Outline StoreRange in the new experimental hasher

PiperOrigin-RevId: 979429901
diff --git a/c/enc/backward_references_inc_opt.h b/c/enc/backward_references_inc_opt.h
index 02e1f4f..9e1ef64 100644
--- a/c/enc/backward_references_inc_opt.h
+++ b/c/enc/backward_references_inc_opt.h
@@ -113,7 +113,7 @@
         ringbuffer, ringbuffer_mask, dist_cache, position, max_length,
         max_distance, dictionary_start + gap, params->dist.max_distance, &sr);
     if (ENABLE_COMPOUND_DICTIONARY) {
-      LookupCompoundDictionaryMatch(&params->dictionary.compound, ringbuffer,
+      LookupCompoundDictionaryMatchOpt(&params->dictionary.compound, ringbuffer,
           ringbuffer_mask, dist_cache, position, max_length,
           dictionary_start, params->dist.max_distance, &sr);
     }
@@ -144,7 +144,7 @@
             max_distance, dictionary_start + gap, params->dist.max_distance,
             &sr2);
         if (ENABLE_COMPOUND_DICTIONARY) {
-          LookupCompoundDictionaryMatch(
+          LookupCompoundDictionaryMatchOpt(
               &params->dictionary.compound, ringbuffer,
               ringbuffer_mask, dist_cache, position + 1, max_length,
               dictionary_start, params->dist.max_distance, &sr2);
diff --git a/c/enc/compound_dictionary.c b/c/enc/compound_dictionary.c
index 878b6bc..23166c8 100644
--- a/c/enc/compound_dictionary.c
+++ b/c/enc/compound_dictionary.c
@@ -205,6 +205,17 @@
       compound->chunk_source[index] =
           (const uint8_t*)BROTLI_UNALIGNED_LOAD_PTR((const uint8_t**)tail);
     }
+    {
+      PreparedDictionaryView* view = &compound->chunk_views[index];
+      view->slot_offsets = slot_offsets;
+      view->heads = heads;
+      view->items = items;
+      view->source = compound->chunk_source[index];
+      view->source_size = dictionary->source_size;
+      view->hash_shift = 64u - dictionary->bucket_bits;
+      view->slot_mask = (~((uint32_t)0U)) >> (32 - dictionary->slot_bits);
+      view->hash_mask = (~((uint64_t)0U)) >> (64 - dictionary->hash_bits);
+    }
   }
   compound->num_chunks++;
   return BROTLI_TRUE;
diff --git a/c/enc/compound_dictionary.h b/c/enc/compound_dictionary.h
index 686f92f..240a7f9 100644
--- a/c/enc/compound_dictionary.h
+++ b/c/enc/compound_dictionary.h
@@ -51,12 +51,28 @@
 BROTLI_INTERNAL void DestroyPreparedDictionary(MemoryManager* m,
     PreparedDictionary* dictionary);
 
+/* A derived view of a PreparedDictionary, optimized for faster lookup.  Used by
+ * the optimized H59 hasher.  Its fields are a pure function of
+ * PreparedDictionary's so we can generate it at attach time. */
+typedef struct PreparedDictionaryView {
+  const uint32_t* slot_offsets;
+  const uint16_t* heads;
+  const uint32_t* items;
+  const uint8_t* source;
+  uint32_t source_size;
+  uint32_t hash_shift;  /* 64 - bucket_bits */
+  uint32_t slot_mask;
+  uint64_t hash_mask;
+} PreparedDictionaryView;
+
 typedef struct CompoundDictionary {
   /* LZ77 prefix, compound dictionary */
   size_t num_chunks;
   size_t total_size;
   /* Client instances. */
   const PreparedDictionary* chunks[SHARED_BROTLI_MAX_COMPOUND_DICTS + 1];
+  /* TODO: const? */
+  PreparedDictionaryView chunk_views[SHARED_BROTLI_MAX_COMPOUND_DICTS + 1];
   const uint8_t* chunk_source[SHARED_BROTLI_MAX_COMPOUND_DICTS + 1];
   size_t chunk_offsets[SHARED_BROTLI_MAX_COMPOUND_DICTS + 1];
 
diff --git a/c/enc/hash.h b/c/enc/hash.h
index 5cca3c1..c88a6c5 100644
--- a/c/enc/hash.h
+++ b/c/enc/hash.h
@@ -292,7 +292,7 @@
 #undef HASHER
 
 #define HASHER() H59
-#include "hash_longest_match_simd_inc.h" /* NOLINT(build/include) */
+#include "hash_longest_match_simd_opt_inc.h" /* NOLINT(build/include) */
 #undef HASHER
 
