1 /* deflate.c -- compress data using the deflation algorithm
2 * Copyright (C) 1995-2003 Jean-loup Gailly.
3 * For conditions of distribution and use, see copyright notice in zlib.h
9 * The "deflation" process depends on being able to identify portions
10 * of the input text which are identical to earlier input (within a
11 * sliding window trailing behind the input currently being processed).
13 * The most straightforward technique turns out to be the fastest for
14 * most input files: try all possible matches and select the longest.
15 * The key feature of this algorithm is that insertions into the string
16 * dictionary are very simple and thus fast, and deletions are avoided
17 * completely. Insertions are performed at each input character, whereas
18 * string matches are performed only when the previous match ends. So it
19 * is preferable to spend more time in matches to allow very fast string
20 * insertions and avoid deletions. The matching algorithm for small
21 * strings is inspired from that of Rabin & Karp. A brute force approach
22 * is used to find longer strings when a small match has been found.
23 * A similar algorithm is used in comic (by Jan-Mark Wams) and freeze
24 * (by Leonid Broukhis).
25 * A previous version of this file used a more sophisticated algorithm
26 * (by Fiala and Greene) which is guaranteed to run in linear amortized
27 * time, but has a larger average cost, uses more memory and is patented.
28 * However the F&G algorithm may be faster for some highly redundant
29 * files if the parameter max_chain_length (described below) is too large.
33 * The idea of lazy evaluation of matches is due to Jan-Mark Wams, and
34 * I found it in 'freeze' written by Leonid Broukhis.
35 * Thanks to many people for bug reports and testing.
39 * Deutsch, L.P.,"DEFLATE Compressed Data Format Specification".
40 * Available in ftp://ds.internic.net/rfc/rfc1951.txt
42 * A description of the Rabin and Karp algorithm is given in the book
43 * "Algorithms" by R. Sedgewick, Addison-Wesley, p252.
45 * Fiala,E.R., and Greene,D.H.
46 * Data Compression with Finite Windows, Comm.ACM, 32,4 (1989) 490-595
54 const char deflate_copyright[] =
55 " deflate 1.2.0.1 Copyright 1995-2003 Jean-loup Gailly ";
57 If you use the zlib library in a product, an acknowledgment is welcome
58 in the documentation of your product. If for some reason you cannot
59 include such an acknowledgment, I would appreciate that you keep this
60 copyright string in the executable of your product.
63 /* ===========================================================================
64 * Function prototypes.
67 need_more, /* block not completed, need more input or more output */
68 block_done, /* block flush performed */
69 finish_started, /* finish started, need only more output at next deflate */
70 finish_done /* finish done, accept no more input or output */
73 typedef block_state (*compress_func) OF((deflate_state *s, int flush));
74 /* Compression function. Returns the block state after the call. */
76 local void fill_window OF((deflate_state *s));
77 local block_state deflate_stored OF((deflate_state *s, int flush));
78 local block_state deflate_fast OF((deflate_state *s, int flush));
80 local block_state deflate_slow OF((deflate_state *s, int flush));
82 local void lm_init OF((deflate_state *s));
83 local void putShortMSB OF((deflate_state *s, uInt b));
84 local void flush_pending OF((z_streamp strm));
85 local int read_buf OF((z_streamp strm, Bytef *buf, unsigned size));
88 void match_init OF((void)); /* asm code initialization */
89 uInt longest_match OF((deflate_state *s, IPos cur_match));
91 local uInt longest_match OF((deflate_state *s, IPos cur_match));
94 local uInt longest_match_fast OF((deflate_state *s, IPos cur_match));
97 local void check_match OF((deflate_state *s, IPos start, IPos match,
101 /* ===========================================================================
106 /* Tail of hash chains */
109 # define TOO_FAR 4096
111 /* Matches of length 3 are discarded if their distance exceeds TOO_FAR */
113 #define MIN_LOOKAHEAD (MAX_MATCH+MIN_MATCH+1)
114 /* Minimum amount of lookahead, except at the end of the input file.
115 * See deflate.c for comments about the MIN_MATCH+1.
118 /* Values for max_lazy_match, good_match and max_chain_length, depending on
119 * the desired pack level (0..9). The values given below have been tuned to
120 * exclude worst case performance for pathological files. Better values may be
121 * found for specific files.
123 typedef struct config_s {
124 ush good_length; /* reduce lazy search above this match length */
125 ush max_lazy; /* do not perform lazy search above this match length */
126 ush nice_length; /* quit search above this match length */
132 local const config configuration_table[2] = {
133 /* good lazy nice chain */
134 /* 0 */ {0, 0, 0, 0, deflate_stored}, /* store only */
135 /* 1 */ {4, 4, 8, 4, deflate_fast}}; /* maximum speed, no lazy matches */
137 local const config configuration_table[10] = {
138 /* good lazy nice chain */
139 /* 0 */ {0, 0, 0, 0, deflate_stored}, /* store only */
140 /* 1 */ {4, 4, 8, 4, deflate_fast}, /* maximum speed, no lazy matches */
141 /* 2 */ {4, 5, 16, 8, deflate_fast},
142 /* 3 */ {4, 6, 32, 32, deflate_fast},
144 /* 4 */ {4, 4, 16, 16, deflate_slow}, /* lazy matches */
145 /* 5 */ {8, 16, 32, 32, deflate_slow},
146 /* 6 */ {8, 16, 128, 128, deflate_slow},
147 /* 7 */ {8, 32, 128, 256, deflate_slow},
148 /* 8 */ {32, 128, 258, 1024, deflate_slow},
149 /* 9 */ {32, 258, 258, 4096, deflate_slow}}; /* maximum compression */
152 /* Note: the deflate() code requires max_lazy >= MIN_MATCH and max_chain >= 4
153 * For deflate_fast() (levels <= 3) good is ignored and lazy has a different
158 /* result of memcmp for equal strings */
160 struct static_tree_desc_s {int dummy;}; /* for buggy compilers */
162 /* ===========================================================================
163 * Update a hash value with the given input byte
164 * IN assertion: all calls to to UPDATE_HASH are made with consecutive
165 * input characters, so that a running hash key can be computed from the
166 * previous key instead of complete recalculation each time.
