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. 297 // 298 // NOTE: Since 5.3.5beta the LZMA encoder ensures that nice_len 299 // is big enough for the selected match finder. This makes it 300 // easier for applications as nice_len = 2 will always be accepted 301 // even though the effective value can be slightly bigger. 302 const uint32_t hash_bytes 303 = mf_get_hash_bytes(lz_options->match_finder); 304 assert(hash_bytes <= mf->nice_len); 305 306 const bool is_bt = (lz_options->match_finder & 0x10) != 0; 307 uint32_t hs; 308 309 if (hash_bytes == 2) { 310 hs = 0xFFFF; 311 } else { 312 // Round dictionary size up to the next 2^n - 1 so it can 313 // be used as a hash mask. 314 hs = lz_options->dict_size - 1; 315 hs |= hs >> 1; 316 hs |= hs >> 2; 317 hs |= hs >> 4; 318 hs |= hs >> 8; 319 hs >>= 1; 320 hs |= 0xFFFF; 321 322 if (hs > (UINT32_C(1) << 24)) { 323 if (hash_bytes == 3) 324 hs = (UINT32_C(1) << 24) - 1; 325 else 326 hs >>= 1; 327 } 328 } 329 330 mf->hash_mask = hs; 331 332 ++hs; 333 if (hash_bytes > 2) 334 hs += HASH_2_SIZE; 335 if (hash_bytes > 3) 336 hs += HASH_3_SIZE; 337 /* 338 No match finder uses this at the moment. 339 if (mf->hash_bytes > 4) 340 hs += HASH_4_SIZE; 341 */ 342 343 const uint32_t old_hash_count = mf->hash_count; 344 const uint32_t old_sons_count = mf->sons_count; 345 mf->hash_count = hs; 346 mf->sons_count = mf->cyclic_size; 347 if (is_bt) 348 mf->sons_count *= 2; 349 350 // Deallocate the old hash array if it exists and has different size 351 // than what is needed now. 352 if (old_hash_count != mf->hash_count 353 || old_sons_count != mf->sons_count) { 354 lzma_free(mf->hash, allocator); 355 mf->hash = NULL; 356 357 lzma_free(mf->son, allocator); 358 mf->son = NULL; 359 } 360 361 // Maximum number of match finder cycles 362 mf->depth = lz_options->depth; 363 if (mf->depth == 0) { 364 if (is_bt) 365 mf->depth = 16 + mf->nice_len / 2; 366 else 367 mf->depth = 4 + mf->nice_len / 4; 368 } 369 370 return false; 371 } 372 373 374 static bool 375 lz_encoder_init(lzma_mf *mf, const lzma_allocator *allocator, 376 const lzma_lz_options *lz_options) 377 { 378 // Allocate the history buffer. 379 if (mf->buffer == NULL) { 380 // lzma_memcmplen() is used for the dictionary buffer 381 // so we need to allocate a few extra bytes to prevent 382 // it from reading past the end of the buffer. 383 mf->buffer = lzma_alloc(mf->size + LZMA_MEMCMPLEN_EXTRA, 384 allocator); 385 if (mf->buffer == NULL) 386 return true; 387 388 // Keep Valgrind happy with lzma_memcmplen() and initialize 389 // the extra bytes whose value may get read but which will 390 // effectively get ignored. 391 memzero(mf->buffer + mf->size, LZMA_MEMCMPLEN_EXTRA); 392 } 393 394 // Use cyclic_size as initial mf->offset. This allows 395 // avoiding a few branches in the match finders. The downside is 396 // that match finder needs to be normalized more often, which may 397 // hurt performance with huge dictionaries. 