4 * A compressor that produces output compatible with the LZX compression format.
8 * Copyright (C) 2012, 2013, 2014 Eric Biggers
10 * This file is part of wimlib, a library for working with WIM files.
12 * wimlib is free software; you can redistribute it and/or modify it under the
13 * terms of the GNU General Public License as published by the Free
14 * Software Foundation; either version 3 of the License, or (at your option)
17 * wimlib is distributed in the hope that it will be useful, but WITHOUT ANY
18 * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
19 * A PARTICULAR PURPOSE. See the GNU General Public License for more
22 * You should have received a copy of the GNU General Public License
23 * along with wimlib; if not, see http://www.gnu.org/licenses/.
28 * This file contains a compressor for the LZX ("Lempel-Ziv eXtended"?)
29 * compression format, as used in the WIM (Windows IMaging) file format. This
30 * code may need some slight modifications to be used outside of the WIM format.
31 * In particular, in other situations the LZX block header might be slightly
32 * different, and a sliding window rather than a fixed-size window might be
35 * ----------------------------------------------------------------------------
39 * The primary reference for LZX is the specification released by Microsoft.
40 * However, the comments in lzx-decompress.c provide more information about LZX
41 * and note some errors in the Microsoft specification.
43 * LZX shares many similarities with DEFLATE, the format used by zlib and gzip.
44 * Both LZX and DEFLATE use LZ77 matching and Huffman coding. Certain details
45 * are quite similar, such as the method for storing Huffman codes. However,
46 * the main differences are:
48 * - LZX preprocesses the data to attempt to make x86 machine code slightly more
49 * compressible before attempting to compress it further.
51 * - LZX uses a "main" alphabet which combines literals and matches, with the
52 * match symbols containing a "length header" (giving all or part of the match
53 * length) and a "position slot" (giving, roughly speaking, the order of
54 * magnitude of the match offset).
56 * - LZX does not have static Huffman blocks (that is, the kind with preset
57 * Huffman codes); however it does have two types of dynamic Huffman blocks
58 * ("verbatim" and "aligned").
60 * - LZX has a minimum match length of 2 rather than 3.
62 * - In LZX, match offsets 0 through 2 actually represent entries in an LRU
63 * queue of match offsets. This is very useful for certain types of files,
64 * such as binary files that have repeating records.
66 * ----------------------------------------------------------------------------
68 * Algorithmic Overview
70 * At a high level, any implementation of LZX compression must operate as
73 * 1. Preprocess the input data to translate the targets of 32-bit x86 call
74 * instructions to absolute offsets. (Actually, this is required for WIM,
75 * but might not be in other places LZX is used.)
77 * 2. Find a sequence of LZ77-style matches and literal bytes that expands to
78 * the preprocessed data.
80 * 3. Divide the match/literal sequence into one or more LZX blocks, each of
81 * which may be "uncompressed", "verbatim", or "aligned".
83 * 4. Output each LZX block.
85 * Step (1) is fairly straightforward. It requires looking for 0xe8 bytes in
86 * the input data and performing a translation on the 4 bytes following each
89 * Step (4) is complicated, but it is mostly determined by the LZX format. The
90 * only real choice we have is what algorithm to use to build the length-limited
91 * canonical Huffman codes. See lzx_write_all_blocks() for details.
93 * That leaves steps (2) and (3) as where all the hard stuff happens. Focusing
94 * on step (2), we need to do LZ77-style parsing on the input data, or "window",
95 * to divide it into a sequence of matches and literals. Each position in the
96 * window might have multiple matches associated with it, and we need to choose
97 * which one, if any, to actually use. Therefore, the problem can really be
98 * divided into two areas of concern: (a) finding matches at a given position,
99 * which we shall call "match-finding", and (b) choosing whether to use a
100 * match or a literal at a given position, and if using a match, which one (if
101 * there is more than one available). We shall call this "match-choosing". We
102 * first consider match-finding, then match-choosing.
104 * ----------------------------------------------------------------------------
108 * Given a position in the window, we want to find LZ77-style "matches" with
109 * that position at previous positions in the window. With LZX, the minimum
110 * match length is 2 and the maximum match length is 257. The only restriction
111 * on offsets is that LZX does not allow the last 2 bytes of the window to match
112 * the beginning of the window.
114 * There are a number of algorithms that can be used for this, including hash
115 * chains, binary trees, and suffix arrays. Binary trees generally work well
116 * for LZX compression since it uses medium-size windows (2^15 to 2^21 bytes).
117 * However, when compressing in a fast mode where many positions are skipped
118 * (not searched for matches), hash chains are faster.
120 * Since the match-finders are not specific to LZX, I will not explain them in
121 * detail here. Instead, see lz_hash_chains.c and lz_binary_trees.c.
123 * ----------------------------------------------------------------------------
127 * Usually, choosing the longest match is best because it encodes the most data
128 * in that one item. However, sometimes the longest match is not optimal
129 * because (a) choosing a long match now might prevent using an even longer
130 * match later, or (b) more generally, what we actually care about is the number
131 * of bits it will ultimately take to output each match or literal, which is
132 * actually dependent on the entropy encoding using by the underlying
133 * compression format. Consequently, a longer match usually, but not always,
134 * takes fewer bits to encode than multiple shorter matches or literals that
135 * cover the same data.
137 * This problem of choosing the truly best match/literal sequence is probably
138 * impossible to solve efficiently when combined with entropy encoding. If we
139 * knew how many bits it takes to output each match/literal, then we could
140 * choose the optimal sequence using shortest-path search a la Dijkstra's
141 * algorithm. However, with entropy encoding, the chosen match/literal sequence
142 * affects its own encoding. Therefore, we can't know how many bits it will
143 * take to actually output any one match or literal until we have actually
144 * chosen the full sequence of matches and literals.
146 * Notwithstanding the entropy encoding problem, we also aren't guaranteed to
147 * choose the optimal match/literal sequence unless the match-finder (see
148 * section "Match-finder") provides the match-chooser with all possible matches
149 * at each position. However, this is not computationally efficient. For
150 * example, there might be many matches of the same length, and usually (but not
151 * always) the best choice is the one with the smallest offset. So in practice,
152 * it's fine to only consider the smallest offset for a given match length at a
153 * given position. (Actually, for LZX, it's also worth considering repeat
156 * In addition, as mentioned earlier, in LZX we have the choice of using
157 * multiple blocks, each of which resets the Huffman codes. This expands the
158 * search space even further. Therefore, to simplify the problem, we currently
159 * we don't attempt to actually choose the LZX blocks based on the data.
160 * Instead, we just divide the data into fixed-size blocks of LZX_DIV_BLOCK_SIZE
161 * bytes each, and always use verbatim or aligned blocks (never uncompressed).
162 * A previous version of this code recursively split the input data into
163 * equal-sized blocks, up to a maximum depth, and chose the lowest-cost block
164 * divisions. However, this made compression much slower and did not actually
165 * help very much. It remains an open question whether a sufficiently fast and
166 * useful block-splitting algorithm is possible for LZX. Essentially the same
167 * problem also applies to DEFLATE. The Microsoft LZX compressor seemingly does
168 * do block splitting, although I don't know how fast or useful it is,
171 * Now, back to the entropy encoding problem. The "solution" is to use an
172 * iterative approach to compute a good, but not necessarily optimal,
173 * match/literal sequence. Start with a fixed assignment of symbol costs and
174 * choose an "optimal" match/literal sequence based on those costs, using
175 * shortest-path seach a la Dijkstra's algorithm. Then, for each iteration of
176 * the optimization, update the costs based on the entropy encoding of the
177 * current match/literal sequence, then choose a new match/literal sequence
178 * based on the updated costs. Usually, the actual cost to output the current
179 * match/literal sequence will decrease in each iteration until it converges on
180 * a fixed point. This result may not be the truly optimal match/literal
181 * sequence, but it usually is much better than one chosen by doing a "greedy"
182 * parse where we always chooe the longest match.
