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, char *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 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.
31 * pglz_decompress(const char *source, int32 slen, char *dest,
34 * source is the compressed input.
36 * slen is the length of the compressed input.
38 * dest is the area where the uncompressed data will be
39 * written to. It is the callers responsibility to
40 * provide enough space.
42 * The data is written to buff exactly as it was handed
43 * to pglz_compress(). No terminating zero byte is added.
45 * rawsize is the length of the uncompressed data.
47 * The return value is the number of bytes written in the
48 * buffer dest, or -1 if decompression fails.
50 * The decompression algorithm and internal data format:
52 * It is made with the compressed data itself.
54 * The data representation is easiest explained by describing
55 * the process of decompression.
57 * If compressed_size == rawsize, then the data
58 * is stored uncompressed as plain bytes. Thus, the decompressor
59 * simply copies rawsize bytes to the destination.
61 * Otherwise the first byte tells what to do the next 8 times.
62 * We call this the control byte.
64 * An unset bit in the control byte means, that one uncompressed
65 * byte follows, which is copied from input to output.
67 * A set bit in the control byte means, that a tag of 2-3 bytes
68 * follows. A tag contains information to copy some bytes, that
69 * are already in the output buffer, to the current location in
70 * the output. Let's call the three tag bytes T1, T2 and T3. The
71 * position of the data to copy is coded as an offset from the
72 * actual output position.
74 * The offset is in the upper nibble of T1 and in T2.
75 * The length is in the lower nibble of T1.
77 * So the 16 bits of a 2 byte tag are coded as
82 * This limits the offset to 1-4095 (12 bits) and the length
83 * to 3-18 (4 bits) because 3 is always added to it. To emit
84 * a tag of 2 bytes with a length of 2 only saves one control
85 * bit. But we lose one byte in the possible length of a tag.
87 * In the actual implementation, the 2 byte tag's length is
88 * limited to 3-17, because the value 0xF in the length nibble
89 * has special meaning. It means, that the next following
90 * byte (T3) has to be added to the length value of 18. That
91 * makes total limits of 1-4095 for offset and 3-273 for length.
93 * Now that we have successfully decoded a tag. We simply copy
94 * the output that occurred <offset> bytes back to the current
95 * output location in the specified <length>. Thus, a
96 * sequence of 200 spaces (think about bpchar fields) could be
97 * coded in 4 bytes. One literal space and a three byte tag to
98 * copy 199 bytes with a -1 offset. Whow - that's a compression
99 * rate of 98%! Well, the implementation needs to save the
100 * original data size too, so we need another 4 bytes for it
101 * and end up with a total compression rate of 96%, what's still
104 * The compression algorithm
106 * The following uses numbers used in the default strategy.
108 * The compressor works best for attributes of a size between
109 * 1K and 1M. For smaller items there's not that much chance of
110 * redundancy in the character sequence (except for large areas
111 * of identical bytes like trailing spaces) and for bigger ones
112 * our 4K maximum look-back distance is too small.
114 * The compressor creates a table for lists of positions.
115 * For each input position (except the last 3), a hash key is
116 * built from the 4 next input bytes and the position remembered
117 * in the appropriate list. Thus, the table points to linked
118 * lists of likely to be at least in the first 4 characters
119 * matching strings. This is done on the fly while the input
120 * is compressed into the output area. Table entries are only
121 * kept for the last 4096 input positions, since we cannot use
122 * back-pointers larger than that anyway. The size of the hash
123 * table is chosen based on the size of the input - a larger table
124 * has a larger startup cost, as it needs to be initialized to
125 * zero, but reduces the number of hash collisions on long inputs.
127 * For each byte in the input, its hash key (built from this
128 * byte and the next 3) is used to find the appropriate list
129 * in the table. The lists remember the positions of all bytes
130 * that had the same hash key in the past in increasing backward
131 * offset order. Now for all entries in the used lists, the
132 * match length is computed by comparing the characters from the
133 * entries position with the characters from the actual input
136 * The compressor starts with a so called "good_match" of 128.
137 * It is a "prefer speed against compression ratio" optimizer.
138 * So if the first entry looked at already has 128 or more
139 * matching characters, the lookup stops and that position is
140 * used for the next tag in the output.
142 * For each subsequent entry in the history list, the "good_match"
143 * is lowered by 10%. So the compressor will be more happy with
144 * short matches the farer it has to go back in the history.
145 * Another "speed against ratio" preference characteristic of
148 * Thus there are 3 stop conditions for the lookup of matches:
150 * - a match >= good_match is found
151 * - there are no more history entries to look at
152 * - the next history entry is already too far back
153 * to be coded into a tag.
