/src/h2o/deps/brotli/c/enc/compress_fragment.c
Line | Count | Source (jump to first uncovered line) |
1 | | /* Copyright 2015 Google Inc. All Rights Reserved. |
2 | | |
3 | | Distributed under MIT license. |
4 | | See file LICENSE for detail or copy at https://opensource.org/licenses/MIT |
5 | | */ |
6 | | |
7 | | /* Function for fast encoding of an input fragment, independently from the input |
8 | | history. This function uses one-pass processing: when we find a backward |
9 | | match, we immediately emit the corresponding command and literal codes to |
10 | | the bit stream. |
11 | | |
12 | | Adapted from the CompressFragment() function in |
13 | | https://github.com/google/snappy/blob/master/snappy.cc */ |
14 | | |
15 | | #include "./compress_fragment.h" |
16 | | |
17 | | #include <string.h> /* memcmp, memcpy, memset */ |
18 | | |
19 | | #include "../common/constants.h" |
20 | | #include <brotli/types.h> |
21 | | #include "./brotli_bit_stream.h" |
22 | | #include "./entropy_encode.h" |
23 | | #include "./fast_log.h" |
24 | | #include "./find_match_length.h" |
25 | | #include "./memory.h" |
26 | | #include "./port.h" |
27 | | #include "./write_bits.h" |
28 | | |
29 | | |
30 | | #if defined(__cplusplus) || defined(c_plusplus) |
31 | | extern "C" { |
32 | | #endif |
33 | | |
34 | 0 | #define MAX_DISTANCE (long)BROTLI_MAX_BACKWARD_LIMIT(18) |
35 | | |
36 | | /* kHashMul32 multiplier has these properties: |
37 | | * The multiplier must be odd. Otherwise we may lose the highest bit. |
38 | | * No long streaks of ones or zeros. |
39 | | * There is no effort to ensure that it is a prime, the oddity is enough |
40 | | for this use. |
41 | | * The number has been tuned heuristically against compression benchmarks. */ |
42 | | static const uint32_t kHashMul32 = 0x1e35a7bd; |
43 | | |
44 | 0 | static BROTLI_INLINE uint32_t Hash(const uint8_t* p, size_t shift) { |
45 | 0 | const uint64_t h = (BROTLI_UNALIGNED_LOAD64LE(p) << 24) * kHashMul32; |
46 | 0 | return (uint32_t)(h >> shift); |
47 | 0 | } |
48 | | |
49 | | static BROTLI_INLINE uint32_t HashBytesAtOffset( |
50 | 0 | uint64_t v, int offset, size_t shift) { |
51 | 0 | assert(offset >= 0); |
52 | 0 | assert(offset <= 3); |
53 | 0 | { |
54 | 0 | const uint64_t h = ((v >> (8 * offset)) << 24) * kHashMul32; |
55 | 0 | return (uint32_t)(h >> shift); |
56 | 0 | } |
57 | 0 | } |
58 | | |
59 | 0 | static BROTLI_INLINE BROTLI_BOOL IsMatch(const uint8_t* p1, const uint8_t* p2) { |
60 | 0 | return TO_BROTLI_BOOL( |
61 | 0 | BROTLI_UNALIGNED_LOAD32(p1) == BROTLI_UNALIGNED_LOAD32(p2) && |
62 | 0 | p1[4] == p2[4]); |
63 | 0 | } |
64 | | |
65 | | /* Builds a literal prefix code into "depths" and "bits" based on the statistics |
66 | | of the "input" string and stores it into the bit stream. |
67 | | Note that the prefix code here is built from the pre-LZ77 input, therefore |
68 | | we can only approximate the statistics of the actual literal stream. |
69 | | Moreover, for long inputs we build a histogram from a sample of the input |
70 | | and thus have to assign a non-zero depth for each literal. |
71 | | Returns estimated compression ratio millibytes/char for encoding given input |
72 | | with generated code. */ |
73 | | static size_t BuildAndStoreLiteralPrefixCode(MemoryManager* m, |
74 | | const uint8_t* input, |
75 | | const size_t input_size, |
76 | | uint8_t depths[256], |
77 | | uint16_t bits[256], |
78 | | size_t* storage_ix, |
79 | 0 | uint8_t* storage) { |
80 | 0 | uint32_t histogram[256] = { 0 }; |
81 | 0 | size_t histogram_total; |
82 | 0 | size_t i; |
83 | 0 | if (input_size < (1 << 15)) { |
84 | 0 | for (i = 0; i < input_size; ++i) { |
85 | 0 | ++histogram[input[i]]; |
86 | 0 | } |
87 | 0 | histogram_total = input_size; |
88 | 0 | for (i = 0; i < 256; ++i) { |
89 | | /* We weigh the first 11 samples with weight 3 to account for the |
90 | | balancing effect of the LZ77 phase on the histogram. */ |
91 | 0 | const uint32_t adjust = 2 * BROTLI_MIN(uint32_t, histogram[i], 11u); |
92 | 0 | histogram[i] += adjust; |
93 | 0 | histogram_total += adjust; |
94 | 0 | } |
95 | 0 | } else { |
96 | 0 | static const size_t kSampleRate = 29; |
97 | 0 | for (i = 0; i < input_size; i += kSampleRate) { |
98 | 0 | ++histogram[input[i]]; |
99 | 0 | } |
100 | 0 | histogram_total = (input_size + kSampleRate - 1) / kSampleRate; |
101 | 0 | for (i = 0; i < 256; ++i) { |
102 | | /* We add 1 to each population count to avoid 0 bit depths (since this is |
103 | | only a sample and we don't know if the symbol appears or not), and we |