168 #define UPDATE_HASH(s,h,c) (h = (((h)<<s->hash_shift) ^ (c)) & s->hash_mask)
171 /* ===========================================================================
172 * Insert string str in the dictionary and set match_head to the previous head
173 * of the hash chain (the most recent string with same hash key). Return
174 * the previous length of the hash chain.
175 * If this file is compiled with -DFASTEST, the compression level is forced
176 * to 1, and no hash chains are maintained.
177 * IN assertion: all calls to to INSERT_STRING are made with consecutive
178 * input characters and the first MIN_MATCH bytes of str are valid
179 * (except for the last MIN_MATCH-1 bytes of the input file).
182 #define INSERT_STRING(s, str, match_head) \
183 (UPDATE_HASH(s, s->ins_h, s->window[(str) + (MIN_MATCH-1)]), \
184 match_head = s->head[s->ins_h], \
185 s->head[s->ins_h] = (Pos)(str))
187 #define INSERT_STRING(s, str, match_head) \
188 (UPDATE_HASH(s, s->ins_h, s->window[(str) + (MIN_MATCH-1)]), \
189 s->prev[(str) & s->w_mask] = match_head = s->head[s->ins_h], \
190 s->head[s->ins_h] = (Pos)(str))
193 /* ===========================================================================
194 * Initialize the hash table (avoiding 64K overflow for 16 bit systems).
195 * prev[] will be initialized on the fly.
197 #define CLEAR_HASH(s) \
198 s->head[s->hash_size-1] = NIL; \
199 zmemzero((Bytef *)s->head, (unsigned)(s->hash_size-1)*sizeof(*s->head));
201 /* ========================================================================= */
202 int ZEXPORT deflateInit_(strm, level, version, stream_size)
208 return deflateInit2_(strm, level, Z_DEFLATED, MAX_WBITS, DEF_MEM_LEVEL,
209 Z_DEFAULT_STRATEGY, version, stream_size);
210 /* To do: ignore strm->next_in if we use it as window */
213 /* ========================================================================= */
214 int ZEXPORT deflateInit2_(strm, level, method, windowBits, memLevel, strategy,
215 version, stream_size)
227 static const char my_version[] = ZLIB_VERSION;
230 /* We overlay pending_buf and d_buf+l_buf. This works since the average
231 * output size for (length,distance) codes is <= 24 bits.
234 if (version == Z_NULL || version[0] != my_version[0] ||
235 stream_size != sizeof(z_stream)) {
236 return Z_VERSION_ERROR;
238 if (strm == Z_NULL) return Z_STREAM_ERROR;
241 if (strm->zalloc == Z_NULL) {
242 strm->zalloc = zcalloc;
243 strm->opaque = (voidpf)0;
245 if (strm->zfree == Z_NULL) strm->zfree = zcfree;
248 if (level != 0) level = 1;
250 if (level == Z_DEFAULT_COMPRESSION) level = 6;
253 if (windowBits < 0) { /* undocumented feature: suppress zlib header */
255 windowBits = -windowBits;
257 if (memLevel < 1 || memLevel > MAX_MEM_LEVEL || method != Z_DEFLATED ||
258 windowBits < 9 || windowBits > 15 || level < 0 || level > 9 ||
259 strategy < 0 || strategy > Z_RLE) {
260 return Z_STREAM_ERROR;
262 s = (deflate_state *) ZALLOC(strm, 1, sizeof(deflate_state));
263 if (s == Z_NULL) return Z_MEM_ERROR;
264 strm->state = (struct internal_state FAR *)s;
267 s->noheader = noheader;
268 s->w_bits = windowBits;
269 s->w_size = 1 << s->w_bits;
270 s->w_mask = s->w_size - 1;
272 s->hash_bits = memLevel + 7;
273 s->hash_size = 1 << s->hash_bits;
274 s->hash_mask = s->hash_size - 1;
275 s->hash_shift = ((s->hash_bits+MIN_MATCH-1)/MIN_MATCH);
277 s->window = (Bytef *) ZALLOC(strm, s->w_size, 2*sizeof(Byte));
278 s->prev = (Posf *) ZALLOC(strm, s->w_size, sizeof(Pos));
279 s->head = (Posf *) ZALLOC(strm, s->hash_size, sizeof(Pos));
281 s->lit_bufsize = 1 << (memLevel + 6); /* 16K elements by default */
283 overlay = (ushf *) ZALLOC(strm, s->lit_bufsize, sizeof(ush)+2);
284 s->pending_buf = (uchf *) overlay;
285 s->pending_buf_size = (ulg)s->lit_bufsize * (sizeof(ush)+2L);
287 if (s->window == Z_NULL || s->prev == Z_NULL || s->head == Z_NULL ||
288 s->pending_buf == Z_NULL) {
289 s->status = FINISH_STATE;
290 strm->msg = (char*)ERR_MSG(Z_MEM_ERROR);
294 s->d_buf = overlay + s->lit_bufsize/sizeof(ush);
295 s->l_buf = s->pending_buf + (1+sizeof(ush))*s->lit_bufsize;
298 s->strategy = strategy;
299 s->method = (Byte)method;
301 return deflateReset(strm);
304 /* ========================================================================= */
305 int ZEXPORT deflateSetDictionary (strm, dictionary, dictLength)
307 const Bytef *dictionary;
311 uInt length = dictLength;
315 if (strm == Z_NULL || strm->state == Z_NULL || dictionary == Z_NULL ||
316 (!strm->state->noheader && strm->state->status != INIT_STATE))
317 return Z_STREAM_ERROR;
321 strm->adler = adler32(strm->adler, dictionary, dictLength);
323 if (length < MIN_MATCH) return Z_OK;
324 if (length > MAX_DIST(s)) {
325 length = MAX_DIST(s);
326 #ifndef USE_DICT_HEAD
327 dictionary += dictLength - length; /* use the tail of the dictionary */
330 zmemcpy(s->window, dictionary, length);
331 s->strstart = length;
332 s->block_start = (long)length;
334 /* Insert all strings in the hash table (except for the last two bytes).