184 * An alternative to both greedy parsing and iterative, near-optimal parsing is
185 * "lazy" parsing. Briefly, "lazy" parsing considers just the longest match at
186 * each position, but it waits to choose that match until it has also examined
187 * the next position. This is actually a useful approach; it's used by zlib,
188 * for example. Therefore, for fast compression we combine lazy parsing with
189 * the hash chain max-finder. For normal/high compression we combine
190 * near-optimal parsing with the binary tree match-finder.
197 #include "wimlib/compressor_ops.h"
198 #include "wimlib/compress_common.h"
199 #include "wimlib/endianness.h"
200 #include "wimlib/error.h"
201 #include "wimlib/lz_mf.h"
202 #include "wimlib/lzx.h"
203 #include "wimlib/util.h"
206 #define LZX_OPTIM_ARRAY_LENGTH 4096
208 #define LZX_DIV_BLOCK_SIZE 32768
210 #define LZX_CACHE_PER_POS 8
212 #define LZX_MAX_MATCHES_PER_POS (LZX_MAX_MATCH_LEN - LZX_MIN_MATCH_LEN + 1)
214 #define LZX_CACHE_LEN (LZX_DIV_BLOCK_SIZE * (LZX_CACHE_PER_POS + 1))
216 /* Codewords for the LZX main, length, and aligned offset Huffman codes */
217 struct lzx_codewords {
218 u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
219 u32 len[LZX_LENCODE_NUM_SYMBOLS];
220 u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
223 /* Codeword lengths (in bits) for the LZX main, length, and aligned offset
226 * A 0 length means the codeword has zero frequency.
229 u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
230 u8 len[LZX_LENCODE_NUM_SYMBOLS];
231 u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
234 /* Costs for the LZX main, length, and aligned offset Huffman symbols.
236 * If a codeword has zero frequency, it must still be assigned some nonzero cost
237 * --- generally a high cost, since even if it gets used in the next iteration,
238 * it probably will not be used very many times. */
240 u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
241 u8 len[LZX_LENCODE_NUM_SYMBOLS];
242 u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
245 /* The LZX main, length, and aligned offset Huffman codes */
247 struct lzx_codewords codewords;
248 struct lzx_lens lens;
251 /* Tables for tallying symbol frequencies in the three LZX alphabets */
253 u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
254 u32 len[LZX_LENCODE_NUM_SYMBOLS];
255 u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
258 /* LZX intermediate match/literal format */
262 * 31 1 if a match, 0 if a literal.
264 * 30-25 position slot. This can be at most 50, so it will fit in 6
267 * 8-24 position footer. This is the offset of the real formatted
268 * offset from the position base. This can be at most 17 bits
269 * (since lzx_extra_bits[LZX_MAX_POSITION_SLOTS - 1] is 17).
271 * 0-7 length of match, minus 2. This can be at most
272 * (LZX_MAX_MATCH_LEN - 2) == 255, so it will fit in 8 bits. */
276 /* Specification for an LZX block. */
277 struct lzx_block_spec {
279 /* One of the LZX_BLOCKTYPE_* constants indicating which type of this
283 /* 0-based position in the window at which this block starts. */
286 /* The number of bytes of uncompressed data this block represents. */
289 /* The match/literal sequence for this block. */
290 struct lzx_item *chosen_items;
292 /* The length of the @chosen_items sequence. */
293 u32 num_chosen_items;
295 /* Huffman codes for this block. */
296 struct lzx_codes codes;
299 struct lzx_compressor;
301 struct lzx_compressor_params {
302 struct lz_match (*choose_item_func)(struct lzx_compressor *);
303 enum lz_mf_algo mf_algo;
304 u32 num_optim_passes;
305 u32 min_match_length;
306 u32 nice_match_length;
307 u32 max_search_depth;
310 /* State of the LZX compressor. */
311 struct lzx_compressor {
313 /* The buffer of data to be compressed.
315 * 0xe8 byte preprocessing is done directly on the data here before
316 * further compression.
318 * Note that this compressor does *not* use a real sliding window!!!!
319 * It's not needed in the WIM format, since every chunk is compressed
320 * independently. This is by design, to allow random access to the
324 /* Number of bytes of data to be compressed, which is the number of
325 * bytes of data in @cur_window that are actually valid. */
328 /* Allocated size of @cur_window. */
331 /* Compression parameters. */
332 struct lzx_compressor_params params;
334 unsigned (*get_matches_func)(struct lzx_compressor *, const struct lz_match **);
335 void (*skip_bytes_func)(struct lzx_compressor *, unsigned n);
337 /* Number of symbols in the main alphabet (depends on the
338 * @max_window_size since it determines the maximum allowed offset). */
339 unsigned num_main_syms;
341 /* The current match offset LRU queue. */
342 struct lzx_lru_queue queue;
344 /* Space for the sequences of matches/literals that were chosen for each
346 struct lzx_item *chosen_items;
348 /* Information about the LZX blocks the preprocessed input was divided
350 struct lzx_block_spec *block_specs;
352 /* Number of LZX blocks the input was divided into; a.k.a. the number of
353 * elements of @block_specs that are valid. */
356 /* This is simply filled in with zeroes and used to avoid special-casing
357 * the output of the first compressed Huffman code, which conceptually
358 * has a delta taken from a code with all symbols having zero-length
360 struct lzx_codes zero_codes;
362 /* The current cost model. */
363 struct lzx_costs costs;
365 /* Lempel-Ziv match-finder. */
368 /* Position in window of next match to return. */
369 u32 match_window_pos;
371 /* The end-of-block position. We can't allow any matches to span this
373 u32 match_window_end;
375 /* When doing more than one match-choosing pass over the data, matches
376 * found by the match-finder are cached in the following array to
377 * achieve a slight speedup when the same matches are needed on
378 * subsequent passes. This is suboptimal because different matches may
379 * be preferred with different cost models, but seems to be a worthwhile
381 struct lz_match *cached_matches;
382 struct lz_match *cache_ptr;
383 struct lz_match *cache_limit;
385 /* Match-chooser state, used when doing near-optimal parsing.
387 * When matches have been chosen, optimum_cur_idx is set to the position
388 * in the window of the next match/literal to return and optimum_end_idx
389 * is set to the position in the window at the end of the last
390 * match/literal to return. */
391 struct lzx_mc_pos_data *optimum;
392 unsigned optimum_cur_idx;
393 unsigned optimum_end_idx;
395 /* Previous match, used when doing lazy parsing. */
396 struct lz_match prev_match;
400 * Match chooser position data:
402 * An array of these structures is used during the match-choosing algorithm.
403 * They correspond to consecutive positions in the window and are used to keep
404 * track of the cost to reach each position, and the match/literal choices that
405 * need to be chosen to reach that position.
407 struct lzx_mc_pos_data {
408 /* The approximate minimum cost, in bits, to reach this position in the
409 * window which has been found so far. */
411 #define MC_INFINITE_COST ((u32)~0UL)
413 /* The union here is just for clarity, since the fields are used in two
414 * slightly different ways. Initially, the @prev structure is filled in
415 * first, and links go from later in the window to earlier in the
416 * window. Later, @next structure is filled in and links go from
417 * earlier in the window to later in the window. */
420 /* Position of the start of the match or literal that
421 * was taken to get to this position in the approximate
422 * minimum-cost parse. */
425 /* Offset (as in an LZ (length, offset) pair) of the
426 * match or literal that was taken to get to this
427 * position in the approximate minimum-cost parse. */
431 /* Position at which the match or literal starting at
432 * this position ends in the minimum-cost parse. */
435 /* Offset (as in an LZ (length, offset) pair) of the
436 * match or literal starting at this position in the
437 * approximate minimum-cost parse. */
442 /* Adaptive state that exists after an approximate minimum-cost path to
443 * reach this position is taken.
