6 * Copyright (C) 2012, 2013, 2014 Eric Biggers
8 * This file is part of wimlib, a library for working with WIM files.
10 * wimlib is free software; you can redistribute it and/or modify it under the
11 * terms of the GNU General Public License as published by the Free
12 * Software Foundation; either version 3 of the License, or (at your option)
15 * wimlib is distributed in the hope that it will be useful, but WITHOUT ANY
16 * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
17 * A PARTICULAR PURPOSE. See the GNU General Public License for more
20 * You should have received a copy of the GNU General Public License
21 * along with wimlib; if not, see http://www.gnu.org/licenses/.
26 * This file contains a compressor for the LZX ("Lempel-Ziv eXtended"?)
27 * compression format, as used in the WIM (Windows IMaging) file format. This
28 * code may need some slight modifications to be used outside of the WIM format.
29 * In particular, in other situations the LZX block header might be slightly
30 * different, and a sliding window rather than a fixed-size window might be
33 * ----------------------------------------------------------------------------
37 * The primary reference for LZX is the specification released by Microsoft.
38 * However, the comments in lzx-decompress.c provide more information about LZX
39 * and note some errors in the Microsoft specification.
41 * LZX shares many similarities with DEFLATE, the format used by zlib and gzip.
42 * Both LZX and DEFLATE use LZ77 matching and Huffman coding. Certain details
43 * are quite similar, such as the method for storing Huffman codes. However,
44 * the main differences are:
46 * - LZX preprocesses the data to attempt to make x86 machine code slightly more
47 * compressible before attempting to compress it further.
49 * - LZX uses a "main" alphabet which combines literals and matches, with the
50 * match symbols containing a "length header" (giving all or part of the match
51 * length) and a "position slot" (giving, roughly speaking, the order of
52 * magnitude of the match offset).
54 * - LZX does not have static Huffman blocks (that is, the kind with preset
55 * Huffman codes); however it does have two types of dynamic Huffman blocks
56 * ("verbatim" and "aligned").
58 * - LZX has a minimum match length of 2 rather than 3.
60 * - In LZX, match offsets 0 through 2 actually represent entries in an LRU
61 * queue of match offsets. This is very useful for certain types of files,
62 * such as binary files that have repeating records.
64 * ----------------------------------------------------------------------------
66 * Algorithmic Overview
68 * At a high level, any implementation of LZX compression must operate as
71 * 1. Preprocess the input data to translate the targets of 32-bit x86 call
72 * instructions to absolute offsets. (Actually, this is required for WIM,
73 * but might not be in other places LZX is used.)
75 * 2. Find a sequence of LZ77-style matches and literal bytes that expands to
76 * the preprocessed data.
78 * 3. Divide the match/literal sequence into one or more LZX blocks, each of
79 * which may be "uncompressed", "verbatim", or "aligned".
81 * 4. Output each LZX block.
83 * Step (1) is fairly straightforward. It requires looking for 0xe8 bytes in
84 * the input data and performing a translation on the 4 bytes following each
87 * Step (4) is complicated, but it is mostly determined by the LZX format. The
88 * only real choice we have is what algorithm to use to build the length-limited
89 * canonical Huffman codes. See lzx_write_all_blocks() for details.
91 * That leaves steps (2) and (3) as where all the hard stuff happens. Focusing
92 * on step (2), we need to do LZ77-style parsing on the input data, or "window",
93 * to divide it into a sequence of matches and literals. Each position in the
94 * window might have multiple matches associated with it, and we need to choose
95 * which one, if any, to actually use. Therefore, the problem can really be
96 * divided into two areas of concern: (a) finding matches at a given position,
97 * which we shall call "match-finding", and (b) choosing whether to use a
98 * match or a literal at a given position, and if using a match, which one (if
99 * there is more than one available). We shall call this "match-choosing". We
100 * first consider match-finding, then match-choosing.
102 * ----------------------------------------------------------------------------
106 * Given a position in the window, we want to find LZ77-style "matches" with
107 * that position at previous positions in the window. With LZX, the minimum
108 * match length is 2 and the maximum match length is 257. The only restriction
109 * on offsets is that LZX does not allow the last 2 bytes of the window to match
110 * the the beginning of the window.
112 * Depending on how good a compression ratio we want (see the "Match-choosing"
113 * section), we may want to find: (a) all matches, or (b) just the longest
114 * match, or (c) just some "promising" matches that we are able to find quickly,
115 * or (d) just the longest match that we're able to find quickly. Below we
116 * introduce the match-finding methods that the code currently uses or has
119 * - Hash chains. Maintain a table that maps hash codes, computed from
120 * fixed-length byte sequences, to linked lists containing previous window
121 * positions. To search for matches, compute the hash for the current
122 * position in the window and search the appropriate hash chain. When
123 * advancing to the next position, prepend the current position to the
124 * appropriate hash list. This is a good approach for producing matches with
125 * stategy (d) and is useful for fast compression. Therefore, we provide an
126 * option to use this method for LZX compression. See lz_hash.c for the
129 * - Binary trees. Similar to hash chains, but each hash bucket contains a
130 * binary tree of previous window positions rather than a linked list. This
131 * is a good approach for producing matches with stategy (c) and is useful for
132 * achieving a good compression ratio. Therefore, we provide an option to use
133 * this method; see lz_bt.c for the implementation.
135 * - Suffix arrays. This code previously used this method to produce matches
136 * with stategy (c), but I've dropped it because it was slower than the binary
137 * trees approach, used more memory, and did not improve the compression ratio
138 * enough to compensate. Download wimlib v1.6.2 if you want the code.
139 * However, the suffix array method was basically as follows. Build the
140 * suffix array for the entire window. The suffix array contains each
141 * possible window position, sorted by the lexicographic order of the strings
142 * that begin at those positions. Find the matches at a given position by
143 * searching the suffix array outwards, in both directions, from the suffix
144 * array slot for that position. This produces the longest matches first, but
145 * "matches" that actually occur at later positions in the window must be
146 * skipped. To do this skipping, use an auxiliary array with dynamically
147 * constructed linked lists. Also, use the inverse suffix array to quickly
148 * find the suffix array slot for a given position without doing a binary
151 * ----------------------------------------------------------------------------
155 * Usually, choosing the longest match is best because it encodes the most data
156 * in that one item. However, sometimes the longest match is not optimal
157 * because (a) choosing a long match now might prevent using an even longer
158 * match later, or (b) more generally, what we actually care about is the number
159 * of bits it will ultimately take to output each match or literal, which is
160 * actually dependent on the entropy encoding using by the underlying
161 * compression format. Consequently, a longer match usually, but not always,
162 * takes fewer bits to encode than multiple shorter matches or literals that
163 * cover the same data.
165 * This problem of choosing the truly best match/literal sequence is probably
166 * impossible to solve efficiently when combined with entropy encoding. If we
167 * knew how many bits it takes to output each match/literal, then we could
168 * choose the optimal sequence using shortest-path search a la Dijkstra's
169 * algorithm. However, with entropy encoding, the chosen match/literal sequence
170 * affects its own encoding. Therefore, we can't know how many bits it will
171 * take to actually output any one match or literal until we have actually
172 * chosen the full sequence of matches and literals.
174 * Notwithstanding the entropy encoding problem, we also aren't guaranteed to
175 * choose the optimal match/literal sequence unless the match-finder (see
176 * section "Match-finder") provides the match-chooser with all possible matches
177 * at each position. However, this is not computationally efficient. For
178 * example, there might be many matches of the same length, and usually (but not
179 * always) the best choice is the one with the smallest offset. So in practice,
180 * it's fine to only consider the smallest offset for a given match length at a
181 * given position. (Actually, for LZX, it's also worth considering repeat
184 * In addition, as mentioned earlier, in LZX we have the choice of using
185 * multiple blocks, each of which resets the Huffman codes. This expands the
186 * search space even further. Therefore, to simplify the problem, we currently
187 * we don't attempt to actually choose the LZX blocks based on the data.
188 * Instead, we just divide the data into fixed-size blocks of LZX_DIV_BLOCK_SIZE
189 * bytes each, and always use verbatim or aligned blocks (never uncompressed).
190 * A previous version of this code recursively split the input data into
191 * equal-sized blocks, up to a maximum depth, and chose the lowest-cost block
192 * divisions. However, this made compression much slower and did not actually
193 * help very much. It remains an open question whether a sufficiently fast and
194 * useful block-splitting algorithm is possible for LZX. Essentially the same
195 * problem also applies to DEFLATE. The Microsoft LZX compressor seemingly does
196 * do block splitting, although I don't know how fast or useful it is,
199 * Now, back to the entropy encoding problem. The "solution" is to use an
200 * iterative approach to compute a good, but not necessarily optimal,
201 * match/literal sequence. Start with a fixed assignment of symbol costs and
202 * choose an "optimal" match/literal sequence based on those costs, using
203 * shortest-path seach a la Dijkstra's algorithm. Then, for each iteration of
204 * the optimization, update the costs based on the entropy encoding of the
205 * current match/literal sequence, then choose a new match/literal sequence
206 * based on the updated costs. Usually, the actual cost to output the current
207 * match/literal sequence will decrease in each iteration until it converges on
208 * a fixed point. This result may not be the truly optimal match/literal
209 * sequence, but it usually is much better than one chosen by doing a "greedy"
210 * parse where we always chooe the longest match.
212 * An alternative to both greedy parsing and iterative, near-optimal parsing is
213 * "lazy" parsing. Briefly, "lazy" parsing considers just the longest match at
214 * each position, but it waits to choose that match until it has also examined
215 * the next position. This is actually a useful approach; it's used by zlib,
216 * for example. Therefore, for fast compression we combine lazy parsing with
217 * the hash chain max-finder. For normal/high compression we combine
218 * near-optimal parsing with the binary tree match-finder.