335 * s->lookahead stays null, so s->ins_h will be recomputed at the next
336 * call of fill_window.
338 s->ins_h = s->window[0];
339 UPDATE_HASH(s, s->ins_h, s->window[1]);
340 for (n = 0; n <= length - MIN_MATCH; n++) {
341 INSERT_STRING(s, n, hash_head);
343 if (hash_head) hash_head = 0; /* to make compiler happy */
347 /* ========================================================================= */
348 int ZEXPORT deflateReset (strm)
353 if (strm == Z_NULL || strm->state == Z_NULL ||
354 strm->zalloc == Z_NULL || strm->zfree == Z_NULL) return Z_STREAM_ERROR;
356 strm->total_in = strm->total_out = 0;
357 strm->msg = Z_NULL; /* use zfree if we ever allocate msg dynamically */
358 strm->data_type = Z_UNKNOWN;
360 s = (deflate_state *)strm->state;
362 s->pending_out = s->pending_buf;
364 if (s->noheader < 0) {
365 s->noheader = 0; /* was set to -1 by deflate(..., Z_FINISH); */
367 s->status = s->noheader ? BUSY_STATE : INIT_STATE;
369 s->last_flush = Z_NO_FLUSH;
377 /* ========================================================================= */
378 int ZEXPORT deflateParams(strm, level, strategy)
387 if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
391 if (level != 0) level = 1;
393 if (level == Z_DEFAULT_COMPRESSION) level = 6;
395 if (level < 0 || level > 9 || strategy < 0 || strategy > Z_RLE) {
396 return Z_STREAM_ERROR;
398 func = configuration_table[s->level].func;
400 if (func != configuration_table[level].func && strm->total_in != 0) {
401 /* Flush the last buffer: */
402 err = deflate(strm, Z_PARTIAL_FLUSH);
404 if (s->level != level) {
406 s->max_lazy_match = configuration_table[level].max_lazy;
407 s->good_match = configuration_table[level].good_length;
408 s->nice_match = configuration_table[level].nice_length;
409 s->max_chain_length = configuration_table[level].max_chain;
411 s->strategy = strategy;
415 /* =========================================================================
416 * For the default windowBits of 15 and memLevel of 8, this function returns
417 * a close to exact, as well as small, upper bound on the compressed size.
418 * They are coded as constants here for a reason--if the #define's are
419 * changed, then this function needs to be changed as well. The return
420 * value for 15 and 8 only works for those exact settings.
422 * For any setting other than those defaults for windowBits and memLevel,
423 * the value returned is a conservative worst case for the maximum expansion
424 * resulting from using fixed blocks instead of stored blocks, which deflate
425 * can emit on compressed data for some combinations of the parameters.
427 * This function could be more sophisticated to provide closer upper bounds
428 * for every combination of windowBits and memLevel, as well as noheader.
429 * But even the conservative upper bound of about 14% expansion does not
430 * seem onerous for output buffer allocation.
432 uLong ZEXPORT deflateBound(strm, sourceLen)
439 /* conservative upper bound */
440 destLen = sourceLen +
441 ((sourceLen + 7) >> 3) + ((sourceLen + 63) >> 6) + 11;
443 /* if can't get parameters, return conservative bound */
444 if (strm == Z_NULL || strm->state == Z_NULL)
447 /* if not default parameters, return conservative bound */
449 if (s->w_bits != 15 || s->hash_bits != 8 + 7)
452 /* default settings: return tight bound for that case */
453 return compressBound(sourceLen);
456 /* =========================================================================
457 * Put a short in the pending buffer. The 16-bit value is put in MSB order.
458 * IN assertion: the stream state is correct and there is enough room in
461 local void putShortMSB (s, b)
465 put_byte(s, (Byte)(b >> 8));
466 put_byte(s, (Byte)(b & 0xff));
469 /* =========================================================================
470 * Flush as much pending output as possible. All deflate() output goes
471 * through this function so some applications may wish to modify it
472 * to avoid allocating a large strm->next_out buffer and copying into it.
473 * (See also read_buf()).
475 local void flush_pending(strm)
478 unsigned len = strm->state->pending;
480 if (len > strm->avail_out) len = strm->avail_out;
481 if (len == 0) return;
483 zmemcpy(strm->next_out, strm->state->pending_out, len);
484 strm->next_out += len;
485 strm->state->pending_out += len;
486 strm->total_out += len;
487 strm->avail_out -= len;
488 strm->state->pending -= len;
489 if (strm->state->pending == 0) {
490 strm->state->pending_out = strm->state->pending_buf;
494 /* ========================================================================= */
495 int ZEXPORT deflate (strm, flush)
499 int old_flush; /* value of flush param for previous deflate call */
502 if (strm == Z_NULL || strm->state == Z_NULL ||
503 flush > Z_FINISH || flush < 0) {
504 return Z_STREAM_ERROR;
508 if (strm->next_out == Z_NULL ||
509 (strm->next_in == Z_NULL && strm->avail_in != 0) ||
510 (s->status == FINISH_STATE && flush != Z_FINISH)) {
511 ERR_RETURN(strm, Z_STREAM_ERROR);
513 if (strm->avail_out == 0) ERR_RETURN(strm, Z_BUF_ERROR);
515 s->strm = strm; /* just in case */
516 old_flush = s->last_flush;
517 s->last_flush = flush;
519 /* Write the zlib header */
520 if (s->status == INIT_STATE) {
522 uInt header = (Z_DEFLATED + ((s->w_bits-8)<<4)) << 8;
523 uInt level_flags = (s->level-1) >> 1;
525 if (level_flags > 3) level_flags = 3;
526 header |= (level_flags << 6);
527 if (s->strstart != 0) header |= PRESET_DICT;
528 header += 31 - (header % 31);
530 s->status = BUSY_STATE;
531 putShortMSB(s, header);
533 /* Save the adler32 of the preset dictionary: */
534 if (s->strstart != 0) {
535 putShortMSB(s, (uInt)(strm->adler >> 16));
536 putShortMSB(s, (uInt)(strm->adler & 0xffff));
541 /* Flush as much pending output as possible */
542 if (s->pending != 0) {
544 if (strm->avail_out == 0) {
545 /* Since avail_out is 0, deflate will be called again with
546 * more output space, but possibly with both pending and
547 * avail_in equal to zero. There won't be anything to do,
548 * but this is not an error situation so make sure we
549 * return OK instead of BUF_ERROR at next call of deflate:
555 /* Make sure there is something to do and avoid duplicate consecutive
556 * flushes. For repeated and useless calls with Z_FINISH, we keep
557 * returning Z_STREAM_END instead of Z_BUF_ERROR.
