1/*
2 * Copyright (c) Meta Platforms, Inc. and affiliates.
3 * All rights reserved.
4 *
5 * This source code is licensed under both the BSD-style license (found in the
6 * LICENSE file in the root directory of this source tree) and the GPLv2 (found
7 * in the COPYING file in the root directory of this source tree).
8 * You may select, at your option, one of the above-listed licenses.
9 */
10
11/* zstd_decompress_block :
12 * this module takes care of decompressing _compressed_ block */
13
14/*-*******************************************************
15* Dependencies
16*********************************************************/
17#include "../common/zstd_deps.h" /* ZSTD_memcpy, ZSTD_memmove, ZSTD_memset */
18#include "../common/compiler.h" /* prefetch */
19#include "../common/cpu.h" /* bmi2 */
20#include "../common/mem.h" /* low level memory routines */
21#define FSE_STATIC_LINKING_ONLY
22#include "../common/fse.h"
23#include "../common/huf.h"
24#include "../common/zstd_internal.h"
25#include "zstd_decompress_internal.h" /* ZSTD_DCtx */
26#include "zstd_ddict.h" /* ZSTD_DDictDictContent */
27#include "zstd_decompress_block.h"
28#include "../common/bits.h" /* ZSTD_highbit32 */
29
30/*_*******************************************************
31* Macros
32**********************************************************/
33
34/* These two optional macros force the use one way or another of the two
35 * ZSTD_decompressSequences implementations. You can't force in both directions
36 * at the same time.
37 */
38#if defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
39 defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
40#error "Cannot force the use of the short and the long ZSTD_decompressSequences variants!"
41#endif
42
43
44/*_*******************************************************
45* Memory operations
46**********************************************************/
47static void ZSTD_copy4(void* dst, const void* src) { ZSTD_memcpy(dst, src, 4); }
48
49
50/*-*************************************************************
51 * Block decoding
52 ***************************************************************/
53
54/*! ZSTD_getcBlockSize() :
55 * Provides the size of compressed block from block header `src` */
56size_t ZSTD_getcBlockSize(const void* src, size_t srcSize,
57 blockProperties_t* bpPtr)
58{
59 RETURN_ERROR_IF(srcSize < ZSTD_blockHeaderSize, srcSize_wrong, "");
60
61 { U32 const cBlockHeader = MEM_readLE24(src);
62 U32 const cSize = cBlockHeader >> 3;
63 bpPtr->lastBlock = cBlockHeader & 1;
64 bpPtr->blockType = (blockType_e)((cBlockHeader >> 1) & 3);
65 bpPtr->origSize = cSize; /* only useful for RLE */
66 if (bpPtr->blockType == bt_rle) return 1;
67 RETURN_ERROR_IF(bpPtr->blockType == bt_reserved, corruption_detected, "");
68 return cSize;
69 }
70}
71
72/* Allocate buffer for literals, either overlapping current dst, or split between dst and litExtraBuffer, or stored entirely within litExtraBuffer */
73static void ZSTD_allocateLiteralsBuffer(ZSTD_DCtx* dctx, void* const dst, const size_t dstCapacity, const size_t litSize,
74 const streaming_operation streaming, const size_t expectedWriteSize, const unsigned splitImmediately)
75{
76 if (streaming == not_streaming && dstCapacity > ZSTD_BLOCKSIZE_MAX + WILDCOPY_OVERLENGTH + litSize + WILDCOPY_OVERLENGTH)
77 {
78 /* room for litbuffer to fit without read faulting */
79 dctx->litBuffer = (BYTE*)dst + ZSTD_BLOCKSIZE_MAX + WILDCOPY_OVERLENGTH;
80 dctx->litBufferEnd = dctx->litBuffer + litSize;
81 dctx->litBufferLocation = ZSTD_in_dst;
82 }
83 else if (litSize > ZSTD_LITBUFFEREXTRASIZE)
84 {
85 /* won't fit in litExtraBuffer, so it will be split between end of dst and extra buffer */
86 if (splitImmediately) {
87 /* won't fit in litExtraBuffer, so it will be split between end of dst and extra buffer */
88 dctx->litBuffer = (BYTE*)dst + expectedWriteSize - litSize + ZSTD_LITBUFFEREXTRASIZE - WILDCOPY_OVERLENGTH;
89 dctx->litBufferEnd = dctx->litBuffer + litSize - ZSTD_LITBUFFEREXTRASIZE;
90 }
91 else {
92 /* initially this will be stored entirely in dst during huffman decoding, it will partially be shifted to litExtraBuffer after */
93 dctx->litBuffer = (BYTE*)dst + expectedWriteSize - litSize;
94 dctx->litBufferEnd = (BYTE*)dst + expectedWriteSize;
95 }
96 dctx->litBufferLocation = ZSTD_split;
97 }
98 else
99 {
100 /* fits entirely within litExtraBuffer, so no split is necessary */
101 dctx->litBuffer = dctx->litExtraBuffer;
102 dctx->litBufferEnd = dctx->litBuffer + litSize;
103 dctx->litBufferLocation = ZSTD_not_in_dst;
104 }
105}
106
107/* Hidden declaration for fullbench */
108size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
109 const void* src, size_t srcSize,
110 void* dst, size_t dstCapacity, const streaming_operation streaming);
111/*! ZSTD_decodeLiteralsBlock() :
112 * Where it is possible to do so without being stomped by the output during decompression, the literals block will be stored
113 * in the dstBuffer. If there is room to do so, it will be stored in full in the excess dst space after where the current
114 * block will be output. Otherwise it will be stored at the end of the current dst blockspace, with a small portion being
115 * stored in dctx->litExtraBuffer to help keep it "ahead" of the current output write.
116 *
117 * @return : nb of bytes read from src (< srcSize )
118 * note : symbol not declared but exposed for fullbench */
119size_t ZSTD_decodeLiteralsBlock(ZSTD_DCtx* dctx,
120 const void* src, size_t srcSize, /* note : srcSize < BLOCKSIZE */
121 void* dst, size_t dstCapacity, const streaming_operation streaming)
122{
123 DEBUGLOG(5, "ZSTD_decodeLiteralsBlock");
124 RETURN_ERROR_IF(srcSize < MIN_CBLOCK_SIZE, corruption_detected, "");
125
126 { const BYTE* const istart = (const BYTE*) src;
127 symbolEncodingType_e const litEncType = (symbolEncodingType_e)(istart[0] & 3);
128
129 switch(litEncType)
130 {
131 case set_repeat:
132 DEBUGLOG(5, "set_repeat flag : re-using stats from previous compressed literals block");
133 RETURN_ERROR_IF(dctx->litEntropy==0, dictionary_corrupted, "");
134 ZSTD_FALLTHROUGH;
135
136 case set_compressed:
137 RETURN_ERROR_IF(srcSize < 5, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need up to 5 for case 3");
138 { size_t lhSize, litSize, litCSize;
139 U32 singleStream=0;
140 U32 const lhlCode = (istart[0] >> 2) & 3;
141 U32 const lhc = MEM_readLE32(istart);
142 size_t hufSuccess;
143 size_t expectedWriteSize = MIN(ZSTD_BLOCKSIZE_MAX, dstCapacity);
144 int const flags = 0
145 | (ZSTD_DCtx_get_bmi2(dctx) ? HUF_flags_bmi2 : 0)
146 | (dctx->disableHufAsm ? HUF_flags_disableAsm : 0);
147 switch(lhlCode)
148 {
149 case 0: case 1: default: /* note : default is impossible, since lhlCode into [0..3] */
150 /* 2 - 2 - 10 - 10 */
151 singleStream = !lhlCode;
152 lhSize = 3;
153 litSize = (lhc >> 4) & 0x3FF;
154 litCSize = (lhc >> 14) & 0x3FF;
155 break;
156 case 2:
157 /* 2 - 2 - 14 - 14 */
158 lhSize = 4;
159 litSize = (lhc >> 4) & 0x3FFF;
160 litCSize = lhc >> 18;
161 break;
162 case 3:
163 /* 2 - 2 - 18 - 18 */
164 lhSize = 5;
165 litSize = (lhc >> 4) & 0x3FFFF;
166 litCSize = (lhc >> 22) + ((size_t)istart[4] << 10);
167 break;
168 }
169 RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
170 RETURN_ERROR_IF(litSize > ZSTD_BLOCKSIZE_MAX, corruption_detected, "");
171 if (!singleStream)
172 RETURN_ERROR_IF(litSize < MIN_LITERALS_FOR_4_STREAMS, literals_headerWrong,
173 "Not enough literals (%zu) for the 4-streams mode (min %u)",
174 litSize, MIN_LITERALS_FOR_4_STREAMS);
175 RETURN_ERROR_IF(litCSize + lhSize > srcSize, corruption_detected, "");
176 RETURN_ERROR_IF(expectedWriteSize < litSize , dstSize_tooSmall, "");
177 ZSTD_allocateLiteralsBuffer(dctx, dst, dstCapacity, litSize, streaming, expectedWriteSize, 0);
178
179 /* prefetch huffman table if cold */
180 if (dctx->ddictIsCold && (litSize > 768 /* heuristic */)) {
181 PREFETCH_AREA(dctx->HUFptr, sizeof(dctx->entropy.hufTable));
182 }
183
184 if (litEncType==set_repeat) {
185 if (singleStream) {
186 hufSuccess = HUF_decompress1X_usingDTable(
187 dctx->litBuffer, litSize, istart+lhSize, litCSize,
188 dctx->HUFptr, flags);
189 } else {
190 assert(litSize >= MIN_LITERALS_FOR_4_STREAMS);
191 hufSuccess = HUF_decompress4X_usingDTable(
192 dctx->litBuffer, litSize, istart+lhSize, litCSize,
193 dctx->HUFptr, flags);
194 }
195 } else {
196 if (singleStream) {
197#if defined(HUF_FORCE_DECOMPRESS_X2)
198 hufSuccess = HUF_decompress1X_DCtx_wksp(
199 dctx->entropy.hufTable, dctx->litBuffer, litSize,
200 istart+lhSize, litCSize, dctx->workspace,
201 sizeof(dctx->workspace), flags);
202#else
203 hufSuccess = HUF_decompress1X1_DCtx_wksp(
204 dctx->entropy.hufTable, dctx->litBuffer, litSize,
205 istart+lhSize, litCSize, dctx->workspace,
206 sizeof(dctx->workspace), flags);
207#endif
208 } else {
209 hufSuccess = HUF_decompress4X_hufOnly_wksp(
210 dctx->entropy.hufTable, dctx->litBuffer, litSize,
211 istart+lhSize, litCSize, dctx->workspace,
212 sizeof(dctx->workspace), flags);
213 }
214 }
215 if (dctx->litBufferLocation == ZSTD_split)
216 {
217 ZSTD_memcpy(dctx->litExtraBuffer, dctx->litBufferEnd - ZSTD_LITBUFFEREXTRASIZE, ZSTD_LITBUFFEREXTRASIZE);
218 ZSTD_memmove(dctx->litBuffer + ZSTD_LITBUFFEREXTRASIZE - WILDCOPY_OVERLENGTH, dctx->litBuffer, litSize - ZSTD_LITBUFFEREXTRASIZE);
219 dctx->litBuffer += ZSTD_LITBUFFEREXTRASIZE - WILDCOPY_OVERLENGTH;
220 dctx->litBufferEnd -= WILDCOPY_OVERLENGTH;
221 }
222
223 RETURN_ERROR_IF(HUF_isError(hufSuccess), corruption_detected, "");
224
225 dctx->litPtr = dctx->litBuffer;
226 dctx->litSize = litSize;
227 dctx->litEntropy = 1;
228 if (litEncType==set_compressed) dctx->HUFptr = dctx->entropy.hufTable;
229 return litCSize + lhSize;
230 }
231
232 case set_basic:
233 { size_t litSize, lhSize;
234 U32 const lhlCode = ((istart[0]) >> 2) & 3;
235 size_t expectedWriteSize = MIN(ZSTD_BLOCKSIZE_MAX, dstCapacity);
236 switch(lhlCode)
237 {
238 case 0: case 2: default: /* note : default is impossible, since lhlCode into [0..3] */
239 lhSize = 1;
240 litSize = istart[0] >> 3;
241 break;
242 case 1:
243 lhSize = 2;
244 litSize = MEM_readLE16(istart) >> 4;
245 break;
246 case 3:
247 lhSize = 3;
248 RETURN_ERROR_IF(srcSize<3, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize = 3");
249 litSize = MEM_readLE24(istart) >> 4;
250 break;
251 }
252
253 RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
254 RETURN_ERROR_IF(expectedWriteSize < litSize, dstSize_tooSmall, "");
255 ZSTD_allocateLiteralsBuffer(dctx, dst, dstCapacity, litSize, streaming, expectedWriteSize, 1);
256 if (lhSize+litSize+WILDCOPY_OVERLENGTH > srcSize) { /* risk reading beyond src buffer with wildcopy */
257 RETURN_ERROR_IF(litSize+lhSize > srcSize, corruption_detected, "");
258 if (dctx->litBufferLocation == ZSTD_split)
259 {
260 ZSTD_memcpy(dctx->litBuffer, istart + lhSize, litSize - ZSTD_LITBUFFEREXTRASIZE);
261 ZSTD_memcpy(dctx->litExtraBuffer, istart + lhSize + litSize - ZSTD_LITBUFFEREXTRASIZE, ZSTD_LITBUFFEREXTRASIZE);
262 }
263 else
264 {
265 ZSTD_memcpy(dctx->litBuffer, istart + lhSize, litSize);
266 }
267 dctx->litPtr = dctx->litBuffer;
268 dctx->litSize = litSize;
269 return lhSize+litSize;
270 }
271 /* direct reference into compressed stream */
272 dctx->litPtr = istart+lhSize;
273 dctx->litSize = litSize;
274 dctx->litBufferEnd = dctx->litPtr + litSize;
275 dctx->litBufferLocation = ZSTD_not_in_dst;
276 return lhSize+litSize;
277 }
278
279 case set_rle:
280 { U32 const lhlCode = ((istart[0]) >> 2) & 3;
281 size_t litSize, lhSize;
282 size_t expectedWriteSize = MIN(ZSTD_BLOCKSIZE_MAX, dstCapacity);
283 switch(lhlCode)
284 {
285 case 0: case 2: default: /* note : default is impossible, since lhlCode into [0..3] */
286 lhSize = 1;
287 litSize = istart[0] >> 3;
288 break;
289 case 1:
290 lhSize = 2;
291 RETURN_ERROR_IF(srcSize<3, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize+1 = 3");
292 litSize = MEM_readLE16(istart) >> 4;
293 break;
294 case 3:
295 lhSize = 3;
296 RETURN_ERROR_IF(srcSize<4, corruption_detected, "srcSize >= MIN_CBLOCK_SIZE == 2; here we need lhSize+1 = 4");
297 litSize = MEM_readLE24(istart) >> 4;
298 break;
299 }
300 RETURN_ERROR_IF(litSize > 0 && dst == NULL, dstSize_tooSmall, "NULL not handled");
301 RETURN_ERROR_IF(litSize > ZSTD_BLOCKSIZE_MAX, corruption_detected, "");
302 RETURN_ERROR_IF(expectedWriteSize < litSize, dstSize_tooSmall, "");
303 ZSTD_allocateLiteralsBuffer(dctx, dst, dstCapacity, litSize, streaming, expectedWriteSize, 1);
304 if (dctx->litBufferLocation == ZSTD_split)
305 {
306 ZSTD_memset(dctx->litBuffer, istart[lhSize], litSize - ZSTD_LITBUFFEREXTRASIZE);
307 ZSTD_memset(dctx->litExtraBuffer, istart[lhSize], ZSTD_LITBUFFEREXTRASIZE);
308 }
309 else
310 {
311 ZSTD_memset(dctx->litBuffer, istart[lhSize], litSize);
312 }
313 dctx->litPtr = dctx->litBuffer;
314 dctx->litSize = litSize;
315 return lhSize+1;
316 }
317 default:
318 RETURN_ERROR(corruption_detected, "impossible");
319 }
320 }
321}
322
323/* Default FSE distribution tables.
