
 
-    if (cur_ix_masked + best_len > ring_buffer_mask ||
-        prev_ix + best_len > ring_buffer_mask ||
+    if (cur_ix_masked + best_len > ring_buffer_mask) {
+      break;
+    }
+    if (prev_ix + best_len > ring_buffer_mask ||
         data[cur_ix_masked + best_len] != data[prev_ix + best_len]) {
       continue;
     }
@@ -208,10 +221,12 @@ static BROTLI_INLINE void FN(FindLongestMatch)(
       }
     }
   }
+  /* we require matches of len >4, so increase best_len to 3, so we can compare
+   * 4 bytes all the time. */
+  if (best_len < 3) {
+    best_len = 3;
+  }
   {
-    const uint32_t key =
-        FN(HashBytes)(&data[cur_ix_masked], self->hash_shift_);
-    uint32_t* BROTLI_RESTRICT bucket = &buckets[key << self->block_bits_];
     const size_t down =
         (num[key] > self->block_size_) ? (num[key] - self->block_size_) : 0u;
     for (i = num[key]; i > down;) {
@@ -221,9 +236,13 @@ static BROTLI_INLINE void FN(FindLongestMatch)(
         break;
       }
       prev_ix &= ring_buffer_mask;
-      if (cur_ix_masked + best_len > ring_buffer_mask ||
-          prev_ix + best_len > ring_buffer_mask ||
-          data[cur_ix_masked + best_len] != data[prev_ix + best_len]) {
+      if (cur_ix_masked + best_len > ring_buffer_mask) {
+        break;
+      }
+      if (prev_ix + best_len > ring_buffer_mask ||
+          /* compare 4 bytes ending at best_len + 1 */
+          BrotliUnalignedRead32(&data[cur_ix_masked + best_len - 3]) !=
+              BrotliUnalignedRead32(&data[prev_ix + best_len - 3])) {
         continue;
       }
       {

data/ext/multi_compress/vendor/brotli/c/enc/hash_longest_match_quickly_inc.h

@@ -155,6 +155,7 @@ static BROTLI_INLINE void FN(FindLongestMatch)(
   uint32_t* BROTLI_RESTRICT buckets = self->buckets_;
   const size_t best_len_in = out->len;
   const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
+  /* TODO: compare 4 bytes at once (and set the minimum best len to 4) */
   int compare_char = data[cur_ix_masked + best_len_in];
   size_t key = FN(HashBytes)(&data[cur_ix_masked]);
   size_t key_out;
@@ -163,6 +164,9 @@ static BROTLI_INLINE void FN(FindLongestMatch)(
   size_t best_len = best_len_in;
   size_t cached_backward = (size_t)distance_cache[0];
   size_t prev_ix = cur_ix - cached_backward;
+
+  BROTLI_DCHECK(cur_ix_masked + max_length <= ring_buffer_mask);
+
   out->len_code_delta = 0;
   if (prev_ix < cur_ix) {
     prev_ix &= (uint32_t)ring_buffer_mask;