1 /////////////////////////////////////////////////////////////////////////////// 2 // 3 /// \file lz_encoder.c 4 /// \brief LZ in window 5 /// 6 // Authors: Igor Pavlov 7 // Lasse Collin 8 // 9 // This file has been put into the public domain. 10 // You can do whatever you want with this file. 11 // 12 /////////////////////////////////////////////////////////////////////////////// 13 14 #include "lz_encoder.h" 15 #include "lz_encoder_hash.h" 16 17 // See lz_encoder_hash.h. This is a bit hackish but avoids making 18 // endianness a conditional in makefiles. 19 #if defined(WORDS_BIGENDIAN) && !defined(HAVE_SMALL) 20 # include "lz_encoder_hash_table.h" 21 #endif 22 23 #include "memcmplen.h" 24 25 26 typedef struct { 27 /// LZ-based encoder e.g. LZMA 28 lzma_lz_encoder lz; 29 30 /// History buffer and match finder 31 lzma_mf mf; 32 33 /// Next coder in the chain 34 lzma_next_coder next; 35 } lzma_coder; 36 37 38 /// \brief Moves the data in the input window to free space for new data 39 /// 40 /// mf->buffer is a sliding input window, which keeps mf->keep_size_before 41 /// bytes of input history available all the time. Now and then we need to 42 /// "slide" the buffer to make space for the new data to the end of the 43 /// buffer. At the same time, data older than keep_size_before is dropped. 44 /// 45 static void 46 move_window(lzma_mf *mf) 47 { 48 // Align the move to a multiple of 16 bytes. Some LZ-based encoders 49 // like LZMA use the lowest bits of mf->read_pos to know the 50 // alignment of the uncompressed data. We also get better speed 51 // for memmove() with aligned buffers. 52 assert(mf->read_pos > mf->keep_size_before); 53 const uint32_t move_offset 54 = (mf->read_pos - mf->keep_size_before) & ~UINT32_C(15); 55 56 assert(mf->write_pos > move_offset); 57 const size_t move_size = mf->write_pos - move_offset; 58 59 assert(move_offset + move_size <= mf->size); 60 61 memmove(mf->buffer, mf->buffer + move_offset, move_size); 62 63 mf->offset += move_offset; 64 mf->read_pos -= move_offset; 65 mf->read_limit -= move_offset; 66 mf->write_pos -= move_offset; 67 68 return; 69 } 70 71 72 /// \brief Tries to fill the input window (mf->buffer) 73 /// 74 /// If we are the last encoder in the chain, our input data is in in[]. 75 /// Otherwise we call the next filter in the chain to process in[] and 76 /// write its output to mf->buffer. 77 /// 78 /// This function must not be called once it has returned LZMA_STREAM_END. 79 /// 80 static lzma_ret 81 fill_window(lzma_coder *coder, const lzma_allocator *allocator, 82 const uint8_t *in, size_t *in_pos, size_t in_size, 83 lzma_action action) 84 { 85 assert(coder->mf.read_pos <= coder->mf.write_pos); 86 87 // Move the sliding window if needed. 88 if (coder->mf.read_pos >= coder->mf.size - coder->mf.keep_size_after) 89 move_window(&coder->mf); 90 91 // Maybe this is ugly, but lzma_mf uses uint32_t for most things 92 // (which I find cleanest), but we need size_t here when filling 93 // the history window. 94 size_t write_pos = coder->mf.write_pos; 95 lzma_ret ret; 96 if (coder->next.code == NULL) { 97 // Not using a filter, simply memcpy() as much as possible. 98 lzma_bufcpy(in, in_pos, in_size, coder->mf.buffer, 99 &write_pos, coder->mf.size); 100 101 ret = action != LZMA_RUN && *in_pos == in_size 102 ? LZMA_STREAM_END : LZMA_OK; 103 104 } else { 105 ret = coder->next.code(coder->next.coder, allocator, 106 in, in_pos, in_size, 107 coder->mf.buffer, &write_pos, 108 coder->mf.size, action); 109 } 110 111 coder->mf.write_pos = write_pos; 112 113 // Silence Valgrind. lzma_memcmplen() can read extra bytes 114 // and Valgrind will give warnings if those bytes are uninitialized 115 // because Valgrind cannot see that the values of the uninitialized 116 // bytes are eventually ignored. 