1 /* NOLINT(build/header_guard) */
2 /* Copyright 2010 Google Inc. All Rights Reserved.
3
4 Distributed under MIT license.
5 See file LICENSE for detail or copy at https://opensource.org/licenses/MIT
6 */
7
8 /* template parameters: FN, BUCKET_BITS, BUCKET_SWEEP, HASH_LEN,
9 USE_DICTIONARY
10 */
11
12 #define HashLongestMatchQuickly HASHER()
13
14 #define BUCKET_SIZE (1 << BUCKET_BITS)
15
16 #define HASH_MAP_SIZE (4 << BUCKET_BITS)
17
FN(HashTypeLength)18 static BROTLI_INLINE size_t FN(HashTypeLength)(void) { return 8; }
FN(StoreLookahead)19 static BROTLI_INLINE size_t FN(StoreLookahead)(void) { return 8; }
20
21 /* HashBytes is the function that chooses the bucket to place
22 the address in. The HashLongestMatch and HashLongestMatchQuickly
23 classes have separate, different implementations of hashing. */
FN(HashBytes)24 static uint32_t FN(HashBytes)(const uint8_t* data) {
25 const uint64_t h = ((BROTLI_UNALIGNED_LOAD64LE(data) << (64 - 8 * HASH_LEN)) *
26 kHashMul64);
27 /* The higher bits contain more mixture from the multiplication,
28 so we take our results from there. */
29 return (uint32_t)(h >> (64 - BUCKET_BITS));
30 }
31
32 /* A (forgetful) hash table to the data seen by the compressor, to
33 help create backward references to previous data.
34
35 This is a hash map of fixed size (BUCKET_SIZE). Starting from the
36 given index, BUCKET_SWEEP buckets are used to store values of a key. */
37 typedef struct HashLongestMatchQuickly {
38 uint32_t buckets_[BUCKET_SIZE + BUCKET_SWEEP];
39 } HashLongestMatchQuickly;
40
FN(Self)41 static BROTLI_INLINE HashLongestMatchQuickly* FN(Self)(HasherHandle handle) {
42 return (HashLongestMatchQuickly*)&(GetHasherCommon(handle)[1]);
43 }
44
FN(Initialize)45 static void FN(Initialize)(
46 HasherHandle handle, const BrotliEncoderParams* params) {
47 BROTLI_UNUSED(handle);
48 BROTLI_UNUSED(params);
49 }
50
FN(Prepare)51 static void FN(Prepare)(HasherHandle handle, BROTLI_BOOL one_shot,
52 size_t input_size, const uint8_t* data) {
53 HashLongestMatchQuickly* self = FN(Self)(handle);
54 /* Partial preparation is 100 times slower (per socket). */
55 size_t partial_prepare_threshold = HASH_MAP_SIZE >> 7;
56 if (one_shot && input_size <= partial_prepare_threshold) {
57 size_t i;
58 for (i = 0; i < input_size; ++i) {
59 const uint32_t key = FN(HashBytes)(&data[i]);
60 memset(&self->buckets_[key], 0, BUCKET_SWEEP * sizeof(self->buckets_[0]));
61 }
62 } else {
63 /* It is not strictly necessary to fill this buffer here, but
64 not filling will make the results of the compression stochastic
65 (but correct). This is because random data would cause the
66 system to find accidentally good backward references here and there. */
67 memset(&self->buckets_[0], 0, sizeof(self->buckets_));
68 }
69 }
70
FN(HashMemAllocInBytes)71 static BROTLI_INLINE size_t FN(HashMemAllocInBytes)(
72 const BrotliEncoderParams* params, BROTLI_BOOL one_shot,
73 size_t input_size) {
74 BROTLI_UNUSED(params);
75 BROTLI_UNUSED(one_shot);
76 BROTLI_UNUSED(input_size);
77 return sizeof(HashLongestMatchQuickly);
78 }
79
80 /* Look at 5 bytes at &data[ix & mask].
81 Compute a hash from these, and store the value somewhere within
82 [ix .. ix+3]. */
FN(Store)83 static BROTLI_INLINE void FN(Store)(HasherHandle handle,
84 const uint8_t *data, const size_t mask, const size_t ix) {
85 const uint32_t key = FN(HashBytes)(&data[ix & mask]);
86 /* Wiggle the value with the bucket sweep range. */
87 const uint32_t off = (ix >> 3) % BUCKET_SWEEP;
88 FN(Self)(handle)->buckets_[key + off] = (uint32_t)ix;
89 }
90
FN(StoreRange)91 static BROTLI_INLINE void FN(StoreRange)(HasherHandle handle,
92 const uint8_t *data, const size_t mask, const size_t ix_start,
93 const size_t ix_end) {
94 size_t i;
95 for (i = ix_start; i < ix_end; ++i) {
96 FN(Store)(handle, data, mask, i);
97 }
98 }
99
FN(StitchToPreviousBlock)100 static BROTLI_INLINE void FN(StitchToPreviousBlock)(
101 HasherHandle handle, size_t num_bytes, size_t position,
102 const uint8_t* ringbuffer, size_t ringbuffer_mask) {
103 if (num_bytes >= FN(HashTypeLength)() - 1 && position >= 3) {
104 /* Prepare the hashes for three last bytes of the last write.