104 | | weigh the first 11 samples with weight 3 to account for the balancing |
105 | | effect of the LZ77 phase on the histogram (more frequent symbols are |
106 | | more likely to be in backward references instead as literals). */ |
107 | 0 | const uint32_t adjust = 1 + 2 * BROTLI_MIN(uint32_t, histogram[i], 11u); |
108 | 0 | histogram[i] += adjust; |
109 | 0 | histogram_total += adjust; |
110 | 0 | } |
111 | 0 | } |
112 | 0 | BrotliBuildAndStoreHuffmanTreeFast(m, histogram, histogram_total, |
113 | 0 | /* max_bits = */ 8, |
114 | 0 | depths, bits, storage_ix, storage); |
115 | 0 | if (BROTLI_IS_OOM(m)) return 0; |
116 | 0 | { |
117 | 0 | size_t literal_ratio = 0; |
118 | 0 | for (i = 0; i < 256; ++i) { |
119 | 0 | if (histogram[i]) literal_ratio += histogram[i] * depths[i]; |
120 | 0 | } |
121 | | /* Estimated encoding ratio, millibytes per symbol. */ |
122 | 0 | return (literal_ratio * 125) / histogram_total; |
123 | 0 | } |
124 | 0 | } |
125 | | |
126 | | /* Builds a command and distance prefix code (each 64 symbols) into "depth" and |
127 | | "bits" based on "histogram" and stores it into the bit stream. */ |
128 | | static void BuildAndStoreCommandPrefixCode(const uint32_t histogram[128], |
129 | | uint8_t depth[128], uint16_t bits[128], size_t* storage_ix, |
130 | 0 | uint8_t* storage) { |
131 | | /* Tree size for building a tree over 64 symbols is 2 * 64 + 1. */ |
132 | 0 | HuffmanTree tree[129]; |
133 | 0 | uint8_t cmd_depth[BROTLI_NUM_COMMAND_SYMBOLS] = { 0 }; |
134 | 0 | uint16_t cmd_bits[64]; |
135 | |
|
136 | 0 | BrotliCreateHuffmanTree(histogram, 64, 15, tree, depth); |
137 | 0 | BrotliCreateHuffmanTree(&histogram[64], 64, 14, tree, &depth[64]); |
138 | | /* We have to jump through a few hoops here in order to compute |
139 | | the command bits because the symbols are in a different order than in |
140 | | the full alphabet. This looks complicated, but having the symbols |
141 | | in this order in the command bits saves a few branches in the Emit* |
142 | | functions. */ |
143 | 0 | memcpy(cmd_depth, depth, 24); |
144 | 0 | memcpy(cmd_depth + 24, depth + 40, 8); |
145 | 0 | memcpy(cmd_depth + 32, depth + 24, 8); |
146 | 0 | memcpy(cmd_depth + 40, depth + 48, 8); |
147 | 0 | memcpy(cmd_depth + 48, depth + 32, 8); |
148 | 0 | memcpy(cmd_depth + 56, depth + 56, 8); |
149 | 0 | BrotliConvertBitDepthsToSymbols(cmd_depth, 64, cmd_bits); |
150 | 0 | memcpy(bits, cmd_bits, 48); |
151 | 0 | memcpy(bits + 24, cmd_bits + 32, 16); |
152 | 0 | memcpy(bits + 32, cmd_bits + 48, 16); |
153 | 0 | memcpy(bits + 40, cmd_bits + 24, 16); |
154 | 0 | memcpy(bits + 48, cmd_bits + 40, 16); |
155 | 0 | memcpy(bits + 56, cmd_bits + 56, 16); |
156 | 0 | BrotliConvertBitDepthsToSymbols(&depth[64], 64, &bits[64]); |
157 | 0 | { |
158 | | /* Create the bit length array for the full command alphabet. */ |
159 | 0 | size_t i; |
160 | 0 | memset(cmd_depth, 0, 64); /* only 64 first values were used */ |
161 | 0 | memcpy(cmd_depth, depth, 8); |
162 | 0 | memcpy(cmd_depth + 64, depth + 8, 8); |
163 | 0 | memcpy(cmd_depth + 128, depth + 16, 8); |
164 | 0 | memcpy(cmd_depth + 192, depth + 24, 8); |
165 | 0 | memcpy(cmd_depth + 384, depth + 32, 8); |
166 | 0 | for (i = 0; i < 8; ++i) { |
167 | 0 | cmd_depth[128 + 8 * i] = depth[40 + i]; |
168 | 0 | cmd_depth[256 + 8 * i] = depth[48 + i]; |
169 | 0 | cmd_depth[448 + 8 * i] = depth[56 + i]; |
170 | 0 | } |
171 | 0 | BrotliStoreHuffmanTree( |
172 | 0 | cmd_depth, BROTLI_NUM_COMMAND_SYMBOLS, tree, storage_ix, storage); |
173 | 0 | } |
174 | 0 | BrotliStoreHuffmanTree(&depth[64], 64, tree, storage_ix, storage); |
175 | 0 | } |
176 | | |
177 | | /* REQUIRES: insertlen < 6210 */ |
178 | | static BROTLI_INLINE void EmitInsertLen(size_t insertlen, |
179 | | const uint8_t depth[128], |
180 | | const uint16_t bits[128], |
181 | | uint32_t histo[128], |
182 | | size_t* storage_ix, |
183 | 0 | uint8_t* storage) { |
184 | 0 | if (insertlen < 6) { |
185 | 0 | const size_t code = insertlen + 40; |
186 | 0 | BrotliWriteBits(depth[code], bits[code], storage_ix, storage); |
187 | 0 | ++histo[code]; |
188 | 0 | } else if (insertlen < 130) { |
189 | 0 | const size_t tail = insertlen - 2; |
190 | 0 | const uint32_t nbits = Log2FloorNonZero(tail) - 1u; |
191 | 0 | const size_t prefix = tail >> nbits; |
192 | 0 | const size_t inscode = (nbits << 1) + prefix + 42; |
193 | 0 | BrotliWriteBits(depth[inscode], bits[inscode], storage_ix, storage); |
194 | 0 | BrotliWriteBits(nbits, tail - (prefix << nbits), storage_ix, storage); |
195 | 0 | ++histo[inscode]; |
