Age Owner TLA Line data Source code
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 : *
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 further 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) 1999-2023, 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
3569 heikki.linnakangas 399 CBC 133805106 : 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;
8397 bruce 404 133805106 : int32 len = 0;
405 133805106 : int32 off = 0;
406 :
407 : /*
408 : * Traverse the linked history list until a good enough match is found.
409 : */
3569 heikki.linnakangas 410 133805106 : hentno = hstart[pglz_hist_idx(input, end, mask)];
411 133805106 : hent = &hist_entries[hentno];
412 534244153 : while (hent != INVALID_ENTRY_PTR)
413 : {
6030 tgl 414 406259305 : const char *ip = input;
415 406259305 : 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 : */
7813 422 406259305 : thisoff = ip - hp;
8544 JanWieck 423 406259305 : if (thisoff >= 0x0fff)
424 111295 : 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 : */
7813 tgl 434 406148010 : thislen = 0;
8544 JanWieck 435 406148010 : if (len >= 16)
436 : {
7813 tgl 437 38662839 : if (memcmp(ip, hp, len) == 0)
438 : {
439 12422547 : thislen = len;
440 12422547 : ip += len;
441 12422547 : hp += len;
442 41333099 : while (ip < end && *ip == *hp && thislen < PGLZ_MAX_MATCH)
443 : {
444 28910552 : thislen++;
445 28910552 : ip++;
446 28910552 : hp++;
447 : }
448 : }
449 : }
450 : else
451 : {
452 2584527849 : while (ip < end && *ip == *hp && thislen < PGLZ_MAX_MATCH)
453 : {
454 2217042678 : thislen++;
455 2217042678 : ip++;
456 2217042678 : hp++;
457 : }
458 : }
459 :
460 : /*
461 : * Remember this match as the best (if it is)
462 : */
8544 JanWieck 463 406148010 : if (thislen > len)
464 : {
465 59911119 : len = thislen;
466 59911119 : off = thisoff;
467 : }
468 :
469 : /*
470 : * Advance to the next history entry
471 : */
472 406148010 : 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 : */
3553 heikki.linnakangas 478 406148010 : if (hent != INVALID_ENTRY_PTR)
479 : {
7813 tgl 480 352031520 : if (len >= good_match)
481 5708963 : break;
482 346322557 : 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 : */
8544 JanWieck 490 133805106 : if (len > 2)
491 : {
492 39739735 : *lenp = len;
493 39739735 : *offp = off;
494 39739735 : return 1;
495 : }
496 :
497 94065371 : 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
2981 fujii 509 69161 : pglz_compress(const char *source, int32 slen, char *dest,
510 : const PGLZ_Strategy *strategy)
511 : {
512 69161 : unsigned char *bp = (unsigned char *) dest;
8397 bruce 513 69161 : unsigned char *bstart = bp;
3569 heikki.linnakangas 514 69161 : int hist_next = 1;
7813 tgl 515 69161 : bool hist_recycle = false;
6030 516 69161 : const char *dp = source;
517 69161 : const char *dend = source + slen;
8397 bruce 518 69161 : unsigned char ctrl_dummy = 0;
519 69161 : unsigned char *ctrlp = &ctrl_dummy;
520 69161 : unsigned char ctrlb = 0;
521 69161 : unsigned char ctrl = 0;
5511 tgl 522 69161 : 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 : */
8544 JanWieck 536 69161 : if (strategy == NULL)
537 69161 : 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 : */
5727 tgl 543 69161 : if (strategy->match_size_good <= 0 ||
5511 544 69161 : slen < strategy->min_input_size ||
545 69161 : slen > strategy->max_input_size)
2981 fujii 546 UBC 0 : return -1;
547 :
548 : /*
549 : * Limit the match parameters to the supported range.
550 : */
5511 tgl 551 CBC 69161 : good_match = strategy->match_size_good;
552 69161 : if (good_match > PGLZ_MAX_MATCH)
8544 JanWieck 553 UBC 0 : good_match = PGLZ_MAX_MATCH;
5511 tgl 554 CBC 69161 : else if (good_match < 17)
8544 JanWieck 555 UBC 0 : good_match = 17;
556 :
5511 tgl 557 CBC 69161 : good_drop = strategy->match_size_drop;
558 69161 : if (good_drop < 0)
8544 JanWieck 559 UBC 0 : good_drop = 0;
5511 tgl 560 CBC 69161 : else if (good_drop > 100)
8544 JanWieck 561 UBC 0 : good_drop = 100;
562 :
5511 tgl 563 CBC 69161 : need_rate = strategy->min_comp_rate;
564 69161 : if (need_rate < 0)
5511 tgl 565 UBC 0 : need_rate = 0;
5511 tgl 566 CBC 69161 : else if (need_rate > 99)
5511 tgl 567 UBC 0 : 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 : */
5050 bruce 574 CBC 69161 : if (slen > (INT_MAX / 100))
575 : {
576 : /* Approximate to avoid overflow */
5511 tgl 577 UBC 0 : result_max = (slen / 100) * (100 - need_rate);
578 : }
579 : else
5511 tgl 580 CBC 69161 : 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 : */
3569 heikki.linnakangas 588 69161 : if (slen < 128)
589 320 : hashsz = 512;
590 68841 : else if (slen < 256)
591 11 : hashsz = 1024;
592 68830 : else if (slen < 512)
593 2018 : hashsz = 2048;
594 66812 : else if (slen < 1024)
595 1588 : hashsz = 4096;
596 : else
597 65224 : hashsz = 8192;
598 69161 : 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 69161 : memset(hist_start, 0, hashsz * sizeof(int16));
605 :
606 : /*
607 : * Compress the source directly into the output buffer.
