brotli 0.1.1 → 0.1.2
Sign up to get free protection for your applications and to get access to all the features.
- checksums.yaml +4 -4
- data/ext/brotli/brotli.cc +114 -24
- data/ext/brotli/brotli.h +0 -1
- data/ext/brotli/extconf.rb +30 -23
- data/lib/brotli/version.rb +1 -1
- data/vendor/brotli/LICENSE +1 -1
- data/vendor/brotli/dec/Makefile +1 -1
- data/vendor/brotli/dec/bit_reader.c +3 -3
- data/vendor/brotli/dec/bit_reader.h +25 -27
- data/vendor/brotli/dec/context.h +4 -4
- data/vendor/brotli/dec/decode.c +410 -486
- data/vendor/brotli/dec/decode.h +101 -105
- data/vendor/brotli/dec/dictionary.c +1 -1
- data/vendor/brotli/dec/dictionary.h +7 -8
- data/vendor/brotli/dec/huffman.c +103 -105
- data/vendor/brotli/dec/huffman.h +18 -18
- data/vendor/brotli/dec/port.h +52 -40
- data/vendor/brotli/dec/prefix.h +2 -0
- data/vendor/brotli/dec/state.c +13 -19
- data/vendor/brotli/dec/state.h +25 -39
- data/vendor/brotli/dec/transform.h +38 -44
- data/vendor/brotli/dec/types.h +2 -2
- data/vendor/brotli/enc/Makefile +1 -1
- data/vendor/brotli/enc/backward_references.cc +455 -359
- data/vendor/brotli/enc/backward_references.h +79 -3
- data/vendor/brotli/enc/bit_cost.h +54 -32
- data/vendor/brotli/enc/block_splitter.cc +285 -193
- data/vendor/brotli/enc/block_splitter.h +4 -12
- data/vendor/brotli/enc/brotli_bit_stream.cc +623 -324
- data/vendor/brotli/enc/brotli_bit_stream.h +76 -37
- data/vendor/brotli/enc/cluster.h +161 -120
- data/vendor/brotli/enc/command.h +60 -37
- data/vendor/brotli/enc/compress_fragment.cc +701 -0
- data/vendor/brotli/enc/compress_fragment.h +47 -0
- data/vendor/brotli/enc/compress_fragment_two_pass.cc +524 -0
- data/vendor/brotli/enc/compress_fragment_two_pass.h +40 -0
- data/vendor/brotli/enc/compressor.h +15 -0
- data/vendor/brotli/enc/context.h +1 -1
- data/vendor/brotli/enc/dictionary.h +2 -2
- data/vendor/brotli/enc/encode.cc +819 -286
- data/vendor/brotli/enc/encode.h +38 -15
- data/vendor/brotli/enc/encode_parallel.cc +40 -42
- data/vendor/brotli/enc/entropy_encode.cc +144 -147
- data/vendor/brotli/enc/entropy_encode.h +32 -8
- data/vendor/brotli/enc/entropy_encode_static.h +572 -0
- data/vendor/brotli/enc/fast_log.h +7 -40
- data/vendor/brotli/enc/find_match_length.h +9 -9
- data/vendor/brotli/enc/hash.h +462 -154
- data/vendor/brotli/enc/histogram.cc +6 -6
- data/vendor/brotli/enc/histogram.h +13 -13
- data/vendor/brotli/enc/literal_cost.cc +45 -45
- data/vendor/brotli/enc/metablock.cc +92 -89
- data/vendor/brotli/enc/metablock.h +12 -12
- data/vendor/brotli/enc/port.h +7 -16
- data/vendor/brotli/enc/prefix.h +23 -22
- data/vendor/brotli/enc/ringbuffer.h +75 -29
- data/vendor/brotli/enc/static_dict.cc +56 -48
- data/vendor/brotli/enc/static_dict.h +5 -5
- data/vendor/brotli/enc/streams.cc +1 -1
- data/vendor/brotli/enc/streams.h +5 -5
- data/vendor/brotli/enc/transform.h +40 -35
- data/vendor/brotli/enc/types.h +2 -0
- data/vendor/brotli/enc/utf8_util.cc +3 -2
- data/vendor/brotli/enc/write_bits.h +6 -6
- metadata +9 -5
- data/vendor/brotli/dec/streams.c +0 -102
- data/vendor/brotli/dec/streams.h +0 -95
data/vendor/brotli/enc/hash.h
CHANGED
@@ -10,12 +10,10 @@
|
|
10
10
|
#ifndef BROTLI_ENC_HASH_H_
|
11
11
|
#define BROTLI_ENC_HASH_H_
|
12
12
|
|
13
|
-
#include <string.h>
|
14
13
|
#include <sys/types.h>
|
15
14
|
#include <algorithm>
|
16
|
-
#include <
|
17
|
-
#include <
|
18
|
-
#include <string>
|
15
|
+
#include <cstring>
|
16
|
+
#include <limits>
|
19
17
|
|
20
18
|
#include "./dictionary_hash.h"
|
21
19
|
#include "./fast_log.h"
|
@@ -28,15 +26,20 @@
|
|
28
26
|
|
29
27
|
namespace brotli {
|
30
28
|
|
31
|
-
static const
|
29
|
+
static const size_t kMaxTreeSearchDepth = 64;
|
30
|
+
static const size_t kMaxTreeCompLength = 128;
|
31
|
+
|
32
|
+
static const uint32_t kDistanceCacheIndex[] = {
|
32
33
|
0, 1, 2, 3, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1,
|
33
34
|
};
|
34
35
|
static const int kDistanceCacheOffset[] = {
|
35
36
|
0, 0, 0, 0, -1, 1, -2, 2, -3, 3, -1, 1, -2, 2, -3, 3
|
36
37
|
};
|
37
38
|
|
38
|
-
static const
|
39
|
-
static const
|
39
|
+
static const uint32_t kCutoffTransformsCount = 10;
|
40
|
+
static const uint8_t kCutoffTransforms[] = {
|
41
|
+
0, 12, 27, 23, 42, 63, 56, 48, 59, 64
|
42
|
+
};
|
40
43
|
|
41
44
|
// kHashMul32 multiplier has these properties:
|
42
45
|
// * The multiplier must be odd. Otherwise we may lose the highest bit.
|
@@ -68,41 +71,47 @@ inline uint32_t Hash(const uint8_t *data) {
|
|
68
71
|
// This function is used to sometimes discard a longer backward reference
|
69
72
|
// when it is not much longer and the bit cost for encoding it is more
|
70
73
|
// than the saved literals.
|
71
|
-
|
72
|
-
|
73
|
-
|
74
|
+
//
|
75
|
+
// backward_reference_offset MUST be positive.