559 } else if (strm->avail_in == 0 && flush <= old_flush &&
561 ERR_RETURN(strm, Z_BUF_ERROR);
564 /* User must not provide more input after the first FINISH: */
565 if (s->status == FINISH_STATE && strm->avail_in != 0) {
566 ERR_RETURN(strm, Z_BUF_ERROR);
569 /* Start a new block or continue the current one.
571 if (strm->avail_in != 0 || s->lookahead != 0 ||
572 (flush != Z_NO_FLUSH && s->status != FINISH_STATE)) {
575 bstate = (*(configuration_table[s->level].func))(s, flush);
577 if (bstate == finish_started || bstate == finish_done) {
578 s->status = FINISH_STATE;
580 if (bstate == need_more || bstate == finish_started) {
581 if (strm->avail_out == 0) {
582 s->last_flush = -1; /* avoid BUF_ERROR next call, see above */
585 /* If flush != Z_NO_FLUSH && avail_out == 0, the next call
586 * of deflate should use the same flush parameter to make sure
587 * that the flush is complete. So we don't have to output an
588 * empty block here, this will be done at next call. This also
589 * ensures that for a very small output buffer, we emit at most
593 if (bstate == block_done) {
594 if (flush == Z_PARTIAL_FLUSH) {
596 } else { /* FULL_FLUSH or SYNC_FLUSH */
597 _tr_stored_block(s, (char*)0, 0L, 0);
598 /* For a full flush, this empty block will be recognized
599 * as a special marker by inflate_sync().
601 if (flush == Z_FULL_FLUSH) {
602 CLEAR_HASH(s); /* forget history */
606 if (strm->avail_out == 0) {
607 s->last_flush = -1; /* avoid BUF_ERROR at next call, see above */
612 Assert(strm->avail_out > 0, "bug2");
614 if (flush != Z_FINISH) return Z_OK;
615 if (s->noheader) return Z_STREAM_END;
617 /* Write the zlib trailer (adler32) */
618 putShortMSB(s, (uInt)(strm->adler >> 16));
619 putShortMSB(s, (uInt)(strm->adler & 0xffff));
621 /* If avail_out is zero, the application will call deflate again
624 s->noheader = -1; /* write the trailer only once! */
625 return s->pending != 0 ? Z_OK : Z_STREAM_END;
628 /* ========================================================================= */
629 int ZEXPORT deflateEnd (strm)
634 if (strm == Z_NULL || strm->state == Z_NULL) return Z_STREAM_ERROR;
636 status = strm->state->status;
637 if (status != INIT_STATE && status != BUSY_STATE &&
638 status != FINISH_STATE) {
639 return Z_STREAM_ERROR;
642 /* Deallocate in reverse order of allocations: */
643 TRY_FREE(strm, strm->state->pending_buf);
644 TRY_FREE(strm, strm->state->head);
645 TRY_FREE(strm, strm->state->prev);
646 TRY_FREE(strm, strm->state->window);
648 ZFREE(strm, strm->state);
649 strm->state = Z_NULL;
651 return status == BUSY_STATE ? Z_DATA_ERROR : Z_OK;
654 /* =========================================================================
655 * Copy the source state to the destination state.
656 * To simplify the source, this is not supported for 16-bit MSDOS (which
657 * doesn't have enough memory anyway to duplicate compression states).
659 int ZEXPORT deflateCopy (dest, source)
664 return Z_STREAM_ERROR;
671 if (source == Z_NULL || dest == Z_NULL || source->state == Z_NULL) {
672 return Z_STREAM_ERROR;
679 ds = (deflate_state *) ZALLOC(dest, 1, sizeof(deflate_state));
680 if (ds == Z_NULL) return Z_MEM_ERROR;
681 dest->state = (struct internal_state FAR *) ds;
685 ds->window = (Bytef *) ZALLOC(dest, ds->w_size, 2*sizeof(Byte));
686 ds->prev = (Posf *) ZALLOC(dest, ds->w_size, sizeof(Pos));
687 ds->head = (Posf *) ZALLOC(dest, ds->hash_size, sizeof(Pos));
688 overlay = (ushf *) ZALLOC(dest, ds->lit_bufsize, sizeof(ush)+2);
689 ds->pending_buf = (uchf *) overlay;
691 if (ds->window == Z_NULL || ds->prev == Z_NULL || ds->head == Z_NULL ||
692 ds->pending_buf == Z_NULL) {
696 /* following zmemcpy do not work for 16-bit MSDOS */
697 zmemcpy(ds->window, ss->window, ds->w_size * 2 * sizeof(Byte));
698 zmemcpy(ds->prev, ss->prev, ds->w_size * sizeof(Pos));
699 zmemcpy(ds->head, ss->head, ds->hash_size * sizeof(Pos));
700 zmemcpy(ds->pending_buf, ss->pending_buf, (uInt)ds->pending_buf_size);
702 ds->pending_out = ds->pending_buf + (ss->pending_out - ss->pending_buf);
703 ds->d_buf = overlay + ds->lit_bufsize/sizeof(ush);
704 ds->l_buf = ds->pending_buf + (1+sizeof(ush))*ds->lit_bufsize;
706 ds->l_desc.dyn_tree = ds->dyn_ltree;
707 ds->d_desc.dyn_tree = ds->dyn_dtree;
708 ds->bl_desc.dyn_tree = ds->bl_tree;
711 #endif /* MAXSEG_64K */
714 /* ===========================================================================
715 * Read a new buffer from the current input stream, update the adler32
716 * and total number of bytes read. All deflate() input goes through
717 * this function so some applications may wish to modify it to avoid
718 * allocating a large strm->next_in buffer and copying from it.