117 memzero(coder->mf.buffer + write_pos, LZMA_MEMCMPLEN_EXTRA); 118 119 // If end of stream has been reached or flushing completed, we allow 120 // the encoder to process all the input (that is, read_pos is allowed 121 // to reach write_pos). Otherwise we keep keep_size_after bytes 122 // available as prebuffer. 123 if (ret == LZMA_STREAM_END) { 124 assert(*in_pos == in_size); 125 ret = LZMA_OK; 126 coder->mf.action = action; 127 coder->mf.read_limit = coder->mf.write_pos; 128 129 } else if (coder->mf.write_pos > coder->mf.keep_size_after) { 130 // This needs to be done conditionally, because if we got 131 // only little new input, there may be too little input 132 // to do any encoding yet. 133 coder->mf.read_limit = coder->mf.write_pos 134 - coder->mf.keep_size_after; 135 } 136 137 // Restart the match finder after finished LZMA_SYNC_FLUSH. 138 if (coder->mf.pending > 0 139 && coder->mf.read_pos < coder->mf.read_limit) { 140 // Match finder may update coder->pending and expects it to 141 // start from zero, so use a temporary variable. 142 const uint32_t pending = coder->mf.pending; 143 coder->mf.pending = 0; 144 145 // Rewind read_pos so that the match finder can hash 146 // the pending bytes. 147 assert(coder->mf.read_pos >= pending); 148 coder->mf.read_pos -= pending; 149 150 // Call the skip function directly instead of using 151 // mf_skip(), since we don't want to touch mf->read_ahead. 152 coder->mf.skip(&coder->mf, pending); 153 } 154 155 return ret; 156 } 157 158 159 static lzma_ret 160 lz_encode(void *coder_ptr, const lzma_allocator *allocator, 161 const uint8_t *restrict in, size_t *restrict in_pos, 162 size_t in_size, 163 uint8_t *restrict out, size_t *restrict out_pos, 164 size_t out_size, lzma_action action) 165 { 166 lzma_coder *coder = coder_ptr; 167 168 while (*out_pos < out_size 169 && (*in_pos < in_size || action != LZMA_RUN)) { 170 // Read more data to coder->mf.buffer if needed. 171 if (coder->mf.action == LZMA_RUN && coder->mf.read_pos 172 >= coder->mf.read_limit) 173 return_if_error(fill_window(coder, allocator, 174 in, in_pos, in_size, action)); 175 176 // Encode 177 const lzma_ret ret = coder->lz.code(coder->lz.coder, 178 &coder->mf, out, out_pos, out_size); 179 if (ret != LZMA_OK) { 180 // Setting this to LZMA_RUN for cases when we are 181 // flushing. It doesn't matter when finishing or if 182 // an error occurred. 183 coder->mf.action = LZMA_RUN; 184 return ret; 185 } 186 } 187 188 return LZMA_OK; 189 } 190 191 192 static bool 193 lz_encoder_prepare(lzma_mf *mf, const lzma_allocator *allocator, 194 const lzma_lz_options *lz_options) 195 { 196 // For now, the dictionary size is limited to 1.5 GiB. This may grow 197 // in the future if needed, but it needs a little more work than just 198 // changing this check. 199 if (lz_options->dict_size < LZMA_DICT_SIZE_MIN 200 || lz_options->dict_size 201 > (UINT32_C(1) << 30) + (UINT32_C(1) << 29) 202 || lz_options->nice_len > lz_options->match_len_max) 203 return true; 204 205 mf->keep_size_before = lz_options->before_size + lz_options->dict_size; 206 207 mf->keep_size_after = lz_options->after_size 208 + lz_options->match_len_max; 209 210 // To avoid constant memmove()s, allocate some extra space. Since 211 // memmove()s become more expensive when the size of the buffer 212 // increases, we reserve more space when a large dictionary is 213 // used to make the memmove() calls rarer. 