|
76
|
+
inline double BackwardReferenceScore(size_t copy_length,
|
77
|
+
size_t backward_reference_offset) {
|
78
|
+
return 5.4 * static_cast<double>(copy_length) -
|
79
|
+
1.20 * Log2FloorNonZero(backward_reference_offset);
|
74
80
|
}
|
75
81
|
|
76
|
-
inline double BackwardReferenceScoreUsingLastDistance(
|
77
|
-
|
82
|
+
inline double BackwardReferenceScoreUsingLastDistance(size_t copy_length,
|
83
|
+
size_t distance_short_code) {
|
78
84
|
static const double kDistanceShortCodeBitCost[16] = {
|
79
85
|
-0.6, 0.95, 1.17, 1.27,
|
80
86
|
0.93, 0.93, 0.96, 0.96, 0.99, 0.99,
|
81
87
|
1.05, 1.05, 1.15, 1.15, 1.25, 1.25
|
82
88
|
};
|
83
|
-
return 5.4 * copy_length -
|
89
|
+
return 5.4 * static_cast<double>(copy_length) -
|
90
|
+
kDistanceShortCodeBitCost[distance_short_code];
|
84
91
|
}
|
85
92
|
|
86
93
|
struct BackwardMatch {
|
87
|
-
BackwardMatch() : distance(0), length_and_code(0) {}
|
94
|
+
BackwardMatch(void) : distance(0), length_and_code(0) {}
|
88
95
|
|
89
|
-
BackwardMatch(
|
90
|
-
: distance(dist)
|
96
|
+
BackwardMatch(size_t dist, size_t len)
|
97
|
+
: distance(static_cast<uint32_t>(dist))
|
98
|
+
, length_and_code(static_cast<uint32_t>(len << 5)) {}
|
91
99
|
|
92
|
-
BackwardMatch(
|
93
|
-
: distance(dist)
|
94
|
-
|
100
|
+
BackwardMatch(size_t dist, size_t len, size_t len_code)
|
101
|
+
: distance(static_cast<uint32_t>(dist))
|
102
|
+
, length_and_code(static_cast<uint32_t>(
|
103
|
+
(len << 5) | (len == len_code ? 0 : len_code))) {}
|
95
104
|
|
96
|
-
|
105
|
+
size_t length(void) const {
|
97
106
|
return length_and_code >> 5;
|
98
107
|
}
|
99
|
-
|
100
|
-
|
108
|
+
size_t length_code(void) const {
|
109
|
+
size_t code = length_and_code & 31;
|
101
110
|
return code ? code : length();
|
102
111
|
}
|
103
112
|
|
104
|
-
|
105
|
-
|
113
|
+
uint32_t distance;
|
114
|
+
uint32_t length_and_code;
|
106
115
|
};
|
107
116
|
|
108
117
|
// A (forgetful) hash table to the data seen by the compressor, to
|
@@ -113,18 +122,31 @@ struct BackwardMatch {
|
|
113
122
|
template <int kBucketBits, int kBucketSweep, bool kUseDictionary>
|
114
123
|
class HashLongestMatchQuickly {
|
115
124
|
public:
|
116
|
-
HashLongestMatchQuickly() {
|
125
|
+
HashLongestMatchQuickly(void) {
|
117
126
|
Reset();
|
118
127
|
}
|
119
|
-
void Reset() {
|
120
|
-
|
121
|
-
// not filling will make the results of the compression stochastic
|
122
|
-
// (but correct). This is because random data would cause the
|
123
|
-
// system to find accidentally good backward references here and there.
|
124
|
-
memset(&buckets_[0], 0, sizeof(buckets_));
|
128
|
+
void Reset(void) {
|
129
|
+
need_init_ = true;
|
125
130
|
num_dict_lookups_ = 0;
|
126
131
|
num_dict_matches_ = 0;
|
127
132
|
}
|
133
|
+
void Init(void) {
|
134
|
+
if (need_init_) {
|
135
|
+
// It is not strictly necessary to fill this buffer here, but
|
136
|
+
// not filling will make the results of the compression stochastic
|
137
|
+
// (but correct). This is because random data would cause the
|
138
|
+
// system to find accidentally good backward references here and there.
|
139
|
+
memset(&buckets_[0], 0, sizeof(buckets_));
|
140
|
+
need_init_ = false;
|
141
|
+
}
|
142
|
+
}
|
143
|
+
void InitForData(const uint8_t* data, size_t num) {
|
144
|
+
for (size_t i = 0; i < num; ++i) {
|
145
|
+
const uint32_t key = HashBytes(&data[i]);
|
146
|
+
memset(&buckets_[key], 0, kBucketSweep * sizeof(buckets_[0]));
|
147
|
+
need_init_ = false;
|
148
|
+
}
|
149
|
+
}
|
128
150
|
// Look at 4 bytes at data.
|
129
151
|
// Compute a hash from these, and store the value somewhere within
|
130
152
|
// [ix .. ix+3].
|
@@ -136,7 +158,8 @@ class HashLongestMatchQuickly {
|
|
136
158
|
}
|
137
159
|
|
138
160
|
// Find a longest backward match of &ring_buffer[cur_ix & ring_buffer_mask]
|
139
|
-
// up to the length of max_length
|
161
|
+
// up to the length of max_length and stores the position cur_ix in the
|
162
|
+
// hash table.
|
140
163
|
//
|
141
164
|
// Does not look for matches longer than max_length.
|
142
165
|
// Does not look for matches further away than max_backward.
|
@@ -146,27 +169,28 @@ class HashLongestMatchQuickly {
|
|
146
169
|
inline bool FindLongestMatch(const uint8_t * __restrict ring_buffer,
|
147
170
|
const size_t ring_buffer_mask,
|
148
171
|
const int* __restrict distance_cache,
|
149
|
-
const
|
150
|
-
const
|
151
|
-
const
|
152
|
-
|
153
|
-
|
154
|
-
|
172
|
+
const size_t cur_ix,
|
173
|
+
const size_t max_length,
|
174
|
+
const size_t max_backward,
|
175
|
+
size_t * __restrict best_len_out,
|
176
|
+
size_t * __restrict best_len_code_out,
|
177
|
+
size_t * __restrict best_distance_out,
|
155
178
|
double* __restrict best_score_out) {
|
156
|
-
const
|
179
|
+
const size_t best_len_in = *best_len_out;
|
157
180
|
const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
|
181
|
+
const uint32_t key = HashBytes(&ring_buffer[cur_ix_masked]);
|
158
182
|
int compare_char = ring_buffer[cur_ix_masked + best_len_in];
|
159
183
|
double best_score = *best_score_out;
|
160
|
-
|
161
|
-
|
162
|
-
|
184
|
+
size_t best_len = best_len_in;
|
185
|
+
size_t cached_backward = static_cast<size_t>(distance_cache[0]);
|
186
|
+
size_t prev_ix = cur_ix - cached_backward;
|
163
187
|
bool match_found = false;
|
164
188
|
if (prev_ix < cur_ix) {
|
165
189
|
prev_ix &= static_cast<uint32_t>(ring_buffer_mask);
|
166
190
|
if (compare_char == ring_buffer[prev_ix + best_len]) {
|
167
|
-
|
168
|
-
|
169
|
-
|
191
|
+
size_t len = FindMatchLengthWithLimit(&ring_buffer[prev_ix],
|
192
|
+
&ring_buffer[cur_ix_masked],
|
193
|
+
max_length);
|
170
194
|
if (len >= 4) {
|
171
195
|
best_score = BackwardReferenceScoreUsingLastDistance(len, 0);
|
172
196
|
best_len = len;
|
@@ -176,6 +200,7 @@ class HashLongestMatchQuickly {
|
|
176
200
|
*best_score_out = best_score;
|
177
201
|
compare_char = ring_buffer[cur_ix_masked + best_len];
|
178
202
|
if (kBucketSweep == 1) {
|
203
|
+
buckets_[key] = static_cast<uint32_t>(cur_ix);
|
179
204
|
return true;
|
180
205
|
} else {
|
181
206
|
match_found = true;
|
@@ -183,11 +208,11 @@ class HashLongestMatchQuickly {
|
|
183
208
|
}
|
184
209
|
}
|
185
210
|
}
|
186
|
-
const uint32_t key = HashBytes(&ring_buffer[cur_ix_masked]);
|
187
211
|
if (kBucketSweep == 1) {
|
188
212
|
// Only one to look for, don't bother to prepare for a loop.