196 | 0 | } else if (insertlen < 2114) { |
197 | 0 | const size_t tail = insertlen - 66; |
198 | 0 | const uint32_t nbits = Log2FloorNonZero(tail); |
199 | 0 | const size_t code = nbits + 50; |
200 | 0 | BrotliWriteBits(depth[code], bits[code], storage_ix, storage); |
201 | 0 | BrotliWriteBits(nbits, tail - ((size_t)1 << nbits), storage_ix, storage); |
202 | 0 | ++histo[code]; |
203 | 0 | } else { |
204 | 0 | BrotliWriteBits(depth[61], bits[61], storage_ix, storage); |
205 | 0 | BrotliWriteBits(12, insertlen - 2114, storage_ix, storage); |
206 | 0 | ++histo[21]; |
207 | 0 | } |
208 | 0 | } |
209 | | |
210 | | static BROTLI_INLINE void EmitLongInsertLen(size_t insertlen, |
211 | | const uint8_t depth[128], |
212 | | const uint16_t bits[128], |
213 | | uint32_t histo[128], |
214 | | size_t* storage_ix, |
215 | 0 | uint8_t* storage) { |
216 | 0 | if (insertlen < 22594) { |
217 | 0 | BrotliWriteBits(depth[62], bits[62], storage_ix, storage); |
218 | 0 | BrotliWriteBits(14, insertlen - 6210, storage_ix, storage); |
219 | 0 | ++histo[22]; |
220 | 0 | } else { |
221 | 0 | BrotliWriteBits(depth[63], bits[63], storage_ix, storage); |
222 | 0 | BrotliWriteBits(24, insertlen - 22594, storage_ix, storage); |
223 | 0 | ++histo[23]; |
224 | 0 | } |
225 | 0 | } |
226 | | |
227 | | static BROTLI_INLINE void EmitCopyLen(size_t copylen, |
228 | | const uint8_t depth[128], |
229 | | const uint16_t bits[128], |
230 | | uint32_t histo[128], |
231 | | size_t* storage_ix, |
232 | 0 | uint8_t* storage) { |
233 | 0 | if (copylen < 10) { |
234 | 0 | BrotliWriteBits( |
235 | 0 | depth[copylen + 14], bits[copylen + 14], storage_ix, storage); |
236 | 0 | ++histo[copylen + 14]; |
237 | 0 | } else if (copylen < 134) { |
238 | 0 | const size_t tail = copylen - 6; |
239 | 0 | const uint32_t nbits = Log2FloorNonZero(tail) - 1u; |
240 | 0 | const size_t prefix = tail >> nbits; |
241 | 0 | const size_t code = (nbits << 1) + prefix + 20; |
242 | 0 | BrotliWriteBits(depth[code], bits[code], storage_ix, storage); |
243 | 0 | BrotliWriteBits(nbits, tail - (prefix << nbits), storage_ix, storage); |
244 | 0 | ++histo[code]; |
245 | 0 | } else if (copylen < 2118) { |
246 | 0 | const size_t tail = copylen - 70; |
247 | 0 | const uint32_t nbits = Log2FloorNonZero(tail); |
248 | 0 | const size_t code = nbits + 28; |
249 | 0 | BrotliWriteBits(depth[code], bits[code], storage_ix, storage); |
250 | 0 | BrotliWriteBits(nbits, tail - ((size_t)1 << nbits), storage_ix, storage); |
251 | 0 | ++histo[code]; |
252 | 0 | } else { |
253 | 0 | BrotliWriteBits(depth[39], bits[39], storage_ix, storage); |
254 | 0 | BrotliWriteBits(24, copylen - 2118, storage_ix, storage); |
255 | 0 | ++histo[47]; |
256 | 0 | } |
257 | 0 | } |
258 | | |
259 | | static BROTLI_INLINE void EmitCopyLenLastDistance(size_t copylen, |
260 | | const uint8_t depth[128], |
261 | | const uint16_t bits[128], |
262 | | uint32_t histo[128], |
263 | | size_t* storage_ix, |
264 | 0 | uint8_t* storage) { |
265 | 0 | if (copylen < 12) { |
266 | 0 | BrotliWriteBits(depth[copylen - 4], bits[copylen - 4], storage_ix, storage); |
267 | 0 | ++histo[copylen - 4]; |
268 | 0 | } else if (copylen < 72) { |
269 | 0 | const size_t tail = copylen - 8; |
270 | 0 | const uint32_t nbits = Log2FloorNonZero(tail) - 1; |
271 | 0 | const size_t prefix = tail >> nbits; |
272 | 0 | const size_t code = (nbits << 1) + prefix + 4; |
273 | 0 | BrotliWriteBits(depth[code], bits[code], storage_ix, storage); |
274 | 0 | BrotliWriteBits(nbits, tail - (prefix << nbits), storage_ix, storage); |
275 | 0 | ++histo[code]; |
276 | 0 | } else if (copylen < 136) { |
277 | 0 | const size_t tail = copylen - 8; |
278 | 0 | const size_t code = (tail >> 5) + 30; |
279 | 0 | BrotliWriteBits(depth[code], bits[code], storage_ix, storage); |
280 | 0 | BrotliWriteBits(5, tail & 31, storage_ix, storage); |
281 | 0 | BrotliWriteBits(depth[64], bits[64], storage_ix, storage); |
282 | 0 | ++histo[code]; |
283 | 0 | ++histo[64]; |
284 | 0 | } else if (copylen < 2120) { |
285 | 0 | const size_t tail = copylen - 72; |
286 | 0 | const uint32_t nbits = Log2FloorNonZero(tail); |
287 | 0 | const size_t code = nbits + 28; |
288 | 0 | BrotliWriteBits(depth[code], bits[code], storage_ix, storage); |
289 | 0 | BrotliWriteBits(nbits, tail - ((size_t)1 << nbits), storage_ix, storage); |
290 | 0 | BrotliWriteBits(depth[64], bits[64], storage_ix, storage); |
291 | 0 | ++histo[code]; |
292 | 0 | ++histo[64]; |
293 | 0 | } else { |
294 | 0 | BrotliWriteBits(depth[39], bits[39], storage_ix, storage); |