719 * (See also flush_pending()).
721 local int read_buf(strm, buf, size)
726 unsigned len = strm->avail_in;
728 if (len > size) len = size;
729 if (len == 0) return 0;
731 strm->avail_in -= len;
733 if (!strm->state->noheader) {
734 strm->adler = adler32(strm->adler, strm->next_in, len);
736 zmemcpy(buf, strm->next_in, len);
737 strm->next_in += len;
738 strm->total_in += len;
743 /* ===========================================================================
744 * Initialize the "longest match" routines for a new zlib stream
746 local void lm_init (s)
749 s->window_size = (ulg)2L*s->w_size;
753 /* Set the default configuration parameters:
755 s->max_lazy_match = configuration_table[s->level].max_lazy;
756 s->good_match = configuration_table[s->level].good_length;
757 s->nice_match = configuration_table[s->level].nice_length;
758 s->max_chain_length = configuration_table[s->level].max_chain;
763 s->match_length = s->prev_length = MIN_MATCH-1;
764 s->match_available = 0;
767 match_init(); /* initialize the asm code */
772 /* ===========================================================================
773 * Set match_start to the longest match starting at the given string and
774 * return its length. Matches shorter or equal to prev_length are discarded,
775 * in which case the result is equal to prev_length and match_start is
777 * IN assertions: cur_match is the head of the hash chain for the current
778 * string (strstart) and its distance is <= MAX_DIST, and prev_length >= 1
779 * OUT assertion: the match length is not greater than s->lookahead.
782 /* For 80x86 and 680x0, an optimized version will be provided in match.asm or
783 * match.S. The code will be functionally equivalent.
785 local uInt longest_match(s, cur_match)
787 IPos cur_match; /* current match */
789 unsigned chain_length = s->max_chain_length;/* max hash chain length */
790 register Bytef *scan = s->window + s->strstart; /* current string */
791 register Bytef *match; /* matched string */
792 register int len; /* length of current match */
793 int best_len = s->prev_length; /* best match length so far */
794 int nice_match = s->nice_match; /* stop if match long enough */
795 IPos limit = s->strstart > (IPos)MAX_DIST(s) ?
796 s->strstart - (IPos)MAX_DIST(s) : NIL;
797 /* Stop when cur_match becomes <= limit. To simplify the code,
798 * we prevent matches with the string of window index 0.
800 Posf *prev = s->prev;
801 uInt wmask = s->w_mask;
804 /* Compare two bytes at a time. Note: this is not always beneficial.
805 * Try with and without -DUNALIGNED_OK to check.
807 register Bytef *strend = s->window + s->strstart + MAX_MATCH - 1;
808 register ush scan_start = *(ushf*)scan;
809 register ush scan_end = *(ushf*)(scan+best_len-1);
811 register Bytef *strend = s->window + s->strstart + MAX_MATCH;
812 register Byte scan_end1 = scan[best_len-1];
813 register Byte scan_end = scan[best_len];
816 /* The code is optimized for HASH_BITS >= 8 and MAX_MATCH-2 multiple of 16.
817 * It is easy to get rid of this optimization if necessary.
819 Assert(s->hash_bits >= 8 && MAX_MATCH == 258, "Code too clever");
821 /* Do not waste too much time if we already have a good match: */
822 if (s->prev_length >= s->good_match) {
825 /* Do not look for matches beyond the end of the input. This is necessary
826 * to make deflate deterministic.
828 if ((uInt)nice_match > s->lookahead) nice_match = s->lookahead;
830 Assert((ulg)s->strstart <= s->window_size-MIN_LOOKAHEAD, "need lookahead");
833 Assert(cur_match < s->strstart, "no future");
834 match = s->window + cur_match;
836 /* Skip to next match if the match length cannot increase
837 * or if the match length is less than 2:
839 #if (defined(UNALIGNED_OK) && MAX_MATCH == 258)
840 /* This code assumes sizeof(unsigned short) == 2. Do not use
841 * UNALIGNED_OK if your compiler uses a different size.
843 if (*(ushf*)(match+best_len-1) != scan_end ||
844 *(ushf*)match != scan_start) continue;
846 /* It is not necessary to compare scan[2] and match[2] since they are
847 * always equal when the other bytes match, given that the hash keys
848 * are equal and that HASH_BITS >= 8. Compare 2 bytes at a time at
849 * strstart+3, +5, ... up to strstart+257. We check for insufficient
850 * lookahead only every 4th comparison; the 128th check will be made
851 * at strstart+257. If MAX_MATCH-2 is not a multiple of 8, it is
852 * necessary to put more guard bytes at the end of the window, or
853 * to check more often for insufficient lookahead.
855 Assert(scan[2] == match[2], "scan[2]?");
858 } while (*(ushf*)(scan+=2) == *(ushf*)(match+=2) &&
859 *(ushf*)(scan+=2) == *(ushf*)(match+=2) &&
860 *(ushf*)(scan+=2) == *(ushf*)(match+=2) &&
861 *(ushf*)(scan+=2) == *(ushf*)(match+=2) &&
863 /* The funny "do {}" generates better code on most compilers */
865 /* Here, scan <= window+strstart+257 */
866 Assert(scan <= s->window+(unsigned)(s->window_size-1), "wild scan");
867 if (*scan == *match) scan++;
869 len = (MAX_MATCH - 1) - (int)(strend-scan);
870 scan = strend - (MAX_MATCH-1);
872 #else /* UNALIGNED_OK */
874 if (match[best_len] != scan_end ||
875 match[best_len-1] != scan_end1 ||
877 *++match != scan[1]) continue;
879 /* The check at best_len-1 can be removed because it will be made
880 * again later. (This heuristic is not always a win.)
881 * It is not necessary to compare scan[2] and match[2] since they
882 * are always equal when the other bytes match, given that
883 * the hash keys are equal and that HASH_BITS >= 8.
886 Assert(*scan == *match, "match[2]?");
888 /* We check for insufficient lookahead only every 8th comparison;
889 * the 256th check will be made at strstart+258.