|
189
213
|
prev_ix = buckets_[key];
|
190
|
-
|
214
|
+
buckets_[key] = static_cast<uint32_t>(cur_ix);
|
215
|
+
size_t backward = cur_ix - prev_ix;
|
191
216
|
prev_ix &= static_cast<uint32_t>(ring_buffer_mask);
|
192
217
|
if (compare_char != ring_buffer[prev_ix + best_len_in]) {
|
193
218
|
return false;
|
@@ -195,9 +220,9 @@ class HashLongestMatchQuickly {
|
|
195
220
|
if (PREDICT_FALSE(backward == 0 || backward > max_backward)) {
|
196
221
|
return false;
|
197
222
|
}
|
198
|
-
const
|
199
|
-
|
200
|
-
|
223
|
+
const size_t len = FindMatchLengthWithLimit(&ring_buffer[prev_ix],
|
224
|
+
&ring_buffer[cur_ix_masked],
|
225
|
+
max_length);
|
201
226
|
if (len >= 4) {
|
202
227
|
*best_len_out = len;
|
203
228
|
*best_len_code_out = len;
|
@@ -209,7 +234,7 @@ class HashLongestMatchQuickly {
|
|
209
234
|
uint32_t *bucket = buckets_ + key;
|
210
235
|
prev_ix = *bucket++;
|
211
236
|
for (int i = 0; i < kBucketSweep; ++i, prev_ix = *bucket++) {
|
212
|
-
const
|
237
|
+
const size_t backward = cur_ix - prev_ix;
|
213
238
|
prev_ix &= static_cast<uint32_t>(ring_buffer_mask);
|
214
239
|
if (compare_char != ring_buffer[prev_ix + best_len]) {
|
215
240
|
continue;
|
@@ -217,10 +242,9 @@ class HashLongestMatchQuickly {
|
|
217
242
|
if (PREDICT_FALSE(backward == 0 || backward > max_backward)) {
|
218
243
|
continue;
|
219
244
|
}
|
220
|
-
const
|
221
|
-
|
222
|
-
|
223
|
-
max_length);
|
245
|
+
const size_t len = FindMatchLengthWithLimit(&ring_buffer[prev_ix],
|
246
|
+
&ring_buffer[cur_ix_masked],
|
247
|
+
max_length);
|
224
248
|
if (len >= 4) {
|
225
249
|
const double score = BackwardReferenceScore(len, backward);
|
226
250
|
if (best_score < score) {
|
@@ -242,19 +266,20 @@ class HashLongestMatchQuickly {
|
|
242
266
|
const uint32_t dict_key = Hash<14>(&ring_buffer[cur_ix_masked]) << 1;
|
243
267
|
const uint16_t v = kStaticDictionaryHash[dict_key];
|
244
268
|
if (v > 0) {
|
245
|
-
const
|
246
|
-
const
|
247
|
-
const
|
269
|
+
const uint32_t len = v & 31;
|
270
|
+
const uint32_t dist = v >> 5;
|
271
|
+
const size_t offset =
|
272
|
+
kBrotliDictionaryOffsetsByLength[len] + len * dist;
|
248
273
|
if (len <= max_length) {
|
249
|
-
const
|
274
|
+
const size_t matchlen =
|
250
275
|
FindMatchLengthWithLimit(&ring_buffer[cur_ix_masked],
|
251
276
|
&kBrotliDictionary[offset], len);
|
252
|
-
if (matchlen > len
|
253
|
-
const
|
254
|
-
const
|
255
|
-
transform_id * (
|
277
|
+
if (matchlen + kCutoffTransformsCount > len && matchlen > 0) {
|
278
|
+
const size_t transform_id = kCutoffTransforms[len - matchlen];
|
279
|
+
const size_t word_id =
|
280
|
+
transform_id * (1u << kBrotliDictionarySizeBitsByLength[len]) +
|
256
281
|
dist;
|
257
|
-
const
|
282
|
+
const size_t backward = max_backward + word_id + 1;
|
258
283
|
const double score = BackwardReferenceScore(matchlen, backward);
|
259
284
|
if (best_score < score) {
|
260
285
|
++num_dict_matches_;
|
@@ -264,12 +289,14 @@ class HashLongestMatchQuickly {
|
|
264
289
|
*best_len_code_out = len;
|
265
290
|
*best_distance_out = backward;
|
266
291
|
*best_score_out = best_score;
|
267
|
-
|
292
|
+
match_found = true;
|
268
293
|
}
|
269
294
|
}
|
270
295
|
}
|
271
296
|
}
|
272
297
|
}
|
298
|
+
const uint32_t off = (cur_ix >> 3) % kBucketSweep;
|
299
|
+
buckets_[key + off] = static_cast<uint32_t>(cur_ix);
|
273
300
|
return match_found;
|
274
301
|
}
|
275
302
|
|
@@ -287,16 +314,17 @@ class HashLongestMatchQuickly {
|
|
287
314
|
return static_cast<uint32_t>(h >> (64 - kBucketBits));
|
288
315
|
}
|
289
316
|
|
317
|
+
enum { kHashMapSize = 4 << kBucketBits };
|
318
|
+
|
290
319
|
private:
|
291
320
|
static const uint32_t kBucketSize = 1 << kBucketBits;
|
292
321
|
uint32_t buckets_[kBucketSize + kBucketSweep];
|
322
|
+
// True if buckets_ array needs to be initialized.
|
323
|
+
bool need_init_;
|
293
324
|
size_t num_dict_lookups_;
|
294
325
|
size_t num_dict_matches_;
|
295
326
|
};
|
296
327
|
|
297
|
-
// The maximum length for which the zopflification uses distinct distances.
|
298
|
-
static const int kMaxZopfliLen = 325;
|
299
|
-
|
300
328
|
// A (forgetful) hash table to the data seen by the compressor, to
|
301
329
|
// help create backward references to previous data.