155 * Finally the match algorithm checks that at least a match
156 * of 3 or more bytes has been found, because thats the smallest
157 * amount of copy information to code into a tag. If so, a tag
158 * is omitted and all the input bytes covered by that are just
159 * scanned for the history add's, otherwise a literal character
160 * is omitted and only his history entry added.
164 * Many thanks to Adisak Pochanayon, who's article about SLZ
165 * inspired me to write the PostgreSQL compression this way.
169 * Copyright (c) 1999-2016, PostgreSQL Global Development Group
171 * src/common/pg_lzcompress.c
175 #include "postgres.h"
177 #include "postgres_fe.h"
182 #include "common/pg_lzcompress.h"
189 #define PGLZ_MAX_HISTORY_LISTS 8192 /* must be power of 2 */
190 #define PGLZ_HISTORY_SIZE 4096
191 #define PGLZ_MAX_MATCH 273
197 * Linked list for the backward history lookup
199 * All the entries sharing a hash key are linked in a doubly linked list.
200 * This makes it easy to remove an entry when it's time to recycle it
201 * (because it's more than 4K positions old).
204 typedef struct PGLZ_HistEntry
206 struct PGLZ_HistEntry *next; /* links for my hash key's list */
207 struct PGLZ_HistEntry *prev;
208 int hindex; /* my current hash key */
209 const char *pos; /* my input position */
214 * The provided standard strategies
217 static const PGLZ_Strategy strategy_default_data = {
218 32, /* Data chunks less than 32 bytes are not
220 INT_MAX, /* No upper limit on what we'll try to
222 25, /* Require 25% compression rate, or not worth
224 1024, /* Give up if no compression in the first 1KB */
225 128, /* Stop history lookup if a match of 128 bytes
227 10 /* Lower good match size by 10% at every loop
230 const PGLZ_Strategy *const PGLZ_strategy_default = &strategy_default_data;
233 static const PGLZ_Strategy strategy_always_data = {
234 0, /* Chunks of any size are compressed */
236 0, /* It's enough to save one single byte */
237 INT_MAX, /* Never give up early */
238 128, /* Stop history lookup if a match of 128 bytes
240 6 /* Look harder for a good match */
242 const PGLZ_Strategy *const PGLZ_strategy_always = &strategy_always_data;
246 * Statically allocated work arrays for history
249 static int16 hist_start[PGLZ_MAX_HISTORY_LISTS];
250 static PGLZ_HistEntry hist_entries[PGLZ_HISTORY_SIZE + 1];
253 * Element 0 in hist_entries is unused, and means 'invalid'. Likewise,
254 * INVALID_ENTRY_PTR in next/prev pointers mean 'invalid'.
256 #define INVALID_ENTRY 0
257 #define INVALID_ENTRY_PTR (&hist_entries[INVALID_ENTRY])
262 * Computes the history table slot for the lookup by the next 4
263 * characters in the input.
265 * NB: because we use the next 4 characters, we are not guaranteed to
266 * find 3-character matches; they very possibly will be in the wrong
267 * hash list. This seems an acceptable tradeoff for spreading out the
271 #define pglz_hist_idx(_s,_e, _mask) ( \
272 ((((_e) - (_s)) < 4) ? (int) (_s)[0] : \
273 (((_s)[0] << 6) ^ ((_s)[1] << 4) ^ \
274 ((_s)[2] << 2) ^ (_s)[3])) & (_mask) \
281 * Adds a new entry to the history table.
283 * If _recycle is true, then we are recycling a previously used entry,
284 * and must first delink it from its old hashcode's linked list.
286 * NOTE: beware of multiple evaluations of macro's arguments, and note that
287 * _hn and _recycle are modified in the macro.
290 #define pglz_hist_add(_hs,_he,_hn,_recycle,_s,_e, _mask) \
292 int __hindex = pglz_hist_idx((_s),(_e), (_mask)); \
293 int16 *__myhsp = &(_hs)[__hindex]; \
294 PGLZ_HistEntry *__myhe = &(_he)[_hn]; \
296 if (__myhe->prev == NULL) \
297 (_hs)[__myhe->hindex] = __myhe->next - (_he); \
299 __myhe->prev->next = __myhe->next; \
300 if (__myhe->next != NULL) \
301 __myhe->next->prev = __myhe->prev; \
303 __myhe->next = &(_he)[*__myhsp]; \
304 __myhe->prev = NULL; \
305 __myhe->hindex = __hindex; \
306 __myhe->pos = (_s); \
307 /* If there was an existing entry in this hash slot, link */ \
308 /* this new entry to it. However, the 0th entry in the */ \
309 /* entries table is unused, so we can freely scribble on it. */ \
310 /* So don't bother checking if the slot was used - we'll */ \
311 /* scribble on the unused entry if it was not, but that's */ \
312 /* harmless. Avoiding the branch in this critical path */ \
313 /* speeds this up a little bit. */ \
314 /* if (*__myhsp != INVALID_ENTRY) */ \
315 (_he)[(*__myhsp)].prev = __myhe; \
317 if (++(_hn) >= PGLZ_HISTORY_SIZE + 1) { \
327 * Outputs the last and allocates a new control byte if needed.
