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 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
pglz_find_match(int16 * hstart,const char * input,const char * end,int * lenp,int * offp,int good_match,int good_drop,int mask)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
pglz_compress(const char * source,int32 slen,char * dest,const PGLZ_Strategy * strategy)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
pglz_decompress(const char * source,int32 slen,char * dest,int32 rawsize,bool check_complete)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
pglz_maximum_compressed_size(int32 rawsize,int32 total_compressed_size)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