214 // 215 // This works with dictionaries up to about 3 GiB. If bigger 216 // dictionary is wanted, some extra work is needed: 217 // - Several variables in lzma_mf have to be changed from uint32_t 218 // to size_t. 219 // - Memory usage calculation needs something too, e.g. use uint64_t 220 // for mf->size. 221 uint32_t reserve = lz_options->dict_size / 2; 222 if (reserve > (UINT32_C(1) << 30)) 223 reserve /= 2; 224 225 reserve += (lz_options->before_size + lz_options->match_len_max 226 + lz_options->after_size) / 2 + (UINT32_C(1) << 19); 227 228 const uint32_t old_size = mf->size; 229 mf->size = mf->keep_size_before + reserve + mf->keep_size_after; 230 231 // Deallocate the old history buffer if it exists but has different 232 // size than what is needed now. 233 if (mf->buffer != NULL && old_size != mf->size) { 234 lzma_free(mf->buffer, allocator); 235 mf->buffer = NULL; 236 } 237 238 // Match finder options 239 mf->match_len_max = lz_options->match_len_max; 240 mf->nice_len = lz_options->nice_len; 241 242 // cyclic_size has to stay smaller than 2 Gi. Note that this doesn't 243 // mean limiting dictionary size to less than 2 GiB. With a match 244 // finder that uses multibyte resolution (hashes start at e.g. every 245 // fourth byte), cyclic_size would stay below 2 Gi even when 246 // dictionary size is greater than 2 GiB. 247 // 248 // It would be possible to allow cyclic_size >= 2 Gi, but then we 249 // would need to be careful to use 64-bit types in various places 250 // (size_t could do since we would need bigger than 32-bit address 251 // space anyway). It would also require either zeroing a multigigabyte 252 // buffer at initialization (waste of time and RAM) or allow 253 // normalization in lz_encoder_mf.c to access uninitialized 254 // memory to keep the code simpler. The current way is simple and 255 // still allows pretty big dictionaries, so I don't expect these 256 // limits to change. 257 mf->cyclic_size = lz_options->dict_size + 1; 258 259 // Validate the match finder ID and setup the function pointers. 260 switch (lz_options->match_finder) { 261 #ifdef HAVE_MF_HC3 262 case LZMA_MF_HC3: 263 mf->find = &lzma_mf_hc3_find; 264 mf->skip = &lzma_mf_hc3_skip; 265 break; 266 #endif 267 #ifdef HAVE_MF_HC4 268 case LZMA_MF_HC4: 269 mf->find = &lzma_mf_hc4_find; 270 mf->skip = &lzma_mf_hc4_skip; 271 break; 272 #endif 273 #ifdef HAVE_MF_BT2 274 case LZMA_MF_BT2: 275 mf->find = &lzma_mf_bt2_find; 276 mf->skip = &lzma_mf_bt2_skip; 277 break; 278 #endif 279 #ifdef HAVE_MF_BT3 280 case LZMA_MF_BT3: 281 mf->find = &lzma_mf_bt3_find; 282 mf->skip = &lzma_mf_bt3_skip; 283 break; 284 #endif 285 #ifdef HAVE_MF_BT4 286 case LZMA_MF_BT4: 287 mf->find = &lzma_mf_bt4_find; 288 mf->skip = &lzma_mf_bt4_skip; 289 break; 290 #endif 291 292 default: 293 return true; 294 } 295 296 // Calculate the sizes of mf->hash and mf->son and check that 297 // nice_len is big enough for the selected match finder. 298 const uint32_t hash_bytes = lz_options->match_finder & 0x0F; 299 if (hash_bytes > mf->nice_len) 300 return true; 301 302 const bool is_bt = (lz_options->match_finder & 0x10) != 0; 303 uint32_t hs; 304 305 if (hash_bytes == 2) { 306 hs = 0xFFFF; 307 } else { 308 // Round dictionary size up to the next 2^n - 1 so it can 309 // be used as a hash mask. 310 hs = lz_options->dict_size - 1; 311 hs |= hs >> 1; 312 hs |= hs >> 2; 313 hs |= hs >> 4; 314 hs |= hs >> 8; 315 hs >>= 1; 316 hs |= 0xFFFF; 317 318 if (hs > (UINT32_C(1) << 24)) { 319 if (hash_bytes == 3) 320 hs = (UINT32_C(1) << 24) - 1; 321 else 322 hs >>= 1; 323 } 324 } 325 326 mf->hash_mask = hs; 327 328 ++hs; 329 if (hash_bytes > 2) 330 hs += HASH_2_SIZE; 331 if (hash_bytes > 3) 332 hs += HASH_3_SIZE; 333 /* 334 No match finder uses this at the moment. 