1 /* ---------- 2 * pg_lzcompress.c - 3 * 4 * This is an implementation of LZ compression for PostgreSQL. 5 * It uses a simple history table and generates 2-3 byte tags 6 * capable of backward copy information for 3-273 bytes with 7 * a max offset of 4095. 8 * 9 * Entry routines: 10 * 11 * int32 12 * pglz_compress(const char *source, int32 slen, char *dest, 13 * const PGLZ_Strategy *strategy); 14 * 15 * source is the input data to be compressed. 16 * 17 * slen is the length of the input data. 18 * 19 * dest is the output area for the compressed result. 20 * It must be at least as big as PGLZ_MAX_OUTPUT(slen). 21 * 22 * strategy is a pointer to some information controlling 23 * the compression algorithm. If NULL, the compiled 24 * in default strategy is used. 25 * 26 * The return value is the number of bytes written in the 27 * buffer dest, or -1 if compression fails; in the latter 28 * case the contents of dest are undefined. 29 * 30 * int32 31 * pglz_decompress(const char *source, int32 slen, char *dest, 32 * int32 rawsize, bool check_complete) 33 * 34 * source is the compressed input. 35 * 36 * slen is the length of the compressed input. 37 * 38 * dest is the area where the uncompressed data will be 39 * written to. It is the callers responsibility to 40 * provide enough space. 41 * 42 * The data is written to buff exactly as it was handed 43 * to pglz_compress(). No terminating zero byte is added. 44 * 45 * rawsize is the length of the uncompressed data. 46 * enable_vt_processing(void)47 * check_complete is a flag to let us know if -1 should be 48 * returned in cases where we don't reach the end of the 49 * source or dest buffers, or not. This should be false 50 * if the caller is asking for only a partial result and 51 * true otherwise. 52 * 53 * The return value is the number of bytes written in the 54 * buffer dest, or -1 if decompression fails. 55 * 56 * The decompression algorithm and internal data format: 57 * 58 * It is made with the compressed data itself. 59 * 60 * The data representation is easiest explained by describing 61 * the process of decompression. 62 * 63 * If compressed_size == rawsize, then the data 64 * is stored uncompressed as plain bytes. Thus, the decompressor 65 * simply copies rawsize bytes to the destination. 66 * 67 * Otherwise the first byte tells what to do the next 8 times. 68 * We call this the control byte. 69 * 70 * An unset bit in the control byte means, that one uncompressed 71 * byte follows, which is copied from input to output. 72 * 73 * A set bit in the control byte means, that a tag of 2-3 bytes 74 * follows. A tag contains information to copy some bytes, that 75 * are already in the output buffer, to the current location in 76 * the output. Let's call the three tag bytes T1, T2 and T3. The 77 * position of the data to copy is coded as an offset from the 78 * actual output position. 79 * 80 * The offset is in the upper nibble of T1 and in T2. 81 * The length is in the lower nibble of T1. 82 * 83 * So the 16 bits of a 2 byte tag are coded as 84 * 85 * 7---T1--0 7---T2--0 86 * OOOO LLLL OOOO OOOO 87 * 88 * This limits the offset to 1-4095 (12 bits) and the length 89 * to 3-18 (4 bits) because 3 is always added to it. To emit 90 * a tag of 2 bytes with a length of 2 only saves one control 91 * bit. But we lose one byte in the possible length of a tag. 92 * 93 * In the actual implementation, the 2 byte tag's length is 94 * limited to 3-17, because the value 0xF in the length nibble 95 * has special meaning. It means, that the next following 96 * byte (T3) has to be added to the length value of 18. That 97 * makes total limits of 1-4095 for offset and 3-273 for length. 