220 * Anyway, if you've read through this comment, you hopefully should have a
221 * better idea of why things are done in a certain way in this LZX compressor,
222 * as well as in other compressors for LZ77-based formats (including third-party
223 * ones). In my opinion, the phrase "compression algorithm" is often mis-used
224 * in place of "compression format", since there can be many different
225 * algorithms that all generate compressed data in the same format. The
226 * challenge is to design an algorithm that is efficient but still gives a good
235 #include "wimlib/compressor_ops.h"
236 #include "wimlib/compress_common.h"
237 #include "wimlib/endianness.h"
238 #include "wimlib/error.h"
239 #include "wimlib/lz.h"
240 #include "wimlib/lz_hash.h"
241 #include "wimlib/lz_bt.h"
242 #include "wimlib/lzx.h"
243 #include "wimlib/util.h"
246 #ifdef ENABLE_LZX_DEBUG
247 # include "wimlib/decompress_common.h"
250 #define LZX_OPTIM_ARRAY_SIZE 4096
252 #define LZX_DIV_BLOCK_SIZE 32768
254 #define LZX_CACHE_PER_POS 10
256 #define LZX_CACHE_LEN (LZX_DIV_BLOCK_SIZE * (LZX_CACHE_PER_POS + 1))
257 #define LZX_CACHE_SIZE (LZX_CACHE_LEN * sizeof(struct raw_match))
259 /* Dependent on behavior of lz_bt_get_matches(). */
260 #define LZX_MAX_MATCHES_PER_POS (LZX_MAX_MATCH_LEN - LZX_MIN_MATCH_LEN + 1)
262 /* Codewords for the LZX main, length, and aligned offset Huffman codes */
263 struct lzx_codewords {
264 u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
265 u32 len[LZX_LENCODE_NUM_SYMBOLS];
266 u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
269 /* Codeword lengths (in bits) for the LZX main, length, and aligned offset
272 * A 0 length means the codeword has zero frequency.
275 u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
276 u8 len[LZX_LENCODE_NUM_SYMBOLS];
277 u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
280 /* Costs for the LZX main, length, and aligned offset Huffman symbols.
282 * If a codeword has zero frequency, it must still be assigned some nonzero cost
283 * --- generally a high cost, since even if it gets used in the next iteration,
284 * it probably will not be used very times. */
286 u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
287 u8 len[LZX_LENCODE_NUM_SYMBOLS];
288 u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
291 /* The LZX main, length, and aligned offset Huffman codes */
293 struct lzx_codewords codewords;
294 struct lzx_lens lens;
297 /* Tables for tallying symbol frequencies in the three LZX alphabets */
299 input_idx_t main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
300 input_idx_t len[LZX_LENCODE_NUM_SYMBOLS];
301 input_idx_t aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
304 /* LZX intermediate match/literal format */
308 * 31 1 if a match, 0 if a literal.
310 * 30-25 position slot. This can be at most 50, so it will fit in 6
313 * 8-24 position footer. This is the offset of the real formatted
314 * offset from the position base. This can be at most 17 bits
315 * (since lzx_extra_bits[LZX_MAX_POSITION_SLOTS - 1] is 17).
317 * 0-7 length of match, minus 2. This can be at most
318 * (LZX_MAX_MATCH_LEN - 2) == 255, so it will fit in 8 bits. */
322 /* Specification for an LZX block. */
323 struct lzx_block_spec {
325 /* One of the LZX_BLOCKTYPE_* constants indicating which type of this
329 /* 0-based position in the window at which this block starts. */
330 input_idx_t window_pos;
332 /* The number of bytes of uncompressed data this block represents. */
333 input_idx_t block_size;
335 /* The match/literal sequence for this block. */
336 struct lzx_match *chosen_matches;
338 /* The length of the @chosen_matches sequence. */
339 input_idx_t num_chosen_matches;
341 /* Huffman codes for this block. */
342 struct lzx_codes codes;
345 /* State of the LZX compressor. */
346 struct lzx_compressor {
348 /* The parameters that were used to create the compressor. */
349 struct wimlib_lzx_compressor_params params;
351 /* The buffer of data to be compressed.
353 * 0xe8 byte preprocessing is done directly on the data here before
354 * further compression.
356 * Note that this compressor does *not* use a real sliding window!!!!
357 * It's not needed in the WIM format, since every chunk is compressed
358 * independently. This is by design, to allow random access to the
361 * We reserve a few extra bytes to potentially allow reading off the end
362 * of the array in the match-finding code for optimization purposes
363 * (currently only needed for the hash chain match-finder). */
366 /* Number of bytes of data to be compressed, which is the number of
367 * bytes of data in @window that are actually valid. */
368 input_idx_t window_size;
370 /* Allocated size of the @window. */
371 input_idx_t max_window_size;
373 /* Number of symbols in the main alphabet (depends on the
374 * @max_window_size since it determines the maximum allowed offset). */
375 unsigned num_main_syms;
377 /* The current match offset LRU queue. */
378 struct lzx_lru_queue queue;
380 /* Space for the sequences of matches/literals that were chosen for each
382 struct lzx_match *chosen_matches;
384 /* Information about the LZX blocks the preprocessed input was divided
386 struct lzx_block_spec *block_specs;
388 /* Number of LZX blocks the input was divided into; a.k.a. the number of
389 * elements of @block_specs that are valid. */
392 /* This is simply filled in with zeroes and used to avoid special-casing
393 * the output of the first compressed Huffman code, which conceptually
394 * has a delta taken from a code with all symbols having zero-length
396 struct lzx_codes zero_codes;
398 /* The current cost model. */
399 struct lzx_costs costs;
401 /* Fast algorithm only: Array of hash table links. */
402 input_idx_t *prev_tab;
404 /* Slow algorithm only: Binary tree match-finder. */
407 /* Position in window of next match to return. */
408 input_idx_t match_window_pos;
410 /* The end-of-block position. We can't allow any matches to span this
412 input_idx_t match_window_end;
414 /* Matches found by the match-finder are cached in the following array
415 * to achieve a slight speedup when the same matches are needed on
416 * subsequent passes. This is suboptimal because different matches may
417 * be preferred with different cost models, but seems to be a worthwhile
419 struct raw_match *cached_matches;
420 struct raw_match *cache_ptr;
422 struct raw_match *cache_limit;
424 /* Match-chooser state.
425 * When matches have been chosen, optimum_cur_idx is set to the position
426 * in the window of the next match/literal to return and optimum_end_idx
427 * is set to the position in the window at the end of the last
428 * match/literal to return. */
429 struct lzx_mc_pos_data *optimum;
430 unsigned optimum_cur_idx;
431 unsigned optimum_end_idx;
435 * Match chooser position data:
437 * An array of these structures is used during the match-choosing algorithm.
438 * They correspond to consecutive positions in the window and are used to keep
439 * track of the cost to reach each position, and the match/literal choices that
440 * need to be chosen to reach that position.
442 struct lzx_mc_pos_data {
443 /* The approximate minimum cost, in bits, to reach this position in the
444 * window which has been found so far. */
446 #define MC_INFINITE_COST ((u32)~0UL)
448 /* The union here is just for clarity, since the fields are used in two
449 * slightly different ways. Initially, the @prev structure is filled in
450 * first, and links go from later in the window to earlier in the
451 * window. Later, @next structure is filled in and links go from
452 * earlier in the window to later in the window. */
455 /* Position of the start of the match or literal that
456 * was taken to get to this position in the approximate
457 * minimum-cost parse. */
460 /* Offset (as in an LZ (length, offset) pair) of the
461 * match or literal that was taken to get to this
462 * position in the approximate minimum-cost parse. */
463 input_idx_t match_offset;
466 /* Position at which the match or literal starting at
467 * this position ends in the minimum-cost parse. */
470 /* Offset (as in an LZ (length, offset) pair) of the
471 * match or literal starting at this position in the
472 * approximate minimum-cost parse. */
473 input_idx_t match_offset;
477 /* Adaptive state that exists after an approximate minimum-cost path to
478 * reach this position is taken. */
479 struct lzx_lru_queue queue;
482 /* Returns the LZX position slot that corresponds to a given match offset,
483 * taking into account the recent offset queue and updating it if the offset is
486 lzx_get_position_slot(u32 offset, struct lzx_lru_queue *queue)
488 unsigned position_slot;
490 /* See if the offset was recently used. */
491 for (int i = 0; i < LZX_NUM_RECENT_OFFSETS; i++) {
492 if (offset == queue->R[i]) {
495 /* Bring the repeat offset to the front of the
496 * queue. Note: this is, in fact, not a real
497 * LRU queue because repeat matches are simply
498 * swapped to the front. */
499 swap(queue->R[0], queue->R[i]);
501 /* The resulting position slot is simply the first index
502 * at which the offset was found in the queue. */
507 /* The offset was not recently used; look up its real position slot. */
508 position_slot = lzx_get_position_slot_raw(offset + LZX_OFFSET_OFFSET);
510 /* Bring the new offset to the front of the queue. */
511 for (int i = LZX_NUM_RECENT_OFFSETS - 1; i > 0; i--)
512 queue->R[i] = queue->R[i - 1];
513 queue->R[0] = offset;
515 return position_slot;
518 /* Build the main, length, and aligned offset Huffman codes used in LZX.
520 * This takes as input the frequency tables for each code and produces as output
521 * a set of tables that map symbols to codewords and codeword lengths. */
523 lzx_make_huffman_codes(const struct lzx_freqs *freqs,
524 struct lzx_codes *codes,
525 unsigned num_main_syms)
527 make_canonical_huffman_code(num_main_syms,
528 LZX_MAX_MAIN_CODEWORD_LEN,
531 codes->codewords.main);
533 make_canonical_huffman_code(LZX_LENCODE_NUM_SYMBOLS,
534 LZX_MAX_LEN_CODEWORD_LEN,
537 codes->codewords.len);
539 make_canonical_huffman_code(LZX_ALIGNEDCODE_NUM_SYMBOLS,
540 LZX_MAX_ALIGNED_CODEWORD_LEN,
543 codes->codewords.aligned);
547 * Output a precomputed LZX match.
550 * The bitstream to which to write the match.
552 * The type of the LZX block (LZX_BLOCKTYPE_ALIGNED or
553 * LZX_BLOCKTYPE_VERBATIM)
555 * The match, as a (length, offset) pair.
557 * Pointer to a structure that contains the codewords for the main, length,
558 * and aligned offset Huffman codes for the current LZX compressed block.