398 mf->offset = mf->cyclic_size; 399 mf->read_pos = 0; 400 mf->read_ahead = 0; 401 mf->read_limit = 0; 402 mf->write_pos = 0; 403 mf->pending = 0; 404 405 #if UINT32_MAX >= SIZE_MAX / 4 406 // Check for integer overflow. (Huge dictionaries are not 407 // possible on 32-bit CPU.) 408 if (mf->hash_count > SIZE_MAX / sizeof(uint32_t) 409 || mf->sons_count > SIZE_MAX / sizeof(uint32_t)) 410 return true; 411 #endif 412 413 // Allocate and initialize the hash table. Since EMPTY_HASH_VALUE 414 // is zero, we can use lzma_alloc_zero() or memzero() for mf->hash. 415 // 416 // We don't need to initialize mf->son, but not doing that may 417 // make Valgrind complain in normalization (see normalize() in 418 // lz_encoder_mf.c). Skipping the initialization is *very* good 419 // when big dictionary is used but only small amount of data gets 420 // actually compressed: most of the mf->son won't get actually 421 // allocated by the kernel, so we avoid wasting RAM and improve 422 // initialization speed a lot. 423 if (mf->hash == NULL) { 424 mf->hash = lzma_alloc_zero(mf->hash_count * sizeof(uint32_t), 425 allocator); 426 mf->son = lzma_alloc(mf->sons_count * sizeof(uint32_t), 427 allocator); 428 429 if (mf->hash == NULL || mf->son == NULL) { 430 lzma_free(mf->hash, allocator); 431 mf->hash = NULL; 432 433 lzma_free(mf->son, allocator); 434 mf->son = NULL; 435 436 return true; 437 } 438 } else { 439 /* 440 for (uint32_t i = 0; i < mf->hash_count; ++i) 441 mf->hash[i] = EMPTY_HASH_VALUE; 442 */ 443 memzero(mf->hash, mf->hash_count * sizeof(uint32_t)); 444 } 445 446 mf->cyclic_pos = 0; 447 448 // Handle preset dictionary. 449 if (lz_options->preset_dict != NULL 450 && lz_options->preset_dict_size > 0) { 451 // If the preset dictionary is bigger than the actual 452 // dictionary, use only the tail. 453 mf->write_pos = my_min(lz_options->preset_dict_size, mf->size); 454 memcpy(mf->buffer, lz_options->preset_dict 455 + lz_options->preset_dict_size - mf->write_pos, 456 mf->write_pos); 457 mf->action = LZMA_SYNC_FLUSH; 458 mf->skip(mf, mf->write_pos); 459 } 460 461 mf->action = LZMA_RUN; 462 463 return false; 464 } 465 466 467 extern uint64_t 468 lzma_lz_encoder_memusage(const lzma_lz_options *lz_options) 469 { 470 // Old buffers must not exist when calling lz_encoder_prepare(). 471 lzma_mf mf = { 472 .buffer = NULL, 473 .hash = NULL, 474 .son = NULL, 475 .hash_count = 0, 476 .sons_count = 0, 477 }; 478 479 // Setup the size information into mf. 480 if (lz_encoder_prepare(&mf, NULL, lz_options)) 481 return UINT64_MAX; 482 483 // Calculate the memory usage. 484 return ((uint64_t)(mf.hash_count) + mf.sons_count) * sizeof(uint32_t) 485 + mf.size + sizeof(lzma_coder); 486 } 487 488 489 static void 490 lz_encoder_end(void *coder_ptr, const lzma_allocator *allocator) 491 { 492 lzma_coder *coder = coder_ptr; 493 494 lzma_next_end(&coder->next, allocator); 495 496 lzma_free(coder->mf.son, allocator); 497 lzma_free(coder->mf.hash, allocator); 498 lzma_free(coder->mf.buffer, allocator); 499 500 if (coder->lz.end != NULL) 501 coder->lz.end(coder->lz.coder, allocator); 502 else 503 lzma_free(coder->lz.coder, allocator); 504 505 lzma_free(coder, allocator); 506 return; 507 } 508 509 510 static lzma_ret 511 lz_encoder_update(void *coder_ptr, const lzma_allocator *allocator, 512 const lzma_filter *filters_null lzma_attribute((__unused__)), 513 const lzma_filter *reversed_filters) 514 { 515 lzma_coder *coder = coder_ptr; 516 517 if (coder->lz.options_update == NULL) 518 return LZMA_PROG_ERROR; 519 520 return_if_error(coder->lz.options_update( 521 coder->lz.coder, reversed_filters)); 522 523 return lzma_next_filter_update( 524 &coder->next, allocator, reversed_filters + 1); 525 } 526 527 528 static lzma_ret 529 lz_encoder_set_out_limit(void *coder_ptr, uint64_t *uncomp_size, 530 uint64_t out_limit) 531 { 532 lzma_coder *coder = coder_ptr; 533 534 // This is supported only if there are no other filters chained. 