295 | 0 | BrotliWriteBits(24, copylen - 2120, storage_ix, storage); |
296 | 0 | BrotliWriteBits(depth[64], bits[64], storage_ix, storage); |
297 | 0 | ++histo[47]; |
298 | 0 | ++histo[64]; |
299 | 0 | } |
300 | 0 | } |
301 | | |
302 | | static BROTLI_INLINE void EmitDistance(size_t distance, |
303 | | const uint8_t depth[128], |
304 | | const uint16_t bits[128], |
305 | | uint32_t histo[128], |
306 | 0 | size_t* storage_ix, uint8_t* storage) { |
307 | 0 | const size_t d = distance + 3; |
308 | 0 | const uint32_t nbits = Log2FloorNonZero(d) - 1u; |
309 | 0 | const size_t prefix = (d >> nbits) & 1; |
310 | 0 | const size_t offset = (2 + prefix) << nbits; |
311 | 0 | const size_t distcode = 2 * (nbits - 1) + prefix + 80; |
312 | 0 | BrotliWriteBits(depth[distcode], bits[distcode], storage_ix, storage); |
313 | 0 | BrotliWriteBits(nbits, d - offset, storage_ix, storage); |
314 | 0 | ++histo[distcode]; |
315 | 0 | } |
316 | | |
317 | | static BROTLI_INLINE void EmitLiterals(const uint8_t* input, const size_t len, |
318 | | const uint8_t depth[256], |
319 | | const uint16_t bits[256], |
320 | 0 | size_t* storage_ix, uint8_t* storage) { |
321 | 0 | size_t j; |
322 | 0 | for (j = 0; j < len; j++) { |
323 | 0 | const uint8_t lit = input[j]; |
324 | 0 | BrotliWriteBits(depth[lit], bits[lit], storage_ix, storage); |
325 | 0 | } |
326 | 0 | } |
327 | | |
328 | | /* REQUIRES: len <= 1 << 24. */ |
329 | | static void BrotliStoreMetaBlockHeader( |
330 | | size_t len, BROTLI_BOOL is_uncompressed, size_t* storage_ix, |
331 | 0 | uint8_t* storage) { |
332 | 0 | size_t nibbles = 6; |
333 | | /* ISLAST */ |
334 | 0 | BrotliWriteBits(1, 0, storage_ix, storage); |
335 | 0 | if (len <= (1U << 16)) { |
336 | 0 | nibbles = 4; |
337 | 0 | } else if (len <= (1U << 20)) { |
338 | 0 | nibbles = 5; |
339 | 0 | } |
340 | 0 | BrotliWriteBits(2, nibbles - 4, storage_ix, storage); |
341 | 0 | BrotliWriteBits(nibbles * 4, len - 1, storage_ix, storage); |
342 | | /* ISUNCOMPRESSED */ |
343 | 0 | BrotliWriteBits(1, (uint64_t)is_uncompressed, storage_ix, storage); |
344 | 0 | } |
345 | | |
346 | | static void UpdateBits(size_t n_bits, uint32_t bits, size_t pos, |
347 | 0 | uint8_t *array) { |
348 | 0 | while (n_bits > 0) { |
349 | 0 | size_t byte_pos = pos >> 3; |
350 | 0 | size_t n_unchanged_bits = pos & 7; |
351 | 0 | size_t n_changed_bits = BROTLI_MIN(size_t, n_bits, 8 - n_unchanged_bits); |
352 | 0 | size_t total_bits = n_unchanged_bits + n_changed_bits; |
353 | 0 | uint32_t mask = |
354 | 0 | (~((1u << total_bits) - 1u)) | ((1u << n_unchanged_bits) - 1u); |
355 | 0 | uint32_t unchanged_bits = array[byte_pos] & mask; |
356 | 0 | uint32_t changed_bits = bits & ((1u << n_changed_bits) - 1u); |
357 | 0 | array[byte_pos] = |
358 | 0 | (uint8_t)((changed_bits << n_unchanged_bits) | unchanged_bits); |
359 | 0 | n_bits -= n_changed_bits; |
360 | 0 | bits >>= n_changed_bits; |
361 | 0 | pos += n_changed_bits; |
362 | 0 | } |
363 | 0 | } |
364 | | |
365 | | static void RewindBitPosition(const size_t new_storage_ix, |
366 | 0 | size_t* storage_ix, uint8_t* storage) { |
367 | 0 | const size_t bitpos = new_storage_ix & 7; |
368 | 0 | const size_t mask = (1u << bitpos) - 1; |
369 | 0 | storage[new_storage_ix >> 3] &= (uint8_t)mask; |
370 | 0 | *storage_ix = new_storage_ix; |
371 | 0 | } |
372 | | |
373 | | static BROTLI_BOOL ShouldMergeBlock( |
374 | 0 | const uint8_t* data, size_t len, const uint8_t* depths) { |
375 | 0 | size_t histo[256] = { 0 }; |
376 | 0 | static const size_t kSampleRate = 43; |
377 | 0 | size_t i; |
378 | 0 | for (i = 0; i < len; i += kSampleRate) { |
379 | 0 | ++histo[data[i]]; |
380 | 0 | } |
381 | 0 | { |
382 | 0 | const size_t total = (len + kSampleRate - 1) / kSampleRate; |
383 | 0 | double r = (FastLog2(total) + 0.5) * (double)total + 200; |
384 | 0 | for (i = 0; i < 256; ++i) { |
385 | 0 | r -= (double)histo[i] * (depths[i] + FastLog2(histo[i])); |
386 | 0 | } |
387 | 0 | return TO_BROTLI_BOOL(r >= 0.0); |
388 | 0 | } |
389 | 0 | } |
390 | | |
391 | | /* Acceptable loss for uncompressible speedup is 2% */ |
392 | | #define MIN_RATIO 980 |
393 | | |
394 | | static BROTLI_INLINE BROTLI_BOOL ShouldUseUncompressedMode( |
395 | | const uint8_t* metablock_start, const uint8_t* next_emit, |
396 | 0 | const size_t insertlen, const size_t literal_ratio) { |
397 | 0 | const size_t compressed = (size_t)(next_emit - metablock_start); |
398 | 0 | if (compressed * 50 > insertlen) { |
399 | 0 | return BROTLI_FALSE; |
400 | 0 | } else { |
401 | 0 | return TO_BROTLI_BOOL(literal_ratio > MIN_RATIO); |
402 | 0 | } |
403 | 0 | } |