98 * 99 * Now that we have successfully decoded a tag. We simply copy 100 * the output that occurred <offset> bytes back to the current 101 * output location in the specified <length>. Thus, a 102 * sequence of 200 spaces (think about bpchar fields) could be 103 * coded in 4 bytes. One literal space and a three byte tag to 104 * copy 199 bytes with a -1 offset. Whow - that's a compression 105 * rate of 98%! Well, the implementation needs to save the 106 * original data size too, so we need another 4 bytes for it 107 * and end up with a total compression rate of 96%, what's still 108 * worth a Whow. 109 * 110 * The compression algorithm 111 * 112 * The following uses numbers used in the default strategy. 113 * 114 * The compressor works best for attributes of a size between 115 * 1K and 1M. For smaller items there's not that much chance of 116 * redundancy in the character sequence (except for large areas 117 * of identical bytes like trailing spaces) and for bigger ones 118 * our 4K maximum look-back distance is too small. 119 * 120 * The compressor creates a table for lists of positions. 121 * For each input position (except the last 3), a hash key is 122 * built from the 4 next input bytes and the position remembered 123 * in the appropriate list. Thus, the table points to linked 124 * lists of likely to be at least in the first 4 characters 125 * matching strings. This is done on the fly while the input 126 * is compressed into the output area. Table entries are only 127 * kept for the last 4096 input positions, since we cannot use 128 * back-pointers larger than that anyway. The size of the hash 129 * table is chosen based on the size of the input - a larger table 130 * has a larger startup cost, as it needs to be initialized to 131 * zero, but reduces the number of hash collisions on long inputs. 132 * 133 * For each byte in the input, its hash key (built from this 134 * byte and the next 3) is used to find the appropriate list 135 * in the table. The lists remember the positions of all bytes 136 * that had the same hash key in the past in increasing backward 137 * offset order. Now for all entries in the used lists, the 138 * match length is computed by comparing the characters from the 139 * entries position with the characters from the actual input 140 * position. 141 * 142 * The compressor starts with a so called "good_match" of 128. 143 * It is a "prefer speed against compression ratio" optimizer. 144 * So if the first entry looked at already has 128 or more 145 * matching characters, the lookup stops and that position is 146 * used for the next tag in the output. 147 * 148 * For each subsequent entry in the history list, the "good_match" 149 * is lowered by 10%. So the compressor will be more happy with 150 * short matches the farer it has to go back in the history. 151 * Another "speed against ratio" preference characteristic of 152 * the algorithm. 153 * 154 * Thus there are 3 stop conditions for the lookup of matches: 155 * 156 * - a match >= good_match is found 157 * - there are no more history entries to look at 158 * - the next history entry is already too far back 159 * to be coded into a tag. 160 * 161 * Finally the match algorithm checks that at least a match 162 * of 3 or more bytes has been found, because that is the smallest 163 * amount of copy information to code into a tag. If so, a tag 164 * is omitted and all the input bytes covered by that are just 165 * scanned for the history add's, otherwise a literal character 166 * is omitted and only his history entry added. 167 * 168 * Acknowledgments: 169 * 170 * Many thanks to Adisak Pochanayon, who's article about SLZ 171 * inspired me to write the PostgreSQL compression this way. 172 * 173 * Jan Wieck 174 * 175 * Copyright (c) 1999-2020, PostgreSQL Global Development Group 176 * 177 * src/common/pg_lzcompress.c 178 * ---------- 179 */ 180 #ifndef FRONTEND 181 #include "postgres.h" 182 #else 183 #include "postgres_fe.h" 184 #endif 185 186 #include <limits.h> 187 188 #include "common/pg_lzcompress.h" 189 190 191 /* ---------- 192 * Local definitions 193 * ---------- 194 */ 195 #define PGLZ_MAX_HISTORY_LISTS 8192 /* must be power of 2 */ 196 #define PGLZ_HISTORY_SIZE 4096 197 #define PGLZ_MAX_MATCH 273 198 199 200 /* ---------- 201 * PGLZ_HistEntry - 202 * 203 * Linked list for the backward history lookup 204 * 205 * All the entries sharing a hash key are linked in a doubly linked list. 206 * This makes it easy to remove an entry when it's time to recycle it 207 * (because it's more than 4K positions old). 