330 #define pglz_out_ctrl(__ctrlp,__ctrlb,__ctrl,__buf) \
332 if ((__ctrl & 0xff) == 0) \
334 *(__ctrlp) = __ctrlb; \
335 __ctrlp = (__buf)++; \
345 * Outputs a literal byte to the destination buffer including the
346 * appropriate control bit.
349 #define pglz_out_literal(_ctrlp,_ctrlb,_ctrl,_buf,_byte) \
351 pglz_out_ctrl(_ctrlp,_ctrlb,_ctrl,_buf); \
352 *(_buf)++ = (unsigned char)(_byte); \
360 * Outputs a backward reference tag of 2-4 bytes (depending on
361 * offset and length) to the destination buffer including the
362 * appropriate control bit.
365 #define pglz_out_tag(_ctrlp,_ctrlb,_ctrl,_buf,_len,_off) \
367 pglz_out_ctrl(_ctrlp,_ctrlb,_ctrl,_buf); \
372 (_buf)[0] = (unsigned char)((((_off) & 0xf00) >> 4) | 0x0f); \
373 (_buf)[1] = (unsigned char)(((_off) & 0xff)); \
374 (_buf)[2] = (unsigned char)((_len) - 18); \
377 (_buf)[0] = (unsigned char)((((_off) & 0xf00) >> 4) | ((_len) - 3)); \
378 (_buf)[1] = (unsigned char)((_off) & 0xff); \
387 * Lookup the history table if the actual input stream matches
388 * another sequence of characters, starting somewhere earlier
389 * in the input buffer.
393 pglz_find_match(int16 *hstart, const char *input, const char *end,
394 int *lenp, int *offp, int good_match, int good_drop, int mask)
396 PGLZ_HistEntry *hent;
402 * Traverse the linked history list until a good enough match is found.
404 hentno = hstart[pglz_hist_idx(input, end, mask)];
405 hent = &hist_entries[hentno];
406 while (hent != INVALID_ENTRY_PTR)
408 const char *ip = input;
409 const char *hp = hent->pos;
414 * Stop if the offset does not fit into our tag anymore.
417 if (thisoff >= 0x0fff)
421 * Determine length of match. A better match must be larger than the
422 * best so far. And if we already have a match of 16 or more bytes,
423 * it's worth the call overhead to use memcmp() to check if this match
424 * is equal for the same size. After that we must fallback to
425 * character by character comparison to know the exact position where
431 if (memcmp(ip, hp, len) == 0)
436 while (ip < end && *ip == *hp && thislen < PGLZ_MAX_MATCH)
446 while (ip < end && *ip == *hp && thislen < PGLZ_MAX_MATCH)
455 * Remember this match as the best (if it is)
464 * Advance to the next history entry
469 * Be happy with lesser good matches the more entries we visited. But
470 * no point in doing calculation if we're at end of list.
472 if (hent != INVALID_ENTRY_PTR)
474 if (len >= good_match)
476 good_match -= (good_match * good_drop) / 100;
481 * Return match information only if it results at least in one byte
498 * Compresses source into dest using strategy. Returns the number of
499 * bytes written in buffer dest, or -1 if compression fails.
503 pglz_compress(const char *source, int32 slen, char *dest,
504 const PGLZ_Strategy *strategy)
506 unsigned char *bp = (unsigned char *) dest;
507 unsigned char *bstart = bp;
509 bool hist_recycle = false;
510 const char *dp = source;
511 const char *dend = source + slen;
512 unsigned char ctrl_dummy = 0;
513 unsigned char *ctrlp = &ctrl_dummy;
514 unsigned char ctrlb = 0;
515 unsigned char ctrl = 0;
516 bool found_match = false;
528 * Our fallback strategy is the default.
530 if (strategy == NULL)
531 strategy = PGLZ_strategy_default;
534 * If the strategy forbids compression (at all or if source chunk size out
537 if (strategy->match_size_good <= 0 ||
538 slen < strategy->min_input_size ||
539 slen > strategy->max_input_size)
543 * Limit the match parameters to the supported range.