561 lzx_write_match(struct output_bitstream *out, int block_type,
562 struct lzx_match match, const struct lzx_codes *codes)
564 /* low 8 bits are the match length minus 2 */
565 unsigned match_len_minus_2 = match.data & 0xff;
566 /* Next 17 bits are the position footer */
567 unsigned position_footer = (match.data >> 8) & 0x1ffff; /* 17 bits */
568 /* Next 6 bits are the position slot. */
569 unsigned position_slot = (match.data >> 25) & 0x3f; /* 6 bits */
572 unsigned main_symbol;
573 unsigned num_extra_bits;
574 unsigned verbatim_bits;
575 unsigned aligned_bits;
577 /* If the match length is less than MIN_MATCH_LEN (= 2) +
578 * NUM_PRIMARY_LENS (= 7), the length header contains
579 * the match length minus MIN_MATCH_LEN, and there is no
582 * Otherwise, the length header contains
583 * NUM_PRIMARY_LENS, and the length footer contains
584 * the match length minus NUM_PRIMARY_LENS minus
586 if (match_len_minus_2 < LZX_NUM_PRIMARY_LENS) {
587 len_header = match_len_minus_2;
589 len_header = LZX_NUM_PRIMARY_LENS;
590 len_footer = match_len_minus_2 - LZX_NUM_PRIMARY_LENS;
593 /* Combine the position slot with the length header into a single symbol
594 * that will be encoded with the main code.
596 * The actual main symbol is offset by LZX_NUM_CHARS because values
597 * under LZX_NUM_CHARS are used to indicate a literal byte rather than a
599 main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS;
601 /* Output main symbol. */
602 bitstream_put_bits(out, codes->codewords.main[main_symbol],
603 codes->lens.main[main_symbol]);
605 /* If there is a length footer, output it using the
606 * length Huffman code. */
607 if (len_header == LZX_NUM_PRIMARY_LENS)
608 bitstream_put_bits(out, codes->codewords.len[len_footer],
609 codes->lens.len[len_footer]);
611 num_extra_bits = lzx_get_num_extra_bits(position_slot);
613 /* For aligned offset blocks with at least 3 extra bits, output the
614 * verbatim bits literally, then the aligned bits encoded using the
615 * aligned offset code. Otherwise, only the verbatim bits need to be
617 if ((block_type == LZX_BLOCKTYPE_ALIGNED) && (num_extra_bits >= 3)) {
619 verbatim_bits = position_footer >> 3;
620 bitstream_put_bits(out, verbatim_bits,
623 aligned_bits = (position_footer & 7);
624 bitstream_put_bits(out,
625 codes->codewords.aligned[aligned_bits],
626 codes->lens.aligned[aligned_bits]);
628 /* verbatim bits is the same as the position
629 * footer, in this case. */
630 bitstream_put_bits(out, position_footer, num_extra_bits);
634 /* Output an LZX literal (encoded with the main Huffman code). */
636 lzx_write_literal(struct output_bitstream *out, u8 literal,
637 const struct lzx_codes *codes)
639 bitstream_put_bits(out,
640 codes->codewords.main[literal],
641 codes->lens.main[literal]);
645 lzx_build_precode(const u8 lens[restrict],
646 const u8 prev_lens[restrict],
647 const unsigned num_syms,
648 input_idx_t precode_freqs[restrict LZX_PRECODE_NUM_SYMBOLS],
649 u8 output_syms[restrict num_syms],
650 u8 precode_lens[restrict LZX_PRECODE_NUM_SYMBOLS],
651 u32 precode_codewords[restrict LZX_PRECODE_NUM_SYMBOLS],
652 unsigned *num_additional_bits_ret)
654 memset(precode_freqs, 0,
655 LZX_PRECODE_NUM_SYMBOLS * sizeof(precode_freqs[0]));
657 /* Since the code word lengths use a form of RLE encoding, the goal here
658 * is to find each run of identical lengths when going through them in
659 * symbol order (including runs of length 1). For each run, as many
660 * lengths are encoded using RLE as possible, and the rest are output
663 * output_syms[] will be filled in with the length symbols that will be
664 * output, including RLE codes, not yet encoded using the precode.
666 * cur_run_len keeps track of how many code word lengths are in the
667 * current run of identical lengths. */
668 unsigned output_syms_idx = 0;
669 unsigned cur_run_len = 1;
670 unsigned num_additional_bits = 0;
671 for (unsigned i = 1; i <= num_syms; i++) {
673 if (i != num_syms && lens[i] == lens[i - 1]) {
674 /* Still in a run--- keep going. */
679 /* Run ended! Check if it is a run of zeroes or a run of
682 /* The symbol that was repeated in the run--- not to be confused
683 * with the length *of* the run (cur_run_len) */
684 unsigned len_in_run = lens[i - 1];
686 if (len_in_run == 0) {
687 /* A run of 0's. Encode it in as few length
688 * codes as we can. */
690 /* The magic length 18 indicates a run of 20 + n zeroes,
691 * where n is an uncompressed literal 5-bit integer that
692 * follows the magic length. */
693 while (cur_run_len >= 20) {
694 unsigned additional_bits;
696 additional_bits = min(cur_run_len - 20, 0x1f);
697 num_additional_bits += 5;
699 output_syms[output_syms_idx++] = 18;
700 output_syms[output_syms_idx++] = additional_bits;
701 cur_run_len -= 20 + additional_bits;
704 /* The magic length 17 indicates a run of 4 + n zeroes,
705 * where n is an uncompressed literal 4-bit integer that
706 * follows the magic length. */
707 while (cur_run_len >= 4) {
708 unsigned additional_bits;
710 additional_bits = min(cur_run_len - 4, 0xf);
711 num_additional_bits += 4;
713 output_syms[output_syms_idx++] = 17;
714 output_syms[output_syms_idx++] = additional_bits;
715 cur_run_len -= 4 + additional_bits;
720 /* A run of nonzero lengths. */
722 /* The magic length 19 indicates a run of 4 + n
723 * nonzeroes, where n is a literal bit that follows the
724 * magic length, and where the value of the lengths in
725 * the run is given by an extra length symbol, encoded
726 * with the precode, that follows the literal bit.
728 * The extra length symbol is encoded as a difference
729 * from the length of the codeword for the first symbol
730 * in the run in the previous code.
732 while (cur_run_len >= 4) {
733 unsigned additional_bits;
736 additional_bits = (cur_run_len > 4);
737 num_additional_bits += 1;
738 delta = (signed char)prev_lens[i - cur_run_len] -
739 (signed char)len_in_run;
743 precode_freqs[(unsigned char)delta]++;
744 output_syms[output_syms_idx++] = 19;
745 output_syms[output_syms_idx++] = additional_bits;
746 output_syms[output_syms_idx++] = delta;
747 cur_run_len -= 4 + additional_bits;
751 /* Any remaining lengths in the run are outputted without RLE,
752 * as a difference from the length of that codeword in the
754 while (cur_run_len > 0) {
757 delta = (signed char)prev_lens[i - cur_run_len] -
758 (signed char)len_in_run;
762 precode_freqs[(unsigned char)delta]++;
763 output_syms[output_syms_idx++] = delta;
770 /* Build the precode from the frequencies of the length symbols. */
772 make_canonical_huffman_code(LZX_PRECODE_NUM_SYMBOLS,
773 LZX_MAX_PRE_CODEWORD_LEN,
774 precode_freqs, precode_lens,
777 *num_additional_bits_ret = num_additional_bits;
779 return output_syms_idx;
783 * Output a Huffman code in the compressed form used in LZX.
785 * The Huffman code is represented in the output as a logical series of codeword
786 * lengths from which the Huffman code, which must be in canonical form, can be
789 * The codeword lengths are themselves compressed using a separate Huffman code,
790 * the "precode", which contains a symbol for each possible codeword length in
791 * the larger code as well as several special symbols to represent repeated
792 * codeword lengths (a form of run-length encoding). The precode is itself
793 * constructed in canonical form, and its codeword lengths are represented
794 * literally in 20 4-bit fields that immediately precede the compressed codeword
795 * lengths of the larger code.
797 * Furthermore, the codeword lengths of the larger code are actually represented
798 * as deltas from the codeword lengths of the corresponding code in the previous
802 * Bitstream to which to write the compressed Huffman code.
804 * The codeword lengths, indexed by symbol, in the Huffman code.
806 * The codeword lengths, indexed by symbol, in the corresponding Huffman
807 * code in the previous block, or all zeroes if this is the first block.
809 * The number of symbols in the Huffman code.
812 lzx_write_compressed_code(struct output_bitstream *out,
813 const u8 lens[restrict],
814 const u8 prev_lens[restrict],
817 input_idx_t precode_freqs[LZX_PRECODE_NUM_SYMBOLS];
818 u8 output_syms[num_syms];
819 u8 precode_lens[LZX_PRECODE_NUM_SYMBOLS];
820 u32 precode_codewords[LZX_PRECODE_NUM_SYMBOLS];
822 unsigned num_output_syms;
826 num_output_syms = lzx_build_precode(lens,
835 /* Write the lengths of the precode codes to the output. */
836 for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++)
837 bitstream_put_bits(out, precode_lens[i],
838 LZX_PRECODE_ELEMENT_SIZE);
840 /* Write the length symbols, encoded with the precode, to the output. */
842 for (i = 0; i < num_output_syms; ) {
843 precode_sym = output_syms[i++];
845 bitstream_put_bits(out, precode_codewords[precode_sym],
846 precode_lens[precode_sym]);
847 switch (precode_sym) {
849 bitstream_put_bits(out, output_syms[i++], 4);
852 bitstream_put_bits(out, output_syms[i++], 5);
855 bitstream_put_bits(out, output_syms[i++], 1);
856 bitstream_put_bits(out,
857 precode_codewords[output_syms[i]],
858 precode_lens[output_syms[i]]);
868 * Write all matches and literal bytes (which were precomputed) in an LZX
869 * compressed block to the output bitstream in the final compressed
873 * The output bitstream.
875 * The chosen type of the LZX compressed block (LZX_BLOCKTYPE_ALIGNED or
876 * LZX_BLOCKTYPE_VERBATIM).