208 * ---------- 209 */ 210 typedef struct PGLZ_HistEntry 211 { 212 struct PGLZ_HistEntry *next; /* links for my hash key's list */ 213 struct PGLZ_HistEntry *prev; 214 int hindex; /* my current hash key */ 215 const char *pos; /* my input position */ 216 } PGLZ_HistEntry; 217 218 219 /* ---------- 220 * The provided standard strategies 221 * ---------- 222 */ 223 static const PGLZ_Strategy strategy_default_data = { 224 32, /* Data chunks less than 32 bytes are not 225 * compressed */ 226 INT_MAX, /* No upper limit on what we'll try to 227 * compress */ 228 25, /* Require 25% compression rate, or not worth 229 * it */ 230 1024, /* Give up if no compression in the first 1KB */ 231 128, /* Stop history lookup if a match of 128 bytes 232 * is found */ 233 10 /* Lower good match size by 10% at every loop 234 * iteration */ 235 }; 236 const PGLZ_Strategy *const PGLZ_strategy_default = &strategy_default_data; 237 238 239 static const PGLZ_Strategy strategy_always_data = { 240 0, /* Chunks of any size are compressed */ 241 INT_MAX, 242 0, /* It's enough to save one single byte */ 243 INT_MAX, /* Never give up early */ 244 128, /* Stop history lookup if a match of 128 bytes 245 * is found */ 246 6 /* Look harder for a good match */ 247 }; 248 const PGLZ_Strategy *const PGLZ_strategy_always = &strategy_always_data; 249 250 251 /* ---------- 252 * Statically allocated work arrays for history 253 * ---------- 254 */ 255 static int16 hist_start[PGLZ_MAX_HISTORY_LISTS]; 256 static PGLZ_HistEntry hist_entries[PGLZ_HISTORY_SIZE + 1]; 257 258 /* 259 * Element 0 in hist_entries is unused, and means 'invalid'. Likewise, 260 * INVALID_ENTRY_PTR in next/prev pointers mean 'invalid'. 261 */ 262 #define INVALID_ENTRY 0 263 #define INVALID_ENTRY_PTR (&hist_entries[INVALID_ENTRY]) 264 265 /* ---------- 266 * pglz_hist_idx - 267 * 268 * Computes the history table slot for the lookup by the next 4 269 * characters in the input. 270 * 271 * NB: because we use the next 4 characters, we are not guaranteed to 272 * find 3-character matches; they very possibly will be in the wrong 273 * hash list. This seems an acceptable tradeoff for spreading out the 274 * hash keys more. 275 * ---------- 276 */ 277 #define pglz_hist_idx(_s,_e, _mask) ( \ 278 ((((_e) - (_s)) < 4) ? (int) (_s)[0] : \ 279 (((_s)[0] << 6) ^ ((_s)[1] << 4) ^ \ 280 ((_s)[2] << 2) ^ (_s)[3])) & (_mask) \ 281 ) 282 283 284 /* ---------- 285 * pglz_hist_add - 286 * 287 * Adds a new entry to the history table. 288 * 289 * If _recycle is true, then we are recycling a previously used entry, 290 * and must first delink it from its old hashcode's linked list. 291 * 292 * NOTE: beware of multiple evaluations of macro's arguments, and note that 293 * _hn and _recycle are modified in the macro. 294 * ---------- 295 */ 296 #define pglz_hist_add(_hs,_he,_hn,_recycle,_s,_e, _mask) \ 297 do { \ 298 int __hindex = pglz_hist_idx((_s),(_e), (_mask)); \ 299 int16 *__myhsp = &(_hs)[__hindex]; \ 300 PGLZ_HistEntry *__myhe = &(_he)[_hn]; \ 301 if (_recycle) { \ 302 if (__myhe->prev == NULL) \ 303 (_hs)[__myhe->hindex] = __myhe->next - (_he); \ 304 else \ 305 __myhe->prev->next = __myhe->next; \ 306 if (__myhe->next != NULL) \ 307 __myhe->next->prev = __myhe->prev; \ 308 } \ 309 __myhe->next = &(_he)[*__myhsp]; \ 310 __myhe->prev = NULL; \ 311 __myhe->hindex = __hindex; \ 312 __myhe->pos = (_s); \ 313 /* If there was an existing entry in this hash slot, link */ \ 314 /* this new entry to it. However, the 