324 * These are pre-calculated FSE decoding tables using default distributions as defined in specification :
325 * https://github.com/facebook/zstd/blob/release/doc/zstd_compression_format.md#default-distributions
326 * They were generated programmatically with following method :
327 * - start from default distributions, present in /lib/common/zstd_internal.h
328 * - generate tables normally, using ZSTD_buildFSETable()
329 * - printout the content of tables
330 * - pretify output, report below, test with fuzzer to ensure it's correct */
331
332/* Default FSE distribution table for Literal Lengths */
333static const ZSTD_seqSymbol LL_defaultDTable[(1<<LL_DEFAULTNORMLOG)+1] = {
334 { 1, 1, 1, LL_DEFAULTNORMLOG}, /* header : fastMode, tableLog */
335 /* nextState, nbAddBits, nbBits, baseVal */
336 { 0, 0, 4, 0}, { 16, 0, 4, 0},
337 { 32, 0, 5, 1}, { 0, 0, 5, 3},
338 { 0, 0, 5, 4}, { 0, 0, 5, 6},
339 { 0, 0, 5, 7}, { 0, 0, 5, 9},
340 { 0, 0, 5, 10}, { 0, 0, 5, 12},
341 { 0, 0, 6, 14}, { 0, 1, 5, 16},
342 { 0, 1, 5, 20}, { 0, 1, 5, 22},
343 { 0, 2, 5, 28}, { 0, 3, 5, 32},
344 { 0, 4, 5, 48}, { 32, 6, 5, 64},
345 { 0, 7, 5, 128}, { 0, 8, 6, 256},
346 { 0, 10, 6, 1024}, { 0, 12, 6, 4096},
347 { 32, 0, 4, 0}, { 0, 0, 4, 1},
348 { 0, 0, 5, 2}, { 32, 0, 5, 4},
349 { 0, 0, 5, 5}, { 32, 0, 5, 7},
350 { 0, 0, 5, 8}, { 32, 0, 5, 10},
351 { 0, 0, 5, 11}, { 0, 0, 6, 13},
352 { 32, 1, 5, 16}, { 0, 1, 5, 18},
353 { 32, 1, 5, 22}, { 0, 2, 5, 24},
354 { 32, 3, 5, 32}, { 0, 3, 5, 40},
355 { 0, 6, 4, 64}, { 16, 6, 4, 64},
356 { 32, 7, 5, 128}, { 0, 9, 6, 512},
357 { 0, 11, 6, 2048}, { 48, 0, 4, 0},
358 { 16, 0, 4, 1}, { 32, 0, 5, 2},
359 { 32, 0, 5, 3}, { 32, 0, 5, 5},
360 { 32, 0, 5, 6}, { 32, 0, 5, 8},
361 { 32, 0, 5, 9}, { 32, 0, 5, 11},
362 { 32, 0, 5, 12}, { 0, 0, 6, 15},
363 { 32, 1, 5, 18}, { 32, 1, 5, 20},
364 { 32, 2, 5, 24}, { 32, 2, 5, 28},
365 { 32, 3, 5, 40}, { 32, 4, 5, 48},
366 { 0, 16, 6,65536}, { 0, 15, 6,32768},
367 { 0, 14, 6,16384}, { 0, 13, 6, 8192},
368}; /* LL_defaultDTable */
369
370/* Default FSE distribution table for Offset Codes */
371static const ZSTD_seqSymbol OF_defaultDTable[(1<<OF_DEFAULTNORMLOG)+1] = {
372 { 1, 1, 1, OF_DEFAULTNORMLOG}, /* header : fastMode, tableLog */
373 /* nextState, nbAddBits, nbBits, baseVal */
374 { 0, 0, 5, 0}, { 0, 6, 4, 61},
375 { 0, 9, 5, 509}, { 0, 15, 5,32765},
376 { 0, 21, 5,2097149}, { 0, 3, 5, 5},
377 { 0, 7, 4, 125}, { 0, 12, 5, 4093},
378 { 0, 18, 5,262141}, { 0, 23, 5,8388605},
379 { 0, 5, 5, 29}, { 0, 8, 4, 253},
380 { 0, 14, 5,16381}, { 0, 20, 5,1048573},
381 { 0, 2, 5, 1}, { 16, 7, 4, 125},
382 { 0, 11, 5, 2045}, { 0, 17, 5,131069},
383 { 0, 22, 5,4194301}, { 0, 4, 5, 13},
384 { 16, 8, 4, 253}, { 0, 13, 5, 8189},
385 { 0, 19, 5,524285}, { 0, 1, 5, 1},
386 { 16, 6, 4, 61}, { 0, 10, 5, 1021},
387 { 0, 16, 5,65533}, { 0, 28, 5,268435453},
388 { 0, 27, 5,134217725}, { 0, 26, 5,67108861},
389 { 0, 25, 5,33554429}, { 0, 24, 5,16777213},
390}; /* OF_defaultDTable */
391
392
393/* Default FSE distribution table for Match Lengths */
394static const ZSTD_seqSymbol ML_defaultDTable[(1<<ML_DEFAULTNORMLOG)+1] = {
395 { 1, 1, 1, ML_DEFAULTNORMLOG}, /* header : fastMode, tableLog */
396 /* nextState, nbAddBits, nbBits, baseVal */
397 { 0, 0, 6, 3}, { 0, 0, 4, 4},
398 { 32, 0, 5, 5}, { 0, 0, 5, 6},
399 { 0, 0, 5, 8}, { 0, 0, 5, 9},
400 { 0, 0, 5, 11}, { 0, 0, 6, 13},
401 { 0, 0, 6, 16}, { 0, 0, 6, 19},
402 { 0, 0, 6, 22}, { 0, 0, 6, 25},
403 { 0, 0, 6, 28}, { 0, 0, 6, 31},
404 { 0, 0, 6, 34}, { 0, 1, 6, 37},
405 { 0, 1, 6, 41}, { 0, 2, 6, 47},
406 { 0, 3, 6, 59}, { 0, 4, 6, 83},
407 { 0, 7, 6, 131}, { 0, 9, 6, 515},
408 { 16, 0, 4, 4}, { 0, 0, 4, 5},
409 { 32, 0, 5, 6}, { 0, 0, 5, 7},
410 { 32, 0, 5, 9}, { 0, 0, 5, 10},
411 { 0, 0, 6, 12}, { 0, 0, 6, 15},
412 { 0, 0, 6, 18}, { 0, 0, 6, 21},
413 { 0, 0, 6, 24}, { 0, 0, 6, 27},
414 { 0, 0, 6, 30}, { 0, 0, 6, 33},
415 { 0, 1, 6, 35}, { 0, 1, 6, 39},
416 { 0, 2, 6, 43}, { 0, 3, 6, 51},
417 { 0, 4, 6, 67}, { 0, 5, 6, 99},
418 { 0, 8, 6, 259}, { 32, 0, 4, 4},
419 { 48, 0, 4, 4}, { 16, 0, 4, 5},
420 { 32, 0, 5, 7}, { 32, 0, 5, 8},
421 { 32, 0, 5, 10}, { 32, 0, 5, 11},
422 { 0, 0, 6, 14}, { 0, 0, 6, 17},
423 { 0, 0, 6, 20}, { 0, 0, 6, 23},
424 { 0, 0, 6, 26}, { 0, 0, 6, 29},
425 { 0, 0, 6, 32}, { 0, 16, 6,65539},
426 { 0, 15, 6,32771}, { 0, 14, 6,16387},
427 { 0, 13, 6, 8195}, { 0, 12, 6, 4099},
428 { 0, 11, 6, 2051}, { 0, 10, 6, 1027},
429}; /* ML_defaultDTable */
430
431
432static void ZSTD_buildSeqTable_rle(ZSTD_seqSymbol* dt, U32 baseValue, U8 nbAddBits)
433{
434 void* ptr = dt;
435 ZSTD_seqSymbol_header* const DTableH = (ZSTD_seqSymbol_header*)ptr;
436 ZSTD_seqSymbol* const cell = dt + 1;
437
438 DTableH->tableLog = 0;
439 DTableH->fastMode = 0;
440
441 cell->nbBits = 0;
442 cell->nextState = 0;
443 assert(nbAddBits < 255);
444 cell->nbAdditionalBits = nbAddBits;
445 cell->baseValue = baseValue;
446}
447
448
449/* ZSTD_buildFSETable() :
450 * generate FSE decoding table for one symbol (ll, ml or off)
451 * cannot fail if input is valid =>
452 * all inputs are presumed validated at this stage */
453FORCE_INLINE_TEMPLATE
454void ZSTD_buildFSETable_body(ZSTD_seqSymbol* dt,
455 const short* normalizedCounter, unsigned maxSymbolValue,
456 const U32* baseValue, const U8* nbAdditionalBits,
457 unsigned tableLog, void* wksp, size_t wkspSize)
458{
459 ZSTD_seqSymbol* const tableDecode = dt+1;
460 U32 const maxSV1 = maxSymbolValue + 1;
461 U32 const tableSize = 1 << tableLog;
462
463 U16* symbolNext = (U16*)wksp;
464 BYTE* spread = (BYTE*)(symbolNext + MaxSeq + 1);
465 U32 highThreshold = tableSize - 1;
466
467
468 /* Sanity Checks */
469 assert(maxSymbolValue <= MaxSeq);
470 assert(tableLog <= MaxFSELog);
471 assert(wkspSize >= ZSTD_BUILD_FSE_TABLE_WKSP_SIZE);
472 (void)wkspSize;
473 /* Init, lay down lowprob symbols */
474 { ZSTD_seqSymbol_header DTableH;
475 DTableH.tableLog = tableLog;
476 DTableH.fastMode = 1;
477 { S16 const largeLimit= (S16)(1 << (tableLog-1));
478 U32 s;
479 for (s=0; s<maxSV1; s++) {
480 if (normalizedCounter[s]==-1) {
481 tableDecode[highThreshold--].baseValue = s;
482 symbolNext[s] = 1;
483 } else {
484 if (normalizedCounter[s] >= largeLimit) DTableH.fastMode=0;
485 assert(normalizedCounter[s]>=0);
486 symbolNext[s] = (U16)normalizedCounter[s];
487 } } }
488 ZSTD_memcpy(dt, &DTableH, sizeof(DTableH));
489 }
490
491 /* Spread symbols */
492 assert(tableSize <= 512);
493 /* Specialized symbol spreading for the case when there are
494 * no low probability (-1 count) symbols. When compressing
495 * small blocks we avoid low probability symbols to hit this
496 * case, since header decoding speed matters more.
497 */
498 if (highThreshold == tableSize - 1) {
499 size_t const tableMask = tableSize-1;
500 size_t const step = FSE_TABLESTEP(tableSize);
501 /* First lay down the symbols in order.
502 * We use a uint64_t to lay down 8 bytes at a time. This reduces branch
503 * misses since small blocks generally have small table logs, so nearly
504 * all symbols have counts <= 8. We ensure we have 8 bytes at the end of
505 * our buffer to handle the over-write.
506 */
507 {
508 U64 const add = 0x0101010101010101ull;
509 size_t pos = 0;
510 U64 sv = 0;
511 U32 s;
512 for (s=0; s<maxSV1; ++s, sv += add) {
513 int i;
514 int const n = normalizedCounter[s];
515 MEM_write64(spread + pos, sv);
516 for (i = 8; i < n; i += 8) {
517 MEM_write64(spread + pos + i, sv);
518 }
519 assert(n>=0);
520 pos += (size_t)n;
521 }
522 }
523 /* Now we spread those positions across the table.
524 * The benefit of doing it in two stages is that we avoid the
525 * variable size inner loop, which caused lots of branch misses.
526 * Now we can run through all the positions without any branch misses.