878 * The array of matches/literals to output.
880 * Number of matches/literals to output (length of @match_tab).
882 * The main, length, and aligned offset Huffman codes for the current
883 * LZX compressed block.
886 lzx_write_matches_and_literals(struct output_bitstream *ostream,
888 const struct lzx_match match_tab[],
889 unsigned match_count,
890 const struct lzx_codes *codes)
892 for (unsigned i = 0; i < match_count; i++) {
893 struct lzx_match match = match_tab[i];
895 /* The high bit of the 32-bit intermediate representation
896 * indicates whether the item is an actual LZ-style match (1) or
897 * a literal byte (0). */
898 if (match.data & 0x80000000)
899 lzx_write_match(ostream, block_type, match, codes);
901 lzx_write_literal(ostream, match.data, codes);
906 lzx_assert_codes_valid(const struct lzx_codes * codes, unsigned num_main_syms)
908 #ifdef ENABLE_LZX_DEBUG
911 for (i = 0; i < num_main_syms; i++)
912 LZX_ASSERT(codes->lens.main[i] <= LZX_MAX_MAIN_CODEWORD_LEN);
914 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
915 LZX_ASSERT(codes->lens.len[i] <= LZX_MAX_LEN_CODEWORD_LEN);
917 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
918 LZX_ASSERT(codes->lens.aligned[i] <= LZX_MAX_ALIGNED_CODEWORD_LEN);
920 const unsigned tablebits = 10;
921 u16 decode_table[(1 << tablebits) +
922 (2 * max(num_main_syms, LZX_LENCODE_NUM_SYMBOLS))]
923 _aligned_attribute(DECODE_TABLE_ALIGNMENT);
924 LZX_ASSERT(0 == make_huffman_decode_table(decode_table,
926 min(tablebits, LZX_MAINCODE_TABLEBITS),
928 LZX_MAX_MAIN_CODEWORD_LEN));
929 LZX_ASSERT(0 == make_huffman_decode_table(decode_table,
930 LZX_LENCODE_NUM_SYMBOLS,
931 min(tablebits, LZX_LENCODE_TABLEBITS),
933 LZX_MAX_LEN_CODEWORD_LEN));
934 LZX_ASSERT(0 == make_huffman_decode_table(decode_table,
935 LZX_ALIGNEDCODE_NUM_SYMBOLS,
936 min(tablebits, LZX_ALIGNEDCODE_TABLEBITS),
938 LZX_MAX_ALIGNED_CODEWORD_LEN));
939 #endif /* ENABLE_LZX_DEBUG */
942 /* Write an LZX aligned offset or verbatim block to the output. */
944 lzx_write_compressed_block(int block_type,
946 unsigned max_window_size,
947 unsigned num_main_syms,
948 struct lzx_match * chosen_matches,
949 unsigned num_chosen_matches,
950 const struct lzx_codes * codes,
951 const struct lzx_codes * prev_codes,
952 struct output_bitstream * ostream)
956 LZX_ASSERT(block_type == LZX_BLOCKTYPE_ALIGNED ||
957 block_type == LZX_BLOCKTYPE_VERBATIM);
958 lzx_assert_codes_valid(codes, num_main_syms);
960 /* The first three bits indicate the type of block and are one of the
961 * LZX_BLOCKTYPE_* constants. */
962 bitstream_put_bits(ostream, block_type, 3);
964 /* Output the block size.
966 * The original LZX format seemed to always encode the block size in 3
967 * bytes. However, the implementation in WIMGAPI, as used in WIM files,
968 * uses the first bit to indicate whether the block is the default size
969 * (32768) or a different size given explicitly by the next 16 bits.
971 * By default, this compressor uses a window size of 32768 and therefore
972 * follows the WIMGAPI behavior. However, this compressor also supports
973 * window sizes greater than 32768 bytes, which do not appear to be
974 * supported by WIMGAPI. In such cases, we retain the default size bit
975 * to mean a size of 32768 bytes but output non-default block size in 24
976 * bits rather than 16. The compatibility of this behavior is unknown
977 * because WIMs created with chunk size greater than 32768 can seemingly
978 * only be opened by wimlib anyway. */
979 if (block_size == LZX_DEFAULT_BLOCK_SIZE) {
980 bitstream_put_bits(ostream, 1, 1);
982 bitstream_put_bits(ostream, 0, 1);
984 if (max_window_size >= 65536)
985 bitstream_put_bits(ostream, block_size >> 16, 8);
987 bitstream_put_bits(ostream, block_size, 16);
990 /* Write out lengths of the main code. Note that the LZX specification
991 * incorrectly states that the aligned offset code comes after the
992 * length code, but in fact it is the very first code to be written
993 * (before the main code). */
994 if (block_type == LZX_BLOCKTYPE_ALIGNED)
995 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
996 bitstream_put_bits(ostream, codes->lens.aligned[i],
997 LZX_ALIGNEDCODE_ELEMENT_SIZE);
999 LZX_DEBUG("Writing main code...");
1001 /* Write the precode and lengths for the first LZX_NUM_CHARS symbols in
1002 * the main code, which are the codewords for literal bytes. */
1003 lzx_write_compressed_code(ostream,
1005 prev_codes->lens.main,
1008 /* Write the precode and lengths for the rest of the main code, which
1009 * are the codewords for match headers. */
1010 lzx_write_compressed_code(ostream,
1011 codes->lens.main + LZX_NUM_CHARS,
1012 prev_codes->lens.main + LZX_NUM_CHARS,
1013 num_main_syms - LZX_NUM_CHARS);
1015 LZX_DEBUG("Writing length code...");
1017 /* Write the precode and lengths for the length code. */
1018 lzx_write_compressed_code(ostream,
1020 prev_codes->lens.len,
1021 LZX_LENCODE_NUM_SYMBOLS);
1023 LZX_DEBUG("Writing matches and literals...");
1025 /* Write the actual matches and literals. */
1026 lzx_write_matches_and_literals(ostream, block_type,
1027 chosen_matches, num_chosen_matches,
1030 LZX_DEBUG("Done writing block.");
1033 /* Write out the LZX blocks that were computed. */
1035 lzx_write_all_blocks(struct lzx_compressor *ctx, struct output_bitstream *ostream)
1038 const struct lzx_codes *prev_codes = &ctx->zero_codes;
1039 for (unsigned i = 0; i < ctx->num_blocks; i++) {
1040 const struct lzx_block_spec *spec = &ctx->block_specs[i];
1042 LZX_DEBUG("Writing block %u/%u (type=%d, size=%u, num_chosen_matches=%u)...",
1043 i + 1, ctx->num_blocks,
1044 spec->block_type, spec->block_size,
1045 spec->num_chosen_matches);
1047 lzx_write_compressed_block(spec->block_type,
1049 ctx->max_window_size,
1051 spec->chosen_matches,
1052 spec->num_chosen_matches,
1057 prev_codes = &spec->codes;
1061 /* Constructs an LZX match from a literal byte and updates the main code symbol
1064 lzx_tally_literal(u8 lit, struct lzx_freqs *freqs)
1070 /* Constructs an LZX match from an offset and a length, and updates the LRU
1071 * queue and the frequency of symbols in the main, length, and aligned offset
1072 * alphabets. The return value is a 32-bit number that provides the match in an
1073 * intermediate representation documented below. */
1075 lzx_tally_match(unsigned match_len, u32 match_offset,
1076 struct lzx_freqs *freqs, struct lzx_lru_queue *queue)
1078 unsigned position_slot;
1079 unsigned position_footer;
1081 unsigned main_symbol;
1082 unsigned len_footer;
1083 unsigned adjusted_match_len;
1085 LZX_ASSERT(match_len >= LZX_MIN_MATCH_LEN && match_len <= LZX_MAX_MATCH_LEN);
1087 /* The match offset shall be encoded as a position slot (itself encoded
1088 * as part of the main symbol) and a position footer. */
1089 position_slot = lzx_get_position_slot(match_offset, queue);
1090 position_footer = (match_offset + LZX_OFFSET_OFFSET) &
1091 ((1U << lzx_get_num_extra_bits(position_slot)) - 1);
1093 /* The match length shall be encoded as a length header (itself encoded
1094 * as part of the main symbol) and an optional length footer. */
1095 adjusted_match_len = match_len - LZX_MIN_MATCH_LEN;
1096 if (adjusted_match_len < LZX_NUM_PRIMARY_LENS) {
1097 /* No length footer needed. */
1098 len_header = adjusted_match_len;
1100 /* Length footer needed. It will be encoded using the length
1102 len_header = LZX_NUM_PRIMARY_LENS;
1103 len_footer = adjusted_match_len - LZX_NUM_PRIMARY_LENS;
1104 freqs->len[len_footer]++;
1107 /* Account for the main symbol. */
1108 main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS;
1110 freqs->main[main_symbol]++;
1112 /* In an aligned offset block, 3 bits of the position footer are output
1113 * as an aligned offset symbol. Account for this, although we may
1114 * ultimately decide to output the block as verbatim. */
1116 /* The following check is equivalent to:
1118 * if (lzx_extra_bits[position_slot] >= 3)
1120 * Note that this correctly excludes position slots that correspond to
1121 * recent offsets. */
1122 if (position_slot >= 8)
1123 freqs->aligned[position_footer & 7]++;
1125 /* Pack the position slot, position footer, and match length into an
1126 * intermediate representation. See `struct lzx_match' for details.