|
302
330
|
//
|
@@ -308,16 +336,31 @@ template <int kBucketBits,
|
|
308
336
|
int kNumLastDistancesToCheck>
|
309
337
|
class HashLongestMatch {
|
310
338
|
public:
|
311
|
-
HashLongestMatch() {
|
339
|
+
HashLongestMatch(void) {
|
312
340
|
Reset();
|
313
341
|
}
|
314
342
|
|
315
|
-
void Reset() {
|
316
|
-
|
343
|
+
void Reset(void) {
|
344
|
+
need_init_ = true;
|
317
345
|
num_dict_lookups_ = 0;
|
318
346
|
num_dict_matches_ = 0;
|
319
347
|
}
|
320
348
|
|
349
|
+
void Init(void) {
|
350
|
+
if (need_init_) {
|
351
|
+
memset(&num_[0], 0, sizeof(num_));
|
352
|
+
need_init_ = false;
|
353
|
+
}
|
354
|
+
}
|
355
|
+
|
356
|
+
void InitForData(const uint8_t* data, size_t num) {
|
357
|
+
for (size_t i = 0; i < num; ++i) {
|
358
|
+
const uint32_t key = HashBytes(&data[i]);
|
359
|
+
num_[key] = 0;
|
360
|
+
need_init_ = false;
|
361
|
+
}
|
362
|
+
}
|
363
|
+
|
321
364
|
// Look at 3 bytes at data.
|
322
365
|
// Compute a hash from these, and store the value of ix at that position.
|
323
366
|
inline void Store(const uint8_t *data, const uint32_t ix) {
|
@@ -328,7 +371,7 @@ class HashLongestMatch {
|
|
328
371
|
}
|
329
372
|
|
330
373
|
// Find a longest backward match of &data[cur_ix] up to the length of
|
331
|
-
// max_length.
|
374
|
+
// max_length and stores the position cur_ix in the hash table.
|
332
375
|
//
|
333
376
|
// Does not look for matches longer than max_length.
|
334
377
|
// Does not look for matches further away than max_backward.
|
@@ -339,41 +382,42 @@ class HashLongestMatch {
|
|
339
382
|
bool FindLongestMatch(const uint8_t * __restrict data,
|
340
383
|
const size_t ring_buffer_mask,
|
341
384
|
const int* __restrict distance_cache,
|
342
|
-
const
|
343
|
-
const
|
344
|
-
const
|
345
|
-
|
346
|
-
|
347
|
-
|
385
|
+
const size_t cur_ix,
|
386
|
+
const size_t max_length,
|
387
|
+
const size_t max_backward,
|
388
|
+
size_t * __restrict best_len_out,
|
389
|
+
size_t * __restrict best_len_code_out,
|
390
|
+
size_t * __restrict best_distance_out,
|
348
391
|
double * __restrict best_score_out) {
|
349
392
|
*best_len_code_out = 0;
|
350
393
|
const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
|
351
394
|
bool match_found = false;
|
352
395
|
// Don't accept a short copy from far away.
|
353
396
|
double best_score = *best_score_out;
|
354
|
-
|
397
|
+
size_t best_len = *best_len_out;
|
355
398
|
*best_len_out = 0;
|
356
399
|
// Try last distance first.
|
357
|
-
for (
|
358
|
-
const
|
359
|
-
const
|
360
|
-
|
400
|
+
for (size_t i = 0; i < kNumLastDistancesToCheck; ++i) {
|
401
|
+
const size_t idx = kDistanceCacheIndex[i];
|
402
|
+
const size_t backward =
|
403
|
+
static_cast<size_t>(distance_cache[idx] + kDistanceCacheOffset[i]);
|
404
|
+
size_t prev_ix = static_cast<size_t>(cur_ix - backward);
|
361
405
|
if (prev_ix >= cur_ix) {
|
362
406
|
continue;
|
363
407
|
}
|
364
|
-
if (PREDICT_FALSE(backward >
|
408
|
+
if (PREDICT_FALSE(backward > max_backward)) {
|
365
409
|
continue;
|
366
410
|
}
|
367
|
-
prev_ix &=
|
411
|
+
prev_ix &= ring_buffer_mask;
|
368
412
|
|
369
413
|
if (cur_ix_masked + best_len > ring_buffer_mask ||
|
370
414
|
prev_ix + best_len > ring_buffer_mask ||
|
371
415
|
data[cur_ix_masked + best_len] != data[prev_ix + best_len]) {
|
372
416
|
continue;
|
373
417
|
}
|
374
|
-
const
|
375
|
-
|
376
|
-
|
418
|
+
const size_t len = FindMatchLengthWithLimit(&data[prev_ix],
|
419
|
+
&data[cur_ix_masked],
|
420
|
+
max_length);
|
377
421
|
if (len >= 3 || (len == 2 && i < 2)) {
|
378
422
|
// Comparing for >= 2 does not change the semantics, but just saves for
|
379
423
|
// a few unnecessary binary logarithms in backward reference score,
|
@@ -392,22 +436,23 @@ class HashLongestMatch {
|
|
392
436
|
}
|
393
437
|
const uint32_t key = HashBytes(&data[cur_ix_masked]);
|
394
438
|
const uint32_t * __restrict const bucket = &buckets_[key][0];
|
395
|
-
const
|
396
|
-
for (
|
397
|
-
|
398
|
-
|
439
|
+
const size_t down = (num_[key] > kBlockSize) ? (num_[key] - kBlockSize) : 0;
|
440
|
+
for (size_t i = num_[key]; i > down;) {
|
441
|
+
--i;
|
442
|
+
size_t prev_ix = bucket[i & kBlockMask];
|
443
|
+
const size_t backward = cur_ix - prev_ix;
|
399
444
|
if (PREDICT_FALSE(backward == 0 || backward > max_backward)) {
|
400
445
|
break;
|
401
446
|
}
|
402
|
-
prev_ix &=
|
447
|
+
prev_ix &= ring_buffer_mask;
|
403
448
|
if (cur_ix_masked + best_len > ring_buffer_mask ||
|
404
449
|
prev_ix + best_len > ring_buffer_mask ||
|
405
450
|
data[cur_ix_masked + best_len] != data[prev_ix + best_len]) {
|
406
451
|
continue;
|
407
452
|
}
|
408
|
-
const
|
409
|
-
|
410
|
-
|
453
|
+
const size_t len = FindMatchLengthWithLimit(&data[prev_ix],
|
454
|
+
&data[cur_ix_masked],
|
455
|
+
max_length);
|
411
456
|
if (len >= 4) {
|
412
457
|
// Comparing for >= 3 does not change the semantics, but just saves
|
413
458
|
// for a few unnecessary binary logarithms in backward reference
|
@@ -424,25 +469,28 @@ class HashLongestMatch {
|
|
424
469
|
}
|
425
470
|
}
|
426
471
|
}
|
472
|
+
buckets_[key][num_[key] & kBlockMask] = static_cast<uint32_t>(cur_ix);
|
473
|
+
++num_[key];
|
427
474
|
if (!match_found && num_dict_matches_ >= (num_dict_lookups_ >> 7)) {
|
428
|
-
|
475
|
+
size_t dict_key = Hash<14>(&data[cur_ix_masked]) << 1;
|
429
476
|
for (int k = 0; k < 2; ++k, ++dict_key) {
|
430
477
|
++num_dict_lookups_;
|
431
478
|
const uint16_t v = kStaticDictionaryHash[dict_key];
|
432
479
|
if (v > 0) {
|
433
|
-
const
|
434
|
-
const
|
435
|
-
const
|
480
|
+
const size_t len = v & 31;
|
481
|
+
const size_t dist = v >> 5;
|
482
|
+
const size_t offset =
|
483
|
+
kBrotliDictionaryOffsetsByLength[len] + len * dist;
|
436
484
|
if (len <= max_length) {
|
437
|
-
const
|
485
|
+
const size_t matchlen =
|
438
486
|
FindMatchLengthWithLimit(&data[cur_ix_masked],
|
439
487
|
&kBrotliDictionary[offset], len);
|
440
|
-
if (matchlen > len
|
441
|
-
const
|
442
|
-
const
|
488
|
+
if (matchlen + kCutoffTransformsCount > len && matchlen > 0) {
|
489
|
+
const size_t transform_id = kCutoffTransforms[len - matchlen];
|
490
|
+
const size_t word_id =
|
443
491
|
transform_id * (1 << kBrotliDictionarySizeBitsByLength[len]) +
|
444
492
|
dist;
|
445
|
-
const
|
493
|
+
const size_t backward = max_backward + word_id + 1;
|
446
494
|
double score = BackwardReferenceScore(matchlen, backward);
|
447
495
|
if (best_score < score) {
|
448
496
|
++num_dict_matches_;
|
@@ -462,28 +510,25 @@ class HashLongestMatch {
|
|
462
510
|
return match_found;
|
463
511
|
}
|
464
512
|
|
465
|
-
//
|
513
|
+
// Finds all backward matches of &data[cur_ix & ring_buffer_mask] up to the
|
514
|
+
// length of max_length and stores the position cur_ix in the hash table.