335 if (mf->hash_bytes > 4) 336 hs += HASH_4_SIZE; 337 */ 338 339 const uint32_t old_hash_count = mf->hash_count; 340 const uint32_t old_sons_count = mf->sons_count; 341 mf->hash_count = hs; 342 mf->sons_count = mf->cyclic_size; 343 if (is_bt) 344 mf->sons_count *= 2; 345 346 // Deallocate the old hash array if it exists and has different size 347 // than what is needed now. 348 if (old_hash_count != mf->hash_count 349 || old_sons_count != mf->sons_count) { 350 lzma_free(mf->hash, allocator); 351 mf->hash = NULL; 352 353 lzma_free(mf->son, allocator); 354 mf->son = NULL; 355 } 356 357 // Maximum number of match finder cycles 358 mf->depth = lz_options->depth; 359 if (mf->depth == 0) { 360 if (is_bt) 361 mf->depth = 16 + mf->nice_len / 2; 362 else 363 mf->depth = 4 + mf->nice_len / 4; 364 } 365 366 return false; 367 } 368 369 370 static bool 371 lz_encoder_init(lzma_mf *mf, const lzma_allocator *allocator, 372 const lzma_lz_options *lz_options) 373 { 374 // Allocate the history buffer. 375 if (mf->buffer == NULL) { 376 // lzma_memcmplen() is used for the dictionary buffer 377 // so we need to allocate a few extra bytes to prevent 378 // it from reading past the end of the buffer. 379 mf->buffer = lzma_alloc(mf->size + LZMA_MEMCMPLEN_EXTRA, 380 allocator); 381 if (mf->buffer == NULL) 382 return true; 383 384 // Keep Valgrind happy with lzma_memcmplen() and initialize 385 // the extra bytes whose value may get read but which will 386 // effectively get ignored. 387 memzero(mf->buffer + mf->size, LZMA_MEMCMPLEN_EXTRA); 388 } 389 390 // Use cyclic_size as initial mf->offset. This allows 391 // avoiding a few branches in the match finders. The downside is 392 // that match finder needs to be normalized more often, which may 393 // hurt performance with huge dictionaries. 394 mf->offset = mf->cyclic_size; 395 mf->read_pos = 0; 396 mf->read_ahead = 0; 397 mf->read_limit = 0; 398 mf->write_pos = 0; 399 mf->pending = 0; 400 401 #if UINT32_MAX >= SIZE_MAX / 4 402 // Check for integer overflow. (Huge dictionaries are not 403 // possible on 32-bit CPU.) 404 if (mf->hash_count > SIZE_MAX / sizeof(uint32_t) 405 || mf->sons_count > SIZE_MAX / sizeof(uint32_t)) 406 return true; 407 #endif 408 409 // Allocate and initialize the hash table. Since EMPTY_HASH_VALUE 410 // is zero, we can use lzma_alloc_zero() or memzero() for mf->hash. 411 // 412 // We don't need to initialize mf->son, but not doing that may 413 // make Valgrind complain in normalization (see normalize() in 414 // lz_encoder_mf.c). Skipping the initialization is *very* good 415 // when big dictionary is used but only small amount of data gets 416 // actually compressed: most of the mf->son won't get actually 417 // allocated by the kernel, so we avoid wasting RAM and improve 418 // initialization speed a lot. 419 if (mf->hash == NULL) { 420 mf->hash = lzma_alloc_zero(mf->hash_count * sizeof(uint32_t), 421 allocator); 422 mf->son = lzma_alloc(mf->sons_count * sizeof(uint32_t), 423 allocator); 424 425 if (mf->hash == NULL || mf->son == NULL) { 426 lzma_free(mf->hash, allocator); 427 mf->hash = NULL; 428 429 lzma_free(mf->son, allocator); 430 mf->son = NULL; 431 432 return true; 433 } 434 } else { 435 /* 436 for (uint32_t i = 0; i < mf->hash_count; ++i) 437 mf->hash[i] = EMPTY_HASH_VALUE; 438 */ 439 memzero(mf->hash, mf->hash_count * sizeof(uint32_t)); 440 } 441 442 mf->cyclic_pos = 0; 443 444 // Handle preset dictionary. 445 if (lz_options->preset_dict != NULL 446 && lz_options->preset_dict_size > 0) { 447 // If the preset dictionary is bigger than the actual 448 // dictionary, use only the tail. 449 mf->write_pos = my_min(lz_options->preset_dict_size, mf->size); 450 memcpy(mf->buffer, lz_options->preset_dict 451 + lz_options->preset_dict_size - mf->write_pos, 452 mf->write_pos); 453 mf->action = LZMA_SYNC_FLUSH; 454 mf->skip(mf, mf->write_pos); 455 } 456 457 mf->action = LZMA_RUN; 458 459 return false; 460 } 461 462 463 extern uint64_t 464 lzma_lz_encoder_memusage(const lzma_lz_options *lz_options) 465 { 466 // Old buffers must not exist when calling lz_encoder_prepare(). 