445 * Note: we update this whenever we update the pending minimum-cost
446 * path. This is in contrast to LZMA, which also has an optimal parser
447 * that maintains a repeat offset queue per position, but will only
448 * compute the queue once that position is actually reached in the
449 * parse, meaning that matches are being considered *starting* at that
450 * position. However, the two methods seem to have approximately the
451 * same performance if appropriate optimizations are used. Intuitively
452 * the LZMA method seems faster, but it actually suffers from 1-2 extra
453 * hard-to-predict branches at each position. Probably it works better
454 * for LZMA than LZX because LZMA has a larger adaptive state than LZX,
455 * and the LZMA encoder considers more possibilities. */
456 struct lzx_lru_queue queue;
461 * Structure to keep track of the current state of sending bits to the
462 * compressed output buffer.
464 * The LZX bitstream is encoded as a sequence of 16-bit coding units.
466 struct lzx_output_bitstream {
468 /* Bits that haven't yet been written to the output buffer. */
471 /* Number of bits currently held in @bitbuf. */
474 /* Pointer to the start of the output buffer. */
477 /* Pointer to the position in the output buffer at which the next coding
478 * unit should be written. */
481 /* Pointer past the end of the output buffer. */
486 * Initialize the output bitstream.
489 * The output bitstream structure to initialize.
491 * The buffer being written to.
493 * Size of @buffer, in bytes.
496 lzx_init_output(struct lzx_output_bitstream *os, void *buffer, u32 size)
501 os->next = os->start;
502 os->end = os->start + size / sizeof(le16);
506 * Write some bits to the output bitstream.
508 * The bits are given by the low-order @num_bits bits of @bits. Higher-order
509 * bits in @bits cannot be set. At most 17 bits can be written at once.
511 * @max_bits is a compile-time constant that specifies the maximum number of
512 * bits that can ever be written at the call site. Currently, it is used to
513 * optimize away the conditional code for writing a second 16-bit coding unit
514 * when writing fewer than 17 bits.
516 * If the output buffer space is exhausted, then the bits will be ignored, and
517 * lzx_flush_output() will return 0 when it gets called.
519 static _always_inline_attribute void
520 lzx_write_varbits(struct lzx_output_bitstream *os,
521 const u32 bits, const unsigned int num_bits,
522 const unsigned int max_num_bits)
524 /* This code is optimized for LZX, which never needs to write more than
525 * 17 bits at once. */
526 LZX_ASSERT(num_bits <= 17);
527 LZX_ASSERT(num_bits <= max_num_bits);
528 LZX_ASSERT(os->bitcount <= 15);
530 /* Add the bits to the bit buffer variable. @bitcount will be at most
531 * 15, so there will be just enough space for the maximum possible
532 * @num_bits of 17. */
533 os->bitcount += num_bits;
534 os->bitbuf = (os->bitbuf << num_bits) | bits;
536 /* Check whether any coding units need to be written. */
537 if (os->bitcount >= 16) {
541 /* Write a coding unit, unless it would overflow the buffer. */
542 if (os->next != os->end)
543 *os->next++ = cpu_to_le16(os->bitbuf >> os->bitcount);
545 /* If writing 17 bits, a second coding unit might need to be
546 * written. But because 'max_num_bits' is a compile-time
547 * constant, the compiler will optimize away this code at most
549 if (max_num_bits == 17 && os->bitcount == 16) {
550 if (os->next != os->end)
551 *os->next++ = cpu_to_le16(os->bitbuf);
557 /* Use when @num_bits is a compile-time constant. Otherwise use
558 * lzx_write_varbits(). */
559 static _always_inline_attribute void
560 lzx_write_bits(struct lzx_output_bitstream *os,
561 const u32 bits, const unsigned int num_bits)
563 lzx_write_varbits(os, bits, num_bits, num_bits);
567 * Flush the last coding unit to the output buffer if needed. Return the total
568 * number of bytes written to the output buffer, or 0 if an overflow occurred.
571 lzx_flush_output(struct lzx_output_bitstream *os)
573 if (os->next == os->end)
576 if (os->bitcount != 0)
577 *os->next++ = cpu_to_le16(os->bitbuf << (16 - os->bitcount));
579 return (const u8 *)os->next - (const u8 *)os->start;
582 /* Returns the LZX position slot that corresponds to a given match offset,
583 * taking into account the recent offset queue and updating it if the offset is
586 lzx_get_position_slot(u32 offset, struct lzx_lru_queue *queue)
588 unsigned position_slot;
590 /* See if the offset was recently used. */
591 for (int i = 0; i < LZX_NUM_RECENT_OFFSETS; i++) {
592 if (offset == queue->R[i]) {
595 /* Bring the repeat offset to the front of the
596 * queue. Note: this is, in fact, not a real
597 * LRU queue because repeat matches are simply
598 * swapped to the front. */
599 swap(queue->R[0], queue->R[i]);
601 /* The resulting position slot is simply the first index
602 * at which the offset was found in the queue. */
607 /* The offset was not recently used; look up its real position slot. */
608 position_slot = lzx_get_position_slot_raw(offset + LZX_OFFSET_OFFSET);
610 /* Bring the new offset to the front of the queue. */
611 for (int i = LZX_NUM_RECENT_OFFSETS - 1; i > 0; i--)
612 queue->R[i] = queue->R[i - 1];
613 queue->R[0] = offset;
615 return position_slot;
618 /* Build the main, length, and aligned offset Huffman codes used in LZX.
620 * This takes as input the frequency tables for each code and produces as output
621 * a set of tables that map symbols to codewords and codeword lengths. */
623 lzx_make_huffman_codes(const struct lzx_freqs *freqs,
624 struct lzx_codes *codes,
625 unsigned num_main_syms)
627 make_canonical_huffman_code(num_main_syms,
628 LZX_MAX_MAIN_CODEWORD_LEN,
631 codes->codewords.main);
633 make_canonical_huffman_code(LZX_LENCODE_NUM_SYMBOLS,
634 LZX_MAX_LEN_CODEWORD_LEN,
637 codes->codewords.len);
639 make_canonical_huffman_code(LZX_ALIGNEDCODE_NUM_SYMBOLS,
640 LZX_MAX_ALIGNED_CODEWORD_LEN,
643 codes->codewords.aligned);
647 * Output a precomputed LZX match.
650 * The bitstream to which to write the match.
652 * The type of the LZX block (LZX_BLOCKTYPE_ALIGNED or
653 * LZX_BLOCKTYPE_VERBATIM)
657 * Pointer to a structure that contains the codewords for the main, length,
658 * and aligned offset Huffman codes for the current LZX compressed block.
661 lzx_write_match(struct lzx_output_bitstream *os, int block_type,
662 struct lzx_item match, const struct lzx_codes *codes)
664 unsigned match_len_minus_2 = match.data & 0xff;
665 u32 position_footer = (match.data >> 8) & 0x1ffff;
666 unsigned position_slot = (match.data >> 25) & 0x3f;
669 unsigned main_symbol;
670 unsigned num_extra_bits;
672 /* If the match length is less than MIN_MATCH_LEN (= 2) +
673 * NUM_PRIMARY_LENS (= 7), the length header contains the match length
674 * minus MIN_MATCH_LEN, and there is no length footer.
