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.
12 * pglz_compress(const char *source, int32 slen, PGLZ_Header *dest,
13 * const PGLZ_Strategy *strategy);
15 * source is the input data to be compressed.
17 * slen is the length of the input data.
19 * dest is the output area for the compressed result.
20 * It must be at least as big as PGLZ_MAX_OUTPUT(slen).
22 * strategy is a pointer to some information controlling
23 * the compression algorithm. If NULL, the compiled
24 * in default strategy is used.
26 * The return value is TRUE if compression succeeded,
27 * FALSE if not; in the latter case the contents of dest
31 * pglz_decompress(const PGLZ_Header *source, char *dest)
33 * source is the compressed input.
35 * dest is the area where the uncompressed data will be
36 * written to. It is the callers responsibility to
37 * provide enough space. The required amount can be
38 * obtained with the macro PGLZ_RAW_SIZE(source).
40 * The data is written to buff exactly as it was handed
41 * to pglz_compress(). No terminating zero byte is added.
43 * The decompression algorithm and internal data format:
45 * PGLZ_Header is defined as
47 * typedef struct PGLZ_Header {
52 * The header is followed by the compressed data itself.
54 * The data representation is easiest explained by describing
55 * the process of decompression.
57 * If VARSIZE(x) == rawsize + sizeof(PGLZ_Header), then the data
58 * is stored uncompressed as plain bytes. Thus, the decompressor
59 * simply copies rawsize bytes from the location after the
60 * header to the destination.
62 * Otherwise the first byte after the header tells what to do
63 * the next 8 times. We call this the control byte.
65 * An unset bit in the control byte means, that one uncompressed
66 * byte follows, which is copied from input to output.
68 * A set bit in the control byte means, that a tag of 2-3 bytes
69 * follows. A tag contains information to copy some bytes, that
70 * are already in the output buffer, to the current location in
71 * the output. Let's call the three tag bytes T1, T2 and T3. The
72 * position of the data to copy is coded as an offset from the
73 * actual output position.
75 * The offset is in the upper nibble of T1 and in T2.
76 * The length is in the lower nibble of T1.
78 * So the 16 bits of a 2 byte tag are coded as
83 * This limits the offset to 1-4095 (12 bits) and the length
84 * to 3-18 (4 bits) because 3 is always added to it. To emit
85 * a tag of 2 bytes with a length of 2 only saves one control
86 * bit. But we lose one byte in the possible length of a tag.
88 * In the actual implementation, the 2 byte tag's length is
89 * limited to 3-17, because the value 0xF in the length nibble
90 * has special meaning. It means, that the next following
91 * byte (T3) has to be added to the length value of 18. That
92 * makes total limits of 1-4095 for offset and 3-273 for length.
94 * Now that we have successfully decoded a tag. We simply copy
95 * the output that occurred <offset> bytes back to the current
96 * output location in the specified <length>. Thus, a
97 * sequence of 200 spaces (think about bpchar fields) could be
98 * coded in 4 bytes. One literal space and a three byte tag to
99 * copy 199 bytes with a -1 offset. Whow - that's a compression
100 * rate of 98%! Well, the implementation needs to save the
101 * original data size too, so we need another 4 bytes for it
102 * and end up with a total compression rate of 96%, what's still
105 * The compression algorithm
107 * The following uses numbers used in the default strategy.
109 * The compressor works best for attributes of a size between
110 * 1K and 1M. For smaller items there's not that much chance of
111 * redundancy in the character sequence (except for large areas
112 * of identical bytes like trailing spaces) and for bigger ones
113 * our 4K maximum look-back distance is too small.
115 * The compressor creates a table for lists of positions.