535 if (coder->next.code == NULL && coder->lz.set_out_limit != NULL) 536 return coder->lz.set_out_limit( 537 coder->lz.coder, uncomp_size, out_limit); 538 539 return LZMA_OPTIONS_ERROR; 540 } 541 542 543 extern lzma_ret 544 lzma_lz_encoder_init(lzma_next_coder *next, const lzma_allocator *allocator, 545 const lzma_filter_info *filters, 546 lzma_ret (*lz_init)(lzma_lz_encoder *lz, 547 const lzma_allocator *allocator, 548 lzma_vli id, const void *options, 549 lzma_lz_options *lz_options)) 550 { 551 #if defined(HAVE_SMALL) && !defined(HAVE_FUNC_ATTRIBUTE_CONSTRUCTOR) 552 // We need that the CRC32 table has been initialized. 553 lzma_crc32_init(); 554 #endif 555 556 // Allocate and initialize the base data structure. 557 lzma_coder *coder = next->coder; 558 if (coder == NULL) { 559 coder = lzma_alloc(sizeof(lzma_coder), allocator); 560 if (coder == NULL) 561 return LZMA_MEM_ERROR; 562 563 next->coder = coder; 564 next->code = &lz_encode; 565 next->end = &lz_encoder_end; 566 next->update = &lz_encoder_update; 567 next->set_out_limit = &lz_encoder_set_out_limit; 568 569 coder->lz.coder = NULL; 570 coder->lz.code = NULL; 571 coder->lz.end = NULL; 572 573 // mf.size is initialized to silence Valgrind 574 // when used on optimized binaries (GCC may reorder 575 // code in a way that Valgrind gets unhappy). 576 coder->mf.buffer = NULL; 577 coder->mf.size = 0; 578 coder->mf.hash = NULL; 579 coder->mf.son = NULL; 580 coder->mf.hash_count = 0; 581 coder->mf.sons_count = 0; 582 583 coder->next = LZMA_NEXT_CODER_INIT; 584 } 585 586 // Initialize the LZ-based encoder. 587 lzma_lz_options lz_options; 588 return_if_error(lz_init(&coder->lz, allocator, 589 filters[0].id, filters[0].options, &lz_options)); 590 591 // Setup the size information into coder->mf and deallocate 592 // old buffers if they have wrong size. 593 if (lz_encoder_prepare(&coder->mf, allocator, &lz_options)) 594 return LZMA_OPTIONS_ERROR; 595 596 // Allocate new buffers if needed, and do the rest of 597 // the initialization. 598 if (lz_encoder_init(&coder->mf, allocator, &lz_options)) 599 return LZMA_MEM_ERROR; 600 601 // Initialize the next filter in the chain, if any. 602 return lzma_next_filter_init(&coder->next, allocator, filters + 1); 603 } 604 605 606 extern LZMA_API(lzma_bool) 607 lzma_mf_is_supported(lzma_match_finder mf) 608 { 609 switch (mf) { 610 #ifdef HAVE_MF_HC3 611 case LZMA_MF_HC3: 612 return true; 613 #endif 614 #ifdef HAVE_MF_HC4 615 case LZMA_MF_HC4: 616 return true; 617 #endif 618 #ifdef HAVE_MF_BT2 619 case LZMA_MF_BT2: 620 return true; 621 #endif 622 #ifdef HAVE_MF_BT3 623 case LZMA_MF_BT3: 624 return true; 625 #endif 626 #ifdef HAVE_MF_BT4 627 case LZMA_MF_BT4: 628 return true; 629 #endif 630 default: 631 return false; 632 } 633 } 634