676 * Otherwise, the length header contains NUM_PRIMARY_LENS, and the
677 * length footer contains the match length minus NUM_PRIMARY_LENS minus
679 if (match_len_minus_2 < LZX_NUM_PRIMARY_LENS) {
680 len_header = match_len_minus_2;
682 len_header = LZX_NUM_PRIMARY_LENS;
683 len_footer = match_len_minus_2 - LZX_NUM_PRIMARY_LENS;
686 /* Combine the position slot with the length header into a single symbol
687 * that will be encoded with the main code.
689 * The actual main symbol is offset by LZX_NUM_CHARS because values
690 * under LZX_NUM_CHARS are used to indicate a literal byte rather than a
692 main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS;
694 /* Output main symbol. */
695 lzx_write_varbits(os, codes->codewords.main[main_symbol],
696 codes->lens.main[main_symbol],
697 LZX_MAX_MAIN_CODEWORD_LEN);
699 /* If there is a length footer, output it using the
700 * length Huffman code. */
701 if (len_header == LZX_NUM_PRIMARY_LENS) {
702 lzx_write_varbits(os, codes->codewords.len[len_footer],
703 codes->lens.len[len_footer],
704 LZX_MAX_LEN_CODEWORD_LEN);
707 /* Output the position footer. */
709 num_extra_bits = lzx_get_num_extra_bits(position_slot);
711 if ((block_type == LZX_BLOCKTYPE_ALIGNED) && (num_extra_bits >= 3)) {
713 /* Aligned offset blocks: The low 3 bits of the position footer
714 * are Huffman-encoded using the aligned offset code. The
715 * remaining bits are output literally. */
717 lzx_write_varbits(os,
718 position_footer >> 3, num_extra_bits - 3, 14);
720 lzx_write_varbits(os,
721 codes->codewords.aligned[position_footer & 7],
722 codes->lens.aligned[position_footer & 7],
723 LZX_MAX_ALIGNED_CODEWORD_LEN);
725 /* Verbatim blocks, or fewer than 3 extra bits: All position
726 * footer bits are output literally. */
727 lzx_write_varbits(os, position_footer, num_extra_bits, 17);
731 /* Output an LZX literal (encoded with the main Huffman code). */
733 lzx_write_literal(struct lzx_output_bitstream *os, unsigned literal,
734 const struct lzx_codes *codes)
736 lzx_write_varbits(os, codes->codewords.main[literal],
737 codes->lens.main[literal], LZX_MAX_MAIN_CODEWORD_LEN);
741 lzx_build_precode(const u8 lens[restrict],
742 const u8 prev_lens[restrict],
743 const unsigned num_syms,
744 u32 precode_freqs[restrict LZX_PRECODE_NUM_SYMBOLS],
745 u8 output_syms[restrict num_syms],
746 u8 precode_lens[restrict LZX_PRECODE_NUM_SYMBOLS],
747 u32 precode_codewords[restrict LZX_PRECODE_NUM_SYMBOLS],
748 unsigned *num_additional_bits_ret)
750 memset(precode_freqs, 0,
751 LZX_PRECODE_NUM_SYMBOLS * sizeof(precode_freqs[0]));
753 /* Since the code word lengths use a form of RLE encoding, the goal here
754 * is to find each run of identical lengths when going through them in
755 * symbol order (including runs of length 1). For each run, as many
756 * lengths are encoded using RLE as possible, and the rest are output
759 * output_syms[] will be filled in with the length symbols that will be
760 * output, including RLE codes, not yet encoded using the precode.
762 * cur_run_len keeps track of how many code word lengths are in the
763 * current run of identical lengths. */
764 unsigned output_syms_idx = 0;
765 unsigned cur_run_len = 1;
766 unsigned num_additional_bits = 0;
767 for (unsigned i = 1; i <= num_syms; i++) {
769 if (i != num_syms && lens[i] == lens[i - 1]) {
770 /* Still in a run--- keep going. */
775 /* Run ended! Check if it is a run of zeroes or a run of
778 /* The symbol that was repeated in the run--- not to be confused
779 * with the length *of* the run (cur_run_len) */
780 unsigned len_in_run = lens[i - 1];
782 if (len_in_run == 0) {
783 /* A run of 0's. Encode it in as few length
784 * codes as we can. */
786 /* The magic length 18 indicates a run of 20 + n zeroes,
787 * where n is an uncompressed literal 5-bit integer that
788 * follows the magic length. */
789 while (cur_run_len >= 20) {
790 unsigned additional_bits;
792 additional_bits = min(cur_run_len - 20, 0x1f);
793 num_additional_bits += 5;
795 output_syms[output_syms_idx++] = 18;
796 output_syms[output_syms_idx++] = additional_bits;
797 cur_run_len -= 20 + additional_bits;
800 /* The magic length 17 indicates a run of 4 + n zeroes,
801 * where n is an uncompressed literal 4-bit integer that
802 * follows the magic length. */
803 while (cur_run_len >= 4) {
804 unsigned additional_bits;
806 additional_bits = min(cur_run_len - 4, 0xf);
807 num_additional_bits += 4;
809 output_syms[output_syms_idx++] = 17;
810 output_syms[output_syms_idx++] = additional_bits;
811 cur_run_len -= 4 + additional_bits;
816 /* A run of nonzero lengths. */
818 /* The magic length 19 indicates a run of 4 + n
819 * nonzeroes, where n is a literal bit that follows the
820 * magic length, and where the value of the lengths in
821 * the run is given by an extra length symbol, encoded
822 * with the precode, that follows the literal bit.
824 * The extra length symbol is encoded as a difference
825 * from the length of the codeword for the first symbol
826 * in the run in the previous code.
828 while (cur_run_len >= 4) {
829 unsigned additional_bits;
832 additional_bits = (cur_run_len > 4);
833 num_additional_bits += 1;
834 delta = (signed char)prev_lens[i - cur_run_len] -
835 (signed char)len_in_run;
839 precode_freqs[(unsigned char)delta]++;
840 output_syms[output_syms_idx++] = 19;
841 output_syms[output_syms_idx++] = additional_bits;
842 output_syms[output_syms_idx++] = delta;
843 cur_run_len -= 4 + additional_bits;
847 /* Any remaining lengths in the run are outputted without RLE,
848 * as a difference from the length of that codeword in the
850 while (cur_run_len > 0) {
853 delta = (signed char)prev_lens[i - cur_run_len] -
854 (signed char)len_in_run;
858 precode_freqs[(unsigned char)delta]++;
859 output_syms[output_syms_idx++] = delta;
866 /* Build the precode from the frequencies of the length symbols. */
868 make_canonical_huffman_code(LZX_PRECODE_NUM_SYMBOLS,
869 LZX_MAX_PRE_CODEWORD_LEN,
870 precode_freqs, precode_lens,
873 *num_additional_bits_ret = num_additional_bits;
875 return output_syms_idx;
879 * Output a Huffman code in the compressed form used in LZX.
881 * The Huffman code is represented in the output as a logical series of codeword
882 * lengths from which the Huffman code, which must be in canonical form, can be
885 * The codeword lengths are themselves compressed using a separate Huffman code,
886 * the "precode", which contains a symbol for each possible codeword length in
887 * the larger code as well as several special symbols to represent repeated
888 * codeword lengths (a form of run-length encoding). The precode is itself
889 * constructed in canonical form, and its codeword lengths are represented
890 * literally in 20 4-bit fields that immediately precede the compressed codeword
891 * lengths of the larger code.