1128 LZX_ASSERT(LZX_MAX_POSITION_SLOTS <= 64);
1129 LZX_ASSERT(lzx_get_num_extra_bits(LZX_MAX_POSITION_SLOTS - 1) <= 17);
1130 LZX_ASSERT(LZX_MAX_MATCH_LEN - LZX_MIN_MATCH_LEN + 1 <= 256);
1132 LZX_ASSERT(position_slot <= (1U << (31 - 25)) - 1);
1133 LZX_ASSERT(position_footer <= (1U << (25 - 8)) - 1);
1134 LZX_ASSERT(adjusted_match_len <= (1U << (8 - 0)) - 1);
1136 (position_slot << 25) |
1137 (position_footer << 8) |
1138 (adjusted_match_len);
1141 struct lzx_record_ctx {
1142 struct lzx_freqs freqs;
1143 struct lzx_lru_queue queue;
1144 struct lzx_match *matches;
1148 lzx_record_match(unsigned len, unsigned offset, void *_ctx)
1150 struct lzx_record_ctx *ctx = _ctx;
1152 (ctx->matches++)->data = lzx_tally_match(len, offset, &ctx->freqs, &ctx->queue);
1156 lzx_record_literal(u8 lit, void *_ctx)
1158 struct lzx_record_ctx *ctx = _ctx;
1160 (ctx->matches++)->data = lzx_tally_literal(lit, &ctx->freqs);
1163 /* Returns the cost, in bits, to output a literal byte using the specified cost
1166 lzx_literal_cost(u8 c, const struct lzx_costs * costs)
1168 return costs->main[c];
1171 /* Given a (length, offset) pair that could be turned into a valid LZX match as
1172 * well as costs for the codewords in the main, length, and aligned Huffman
1173 * codes, return the approximate number of bits it will take to represent this
1174 * match in the compressed output. Take into account the match offset LRU
1175 * queue and optionally update it. */
1177 lzx_match_cost(unsigned length, u32 offset, const struct lzx_costs *costs,
1178 struct lzx_lru_queue *queue)
1180 unsigned position_slot;
1181 unsigned len_header, main_symbol;
1182 unsigned num_extra_bits;
1185 position_slot = lzx_get_position_slot(offset, queue);
1187 len_header = min(length - LZX_MIN_MATCH_LEN, LZX_NUM_PRIMARY_LENS);
1188 main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS;
1190 /* Account for main symbol. */
1191 cost += costs->main[main_symbol];
1193 /* Account for extra position information. */
1194 num_extra_bits = lzx_get_num_extra_bits(position_slot);
1195 if (num_extra_bits >= 3) {
1196 cost += num_extra_bits - 3;
1197 cost += costs->aligned[(offset + LZX_OFFSET_OFFSET) & 7];
1199 cost += num_extra_bits;
1202 /* Account for extra length information. */
1203 if (len_header == LZX_NUM_PRIMARY_LENS)
1204 cost += costs->len[length - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS];
1210 /* Set the cost model @ctx->costs from the Huffman codeword lengths specified in
1213 * The cost model and codeword lengths are almost the same thing, but the
1214 * Huffman codewords with length 0 correspond to symbols with zero frequency
1215 * that still need to be assigned actual costs. The specific values assigned
1216 * are arbitrary, but they should be fairly high (near the maximum codeword
1217 * length) to take into account the fact that uses of these symbols are expected
1220 lzx_set_costs(struct lzx_compressor * ctx, const struct lzx_lens * lens)
1223 unsigned num_main_syms = ctx->num_main_syms;
1226 for (i = 0; i < num_main_syms; i++) {
1227 ctx->costs.main[i] = lens->main[i];
1228 if (ctx->costs.main[i] == 0)
1229 ctx->costs.main[i] = ctx->params.alg_params.slow.main_nostat_cost;
1233 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) {
1234 ctx->costs.len[i] = lens->len[i];
1235 if (ctx->costs.len[i] == 0)
1236 ctx->costs.len[i] = ctx->params.alg_params.slow.len_nostat_cost;
1239 /* Aligned offset code */
1240 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1241 ctx->costs.aligned[i] = lens->aligned[i];
1242 if (ctx->costs.aligned[i] == 0)
1243 ctx->costs.aligned[i] = ctx->params.alg_params.slow.aligned_nostat_cost;
1247 /* Retrieve a list of matches available at the next position in the input.
1249 * A pointer to the matches array is written into @matches_ret, and the return
1250 * value is the number of matches found. */
1252 lzx_get_matches(struct lzx_compressor *ctx,
1253 const struct raw_match **matches_ret)
1255 struct raw_match *cache_ptr;
1256 struct raw_match *matches;
1257 unsigned num_matches;
1259 LZX_ASSERT(ctx->match_window_pos < ctx->match_window_end);
1261 cache_ptr = ctx->cache_ptr;
1262 matches = cache_ptr + 1;
1263 if (ctx->matches_cached) {
1264 num_matches = cache_ptr->len;
1266 num_matches = lz_bt_get_matches(&ctx->mf, matches);
1267 cache_ptr->len = num_matches;
1270 /* Don't allow matches to span the end of an LZX block. */
1271 if (ctx->match_window_end < ctx->window_size && num_matches != 0) {
1272 unsigned limit = ctx->match_window_end - ctx->match_window_pos;
1274 if (limit >= LZX_MIN_MATCH_LEN) {
1276 unsigned i = num_matches - 1;
1278 if (matches[i].len >= limit) {
1279 matches[i].len = limit;
1281 /* Truncation might produce multiple
1282 * matches with length 'limit'. Keep at
1284 num_matches = i + 1;
1290 cache_ptr->len = num_matches;
1294 fprintf(stderr, "Pos %u/%u: %u matches\n",
1295 ctx->match_window_pos, ctx->window_size, num_matches);
1296 for (unsigned i = 0; i < num_matches; i++)
1297 fprintf(stderr, "\tLen %u Offset %u\n", matches[i].len, matches[i].offset);
1300 #ifdef ENABLE_LZX_DEBUG
1301 for (unsigned i = 0; i < num_matches; i++) {
1302 LZX_ASSERT(matches[i].len >= LZX_MIN_MATCH_LEN);
1303 LZX_ASSERT(matches[i].len <= LZX_MAX_MATCH_LEN);
1304 LZX_ASSERT(matches[i].len <= ctx->match_window_end - ctx->match_window_pos);
1305 LZX_ASSERT(matches[i].offset > 0);
1306 LZX_ASSERT(matches[i].offset <= ctx->match_window_pos);
1307 LZX_ASSERT(!memcmp(&ctx->window[ctx->match_window_pos],
1308 &ctx->window[ctx->match_window_pos - matches[i].offset],
1311 LZX_ASSERT(matches[i].len > matches[i - 1].len);
1312 LZX_ASSERT(matches[i].offset > matches[i - 1].offset);
1316 ctx->match_window_pos++;
1317 ctx->cache_ptr = matches + num_matches;
1318 *matches_ret = matches;
1323 lzx_skip_bytes(struct lzx_compressor *ctx, unsigned n)
1325 struct raw_match *cache_ptr;
1327 LZX_ASSERT(n <= ctx->match_window_end - ctx->match_window_pos);
1329 cache_ptr = ctx->cache_ptr;
1330 ctx->match_window_pos += n;
1331 if (ctx->matches_cached) {
1333 cache_ptr += 1 + cache_ptr->len;
1335 lz_bt_skip_positions(&ctx->mf, n);
1341 ctx->cache_ptr = cache_ptr;
1345 * Reverse the linked list of near-optimal matches so that they can be returned
1346 * in forwards order.
1348 * Returns the first match in the list.
1350 static struct raw_match
1351 lzx_match_chooser_reverse_list(struct lzx_compressor *ctx, unsigned cur_pos)
1353 unsigned prev_link, saved_prev_link;
1354 unsigned prev_match_offset, saved_prev_match_offset;
1356 ctx->optimum_end_idx = cur_pos;
1358 saved_prev_link = ctx->optimum[cur_pos].prev.link;
1359 saved_prev_match_offset = ctx->optimum[cur_pos].prev.match_offset;
1362 prev_link = saved_prev_link;
1363 prev_match_offset = saved_prev_match_offset;
1365 saved_prev_link = ctx->optimum[prev_link].prev.link;
1366 saved_prev_match_offset = ctx->optimum[prev_link].prev.match_offset;
1368 ctx->optimum[prev_link].next.link = cur_pos;
1369 ctx->optimum[prev_link].next.match_offset = prev_match_offset;
1371 cur_pos = prev_link;
1372 } while (cur_pos != 0);
1374 ctx->optimum_cur_idx = ctx->optimum[0].next.link;
1376 return (struct raw_match)
1377 { .len = ctx->optimum_cur_idx,
1378 .offset = ctx->optimum[0].next.match_offset,
1383 * lzx_get_near_optimal_match() -
1385 * Choose an approximately optimal match or literal to use at the next position
1386 * in the string, or "window", being LZ-encoded.
1388 * This is based on algorithms used in 7-Zip, including the DEFLATE encoder
1389 * and the LZMA encoder, written by Igor Pavlov.
1391 * Unlike a greedy parser that always takes the longest match, or even a "lazy"
1392 * parser with one match/literal look-ahead like zlib, the algorithm used here
1393 * may look ahead many matches/literals to determine the approximately optimal
1394 * match/literal to code next. The motivation is that the compression ratio is
1395 * improved if the compressor can do things like use a shorter-than-possible
1396 * match in order to allow a longer match later, and also take into account the
1397 * estimated real cost of coding each match/literal based on the underlying
1400 * Still, this is not a true optimal parser for several reasons:
1402 * - Real compression formats use entropy encoding of the literal/match
1403 * sequence, so the real cost of coding each match or literal is unknown until
1404 * the parse is fully determined. It can be approximated based on iterative
1405 * parses, but the end result is not guaranteed to be globally optimal.
1407 * - Very long matches are chosen immediately. This is because locations with
1408 * long matches are likely to have many possible alternatives that would cause
1409 * slow optimal parsing, but also such locations are already highly
1410 * compressible so it is not too harmful to just grab the longest match.
1412 * - Not all possible matches at each location are considered because the
1413 * underlying match-finder limits the number and type of matches produced at
1414 * each position. For example, for a given match length it's usually not
1415 * worth it to only consider matches other than the lowest-offset match,
1416 * except in the case of a repeat offset.
1418 * - Although we take into account the adaptive state (in LZX, the recent offset
1419 * queue), coding decisions made with respect to the adaptive state will be
1420 * locally optimal but will not necessarily be globally optimal. This is
1421 * because the algorithm only keeps the least-costly path to get to a given
1422 * location and does not take into account that a slightly more costly path
1423 * could result in a different adaptive state that ultimately results in a
1424 * lower global cost.