527 * We unroll the loop twice, since that is what empirically worked best.
528 */
529 {
530 size_t position = 0;
531 size_t s;
532 size_t const unroll = 2;
533 assert(tableSize % unroll == 0); /* FSE_MIN_TABLELOG is 5 */
534 for (s = 0; s < (size_t)tableSize; s += unroll) {
535 size_t u;
536 for (u = 0; u < unroll; ++u) {
537 size_t const uPosition = (position + (u * step)) & tableMask;
538 tableDecode[uPosition].baseValue = spread[s + u];
539 }
540 position = (position + (unroll * step)) & tableMask;
541 }
542 assert(position == 0);
543 }
544 } else {
545 U32 const tableMask = tableSize-1;
546 U32 const step = FSE_TABLESTEP(tableSize);
547 U32 s, position = 0;
548 for (s=0; s<maxSV1; s++) {
549 int i;
550 int const n = normalizedCounter[s];
551 for (i=0; i<n; i++) {
552 tableDecode[position].baseValue = s;
553 position = (position + step) & tableMask;
554 while (UNLIKELY(position > highThreshold)) position = (position + step) & tableMask; /* lowprob area */
555 } }
556 assert(position == 0); /* position must reach all cells once, otherwise normalizedCounter is incorrect */
557 }
558
559 /* Build Decoding table */
560 {
561 U32 u;
562 for (u=0; u<tableSize; u++) {
563 U32 const symbol = tableDecode[u].baseValue;
564 U32 const nextState = symbolNext[symbol]++;
565 tableDecode[u].nbBits = (BYTE) (tableLog - ZSTD_highbit32(nextState) );
566 tableDecode[u].nextState = (U16) ( (nextState << tableDecode[u].nbBits) - tableSize);
567 assert(nbAdditionalBits[symbol] < 255);
568 tableDecode[u].nbAdditionalBits = nbAdditionalBits[symbol];
569 tableDecode[u].baseValue = baseValue[symbol];
570 }
571 }
572}
573
574/* Avoids the FORCE_INLINE of the _body() function. */
575static void ZSTD_buildFSETable_body_default(ZSTD_seqSymbol* dt,
576 const short* normalizedCounter, unsigned maxSymbolValue,
577 const U32* baseValue, const U8* nbAdditionalBits,
578 unsigned tableLog, void* wksp, size_t wkspSize)
579{
580 ZSTD_buildFSETable_body(dt, normalizedCounter, maxSymbolValue,
581 baseValue, nbAdditionalBits, tableLog, wksp, wkspSize);
582}
583
584#if DYNAMIC_BMI2
585BMI2_TARGET_ATTRIBUTE static void ZSTD_buildFSETable_body_bmi2(ZSTD_seqSymbol* dt,
586 const short* normalizedCounter, unsigned maxSymbolValue,
587 const U32* baseValue, const U8* nbAdditionalBits,
588 unsigned tableLog, void* wksp, size_t wkspSize)
589{
590 ZSTD_buildFSETable_body(dt, normalizedCounter, maxSymbolValue,
591 baseValue, nbAdditionalBits, tableLog, wksp, wkspSize);
592}
593#endif
594
595void ZSTD_buildFSETable(ZSTD_seqSymbol* dt,
596 const short* normalizedCounter, unsigned maxSymbolValue,
597 const U32* baseValue, const U8* nbAdditionalBits,
598 unsigned tableLog, void* wksp, size_t wkspSize, int bmi2)
599{
600#if DYNAMIC_BMI2
601 if (bmi2) {
602 ZSTD_buildFSETable_body_bmi2(dt, normalizedCounter, maxSymbolValue,
603 baseValue, nbAdditionalBits, tableLog, wksp, wkspSize);
604 return;
605 }
606#endif
607 (void)bmi2;
608 ZSTD_buildFSETable_body_default(dt, normalizedCounter, maxSymbolValue,
609 baseValue, nbAdditionalBits, tableLog, wksp, wkspSize);
610}
611
612
613/*! ZSTD_buildSeqTable() :
614 * @return : nb bytes read from src,
615 * or an error code if it fails */
616static size_t ZSTD_buildSeqTable(ZSTD_seqSymbol* DTableSpace, const ZSTD_seqSymbol** DTablePtr,
617 symbolEncodingType_e type, unsigned max, U32 maxLog,
618 const void* src, size_t srcSize,
619 const U32* baseValue, const U8* nbAdditionalBits,
620 const ZSTD_seqSymbol* defaultTable, U32 flagRepeatTable,
621 int ddictIsCold, int nbSeq, U32* wksp, size_t wkspSize,
622 int bmi2)
623{
624 switch(type)
625 {
626 case set_rle :
627 RETURN_ERROR_IF(!srcSize, srcSize_wrong, "");
628 RETURN_ERROR_IF((*(const BYTE*)src) > max, corruption_detected, "");
629 { U32 const symbol = *(const BYTE*)src;
630 U32 const baseline = baseValue[symbol];
631 U8 const nbBits = nbAdditionalBits[symbol];
632 ZSTD_buildSeqTable_rle(DTableSpace, baseline, nbBits);
633 }
634 *DTablePtr = DTableSpace;
635 return 1;
636 case set_basic :
637 *DTablePtr = defaultTable;
638 return 0;
639 case set_repeat:
640 RETURN_ERROR_IF(!flagRepeatTable, corruption_detected, "");
641 /* prefetch FSE table if used */
642 if (ddictIsCold && (nbSeq > 24 /* heuristic */)) {
643 const void* const pStart = *DTablePtr;
644 size_t const pSize = sizeof(ZSTD_seqSymbol) * (SEQSYMBOL_TABLE_SIZE(maxLog));
645 PREFETCH_AREA(pStart, pSize);
646 }
647 return 0;
648 case set_compressed :
649 { unsigned tableLog;
650 S16 norm[MaxSeq+1];
651 size_t const headerSize = FSE_readNCount(norm, &max, &tableLog, src, srcSize);
652 RETURN_ERROR_IF(FSE_isError(headerSize), corruption_detected, "");
653 RETURN_ERROR_IF(tableLog > maxLog, corruption_detected, "");
654 ZSTD_buildFSETable(DTableSpace, norm, max, baseValue, nbAdditionalBits, tableLog, wksp, wkspSize, bmi2);
655 *DTablePtr = DTableSpace;
656 return headerSize;
657 }
658 default :
659 assert(0);
660 RETURN_ERROR(GENERIC, "impossible");
661 }
662}
663
664size_t ZSTD_decodeSeqHeaders(ZSTD_DCtx* dctx, int* nbSeqPtr,
665 const void* src, size_t srcSize)
666{
667 const BYTE* const istart = (const BYTE*)src;
668 const BYTE* const iend = istart + srcSize;
669 const BYTE* ip = istart;
670 int nbSeq;
671 DEBUGLOG(5, "ZSTD_decodeSeqHeaders");
672
673 /* check */
674 RETURN_ERROR_IF(srcSize < MIN_SEQUENCES_SIZE, srcSize_wrong, "");
675
676 /* SeqHead */
677 nbSeq = *ip++;
678 if (!nbSeq) {
679 *nbSeqPtr=0;
680 RETURN_ERROR_IF(srcSize != 1, srcSize_wrong, "");
681 return 1;
682 }
683 if (nbSeq > 0x7F) {
684 if (nbSeq == 0xFF) {
685 RETURN_ERROR_IF(ip+2 > iend, srcSize_wrong, "");
686 nbSeq = MEM_readLE16(ip) + LONGNBSEQ;
687 ip+=2;
688 } else {
689 RETURN_ERROR_IF(ip >= iend, srcSize_wrong, "");
690 nbSeq = ((nbSeq-0x80)<<8) + *ip++;
691 }
692 }
693 *nbSeqPtr = nbSeq;
694
695 /* FSE table descriptors */
696 RETURN_ERROR_IF(ip+1 > iend, srcSize_wrong, ""); /* minimum possible size: 1 byte for symbol encoding types */
697 { symbolEncodingType_e const LLtype = (symbolEncodingType_e)(*ip >> 6);
698 symbolEncodingType_e const OFtype = (symbolEncodingType_e)((*ip >> 4) & 3);
699 symbolEncodingType_e const MLtype = (symbolEncodingType_e)((*ip >> 2) & 3);
700 ip++;
701
702 /* Build DTables */
703 { size_t const llhSize = ZSTD_buildSeqTable(dctx->entropy.LLTable, &dctx->LLTptr,
704 LLtype, MaxLL, LLFSELog,
705 ip, iend-ip,
706 LL_base, LL_bits,
707 LL_defaultDTable, dctx->fseEntropy,
708 dctx->ddictIsCold, nbSeq,
709 dctx->workspace, sizeof(dctx->workspace),
710 ZSTD_DCtx_get_bmi2(dctx));
711 RETURN_ERROR_IF(ZSTD_isError(llhSize), corruption_detected, "ZSTD_buildSeqTable failed");
712 ip += llhSize;
713 }
714
715 { size_t const ofhSize = ZSTD_buildSeqTable(dctx->entropy.OFTable, &dctx->OFTptr,
716 OFtype, MaxOff, OffFSELog,
717 ip, iend-ip,
718 OF_base, OF_bits,
719 OF_defaultDTable, dctx->fseEntropy,
720 dctx->ddictIsCold, nbSeq,
721 dctx->workspace, sizeof(dctx->workspace),
722 ZSTD_DCtx_get_bmi2(dctx));
723 RETURN_ERROR_IF(ZSTD_isError(ofhSize), corruption_detected, "ZSTD_buildSeqTable failed");
724 ip += ofhSize;
725 }
726
727 { size_t const mlhSize = ZSTD_buildSeqTable(dctx->entropy.MLTable, &dctx->MLTptr,
728 MLtype, MaxML, MLFSELog,
729 ip, iend-ip,
730 ML_base, ML_bits,
731 ML_defaultDTable, dctx->fseEntropy,
732 dctx->ddictIsCold, nbSeq,
733 dctx->workspace, sizeof(dctx->workspace),
734 ZSTD_DCtx_get_bmi2(dctx));
735 RETURN_ERROR_IF(ZSTD_isError(mlhSize), corruption_detected, "ZSTD_buildSeqTable failed");
736 ip += mlhSize;
737 }
738 }
739
740 return ip-istart;
741}
742
743
744typedef struct {
745 size_t litLength;
746 size_t matchLength;
747 size_t offset;
748} seq_t;
749
750typedef struct {
751 size_t state;
752 const ZSTD_seqSymbol* table;
753} ZSTD_fseState;
754
755typedef struct {
756 BIT_DStream_t DStream;
757 ZSTD_fseState stateLL;
758 ZSTD_fseState stateOffb;
759 ZSTD_fseState stateML;
760 size_t prevOffset[ZSTD_REP_NUM];
761} seqState_t;
762
763/*! ZSTD_overlapCopy8() :
764 * Copies 8 bytes from ip to op and updates op and ip where ip <= op.
765 * If the offset is < 8 then the offset is spread to at least 8 bytes.
766 *
767 * Precondition: *ip <= *op
768 * Postcondition: *op - *op >= 8
769 */
770HINT_INLINE void ZSTD_overlapCopy8(BYTE** op, BYTE const** ip, size_t offset) {
771 assert(*ip <= *op);
772 if (offset < 8) {
773 /* close range match, overlap */
774 static const U32 dec32table[] = { 0, 1, 2, 1, 4, 4, 4, 4 }; /* added */
775 static const int dec64table[] = { 8, 8, 8, 7, 8, 9,10,11 }; /* subtracted */
776 int const sub2 = dec64table[offset];
777 (*op)[0] = (*ip)[0];
778 (*op)[1] = (*ip)[1];
779 (*op)[2] = (*ip)[2];
780 (*op)[3] = (*ip)[3];
781 *ip += dec32table[offset];
782 ZSTD_copy4(*op+4, *ip);
783 *ip -= sub2;
784 } else {
785 ZSTD_copy8(*op, *ip);
786 }
787 *ip += 8;
788 *op += 8;
789 assert(*op - *ip >= 8);
790}
791
792/*! ZSTD_safecopy() :
793 * Specialized version of memcpy() that is allowed to READ up to WILDCOPY_OVERLENGTH past the input buffer
794 * and write up to 16 bytes past oend_w (op >= oend_w is allowed).
795 * This function is only called in the uncommon case where the sequence is near the end of the block. It
796 * should be fast for a single long sequence, but can be slow for several short sequences.
797 *
798 * @param ovtype controls the overlap detection
799 * - ZSTD_no_overlap: The source and destination are guaranteed to be at least WILDCOPY_VECLEN bytes apart.
800 * - ZSTD_overlap_src_before_dst: The src and dst may overlap and may be any distance apart.
801 * The src buffer must be before the dst buffer.