608 : */
8544 JanWieck 609 133874267 : 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 133806881 : if (bp - bstart >= result_max)
2981 fujii 619 1775 : 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 : */
5511 tgl 627 133805106 : if (!found_match && bp - bstart >= strategy->first_success_by)
2981 fujii 628 UBC 0 : return -1;
629 :
630 : /*
631 : * Try to find a match in the history
632 : */
8397 bruce 633 CBC 133805106 : 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 : */
8544 JanWieck 640 39739735 : pglz_out_tag(ctrlp, ctrlb, ctrl, bp, match_len, match_off);
8397 bruce 641 957484232 : while (match_len--)
642 : {
7813 tgl 643 917744497 : pglz_hist_add(hist_start, hist_entries,
644 : hist_next, hist_recycle,
645 : dp, dend, mask);
5511 646 917744497 : dp++; /* Do not do this ++ in the line above! */
647 : /* The macro would do it four times - Jan. */
648 : }
649 39739735 : found_match = true;
650 : }
651 : else
652 : {
653 : /*
654 : * No match found. Copy one literal byte.
655 : */
8544 JanWieck 656 94065371 : pglz_out_literal(ctrlp, ctrlb, ctrl, bp, *dp);
7813 tgl 657 94065371 : pglz_hist_add(hist_start, hist_entries,
658 : hist_next, hist_recycle,
659 : dp, dend, mask);
5511 660 94065371 : 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 : */
6030 669 67386 : *ctrlp = ctrlb;
670 67386 : result_size = bp - bstart;
671 67386 : if (result_size >= result_max)
2981 fujii 672 UBC 0 : return -1;
673 :
674 : /* success */
2981 fujii 675 CBC 67386 : 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 86427 : 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 86427 : sp = (const unsigned char *) source;
701 86427 : srcend = ((const unsigned char *) source) + slen;
5510 tgl 702 86427 : dp = (unsigned char *) dest;
2981 fujii 703 86427 : destend = dp + rawsize;
704 :
5510 tgl 705 22401029 : 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 22314602 : unsigned char ctrl = *sp++;
712 : int ctrlc;
713 :
1468 sfrost 714 200503145 : for (ctrlc = 0; ctrlc < 8 && sp < srcend && dp < destend; ctrlc++)
715 : {
8544 JanWieck 716 178188543 : 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 :
5510 tgl 730 55073229 : len = (sp[0] & 0x0f) + 3;
731 55073229 : off = ((sp[0] & 0xf0) << 4) | sp[1];
732 55073229 : sp += 2;
8544 JanWieck 733 55073229 : if (len == 18)
5510 tgl 734 14456377 : 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 : */
888 742 55073229 : if (unlikely(sp > srcend || off == 0))
888 tgl 743 UBC 0 : return -1;
744 :
745 : /*
746 : * Don't emit more data than requested.
747 : */
1468 sfrost 748 CBC 55073229 : 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 : */
1228 tomas.vondra 759 58021587 : 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 2948358 : memcpy(dp, dp - off, off);
766 2948358 : len -= off;
767 2948358 : 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 2948358 : off += off;
794 : }
795 55073229 : memcpy(dp, dp - off, len);
796 55073229 : 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 : */
5510 tgl 804 123115314 : *dp++ = *sp++;
805 : }
806 :
807 : /*
808 : * Advance the control bit
809 : */
8544 JanWieck 810 178188543 : ctrl >>= 1;
811 : }
812 : }
813 :
814 : /*
815 : * If requested, check we decompressed the right amount.
816 : */
1468 sfrost 817 86427 : if (check_complete && (dp != destend || sp != srcend))
2981 fujii 818 UBC 0 : return -1;
819 :
820 : /*
821 : * That's it.
822 : */
1418 tgl 823 CBC 86427 : return (char *) dp - dest;
824 : }
825 :
826 :
827 : /* ----------
828 : * pglz_maximum_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
1286 tomas.vondra 842 18 : 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 : */
889 tgl 853 18 : 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 18 : 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 : */
1286 tomas.vondra 869 18 : compressed_size = Min(compressed_size, total_compressed_size);
870 :
889 tgl 871 18 : return (int32) compressed_size;
872 : }
|