545 good_match = strategy->match_size_good;
546 if (good_match > PGLZ_MAX_MATCH)
547 good_match = PGLZ_MAX_MATCH;
548 else if (good_match < 17)
551 good_drop = strategy->match_size_drop;
554 else if (good_drop > 100)
557 need_rate = strategy->min_comp_rate;
560 else if (need_rate > 99)
564 * Compute the maximum result size allowed by the strategy, namely the
565 * input size minus the minimum wanted compression rate. This had better
566 * be <= slen, else we might overrun the provided output buffer.
568 if (slen > (INT_MAX / 100))
570 /* Approximate to avoid overflow */
571 result_max = (slen / 100) * (100 - need_rate);
574 result_max = (slen * (100 - need_rate)) / 100;
577 * Experiments suggest that these hash sizes work pretty well. A large
578 * hash table minimizes collision, but has a higher startup cost. For a
579 * small input, the startup cost dominates. The table size must be a power
588 else if (slen < 1024)
595 * Initialize the history lists to empty. We do not need to zero the
596 * hist_entries[] array; its entries are initialized as they are used.
598 memset(hist_start, 0, hashsz * sizeof(int16));
601 * Compress the source directly into the output buffer.
606 * If we already exceeded the maximum result size, fail.
608 * We check once per loop; since the loop body could emit as many as 4
609 * bytes (a control byte and 3-byte tag), PGLZ_MAX_OUTPUT() had better
610 * allow 4 slop bytes.
612 if (bp - bstart >= result_max)
616 * If we've emitted more than first_success_by bytes without finding
617 * anything compressible at all, fail. This lets us fall out
618 * reasonably quickly when looking at incompressible input (such as
619 * pre-compressed data).
621 if (!found_match && bp - bstart >= strategy->first_success_by)
625 * Try to find a match in the history
627 if (pglz_find_match(hist_start, dp, dend, &match_len,
628 &match_off, good_match, good_drop, mask))
631 * Create the tag and add history entries for all matched
634 pglz_out_tag(ctrlp, ctrlb, ctrl, bp, match_len, match_off);
637 pglz_hist_add(hist_start, hist_entries,
638 hist_next, hist_recycle,
640 dp++; /* Do not do this ++ in the line above! */
641 /* The macro would do it four times - Jan. */
648 * No match found. Copy one literal byte.
650 pglz_out_literal(ctrlp, ctrlb, ctrl, bp, *dp);
651 pglz_hist_add(hist_start, hist_entries,
652 hist_next, hist_recycle,
654 dp++; /* Do not do this ++ in the line above! */
655 /* The macro would do it four times - Jan. */
660 * Write out the last control byte and check that we haven't overrun the
661 * output size allowed by the strategy.
664 result_size = bp - bstart;
665 if (result_size >= result_max)
676 * Decompresses source into dest. Returns the number of bytes
677 * decompressed in the destination buffer, or -1 if decompression
682 pglz_decompress(const char *source, int32 slen, char *dest,
685 const unsigned char *sp;
686 const unsigned char *srcend;
688 unsigned char *destend;
690 sp = (const unsigned char *) source;
691 srcend = ((const unsigned char *) source) + slen;
692 dp = (unsigned char *) dest;
693 destend = dp + rawsize;
695 while (sp < srcend && dp < destend)
698 * Read one control byte and process the next 8 items (or as many as
699 * remain in the compressed input).
701 unsigned char ctrl = *sp++;
704 for (ctrlc = 0; ctrlc < 8 && sp < srcend; ctrlc++)
709 * Otherwise it contains the match length minus 3 and the
710 * upper 4 bits of the offset. The next following byte
711 * contains the lower 8 bits of the offset. If the length is
712 * coded as 18, another extension tag byte tells how much
713 * longer the match really was (0-255).
718 len = (sp[0] & 0x0f) + 3;
719 off = ((sp[0] & 0xf0) << 4) | sp[1];
725 * Check for output buffer overrun, to ensure we don't clobber
726 * memory in case of corrupt input. Note: we must advance dp
727 * here to ensure the error is detected below the loop. We
728 * don't simply put the elog inside the loop since that will
729 * probably interfere with optimization.
731 if (dp + len > destend)
738 * Now we copy the bytes specified by the tag from OUTPUT to
739 * OUTPUT. It is dangerous and platform dependent to use
740 * memcpy() here, because the copied areas could overlap
752 * An unset control bit means LITERAL BYTE. So we just copy
753 * one from INPUT to OUTPUT.
755 if (dp >= destend) /* check for buffer overrun */
756 break; /* do not clobber memory */
762 * Advance the control bit
769 * Check we decompressed the right amount.
771 if (dp != destend || sp != srcend)