802 */
803static void ZSTD_safecopy(BYTE* op, const BYTE* const oend_w, BYTE const* ip, ptrdiff_t length, ZSTD_overlap_e ovtype) {
804 ptrdiff_t const diff = op - ip;
805 BYTE* const oend = op + length;
806
807 assert((ovtype == ZSTD_no_overlap && (diff <= -8 || diff >= 8 || op >= oend_w)) ||
808 (ovtype == ZSTD_overlap_src_before_dst && diff >= 0));
809
810 if (length < 8) {
811 /* Handle short lengths. */
812 while (op < oend) *op++ = *ip++;
813 return;
814 }
815 if (ovtype == ZSTD_overlap_src_before_dst) {
816 /* Copy 8 bytes and ensure the offset >= 8 when there can be overlap. */
817 assert(length >= 8);
818 ZSTD_overlapCopy8(&op, &ip, diff);
819 length -= 8;
820 assert(op - ip >= 8);
821 assert(op <= oend);
822 }
823
824 if (oend <= oend_w) {
825 /* No risk of overwrite. */
826 ZSTD_wildcopy(op, ip, length, ovtype);
827 return;
828 }
829 if (op <= oend_w) {
830 /* Wildcopy until we get close to the end. */
831 assert(oend > oend_w);
832 ZSTD_wildcopy(op, ip, oend_w - op, ovtype);
833 ip += oend_w - op;
834 op += oend_w - op;
835 }
836 /* Handle the leftovers. */
837 while (op < oend) *op++ = *ip++;
838}
839
840/* ZSTD_safecopyDstBeforeSrc():
841 * This version allows overlap with dst before src, or handles the non-overlap case with dst after src
842 * Kept separate from more common ZSTD_safecopy case to avoid performance impact to the safecopy common case */
843static void ZSTD_safecopyDstBeforeSrc(BYTE* op, BYTE const* ip, ptrdiff_t length) {
844 ptrdiff_t const diff = op - ip;
845 BYTE* const oend = op + length;
846
847 if (length < 8 || diff > -8) {
848 /* Handle short lengths, close overlaps, and dst not before src. */
849 while (op < oend) *op++ = *ip++;
850 return;
851 }
852
853 if (op <= oend - WILDCOPY_OVERLENGTH && diff < -WILDCOPY_VECLEN) {
854 ZSTD_wildcopy(op, ip, oend - WILDCOPY_OVERLENGTH - op, ZSTD_no_overlap);
855 ip += oend - WILDCOPY_OVERLENGTH - op;
856 op += oend - WILDCOPY_OVERLENGTH - op;
857 }
858
859 /* Handle the leftovers. */
860 while (op < oend) *op++ = *ip++;
861}
862
863/* ZSTD_execSequenceEnd():
864 * This version handles cases that are near the end of the output buffer. It requires
865 * more careful checks to make sure there is no overflow. By separating out these hard
866 * and unlikely cases, we can speed up the common cases.
867 *
868 * NOTE: This function needs to be fast for a single long sequence, but doesn't need
869 * to be optimized for many small sequences, since those fall into ZSTD_execSequence().
870 */
871FORCE_NOINLINE
872size_t ZSTD_execSequenceEnd(BYTE* op,
873 BYTE* const oend, seq_t sequence,
874 const BYTE** litPtr, const BYTE* const litLimit,
875 const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
876{
877 BYTE* const oLitEnd = op + sequence.litLength;
878 size_t const sequenceLength = sequence.litLength + sequence.matchLength;
879 const BYTE* const iLitEnd = *litPtr + sequence.litLength;
880 const BYTE* match = oLitEnd - sequence.offset;
881 BYTE* const oend_w = oend - WILDCOPY_OVERLENGTH;
882
883 /* bounds checks : careful of address space overflow in 32-bit mode */
884 RETURN_ERROR_IF(sequenceLength > (size_t)(oend - op), dstSize_tooSmall, "last match must fit within dstBuffer");
885 RETURN_ERROR_IF(sequence.litLength > (size_t)(litLimit - *litPtr), corruption_detected, "try to read beyond literal buffer");
886 assert(op < op + sequenceLength);
887 assert(oLitEnd < op + sequenceLength);
888
889 /* copy literals */
890 ZSTD_safecopy(op, oend_w, *litPtr, sequence.litLength, ZSTD_no_overlap);
891 op = oLitEnd;
892 *litPtr = iLitEnd;
893
894 /* copy Match */
895 if (sequence.offset > (size_t)(oLitEnd - prefixStart)) {
896 /* offset beyond prefix */
897 RETURN_ERROR_IF(sequence.offset > (size_t)(oLitEnd - virtualStart), corruption_detected, "");
898 match = dictEnd - (prefixStart - match);
899 if (match + sequence.matchLength <= dictEnd) {
900 ZSTD_memmove(oLitEnd, match, sequence.matchLength);
901 return sequenceLength;
902 }
903 /* span extDict & currentPrefixSegment */
904 { size_t const length1 = dictEnd - match;
905 ZSTD_memmove(oLitEnd, match, length1);
906 op = oLitEnd + length1;
907 sequence.matchLength -= length1;
908 match = prefixStart;
909 }
910 }
911 ZSTD_safecopy(op, oend_w, match, sequence.matchLength, ZSTD_overlap_src_before_dst);
912 return sequenceLength;
913}
914
915/* ZSTD_execSequenceEndSplitLitBuffer():
916 * This version is intended to be used during instances where the litBuffer is still split. It is kept separate to avoid performance impact for the good case.
917 */
918FORCE_NOINLINE
919size_t ZSTD_execSequenceEndSplitLitBuffer(BYTE* op,
920 BYTE* const oend, const BYTE* const oend_w, seq_t sequence,
921 const BYTE** litPtr, const BYTE* const litLimit,
922 const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
923{
924 BYTE* const oLitEnd = op + sequence.litLength;
925 size_t const sequenceLength = sequence.litLength + sequence.matchLength;
926 const BYTE* const iLitEnd = *litPtr + sequence.litLength;
927 const BYTE* match = oLitEnd - sequence.offset;
928
929
930 /* bounds checks : careful of address space overflow in 32-bit mode */
931 RETURN_ERROR_IF(sequenceLength > (size_t)(oend - op), dstSize_tooSmall, "last match must fit within dstBuffer");
932 RETURN_ERROR_IF(sequence.litLength > (size_t)(litLimit - *litPtr), corruption_detected, "try to read beyond literal buffer");
933 assert(op < op + sequenceLength);
934 assert(oLitEnd < op + sequenceLength);
935
936 /* copy literals */
937 RETURN_ERROR_IF(op > *litPtr && op < *litPtr + sequence.litLength, dstSize_tooSmall, "output should not catch up to and overwrite literal buffer");
938 ZSTD_safecopyDstBeforeSrc(op, *litPtr, sequence.litLength);
939 op = oLitEnd;
940 *litPtr = iLitEnd;
941
942 /* copy Match */
943 if (sequence.offset > (size_t)(oLitEnd - prefixStart)) {
944 /* offset beyond prefix */
945 RETURN_ERROR_IF(sequence.offset > (size_t)(oLitEnd - virtualStart), corruption_detected, "");
946 match = dictEnd - (prefixStart - match);
947 if (match + sequence.matchLength <= dictEnd) {
948 ZSTD_memmove(oLitEnd, match, sequence.matchLength);
949 return sequenceLength;
950 }
951 /* span extDict & currentPrefixSegment */
952 { size_t const length1 = dictEnd - match;
953 ZSTD_memmove(oLitEnd, match, length1);
954 op = oLitEnd + length1;
955 sequence.matchLength -= length1;
956 match = prefixStart;
957 }
958 }
959 ZSTD_safecopy(op, oend_w, match, sequence.matchLength, ZSTD_overlap_src_before_dst);
960 return sequenceLength;
961}
962
963HINT_INLINE
964size_t ZSTD_execSequence(BYTE* op,
965 BYTE* const oend, seq_t sequence,
966 const BYTE** litPtr, const BYTE* const litLimit,
967 const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
968{
969 BYTE* const oLitEnd = op + sequence.litLength;
970 size_t const sequenceLength = sequence.litLength + sequence.matchLength;
971 BYTE* const oMatchEnd = op + sequenceLength; /* risk : address space overflow (32-bits) */
972 BYTE* const oend_w = oend - WILDCOPY_OVERLENGTH; /* risk : address space underflow on oend=NULL */
973 const BYTE* const iLitEnd = *litPtr + sequence.litLength;
974 const BYTE* match = oLitEnd - sequence.offset;
975
976 assert(op != NULL /* Precondition */);
977 assert(oend_w < oend /* No underflow */);
978
979#if defined(__aarch64__)
980 /* prefetch sequence starting from match that will be used for copy later */
981 PREFETCH_L1(match);
982#endif
983 /* Handle edge cases in a slow path:
984 * - Read beyond end of literals
985 * - Match end is within WILDCOPY_OVERLIMIT of oend
986 * - 32-bit mode and the match length overflows
987 */
988 if (UNLIKELY(
989 iLitEnd > litLimit ||
990 oMatchEnd > oend_w ||
991 (MEM_32bits() && (size_t)(oend - op) < sequenceLength + WILDCOPY_OVERLENGTH)))
992 return ZSTD_execSequenceEnd(op, oend, sequence, litPtr, litLimit, prefixStart, virtualStart, dictEnd);
993
994 /* Assumptions (everything else goes into ZSTD_execSequenceEnd()) */
995 assert(op <= oLitEnd /* No overflow */);
996 assert(oLitEnd < oMatchEnd /* Non-zero match & no overflow */);
997 assert(oMatchEnd <= oend /* No underflow */);
998 assert(iLitEnd <= litLimit /* Literal length is in bounds */);
999 assert(oLitEnd <= oend_w /* Can wildcopy literals */);
1000 assert(oMatchEnd <= oend_w /* Can wildcopy matches */);
1001
1002 /* Copy Literals:
1003 * Split out litLength <= 16 since it is nearly always true. +1.6% on gcc-9.
1004 * We likely don't need the full 32-byte wildcopy.
1005 */
1006 assert(WILDCOPY_OVERLENGTH >= 16);
1007 ZSTD_copy16(op, (*litPtr));
1008 if (UNLIKELY(sequence.litLength > 16)) {
1009 ZSTD_wildcopy(op + 16, (*litPtr) + 16, sequence.litLength - 16, ZSTD_no_overlap);
1010 }
1011 op = oLitEnd;
1012 *litPtr = iLitEnd; /* update for next sequence */
1013
1014 /* Copy Match */
1015 if (sequence.offset > (size_t)(oLitEnd - prefixStart)) {
1016 /* offset beyond prefix -> go into extDict */
1017 RETURN_ERROR_IF(UNLIKELY(sequence.offset > (size_t)(oLitEnd - virtualStart)), corruption_detected, "");
1018 match = dictEnd + (match - prefixStart);
1019 if (match + sequence.matchLength <= dictEnd) {
1020 ZSTD_memmove(oLitEnd, match, sequence.matchLength);
1021 return sequenceLength;
1022 }
1023 /* span extDict & currentPrefixSegment */
1024 { size_t const length1 = dictEnd - match;
1025 ZSTD_memmove(oLitEnd, match, length1);
1026 op = oLitEnd + length1;
1027 sequence.matchLength -= length1;
1028 match = prefixStart;
1029 }
1030 }
1031 /* Match within prefix of 1 or more bytes */
1032 assert(op <= oMatchEnd);
1033 assert(oMatchEnd <= oend_w);
1034 assert(match >= prefixStart);
1035 assert(sequence.matchLength >= 1);
1036
1037 /* Nearly all offsets are >= WILDCOPY_VECLEN bytes, which means we can use wildcopy
1038 * without overlap checking.
1039 */
1040 if (LIKELY(sequence.offset >= WILDCOPY_VECLEN)) {
1041 /* We bet on a full wildcopy for matches, since we expect matches to be
1042 * longer than literals (in general). In silesia, ~10% of matches are longer
1043 * than 16 bytes.
1044 */
1045 ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength, ZSTD_no_overlap);
1046 return sequenceLength;
1047 }
1048 assert(sequence.offset < WILDCOPY_VECLEN);
1049
1050 /* Copy 8 bytes and spread the offset to be >= 8. */
1051 ZSTD_overlapCopy8(&op, &match, sequence.offset);
1052
1053 /* If the match length is > 8 bytes, then continue with the wildcopy. */
1054 if (sequence.matchLength > 8) {
1055 assert(op < oMatchEnd);
1056 ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength - 8, ZSTD_overlap_src_before_dst);
1057 }
1058 return sequenceLength;
1059}
1060
1061HINT_INLINE
1062size_t ZSTD_execSequenceSplitLitBuffer(BYTE* op,
1063 BYTE* const oend, const BYTE* const oend_w, seq_t sequence,
1064 const BYTE** litPtr, const BYTE* const litLimit,
1065 const BYTE* const prefixStart, const BYTE* const virtualStart, const BYTE* const dictEnd)
1066{
1067 BYTE* const oLitEnd = op + sequence.litLength;
1068 size_t const sequenceLength = sequence.litLength + sequence.matchLength;
1069 BYTE* const oMatchEnd = op + sequenceLength; /* risk : address space overflow (32-bits) */
1070 const BYTE* const iLitEnd = *litPtr + sequence.litLength;
1071 const BYTE* match = oLitEnd - sequence.offset;
1072
1073 assert(op != NULL /* Precondition */);
1074 assert(oend_w < oend /* No underflow */);
1075 /* Handle edge cases in a slow path:
1076 * - Read beyond end of literals
1077 * - Match end is within WILDCOPY_OVERLIMIT of oend
1078 * - 32-bit mode and the match length overflows
1079 */
1080 if (UNLIKELY(
1081 iLitEnd > litLimit ||
1082 oMatchEnd > oend_w ||
1083 (MEM_32bits() && (size_t)(oend - op) < sequenceLength + WILDCOPY_OVERLENGTH)))
1084 return ZSTD_execSequenceEndSplitLitBuffer(op, oend, oend_w, sequence, litPtr, litLimit, prefixStart, virtualStart, dictEnd);
1085
1086 /* Assumptions (everything else goes into ZSTD_execSequenceEnd()) */
1087 assert(op <= oLitEnd /* No overflow */);
1088 assert(oLitEnd < oMatchEnd /* Non-zero match & no overflow */);
1089 assert(oMatchEnd <= oend /* No underflow */);
1090 assert(iLitEnd <= litLimit /* Literal length is in bounds */);
1091 assert(oLitEnd <= oend_w /* Can wildcopy literals */);
1092 assert(oMatchEnd <= oend_w /* Can wildcopy matches */);
1093
1094 /* Copy Literals:
1095 * Split out litLength <= 16 since it is nearly always true. +1.6% on gcc-9.
1096 * We likely don't need the full 32-byte wildcopy.