|
466
515
|
//
|
467
516
|
// Sets *num_matches to the number of matches found, and stores the found
|
468
|
-
// matches in matches[0] to matches[*num_matches - 1].
|
469
|
-
//
|
470
|
-
//
|
471
|
-
|
472
|
-
|
473
|
-
|
474
|
-
|
475
|
-
|
476
|
-
|
477
|
-
const int max_length,
|
478
|
-
const uint32_t max_backward,
|
479
|
-
int* num_matches,
|
480
|
-
BackwardMatch* matches) const {
|
517
|
+
// matches in matches[0] to matches[*num_matches - 1]. The matches will be
|
518
|
+
// sorted by strictly increasing length and (non-strictly) increasing
|
519
|
+
// distance.
|
520
|
+
size_t FindAllMatches(const uint8_t* data,
|
521
|
+
const size_t ring_buffer_mask,
|
522
|
+
const size_t cur_ix,
|
523
|
+
const size_t max_length,
|
524
|
+
const size_t max_backward,
|
525
|
+
BackwardMatch* matches) {
|
481
526
|
BackwardMatch* const orig_matches = matches;
|
482
527
|
const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
|
483
|
-
|
484
|
-
|
485
|
-
if (
|
486
|
-
for (
|
528
|
+
size_t best_len = 1;
|
529
|
+
size_t stop = cur_ix - 64;
|
530
|
+
if (cur_ix < 64) { stop = 0; }
|
531
|
+
for (size_t i = cur_ix - 1; i > stop && best_len <= 2; --i) {
|
487
532
|
size_t prev_ix = i;
|
488
533
|
const size_t backward = cur_ix - prev_ix;
|
489
534
|
if (PREDICT_FALSE(backward > max_backward)) {
|
@@ -494,57 +539,57 @@ class HashLongestMatch {
|
|
494
539
|
data[cur_ix_masked + 1] != data[prev_ix + 1]) {
|
495
540
|
continue;
|
496
541
|
}
|
497
|
-
const
|
542
|
+
const size_t len =
|
498
543
|
FindMatchLengthWithLimit(&data[prev_ix], &data[cur_ix_masked],
|
499
544
|
max_length);
|
500
545
|
if (len > best_len) {
|
501
546
|
best_len = len;
|
502
|
-
|
503
|
-
matches = orig_matches;
|
504
|
-
}
|
505
|
-
*matches++ = BackwardMatch(static_cast<int>(backward), len);
|
547
|
+
*matches++ = BackwardMatch(backward, len);
|
506
548
|
}
|
507
549
|
}
|
508
550
|
const uint32_t key = HashBytes(&data[cur_ix_masked]);
|
509
551
|
const uint32_t * __restrict const bucket = &buckets_[key][0];
|
510
|
-
const
|
511
|
-
for (
|
512
|
-
|
513
|
-
|
552
|
+
const size_t down = (num_[key] > kBlockSize) ? (num_[key] - kBlockSize) : 0;
|
553
|
+
for (size_t i = num_[key]; i > down;) {
|
554
|
+
--i;
|
555
|
+
size_t prev_ix = bucket[i & kBlockMask];
|
556
|
+
const size_t backward = cur_ix - prev_ix;
|
514
557
|
if (PREDICT_FALSE(backward == 0 || backward > max_backward)) {
|
515
558
|
break;
|
516
559
|
}
|
517
|
-
prev_ix &=
|
560
|
+
prev_ix &= ring_buffer_mask;
|
518
561
|
if (cur_ix_masked + best_len > ring_buffer_mask ||
|
519
562
|
prev_ix + best_len > ring_buffer_mask ||
|
520
563
|
data[cur_ix_masked + best_len] != data[prev_ix + best_len]) {
|
521
564
|
continue;
|
522
565
|
}
|
523
|
-
const
|
566
|
+
const size_t len =
|
524
567
|
FindMatchLengthWithLimit(&data[prev_ix], &data[cur_ix_masked],
|
525
568
|
max_length);
|
526
569
|
if (len > best_len) {
|
527
570
|
best_len = len;
|
528
|
-
if (len > kMaxZopfliLen) {
|
529
|
-
matches = orig_matches;
|
530
|
-
}
|
531
571
|
*matches++ = BackwardMatch(backward, len);
|
532
572
|
}
|
533
573
|
}
|
534
|
-
|
535
|
-
|
574
|
+
buckets_[key][num_[key] & kBlockMask] = static_cast<uint32_t>(cur_ix);
|
575
|
+
++num_[key];
|
576
|
+
uint32_t dict_matches[kMaxDictionaryMatchLen + 1];
|
577
|
+
for (size_t i = 0; i <= kMaxDictionaryMatchLen; ++i) {
|
578
|
+
dict_matches[i] = kInvalidMatch;
|
579
|
+
}
|
580
|
+
size_t minlen = std::max<size_t>(4, best_len + 1);
|
536
581
|
if (FindAllStaticDictionaryMatches(&data[cur_ix_masked], minlen, max_length,
|
537
582
|
&dict_matches[0])) {
|
538
|
-
|
539
|
-
for (
|
540
|
-
|
583
|
+
size_t maxlen = std::min<size_t>(kMaxDictionaryMatchLen, max_length);
|
584
|
+
for (size_t l = minlen; l <= maxlen; ++l) {
|
585
|
+
uint32_t dict_id = dict_matches[l];
|
541
586
|
if (dict_id < kInvalidMatch) {
|
542
587
|
*matches++ = BackwardMatch(max_backward + (dict_id >> 5) + 1, l,
|
543
588
|
dict_id & 31);
|
544
589
|
}
|
545
590
|
}
|
546
591
|
}
|
547
|
-
|
592
|
+
return static_cast<size_t>(matches - orig_matches);
|
548
593
|
}
|
549
594
|
|
550
595
|
enum { kHashLength = 4 };
|
@@ -560,6 +605,10 @@ class HashLongestMatch {
|
|
560
605
|
return h >> (32 - kBucketBits);
|
561
606
|
}
|
562
607
|
|
608
|
+
enum { kHashMapSize = 2 << kBucketBits };
|
609
|
+
|
610
|
+
static const size_t kMaxNumMatches = 64 + (1 << kBlockBits);
|
611
|
+
|
563
612
|
private:
|
564
613
|
// Number of hash buckets.