116 * For each input position (except the last 3), a hash key is
117 * built from the 4 next input bytes and the position remembered
118 * in the appropriate list. Thus, the table points to linked
119 * lists of likely to be at least in the first 4 characters
120 * matching strings. This is done on the fly while the input
121 * is compressed into the output area. Table entries are only
122 * kept for the last 4096 input positions, since we cannot use
123 * back-pointers larger than that anyway. The size of the hash
124 * table is chosen based on the size of the input - a larger table
125 * has a larger startup cost, as it needs to be initialized to
126 * zero, but reduces the number of hash collisions on long inputs.
128 * For each byte in the input, its hash key (built from this
129 * byte and the next 3) is used to find the appropriate list
130 * in the table. The lists remember the positions of all bytes
131 * that had the same hash key in the past in increasing backward
132 * offset order. Now for all entries in the used lists, the
133 * match length is computed by comparing the characters from the
134 * entries position with the characters from the actual input
137 * The compressor starts with a so called "good_match" of 128.
138 * It is a "prefer speed against compression ratio" optimizer.
139 * So if the first entry looked at already has 128 or more
140 * matching characters, the lookup stops and that position is
141 * used for the next tag in the output.
143 * For each subsequent entry in the history list, the "good_match"
144 * is lowered by 10%. So the compressor will be more happy with
145 * short matches the farer it has to go back in the history.
146 * Another "speed against ratio" preference characteristic of
149 * Thus there are 3 stop conditions for the lookup of matches:
151 * - a match >= good_match is found
152 * - there are no more history entries to look at
153 * - the next history entry is already too far back
154 * to be coded into a tag.
156 * Finally the match algorithm checks that at least a match
157 * of 3 or more bytes has been found, because thats the smallest
158 * amount of copy information to code into a tag. If so, a tag
159 * is omitted and all the input bytes covered by that are just
160 * scanned for the history add's, otherwise a literal character
161 * is omitted and only his history entry added.
165 * Many thanks to Adisak Pochanayon, who's article about SLZ
166 * inspired me to write the PostgreSQL compression this way.
170 * Copyright (c) 1999-2013, PostgreSQL Global Development Group
172 * src/backend/utils/adt/pg_lzcompress.c
175 #include "postgres.h"
179 #include "utils/pg_lzcompress.h"
186 #define PGLZ_MAX_HISTORY_LISTS 8192 /* must be power of 2 */
187 #define PGLZ_HISTORY_SIZE 4096
188 #define PGLZ_MAX_MATCH 273
194 * Linked list for the backward history lookup
196 * All the entries sharing a hash key are linked in a doubly linked list.
197 * This makes it easy to remove an entry when it's time to recycle it
198 * (because it's more than 4K positions old).
201 typedef struct PGLZ_HistEntry
203 struct PGLZ_HistEntry *next; /* links for my hash key's list */
204 struct PGLZ_HistEntry *prev;
205 int hindex; /* my current hash key */
206 const char *pos; /* my input position */
211 * The provided standard strategies
214 static const PGLZ_Strategy strategy_default_data = {
215 32, /* Data chunks less than 32 bytes are not
217 INT_MAX, /* No upper limit on what we'll try to
219 25, /* Require 25% compression rate, or not worth
221 1024, /* Give up if no compression in the first 1KB */
222 128, /* Stop history lookup if a match of 128 bytes
224 10 /* Lower good match size by 10% at every loop
227 const PGLZ_Strategy *const PGLZ_strategy_default = &strategy_default_data;
230 static const PGLZ_Strategy strategy_always_data = {
231 0, /* Chunks of any size are compressed */
233 0, /* It's enough to save one single byte */
234 INT_MAX, /* Never give up early */
235 128, /* Stop history lookup if a match of 128 bytes
237 6 /* Look harder for a good match */
239 const PGLZ_Strategy *const PGLZ_strategy_always = &strategy_always_data;
243 * Statically allocated work arrays for history
246 static int16 hist_start[PGLZ_MAX_HISTORY_LISTS];
247 static PGLZ_HistEntry hist_entries[PGLZ_HISTORY_SIZE + 1];
250 * Element 0 in hist_entries is unused, and means 'invalid'. Likewise,
251 * INVALID_ENTRY_PTR in next/prev pointers mean 'invalid'.