1097 */
1098 assert(WILDCOPY_OVERLENGTH >= 16);
1099 ZSTD_copy16(op, (*litPtr));
1100 if (UNLIKELY(sequence.litLength > 16)) {
1101 ZSTD_wildcopy(op+16, (*litPtr)+16, sequence.litLength-16, ZSTD_no_overlap);
1102 }
1103 op = oLitEnd;
1104 *litPtr = iLitEnd; /* update for next sequence */
1105
1106 /* Copy Match */
1107 if (sequence.offset > (size_t)(oLitEnd - prefixStart)) {
1108 /* offset beyond prefix -> go into extDict */
1109 RETURN_ERROR_IF(UNLIKELY(sequence.offset > (size_t)(oLitEnd - virtualStart)), corruption_detected, "");
1110 match = dictEnd + (match - prefixStart);
1111 if (match + sequence.matchLength <= dictEnd) {
1112 ZSTD_memmove(oLitEnd, match, sequence.matchLength);
1113 return sequenceLength;
1114 }
1115 /* span extDict & currentPrefixSegment */
1116 { size_t const length1 = dictEnd - match;
1117 ZSTD_memmove(oLitEnd, match, length1);
1118 op = oLitEnd + length1;
1119 sequence.matchLength -= length1;
1120 match = prefixStart;
1121 } }
1122 /* Match within prefix of 1 or more bytes */
1123 assert(op <= oMatchEnd);
1124 assert(oMatchEnd <= oend_w);
1125 assert(match >= prefixStart);
1126 assert(sequence.matchLength >= 1);
1127
1128 /* Nearly all offsets are >= WILDCOPY_VECLEN bytes, which means we can use wildcopy
1129 * without overlap checking.
1130 */
1131 if (LIKELY(sequence.offset >= WILDCOPY_VECLEN)) {
1132 /* We bet on a full wildcopy for matches, since we expect matches to be
1133 * longer than literals (in general). In silesia, ~10% of matches are longer
1134 * than 16 bytes.
1135 */
1136 ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength, ZSTD_no_overlap);
1137 return sequenceLength;
1138 }
1139 assert(sequence.offset < WILDCOPY_VECLEN);
1140
1141 /* Copy 8 bytes and spread the offset to be >= 8. */
1142 ZSTD_overlapCopy8(&op, &match, sequence.offset);
1143
1144 /* If the match length is > 8 bytes, then continue with the wildcopy. */
1145 if (sequence.matchLength > 8) {
1146 assert(op < oMatchEnd);
1147 ZSTD_wildcopy(op, match, (ptrdiff_t)sequence.matchLength-8, ZSTD_overlap_src_before_dst);
1148 }
1149 return sequenceLength;
1150}
1151
1152
1153static void
1154ZSTD_initFseState(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, const ZSTD_seqSymbol* dt)
1155{
1156 const void* ptr = dt;
1157 const ZSTD_seqSymbol_header* const DTableH = (const ZSTD_seqSymbol_header*)ptr;
1158 DStatePtr->state = BIT_readBits(bitD, DTableH->tableLog);
1159 DEBUGLOG(6, "ZSTD_initFseState : val=%u using %u bits",
1160 (U32)DStatePtr->state, DTableH->tableLog);
1161 BIT_reloadDStream(bitD);
1162 DStatePtr->table = dt + 1;
1163}
1164
1165FORCE_INLINE_TEMPLATE void
1166ZSTD_updateFseStateWithDInfo(ZSTD_fseState* DStatePtr, BIT_DStream_t* bitD, U16 nextState, U32 nbBits)
1167{
1168 size_t const lowBits = BIT_readBits(bitD, nbBits);
1169 DStatePtr->state = nextState + lowBits;
1170}
1171
1172/* We need to add at most (ZSTD_WINDOWLOG_MAX_32 - 1) bits to read the maximum
1173 * offset bits. But we can only read at most STREAM_ACCUMULATOR_MIN_32
1174 * bits before reloading. This value is the maximum number of bytes we read
1175 * after reloading when we are decoding long offsets.
1176 */
1177#define LONG_OFFSETS_MAX_EXTRA_BITS_32 \
1178 (ZSTD_WINDOWLOG_MAX_32 > STREAM_ACCUMULATOR_MIN_32 \
1179 ? ZSTD_WINDOWLOG_MAX_32 - STREAM_ACCUMULATOR_MIN_32 \
1180 : 0)
1181
1182typedef enum { ZSTD_lo_isRegularOffset, ZSTD_lo_isLongOffset=1 } ZSTD_longOffset_e;
1183
1184FORCE_INLINE_TEMPLATE seq_t
1185ZSTD_decodeSequence(seqState_t* seqState, const ZSTD_longOffset_e longOffsets)
1186{
1187 seq_t seq;
1188 /*
1189 * ZSTD_seqSymbol is a structure with a total of 64 bits wide. So it can be
1190 * loaded in one operation and extracted its fields by simply shifting or
1191 * bit-extracting on aarch64.
1192 * GCC doesn't recognize this and generates more unnecessary ldr/ldrb/ldrh
1193 * operations that cause performance drop. This can be avoided by using this
1194 * ZSTD_memcpy hack.
1195 */
1196#if defined(__aarch64__) && (defined(__GNUC__) && !defined(__clang__))
1197 ZSTD_seqSymbol llDInfoS, mlDInfoS, ofDInfoS;
1198 ZSTD_seqSymbol* const llDInfo = &llDInfoS;
1199 ZSTD_seqSymbol* const mlDInfo = &mlDInfoS;
1200 ZSTD_seqSymbol* const ofDInfo = &ofDInfoS;
1201 ZSTD_memcpy(llDInfo, seqState->stateLL.table + seqState->stateLL.state, sizeof(ZSTD_seqSymbol));
1202 ZSTD_memcpy(mlDInfo, seqState->stateML.table + seqState->stateML.state, sizeof(ZSTD_seqSymbol));
1203 ZSTD_memcpy(ofDInfo, seqState->stateOffb.table + seqState->stateOffb.state, sizeof(ZSTD_seqSymbol));
1204#else
1205 const ZSTD_seqSymbol* const llDInfo = seqState->stateLL.table + seqState->stateLL.state;
1206 const ZSTD_seqSymbol* const mlDInfo = seqState->stateML.table + seqState->stateML.state;
1207 const ZSTD_seqSymbol* const ofDInfo = seqState->stateOffb.table + seqState->stateOffb.state;
1208#endif
1209 seq.matchLength = mlDInfo->baseValue;
1210 seq.litLength = llDInfo->baseValue;
1211 { U32 const ofBase = ofDInfo->baseValue;
1212 BYTE const llBits = llDInfo->nbAdditionalBits;
1213 BYTE const mlBits = mlDInfo->nbAdditionalBits;
1214 BYTE const ofBits = ofDInfo->nbAdditionalBits;
1215 BYTE const totalBits = llBits+mlBits+ofBits;
1216
1217 U16 const llNext = llDInfo->nextState;
1218 U16 const mlNext = mlDInfo->nextState;
1219 U16 const ofNext = ofDInfo->nextState;
1220 U32 const llnbBits = llDInfo->nbBits;
1221 U32 const mlnbBits = mlDInfo->nbBits;
1222 U32 const ofnbBits = ofDInfo->nbBits;
1223
1224 assert(llBits <= MaxLLBits);
1225 assert(mlBits <= MaxMLBits);
1226 assert(ofBits <= MaxOff);
1227 /*
1228 * As gcc has better branch and block analyzers, sometimes it is only
1229 * valuable to mark likeliness for clang, it gives around 3-4% of
1230 * performance.
1231 */
1232
1233 /* sequence */
1234 { size_t offset;
1235 if (ofBits > 1) {
1236 ZSTD_STATIC_ASSERT(ZSTD_lo_isLongOffset == 1);
1237 ZSTD_STATIC_ASSERT(LONG_OFFSETS_MAX_EXTRA_BITS_32 == 5);
1238 ZSTD_STATIC_ASSERT(STREAM_ACCUMULATOR_MIN_32 > LONG_OFFSETS_MAX_EXTRA_BITS_32);
1239 ZSTD_STATIC_ASSERT(STREAM_ACCUMULATOR_MIN_32 - LONG_OFFSETS_MAX_EXTRA_BITS_32 >= MaxMLBits);
1240 if (MEM_32bits() && longOffsets && (ofBits >= STREAM_ACCUMULATOR_MIN_32)) {
1241 /* Always read extra bits, this keeps the logic simple,
1242 * avoids branches, and avoids accidentally reading 0 bits.
1243 */
1244 U32 const extraBits = LONG_OFFSETS_MAX_EXTRA_BITS_32;
1245 offset = ofBase + (BIT_readBitsFast(&seqState->DStream, ofBits - extraBits) << extraBits);
1246 BIT_reloadDStream(&seqState->DStream);
1247 offset += BIT_readBitsFast(&seqState->DStream, extraBits);
1248 } else {
1249 offset = ofBase + BIT_readBitsFast(&seqState->DStream, ofBits/*>0*/); /* <= (ZSTD_WINDOWLOG_MAX-1) bits */
1250 if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream);
1251 }
1252 seqState->prevOffset[2] = seqState->prevOffset[1];
1253 seqState->prevOffset[1] = seqState->prevOffset[0];
1254 seqState->prevOffset[0] = offset;
1255 } else {
1256 U32 const ll0 = (llDInfo->baseValue == 0);
1257 if (LIKELY((ofBits == 0))) {
1258 offset = seqState->prevOffset[ll0];
1259 seqState->prevOffset[1] = seqState->prevOffset[!ll0];
1260 seqState->prevOffset[0] = offset;
1261 } else {
1262 offset = ofBase + ll0 + BIT_readBitsFast(&seqState->DStream, 1);
1263 { size_t temp = (offset==3) ? seqState->prevOffset[0] - 1 : seqState->prevOffset[offset];
1264 temp += !temp; /* 0 is not valid; input is corrupted; force offset to 1 */
1265 if (offset != 1) seqState->prevOffset[2] = seqState->prevOffset[1];
1266 seqState->prevOffset[1] = seqState->prevOffset[0];
1267 seqState->prevOffset[0] = offset = temp;
1268 } } }
1269 seq.offset = offset;
1270 }
1271
1272 if (mlBits > 0)
1273 seq.matchLength += BIT_readBitsFast(&seqState->DStream, mlBits/*>0*/);
1274
1275 if (MEM_32bits() && (mlBits+llBits >= STREAM_ACCUMULATOR_MIN_32-LONG_OFFSETS_MAX_EXTRA_BITS_32))
1276 BIT_reloadDStream(&seqState->DStream);
1277 if (MEM_64bits() && UNLIKELY(totalBits >= STREAM_ACCUMULATOR_MIN_64-(LLFSELog+MLFSELog+OffFSELog)))
1278 BIT_reloadDStream(&seqState->DStream);
1279 /* Ensure there are enough bits to read the rest of data in 64-bit mode. */
1280 ZSTD_STATIC_ASSERT(16+LLFSELog+MLFSELog+OffFSELog < STREAM_ACCUMULATOR_MIN_64);
1281
1282 if (llBits > 0)
1283 seq.litLength += BIT_readBitsFast(&seqState->DStream, llBits/*>0*/);
1284
1285 if (MEM_32bits())
1286 BIT_reloadDStream(&seqState->DStream);
1287
1288 DEBUGLOG(6, "seq: litL=%u, matchL=%u, offset=%u",
1289 (U32)seq.litLength, (U32)seq.matchLength, (U32)seq.offset);
1290
1291 ZSTD_updateFseStateWithDInfo(&seqState->stateLL, &seqState->DStream, llNext, llnbBits); /* <= 9 bits */
1292 ZSTD_updateFseStateWithDInfo(&seqState->stateML, &seqState->DStream, mlNext, mlnbBits); /* <= 9 bits */
1293 if (MEM_32bits()) BIT_reloadDStream(&seqState->DStream); /* <= 18 bits */
1294 ZSTD_updateFseStateWithDInfo(&seqState->stateOffb, &seqState->DStream, ofNext, ofnbBits); /* <= 8 bits */
1295 }
1296
1297 return seq;
1298}
1299
1300#ifdef FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION
1301MEM_STATIC int ZSTD_dictionaryIsActive(ZSTD_DCtx const* dctx, BYTE const* prefixStart, BYTE const* oLitEnd)
1302{
1303 size_t const windowSize = dctx->fParams.windowSize;
1304 /* No dictionary used. */
1305 if (dctx->dictContentEndForFuzzing == NULL) return 0;
1306 /* Dictionary is our prefix. */
1307 if (prefixStart == dctx->dictContentBeginForFuzzing) return 1;
1308 /* Dictionary is not our ext-dict. */
1309 if (dctx->dictEnd != dctx->dictContentEndForFuzzing) return 0;
1310 /* Dictionary is not within our window size. */
1311 if ((size_t)(oLitEnd - prefixStart) >= windowSize) return 0;
1312 /* Dictionary is active. */
1313 return 1;
1314}
1315
1316MEM_STATIC void ZSTD_assertValidSequence(
1317 ZSTD_DCtx const* dctx,
1318 BYTE const* op, BYTE const* oend,
1319 seq_t const seq,
1320 BYTE const* prefixStart, BYTE const* virtualStart)
1321{
1322#if DEBUGLEVEL >= 1
1323 size_t const windowSize = dctx->fParams.windowSize;
1324 size_t const sequenceSize = seq.litLength + seq.matchLength;
1325 BYTE const* const oLitEnd = op + seq.litLength;
1326 DEBUGLOG(6, "Checking sequence: litL=%u matchL=%u offset=%u",
1327 (U32)seq.litLength, (U32)seq.matchLength, (U32)seq.offset);
1328 assert(op <= oend);
1329 assert((size_t)(oend - op) >= sequenceSize);
1330 assert(sequenceSize <= ZSTD_BLOCKSIZE_MAX);
1331 if (ZSTD_dictionaryIsActive(dctx, prefixStart, oLitEnd)) {
1332 size_t const dictSize = (size_t)((char const*)dctx->dictContentEndForFuzzing - (char const*)dctx->dictContentBeginForFuzzing);
1333 /* Offset must be within the dictionary. */
1334 assert(seq.offset <= (size_t)(oLitEnd - virtualStart));
1335 assert(seq.offset <= windowSize + dictSize);
1336 } else {
1337 /* Offset must be within our window. */
1338 assert(seq.offset <= windowSize);
1339 }
1340#else
1341 (void)dctx, (void)op, (void)oend, (void)seq, (void)prefixStart, (void)virtualStart;
1342#endif
1343}
1344#endif
1345
1346#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
1347
1348
1349FORCE_INLINE_TEMPLATE size_t
1350DONT_VECTORIZE
1351ZSTD_decompressSequences_bodySplitLitBuffer( ZSTD_DCtx* dctx,
1352 void* dst, size_t maxDstSize,
1353 const void* seqStart, size_t seqSize, int nbSeq,
1354 const ZSTD_longOffset_e isLongOffset,
1355 const int frame)
1356{
1357 const BYTE* ip = (const BYTE*)seqStart;
1358 const BYTE* const iend = ip + seqSize;
1359 BYTE* const ostart = (BYTE*)dst;
1360 BYTE* const oend = ostart + maxDstSize;
1361 BYTE* op = ostart;
1362 const BYTE* litPtr = dctx->litPtr;
1363 const BYTE* litBufferEnd = dctx->litBufferEnd;
1364 const BYTE* const prefixStart = (const BYTE*) (dctx->prefixStart);
1365 const BYTE* const vBase = (const BYTE*) (dctx->virtualStart);
1366 const BYTE* const dictEnd = (const BYTE*) (dctx->dictEnd);
1367 DEBUGLOG(5, "ZSTD_decompressSequences_bodySplitLitBuffer");
1368 (void)frame;
1369
1370 /* Regen sequences */
1371 if (nbSeq) {
1372 seqState_t seqState;
1373 dctx->fseEntropy = 1;
1374 { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) seqState.prevOffset[i] = dctx->entropy.rep[i]; }
1375 RETURN_ERROR_IF(
1376 ERR_isError(BIT_initDStream(&seqState.DStream, ip, iend-ip)),
1377 corruption_detected, "");
1378 ZSTD_initFseState(&seqState.stateLL, &seqState.DStream, dctx->LLTptr);
1379 ZSTD_initFseState(&seqState.stateOffb, &seqState.DStream, dctx->OFTptr);
1380 ZSTD_initFseState(&seqState.stateML, &seqState.DStream, dctx->MLTptr);
1381 assert(dst != NULL);
1382
1383 ZSTD_STATIC_ASSERT(
1384 BIT_DStream_unfinished < BIT_DStream_completed &&
1385 BIT_DStream_endOfBuffer < BIT_DStream_completed &&
1386 BIT_DStream_completed < BIT_DStream_overflow);
1387
1388 /* decompress without overrunning litPtr begins */
1389 {
1390 seq_t sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
1391 /* Align the decompression loop to 32 + 16 bytes.