893 * Furthermore, the codeword lengths of the larger code are actually represented
894 * as deltas from the codeword lengths of the corresponding code in the previous
898 * Bitstream to which to write the compressed Huffman code.
900 * The codeword lengths, indexed by symbol, in the Huffman code.
902 * The codeword lengths, indexed by symbol, in the corresponding Huffman
903 * code in the previous block, or all zeroes if this is the first block.
905 * The number of symbols in the Huffman code.
908 lzx_write_compressed_code(struct lzx_output_bitstream *os,
909 const u8 lens[restrict],
910 const u8 prev_lens[restrict],
913 u32 precode_freqs[LZX_PRECODE_NUM_SYMBOLS];
914 u8 output_syms[num_syms];
915 u8 precode_lens[LZX_PRECODE_NUM_SYMBOLS];
916 u32 precode_codewords[LZX_PRECODE_NUM_SYMBOLS];
918 unsigned num_output_syms;
922 num_output_syms = lzx_build_precode(lens,
931 /* Write the lengths of the precode codes to the output. */
932 for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++)
933 lzx_write_bits(os, precode_lens[i], LZX_PRECODE_ELEMENT_SIZE);
935 /* Write the length symbols, encoded with the precode, to the output. */
937 for (i = 0; i < num_output_syms; ) {
938 precode_sym = output_syms[i++];
940 lzx_write_varbits(os, precode_codewords[precode_sym],
941 precode_lens[precode_sym],
942 LZX_MAX_PRE_CODEWORD_LEN);
943 switch (precode_sym) {
945 lzx_write_bits(os, output_syms[i++], 4);
948 lzx_write_bits(os, output_syms[i++], 5);
951 lzx_write_bits(os, output_syms[i++], 1);
952 lzx_write_varbits(os, precode_codewords[output_syms[i]],
953 precode_lens[output_syms[i]],
954 LZX_MAX_PRE_CODEWORD_LEN);
964 * Write all matches and literal bytes (which were precomputed) in an LZX
965 * compressed block to the output bitstream in the final compressed
969 * The output bitstream.
971 * The chosen type of the LZX compressed block (LZX_BLOCKTYPE_ALIGNED or
972 * LZX_BLOCKTYPE_VERBATIM).
974 * The array of matches/literals to output.
976 * Number of matches/literals to output (length of @items).
978 * The main, length, and aligned offset Huffman codes for the current
979 * LZX compressed block.
982 lzx_write_items(struct lzx_output_bitstream *os, int block_type,
983 const struct lzx_item items[], u32 num_items,
984 const struct lzx_codes *codes)
986 for (u32 i = 0; i < num_items; i++) {
987 /* The high bit of the 32-bit intermediate representation
988 * indicates whether the item is an actual LZ-style match (1) or
989 * a literal byte (0). */
990 if (items[i].data & 0x80000000)
991 lzx_write_match(os, block_type, items[i], codes);
993 lzx_write_literal(os, items[i].data, codes);
997 /* Write an LZX aligned offset or verbatim block to the output. */
999 lzx_write_compressed_block(int block_type,
1001 u32 max_window_size,
1002 unsigned num_main_syms,
1003 struct lzx_item * chosen_items,
1004 u32 num_chosen_items,
1005 const struct lzx_codes * codes,
1006 const struct lzx_codes * prev_codes,
1007 struct lzx_output_bitstream * os)
1009 LZX_ASSERT(block_type == LZX_BLOCKTYPE_ALIGNED ||
1010 block_type == LZX_BLOCKTYPE_VERBATIM);
1012 /* The first three bits indicate the type of block and are one of the
1013 * LZX_BLOCKTYPE_* constants. */
1014 lzx_write_bits(os, block_type, 3);
1016 /* Output the block size.
1018 * The original LZX format seemed to always encode the block size in 3
1019 * bytes. However, the implementation in WIMGAPI, as used in WIM files,
1020 * uses the first bit to indicate whether the block is the default size
1021 * (32768) or a different size given explicitly by the next 16 bits.
1023 * By default, this compressor uses a window size of 32768 and therefore
1024 * follows the WIMGAPI behavior. However, this compressor also supports
1025 * window sizes greater than 32768 bytes, which do not appear to be
1026 * supported by WIMGAPI. In such cases, we retain the default size bit
1027 * to mean a size of 32768 bytes but output non-default block size in 24
1028 * bits rather than 16. The compatibility of this behavior is unknown
1029 * because WIMs created with chunk size greater than 32768 can seemingly
1030 * only be opened by wimlib anyway. */
1031 if (block_size == LZX_DEFAULT_BLOCK_SIZE) {
1032 lzx_write_bits(os, 1, 1);
1034 lzx_write_bits(os, 0, 1);
1036 if (max_window_size >= 65536)
1037 lzx_write_bits(os, block_size >> 16, 8);
1039 lzx_write_bits(os, block_size & 0xFFFF, 16);
1042 /* Output the aligned offset code. */
1043 if (block_type == LZX_BLOCKTYPE_ALIGNED) {
1044 for (int i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1045 lzx_write_bits(os, codes->lens.aligned[i],
1046 LZX_ALIGNEDCODE_ELEMENT_SIZE);
1050 /* Output the main code (two parts). */
1051 lzx_write_compressed_code(os, codes->lens.main,
1052 prev_codes->lens.main,
1054 lzx_write_compressed_code(os, codes->lens.main + LZX_NUM_CHARS,
1055 prev_codes->lens.main + LZX_NUM_CHARS,
1056 num_main_syms - LZX_NUM_CHARS);
1058 /* Output the length code. */
1059 lzx_write_compressed_code(os, codes->lens.len,
1060 prev_codes->lens.len,
1061 LZX_LENCODE_NUM_SYMBOLS);
1063 /* Output the compressed matches and literals. */
1064 lzx_write_items(os, block_type, chosen_items, num_chosen_items, codes);
1067 /* Write out the LZX blocks that were computed. */
1069 lzx_write_all_blocks(struct lzx_compressor *c, struct lzx_output_bitstream *os)
1072 const struct lzx_codes *prev_codes = &c->zero_codes;
1073 for (unsigned i = 0; i < c->num_blocks; i++) {
1074 const struct lzx_block_spec *spec = &c->block_specs[i];
1076 lzx_write_compressed_block(spec->block_type,
1081 spec->num_chosen_items,
1086 prev_codes = &spec->codes;
1090 /* Constructs an LZX match from a literal byte and updates the main code symbol
1093 lzx_tally_literal(u8 lit, struct lzx_freqs *freqs)
1099 /* Constructs an LZX match from an offset and a length, and updates the LRU
1100 * queue and the frequency of symbols in the main, length, and aligned offset
1101 * alphabets. The return value is a 32-bit number that provides the match in an
1102 * intermediate representation documented below. */
1104 lzx_tally_match(unsigned match_len, u32 match_offset,
1105 struct lzx_freqs *freqs, struct lzx_lru_queue *queue)
1107 unsigned position_slot;
1108 u32 position_footer;
1110 unsigned main_symbol;
1111 unsigned len_footer;
1112 unsigned adjusted_match_len;
1114 LZX_ASSERT(match_len >= LZX_MIN_MATCH_LEN && match_len <= LZX_MAX_MATCH_LEN);
1116 /* The match offset shall be encoded as a position slot (itself encoded
1117 * as part of the main symbol) and a position footer. */
1118 position_slot = lzx_get_position_slot(match_offset, queue);
1119 position_footer = (match_offset + LZX_OFFSET_OFFSET) &
1120 (((u32)1 << lzx_get_num_extra_bits(position_slot)) - 1);
1122 /* The match length shall be encoded as a length header (itself encoded
1123 * as part of the main symbol) and an optional length footer. */
1124 adjusted_match_len = match_len - LZX_MIN_MATCH_LEN;
1125 if (adjusted_match_len < LZX_NUM_PRIMARY_LENS) {
1126 /* No length footer needed. */
1127 len_header = adjusted_match_len;
1129 /* Length footer needed. It will be encoded using the length
1131 len_header = LZX_NUM_PRIMARY_LENS;
1132 len_footer = adjusted_match_len - LZX_NUM_PRIMARY_LENS;
1133 freqs->len[len_footer]++;
1136 /* Account for the main symbol. */
1137 main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS;
1139 freqs->main[main_symbol]++;
1141 /* In an aligned offset block, 3 bits of the position footer are output
1142 * as an aligned offset symbol. Account for this, although we may
1143 * ultimately decide to output the block as verbatim. */
1145 /* The following check is equivalent to:
1147 * if (lzx_extra_bits[position_slot] >= 3)
1149 * Note that this correctly excludes position slots that correspond to
1150 * recent offsets. */
1151 if (position_slot >= 8)
1152 freqs->aligned[position_footer & 7]++;
1154 /* Pack the position slot, position footer, and match length into an
1155 * intermediate representation. See `struct lzx_item' for details.