253 #define INVALID_ENTRY 0
254 #define INVALID_ENTRY_PTR (&hist_entries[INVALID_ENTRY])
259 * Computes the history table slot for the lookup by the next 4
260 * characters in the input.
262 * NB: because we use the next 4 characters, we are not guaranteed to
263 * find 3-character matches; they very possibly will be in the wrong
264 * hash list. This seems an acceptable tradeoff for spreading out the
268 #define pglz_hist_idx(_s,_e, _mask) ( \
269 ((((_e) - (_s)) < 4) ? (int) (_s)[0] : \
270 (((_s)[0] << 6) ^ ((_s)[1] << 4) ^ \
271 ((_s)[2] << 2) ^ (_s)[3])) & (_mask) \
278 * Adds a new entry to the history table.
280 * If _recycle is true, then we are recycling a previously used entry,
281 * and must first delink it from its old hashcode's linked list.
283 * NOTE: beware of multiple evaluations of macro's arguments, and note that
284 * _hn and _recycle are modified in the macro.
287 #define pglz_hist_add(_hs,_he,_hn,_recycle,_s,_e, _mask) \
289 int __hindex = pglz_hist_idx((_s),(_e), (_mask)); \
290 int16 *__myhsp = &(_hs)[__hindex]; \
291 PGLZ_HistEntry *__myhe = &(_he)[_hn]; \
293 if (__myhe->prev == NULL) \
294 (_hs)[__myhe->hindex] = __myhe->next - (_he); \
296 __myhe->prev->next = __myhe->next; \
297 if (__myhe->next != NULL) \
298 __myhe->next->prev = __myhe->prev; \
300 __myhe->next = &(_he)[*__myhsp]; \
301 __myhe->prev = NULL; \
302 __myhe->hindex = __hindex; \
303 __myhe->pos = (_s); \
304 /* If there was an existing entry in this hash slot, link */ \
305 /* this new entry to it. However, the 0th entry in the */ \
306 /* entries table is unused, so we can freely scribble on it. */ \
307 /* So don't bother checking if the slot was used - we'll */ \
308 /* scribble on the unused entry if it was not, but that's */ \
309 /* harmless. Avoiding the branch in this critical path */ \
310 /* speeds this up a little bit. */ \
311 /* if (*__myhsp != INVALID_ENTRY) */ \
312 (_he)[(*__myhsp)].prev = __myhe; \
314 if (++(_hn) >= PGLZ_HISTORY_SIZE + 1) { \
324 * Outputs the last and allocates a new control byte if needed.
327 #define pglz_out_ctrl(__ctrlp,__ctrlb,__ctrl,__buf) \
329 if ((__ctrl & 0xff) == 0) \
331 *(__ctrlp) = __ctrlb; \
332 __ctrlp = (__buf)++; \
342 * Outputs a literal byte to the destination buffer including the
343 * appropriate control bit.
346 #define pglz_out_literal(_ctrlp,_ctrlb,_ctrl,_buf,_byte) \
348 pglz_out_ctrl(_ctrlp,_ctrlb,_ctrl,_buf); \
349 *(_buf)++ = (unsigned char)(_byte); \
357 * Outputs a backward reference tag of 2-4 bytes (depending on
358 * offset and length) to the destination buffer including the
359 * appropriate control bit.
362 #define pglz_out_tag(_ctrlp,_ctrlb,_ctrl,_buf,_len,_off) \
364 pglz_out_ctrl(_ctrlp,_ctrlb,_ctrl,_buf); \
369 (_buf)[0] = (unsigned char)((((_off) & 0xf00) >> 4) | 0x0f); \
370 (_buf)[1] = (unsigned char)(((_off) & 0xff)); \
371 (_buf)[2] = (unsigned char)((_len) - 18); \
374 (_buf)[0] = (unsigned char)((((_off) & 0xf00) >> 4) | ((_len) - 3)); \
375 (_buf)[1] = (unsigned char)((_off) & 0xff); \
384 * Lookup the history table if the actual input stream matches
385 * another sequence of characters, starting somewhere earlier
386 * in the input buffer.