1392 *
1393 * zstd compiled with gcc-9 on an Intel i9-9900k shows 10% decompression
1394 * speed swings based on the alignment of the decompression loop. This
1395 * performance swing is caused by parts of the decompression loop falling
1396 * out of the DSB. The entire decompression loop should fit in the DSB,
1397 * when it can't we get much worse performance. You can measure if you've
1398 * hit the good case or the bad case with this perf command for some
1399 * compressed file test.zst:
1400 *
1401 * perf stat -e cycles -e instructions -e idq.all_dsb_cycles_any_uops \
1402 * -e idq.all_mite_cycles_any_uops -- ./zstd -tq test.zst
1403 *
1404 * If you see most cycles served out of the MITE you've hit the bad case.
1405 * If you see most cycles served out of the DSB you've hit the good case.
1406 * If it is pretty even then you may be in an okay case.
1407 *
1408 * This issue has been reproduced on the following CPUs:
1409 * - Kabylake: Macbook Pro (15-inch, 2019) 2.4 GHz Intel Core i9
1410 * Use Instruments->Counters to get DSB/MITE cycles.
1411 * I never got performance swings, but I was able to
1412 * go from the good case of mostly DSB to half of the
1413 * cycles served from MITE.
1414 * - Coffeelake: Intel i9-9900k
1415 * - Coffeelake: Intel i7-9700k
1416 *
1417 * I haven't been able to reproduce the instability or DSB misses on any
1418 * of the following CPUS:
1419 * - Haswell
1420 * - Broadwell: Intel(R) Xeon(R) CPU E5-2680 v4 @ 2.40GH
1421 * - Skylake
1422 *
1423 * Alignment is done for each of the three major decompression loops:
1424 * - ZSTD_decompressSequences_bodySplitLitBuffer - presplit section of the literal buffer
1425 * - ZSTD_decompressSequences_bodySplitLitBuffer - postsplit section of the literal buffer
1426 * - ZSTD_decompressSequences_body
1427 * Alignment choices are made to minimize large swings on bad cases and influence on performance
1428 * from changes external to this code, rather than to overoptimize on the current commit.
1429 *
1430 * If you are seeing performance stability this script can help test.
1431 * It tests on 4 commits in zstd where I saw performance change.
1432 *
1433 * https://gist.github.com/terrelln/9889fc06a423fd5ca6e99351564473f4
1434 */
1435#if defined(__GNUC__) && defined(__x86_64__)
1436 __asm__(".p2align 6");
1437# if __GNUC__ >= 7
1438 /* good for gcc-7, gcc-9, and gcc-11 */
1439 __asm__("nop");
1440 __asm__(".p2align 5");
1441 __asm__("nop");
1442 __asm__(".p2align 4");
1443# if __GNUC__ == 8 || __GNUC__ == 10
1444 /* good for gcc-8 and gcc-10 */
1445 __asm__("nop");
1446 __asm__(".p2align 3");
1447# endif
1448# endif
1449#endif
1450
1451 /* Handle the initial state where litBuffer is currently split between dst and litExtraBuffer */
1452 for (; litPtr + sequence.litLength <= dctx->litBufferEnd; ) {
1453 size_t const oneSeqSize = ZSTD_execSequenceSplitLitBuffer(op, oend, litPtr + sequence.litLength - WILDCOPY_OVERLENGTH, sequence, &litPtr, litBufferEnd, prefixStart, vBase, dictEnd);
1454#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
1455 assert(!ZSTD_isError(oneSeqSize));
1456 if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
1457#endif
1458 if (UNLIKELY(ZSTD_isError(oneSeqSize)))
1459 return oneSeqSize;
1460 DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
1461 op += oneSeqSize;
1462 if (UNLIKELY(!--nbSeq))
1463 break;
1464 BIT_reloadDStream(&(seqState.DStream));
1465 sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
1466 }
1467
1468 /* If there are more sequences, they will need to read literals from litExtraBuffer; copy over the remainder from dst and update litPtr and litEnd */
1469 if (nbSeq > 0) {
1470 const size_t leftoverLit = dctx->litBufferEnd - litPtr;
1471 if (leftoverLit)
1472 {
1473 RETURN_ERROR_IF(leftoverLit > (size_t)(oend - op), dstSize_tooSmall, "remaining lit must fit within dstBuffer");
1474 ZSTD_safecopyDstBeforeSrc(op, litPtr, leftoverLit);
1475 sequence.litLength -= leftoverLit;
1476 op += leftoverLit;
1477 }
1478 litPtr = dctx->litExtraBuffer;
1479 litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
1480 dctx->litBufferLocation = ZSTD_not_in_dst;
1481 {
1482 size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litBufferEnd, prefixStart, vBase, dictEnd);
1483#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
1484 assert(!ZSTD_isError(oneSeqSize));
1485 if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
1486#endif
1487 if (UNLIKELY(ZSTD_isError(oneSeqSize)))
1488 return oneSeqSize;
1489 DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
1490 op += oneSeqSize;
1491 if (--nbSeq)
1492 BIT_reloadDStream(&(seqState.DStream));
1493 }
1494 }
1495 }
1496
1497 if (nbSeq > 0) /* there is remaining lit from extra buffer */
1498 {
1499
1500#if defined(__GNUC__) && defined(__x86_64__)
1501 __asm__(".p2align 6");
1502 __asm__("nop");
1503# if __GNUC__ != 7
1504 /* worse for gcc-7 better for gcc-8, gcc-9, and gcc-10 and clang */
1505 __asm__(".p2align 4");
1506 __asm__("nop");
1507 __asm__(".p2align 3");
1508# elif __GNUC__ >= 11
1509 __asm__(".p2align 3");
1510# else
1511 __asm__(".p2align 5");
1512 __asm__("nop");
1513 __asm__(".p2align 3");
1514# endif
1515#endif
1516
1517 for (; ; ) {
1518 seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
1519 size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litBufferEnd, prefixStart, vBase, dictEnd);
1520#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
1521 assert(!ZSTD_isError(oneSeqSize));
1522 if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
1523#endif
1524 if (UNLIKELY(ZSTD_isError(oneSeqSize)))
1525 return oneSeqSize;
1526 DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
1527 op += oneSeqSize;
1528 if (UNLIKELY(!--nbSeq))
1529 break;
1530 BIT_reloadDStream(&(seqState.DStream));
1531 }
1532 }
1533
1534 /* check if reached exact end */
1535 DEBUGLOG(5, "ZSTD_decompressSequences_bodySplitLitBuffer: after decode loop, remaining nbSeq : %i", nbSeq);
1536 RETURN_ERROR_IF(nbSeq, corruption_detected, "");
1537 RETURN_ERROR_IF(BIT_reloadDStream(&seqState.DStream) < BIT_DStream_completed, corruption_detected, "");
1538 /* save reps for next block */
1539 { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) dctx->entropy.rep[i] = (U32)(seqState.prevOffset[i]); }
1540 }
1541
1542 /* last literal segment */
1543 if (dctx->litBufferLocation == ZSTD_split) /* split hasn't been reached yet, first get dst then copy litExtraBuffer */
1544 {
1545 size_t const lastLLSize = litBufferEnd - litPtr;
1546 RETURN_ERROR_IF(lastLLSize > (size_t)(oend - op), dstSize_tooSmall, "");
1547 if (op != NULL) {
1548 ZSTD_memmove(op, litPtr, lastLLSize);
1549 op += lastLLSize;
1550 }
1551 litPtr = dctx->litExtraBuffer;
1552 litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
1553 dctx->litBufferLocation = ZSTD_not_in_dst;
1554 }
1555 { size_t const lastLLSize = litBufferEnd - litPtr;
1556 RETURN_ERROR_IF(lastLLSize > (size_t)(oend-op), dstSize_tooSmall, "");
1557 if (op != NULL) {
1558 ZSTD_memcpy(op, litPtr, lastLLSize);
1559 op += lastLLSize;
1560 }
1561 }
1562
1563 return op-ostart;
1564}
1565
1566FORCE_INLINE_TEMPLATE size_t
1567DONT_VECTORIZE
1568ZSTD_decompressSequences_body(ZSTD_DCtx* dctx,
1569 void* dst, size_t maxDstSize,
1570 const void* seqStart, size_t seqSize, int nbSeq,
1571 const ZSTD_longOffset_e isLongOffset,
1572 const int frame)
1573{
1574 const BYTE* ip = (const BYTE*)seqStart;
1575 const BYTE* const iend = ip + seqSize;
1576 BYTE* const ostart = (BYTE*)dst;
1577 BYTE* const oend = dctx->litBufferLocation == ZSTD_not_in_dst ? ostart + maxDstSize : dctx->litBuffer;
1578 BYTE* op = ostart;
1579 const BYTE* litPtr = dctx->litPtr;
1580 const BYTE* const litEnd = litPtr + dctx->litSize;
1581 const BYTE* const prefixStart = (const BYTE*)(dctx->prefixStart);
1582 const BYTE* const vBase = (const BYTE*)(dctx->virtualStart);
1583 const BYTE* const dictEnd = (const BYTE*)(dctx->dictEnd);
1584 DEBUGLOG(5, "ZSTD_decompressSequences_body: nbSeq = %d", nbSeq);
1585 (void)frame;
1586
1587 /* Regen sequences */
1588 if (nbSeq) {
1589 seqState_t seqState;
1590 dctx->fseEntropy = 1;
1591 { U32 i; for (i = 0; i < ZSTD_REP_NUM; i++) seqState.prevOffset[i] = dctx->entropy.rep[i]; }
1592 RETURN_ERROR_IF(
1593 ERR_isError(BIT_initDStream(&seqState.DStream, ip, iend - ip)),
1594 corruption_detected, "");
1595 ZSTD_initFseState(&seqState.stateLL, &seqState.DStream, dctx->LLTptr);
1596 ZSTD_initFseState(&seqState.stateOffb, &seqState.DStream, dctx->OFTptr);
1597 ZSTD_initFseState(&seqState.stateML, &seqState.DStream, dctx->MLTptr);
1598 assert(dst != NULL);
1599
1600 ZSTD_STATIC_ASSERT(
1601 BIT_DStream_unfinished < BIT_DStream_completed &&
1602 BIT_DStream_endOfBuffer < BIT_DStream_completed &&
1603 BIT_DStream_completed < BIT_DStream_overflow);
1604
1605#if defined(__GNUC__) && defined(__x86_64__)
1606 __asm__(".p2align 6");
1607 __asm__("nop");
1608# if __GNUC__ >= 7
1609 __asm__(".p2align 5");
1610 __asm__("nop");
1611 __asm__(".p2align 3");
1612# else
1613 __asm__(".p2align 4");
1614 __asm__("nop");
1615 __asm__(".p2align 3");
1616# endif
1617#endif
1618
1619 for ( ; ; ) {
1620 seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
1621 size_t const oneSeqSize = ZSTD_execSequence(op, oend, sequence, &litPtr, litEnd, prefixStart, vBase, dictEnd);
1622#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
1623 assert(!ZSTD_isError(oneSeqSize));
1624 if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequence, prefixStart, vBase);
1625#endif
1626 if (UNLIKELY(ZSTD_isError(oneSeqSize)))
1627 return oneSeqSize;
1628 DEBUGLOG(6, "regenerated sequence size : %u", (U32)oneSeqSize);
1629 op += oneSeqSize;
1630 if (UNLIKELY(!--nbSeq))
1631 break;
1632 BIT_reloadDStream(&(seqState.DStream));
1633 }
1634
1635 /* check if reached exact end */
1636 DEBUGLOG(5, "ZSTD_decompressSequences_body: after decode loop, remaining nbSeq : %i", nbSeq);
1637 RETURN_ERROR_IF(nbSeq, corruption_detected, "");
1638 RETURN_ERROR_IF(BIT_reloadDStream(&seqState.DStream) < BIT_DStream_completed, corruption_detected, "");
1639 /* save reps for next block */
1640 { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) dctx->entropy.rep[i] = (U32)(seqState.prevOffset[i]); }
1641 }
1642
1643 /* last literal segment */
1644 { size_t const lastLLSize = litEnd - litPtr;
1645 RETURN_ERROR_IF(lastLLSize > (size_t)(oend-op), dstSize_tooSmall, "");
1646 if (op != NULL) {
1647 ZSTD_memcpy(op, litPtr, lastLLSize);
1648 op += lastLLSize;
1649 }
1650 }
1651
1652 return op-ostart;
1653}
1654
1655static size_t
1656ZSTD_decompressSequences_default(ZSTD_DCtx* dctx,
1657 void* dst, size_t maxDstSize,
1658 const void* seqStart, size_t seqSize, int nbSeq,
1659 const ZSTD_longOffset_e isLongOffset,
1660 const int frame)
1661{
1662 return ZSTD_decompressSequences_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
1663}
1664
1665static size_t
1666ZSTD_decompressSequencesSplitLitBuffer_default(ZSTD_DCtx* dctx,
1667 void* dst, size_t maxDstSize,
1668 const void* seqStart, size_t seqSize, int nbSeq,
1669 const ZSTD_longOffset_e isLongOffset,
1670 const int frame)
1671{
1672 return ZSTD_decompressSequences_bodySplitLitBuffer(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
1673}
1674#endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */
1675
1676#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
1677
1678FORCE_INLINE_TEMPLATE size_t
1679ZSTD_prefetchMatch(size_t prefetchPos, seq_t const sequence,
1680 const BYTE* const prefixStart, const BYTE* const dictEnd)
1681{
1682 prefetchPos += sequence.litLength;
1683 { const BYTE* const matchBase = (sequence.offset > prefetchPos) ? dictEnd : prefixStart;
1684 const BYTE* const match = matchBase + prefetchPos - sequence.offset; /* note : this operation can overflow when seq.offset is really too large, which can only happen when input is corrupted.