0th entry in the */ \ 315 /* entries table is unused, so we can freely scribble on it. */ \ 316 /* So don't bother checking if the slot was used - we'll */ \ 317 /* scribble on the unused entry if it was not, but that's */ \ 318 /* harmless. Avoiding the branch in this critical path */ \ 319 /* speeds this up a little bit. */ \ 320 /* if (*__myhsp != INVALID_ENTRY) */ \ 321 (_he)[(*__myhsp)].prev = __myhe; \ 322 *__myhsp = _hn; \ 323 if (++(_hn) >= PGLZ_HISTORY_SIZE + 1) { \ 324 (_hn) = 1; \ 325 (_recycle) = true; \ 326 } \ 327 } while (0) 328 329 330 /* ---------- 331 * pglz_out_ctrl - 332 * 333 * Outputs the last and allocates a new control byte if needed. 334 * ---------- 335 */ 336 #define pglz_out_ctrl(__ctrlp,__ctrlb,__ctrl,__buf) \ 337 do { \ 338 if ((__ctrl & 0xff) == 0) \ 339 { \ 340 *(__ctrlp) = __ctrlb; \ 341 __ctrlp = (__buf)++; \ 342 __ctrlb = 0; \ 343 __ctrl = 1; \ 344 } \ 345 } while (0) 346 347 348 /* ---------- 349 * pglz_out_literal - 350 * 351 * Outputs a literal byte to the destination buffer including the 352 * appropriate control bit. 353 * ---------- 354 */ 355 #define pglz_out_literal(_ctrlp,_ctrlb,_ctrl,_buf,_byte) \ 356 do { \ 357 pglz_out_ctrl(_ctrlp,_ctrlb,_ctrl,_buf); \ 358 *(_buf)++ = (unsigned char)(_byte); \ 359 _ctrl <<= 1; \ 360 } while (0) 361 362 363 /* ---------- 364 * pglz_out_tag - 365 * 366 * Outputs a backward reference tag of 2-4 bytes (depending on 367 * offset and length) to the destination buffer including the 368 * appropriate control bit. 369 * ---------- 370 */ 371 #define pglz_out_tag(_ctrlp,_ctrlb,_ctrl,_buf,_len,_off) \ 372 do { \ 373 pglz_out_ctrl(_ctrlp,_ctrlb,_ctrl,_buf); \ 374 _ctrlb |= _ctrl; \ 375 _ctrl <<= 1; \ 376 if (_len > 17) \ 377 { \ 378 (_buf)[0] = (unsigned char)((((_off) & 0xf00) >> 4) | 0x0f); \ 379 (_buf)[1] = (unsigned char)(((_off) & 0xff)); \ 380 (_buf)[2] = (unsigned char)((_len) - 18); \ 381 (_buf) += 3; \ 382 } else { \ 383 (_buf)[0] = (unsigned char)((((_off) & 0xf00) >> 4) | ((_len) - 3)); \ 384 (_buf)[1] = (unsigned char)((_off) & 0xff); \ 385 (_buf) += 2; \ 386 } \ 387 } while (0) 388 389 390 /* ---------- 391 * pglz_find_match - 392 * 393 * Lookup the history table if the actual input stream matches 394 * another sequence of characters, starting somewhere earlier 395 * in the input buffer. 396 * ---------- 397 */ 398 static inline int 399 pglz_find_match(int16 *hstart, const char *input, const char *end, 400 int *lenp, int *offp, int good_match, int good_drop, int mask) 401 { 402 PGLZ_HistEntry *hent; 403 int16 hentno; 404 int32 len = 0; 405 int32 off = 0; 406 407 /* 408 * Traverse the linked history list until a good enough match is found. 409 */ 410 hentno = hstart[pglz_hist_idx(input, end, mask)]; 411 hent = &hist_entries[hentno]; 412 while (hent != INVALID_ENTRY_PTR) 413 { 414 const char *ip = input; 415 const char *hp = hent->pos; 416 int32 thisoff; 417 int32 thislen; 418 419 /* 420 * Stop if the offset does not fit into our tag anymore. 421 */ 422 thisoff = ip - hp; 423 if (thisoff >= 0x0fff) 424 break; 425 426 /* 427 * Determine length of match. A better match must be larger than the 428 * best so far. And if we already have a match of 16 or more bytes, 429 * it's worth the call overhead to use memcmp() to check if this match 430 * is equal for the same size. After that we must fallback to 431 * character by character comparison to know the exact position where 432 * the diff occurred. 