390 pglz_find_match(int16 *hstart, const char *input, const char *end,
391 int *lenp, int *offp, int good_match, int good_drop, int mask)
393 PGLZ_HistEntry *hent;
399 * Traverse the linked history list until a good enough match is found.
401 hentno = hstart[pglz_hist_idx(input, end, mask)];
402 hent = &hist_entries[hentno];
403 while (hent != INVALID_ENTRY_PTR)
405 const char *ip = input;
406 const char *hp = hent->pos;
411 * Stop if the offset does not fit into our tag anymore.
414 if (thisoff >= 0x0fff)
418 * Determine length of match. A better match must be larger than the
419 * best so far. And if we already have a match of 16 or more bytes,
420 * it's worth the call overhead to use memcmp() to check if this match
421 * is equal for the same size. After that we must fallback to
422 * character by character comparison to know the exact position where
428 if (memcmp(ip, hp, len) == 0)
433 while (ip < end && *ip == *hp && thislen < PGLZ_MAX_MATCH)
443 while (ip < end && *ip == *hp && thislen < PGLZ_MAX_MATCH)
452 * Remember this match as the best (if it is)
461 * Advance to the next history entry
466 * Be happy with lesser good matches the more entries we visited. But
467 * no point in doing calculation if we're at end of list.
469 if (hent != INVALID_ENTRY_PTR)
471 if (len >= good_match)
473 good_match -= (good_match * good_drop) / 100;
478 * Return match information only if it results at least in one byte
495 * Compresses source into dest using strategy.
499 pglz_compress(const char *source, int32 slen, PGLZ_Header *dest,
500 const PGLZ_Strategy *strategy)
502 unsigned char *bp = ((unsigned char *) dest) + sizeof(PGLZ_Header);
503 unsigned char *bstart = bp;
505 bool hist_recycle = false;
506 const char *dp = source;
507 const char *dend = source + slen;
508 unsigned char ctrl_dummy = 0;
509 unsigned char *ctrlp = &ctrl_dummy;
510 unsigned char ctrlb = 0;
511 unsigned char ctrl = 0;
512 bool found_match = false;
524 * Our fallback strategy is the default.
526 if (strategy == NULL)
527 strategy = PGLZ_strategy_default;
530 * If the strategy forbids compression (at all or if source chunk size out
533 if (strategy->match_size_good <= 0 ||
534 slen < strategy->min_input_size ||
535 slen > strategy->max_input_size)
539 * Save the original source size in the header.
541 dest->rawsize = slen;
544 * Limit the match parameters to the supported range.
546 good_match = strategy->match_size_good;
547 if (good_match > PGLZ_MAX_MATCH)
548 good_match = PGLZ_MAX_MATCH;
549 else if (good_match < 17)
552 good_drop = strategy->match_size_drop;
555 else if (good_drop > 100)
558 need_rate = strategy->min_comp_rate;
561 else if (need_rate > 99)
565 * Compute the maximum result size allowed by the strategy, namely the
566 * input size minus the minimum wanted compression rate. This had better
567 * be <= slen, else we might overrun the provided output buffer.
569 if (slen > (INT_MAX / 100))
571 /* Approximate to avoid overflow */
572 result_max = (slen / 100) * (100 - need_rate);
575 result_max = (slen * (100 - need_rate)) / 100;
578 * Experiments suggest that these hash sizes work pretty well. A large
579 * hash table minimizes collision, but has a higher startup cost. For
580 * a small input, the startup cost dominates. The table size must be
589 else if (slen < 1024)
596 * Initialize the history lists to empty. We do not need to zero the
597 * hist_entries[] array; its entries are initialized as they are used.