404 | | |
405 | | static void EmitUncompressedMetaBlock(const uint8_t* begin, const uint8_t* end, |
406 | | const size_t storage_ix_start, |
407 | 0 | size_t* storage_ix, uint8_t* storage) { |
408 | 0 | const size_t len = (size_t)(end - begin); |
409 | 0 | RewindBitPosition(storage_ix_start, storage_ix, storage); |
410 | 0 | BrotliStoreMetaBlockHeader(len, 1, storage_ix, storage); |
411 | 0 | *storage_ix = (*storage_ix + 7u) & ~7u; |
412 | 0 | memcpy(&storage[*storage_ix >> 3], begin, len); |
413 | 0 | *storage_ix += len << 3; |
414 | 0 | storage[*storage_ix >> 3] = 0; |
415 | 0 | } |
416 | | |
417 | | static uint32_t kCmdHistoSeed[128] = { |
418 | | 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 1, 1, 1, 1, 1, |
419 | | 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 1, 1, 1, 1, 1, 1, 1, |
420 | | 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, |
421 | | 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, |
422 | | 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, |
423 | | 1, 1, 1, 1, 0, 0, 0, 0, |
424 | | }; |
425 | | |
426 | | static BROTLI_INLINE void BrotliCompressFragmentFastImpl( |
427 | | MemoryManager* m, const uint8_t* input, size_t input_size, |
428 | | BROTLI_BOOL is_last, int* table, size_t table_bits, uint8_t cmd_depth[128], |
429 | | uint16_t cmd_bits[128], size_t* cmd_code_numbits, uint8_t* cmd_code, |
430 | 0 | size_t* storage_ix, uint8_t* storage) { |
431 | 0 | uint32_t cmd_histo[128]; |
432 | 0 | const uint8_t* ip_end; |
433 | | |
434 | | /* "next_emit" is a pointer to the first byte that is not covered by a |
435 | | previous copy. Bytes between "next_emit" and the start of the next copy or |
436 | | the end of the input will be emitted as literal bytes. */ |
437 | 0 | const uint8_t* next_emit = input; |
438 | | /* Save the start of the first block for position and distance computations. |
439 | | */ |
440 | 0 | const uint8_t* base_ip = input; |
441 | |
|
442 | 0 | static const size_t kFirstBlockSize = 3 << 15; |
443 | 0 | static const size_t kMergeBlockSize = 1 << 16; |
444 | |
|
445 | 0 | const size_t kInputMarginBytes = BROTLI_WINDOW_GAP; |
446 | 0 | const size_t kMinMatchLen = 5; |
447 | |
|
448 | 0 | const uint8_t* metablock_start = input; |
449 | 0 | size_t block_size = BROTLI_MIN(size_t, input_size, kFirstBlockSize); |
450 | 0 | size_t total_block_size = block_size; |
451 | | /* Save the bit position of the MLEN field of the meta-block header, so that |
452 | | we can update it later if we decide to extend this meta-block. */ |
453 | 0 | size_t mlen_storage_ix = *storage_ix + 3; |
454 | |
|
455 | 0 | uint8_t lit_depth[256]; |
456 | 0 | uint16_t lit_bits[256]; |
457 | |
|
458 | 0 | size_t literal_ratio; |
459 | |
|
460 | 0 | const uint8_t* ip; |
461 | 0 | int last_distance; |
462 | |
|
463 | 0 | const size_t shift = 64u - table_bits; |
464 | |
|
465 | 0 | BrotliStoreMetaBlockHeader(block_size, 0, storage_ix, storage); |
466 | | /* No block splits, no contexts. */ |
467 | 0 | BrotliWriteBits(13, 0, storage_ix, storage); |
468 | |
|
469 | 0 | literal_ratio = BuildAndStoreLiteralPrefixCode( |
470 | 0 | m, input, block_size, lit_depth, lit_bits, storage_ix, storage); |
471 | 0 | if (BROTLI_IS_OOM(m)) return; |
472 | | |
473 | 0 | { |
474 | | /* Store the pre-compressed command and distance prefix codes. */ |
475 | 0 | size_t i; |
476 | 0 | for (i = 0; i + 7 < *cmd_code_numbits; i += 8) { |
477 | 0 | BrotliWriteBits(8, cmd_code[i >> 3], storage_ix, storage); |
478 | 0 | } |
479 | 0 | } |
480 | 0 | BrotliWriteBits(*cmd_code_numbits & 7, cmd_code[*cmd_code_numbits >> 3], |
481 | 0 | storage_ix, storage); |
482 | |
|
483 | 0 | emit_commands: |
484 | | /* Initialize the command and distance histograms. We will gather |
485 | | statistics of command and distance codes during the processing |
486 | | of this block and use it to update the command and distance |
487 | | prefix codes for the next block. */ |
488 | 0 | memcpy(cmd_histo, kCmdHistoSeed, sizeof(kCmdHistoSeed)); |
489 | | |
490 | | /* "ip" is the input pointer. */ |
491 | 0 | ip = input; |
492 | 0 | last_distance = -1; |
493 | 0 | ip_end = input + block_size; |
494 | |
|
495 | 0 | if (BROTLI_PREDICT_TRUE(block_size >= kInputMarginBytes)) { |
496 | | /* For the last block, we need to keep a 16 bytes margin so that we can be |
497 | | sure that all distances are at most window size - 16. |
498 | | For all other blocks, we only need to keep a margin of 5 bytes so that |
499 | | we don't go over the block size with a copy. */ |