892 } while (*++scan == *++match && *++scan == *++match &&
893 *++scan == *++match && *++scan == *++match &&
894 *++scan == *++match && *++scan == *++match &&
895 *++scan == *++match && *++scan == *++match &&
898 Assert(scan <= s->window+(unsigned)(s->window_size-1), "wild scan");
900 len = MAX_MATCH - (int)(strend - scan);
901 scan = strend - MAX_MATCH;
903 #endif /* UNALIGNED_OK */
905 if (len > best_len) {
906 s->match_start = cur_match;
908 if (len >= nice_match) break;
910 scan_end = *(ushf*)(scan+best_len-1);
912 scan_end1 = scan[best_len-1];
913 scan_end = scan[best_len];
916 } while ((cur_match = prev[cur_match & wmask]) > limit
917 && --chain_length != 0);
919 if ((uInt)best_len <= s->lookahead) return (uInt)best_len;
925 /* ---------------------------------------------------------------------------
926 * Optimized version for level == 1 or strategy == Z_RLE only
928 local uInt longest_match_fast(s, cur_match)
930 IPos cur_match; /* current match */
932 register Bytef *scan = s->window + s->strstart; /* current string */
933 register Bytef *match; /* matched string */
934 register int len; /* length of current match */
935 register Bytef *strend = s->window + s->strstart + MAX_MATCH;
937 /* The code is optimized for HASH_BITS >= 8 and MAX_MATCH-2 multiple of 16.
938 * It is easy to get rid of this optimization if necessary.
940 Assert(s->hash_bits >= 8 && MAX_MATCH == 258, "Code too clever");
942 Assert((ulg)s->strstart <= s->window_size-MIN_LOOKAHEAD, "need lookahead");
944 Assert(cur_match < s->strstart, "no future");
946 match = s->window + cur_match;
948 /* Return failure if the match length is less than 2:
950 if (match[0] != scan[0] || match[1] != scan[1]) return MIN_MATCH-1;
952 /* The check at best_len-1 can be removed because it will be made
953 * again later. (This heuristic is not always a win.)
954 * It is not necessary to compare scan[2] and match[2] since they
955 * are always equal when the other bytes match, given that
956 * the hash keys are equal and that HASH_BITS >= 8.
958 scan += 2, match += 2;
959 Assert(*scan == *match, "match[2]?");
961 /* We check for insufficient lookahead only every 8th comparison;
962 * the 256th check will be made at strstart+258.
965 } while (*++scan == *++match && *++scan == *++match &&
966 *++scan == *++match && *++scan == *++match &&
967 *++scan == *++match && *++scan == *++match &&
968 *++scan == *++match && *++scan == *++match &&
971 Assert(scan <= s->window+(unsigned)(s->window_size-1), "wild scan");
973 len = MAX_MATCH - (int)(strend - scan);
975 if (len < MIN_MATCH) return MIN_MATCH - 1;
977 s->match_start = cur_match;
978 return len <= s->lookahead ? len : s->lookahead;
982 /* ===========================================================================
983 * Check that the match at match_start is indeed a match.
985 local void check_match(s, start, match, length)
990 /* check that the match is indeed a match */
991 if (zmemcmp(s->window + match,
992 s->window + start, length) != EQUAL) {
993 fprintf(stderr, " start %u, match %u, length %d\n",
994 start, match, length);
996 fprintf(stderr, "%c%c", s->window[match++], s->window[start++]);
997 } while (--length != 0);
998 z_error("invalid match");
1000 if (z_verbose > 1) {
1001 fprintf(stderr,"\\[%d,%d]", start-match, length);
1002 do { putc(s->window[start++], stderr); } while (--length != 0);
1006 # define check_match(s, start, match, length)
1009 /* ===========================================================================
1010 * Fill the window when the lookahead becomes insufficient.
1011 * Updates strstart and lookahead.
1013 * IN assertion: lookahead < MIN_LOOKAHEAD
1014 * OUT assertions: strstart <= window_size-MIN_LOOKAHEAD
1015 * At least one byte has been read, or avail_in == 0; reads are
1016 * performed for at least two bytes (required for the zip translate_eol
1017 * option -- not supported here).
1019 local void fill_window(s)
1022 register unsigned n, m;
1024 unsigned more; /* Amount of free space at the end of the window. */
1025 uInt wsize = s->w_size;
1028 more = (unsigned)(s->window_size -(ulg)s->lookahead -(ulg)s->strstart);
1030 /* Deal with !@#$% 64K limit: */
1031 if (more == 0 && s->strstart == 0 && s->lookahead == 0) {
1034 } else if (more == (unsigned)(-1)) {
1035 /* Very unlikely, but possible on 16 bit machine if strstart == 0
1036 * and lookahead == 1 (input done one byte at time)
1040 /* If the window is almost full and there is insufficient lookahead,
1041 * move the upper half to the lower one to make room in the upper half.
1043 } else if (s->strstart >= wsize+MAX_DIST(s)) {
1045 zmemcpy(s->window, s->window+wsize, (unsigned)wsize);
1046 s->match_start -= wsize;
1047 s->strstart -= wsize; /* we now have strstart >= MAX_DIST */
1048 s->block_start -= (long) wsize;
1050 /* Slide the hash table (could be avoided with 32 bit values
1051 at the expense of memory usage). We slide even when level == 0
1052 to keep the hash table consistent if we switch back to level > 0
1053 later. (Using level 0 permanently is not an optimal usage of
1054 zlib, so we don't care about this pathological case.)
1060 *p = (Pos)(m >= wsize ? m-wsize : NIL);
1068 *p = (Pos)(m >= wsize ? m-wsize : NIL);
1069 /* If n is not on any hash chain, prev[n] is garbage but
1070 * its value will never be used.
1076 if (s->strm->avail_in == 0) return;
1078 /* If there was no sliding:
1079 * strstart <= WSIZE+MAX_DIST-1 && lookahead <= MIN_LOOKAHEAD - 1 &&
1080 * more == window_size - lookahead - strstart
1081 * => more >= window_size - (MIN_LOOKAHEAD-1 + WSIZE + MAX_DIST-1)
1082 * => more >= window_size - 2*WSIZE + 2
1083 * In the BIG_MEM or MMAP case (not yet supported),
1084 * window_size == input_size + MIN_LOOKAHEAD &&
1085 * strstart + s->lookahead <= input_size => more >= MIN_LOOKAHEAD.