433 */ 434 thislen = 0; 435 if (len >= 16) 436 { 437 if (memcmp(ip, hp, len) == 0) 438 { 439 thislen = len; 440 ip += len; 441 hp += len; 442 while (ip < end && *ip == *hp && thislen < PGLZ_MAX_MATCH) 443 { 444 thislen++; 445 ip++; 446 hp++; 447 } 448 } 449 } 450 else 451 { 452 while (ip < end && *ip == *hp && thislen < PGLZ_MAX_MATCH) 453 { 454 thislen++; 455 ip++; 456 hp++; 457 } 458 } 459 460 /* 461 * Remember this match as the best (if it is) 462 */ 463 if (thislen > len) 464 { 465 len = thislen; 466 off = thisoff; 467 } 468 469 /* 470 * Advance to the next history entry 471 */ 472 hent = hent->next; 473 474 /* 475 * Be happy with lesser good matches the more entries we visited. But 476 * no point in doing calculation if we're at end of list. 477 */ 478 if (hent != INVALID_ENTRY_PTR) 479 { 480 if (len >= good_match) 481 break; 482 good_match -= (good_match * good_drop) / 100; 483 } 484 } 485 486 /* 487 * Return match information only if it results at least in one byte 488 * reduction. 489 */ 490 if (len > 2) 491 { 492 *lenp = len; 493 *offp = off; 494 return 1; 495 } 496 497 return 0; 498 } 499 500 501 /* ---------- 502 * pglz_compress - 503 * 504 * Compresses source into dest using strategy. Returns the number of 505 * bytes written in buffer dest, or -1 if compression fails. 506 * ---------- 507 */ 508 int32 509 pglz_compress(const char *source, int32 slen, char *dest, 510 const PGLZ_Strategy *strategy) 511 { 512 unsigned char *bp = (unsigned char *) dest; 513 unsigned char *bstart = bp; 514 int hist_next = 1; 515 bool hist_recycle = false; 516 const char *dp = source; 517 const char *dend = source + slen; 518 unsigned char ctrl_dummy = 0; 519 unsigned char *ctrlp = &ctrl_dummy; 520 unsigned char ctrlb = 0; 521 unsigned char ctrl = 0; 522 bool found_match = false; 523 int32 match_len; 524 int32 match_off; 525 int32 good_match; 526 int32 good_drop; 527 int32 result_size; 528 int32 result_max; 529 int32 need_rate; 530 int hashsz; 531 int mask; 532 533 /* 534 * Our fallback strategy is the default. 535 */ 536 if (strategy == NULL) 537 strategy = PGLZ_strategy_default; 538 539 /* 540 * If the strategy forbids compression (at all or if source chunk size out 541 * of range), fail. 542 */ 543 if (strategy->match_size_good <= 0 || 544 slen < strategy->min_input_size || 545 slen > strategy->max_input_size) 546 return -1; 547 548 /* 549 * Limit the match parameters to the supported range. 550 */ 551 good_match = strategy->match_size_good; 552 if (good_match > PGLZ_MAX_MATCH) 553 good_match = PGLZ_MAX_MATCH; 554 else if (good_match < 17) 555 good_match = 17; 556 557 good_drop = strategy->match_size_drop; 558 if (good_drop < 0) 559 good_drop = 0; 560 else if (good_drop > 100) 561 good_drop = 100; 562 563 need_rate = strategy->min_comp_rate; 564 if (need_rate < 0) 565 need_rate = 0; 566 else if (need_rate > 99) 567 need_rate = 99; 568 569 /* 570 * Compute the maximum result size allowed by the strategy, namely the 571 * input size minus the minimum wanted compression rate. This had better 572 * be <= slen, else we might overrun the provided output buffer. 573 */ 574 if (slen > (INT_MAX / 100)) 575 { 576 /* Approximate to avoid overflow */ 577 result_max = (slen / 100) * (100 - need_rate); 578 } 579 else 580 result_max = (slen * (100 - need_rate)) / 100; 581 582 /* 583 * Experiments suggest that these hash sizes work pretty well. A large 584 * hash table minimizes collision, but has a higher startup cost. For a 585 * small input, the startup cost dominates. The table size must be a power 586 * of two. 587 */ 588 if (slen < 128) 589 hashsz = 512; 590 else if (slen < 256) 591 hashsz = 1024; 592 else if (slen < 512) 593 hashsz = 2048; 594 else if (slen < 1024) 595 hashsz = 4096; 596 else 597 hashsz = 8192; 598 mask = hashsz - 1; 599 600 /* 601 * Initialize the history lists to empty. We do not need to zero the 602 * hist_entries[] array; its entries are initialized as they are used. 603 */ 604 memset(hist_start, 0, hashsz * sizeof(int16)); 605 606 /* 607 * Compress the source directly into the output buffer. 608 */ 609 while (dp < dend) 610 { 611 /* 612 * If we already exceeded the maximum result size, fail. 613 * 614 * We check once per loop; since the loop body could emit as many as 4 615 * bytes (a control byte and 3-byte tag), PGLZ_MAX_OUTPUT() had better 616 * allow 4 slop bytes. 