500 | 0 | const size_t len_limit = BROTLI_MIN(size_t, block_size - kMinMatchLen, |
501 | 0 | input_size - kInputMarginBytes); |
502 | 0 | const uint8_t* ip_limit = input + len_limit; |
503 | |
|
504 | 0 | uint32_t next_hash; |
505 | 0 | for (next_hash = Hash(++ip, shift); ; ) { |
506 | | /* Step 1: Scan forward in the input looking for a 5-byte-long match. |
507 | | If we get close to exhausting the input then goto emit_remainder. |
508 | | |
509 | | Heuristic match skipping: If 32 bytes are scanned with no matches |
510 | | found, start looking only at every other byte. If 32 more bytes are |
511 | | scanned, look at every third byte, etc.. When a match is found, |
512 | | immediately go back to looking at every byte. This is a small loss |
513 | | (~5% performance, ~0.1% density) for compressible data due to more |
514 | | bookkeeping, but for non-compressible data (such as JPEG) it's a huge |
515 | | win since the compressor quickly "realizes" the data is incompressible |
516 | | and doesn't bother looking for matches everywhere. |
517 | | |
518 | | The "skip" variable keeps track of how many bytes there are since the |
519 | | last match; dividing it by 32 (i.e. right-shifting by five) gives the |
520 | | number of bytes to move ahead for each iteration. */ |
521 | 0 | uint32_t skip = 32; |
522 | |
|
523 | 0 | const uint8_t* next_ip = ip; |
524 | 0 | const uint8_t* candidate; |
525 | 0 | assert(next_emit < ip); |
526 | 0 | trawl: |
527 | 0 | do { |
528 | 0 | uint32_t hash = next_hash; |
529 | 0 | uint32_t bytes_between_hash_lookups = skip++ >> 5; |
530 | 0 | assert(hash == Hash(next_ip, shift)); |
531 | 0 | ip = next_ip; |
532 | 0 | next_ip = ip + bytes_between_hash_lookups; |
533 | 0 | if (BROTLI_PREDICT_FALSE(next_ip > ip_limit)) { |
534 | 0 | goto emit_remainder; |
535 | 0 | } |
536 | 0 | next_hash = Hash(next_ip, shift); |
537 | 0 | candidate = ip - last_distance; |
538 | 0 | if (IsMatch(ip, candidate)) { |
539 | 0 | if (BROTLI_PREDICT_TRUE(candidate < ip)) { |
540 | 0 | table[hash] = (int)(ip - base_ip); |
541 | 0 | break; |
542 | 0 | } |
543 | 0 | } |
544 | 0 | candidate = base_ip + table[hash]; |
545 | 0 | assert(candidate >= base_ip); |
546 | 0 | assert(candidate < ip); |
547 | | |
548 | 0 | table[hash] = (int)(ip - base_ip); |
549 | 0 | } while (BROTLI_PREDICT_TRUE(!IsMatch(ip, candidate))); |
550 | | |
551 | | /* Check copy distance. If candidate is not feasible, continue search. |
552 | | Checking is done outside of hot loop to reduce overhead. */ |
553 | 0 | if (ip - candidate > MAX_DISTANCE) goto trawl; |
554 | | |
555 | | /* Step 2: Emit the found match together with the literal bytes from |
556 | | "next_emit" to the bit stream, and then see if we can find a next match |
557 | | immediately afterwards. Repeat until we find no match for the input |
558 | | without emitting some literal bytes. */ |
559 | | |
560 | 0 | { |
561 | | /* We have a 5-byte match at ip, and we need to emit bytes in |
562 | | [next_emit, ip). */ |
563 | 0 | const uint8_t* base = ip; |
564 | 0 | size_t matched = 5 + FindMatchLengthWithLimit( |
565 | 0 | candidate + 5, ip + 5, (size_t)(ip_end - ip) - 5); |
566 | 0 | int distance = (int)(base - candidate); /* > 0 */ |
567 | 0 | size_t insert = (size_t)(base - next_emit); |
568 | 0 | ip += matched; |
569 | 0 | assert(0 == memcmp(base, candidate, matched)); |
570 | 0 | if (BROTLI_PREDICT_TRUE(insert < 6210)) { |
571 | 0 | EmitInsertLen(insert, cmd_depth, cmd_bits, cmd_histo, |
572 | 0 | storage_ix, storage); |
573 | 0 | } else if (ShouldUseUncompressedMode(metablock_start, next_emit, insert, |
574 | 0 | literal_ratio)) { |
575 | 0 | EmitUncompressedMetaBlock(metablock_start, base, mlen_storage_ix - 3, |
576 | 0 | storage_ix, storage); |
577 | 0 | input_size -= (size_t)(base - input); |
578 | 0 | input = base; |
579 | 0 | next_emit = input; |
580 | 0 | goto next_block; |
581 | 0 | } else { |
582 | 0 | EmitLongInsertLen(insert, cmd_depth, cmd_bits, cmd_histo, |
583 | 0 | storage_ix, storage); |
584 | 0 | } |
585 | 0 | EmitLiterals(next_emit, insert, lit_depth, lit_bits, |
586 | 0 | storage_ix, storage); |
587 | 0 | if (distance == last_distance) { |
588 | 0 | BrotliWriteBits(cmd_depth[64], cmd_bits[64], storage_ix, storage); |
589 | 0 | ++cmd_histo[64]; |
590 | 0 | } else { |
591 | 0 | EmitDistance((size_t)distance, cmd_depth, cmd_bits, |
592 | 0 | cmd_histo, storage_ix, storage); |
593 | 0 | last_distance = distance; |
594 | 0 | } |
595 | 0 | EmitCopyLenLastDistance(matched, cmd_depth, cmd_bits, cmd_histo, |
596 | 0 | storage_ix, storage); |
597 | |
|