1086 * Otherwise, window_size == 2*WSIZE so more >= 2.
1087 * If there was sliding, more >= WSIZE. So in all cases, more >= 2.
1089 Assert(more >= 2, "more < 2");
1091 n = read_buf(s->strm, s->window + s->strstart + s->lookahead, more);
1094 /* Initialize the hash value now that we have some input: */
1095 if (s->lookahead >= MIN_MATCH) {
1096 s->ins_h = s->window[s->strstart];
1097 UPDATE_HASH(s, s->ins_h, s->window[s->strstart+1]);
1099 Call UPDATE_HASH() MIN_MATCH-3 more times
1102 /* If the whole input has less than MIN_MATCH bytes, ins_h is garbage,
1103 * but this is not important since only literal bytes will be emitted.
1106 } while (s->lookahead < MIN_LOOKAHEAD && s->strm->avail_in != 0);
1109 /* ===========================================================================
1110 * Flush the current block, with given end-of-file flag.
1111 * IN assertion: strstart is set to the end of the current match.
1113 #define FLUSH_BLOCK_ONLY(s, eof) { \
1114 _tr_flush_block(s, (s->block_start >= 0L ? \
1115 (charf *)&s->window[(unsigned)s->block_start] : \
1117 (ulg)((long)s->strstart - s->block_start), \
1119 s->block_start = s->strstart; \
1120 flush_pending(s->strm); \
1121 Tracev((stderr,"[FLUSH]")); \
1124 /* Same but force premature exit if necessary. */
1125 #define FLUSH_BLOCK(s, eof) { \
1126 FLUSH_BLOCK_ONLY(s, eof); \
1127 if (s->strm->avail_out == 0) return (eof) ? finish_started : need_more; \
1130 /* ===========================================================================
1131 * Copy without compression as much as possible from the input stream, return
1132 * the current block state.
1133 * This function does not insert new strings in the dictionary since
1134 * uncompressible data is probably not useful. This function is used
1135 * only for the level=0 compression option.
1136 * NOTE: this function should be optimized to avoid extra copying from
1137 * window to pending_buf.
1139 local block_state deflate_stored(s, flush)
1143 /* Stored blocks are limited to 0xffff bytes, pending_buf is limited
1144 * to pending_buf_size, and each stored block has a 5 byte header:
1146 ulg max_block_size = 0xffff;
1149 if (max_block_size > s->pending_buf_size - 5) {
1150 max_block_size = s->pending_buf_size - 5;
1153 /* Copy as much as possible from input to output: */
1155 /* Fill the window as much as possible: */
1156 if (s->lookahead <= 1) {
1158 Assert(s->strstart < s->w_size+MAX_DIST(s) ||
1159 s->block_start >= (long)s->w_size, "slide too late");
1162 if (s->lookahead == 0 && flush == Z_NO_FLUSH) return need_more;
1164 if (s->lookahead == 0) break; /* flush the current block */
1166 Assert(s->block_start >= 0L, "block gone");
1168 s->strstart += s->lookahead;
1171 /* Emit a stored block if pending_buf will be full: */
1172 max_start = s->block_start + max_block_size;
1173 if (s->strstart == 0 || (ulg)s->strstart >= max_start) {
1174 /* strstart == 0 is possible when wraparound on 16-bit machine */
1175 s->lookahead = (uInt)(s->strstart - max_start);
1176 s->strstart = (uInt)max_start;
1179 /* Flush if we may have to slide, otherwise block_start may become
1180 * negative and the data will be gone:
1182 if (s->strstart - (uInt)s->block_start >= MAX_DIST(s)) {
1186 FLUSH_BLOCK(s, flush == Z_FINISH);
1187 return flush == Z_FINISH ? finish_done : block_done;
1190 /* ===========================================================================
1191 * Compress as much as possible from the input stream, return the current
1193 * This function does not perform lazy evaluation of matches and inserts
1194 * new strings in the dictionary only for unmatched strings or for short
1195 * matches. It is used only for the fast compression options.
1197 local block_state deflate_fast(s, flush)
1201 IPos hash_head = NIL; /* head of the hash chain */
1202 int bflush; /* set if current block must be flushed */
1205 /* Make sure that we always have enough lookahead, except
1206 * at the end of the input file. We need MAX_MATCH bytes
1207 * for the next match, plus MIN_MATCH bytes to insert the
1208 * string following the next match.
1210 if (s->lookahead < MIN_LOOKAHEAD) {
1212 if (s->lookahead < MIN_LOOKAHEAD && flush == Z_NO_FLUSH) {
1215 if (s->lookahead == 0) break; /* flush the current block */
1218 /* Insert the string window[strstart .. strstart+2] in the
1219 * dictionary, and set hash_head to the head of the hash chain:
1221 if (s->lookahead >= MIN_MATCH) {
1222 INSERT_STRING(s, s->strstart, hash_head);
1225 /* Find the longest match, discarding those <= prev_length.
1226 * At this point we have always match_length < MIN_MATCH
1228 if (hash_head != NIL && s->strstart - hash_head <= MAX_DIST(s)) {
1229 /* To simplify the code, we prevent matches with the string
1230 * of window index 0 (in particular we have to avoid a match
1231 * of the string with itself at the start of the input file).
1234 if ((s->strategy < Z_HUFFMAN_ONLY) ||
1235 (s->strategy == Z_RLE && s->strstart - hash_head == 1)) {
1236 s->match_length = longest_match_fast (s, hash_head);
1239 if (s->strategy < Z_HUFFMAN_ONLY) {
1240 s->match_length = longest_match (s, hash_head);
1241 } else if (s->strategy == Z_RLE && s->strstart - hash_head == 1) {
1242 s->match_length = longest_match_fast (s, hash_head);
1245 /* longest_match() or longest_match_fast() sets match_start */
1247 if (s->match_length >= MIN_MATCH) {
1248 check_match(s, s->strstart, s->match_start, s->match_length);
1250 _tr_tally_dist(s, s->strstart - s->match_start,
1251 s->match_length - MIN_MATCH, bflush);
1253 s->lookahead -= s->match_length;
1255 /* Insert new strings in the hash table only if the match length
1256 * is not too large. This saves time but degrades compression.