1157 LZX_ASSERT(LZX_MAX_POSITION_SLOTS <= 64);
1158 LZX_ASSERT(lzx_get_num_extra_bits(LZX_MAX_POSITION_SLOTS - 1) <= 17);
1159 LZX_ASSERT(LZX_MAX_MATCH_LEN - LZX_MIN_MATCH_LEN + 1 <= 256);
1161 LZX_ASSERT(position_slot <= (1U << (31 - 25)) - 1);
1162 LZX_ASSERT(position_footer <= (1U << (25 - 8)) - 1);
1163 LZX_ASSERT(adjusted_match_len <= (1U << (8 - 0)) - 1);
1165 (position_slot << 25) |
1166 (position_footer << 8) |
1167 (adjusted_match_len);
1170 /* Returns the cost, in bits, to output a literal byte using the specified cost
1173 lzx_literal_cost(u8 c, const struct lzx_costs * costs)
1175 return costs->main[c];
1178 /* Returns the cost, in bits, to output a repeat offset match of the specified
1179 * length and position slot (repeat index) using the specified cost model. */
1181 lzx_repmatch_cost(u32 len, unsigned position_slot, const struct lzx_costs *costs)
1183 unsigned len_header, main_symbol;
1186 len_header = min(len - LZX_MIN_MATCH_LEN, LZX_NUM_PRIMARY_LENS);
1187 main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS;
1189 /* Account for main symbol. */
1190 cost += costs->main[main_symbol];
1192 /* Account for extra length information. */
1193 if (len_header == LZX_NUM_PRIMARY_LENS)
1194 cost += costs->len[len - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS];
1199 /* Set the cost model @c->costs from the Huffman codeword lengths specified in
1202 * The cost model and codeword lengths are almost the same thing, but the
1203 * Huffman codewords with length 0 correspond to symbols with zero frequency
1204 * that still need to be assigned actual costs. The specific values assigned
1205 * are arbitrary, but they should be fairly high (near the maximum codeword
1206 * length) to take into account the fact that uses of these symbols are expected
1209 lzx_set_costs(struct lzx_compressor *c, const struct lzx_lens * lens,
1215 for (i = 0; i < c->num_main_syms; i++)
1216 c->costs.main[i] = lens->main[i] ? lens->main[i] : nostat;
1219 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
1220 c->costs.len[i] = lens->len[i] ? lens->len[i] : nostat;
1222 /* Aligned offset code */
1223 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
1224 c->costs.aligned[i] = lens->aligned[i] ? lens->aligned[i] : nostat / 2;
1227 /* Don't allow matches to span the end of an LZX block. */
1229 maybe_truncate_matches(struct lz_match matches[], u32 num_matches,
1230 struct lzx_compressor *c)
1232 if (c->match_window_end < c->cur_window_size && num_matches != 0) {
1233 u32 limit = c->match_window_end - c->match_window_pos;
1235 if (limit >= LZX_MIN_MATCH_LEN) {
1237 u32 i = num_matches - 1;
1239 if (matches[i].len >= limit) {
1240 matches[i].len = limit;
1242 /* Truncation might produce multiple
1243 * matches with length 'limit'. Keep at
1245 num_matches = i + 1;
1256 lzx_get_matches_fillcache_singleblock(struct lzx_compressor *c,
1257 const struct lz_match **matches_ret)
1259 struct lz_match *cache_ptr;
1260 struct lz_match *matches;
1261 unsigned num_matches;
1263 cache_ptr = c->cache_ptr;
1264 matches = cache_ptr + 1;
1265 if (likely(cache_ptr <= c->cache_limit)) {
1266 num_matches = lz_mf_get_matches(c->mf, matches);
1267 cache_ptr->len = num_matches;
1268 c->cache_ptr = matches + num_matches;
1272 c->match_window_pos++;
1273 *matches_ret = matches;
1278 lzx_get_matches_fillcache_multiblock(struct lzx_compressor *c,
1279 const struct lz_match **matches_ret)
1281 struct lz_match *cache_ptr;
1282 struct lz_match *matches;
1283 unsigned num_matches;
1285 cache_ptr = c->cache_ptr;
1286 matches = cache_ptr + 1;
1287 if (likely(cache_ptr <= c->cache_limit)) {
1288 num_matches = lz_mf_get_matches(c->mf, matches);
1289 num_matches = maybe_truncate_matches(matches, num_matches, c);
1290 cache_ptr->len = num_matches;
1291 c->cache_ptr = matches + num_matches;
1295 c->match_window_pos++;
1296 *matches_ret = matches;
1301 lzx_get_matches_usecache(struct lzx_compressor *c,
1302 const struct lz_match **matches_ret)
1304 struct lz_match *cache_ptr;
1305 struct lz_match *matches;
1306 unsigned num_matches;
1308 cache_ptr = c->cache_ptr;
1309 matches = cache_ptr + 1;
1310 if (cache_ptr <= c->cache_limit) {
1311 num_matches = cache_ptr->len;
1312 c->cache_ptr = matches + num_matches;
1316 c->match_window_pos++;
1317 *matches_ret = matches;
1322 lzx_get_matches_usecache_nocheck(struct lzx_compressor *c,
1323 const struct lz_match **matches_ret)
1325 struct lz_match *cache_ptr;
1326 struct lz_match *matches;
1327 unsigned num_matches;
1329 cache_ptr = c->cache_ptr;
1330 matches = cache_ptr + 1;
1331 num_matches = cache_ptr->len;
1332 c->cache_ptr = matches + num_matches;
1333 c->match_window_pos++;
1334 *matches_ret = matches;
1339 lzx_get_matches_nocache_singleblock(struct lzx_compressor *c,
1340 const struct lz_match **matches_ret)
1342 struct lz_match *matches;
1343 unsigned num_matches;
1345 matches = c->cache_ptr;
1346 num_matches = lz_mf_get_matches(c->mf, matches);
1347 c->match_window_pos++;
1348 *matches_ret = matches;
1353 lzx_get_matches_nocache_multiblock(struct lzx_compressor *c,
1354 const struct lz_match **matches_ret)
1356 struct lz_match *matches;
1357 unsigned num_matches;
1359 matches = c->cache_ptr;
1360 num_matches = lz_mf_get_matches(c->mf, matches);
1361 num_matches = maybe_truncate_matches(matches, num_matches, c);
1362 c->match_window_pos++;
1363 *matches_ret = matches;
1368 * Find matches at the next position in the window.