598 | 0 | next_emit = ip; |
599 | 0 | if (BROTLI_PREDICT_FALSE(ip >= ip_limit)) { |
600 | 0 | goto emit_remainder; |
601 | 0 | } |
602 | | /* We could immediately start working at ip now, but to improve |
603 | | compression we first update "table" with the hashes of some positions |
604 | | within the last copy. */ |
605 | 0 | { |
606 | 0 | uint64_t input_bytes = BROTLI_UNALIGNED_LOAD64LE(ip - 3); |
607 | 0 | uint32_t prev_hash = HashBytesAtOffset(input_bytes, 0, shift); |
608 | 0 | uint32_t cur_hash = HashBytesAtOffset(input_bytes, 3, shift); |
609 | 0 | table[prev_hash] = (int)(ip - base_ip - 3); |
610 | 0 | prev_hash = HashBytesAtOffset(input_bytes, 1, shift); |
611 | 0 | table[prev_hash] = (int)(ip - base_ip - 2); |
612 | 0 | prev_hash = HashBytesAtOffset(input_bytes, 2, shift); |
613 | 0 | table[prev_hash] = (int)(ip - base_ip - 1); |
614 | |
|
615 | 0 | candidate = base_ip + table[cur_hash]; |
616 | 0 | table[cur_hash] = (int)(ip - base_ip); |
617 | 0 | } |
618 | 0 | } |
619 | | |
620 | 0 | while (IsMatch(ip, candidate)) { |
621 | | /* We have a 5-byte match at ip, and no need to emit any literal bytes |
622 | | prior to ip. */ |
623 | 0 | const uint8_t* base = ip; |
624 | 0 | size_t matched = 5 + FindMatchLengthWithLimit( |
625 | 0 | candidate + 5, ip + 5, (size_t)(ip_end - ip) - 5); |
626 | 0 | if (ip - candidate > MAX_DISTANCE) break; |
627 | 0 | ip += matched; |
628 | 0 | last_distance = (int)(base - candidate); /* > 0 */ |
629 | 0 | assert(0 == memcmp(base, candidate, matched)); |
630 | 0 | EmitCopyLen(matched, cmd_depth, cmd_bits, cmd_histo, |
631 | 0 | storage_ix, storage); |
632 | 0 | EmitDistance((size_t)last_distance, cmd_depth, cmd_bits, |
633 | 0 | cmd_histo, storage_ix, storage); |
634 | |
|
635 | 0 | next_emit = ip; |
636 | 0 | if (BROTLI_PREDICT_FALSE(ip >= ip_limit)) { |
637 | 0 | goto emit_remainder; |
638 | 0 | } |
639 | | /* We could immediately start working at ip now, but to improve |
640 | | compression we first update "table" with the hashes of some positions |
641 | | within the last copy. */ |
642 | 0 | { |
643 | 0 | uint64_t input_bytes = BROTLI_UNALIGNED_LOAD64LE(ip - 3); |
644 | 0 | uint32_t prev_hash = HashBytesAtOffset(input_bytes, 0, shift); |
645 | 0 | uint32_t cur_hash = HashBytesAtOffset(input_bytes, 3, shift); |
646 | 0 | table[prev_hash] = (int)(ip - base_ip - 3); |
647 | 0 | prev_hash = HashBytesAtOffset(input_bytes, 1, shift); |
648 | 0 | table[prev_hash] = (int)(ip - base_ip - 2); |
649 | 0 | prev_hash = HashBytesAtOffset(input_bytes, 2, shift); |
650 | 0 | table[prev_hash] = (int)(ip - base_ip - 1); |
651 | |
|
652 | 0 | candidate = base_ip + table[cur_hash]; |
653 | 0 | table[cur_hash] = (int)(ip - base_ip); |
654 | 0 | } |
655 | 0 | } |
656 | | |
657 | 0 | next_hash = Hash(++ip, shift); |
658 | 0 | } |
659 | 0 | } |
660 | | |
661 | 0 | emit_remainder: |
662 | 0 | assert(next_emit <= ip_end); |
663 | 0 | input += block_size; |
664 | 0 | input_size -= block_size; |
665 | 0 | block_size = BROTLI_MIN(size_t, input_size, kMergeBlockSize); |
666 | | |
667 | | /* Decide if we want to continue this meta-block instead of emitting the |
668 | | last insert-only command. */ |
669 | 0 | if (input_size > 0 && |
670 | 0 | total_block_size + block_size <= (1 << 20) && |
671 | 0 | ShouldMergeBlock(input, block_size, lit_depth)) { |
672 | 0 | assert(total_block_size > (1 << 16)); |
673 | | /* Update the size of the current meta-block and continue emitting commands. |
674 | | We can do this because the current size and the new size both have 5 |
675 | | nibbles. */ |
676 | 0 | total_block_size += block_size; |
677 | 0 | UpdateBits(20, (uint32_t)(total_block_size - 1), mlen_storage_ix, storage); |
678 | 0 | goto emit_commands; |
679 | 0 | } |
680 | | |
681 | | /* Emit the remaining bytes as literals. */ |
682 | 0 | if (next_emit < ip_end) { |
683 | 0 | const size_t insert = (size_t)(ip_end - next_emit); |
684 | 0 | if (BROTLI_PREDICT_TRUE(insert < 6210)) { |
685 | 0 | EmitInsertLen(insert, cmd_depth, cmd_bits, cmd_histo, |
686 | 0 | storage_ix, storage); |
687 | 0 | EmitLiterals(next_emit, insert, lit_depth, lit_bits, storage_ix, storage); |
688 | 0 | } else if (ShouldUseUncompressedMode(metablock_start, next_emit, insert, |
689 | 0 | literal_ratio)) { |
690 | 0 | EmitUncompressedMetaBlock(metablock_start, ip_end, mlen_storage_ix - 3, |
691 | 0 | storage_ix, storage); |
692 | 0 | } else { |
693 | 0 | EmitLongInsertLen(insert, cmd_depth, cmd_bits, cmd_histo, |
694 | 0 | storage_ix, storage); |
695 | 0 | EmitLiterals(next_emit, insert, lit_depth, lit_bits, |