1426 * - The array space used by this function is bounded, so in degenerate cases it
1427 * is forced to start returning matches/literals before the algorithm has
1430 * Each call to this function does one of two things:
1432 * 1. Build a sequence of near-optimal matches/literals, up to some point, that
1433 * will be returned by subsequent calls to this function, then return the
1438 * 2. Return the next match/literal previously computed by a call to this
1441 * The return value is a (length, offset) pair specifying the match or literal
1442 * chosen. For literals, the length is 0 or 1 and the offset is meaningless.
1444 static struct raw_match
1445 lzx_get_near_optimal_match(struct lzx_compressor *ctx)
1447 unsigned num_matches;
1448 const struct raw_match *matches;
1449 const struct raw_match *matchptr;
1450 struct raw_match match;
1451 unsigned longest_len;
1452 unsigned longest_rep_len;
1453 u32 longest_rep_offset;
1457 if (ctx->optimum_cur_idx != ctx->optimum_end_idx) {
1458 /* Case 2: Return the next match/literal already found. */
1459 match.len = ctx->optimum[ctx->optimum_cur_idx].next.link -
1460 ctx->optimum_cur_idx;
1461 match.offset = ctx->optimum[ctx->optimum_cur_idx].next.match_offset;
1463 ctx->optimum_cur_idx = ctx->optimum[ctx->optimum_cur_idx].next.link;
1467 /* Case 1: Compute a new list of matches/literals to return. */
1469 ctx->optimum_cur_idx = 0;
1470 ctx->optimum_end_idx = 0;
1472 /* Search for matches at recent offsets. Only keep the one with the
1473 * longest match length. */
1474 longest_rep_len = LZX_MIN_MATCH_LEN - 1;
1475 if (ctx->match_window_pos >= 1) {
1476 unsigned limit = min(LZX_MAX_MATCH_LEN,
1477 ctx->match_window_end - ctx->match_window_pos);
1478 for (int i = 0; i < LZX_NUM_RECENT_OFFSETS; i++) {
1479 u32 offset = ctx->queue.R[i];
1480 const u8 *strptr = &ctx->window[ctx->match_window_pos];
1481 const u8 *matchptr = strptr - offset;
1483 while (len < limit && strptr[len] == matchptr[len])
1485 if (len > longest_rep_len) {
1486 longest_rep_len = len;
1487 longest_rep_offset = offset;
1492 /* If there's a long match with a recent offset, take it. */
1493 if (longest_rep_len >= ctx->params.alg_params.slow.nice_match_length) {
1494 lzx_skip_bytes(ctx, longest_rep_len);
1495 return (struct raw_match) {
1496 .len = longest_rep_len,
1497 .offset = longest_rep_offset,
1501 /* Search other matches. */
1502 num_matches = lzx_get_matches(ctx, &matches);
1504 /* If there's a long match, take it. */
1506 longest_len = matches[num_matches - 1].len;
1507 if (longest_len >= ctx->params.alg_params.slow.nice_match_length) {
1508 lzx_skip_bytes(ctx, longest_len - 1);
1509 return matches[num_matches - 1];
1515 /* Calculate the cost to reach the next position by coding a literal.
1517 ctx->optimum[1].queue = ctx->queue;
1518 ctx->optimum[1].cost = lzx_literal_cost(ctx->window[ctx->match_window_pos - 1],
1520 ctx->optimum[1].prev.link = 0;
1522 /* Calculate the cost to reach any position up to and including that
1523 * reached by the longest match. */
1525 for (unsigned len = 2; len <= longest_len; len++) {
1526 u32 offset = matchptr->offset;
1528 ctx->optimum[len].queue = ctx->queue;
1529 ctx->optimum[len].prev.link = 0;
1530 ctx->optimum[len].prev.match_offset = offset;
1531 ctx->optimum[len].cost = lzx_match_cost(len, offset, &ctx->costs,
1532 &ctx->optimum[len].queue);
1533 if (len == matchptr->len)
1536 end_pos = longest_len;
1538 if (longest_rep_len >= LZX_MIN_MATCH_LEN) {
1539 struct lzx_lru_queue queue;
1542 while (end_pos < longest_rep_len)
1543 ctx->optimum[++end_pos].cost = MC_INFINITE_COST;
1546 cost = lzx_match_cost(longest_rep_len, longest_rep_offset,
1547 &ctx->costs, &queue);
1548 if (cost <= ctx->optimum[longest_rep_len].cost) {
1549 ctx->optimum[longest_rep_len].queue = queue;
1550 ctx->optimum[longest_rep_len].prev.link = 0;
1551 ctx->optimum[longest_rep_len].prev.match_offset = longest_rep_offset;
1552 ctx->optimum[longest_rep_len].cost = cost;
1556 /* Step forward, calculating the estimated minimum cost to reach each
1557 * position. The algorithm may find multiple paths to reach each
1558 * position; only the lowest-cost path is saved.
1560 * The progress of the parse is tracked in the @ctx->optimum array, which
1561 * for each position contains the minimum cost to reach that position,
1562 * the index of the start of the match/literal taken to reach that
1563 * position through the minimum-cost path, the offset of the match taken
1564 * (not relevant for literals), and the adaptive state that will exist
1565 * at that position after the minimum-cost path is taken. The @cur_pos
1566 * variable stores the position at which the algorithm is currently
1567 * considering coding choices, and the @end_pos variable stores the
1568 * greatest position at which the costs of coding choices have been
1569 * saved. (Actually, the algorithm guarantees that all positions up to
1570 * and including @end_pos are reachable by at least one path.)
1572 * The loop terminates when any one of the following conditions occurs:
1574 * 1. A match with length greater than or equal to @nice_match_length is
1575 * found. When this occurs, the algorithm chooses this match
1576 * unconditionally, and consequently the near-optimal match/literal
1577 * sequence up to and including that match is fully determined and it
1578 * can begin returning the match/literal list.
1580 * 2. @cur_pos reaches a position not overlapped by a preceding match.
1581 * In such cases, the near-optimal match/literal sequence up to
1582 * @cur_pos is fully determined and it can begin returning the
1583 * match/literal list.
1585 * 3. Failing either of the above in a degenerate case, the loop
1586 * terminates when space in the @ctx->optimum array is exhausted.
1587 * This terminates the algorithm and forces it to start returning
1588 * matches/literals even though they may not be globally optimal.
1590 * Upon loop termination, a nonempty list of matches/literals will have
1591 * been produced and stored in the @optimum array. These
1592 * matches/literals are linked in reverse order, so the last thing this
1593 * function does is reverse this list and return the first
1594 * match/literal, leaving the rest to be returned immediately by
1595 * subsequent calls to this function.
1601 /* Advance to next position. */
1604 /* Check termination conditions (2) and (3) noted above. */
1605 if (cur_pos == end_pos || cur_pos == LZX_OPTIM_ARRAY_SIZE)
1606 return lzx_match_chooser_reverse_list(ctx, cur_pos);
1608 /* Search for matches at recent offsets. */
1609 longest_rep_len = LZX_MIN_MATCH_LEN - 1;
1610 unsigned limit = min(LZX_MAX_MATCH_LEN,
1611 ctx->match_window_end - ctx->match_window_pos);
1612 for (int i = 0; i < LZX_NUM_RECENT_OFFSETS; i++) {
1613 u32 offset = ctx->optimum[cur_pos].queue.R[i];
1614 const u8 *strptr = &ctx->window[ctx->match_window_pos];
1615 const u8 *matchptr = strptr - offset;
1617 while (len < limit && strptr[len] == matchptr[len])
1619 if (len > longest_rep_len) {
1620 longest_rep_len = len;
1621 longest_rep_offset = offset;
1625 /* If we found a long match at a recent offset, choose it
1627 if (longest_rep_len >= ctx->params.alg_params.slow.nice_match_length) {
1628 /* Build the list of matches to return and get
1630 match = lzx_match_chooser_reverse_list(ctx, cur_pos);
1632 /* Append the long match to the end of the list. */
1633 ctx->optimum[cur_pos].next.match_offset = longest_rep_offset;
1634 ctx->optimum[cur_pos].next.link = cur_pos + longest_rep_len;
1635 ctx->optimum_end_idx = cur_pos + longest_rep_len;
1637 /* Skip over the remaining bytes of the long match. */
1638 lzx_skip_bytes(ctx, longest_rep_len);
1640 /* Return first match in the list. */
1644 /* Search other matches. */
1645 num_matches = lzx_get_matches(ctx, &matches);
1647 /* If there's a long match, take it. */
1649 longest_len = matches[num_matches - 1].len;
1650 if (longest_len >= ctx->params.alg_params.slow.nice_match_length) {
1651 /* Build the list of matches to return and get
1653 match = lzx_match_chooser_reverse_list(ctx, cur_pos);
1655 /* Append the long match to the end of the list. */
1656 ctx->optimum[cur_pos].next.match_offset =
1657 matches[num_matches - 1].offset;
1658 ctx->optimum[cur_pos].next.link = cur_pos + longest_len;
1659 ctx->optimum_end_idx = cur_pos + longest_len;
1661 /* Skip over the remaining bytes of the long match. */
1662 lzx_skip_bytes(ctx, longest_len - 1);
1664 /* Return first match in the list. */
1671 while (end_pos < cur_pos + longest_len)
1672 ctx->optimum[++end_pos].cost = MC_INFINITE_COST;
1674 /* Consider coding a literal. */
1675 cost = ctx->optimum[cur_pos].cost +
1676 lzx_literal_cost(ctx->window[ctx->match_window_pos - 1],
1678 if (cost < ctx->optimum[cur_pos + 1].cost) {
1679 ctx->optimum[cur_pos + 1].queue = ctx->optimum[cur_pos].queue;
1680 ctx->optimum[cur_pos + 1].cost = cost;
1681 ctx->optimum[cur_pos + 1].prev.link = cur_pos;
1684 /* Consider coding a match. */
1686 for (unsigned len = 2; len <= longest_len; len++) {
1688 struct lzx_lru_queue queue;
1690 offset = matchptr->offset;
1691 queue = ctx->optimum[cur_pos].queue;
1693 cost = ctx->optimum[cur_pos].cost +
1694 lzx_match_cost(len, offset, &ctx->costs, &queue);
1695 if (cost < ctx->optimum[cur_pos + len].cost) {
1696 ctx->optimum[cur_pos + len].queue = queue;
1697 ctx->optimum[cur_pos + len].prev.link = cur_pos;
1698 ctx->optimum[cur_pos + len].prev.match_offset = offset;
1699 ctx->optimum[cur_pos + len].cost = cost;
1701 if (len == matchptr->len)
1705 if (longest_rep_len >= LZX_MIN_MATCH_LEN) {