1370 * Returns the number of matches found and sets *matches_ret to point to the
1371 * matches array. The matches will be sorted by strictly increasing length and
1374 static inline unsigned
1375 lzx_get_matches(struct lzx_compressor *c,
1376 const struct lz_match **matches_ret)
1378 return (*c->get_matches_func)(c, matches_ret);
1382 lzx_skip_bytes_fillcache(struct lzx_compressor *c, unsigned n)
1384 struct lz_match *cache_ptr;
1386 cache_ptr = c->cache_ptr;
1387 c->match_window_pos += n;
1388 lz_mf_skip_positions(c->mf, n);
1389 if (cache_ptr <= c->cache_limit) {
1393 } while (--n && cache_ptr <= c->cache_limit);
1395 c->cache_ptr = cache_ptr;
1399 lzx_skip_bytes_usecache(struct lzx_compressor *c, unsigned n)
1401 struct lz_match *cache_ptr;
1403 cache_ptr = c->cache_ptr;
1404 c->match_window_pos += n;
1405 if (cache_ptr <= c->cache_limit) {
1407 cache_ptr += 1 + cache_ptr->len;
1408 } while (--n && cache_ptr <= c->cache_limit);
1410 c->cache_ptr = cache_ptr;
1414 lzx_skip_bytes_usecache_nocheck(struct lzx_compressor *c, unsigned n)
1416 struct lz_match *cache_ptr;
1418 cache_ptr = c->cache_ptr;
1419 c->match_window_pos += n;
1421 cache_ptr += 1 + cache_ptr->len;
1423 c->cache_ptr = cache_ptr;
1427 lzx_skip_bytes_nocache(struct lzx_compressor *c, unsigned n)
1429 c->match_window_pos += n;
1430 lz_mf_skip_positions(c->mf, n);
1434 * Skip the specified number of positions in the window (don't search for
1438 lzx_skip_bytes(struct lzx_compressor *c, unsigned n)
1440 return (*c->skip_bytes_func)(c, n);
1444 * Reverse the linked list of near-optimal matches so that they can be returned
1445 * in forwards order.
1447 * Returns the first match in the list.
1449 static struct lz_match
1450 lzx_match_chooser_reverse_list(struct lzx_compressor *c, unsigned cur_pos)
1452 unsigned prev_link, saved_prev_link;
1453 unsigned prev_match_offset, saved_prev_match_offset;
1455 c->optimum_end_idx = cur_pos;
1457 saved_prev_link = c->optimum[cur_pos].prev.link;
1458 saved_prev_match_offset = c->optimum[cur_pos].prev.match_offset;
1461 prev_link = saved_prev_link;
1462 prev_match_offset = saved_prev_match_offset;
1464 saved_prev_link = c->optimum[prev_link].prev.link;
1465 saved_prev_match_offset = c->optimum[prev_link].prev.match_offset;
1467 c->optimum[prev_link].next.link = cur_pos;
1468 c->optimum[prev_link].next.match_offset = prev_match_offset;
1470 cur_pos = prev_link;
1471 } while (cur_pos != 0);
1473 c->optimum_cur_idx = c->optimum[0].next.link;
1475 return (struct lz_match)
1476 { .len = c->optimum_cur_idx,
1477 .offset = c->optimum[0].next.match_offset,
1482 * lzx_choose_near_optimal_match() -
1484 * Choose an approximately optimal match or literal to use at the next position
1485 * in the string, or "window", being LZ-encoded.
1487 * This is based on algorithms used in 7-Zip, including the DEFLATE encoder
1488 * and the LZMA encoder, written by Igor Pavlov.
1490 * Unlike a greedy parser that always takes the longest match, or even a "lazy"
1491 * parser with one match/literal look-ahead like zlib, the algorithm used here
1492 * may look ahead many matches/literals to determine the approximately optimal
1493 * match/literal to code next. The motivation is that the compression ratio is
1494 * improved if the compressor can do things like use a shorter-than-possible
1495 * match in order to allow a longer match later, and also take into account the
1496 * estimated real cost of coding each match/literal based on the underlying
1499 * Still, this is not a true optimal parser for several reasons:
1501 * - Real compression formats use entropy encoding of the literal/match
1502 * sequence, so the real cost of coding each match or literal is unknown until
1503 * the parse is fully determined. It can be approximated based on iterative
1504 * parses, but the end result is not guaranteed to be globally optimal.
1506 * - Very long matches are chosen immediately. This is because locations with
1507 * long matches are likely to have many possible alternatives that would cause
1508 * slow optimal parsing, but also such locations are already highly
1509 * compressible so it is not too harmful to just grab the longest match.
1511 * - Not all possible matches at each location are considered because the
1512 * underlying match-finder limits the number and type of matches produced at
1513 * each position. For example, for a given match length it's usually not
1514 * worth it to only consider matches other than the lowest-offset match,
1515 * except in the case of a repeat offset.
1517 * - Although we take into account the adaptive state (in LZX, the recent offset
1518 * queue), coding decisions made with respect to the adaptive state will be
1519 * locally optimal but will not necessarily be globally optimal. This is
1520 * because the algorithm only keeps the least-costly path to get to a given
1521 * location and does not take into account that a slightly more costly path
1522 * could result in a different adaptive state that ultimately results in a
1523 * lower global cost.
1525 * - The array space used by this function is bounded, so in degenerate cases it
1526 * is forced to start returning matches/literals before the algorithm has
1529 * Each call to this function does one of two things:
1531 * 1. Build a sequence of near-optimal matches/literals, up to some point, that
1532 * will be returned by subsequent calls to this function, then return the
1537 * 2. Return the next match/literal previously computed by a call to this
1540 * The return value is a (length, offset) pair specifying the match or literal
1541 * chosen. For literals, the length is 0 or 1 and the offset is meaningless.
1543 static struct lz_match
1544 lzx_choose_near_optimal_item(struct lzx_compressor *c)
1546 unsigned num_matches;
1547 const struct lz_match *matches;
1548 struct lz_match match;
1550 u32 longest_rep_len;
1551 unsigned longest_rep_slot;
1554 struct lzx_mc_pos_data *optimum = c->optimum;
1556 if (c->optimum_cur_idx != c->optimum_end_idx) {
1557 /* Case 2: Return the next match/literal already found. */
1558 match.len = optimum[c->optimum_cur_idx].next.link -
1560 match.offset = optimum[c->optimum_cur_idx].next.match_offset;
1562 c->optimum_cur_idx = optimum[c->optimum_cur_idx].next.link;
1566 /* Case 1: Compute a new list of matches/literals to return. */
1568 c->optimum_cur_idx = 0;
1569 c->optimum_end_idx = 0;
1571 /* Search for matches at repeat offsets. As a heuristic, we only keep
1572 * the one with the longest match length. */
1573 longest_rep_len = LZX_MIN_MATCH_LEN - 1;
1574 if (c->match_window_pos >= 1) {
1575 unsigned limit = min(LZX_MAX_MATCH_LEN,
1576 c->match_window_end - c->match_window_pos);
1577 for (int i = 0; i < LZX_NUM_RECENT_OFFSETS; i++) {
1578 u32 offset = c->queue.R[i];
1579 const u8 *strptr = &c->cur_window[c->match_window_pos];
1580 const u8 *matchptr = strptr - offset;
1582 while (len < limit && strptr[len] == matchptr[len])
1584 if (len > longest_rep_len) {
1585 longest_rep_len = len;
1586 longest_rep_slot = i;
1591 /* If there's a long match with a repeat offset, choose it immediately. */
1592 if (longest_rep_len >= c->params.nice_match_length) {
1593 lzx_skip_bytes(c, longest_rep_len);
1594 return (struct lz_match) {
1595 .len = longest_rep_len,
1596 .offset = c->queue.R[longest_rep_slot],
1600 /* Find other matches. */
1601 num_matches = lzx_get_matches(c, &matches);
1603 /* If there's a long match, choose it immediately. */
1605 longest_len = matches[num_matches - 1].len;
1606 if (longest_len >= c->params.nice_match_length) {
1607 lzx_skip_bytes(c, longest_len - 1);
1608 return matches[num_matches - 1];
1614 /* Calculate the cost to reach the next position by coding a literal. */
1615 optimum[1].queue = c->queue;
1616 optimum[1].cost = lzx_literal_cost(c->cur_window[c->match_window_pos - 1],
1618 optimum[1].prev.link = 0;
1620 /* Calculate the cost to reach any position up to and including that
1621 * reached by the longest match.