617 */ 618 if (bp - bstart >= result_max) 619 return -1; 620 621 /* 622 * If we've emitted more than first_success_by bytes without finding 623 * anything compressible at all, fail. This lets us fall out 624 * reasonably quickly when looking at incompressible input (such as 625 * pre-compressed data). 626 */ 627 if (!found_match && bp - bstart >= strategy->first_success_by) 628 return -1; 629 630 /* 631 * Try to find a match in the history 632 */ 633 if (pglz_find_match(hist_start, dp, dend, &match_len, 634 &match_off, good_match, good_drop, mask)) 635 { 636 /* 637 * Create the tag and add history entries for all matched 638 * characters. 639 */ 640 pglz_out_tag(ctrlp, ctrlb, ctrl, bp, match_len, match_off); 641 while (match_len--) 642 { 643 pglz_hist_add(hist_start, hist_entries, 644 hist_next, hist_recycle, 645 dp, dend, mask); 646 dp++; /* Do not do this ++ in the line above! */ 647 /* The macro would do it four times - Jan. */ 648 } 649 found_match = true; 650 } 651 else 652 { 653 /* 654 * No match found. Copy one literal byte. 655 */ 656 pglz_out_literal(ctrlp, ctrlb, ctrl, bp, *dp); 657 pglz_hist_add(hist_start, hist_entries, 658 hist_next, hist_recycle, 659 dp, dend, mask); 660 dp++; /* Do not do this ++ in the line above! */ 661 /* The macro would do it four times - Jan. */ 662 } 663 } 664 665 /* 666 * Write out the last control byte and check that we haven't overrun the 667 * output size allowed by the strategy. 668 */ 669 *ctrlp = ctrlb; 670 result_size = bp - bstart; 671 if (result_size >= result_max) 672 return -1; 673 674 /* success */ 675 return result_size; 676 } 677 678 679 /* ---------- 680 * pglz_decompress - 681 * 682 * Decompresses source into dest. Returns the number of bytes 683 * decompressed into the destination buffer, or -1 if the 684 * compressed data is corrupted. 685 * 686 * If check_complete is true, the data is considered corrupted 687 * if we don't exactly fill the destination buffer. Callers that 688 * are extracting a slice typically can't apply this check. 689 * ---------- 690 */ 691 int32 692 pglz_decompress(const char *source, int32 slen, char *dest, 693 int32 rawsize, bool check_complete) 694 { 695 const unsigned char *sp; 696 const unsigned char *srcend; 697 unsigned char *dp; 698 unsigned char *destend; 699 700 sp = (const unsigned char *) source; 701 srcend = ((const unsigned char *) source) + slen; 702 dp = (unsigned char *) dest; 703 destend = dp + rawsize; 704 705 while (sp < srcend && dp < destend) 706 { 707 /* 708 * Read one control byte and process the next 8 items (or as many as 709 * remain in the compressed input). 710 */ 711 unsigned char ctrl = *sp++; 712 int ctrlc; 713 714 for (ctrlc = 0; ctrlc < 8 && sp < srcend && dp < destend; ctrlc++) 715 { 716 if (ctrl & 1) 717 { 718 /* 719 * Set control bit means we must read a match tag. The match 720 * is coded with two bytes. First byte uses lower nibble to 721 * code length - 3. Higher nibble contains upper 4 bits of the 722 * offset. The next following byte contains the lower 8 bits 723 * of the offset. If the length is coded as 18, another 724 * extension tag byte tells how much longer the match really 725 * was (0-255). 726 */ 727 int32 len; 728 int32 off; 729 730 len = (sp[0] & 0x0f) + 3; 731 off = ((sp[0] & 0xf0) << 4) | sp[1]; 732 sp += 2; 733 if (len == 18) 734 len += *sp++; 735 736 /* 737 * Check for corrupt data: if we fell off the end of the 738 * source, or if we obtained off = 0, we have problems. (We 739 * must check this, else we risk an infinite loop below in the 740 * face of corrupt data.) 741 */ 742 if (unlikely(sp > srcend || off == 0)) 743 return -1; 744 745 /* 746 * Don't emit more data than requested. 747 */ 748 len = Min(len, destend - dp); 749 750 /* 751 * Now we copy the bytes specified by the tag from OUTPUT to 752 * OUTPUT (copy len bytes from dp - off to dp). The copied 753 * areas could overlap, so to avoid undefined behavior in 754 * memcpy(), be careful to copy only non-overlapping regions. 