1706 struct lzx_lru_queue queue;
1708 while (end_pos < cur_pos + longest_rep_len)
1709 ctx->optimum[++end_pos].cost = MC_INFINITE_COST;
1711 queue = ctx->optimum[cur_pos].queue;
1713 cost = ctx->optimum[cur_pos].cost +
1714 lzx_match_cost(longest_rep_len, longest_rep_offset,
1715 &ctx->costs, &queue);
1716 if (cost <= ctx->optimum[cur_pos + longest_rep_len].cost) {
1717 ctx->optimum[cur_pos + longest_rep_len].queue =
1719 ctx->optimum[cur_pos + longest_rep_len].prev.link =
1721 ctx->optimum[cur_pos + longest_rep_len].prev.match_offset =
1723 ctx->optimum[cur_pos + longest_rep_len].cost =
1730 /* Set default symbol costs for the LZX Huffman codes. */
1732 lzx_set_default_costs(struct lzx_costs * costs, unsigned num_main_syms)
1736 /* Main code (part 1): Literal symbols */
1737 for (i = 0; i < LZX_NUM_CHARS; i++)
1740 /* Main code (part 2): Match header symbols */
1741 for (; i < num_main_syms; i++)
1742 costs->main[i] = 10;
1745 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
1748 /* Aligned offset code */
1749 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
1750 costs->aligned[i] = 3;
1753 /* Given the frequencies of symbols in an LZX-compressed block and the
1754 * corresponding Huffman codes, return LZX_BLOCKTYPE_ALIGNED or
1755 * LZX_BLOCKTYPE_VERBATIM if an aligned offset or verbatim block, respectively,
1756 * will take fewer bits to output. */
1758 lzx_choose_verbatim_or_aligned(const struct lzx_freqs * freqs,
1759 const struct lzx_codes * codes)
1761 unsigned aligned_cost = 0;
1762 unsigned verbatim_cost = 0;
1764 /* Verbatim blocks have a constant 3 bits per position footer. Aligned
1765 * offset blocks have an aligned offset symbol per position footer, plus
1766 * an extra 24 bits per block to output the lengths necessary to
1767 * reconstruct the aligned offset code itself. */
1768 for (unsigned i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1769 verbatim_cost += 3 * freqs->aligned[i];
1770 aligned_cost += codes->lens.aligned[i] * freqs->aligned[i];
1772 aligned_cost += LZX_ALIGNEDCODE_ELEMENT_SIZE * LZX_ALIGNEDCODE_NUM_SYMBOLS;
1773 if (aligned_cost < verbatim_cost)
1774 return LZX_BLOCKTYPE_ALIGNED;
1776 return LZX_BLOCKTYPE_VERBATIM;
1779 /* Find a near-optimal sequence of matches/literals with which to output the
1780 * specified LZX block, then set the block's type to that which has the minimum
1781 * cost to output (either verbatim or aligned). */
1783 lzx_optimize_block(struct lzx_compressor *ctx, struct lzx_block_spec *spec,
1784 unsigned num_passes)
1786 const struct lzx_lru_queue orig_queue = ctx->queue;
1787 unsigned num_passes_remaining = num_passes;
1788 struct lzx_freqs freqs;
1790 LZX_ASSERT(num_passes >= 1);
1791 LZX_ASSERT(lz_bt_get_position(&ctx->mf) == spec->window_pos);
1793 ctx->match_window_end = spec->window_pos + spec->block_size;
1794 spec->chosen_matches = &ctx->chosen_matches[spec->window_pos];
1795 ctx->matches_cached = false;
1797 /* The first optimal parsing pass is done using the cost model already
1798 * set in ctx->costs. Each later pass is done using a cost model
1799 * computed from the previous pass. */
1801 const u8 *window_ptr;
1802 const u8 *window_end;
1803 struct lzx_match *next_chosen_match;
1805 --num_passes_remaining;
1806 ctx->match_window_pos = spec->window_pos;
1807 ctx->cache_ptr = ctx->cached_matches;
1808 memset(&freqs, 0, sizeof(freqs));
1809 window_ptr = &ctx->window[spec->window_pos];
1810 window_end = window_ptr + spec->block_size;
1811 next_chosen_match = spec->chosen_matches;
1813 while (window_ptr != window_end) {
1814 struct raw_match raw_match;
1815 struct lzx_match lzx_match;
1817 raw_match = lzx_get_near_optimal_match(ctx);
1819 LZX_ASSERT(!(raw_match.len == LZX_MIN_MATCH_LEN &&
1820 raw_match.offset == ctx->max_window_size -
1821 LZX_MIN_MATCH_LEN));
1822 if (raw_match.len >= LZX_MIN_MATCH_LEN) {
1823 lzx_match.data = lzx_tally_match(raw_match.len,
1827 window_ptr += raw_match.len;
1829 lzx_match.data = lzx_tally_literal(*window_ptr,
1833 *next_chosen_match++ = lzx_match;
1835 spec->num_chosen_matches = next_chosen_match - spec->chosen_matches;
1836 lzx_make_huffman_codes(&freqs, &spec->codes, ctx->num_main_syms);
1837 if (num_passes_remaining) {
1838 lzx_set_costs(ctx, &spec->codes.lens);
1839 ctx->queue = orig_queue;
1840 ctx->matches_cached = true;
1842 } while (num_passes_remaining);
1844 spec->block_type = lzx_choose_verbatim_or_aligned(&freqs, &spec->codes);
1847 /* Prepare the input window into one or more LZX blocks ready to be output. */
1849 lzx_prepare_blocks(struct lzx_compressor * ctx)
1851 /* Set up a default cost model. */
1852 lzx_set_default_costs(&ctx->costs, ctx->num_main_syms);
1854 /* Set up the block specifications.
1855 * TODO: The compression ratio could be slightly improved by performing
1856 * data-dependent block splitting instead of using fixed-size blocks.
1857 * Doing so well is a computationally hard problem, however. */
1858 ctx->num_blocks = DIV_ROUND_UP(ctx->window_size, LZX_DIV_BLOCK_SIZE);
1859 for (unsigned i = 0; i < ctx->num_blocks; i++) {
1860 unsigned pos = LZX_DIV_BLOCK_SIZE * i;
1861 ctx->block_specs[i].window_pos = pos;
1862 ctx->block_specs[i].block_size = min(ctx->window_size - pos,
1863 LZX_DIV_BLOCK_SIZE);
1866 /* Load the window into the match-finder. */
1867 lz_bt_load_window(&ctx->mf, ctx->window, ctx->window_size);
1869 /* Determine sequence of matches/literals to output for each block. */
1870 lzx_lru_queue_init(&ctx->queue);
1871 ctx->optimum_cur_idx = 0;
1872 ctx->optimum_end_idx = 0;
1873 for (unsigned i = 0; i < ctx->num_blocks; i++) {
1874 lzx_optimize_block(ctx, &ctx->block_specs[i],
1875 ctx->params.alg_params.slow.num_optim_passes);
1880 * This is the fast version of lzx_prepare_blocks(). This version "quickly"
1881 * prepares a single compressed block containing the entire input. See the
1882 * description of the "Fast algorithm" at the beginning of this file for more
1885 * Input --- the preprocessed data:
1890 * Output --- the block specification and the corresponding match/literal data:
1892 * ctx->block_specs[]
1894 * ctx->chosen_matches[]
1897 lzx_prepare_block_fast(struct lzx_compressor * ctx)
1899 struct lzx_record_ctx record_ctx;
1900 struct lzx_block_spec *spec;
1902 /* Parameters to hash chain LZ match finder
1903 * (lazy with 1 match lookahead) */
1904 static const struct lz_params lzx_lz_params = {
1905 /* Although LZX_MIN_MATCH_LEN == 2, length 2 matches typically
1906 * aren't worth choosing when using greedy or lazy parsing. */
1908 .max_match = LZX_MAX_MATCH_LEN,
1909 .max_offset = LZX_MAX_WINDOW_SIZE,
1910 .good_match = LZX_MAX_MATCH_LEN,
1911 .nice_match = LZX_MAX_MATCH_LEN,
1912 .max_chain_len = LZX_MAX_MATCH_LEN,
1913 .max_lazy_match = LZX_MAX_MATCH_LEN,
1917 /* Initialize symbol frequencies and match offset LRU queue. */
1918 memset(&record_ctx.freqs, 0, sizeof(struct lzx_freqs));
1919 lzx_lru_queue_init(&record_ctx.queue);
1920 record_ctx.matches = ctx->chosen_matches;
1922 /* Determine series of matches/literals to output. */
1923 lz_analyze_block(ctx->window,
1931 /* Set up block specification. */
1932 spec = &ctx->block_specs[0];
1933 spec->block_type = LZX_BLOCKTYPE_ALIGNED;
1934 spec->window_pos = 0;
1935 spec->block_size = ctx->window_size;
1936 spec->num_chosen_matches = (record_ctx.matches - ctx->chosen_matches);
1937 spec->chosen_matches = ctx->chosen_matches;
1938 lzx_make_huffman_codes(&record_ctx.freqs, &spec->codes,
1939 ctx->num_main_syms);
1940 ctx->num_blocks = 1;
1944 lzx_compress(const void *uncompressed_data, size_t uncompressed_size,
1945 void *compressed_data, size_t compressed_size_avail, void *_ctx)
1947 struct lzx_compressor *ctx = _ctx;
1948 struct output_bitstream ostream;
1949 size_t compressed_size;
1951 if (uncompressed_size < 100) {
1952 LZX_DEBUG("Too small to bother compressing.");
1956 if (uncompressed_size > ctx->max_window_size) {
1957 LZX_DEBUG("Can't compress %zu bytes using window of %u bytes!",
1958 uncompressed_size, ctx->max_window_size);
1962 LZX_DEBUG("Attempting to compress %zu bytes...",
1965 /* The input data must be preprocessed. To avoid changing the original
1966 * input, copy it to a temporary buffer. */
1967 memcpy(ctx->window, uncompressed_data, uncompressed_size);
1968 ctx->window_size = uncompressed_size;
1970 /* This line is unnecessary; it just avoids inconsequential accesses of
1971 * uninitialized memory that would show up in memory-checking tools such
1973 memset(&ctx->window[ctx->window_size], 0, 12);
1975 LZX_DEBUG("Preprocessing data...");
1977 /* Before doing any actual compression, do the call instruction (0xe8
1978 * byte) translation on the uncompressed data. */
1979 lzx_do_e8_preprocessing(ctx->window, ctx->window_size);
1981 LZX_DEBUG("Preparing blocks...");
1983 /* Prepare the compressed data. */
1984 if (ctx->params.algorithm == WIMLIB_LZX_ALGORITHM_FAST)
1985 lzx_prepare_block_fast(ctx);
1987 lzx_prepare_blocks(ctx);
1989 LZX_DEBUG("Writing compressed blocks...");
1991 /* Generate the compressed data. */
1992 init_output_bitstream(&ostream, compressed_data, compressed_size_avail);
1993 lzx_write_all_blocks(ctx, &ostream);
1995 LZX_DEBUG("Flushing bitstream...");
1996 compressed_size = flush_output_bitstream(&ostream);
1997 if (compressed_size == ~(input_idx_t)0) {
1998 LZX_DEBUG("Data did not compress to %zu bytes or less!",
1999 compressed_size_avail);
2003 LZX_DEBUG("Done: compressed %zu => %zu bytes.",
2004 uncompressed_size, compressed_size);
2006 /* Verify that we really get the same thing back when decompressing.