467 lzma_mf mf = { 468 .buffer = NULL, 469 .hash = NULL, 470 .son = NULL, 471 .hash_count = 0, 472 .sons_count = 0, 473 }; 474 475 // Setup the size information into mf. 476 if (lz_encoder_prepare(&mf, NULL, lz_options)) 477 return UINT64_MAX; 478 479 // Calculate the memory usage. 480 return ((uint64_t)(mf.hash_count) + mf.sons_count) * sizeof(uint32_t) 481 + mf.size + sizeof(lzma_coder); 482 } 483 484 485 static void 486 lz_encoder_end(void *coder_ptr, const lzma_allocator *allocator) 487 { 488 lzma_coder *coder = coder_ptr; 489 490 lzma_next_end(&coder->next, allocator); 491 492 lzma_free(coder->mf.son, allocator); 493 lzma_free(coder->mf.hash, allocator); 494 lzma_free(coder->mf.buffer, allocator); 495 496 if (coder->lz.end != NULL) 497 coder->lz.end(coder->lz.coder, allocator); 498 else 499 lzma_free(coder->lz.coder, allocator); 500 501 lzma_free(coder, allocator); 502 return; 503 } 504 505 506 static lzma_ret 507 lz_encoder_update(void *coder_ptr, const lzma_allocator *allocator, 508 const lzma_filter *filters_null lzma_attribute((__unused__)), 509 const lzma_filter *reversed_filters) 510 { 511 lzma_coder *coder = coder_ptr; 512 513 if (coder->lz.options_update == NULL) 514 return LZMA_PROG_ERROR; 515 516 return_if_error(coder->lz.options_update( 517 coder->lz.coder, reversed_filters)); 518 519 return lzma_next_filter_update( 520 &coder->next, allocator, reversed_filters + 1); 521 } 522 523 524 extern lzma_ret 525 lzma_lz_encoder_init(lzma_next_coder *next, const lzma_allocator *allocator, 526 const lzma_filter_info *filters, 527 lzma_ret (*lz_init)(lzma_lz_encoder *lz, 528 const lzma_allocator *allocator, const void *options, 529 lzma_lz_options *lz_options)) 530 { 531 #ifdef HAVE_SMALL 532 // We need that the CRC32 table has been initialized. 533 lzma_crc32_init(); 534 #endif 535 536 // Allocate and initialize the base data structure. 537 lzma_coder *coder = next->coder; 538 if (coder == NULL) { 539 coder = lzma_alloc(sizeof(lzma_coder), allocator); 540 if (coder == NULL) 541 return LZMA_MEM_ERROR; 542 543 next->coder = coder; 544 next->code = &lz_encode; 545 next->end = &lz_encoder_end; 546 next->update = &lz_encoder_update; 547 548 coder->lz.coder = NULL; 549 coder->lz.code = NULL; 550 coder->lz.end = NULL; 551 552 // mf.size is initialized to silence Valgrind 553 // when used on optimized binaries (GCC may reorder 554 // code in a way that Valgrind gets unhappy). 555 coder->mf.buffer = NULL; 556 coder->mf.size = 0; 557 coder->mf.hash = NULL; 558 coder->mf.son = NULL; 559 coder->mf.hash_count = 0; 560 coder->mf.sons_count = 0; 561 562 coder->next = LZMA_NEXT_CODER_INIT; 563 } 564 565 // Initialize the LZ-based encoder. 566 lzma_lz_options lz_options; 567 return_if_error(lz_init(&coder->lz, allocator, 568 filters[0].options, &lz_options)); 569 570 // Setup the size information into coder->mf and deallocate 571 // old buffers if they have wrong size. 572 if (lz_encoder_prepare(&coder->mf, allocator, &lz_options)) 573 return LZMA_OPTIONS_ERROR; 574 575 // Allocate new buffers if needed, and do the rest of 576 // the initialization. 577 if (lz_encoder_init(&coder->mf, allocator, &lz_options)) 578 return LZMA_MEM_ERROR; 579 580 // Initialize the next filter in the chain, if any. 581 return lzma_next_filter_init(&coder->next, allocator, filters + 1); 582 } 583 584 585 extern LZMA_API(lzma_bool) 586 lzma_mf_is_supported(lzma_match_finder mf) 587 { 588 switch (mf) { 589 #ifdef HAVE_MF_HC3 590 case LZMA_MF_HC3: 591 return true; 592 #endif 593 #ifdef HAVE_MF_HC4 594 case LZMA_MF_HC4: 595 return true; 596 #endif 597 #ifdef HAVE_MF_BT2 598 case LZMA_MF_BT2: 599 return true; 600 #endif 601 #ifdef HAVE_MF_BT3 602 case LZMA_MF_BT3: 603 return true; 604 #endif 605 #ifdef HAVE_MF_BT4 606 case LZMA_MF_BT4: 607 return true; 608 #endif 609 default: 610 return false; 611 } 612 } 613