|
565
614
|
static const uint32_t kBucketSize = 1 << kBucketBits;
|
@@ -577,29 +626,281 @@ class HashLongestMatch {
|
|
577
626
|
// Buckets containing kBlockSize of backward references.
|
578
627
|
uint32_t buckets_[kBucketSize][kBlockSize];
|
579
628
|
|
629
|
+
// True if num_ array needs to be initialized.
|
630
|
+
bool need_init_;
|
631
|
+
|
580
632
|
size_t num_dict_lookups_;
|
581
633
|
size_t num_dict_matches_;
|
582
634
|
};
|
583
635
|
|
636
|
+
// A (forgetful) hash table where each hash bucket contains a binary tree of
|
637
|
+
// sequences whose first 4 bytes share the same hash code.
|
638
|
+
// Each sequence is kMaxTreeCompLength long and is identified by its starting
|
639
|
+
// position in the input data. The binary tree is sorted by the lexicographic
|
640
|
+
// order of the sequences, and it is also a max-heap with respect to the
|
641
|
+
// starting positions.
|
642
|
+
class HashToBinaryTree {
|
643
|
+
public:
|
644
|
+
HashToBinaryTree() : forest_(NULL) {
|
645
|
+
Reset();
|
646
|
+
}
|
647
|
+
|
648
|
+
~HashToBinaryTree() {
|
649
|
+
delete[] forest_;
|
650
|
+
}
|
651
|
+
|
652
|
+
void Reset() {
|
653
|
+
need_init_ = true;
|
654
|
+
}
|
655
|
+
|
656
|
+
void Init(int lgwin, size_t position, size_t bytes, bool is_last) {
|
657
|
+
if (need_init_) {
|
658
|
+
window_mask_ = (1u << lgwin) - 1u;
|
659
|
+
invalid_pos_ = static_cast<uint32_t>(-window_mask_);
|
660
|
+
for (uint32_t i = 0; i < kBucketSize; i++) {
|
661
|
+
buckets_[i] = invalid_pos_;
|
662
|
+
}
|
663
|
+
size_t num_nodes = (position == 0 && is_last) ? bytes : window_mask_ + 1;
|
664
|
+
forest_ = new uint32_t[2 * num_nodes];
|
665
|
+
need_init_ = false;
|
666
|
+
}
|
667
|
+
}
|
668
|
+
|
669
|
+
// Finds all backward matches of &data[cur_ix & ring_buffer_mask] up to the
|
670
|
+
// length of max_length and stores the position cur_ix in the hash table.
|
671
|
+
//
|
672
|
+
// Sets *num_matches to the number of matches found, and stores the found
|
673
|
+
// matches in matches[0] to matches[*num_matches - 1]. The matches will be
|
674
|
+
// sorted by strictly increasing length and (non-strictly) increasing
|
675
|
+
// distance.
|
676
|
+
size_t FindAllMatches(const uint8_t* data,
|
677
|
+
const size_t ring_buffer_mask,
|
678
|
+
const size_t cur_ix,
|
679
|
+
const size_t max_length,
|
680
|
+
const size_t max_backward,
|
681
|
+
BackwardMatch* matches) {
|
682
|
+
BackwardMatch* const orig_matches = matches;
|
683
|
+
const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
|
684
|
+
size_t best_len = 1;
|
685
|
+
size_t stop = cur_ix - 64;
|
686
|
+
if (cur_ix < 64) { stop = 0; }
|
687
|
+
for (size_t i = cur_ix - 1; i > stop && best_len <= 2; --i) {
|
688
|
+
size_t prev_ix = i;
|
689
|
+
const size_t backward = cur_ix - prev_ix;
|
690
|
+
if (PREDICT_FALSE(backward > max_backward)) {
|
691
|
+
break;
|
692
|
+
}
|
693
|
+
prev_ix &= ring_buffer_mask;
|
694
|
+
if (data[cur_ix_masked] != data[prev_ix] ||
|
695
|
+
data[cur_ix_masked + 1] != data[prev_ix + 1]) {
|
696
|
+
continue;
|
697
|
+
}
|
698
|
+
const size_t len =
|
699
|
+
FindMatchLengthWithLimit(&data[prev_ix], &data[cur_ix_masked],
|
700
|
+
max_length);
|
701
|
+
if (len > best_len) {
|
702
|
+
best_len = len;
|
703
|
+
*matches++ = BackwardMatch(backward, len);
|
704
|
+
}
|
705
|
+
}
|
706
|
+
if (best_len < max_length) {
|
707
|
+
matches = StoreAndFindMatches(data, cur_ix, ring_buffer_mask,
|
708
|
+
max_length, &best_len, matches);
|
709
|
+
}
|
710
|
+
uint32_t dict_matches[kMaxDictionaryMatchLen + 1];
|
711
|
+
for (size_t i = 0; i <= kMaxDictionaryMatchLen; ++i) {
|
712
|
+
dict_matches[i] = kInvalidMatch;
|
713
|
+
}
|
714
|
+
size_t minlen = std::max<size_t>(4, best_len + 1);
|
715
|
+
if (FindAllStaticDictionaryMatches(&data[cur_ix_masked], minlen, max_length,
|
716
|
+
&dict_matches[0])) {
|
717
|
+
size_t maxlen = std::min<size_t>(kMaxDictionaryMatchLen, max_length);
|
718
|
+
for (size_t l = minlen; l <= maxlen; ++l) {
|
719
|
+
uint32_t dict_id = dict_matches[l];
|
720
|
+
if (dict_id < kInvalidMatch) {
|
721
|
+
*matches++ = BackwardMatch(max_backward + (dict_id >> 5) + 1, l,
|
722
|
+
dict_id & 31);
|
723
|
+
}
|
724
|
+
}
|
725
|
+
}
|
726
|
+
return static_cast<size_t>(matches - orig_matches);
|
727
|
+
}
|
728
|
+
|
729
|
+
// Stores the hash of the next 4 bytes and re-roots the binary tree at the
|
730
|
+
// current sequence, without returning any matches.