2007 * Although this could be disabled by default in all cases, it only
2008 * takes around 2-3% of the running time of the slow algorithm to do the
2010 if (ctx->params.algorithm == WIMLIB_LZX_ALGORITHM_SLOW
2011 #if defined(ENABLE_LZX_DEBUG) || defined(ENABLE_VERIFY_COMPRESSION)
2016 struct wimlib_decompressor *decompressor;
2018 if (0 == wimlib_create_decompressor(WIMLIB_COMPRESSION_TYPE_LZX,
2019 ctx->max_window_size,
2024 ret = wimlib_decompress(compressed_data,
2029 wimlib_free_decompressor(decompressor);
2032 ERROR("Failed to decompress data we "
2033 "compressed using LZX algorithm");
2037 if (memcmp(uncompressed_data, ctx->window, uncompressed_size)) {
2038 ERROR("Data we compressed using LZX algorithm "
2039 "didn't decompress to original");
2044 WARNING("Failed to create decompressor for "
2045 "data verification!");
2048 return compressed_size;
2052 lzx_free_compressor(void *_ctx)
2054 struct lzx_compressor *ctx = _ctx;
2057 FREE(ctx->chosen_matches);
2058 FREE(ctx->cached_matches);
2060 lz_bt_destroy(&ctx->mf);
2061 FREE(ctx->block_specs);
2062 FREE(ctx->prev_tab);
2068 static const struct wimlib_lzx_compressor_params lzx_fast_default = {
2070 .size = sizeof(struct wimlib_lzx_compressor_params),
2072 .algorithm = WIMLIB_LZX_ALGORITHM_FAST,
2079 static const struct wimlib_lzx_compressor_params lzx_slow_default = {
2081 .size = sizeof(struct wimlib_lzx_compressor_params),
2083 .algorithm = WIMLIB_LZX_ALGORITHM_SLOW,
2087 .use_len2_matches = 1,
2088 .nice_match_length = 32,
2089 .num_optim_passes = 2,
2090 .max_search_depth = 50,
2091 .main_nostat_cost = 15,
2092 .len_nostat_cost = 15,
2093 .aligned_nostat_cost = 7,
2098 static const struct wimlib_lzx_compressor_params *
2099 lzx_get_params(const struct wimlib_compressor_params_header *_params)
2101 const struct wimlib_lzx_compressor_params *params =
2102 (const struct wimlib_lzx_compressor_params*)_params;
2104 if (params == NULL) {
2105 LZX_DEBUG("Using default algorithm and parameters.");
2106 params = &lzx_slow_default;
2108 if (params->use_defaults) {
2109 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW)
2110 params = &lzx_slow_default;
2112 params = &lzx_fast_default;
2119 lzx_create_compressor(size_t window_size,
2120 const struct wimlib_compressor_params_header *_params,
2123 const struct wimlib_lzx_compressor_params *params = lzx_get_params(_params);
2124 struct lzx_compressor *ctx;
2126 LZX_DEBUG("Allocating LZX context...");
2128 if (!lzx_window_size_valid(window_size))
2129 return WIMLIB_ERR_INVALID_PARAM;
2131 LZX_DEBUG("Allocating memory.");
2133 ctx = CALLOC(1, sizeof(struct lzx_compressor));
2137 ctx->num_main_syms = lzx_get_num_main_syms(window_size);
2138 ctx->max_window_size = window_size;
2139 ctx->window = MALLOC(window_size + 12);
2140 if (ctx->window == NULL)
2143 if (params->algorithm == WIMLIB_LZX_ALGORITHM_FAST) {
2144 ctx->prev_tab = MALLOC(window_size * sizeof(ctx->prev_tab[0]));
2145 if (ctx->prev_tab == NULL)
2149 size_t block_specs_length = DIV_ROUND_UP(window_size, LZX_DIV_BLOCK_SIZE);
2150 ctx->block_specs = MALLOC(block_specs_length * sizeof(ctx->block_specs[0]));
2151 if (ctx->block_specs == NULL)
2154 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
2155 unsigned min_match_len = LZX_MIN_MATCH_LEN;
2156 if (!params->alg_params.slow.use_len2_matches)
2157 min_match_len = max(min_match_len, 3);
2159 if (!lz_bt_init(&ctx->mf,
2163 params->alg_params.slow.nice_match_length,
2164 params->alg_params.slow.max_search_depth))
2168 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
2169 ctx->optimum = MALLOC((LZX_OPTIM_ARRAY_SIZE +
2170 min(params->alg_params.slow.nice_match_length,
2171 LZX_MAX_MATCH_LEN)) *
2172 sizeof(ctx->optimum[0]));
2177 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
2178 ctx->cached_matches = MALLOC(LZX_CACHE_SIZE);
2179 if (ctx->cached_matches == NULL)
2181 ctx->cache_limit = ctx->cached_matches +
2182 LZX_CACHE_LEN - (LZX_MAX_MATCHES_PER_POS + 1);
2185 ctx->chosen_matches = MALLOC(window_size * sizeof(ctx->chosen_matches[0]));
2186 if (ctx->chosen_matches == NULL)
2189 memcpy(&ctx->params, params, sizeof(struct wimlib_lzx_compressor_params));
2190 memset(&ctx->zero_codes, 0, sizeof(ctx->zero_codes));
2192 LZX_DEBUG("Successfully allocated new LZX context.");
2198 lzx_free_compressor(ctx);
2199 return WIMLIB_ERR_NOMEM;
2203 lzx_get_needed_memory(size_t max_block_size,
2204 const struct wimlib_compressor_params_header *_params)
2206 const struct wimlib_lzx_compressor_params *params = lzx_get_params(_params);
2210 size += sizeof(struct lzx_compressor);
2212 size += max_block_size + 12;
2214 size += DIV_ROUND_UP(max_block_size, LZX_DIV_BLOCK_SIZE) *
2215 sizeof(((struct lzx_compressor*)0)->block_specs[0]);
2217 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
2218 size += max_block_size * sizeof(((struct lzx_compressor*)0)->chosen_matches[0]);
2219 size += lz_bt_get_needed_memory(max_block_size);
2220 size += (LZX_OPTIM_ARRAY_SIZE +
2221 min(params->alg_params.slow.nice_match_length,
2222 LZX_MAX_MATCH_LEN)) *
2223 sizeof(((struct lzx_compressor *)0)->optimum[0]);
2224 size += LZX_CACHE_SIZE;
2226 size += max_block_size * sizeof(((struct lzx_compressor*)0)->prev_tab[0]);
2232 lzx_params_valid(const struct wimlib_compressor_params_header *_params)
2234 const struct wimlib_lzx_compressor_params *params =
2235 (const struct wimlib_lzx_compressor_params*)_params;
2237 if (params->hdr.size != sizeof(struct wimlib_lzx_compressor_params)) {
2238 LZX_DEBUG("Invalid parameter structure size!");
2242 if (params->algorithm != WIMLIB_LZX_ALGORITHM_SLOW &&
2243 params->algorithm != WIMLIB_LZX_ALGORITHM_FAST)
2245 LZX_DEBUG("Invalid algorithm.");
2249 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW &&
2250 !params->use_defaults)
2252 if (params->alg_params.slow.num_optim_passes < 1)
2254 LZX_DEBUG("Invalid number of optimization passes!");
2258 if (params->alg_params.slow.main_nostat_cost < 1 ||
2259 params->alg_params.slow.main_nostat_cost > 16)
2261 LZX_DEBUG("Invalid main_nostat_cost!");
2265 if (params->alg_params.slow.len_nostat_cost < 1 ||
2266 params->alg_params.slow.len_nostat_cost > 16)
2268 LZX_DEBUG("Invalid len_nostat_cost!");
2272 if (params->alg_params.slow.aligned_nostat_cost < 1 ||
2273 params->alg_params.slow.aligned_nostat_cost > 8)
2275 LZX_DEBUG("Invalid aligned_nostat_cost!");
2282 const struct compressor_ops lzx_compressor_ops = {
2283 .params_valid = lzx_params_valid,
2284 .get_needed_memory = lzx_get_needed_memory,
2285 .create_compressor = lzx_create_compressor,
2286 .compress = lzx_compress,
2287 .free_compressor = lzx_free_compressor,