755 * 756 * Note that we cannot use memmove() instead, since while its 757 * behavior is well-defined, it's also not what we want. 758 */ 759 while (off < len) 760 { 761 /* 762 * We can safely copy "off" bytes since that clearly 763 * results in non-overlapping source and destination. 764 */ 765 memcpy(dp, dp - off, off); 766 len -= off; 767 dp += off; 768 769 /*---------- 770 * This bit is less obvious: we can double "off" after 771 * each such step. Consider this raw input: 772 * 112341234123412341234 773 * This will be encoded as 5 literal bytes "11234" and 774 * then a match tag with length 16 and offset 4. After 775 * memcpy'ing the first 4 bytes, we will have emitted 776 * 112341234 777 * so we can double "off" to 8, then after the next step 778 * we have emitted 779 * 11234123412341234 780 * Then we can double "off" again, after which it is more 781 * than the remaining "len" so we fall out of this loop 782 * and finish with a non-overlapping copy of the 783 * remainder. In general, a match tag with off < len 784 * implies that the decoded data has a repeat length of 785 * "off". We can handle 1, 2, 4, etc repetitions of the 786 * repeated string per memcpy until we get to a situation 787 * where the final copy step is non-overlapping. 788 * 789 * (Another way to understand this is that we are keeping 790 * the copy source point dp - off the same throughout.) 791 *---------- 792 */ 793 off += off; 794 } 795 memcpy(dp, dp - off, len); 796 dp += len; 797 } 798 else 799 { 800 /* 801 * An unset control bit means LITERAL BYTE. So we just copy 802 * one from INPUT to OUTPUT. 803 */ 804 *dp++ = *sp++; 805 } 806 807 /* 808 * Advance the control bit 809 */ 810 ctrl >>= 1; 811 } 812 } 813 814 /* 815 * If requested, check we decompressed the right amount. 816 */ 817 if (check_complete && (dp != destend || sp != srcend)) 818 return -1; 819 820 /* 821 * That's it. 822 */ 823 return (char *) dp - dest; 824 } 825 826 827 /* ---------- 828 * pglz_max_compressed_size - 829 * 830 * Calculate the maximum compressed size for a given amount of raw data. 831 * Return the maximum size, or total compressed size if maximum size is 832 * larger than total compressed size. 833 * 834 * We can't use PGLZ_MAX_OUTPUT for this purpose, because that's used to size 835 * the compression buffer (and abort the compression). It does not really say 836 * what's the maximum compressed size for an input of a given length, and it 837 * may happen that while the whole value is compressible (and thus fits into 838 * PGLZ_MAX_OUTPUT nicely), the prefix is not compressible at all. 839 * ---------- 840 */ 841 int32 842 pglz_maximum_compressed_size(int32 rawsize, int32 total_compressed_size) 843 { 844 int64 compressed_size; 845 846 /* 847 * pglz uses one control bit per byte, so if the entire desired prefix is 848 * represented as literal bytes, we'll need (rawsize * 9) bits. We care 849 * about bytes though, so be sure to round up not down. 850 * 851 * Use int64 here to prevent overflow during calculation. 852 */ 853 compressed_size = ((int64) rawsize * 9 + 7) / 8; 854 855 /* 856 * The above fails to account for a corner case: we could have compressed 857 * data that starts with N-1 or N-2 literal bytes and then has a match tag 858 * of 2 or 3 bytes. It's therefore possible that we need to fetch 1 or 2 859 * more bytes in order to have the whole match tag. (Match tags earlier 860 * in the compressed data don't cause a problem, since they should 861 * represent more decompressed bytes than they occupy themselves.) 862 */ 863 compressed_size += 2; 864 865 /* 866 * Maximum compressed size can't be larger than total compressed size. 867 * (This also ensures that our result fits in int32.) 868 */ 869 compressed_size = Min(compressed_size, total_compressed_size); 870 871 return (int32) compressed_size; 872 } 873