 #define HASHER() H68
@@ -751,6 +751,121 @@
   return total_found;
 }
 
+static BROTLI_INLINE void FindCompoundDictionaryMatchOpt(
+    const PreparedDictionaryView* self, 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 distance_offset,
+    const size_t max_distance, HasherSearchResult* BROTLI_RESTRICT out) {
+  const uint32_t source_size = self->source_size;
+  const size_t boundary = distance_offset - source_size;
+  const uint32_t hash_shift = self->hash_shift;
+  const uint32_t slot_mask = self->slot_mask;
+  const uint64_t hash_mask = self->hash_mask;
+
+  const uint32_t* slot_offsets = self->slot_offsets;
+  const uint16_t* heads = self->heads;
+  const uint32_t* items = self->items;
+  const uint8_t* source = self->source;
+
+  const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
+  score_t best_score = out->score;
+  size_t best_len = out->len;
+  size_t i;
+  const uint64_t h =
+      (BROTLI_UNALIGNED_LOAD64LE(&data[cur_ix_masked]) & hash_mask) *
+      kPreparedDictionaryHashMul64Long;
+  const uint32_t key = (uint32_t)(h >> hash_shift);
+  const uint32_t slot = key & slot_mask;
+  const uint32_t head = heads[key];
+  const uint32_t* BROTLI_RESTRICT chain = &items[slot_offsets[slot] + head];
+  uint32_t item = (head == 0xFFFF) ? 1 : 0;
+
+  BROTLI_DCHECK(cur_ix_masked + max_length <= ring_buffer_mask + 1);
+
+  for (i = 0; i < 4; ++i) {
+    const size_t distance = (size_t)distance_cache[i];
+    size_t offset;
+    size_t limit;
+    size_t len;
+    if (distance <= boundary || distance > distance_offset) continue;
+    offset = distance_offset - distance;
+    limit = source_size - offset;
+    limit = limit > max_length ? max_length : limit;
+    len = FindMatchLengthWithLimit(&source[offset], &data[cur_ix_masked],
+                                   limit);
+    if (len >= 2) {
+      score_t score = BackwardReferenceScoreUsingLastDistance(len);
+      if (best_score < score) {
+        if (i != 0) score -= BackwardReferencePenaltyUsingLastDistance(i);
+        if (best_score < score) {
+          best_score = score;
+          if (len > best_len) best_len = len;
+          out->len = len;
+          out->len_code_delta = 0;
+          out->distance = distance;
+          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;
+  }
+  while (item == 0) {
+    size_t offset;
+    size_t distance;
+    size_t limit;
+    item = *chain;
+    chain++;
+    offset = item & 0x7FFFFFFF;
+    item &= 0x80000000;
+    distance = distance_offset - offset;
+    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 ||
+        /* compare 4 bytes ending at best_len + 1 */
+        BrotliUnalignedRead32(&data[cur_ix_masked + best_len - 3]) !=
+            BrotliUnalignedRead32(&source[offset + best_len - 3])) {
+      continue;
+    }
+    {
+      const size_t len = FindMatchLengthWithLimit(&source[offset],
+                                                  &data[cur_ix_masked],
+                                                  limit);
+      if (len >= 4) {
+        score_t score = BackwardReferenceScore(len, distance);
+        if (best_score < score) {
+          best_score = score;
+          best_len = len;
+          out->len = best_len;
+          out->len_code_delta = 0;
+          out->distance = distance;
+          out->score = best_score;
+        }
+      }
+    }
+  }
+}
+
+static BROTLI_INLINE void LookupCompoundDictionaryMatchOpt(
+    const CompoundDictionary* addon, 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_ring_buffer_distance, const size_t max_distance,
+    HasherSearchResult* sr) {
+  size_t base_offset = max_ring_buffer_distance + 1 + addon->total_size - 1;
+  size_t d;
+  for (d = 0; d < addon->num_chunks; ++d) {
+    FindCompoundDictionaryMatchOpt(
+        &addon->chunk_views[d], data, ring_buffer_mask,
+        distance_cache, cur_ix, max_length,
+        base_offset - addon->chunk_offsets[d], max_distance, sr);
+  }
+}
+
 #if defined(__cplusplus) || defined(c_plusplus)
 }  /* extern "C" */
 #endif
diff --git a/c/enc/hash_longest_match_simd_opt_inc.h b/c/enc/hash_longest_match_simd_opt_inc.h
new file mode 100644
index 0000000..8a1cc8a
--- /dev/null
+++ b/c/enc/hash_longest_match_simd_opt_inc.h
@@ -0,0 +1,278 @@
+/* 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 4; }
+static BROTLI_INLINE size_t FN(StoreLookahead)(void) { return 4; }
+/* HashBytes is the function that chooses the bucket to place the address in. */
+static uint32_t FN(HashBytes)(
+    const uint8_t* BROTLI_RESTRICT data, const int shift) {
+  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 (uint32_t)(h >> shift);
+}
+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_;
+  /* Left-shift for computing hash bucket index from hash value. */
+  int hash_shift_;
+  /* 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_shift_ = 32 - common->params.bucket_bits - TAG_HASH_BITS;
+  self->bucket_size_ = (size_t)1 << common->params.bucket_bits;
+  self->block_size_ = (size_t)1 << common->params.block_bits;
+  self->block_mask_ = (uint32_t)(self->block_size_ - 1);
+  self->num_ = (uint16_t*)common->extra[0];
+  self->tags_ = (uint8_t*)common->extra[1];
+  self->buckets_ = (uint32_t*)common->extra[2];
+  self->block_bits_ = common->params.block_bits;
+  self->num_last_distances_to_check_ =
+      common->params.num_last_distances_to_check;
+}
+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 uint32_t hash = FN(HashBytes)(&data[i], self->hash_shift_);
+      const uint32_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_shift_);
+  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_MAYBE_INLINE BROTLI_NOINLINE 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 uint32_t hash =
+      FN(HashBytes)(&data[cur_ix_masked], self->hash_shift_);
+  const uint32_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 + 1);
+
+  /* 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 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];
+      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;
+      }
+      {
+        const size_t len = FindMatchLengthWithLimit(&data[prev_ix],
+                                                    &data[cur_ix_masked],
+                                                    max_length);
+        if (len >= 4) {
+          /* Comparing for >= 3 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 = 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