1685 * No consequence though : memory address is only used for prefetching, not for dereferencing */
1686 PREFETCH_L1(match); PREFETCH_L1(match+CACHELINE_SIZE); /* note : it's safe to invoke PREFETCH() on any memory address, including invalid ones */
1687 }
1688 return prefetchPos + sequence.matchLength;
1689}
1690
1691/* This decoding function employs prefetching
1692 * to reduce latency impact of cache misses.
1693 * It's generally employed when block contains a significant portion of long-distance matches
1694 * or when coupled with a "cold" dictionary */
1695FORCE_INLINE_TEMPLATE size_t
1696ZSTD_decompressSequencesLong_body(
1697 ZSTD_DCtx* dctx,
1698 void* dst, size_t maxDstSize,
1699 const void* seqStart, size_t seqSize, int nbSeq,
1700 const ZSTD_longOffset_e isLongOffset,
1701 const int frame)
1702{
1703 const BYTE* ip = (const BYTE*)seqStart;
1704 const BYTE* const iend = ip + seqSize;
1705 BYTE* const ostart = (BYTE*)dst;
1706 BYTE* const oend = dctx->litBufferLocation == ZSTD_in_dst ? dctx->litBuffer : ostart + maxDstSize;
1707 BYTE* op = ostart;
1708 const BYTE* litPtr = dctx->litPtr;
1709 const BYTE* litBufferEnd = dctx->litBufferEnd;
1710 const BYTE* const prefixStart = (const BYTE*) (dctx->prefixStart);
1711 const BYTE* const dictStart = (const BYTE*) (dctx->virtualStart);
1712 const BYTE* const dictEnd = (const BYTE*) (dctx->dictEnd);
1713 (void)frame;
1714
1715 /* Regen sequences */
1716 if (nbSeq) {
1717#define STORED_SEQS 8
1718#define STORED_SEQS_MASK (STORED_SEQS-1)
1719#define ADVANCED_SEQS STORED_SEQS
1720 seq_t sequences[STORED_SEQS];
1721 int const seqAdvance = MIN(nbSeq, ADVANCED_SEQS);
1722 seqState_t seqState;
1723 int seqNb;
1724 size_t prefetchPos = (size_t)(op-prefixStart); /* track position relative to prefixStart */
1725
1726 dctx->fseEntropy = 1;
1727 { int i; for (i=0; i<ZSTD_REP_NUM; i++) seqState.prevOffset[i] = dctx->entropy.rep[i]; }
1728 assert(dst != NULL);
1729 assert(iend >= ip);
1730 RETURN_ERROR_IF(
1731 ERR_isError(BIT_initDStream(&seqState.DStream, ip, iend-ip)),
1732 corruption_detected, "");
1733 ZSTD_initFseState(&seqState.stateLL, &seqState.DStream, dctx->LLTptr);
1734 ZSTD_initFseState(&seqState.stateOffb, &seqState.DStream, dctx->OFTptr);
1735 ZSTD_initFseState(&seqState.stateML, &seqState.DStream, dctx->MLTptr);
1736
1737 /* prepare in advance */
1738 for (seqNb=0; (BIT_reloadDStream(&seqState.DStream) <= BIT_DStream_completed) && (seqNb<seqAdvance); seqNb++) {
1739 seq_t const sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
1740 prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
1741 sequences[seqNb] = sequence;
1742 }
1743 RETURN_ERROR_IF(seqNb<seqAdvance, corruption_detected, "");
1744
1745 /* decompress without stomping litBuffer */
1746 for (; (BIT_reloadDStream(&(seqState.DStream)) <= BIT_DStream_completed) && (seqNb < nbSeq); seqNb++) {
1747 seq_t sequence = ZSTD_decodeSequence(&seqState, isLongOffset);
1748 size_t oneSeqSize;
1749
1750 if (dctx->litBufferLocation == ZSTD_split && litPtr + sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK].litLength > dctx->litBufferEnd)
1751 {
1752 /* lit buffer is reaching split point, empty out the first buffer and transition to litExtraBuffer */
1753 const size_t leftoverLit = dctx->litBufferEnd - litPtr;
1754 if (leftoverLit)
1755 {
1756 RETURN_ERROR_IF(leftoverLit > (size_t)(oend - op), dstSize_tooSmall, "remaining lit must fit within dstBuffer");
1757 ZSTD_safecopyDstBeforeSrc(op, litPtr, leftoverLit);
1758 sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK].litLength -= leftoverLit;
1759 op += leftoverLit;
1760 }
1761 litPtr = dctx->litExtraBuffer;
1762 litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
1763 dctx->litBufferLocation = ZSTD_not_in_dst;
1764 oneSeqSize = ZSTD_execSequence(op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
1765#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
1766 assert(!ZSTD_isError(oneSeqSize));
1767 if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], prefixStart, dictStart);
1768#endif
1769 if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
1770
1771 prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
1772 sequences[seqNb & STORED_SEQS_MASK] = sequence;
1773 op += oneSeqSize;
1774 }
1775 else
1776 {
1777 /* lit buffer is either wholly contained in first or second split, or not split at all*/
1778 oneSeqSize = dctx->litBufferLocation == ZSTD_split ?
1779 ZSTD_execSequenceSplitLitBuffer(op, oend, litPtr + sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK].litLength - WILDCOPY_OVERLENGTH, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd) :
1780 ZSTD_execSequence(op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
1781#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
1782 assert(!ZSTD_isError(oneSeqSize));
1783 if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[(seqNb - ADVANCED_SEQS) & STORED_SEQS_MASK], prefixStart, dictStart);
1784#endif
1785 if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
1786
1787 prefetchPos = ZSTD_prefetchMatch(prefetchPos, sequence, prefixStart, dictEnd);
1788 sequences[seqNb & STORED_SEQS_MASK] = sequence;
1789 op += oneSeqSize;
1790 }
1791 }
1792 RETURN_ERROR_IF(seqNb<nbSeq, corruption_detected, "");
1793
1794 /* finish queue */
1795 seqNb -= seqAdvance;
1796 for ( ; seqNb<nbSeq ; seqNb++) {
1797 seq_t *sequence = &(sequences[seqNb&STORED_SEQS_MASK]);
1798 if (dctx->litBufferLocation == ZSTD_split && litPtr + sequence->litLength > dctx->litBufferEnd)
1799 {
1800 const size_t leftoverLit = dctx->litBufferEnd - litPtr;
1801 if (leftoverLit)
1802 {
1803 RETURN_ERROR_IF(leftoverLit > (size_t)(oend - op), dstSize_tooSmall, "remaining lit must fit within dstBuffer");
1804 ZSTD_safecopyDstBeforeSrc(op, litPtr, leftoverLit);
1805 sequence->litLength -= leftoverLit;
1806 op += leftoverLit;
1807 }
1808 litPtr = dctx->litExtraBuffer;
1809 litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
1810 dctx->litBufferLocation = ZSTD_not_in_dst;
1811 {
1812 size_t const oneSeqSize = ZSTD_execSequence(op, oend, *sequence, &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
1813#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
1814 assert(!ZSTD_isError(oneSeqSize));
1815 if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[seqNb&STORED_SEQS_MASK], prefixStart, dictStart);
1816#endif
1817 if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
1818 op += oneSeqSize;
1819 }
1820 }
1821 else
1822 {
1823 size_t const oneSeqSize = dctx->litBufferLocation == ZSTD_split ?
1824 ZSTD_execSequenceSplitLitBuffer(op, oend, litPtr + sequence->litLength - WILDCOPY_OVERLENGTH, *sequence, &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd) :
1825 ZSTD_execSequence(op, oend, *sequence, &litPtr, litBufferEnd, prefixStart, dictStart, dictEnd);
1826#if defined(FUZZING_BUILD_MODE_UNSAFE_FOR_PRODUCTION) && defined(FUZZING_ASSERT_VALID_SEQUENCE)
1827 assert(!ZSTD_isError(oneSeqSize));
1828 if (frame) ZSTD_assertValidSequence(dctx, op, oend, sequences[seqNb&STORED_SEQS_MASK], prefixStart, dictStart);
1829#endif
1830 if (ZSTD_isError(oneSeqSize)) return oneSeqSize;
1831 op += oneSeqSize;
1832 }
1833 }
1834
1835 /* save reps for next block */
1836 { U32 i; for (i=0; i<ZSTD_REP_NUM; i++) dctx->entropy.rep[i] = (U32)(seqState.prevOffset[i]); }
1837 }
1838
1839 /* last literal segment */
1840 if (dctx->litBufferLocation == ZSTD_split) /* first deplete literal buffer in dst, then copy litExtraBuffer */
1841 {
1842 size_t const lastLLSize = litBufferEnd - litPtr;
1843 RETURN_ERROR_IF(lastLLSize > (size_t)(oend - op), dstSize_tooSmall, "");
1844 if (op != NULL) {
1845 ZSTD_memmove(op, litPtr, lastLLSize);
1846 op += lastLLSize;
1847 }
1848 litPtr = dctx->litExtraBuffer;
1849 litBufferEnd = dctx->litExtraBuffer + ZSTD_LITBUFFEREXTRASIZE;
1850 }
1851 { size_t const lastLLSize = litBufferEnd - litPtr;
1852 RETURN_ERROR_IF(lastLLSize > (size_t)(oend-op), dstSize_tooSmall, "");
1853 if (op != NULL) {
1854 ZSTD_memmove(op, litPtr, lastLLSize);
1855 op += lastLLSize;
1856 }
1857 }
1858
1859 return op-ostart;
1860}
1861
1862static size_t
1863ZSTD_decompressSequencesLong_default(ZSTD_DCtx* dctx,
1864 void* dst, size_t maxDstSize,
1865 const void* seqStart, size_t seqSize, int nbSeq,
1866 const ZSTD_longOffset_e isLongOffset,
1867 const int frame)
1868{
1869 return ZSTD_decompressSequencesLong_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
1870}
1871#endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT */
1872
1873
1874
1875#if DYNAMIC_BMI2
1876
1877#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
1878static BMI2_TARGET_ATTRIBUTE size_t
1879DONT_VECTORIZE
1880ZSTD_decompressSequences_bmi2(ZSTD_DCtx* dctx,
1881 void* dst, size_t maxDstSize,
1882 const void* seqStart, size_t seqSize, int nbSeq,
1883 const ZSTD_longOffset_e isLongOffset,
1884 const int frame)
1885{
1886 return ZSTD_decompressSequences_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
1887}
1888static BMI2_TARGET_ATTRIBUTE size_t
1889DONT_VECTORIZE
1890ZSTD_decompressSequencesSplitLitBuffer_bmi2(ZSTD_DCtx* dctx,
1891 void* dst, size_t maxDstSize,
1892 const void* seqStart, size_t seqSize, int nbSeq,
1893 const ZSTD_longOffset_e isLongOffset,
1894 const int frame)
1895{
1896 return ZSTD_decompressSequences_bodySplitLitBuffer(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
1897}
1898#endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */
1899
1900#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
1901static BMI2_TARGET_ATTRIBUTE size_t
1902ZSTD_decompressSequencesLong_bmi2(ZSTD_DCtx* dctx,
1903 void* dst, size_t maxDstSize,
1904 const void* seqStart, size_t seqSize, int nbSeq,
1905 const ZSTD_longOffset_e isLongOffset,
1906 const int frame)
1907{
1908 return ZSTD_decompressSequencesLong_body(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
1909}
1910#endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT */
1911
1912#endif /* DYNAMIC_BMI2 */
1913
1914typedef size_t (*ZSTD_decompressSequences_t)(
1915 ZSTD_DCtx* dctx,
1916 void* dst, size_t maxDstSize,
1917 const void* seqStart, size_t seqSize, int nbSeq,
1918 const ZSTD_longOffset_e isLongOffset,
1919 const int frame);
1920
1921#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
1922static size_t
1923ZSTD_decompressSequences(ZSTD_DCtx* dctx, void* dst, size_t maxDstSize,
1924 const void* seqStart, size_t seqSize, int nbSeq,
1925 const ZSTD_longOffset_e isLongOffset,
1926 const int frame)
1927{
1928 DEBUGLOG(5, "ZSTD_decompressSequences");
1929#if DYNAMIC_BMI2
1930 if (ZSTD_DCtx_get_bmi2(dctx)) {
1931 return ZSTD_decompressSequences_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
1932 }
1933#endif
1934 return ZSTD_decompressSequences_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
1935}
1936static size_t
1937ZSTD_decompressSequencesSplitLitBuffer(ZSTD_DCtx* dctx, void* dst, size_t maxDstSize,
1938 const void* seqStart, size_t seqSize, int nbSeq,
1939 const ZSTD_longOffset_e isLongOffset,
1940 const int frame)
1941{
1942 DEBUGLOG(5, "ZSTD_decompressSequencesSplitLitBuffer");
1943#if DYNAMIC_BMI2
1944 if (ZSTD_DCtx_get_bmi2(dctx)) {
1945 return ZSTD_decompressSequencesSplitLitBuffer_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
1946 }
1947#endif
1948 return ZSTD_decompressSequencesSplitLitBuffer_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
1949}
1950#endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG */
1951
1952
1953#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
1954/* ZSTD_decompressSequencesLong() :
1955 * decompression function triggered when a minimum share of offsets is considered "long",
1956 * aka out of cache.