599 memset(hist_start, 0, hashsz * sizeof(int16));
602 * Compress the source directly into the output buffer.
607 * If we already exceeded the maximum result size, fail.
609 * We check once per loop; since the loop body could emit as many as 4
610 * bytes (a control byte and 3-byte tag), PGLZ_MAX_OUTPUT() had better
611 * allow 4 slop bytes.
613 if (bp - bstart >= result_max)
617 * If we've emitted more than first_success_by bytes without finding
618 * anything compressible at all, fail. This lets us fall out
619 * reasonably quickly when looking at incompressible input (such as
620 * pre-compressed data).
622 if (!found_match && bp - bstart >= strategy->first_success_by)
626 * Try to find a match in the history
628 if (pglz_find_match(hist_start, dp, dend, &match_len,
629 &match_off, good_match, good_drop, mask))
632 * Create the tag and add history entries for all matched
635 pglz_out_tag(ctrlp, ctrlb, ctrl, bp, match_len, match_off);
638 pglz_hist_add(hist_start, hist_entries,
639 hist_next, hist_recycle,
641 dp++; /* Do not do this ++ in the line above! */
642 /* The macro would do it four times - Jan. */
649 * No match found. Copy one literal byte.
651 pglz_out_literal(ctrlp, ctrlb, ctrl, bp, *dp);
652 pglz_hist_add(hist_start, hist_entries,
653 hist_next, hist_recycle,
655 dp++; /* Do not do this ++ in the line above! */
656 /* The macro would do it four times - Jan. */
661 * Write out the last control byte and check that we haven't overrun the
662 * output size allowed by the strategy.
665 result_size = bp - bstart;
666 if (result_size >= result_max)
670 * Success - need only fill in the actual length of the compressed datum.
672 SET_VARSIZE_COMPRESSED(dest, result_size + sizeof(PGLZ_Header));
681 * Decompresses source into dest.
685 pglz_decompress(const PGLZ_Header *source, char *dest)
687 const unsigned char *sp;
688 const unsigned char *srcend;
690 unsigned char *destend;
692 sp = ((const unsigned char *) source) + sizeof(PGLZ_Header);
693 srcend = ((const unsigned char *) source) + VARSIZE(source);
694 dp = (unsigned char *) dest;
695 destend = dp + source->rawsize;
697 while (sp < srcend && dp < destend)
700 * Read one control byte and process the next 8 items (or as many as
701 * remain in the compressed input).
703 unsigned char ctrl = *sp++;
706 for (ctrlc = 0; ctrlc < 8 && sp < srcend; ctrlc++)
711 * Otherwise it contains the match length minus 3 and the
712 * upper 4 bits of the offset. The next following byte
713 * contains the lower 8 bits of the offset. If the length is
714 * coded as 18, another extension tag byte tells how much
715 * longer the match really was (0-255).
720 len = (sp[0] & 0x0f) + 3;
721 off = ((sp[0] & 0xf0) << 4) | sp[1];
727 * Check for output buffer overrun, to ensure we don't clobber
728 * memory in case of corrupt input. Note: we must advance dp
729 * here to ensure the error is detected below the loop. We
730 * don't simply put the elog inside the loop since that will
731 * probably interfere with optimization.
733 if (dp + len > destend)
740 * Now we copy the bytes specified by the tag from OUTPUT to
741 * OUTPUT. It is dangerous and platform dependent to use
742 * memcpy() here, because the copied areas could overlap
754 * An unset control bit means LITERAL BYTE. So we just copy
755 * one from INPUT to OUTPUT.
757 if (dp >= destend) /* check for buffer overrun */
758 break; /* do not clobber memory */
764 * Advance the control bit
771 * Check we decompressed the right amount.
773 if (dp != destend || sp != srcend)
774 elog(ERROR, "compressed data is corrupt");