|
731
|
+
// REQUIRES: cur_ix + kMaxTreeCompLength <= end-of-current-block
|
732
|
+
void Store(const uint8_t* data,
|
733
|
+
const size_t ring_buffer_mask,
|
734
|
+
const size_t cur_ix) {
|
735
|
+
size_t best_len = 0;
|
736
|
+
StoreAndFindMatches(data, cur_ix, ring_buffer_mask, kMaxTreeCompLength,
|
737
|
+
&best_len, NULL);
|
738
|
+
}
|
739
|
+
|
740
|
+
void StitchToPreviousBlock(size_t num_bytes,
|
741
|
+
size_t position,
|
742
|
+
const uint8_t* ringbuffer,
|
743
|
+
size_t ringbuffer_mask) {
|
744
|
+
if (num_bytes >= 3 && position >= kMaxTreeCompLength) {
|
745
|
+
// Store the last `kMaxTreeCompLength - 1` positions in the hasher.
|
746
|
+
// These could not be calculated before, since they require knowledge
|
747
|
+
// of both the previous and the current block.
|
748
|
+
const size_t i_start = position - kMaxTreeCompLength + 1;
|
749
|
+
const size_t i_end = std::min(position, i_start + num_bytes);
|
750
|
+
for (size_t i = i_start; i < i_end; ++i) {
|
751
|
+
// We know that i + kMaxTreeCompLength <= position + num_bytes, i.e. the
|
752
|
+
// end of the current block and that we have at least
|
753
|
+
// kMaxTreeCompLength tail in the ringbuffer.
|
754
|
+
Store(ringbuffer, ringbuffer_mask, i);
|
755
|
+
}
|
756
|
+
}
|
757
|
+
}
|
758
|
+
|
759
|
+
static const size_t kMaxNumMatches = 64 + kMaxTreeSearchDepth;
|
760
|
+
|
761
|
+
private:
|
762
|
+
// Stores the hash of the next 4 bytes and in a single tree-traversal, the
|
763
|
+
// hash bucket's binary tree is searched for matches and is re-rooted at the
|
764
|
+
// current position.
|
765
|
+
//
|
766
|
+
// If less than kMaxTreeCompLength data is available, the hash bucket of the
|
767
|
+
// current position is searched for matches, but the state of the hash table
|
768
|
+
// is not changed, since we can not know the final sorting order of the
|
769
|
+
// current (incomplete) sequence.
|
770
|
+
//
|
771
|
+
// This function must be called with increasing cur_ix positions.
|
772
|
+
BackwardMatch* StoreAndFindMatches(const uint8_t* const __restrict data,
|
773
|
+
const size_t cur_ix,
|
774
|
+
const size_t ring_buffer_mask,
|
775
|
+
const size_t max_length,
|
776
|
+
size_t* const __restrict best_len,
|
777
|
+
BackwardMatch* __restrict matches) {
|
778
|
+
const size_t cur_ix_masked = cur_ix & ring_buffer_mask;
|
779
|
+
const size_t max_backward = window_mask_ - 15;
|
780
|
+
const size_t max_comp_len = std::min(max_length, kMaxTreeCompLength);
|
781
|
+
const bool reroot_tree = max_length >= kMaxTreeCompLength;
|
782
|
+
const uint32_t key = HashBytes(&data[cur_ix_masked]);
|
783
|
+
size_t prev_ix = buckets_[key];
|
784
|
+
// The forest index of the rightmost node of the left subtree of the new
|
785
|
+
// root, updated as we traverse and reroot the tree of the hash bucket.
|
786
|
+
size_t node_left = LeftChildIndex(cur_ix);
|
787
|
+
// The forest index of the leftmost node of the right subtree of the new
|
788
|
+
// root, updated as we traverse and reroot the tree of the hash bucket.
|
789
|
+
size_t node_right = RightChildIndex(cur_ix);
|
790
|
+
// The match length of the rightmost node of the left subtree of the new
|
791
|
+
// root, updated as we traverse and reroot the tree of the hash bucket.
|
792
|
+
size_t best_len_left = 0;
|
793
|
+
// The match length of the leftmost node of the right subtree of the new
|
794
|
+
// root, updated as we traverse and reroot the tree of the hash bucket.
|
795
|
+
size_t best_len_right = 0;
|
796
|
+
if (reroot_tree) {
|
797
|
+
buckets_[key] = static_cast<uint32_t>(cur_ix);
|
798
|
+
}
|
799
|
+
for (size_t depth_remaining = kMaxTreeSearchDepth; ; --depth_remaining) {
|
800
|
+
const size_t backward = cur_ix - prev_ix;
|
801
|
+
const size_t prev_ix_masked = prev_ix & ring_buffer_mask;
|
802
|
+
if (backward == 0 || backward > max_backward || depth_remaining == 0) {
|
803
|
+
if (reroot_tree) {
|
804
|
+
forest_[node_left] = invalid_pos_;
|
805
|
+
forest_[node_right] = invalid_pos_;
|
806
|
+
}
|
807
|
+
break;
|
808
|
+
}
|
809
|
+
const size_t cur_len = std::min(best_len_left, best_len_right);
|
810
|
+
const size_t len = cur_len +
|
811
|
+
FindMatchLengthWithLimit(&data[cur_ix_masked + cur_len],
|
812
|
+
&data[prev_ix_masked + cur_len],
|
813
|
+
max_length - cur_len);
|
814
|
+
if (len > *best_len) {
|
815
|
+
*best_len = len;
|
816
|
+
if (matches) {
|
817
|
+
*matches++ = BackwardMatch(backward, len);
|
818
|
+
}
|
819
|
+
if (len >= max_comp_len) {
|
820
|
+
if (reroot_tree) {
|
821
|
+
forest_[node_left] = forest_[LeftChildIndex(prev_ix)];
|
822
|
+
forest_[node_right] = forest_[RightChildIndex(prev_ix)];
|
823
|
+
}
|
824
|
+
break;
|
825
|
+
}
|
826
|
+
}
|
827
|
+
if (data[cur_ix_masked + len] > data[prev_ix_masked + len]) {
|
828
|
+
best_len_left = len;
|
829
|
+
if (reroot_tree) {
|
830
|
+
forest_[node_left] = static_cast<uint32_t>(prev_ix);
|
831
|
+
}
|
832
|
+
node_left = RightChildIndex(prev_ix);
|
833
|
+
prev_ix = forest_[node_left];
|
834
|
+
} else {
|
835
|
+
best_len_right = len;
|
836
|
+
if (reroot_tree) {
|
837
|
+
forest_[node_right] = static_cast<uint32_t>(prev_ix);
|
838
|
+
}
|
839
|
+
node_right = LeftChildIndex(prev_ix);
|
840
|
+
prev_ix = forest_[node_right];
|
841
|
+
}
|
842
|
+
}
|
843
|
+
return matches;
|
844
|
+
}
|
845
|
+
|
846
|
+
inline size_t LeftChildIndex(const size_t pos) {
|
847
|
+
return 2 * (pos & window_mask_);
|
848
|
+
}
|
849
|
+
|
850
|
+
inline size_t RightChildIndex(const size_t pos) {
|
851
|
+
return 2 * (pos & window_mask_) + 1;
|
852
|
+
}
|
853
|
+
|
854
|
+
static uint32_t HashBytes(const uint8_t *data) {
|
855
|
+
uint32_t h = BROTLI_UNALIGNED_LOAD32(data) * kHashMul32;
|
856
|
+
// The higher bits contain more mixture from the multiplication,
|
857
|
+
// so we take our results from there.