1259 if (s->match_length <= s->max_insert_length &&
1260 s->lookahead >= MIN_MATCH) {
1261 s->match_length--; /* string at strstart already in hash table */
1264 INSERT_STRING(s, s->strstart, hash_head);
1265 /* strstart never exceeds WSIZE-MAX_MATCH, so there are
1266 * always MIN_MATCH bytes ahead.
1268 } while (--s->match_length != 0);
1273 s->strstart += s->match_length;
1274 s->match_length = 0;
1275 s->ins_h = s->window[s->strstart];
1276 UPDATE_HASH(s, s->ins_h, s->window[s->strstart+1]);
1278 Call UPDATE_HASH() MIN_MATCH-3 more times
1280 /* If lookahead < MIN_MATCH, ins_h is garbage, but it does not
1281 * matter since it will be recomputed at next deflate call.
1285 /* No match, output a literal byte */
1286 Tracevv((stderr,"%c", s->window[s->strstart]));
1287 _tr_tally_lit (s, s->window[s->strstart], bflush);
1291 if (bflush) FLUSH_BLOCK(s, 0);
1293 FLUSH_BLOCK(s, flush == Z_FINISH);
1294 return flush == Z_FINISH ? finish_done : block_done;
1298 /* ===========================================================================
1299 * Same as above, but achieves better compression. We use a lazy
1300 * evaluation for matches: a match is finally adopted only if there is
1301 * no better match at the next window position.
1303 local block_state deflate_slow(s, flush)
1307 IPos hash_head = NIL; /* head of hash chain */
1308 int bflush; /* set if current block must be flushed */
1310 /* Process the input block. */
1312 /* Make sure that we always have enough lookahead, except
1313 * at the end of the input file. We need MAX_MATCH bytes
1314 * for the next match, plus MIN_MATCH bytes to insert the
1315 * string following the next match.
1317 if (s->lookahead < MIN_LOOKAHEAD) {
1319 if (s->lookahead < MIN_LOOKAHEAD && flush == Z_NO_FLUSH) {
1322 if (s->lookahead == 0) break; /* flush the current block */
1325 /* Insert the string window[strstart .. strstart+2] in the
1326 * dictionary, and set hash_head to the head of the hash chain:
1328 if (s->lookahead >= MIN_MATCH) {
1329 INSERT_STRING(s, s->strstart, hash_head);
1332 /* Find the longest match, discarding those <= prev_length.
1334 s->prev_length = s->match_length, s->prev_match = s->match_start;
1335 s->match_length = MIN_MATCH-1;
1337 if (hash_head != NIL && s->prev_length < s->max_lazy_match &&
1338 s->strstart - hash_head <= MAX_DIST(s)) {
1339 /* To simplify the code, we prevent matches with the string
1340 * of window index 0 (in particular we have to avoid a match
1341 * of the string with itself at the start of the input file).
1343 if (s->strategy < Z_HUFFMAN_ONLY) {
1344 s->match_length = longest_match (s, hash_head);
1345 } else if (s->strategy == Z_RLE && s->strstart - hash_head == 1) {
1346 s->match_length = longest_match_fast (s, hash_head);
1348 /* longest_match() or longest_match_fast() sets match_start */
1350 if (s->match_length <= 5 && (s->strategy == Z_FILTERED ||
1351 (s->match_length == MIN_MATCH &&
1352 s->strstart - s->match_start > TOO_FAR))) {
1354 /* If prev_match is also MIN_MATCH, match_start is garbage
1355 * but we will ignore the current match anyway.
1357 s->match_length = MIN_MATCH-1;
1360 /* If there was a match at the previous step and the current
1361 * match is not better, output the previous match:
1363 if (s->prev_length >= MIN_MATCH && s->match_length <= s->prev_length) {
1364 uInt max_insert = s->strstart + s->lookahead - MIN_MATCH;
1365 /* Do not insert strings in hash table beyond this. */
1367 check_match(s, s->strstart-1, s->prev_match, s->prev_length);
1369 _tr_tally_dist(s, s->strstart -1 - s->prev_match,
1370 s->prev_length - MIN_MATCH, bflush);
1372 /* Insert in hash table all strings up to the end of the match.
1373 * strstart-1 and strstart are already inserted. If there is not
1374 * enough lookahead, the last two strings are not inserted in
1377 s->lookahead -= s->prev_length-1;
1378 s->prev_length -= 2;
1380 if (++s->strstart <= max_insert) {
1381 INSERT_STRING(s, s->strstart, hash_head);
1383 } while (--s->prev_length != 0);
1384 s->match_available = 0;
1385 s->match_length = MIN_MATCH-1;
1388 if (bflush) FLUSH_BLOCK(s, 0);
1390 } else if (s->match_available) {
1391 /* If there was no match at the previous position, output a
1392 * single literal. If there was a match but the current match
1393 * is longer, truncate the previous match to a single literal.
1395 Tracevv((stderr,"%c", s->window[s->strstart-1]));
1396 _tr_tally_lit(s, s->window[s->strstart-1], bflush);
1398 FLUSH_BLOCK_ONLY(s, 0);
1402 if (s->strm->avail_out == 0) return need_more;
1404 /* There is no previous match to compare with, wait for
1405 * the next step to decide.
1407 s->match_available = 1;
1412 Assert (flush != Z_NO_FLUSH, "no flush?");
1413 if (s->match_available) {
1414 Tracevv((stderr,"%c", s->window[s->strstart-1]));
1415 _tr_tally_lit(s, s->window[s->strstart-1], bflush);
1416 s->match_available = 0;
1418 FLUSH_BLOCK(s, flush == Z_FINISH);
1419 return flush == Z_FINISH ? finish_done : block_done;
1421 #endif /* FASTEST */