105 These could not be calculated before, since they require knowledge
106 of both the previous and the current block. */
107 FN(Store)(handle, ringbuffer, ringbuffer_mask, position - 3);
108 FN(Store)(handle, ringbuffer, ringbuffer_mask, position - 2);
109 FN(Store)(handle, ringbuffer, ringbuffer_mask, position - 1);
110 }
111 }
112
FN(PrepareDistanceCache)113 static BROTLI_INLINE void FN(PrepareDistanceCache)(
114 HasherHandle handle, int* BROTLI_RESTRICT distance_cache) {
115 BROTLI_UNUSED(handle);
116 BROTLI_UNUSED(distance_cache);
117 }
118
119 /* Find a longest backward match of &data[cur_ix & ring_buffer_mask]
120 up to the length of max_length and stores the position cur_ix in the
121 hash table.
122
123 Does not look for matches longer than max_length.
124 Does not look for matches further away than max_backward.
125 Writes the best match into |out|.
126 |out|->score is updated only if a better match is found. */
FN(FindLongestMatch)127 static BROTLI_INLINE void FN(FindLongestMatch)(
128 HasherHandle handle, const BrotliDictionary* dictionary,
129 const uint16_t* dictionary_hash, const uint8_t* BROTLI_RESTRICT data,
130 const size_t ring_buffer_mask, const int* BROTLI_RESTRICT distance_cache,
131 const size_t cur_ix, const size_t max_length, const size_t max_backward,
132 const size_t gap, HasherSearchResult* BROTLI_RESTRICT out) {
133 HashLongestMatchQuickly* self = FN(Self)(handle);
134 const size_t best_len_in = out->len;
135 const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
136 const uint32_t key = FN(HashBytes)(&data[cur_ix_masked]);
137 int compare_char = data[cur_ix_masked + best_len_in];
138 score_t min_score = out->score;
139 score_t best_score = out->score;
140 size_t best_len = best_len_in;
141 size_t cached_backward = (size_t)distance_cache[0];
142 size_t prev_ix = cur_ix - cached_backward;
143 out->len_code_delta = 0;
144 if (prev_ix < cur_ix) {
145 prev_ix &= (uint32_t)ring_buffer_mask;
146 if (compare_char == data[prev_ix + best_len]) {
147 size_t len = FindMatchLengthWithLimit(&data[prev_ix],
148 &data[cur_ix_masked],
149 max_length);
150 if (len >= 4) {
151 const score_t score = BackwardReferenceScoreUsingLastDistance(len);
152 if (best_score < score) {
153 best_score = score;
154 best_len = len;
155 out->len = len;
156 out->distance = cached_backward;
157 out->score = best_score;
158 compare_char = data[cur_ix_masked + best_len];
159 if (BUCKET_SWEEP == 1) {
160 self->buckets_[key] = (uint32_t)cur_ix;
161 return;
162 }
163 }
164 }
165 }
166 }
167 if (BUCKET_SWEEP == 1) {
168 size_t backward;
169 size_t len;
170 /* Only one to look for, don't bother to prepare for a loop. */
171 prev_ix = self->buckets_[key];
172 self->buckets_[key] = (uint32_t)cur_ix;
173 backward = cur_ix - prev_ix;
174 prev_ix &= (uint32_t)ring_buffer_mask;
175 if (compare_char != data[prev_ix + best_len_in]) {
176 return;
177 }
178 if (BROTLI_PREDICT_FALSE(backward == 0 || backward > max_backward)) {
179 return;
180 }
181 len = FindMatchLengthWithLimit(&data[prev_ix],
182 &data[cur_ix_masked],
183 max_length);
184 if (len >= 4) {
185 const score_t score = BackwardReferenceScore(len, backward);
186 if (best_score < score) {
187 out->len = len;
188 out->distance = backward;
189 out->score = score;
190 return;
191 }
192 }
193 } else {
194 uint32_t *bucket = self->buckets_ + key;
195 int i;
196 prev_ix = *bucket++;
197 for (i = 0; i < BUCKET_SWEEP; ++i, prev_ix = *bucket++) {
198 const size_t backward = cur_ix - prev_ix;
199 size_t len;
200 prev_ix &= (uint32_t)ring_buffer_mask;
201 if (compare_char != data[prev_ix + best_len]) {
202 continue;
203 }
204 if (BROTLI_PREDICT_FALSE(backward == 0 || backward > max_backward)) {
205 continue;
206 }
207 len = FindMatchLengthWithLimit(&data[prev_ix],
208 &data[cur_ix_masked],
209 max_length);
210 if (len >= 4) {
211 const score_t score = BackwardReferenceScore(len, backward);
212 if (best_score < score) {
213 best_score = score;
214 best_len = len;
215 out->len = best_len;
216 out->distance = backward;
217 out->score = score;
218 compare_char = data[cur_ix_masked + best_len];
219 }
220 }
221 }
222 }
223 if (USE_DICTIONARY && min_score == out->score) {
224 SearchInStaticDictionary(dictionary, dictionary_hash,
225 handle, &data[cur_ix_masked], max_length, max_backward + gap, out,
226 BROTLI_TRUE);
227 }
228 self->buckets_[key + ((cur_ix >> 3) % BUCKET_SWEEP)] = (uint32_t)cur_ix;
229 }
230
231 #undef HASH_MAP_SIZE
232 #undef BUCKET_SIZE
233
234 #undef HashLongestMatchQuickly
235