1957 * note : "long" definition seems overloaded here, sometimes meaning "wider than bitstream register", and sometimes meaning "farther than memory cache distance".
1958 * This function will try to mitigate main memory latency through the use of prefetching */
1959static size_t
1960ZSTD_decompressSequencesLong(ZSTD_DCtx* dctx,
1961 void* dst, size_t maxDstSize,
1962 const void* seqStart, size_t seqSize, int nbSeq,
1963 const ZSTD_longOffset_e isLongOffset,
1964 const int frame)
1965{
1966 DEBUGLOG(5, "ZSTD_decompressSequencesLong");
1967#if DYNAMIC_BMI2
1968 if (ZSTD_DCtx_get_bmi2(dctx)) {
1969 return ZSTD_decompressSequencesLong_bmi2(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
1970 }
1971#endif
1972 return ZSTD_decompressSequencesLong_default(dctx, dst, maxDstSize, seqStart, seqSize, nbSeq, isLongOffset, frame);
1973}
1974#endif /* ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT */
1975
1976
1977/**
1978 * @returns The total size of the history referenceable by zstd, including
1979 * both the prefix and the extDict. At @p op any offset larger than this
1980 * is invalid.
1981 */
1982static size_t ZSTD_totalHistorySize(BYTE* op, BYTE const* virtualStart)
1983{
1984 return (size_t)(op - virtualStart);
1985}
1986
1987typedef struct {
1988 unsigned longOffsetShare;
1989 unsigned maxNbAdditionalBits;
1990} ZSTD_OffsetInfo;
1991
1992/* ZSTD_getOffsetInfo() :
1993 * condition : offTable must be valid
1994 * @return : "share" of long offsets (arbitrarily defined as > (1<<23))
1995 * compared to maximum possible of (1<<OffFSELog),
1996 * as well as the maximum number additional bits required.
1997 */
1998static ZSTD_OffsetInfo
1999ZSTD_getOffsetInfo(const ZSTD_seqSymbol* offTable, int nbSeq)
2000{
2001 ZSTD_OffsetInfo info = {0, 0};
2002 /* If nbSeq == 0, then the offTable is uninitialized, but we have
2003 * no sequences, so both values should be 0.
2004 */
2005 if (nbSeq != 0) {
2006 const void* ptr = offTable;
2007 U32 const tableLog = ((const ZSTD_seqSymbol_header*)ptr)[0].tableLog;
2008 const ZSTD_seqSymbol* table = offTable + 1;
2009 U32 const max = 1 << tableLog;
2010 U32 u;
2011 DEBUGLOG(5, "ZSTD_getLongOffsetsShare: (tableLog=%u)", tableLog);
2012
2013 assert(max <= (1 << OffFSELog)); /* max not too large */
2014 for (u=0; u<max; u++) {
2015 info.maxNbAdditionalBits = MAX(info.maxNbAdditionalBits, table[u].nbAdditionalBits);
2016 if (table[u].nbAdditionalBits > 22) info.longOffsetShare += 1;
2017 }
2018
2019 assert(tableLog <= OffFSELog);
2020 info.longOffsetShare <<= (OffFSELog - tableLog); /* scale to OffFSELog */
2021 }
2022
2023 return info;
2024}
2025
2026/**
2027 * @returns The maximum offset we can decode in one read of our bitstream, without
2028 * reloading more bits in the middle of the offset bits read. Any offsets larger
2029 * than this must use the long offset decoder.
2030 */
2031static size_t ZSTD_maxShortOffset(void)
2032{
2033 if (MEM_64bits()) {
2034 /* We can decode any offset without reloading bits.
2035 * This might change if the max window size grows.
2036 */
2037 ZSTD_STATIC_ASSERT(ZSTD_WINDOWLOG_MAX <= 31);
2038 return (size_t)-1;
2039 } else {
2040 /* The maximum offBase is (1 << (STREAM_ACCUMULATOR_MIN + 1)) - 1.
2041 * This offBase would require STREAM_ACCUMULATOR_MIN extra bits.
2042 * Then we have to subtract ZSTD_REP_NUM to get the maximum possible offset.
2043 */
2044 size_t const maxOffbase = ((size_t)1 << (STREAM_ACCUMULATOR_MIN + 1)) - 1;
2045 size_t const maxOffset = maxOffbase - ZSTD_REP_NUM;
2046 assert(ZSTD_highbit32((U32)maxOffbase) == STREAM_ACCUMULATOR_MIN);
2047 return maxOffset;
2048 }
2049}
2050
2051size_t
2052ZSTD_decompressBlock_internal(ZSTD_DCtx* dctx,
2053 void* dst, size_t dstCapacity,
2054 const void* src, size_t srcSize, const int frame, const streaming_operation streaming)
2055{ /* blockType == blockCompressed */
2056 const BYTE* ip = (const BYTE*)src;
2057 DEBUGLOG(5, "ZSTD_decompressBlock_internal (size : %u)", (U32)srcSize);
2058
2059 /* Note : the wording of the specification
2060 * allows compressed block to be sized exactly ZSTD_BLOCKSIZE_MAX.
2061 * This generally does not happen, as it makes little sense,
2062 * since an uncompressed block would feature same size and have no decompression cost.
2063 * Also, note that decoder from reference libzstd before < v1.5.4
2064 * would consider this edge case as an error.
2065 * As a consequence, avoid generating compressed blocks of size ZSTD_BLOCKSIZE_MAX
2066 * for broader compatibility with the deployed ecosystem of zstd decoders */
2067 RETURN_ERROR_IF(srcSize > ZSTD_BLOCKSIZE_MAX, srcSize_wrong, "");
2068
2069 /* Decode literals section */
2070 { size_t const litCSize = ZSTD_decodeLiteralsBlock(dctx, src, srcSize, dst, dstCapacity, streaming);
2071 DEBUGLOG(5, "ZSTD_decodeLiteralsBlock : cSize=%u, nbLiterals=%zu", (U32)litCSize, dctx->litSize);
2072 if (ZSTD_isError(litCSize)) return litCSize;
2073 ip += litCSize;
2074 srcSize -= litCSize;
2075 }
2076
2077 /* Build Decoding Tables */
2078 {
2079 /* Compute the maximum block size, which must also work when !frame and fParams are unset.
2080 * Additionally, take the min with dstCapacity to ensure that the totalHistorySize fits in a size_t.
2081 */
2082 size_t const blockSizeMax = MIN(dstCapacity, (frame ? dctx->fParams.blockSizeMax : ZSTD_BLOCKSIZE_MAX));
2083 size_t const totalHistorySize = ZSTD_totalHistorySize((BYTE*)dst + blockSizeMax, (BYTE const*)dctx->virtualStart);
2084 /* isLongOffset must be true if there are long offsets.
2085 * Offsets are long if they are larger than ZSTD_maxShortOffset().
2086 * We don't expect that to be the case in 64-bit mode.
2087 *
2088 * We check here to see if our history is large enough to allow long offsets.
2089 * If it isn't, then we can't possible have (valid) long offsets. If the offset
2090 * is invalid, then it is okay to read it incorrectly.
2091 *
2092 * If isLongOffsets is true, then we will later check our decoding table to see
2093 * if it is even possible to generate long offsets.
2094 */
2095 ZSTD_longOffset_e isLongOffset = (ZSTD_longOffset_e)(MEM_32bits() && (totalHistorySize > ZSTD_maxShortOffset()));
2096 /* These macros control at build-time which decompressor implementation
2097 * we use. If neither is defined, we do some inspection and dispatch at
2098 * runtime.
2099 */
2100#if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
2101 !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
2102 int usePrefetchDecoder = dctx->ddictIsCold;
2103#else
2104 /* Set to 1 to avoid computing offset info if we don't need to.
2105 * Otherwise this value is ignored.
2106 */
2107 int usePrefetchDecoder = 1;
2108#endif
2109 int nbSeq;
2110 size_t const seqHSize = ZSTD_decodeSeqHeaders(dctx, &nbSeq, ip, srcSize);
2111 if (ZSTD_isError(seqHSize)) return seqHSize;
2112 ip += seqHSize;
2113 srcSize -= seqHSize;
2114
2115 RETURN_ERROR_IF((dst == NULL || dstCapacity == 0) && nbSeq > 0, dstSize_tooSmall, "NULL not handled");
2116 RETURN_ERROR_IF(MEM_64bits() && sizeof(size_t) == sizeof(void*) && (size_t)(-1) - (size_t)dst < (size_t)(1 << 20), dstSize_tooSmall,
2117 "invalid dst");
2118
2119 /* If we could potentially have long offsets, or we might want to use the prefetch decoder,
2120 * compute information about the share of long offsets, and the maximum nbAdditionalBits.
2121 * NOTE: could probably use a larger nbSeq limit
2122 */
2123 if (isLongOffset || (!usePrefetchDecoder && (totalHistorySize > (1u << 24)) && (nbSeq > 8))) {
2124 ZSTD_OffsetInfo const info = ZSTD_getOffsetInfo(dctx->OFTptr, nbSeq);
2125 if (isLongOffset && info.maxNbAdditionalBits <= STREAM_ACCUMULATOR_MIN) {
2126 /* If isLongOffset, but the maximum number of additional bits that we see in our table is small
2127 * enough, then we know it is impossible to have too long an offset in this block, so we can
2128 * use the regular offset decoder.
2129 */
2130 isLongOffset = ZSTD_lo_isRegularOffset;
2131 }
2132 if (!usePrefetchDecoder) {
2133 U32 const minShare = MEM_64bits() ? 7 : 20; /* heuristic values, correspond to 2.73% and 7.81% */
2134 usePrefetchDecoder = (info.longOffsetShare >= minShare);
2135 }
2136 }
2137
2138 dctx->ddictIsCold = 0;
2139
2140#if !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT) && \
2141 !defined(ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG)
2142 if (usePrefetchDecoder) {
2143#else
2144 (void)usePrefetchDecoder;
2145 {
2146#endif
2147#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_SHORT
2148 return ZSTD_decompressSequencesLong(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame);
2149#endif
2150 }
2151
2152#ifndef ZSTD_FORCE_DECOMPRESS_SEQUENCES_LONG
2153 /* else */
2154 if (dctx->litBufferLocation == ZSTD_split)
2155 return ZSTD_decompressSequencesSplitLitBuffer(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame);
2156 else
2157 return ZSTD_decompressSequences(dctx, dst, dstCapacity, ip, srcSize, nbSeq, isLongOffset, frame);
2158#endif
2159 }
2160}
2161
2162
2163void ZSTD_checkContinuity(ZSTD_DCtx* dctx, const void* dst, size_t dstSize)
2164{
2165 if (dst != dctx->previousDstEnd && dstSize > 0) { /* not contiguous */
2166 dctx->dictEnd = dctx->previousDstEnd;
2167 dctx->virtualStart = (const char*)dst - ((const char*)(dctx->previousDstEnd) - (const char*)(dctx->prefixStart));
2168 dctx->prefixStart = dst;
2169 dctx->previousDstEnd = dst;
2170 }
2171}
2172
2173
2174size_t ZSTD_decompressBlock_deprecated(ZSTD_DCtx* dctx,
2175 void* dst, size_t dstCapacity,
2176 const void* src, size_t srcSize)
2177{
2178 size_t dSize;
2179 ZSTD_checkContinuity(dctx, dst, dstCapacity);
2180 dSize = ZSTD_decompressBlock_internal(dctx, dst, dstCapacity, src, srcSize, /* frame */ 0, not_streaming);
2181 dctx->previousDstEnd = (char*)dst + dSize;
2182 return dSize;
2183}
2184
2185
2186/* NOTE: Must just wrap ZSTD_decompressBlock_deprecated() */
2187size_t ZSTD_decompressBlock(ZSTD_DCtx* dctx,
2188 void* dst, size_t dstCapacity,
2189 const void* src, size_t srcSize)
2190{
2191 return ZSTD_decompressBlock_deprecated(dctx, dst, dstCapacity, src, srcSize);
2192}
2193