|
858
|
+
return h >> (32 - kBucketBits);
|
859
|
+
}
|
860
|
+
|
861
|
+
static const int kBucketBits = 17;
|
862
|
+
static const size_t kBucketSize = 1 << kBucketBits;
|
863
|
+
|
864
|
+
// The window size minus 1
|
865
|
+
size_t window_mask_;
|
866
|
+
|
867
|
+
// Hash table that maps the 4-byte hashes of the sequence to the last
|
868
|
+
// position where this hash was found, which is the root of the binary
|
869
|
+
// tree of sequences that share this hash bucket.
|
870
|
+
uint32_t buckets_[kBucketSize];
|
871
|
+
|
872
|
+
// The union of the binary trees of each hash bucket. The root of the tree
|
873
|
+
// corresponding to a hash is a sequence starting at buckets_[hash] and
|
874
|
+
// the left and right children of a sequence starting at pos are
|
875
|
+
// forest_[2 * pos] and forest_[2 * pos + 1].
|
876
|
+
uint32_t* forest_;
|
877
|
+
|
878
|
+
// A position used to mark a non-existent sequence, i.e. a tree is empty if
|
879
|
+
// its root is at invalid_pos_ and a node is a leaf if both its children
|
880
|
+
// are at invalid_pos_.
|
881
|
+
uint32_t invalid_pos_;
|
882
|
+
|
883
|
+
bool need_init_;
|
884
|
+
};
|
885
|
+
|
584
886
|
struct Hashers {
|
585
887
|
// For kBucketSweep == 1, enabling the dictionary lookup makes compression
|
586
888
|
// a little faster (0.5% - 1%) and it compresses 0.15% better on small text
|
587
889
|
// and html inputs.
|
588
|
-
typedef HashLongestMatchQuickly<16, 1, true>
|
589
|
-
typedef HashLongestMatchQuickly<16, 2, false>
|
590
|
-
typedef HashLongestMatchQuickly<16, 4, false> H3;
|
890
|
+
typedef HashLongestMatchQuickly<16, 1, true> H2;
|
891
|
+
typedef HashLongestMatchQuickly<16, 2, false> H3;
|
591
892
|
typedef HashLongestMatchQuickly<17, 4, true> H4;
|
592
893
|
typedef HashLongestMatch<14, 4, 4> H5;
|
593
894
|
typedef HashLongestMatch<14, 5, 4> H6;
|
594
895
|
typedef HashLongestMatch<15, 6, 10> H7;
|
595
896
|
typedef HashLongestMatch<15, 7, 10> H8;
|
596
897
|
typedef HashLongestMatch<15, 8, 16> H9;
|
898
|
+
typedef HashToBinaryTree H10;
|
597
899
|
|
598
|
-
Hashers() :
|
599
|
-
|
900
|
+
Hashers(void) : hash_h2(0), hash_h3(0), hash_h4(0), hash_h5(0),
|
901
|
+
hash_h6(0), hash_h7(0), hash_h8(0), hash_h9(0), hash_h10(0) {}
|
600
902
|
|
601
|
-
~Hashers() {
|
602
|
-
delete hash_h1;
|
903
|
+
~Hashers(void) {
|
603
904
|
delete hash_h2;
|
604
905
|
delete hash_h3;
|
605
906
|
delete hash_h4;
|
@@ -608,11 +909,11 @@ struct Hashers {
|
|
608
909
|
delete hash_h7;
|
609
910
|
delete hash_h8;
|
610
911
|
delete hash_h9;
|
912
|
+
delete hash_h10;
|
611
913
|
}
|
612
914
|
|
613
915
|
void Init(int type) {
|
614
916
|
switch (type) {
|
615
|
-
case 1: hash_h1 = new H1; break;
|
616
917
|
case 2: hash_h2 = new H2; break;
|
617
918
|
case 3: hash_h3 = new H3; break;
|
618
919
|
case 4: hash_h4 = new H4; break;
|
@@ -621,12 +922,14 @@ struct Hashers {
|
|
621
922
|
case 7: hash_h7 = new H7; break;
|
622
923
|
case 8: hash_h8 = new H8; break;
|
623
924
|
case 9: hash_h9 = new H9; break;
|
925
|
+
case 10: hash_h10 = new H10; break;
|
624
926
|
default: break;
|
625
927
|
}
|
626
928
|
}
|
627
929
|
|
628
930
|
template<typename Hasher>
|
629
931
|
void WarmupHash(const size_t size, const uint8_t* dict, Hasher* hasher) {
|
932
|
+
hasher->Init();
|
630
933
|
for (size_t i = 0; i + Hasher::kHashTypeLength - 1 < size; i++) {
|
631
934
|
hasher->Store(&dict[i], static_cast<uint32_t>(i));
|
632
935
|
}
|
@@ -634,9 +937,8 @@ struct Hashers {
|
|
634
937
|
|
635
938
|
// Custom LZ77 window.
|
636
939
|
void PrependCustomDictionary(
|
637
|
-
int type, const size_t size, const uint8_t* dict) {
|
940
|
+
int type, int lgwin, const size_t size, const uint8_t* dict) {
|
638
941
|
switch (type) {
|
639
|
-
case 1: WarmupHash(size, dict, hash_h1); break;
|
640
942
|
case 2: WarmupHash(size, dict, hash_h2); break;
|
641
943
|
case 3: WarmupHash(size, dict, hash_h3); break;
|
642
944
|
case 4: WarmupHash(size, dict, hash_h4); break;
|
@@ -645,12 +947,17 @@ struct Hashers {
|
|
645
947
|
case 7: WarmupHash(size, dict, hash_h7); break;
|
646
948
|
case 8: WarmupHash(size, dict, hash_h8); break;
|
647
949
|
case 9: WarmupHash(size, dict, hash_h9); break;
|
950
|
+
case 10:
|
951
|
+
hash_h10->Init(lgwin, 0, size, false);
|
952
|
+
for (size_t i = 0; i + kMaxTreeCompLength - 1 < size; ++i) {
|
953
|
+
hash_h10->Store(dict, std::numeric_limits<size_t>::max(), i);
|
954
|
+
}
|
955
|
+
break;
|
648
956
|
default: break;
|
649
957
|
}
|
650
958
|
}
|
651
959
|
|
652
960
|
|
653
|
-
H1* hash_h1;
|
654
961
|
H2* hash_h2;
|
655
962
|
H3* hash_h3;
|
656
963
|
H4* hash_h4;
|
@@ -659,6 +966,7 @@ struct Hashers {
|
|
659
966
|
H7* hash_h7;
|
660
967
|
H8* hash_h8;
|
661
968
|
H9* hash_h9;
|
969
|
+
H10* hash_h10;
|
662
970
|
};
|
663
971
|
|
664
972
|
} // namespace brotli
|