696 | 0 | storage_ix, storage); |
697 | 0 | } |
698 | 0 | } |
699 | 0 | next_emit = ip_end; |
700 | |
|
701 | 0 | next_block: |
702 | | /* If we have more data, write a new meta-block header and prefix codes and |
703 | | then continue emitting commands. */ |
704 | 0 | if (input_size > 0) { |
705 | 0 | metablock_start = input; |
706 | 0 | block_size = BROTLI_MIN(size_t, input_size, kFirstBlockSize); |
707 | 0 | total_block_size = block_size; |
708 | | /* Save the bit position of the MLEN field of the meta-block header, so that |
709 | | we can update it later if we decide to extend this meta-block. */ |
710 | 0 | mlen_storage_ix = *storage_ix + 3; |
711 | 0 | BrotliStoreMetaBlockHeader(block_size, 0, storage_ix, storage); |
712 | | /* No block splits, no contexts. */ |
713 | 0 | BrotliWriteBits(13, 0, storage_ix, storage); |
714 | 0 | literal_ratio = BuildAndStoreLiteralPrefixCode( |
715 | 0 | m, input, block_size, lit_depth, lit_bits, storage_ix, storage); |
716 | 0 | if (BROTLI_IS_OOM(m)) return; |
717 | 0 | BuildAndStoreCommandPrefixCode(cmd_histo, cmd_depth, cmd_bits, |
718 | 0 | storage_ix, storage); |
719 | 0 | goto emit_commands; |
720 | 0 | } |
721 | | |
722 | 0 | if (!is_last) { |
723 | | /* If this is not the last block, update the command and distance prefix |
724 | | codes for the next block and store the compressed forms. */ |
725 | 0 | cmd_code[0] = 0; |
726 | 0 | *cmd_code_numbits = 0; |
727 | 0 | BuildAndStoreCommandPrefixCode(cmd_histo, cmd_depth, cmd_bits, |
728 | 0 | cmd_code_numbits, cmd_code); |
729 | 0 | } |
730 | 0 | } |
731 | | |
732 | 0 | #define FOR_TABLE_BITS_(X) X(9) X(11) X(13) X(15) |
733 | | |
734 | | #define BAKE_METHOD_PARAM_(B) \ |
735 | | static BROTLI_NOINLINE void BrotliCompressFragmentFastImpl ## B( \ |
736 | | MemoryManager* m, const uint8_t* input, size_t input_size, \ |
737 | | BROTLI_BOOL is_last, int* table, uint8_t cmd_depth[128], \ |
738 | | uint16_t cmd_bits[128], size_t* cmd_code_numbits, uint8_t* cmd_code, \ |
739 | 0 | size_t* storage_ix, uint8_t* storage) { \ |
740 | 0 | BrotliCompressFragmentFastImpl(m, input, input_size, is_last, table, B, \ |
741 | 0 | cmd_depth, cmd_bits, cmd_code_numbits, cmd_code, storage_ix, storage); \ |
742 | 0 | } Unexecuted instantiation: compress_fragment.c:BrotliCompressFragmentFastImpl9 Unexecuted instantiation: compress_fragment.c:BrotliCompressFragmentFastImpl11 Unexecuted instantiation: compress_fragment.c:BrotliCompressFragmentFastImpl13 Unexecuted instantiation: compress_fragment.c:BrotliCompressFragmentFastImpl15 |
743 | | FOR_TABLE_BITS_(BAKE_METHOD_PARAM_) |
744 | | #undef BAKE_METHOD_PARAM_ |
745 | | |
746 | | void BrotliCompressFragmentFast( |
747 | | MemoryManager* m, const uint8_t* input, size_t input_size, |
748 | | BROTLI_BOOL is_last, int* table, size_t table_size, uint8_t cmd_depth[128], |
749 | | uint16_t cmd_bits[128], size_t* cmd_code_numbits, uint8_t* cmd_code, |
750 | 0 | size_t* storage_ix, uint8_t* storage) { |
751 | 0 | const size_t initial_storage_ix = *storage_ix; |
752 | 0 | const size_t table_bits = Log2FloorNonZero(table_size); |
753 | |
|
754 | 0 | if (input_size == 0) { |
755 | 0 | assert(is_last); |
756 | 0 | BrotliWriteBits(1, 1, storage_ix, storage); /* islast */ |
757 | 0 | BrotliWriteBits(1, 1, storage_ix, storage); /* isempty */ |
758 | 0 | *storage_ix = (*storage_ix + 7u) & ~7u; |
759 | 0 | return; |
760 | 0 | } |
761 | | |
762 | 0 | switch (table_bits) { |
763 | 0 | #define CASE_(B) \ |
764 | 0 | case B: \ |
765 | 0 | BrotliCompressFragmentFastImpl ## B( \ |
766 | 0 | m, input, input_size, is_last, table, cmd_depth, cmd_bits, \ |
767 | 0 | cmd_code_numbits, cmd_code, storage_ix, storage); \ |
768 | 0 | break; |
769 | 0 | FOR_TABLE_BITS_(CASE_) |
770 | 0 | #undef CASE_ |
771 | 0 | default: assert(0); break; |
772 | 0 | } |
773 | | |
774 | | /* If output is larger than single uncompressed block, rewrite it. */ |
775 | 0 | if (*storage_ix - initial_storage_ix > 31 + (input_size << 3)) { |
776 | 0 | EmitUncompressedMetaBlock(input, input + input_size, initial_storage_ix, |
777 | 0 | storage_ix, storage); |
778 | 0 | } |
779 | |
|
780 | 0 | if (is_last) { |
781 | 0 | BrotliWriteBits(1, 1, storage_ix, storage); /* islast */ |
782 | 0 | BrotliWriteBits(1, 1, storage_ix, storage); /* isempty */ |
783 | 0 | *storage_ix = (*storage_ix + 7u) & ~7u; |
784 | 0 | } |
785 | 0 | } |
786 | | |
787 | | #undef FOR_TABLE_BITS_ |
788 | | |
789 | | #if defined(__cplusplus) || defined(c_plusplus) |
790 | | } /* extern "C" */ |
791 | | #endif |