1623 * Note: We consider only the lowest-offset match that reaches each
1626 * Note: Some of the cost calculation stays the same for each offset,
1627 * regardless of how many lengths it gets used for. Therefore, to
1628 * improve performance, we hand-code the cost calculation instead of
1629 * calling lzx_match_cost() to do a from-scratch cost evaluation at each
1631 for (unsigned i = 0, len = 2; i < num_matches; i++) {
1633 struct lzx_lru_queue queue;
1635 unsigned position_slot;
1636 unsigned num_extra_bits;
1638 offset = matches[i].offset;
1642 position_slot = lzx_get_position_slot(offset, &queue);
1643 num_extra_bits = lzx_get_num_extra_bits(position_slot);
1644 if (num_extra_bits >= 3) {
1645 position_cost += num_extra_bits - 3;
1646 position_cost += c->costs.aligned[(offset + LZX_OFFSET_OFFSET) & 7];
1648 position_cost += num_extra_bits;
1652 unsigned len_header;
1653 unsigned main_symbol;
1656 cost = position_cost;
1658 len_header = min(len - LZX_MIN_MATCH_LEN, LZX_NUM_PRIMARY_LENS);
1659 main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS;
1660 cost += c->costs.main[main_symbol];
1661 if (len_header == LZX_NUM_PRIMARY_LENS)
1662 cost += c->costs.len[len - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS];
1664 optimum[len].queue = queue;
1665 optimum[len].prev.link = 0;
1666 optimum[len].prev.match_offset = offset;
1667 optimum[len].cost = cost;
1668 } while (++len <= matches[i].len);
1670 end_pos = longest_len;
1672 if (longest_rep_len >= LZX_MIN_MATCH_LEN) {
1675 while (end_pos < longest_rep_len)
1676 optimum[++end_pos].cost = MC_INFINITE_COST;
1678 cost = lzx_repmatch_cost(longest_rep_len, longest_rep_slot,
1680 if (cost <= optimum[longest_rep_len].cost) {
1681 optimum[longest_rep_len].queue = c->queue;
1682 swap(optimum[longest_rep_len].queue.R[0],
1683 optimum[longest_rep_len].queue.R[longest_rep_slot]);
1684 optimum[longest_rep_len].prev.link = 0;
1685 optimum[longest_rep_len].prev.match_offset =
1686 optimum[longest_rep_len].queue.R[0];
1687 optimum[longest_rep_len].cost = cost;
1691 /* Step forward, calculating the estimated minimum cost to reach each
1692 * position. The algorithm may find multiple paths to reach each
1693 * position; only the lowest-cost path is saved.
1695 * The progress of the parse is tracked in the @optimum array, which for
1696 * each position contains the minimum cost to reach that position, the
1697 * index of the start of the match/literal taken to reach that position
1698 * through the minimum-cost path, the offset of the match taken (not
1699 * relevant for literals), and the adaptive state that will exist at
1700 * that position after the minimum-cost path is taken. The @cur_pos
1701 * variable stores the position at which the algorithm is currently
1702 * considering coding choices, and the @end_pos variable stores the
1703 * greatest position at which the costs of coding choices have been
1706 * The loop terminates when any one of the following conditions occurs:
1708 * 1. A match with length greater than or equal to @nice_match_length is
1709 * found. When this occurs, the algorithm chooses this match
1710 * unconditionally, and consequently the near-optimal match/literal
1711 * sequence up to and including that match is fully determined and it
1712 * can begin returning the match/literal list.
1714 * 2. @cur_pos reaches a position not overlapped by a preceding match.
1715 * In such cases, the near-optimal match/literal sequence up to
1716 * @cur_pos is fully determined and it can begin returning the
1717 * match/literal list.
1719 * 3. Failing either of the above in a degenerate case, the loop
1720 * terminates when space in the @optimum array is exhausted.
1721 * This terminates the algorithm and forces it to start returning
1722 * matches/literals even though they may not be globally optimal.
1724 * Upon loop termination, a nonempty list of matches/literals will have
1725 * been produced and stored in the @optimum array. These
1726 * matches/literals are linked in reverse order, so the last thing this
1727 * function does is reverse this list and return the first
1728 * match/literal, leaving the rest to be returned immediately by
1729 * subsequent calls to this function.
1735 /* Advance to next position. */
1738 /* Check termination conditions (2) and (3) noted above. */
1739 if (cur_pos == end_pos || cur_pos == LZX_OPTIM_ARRAY_LENGTH)
1740 return lzx_match_chooser_reverse_list(c, cur_pos);
1742 /* Search for matches at repeat offsets. Again, as a heuristic
1743 * we only keep the longest one. */
1744 longest_rep_len = LZX_MIN_MATCH_LEN - 1;
1745 unsigned limit = min(LZX_MAX_MATCH_LEN,
1746 c->match_window_end - c->match_window_pos);
1747 for (int i = 0; i < LZX_NUM_RECENT_OFFSETS; i++) {
1748 u32 offset = optimum[cur_pos].queue.R[i];
1749 const u8 *strptr = &c->cur_window[c->match_window_pos];
1750 const u8 *matchptr = strptr - offset;
1752 while (len < limit && strptr[len] == matchptr[len])
1754 if (len > longest_rep_len) {
1755 longest_rep_len = len;
1756 longest_rep_slot = i;
1760 /* If we found a long match at a repeat offset, choose it
1762 if (longest_rep_len >= c->params.nice_match_length) {
1763 /* Build the list of matches to return and get
1765 match = lzx_match_chooser_reverse_list(c, cur_pos);
1767 /* Append the long match to the end of the list. */
1768 optimum[cur_pos].next.match_offset =
1769 optimum[cur_pos].queue.R[longest_rep_slot];
1770 optimum[cur_pos].next.link = cur_pos + longest_rep_len;
1771 c->optimum_end_idx = cur_pos + longest_rep_len;
1773 /* Skip over the remaining bytes of the long match. */
1774 lzx_skip_bytes(c, longest_rep_len);
1776 /* Return first match in the list. */
1780 /* Find other matches. */
1781 num_matches = lzx_get_matches(c, &matches);
1783 /* If there's a long match, choose it immediately. */
1785 longest_len = matches[num_matches - 1].len;
1786 if (longest_len >= c->params.nice_match_length) {
1787 /* Build the list of matches to return and get
1789 match = lzx_match_chooser_reverse_list(c, cur_pos);
1791 /* Append the long match to the end of the list. */
1792 optimum[cur_pos].next.match_offset =
1793 matches[num_matches - 1].offset;
1794 optimum[cur_pos].next.link = cur_pos + longest_len;
1795 c->optimum_end_idx = cur_pos + longest_len;
1797 /* Skip over the remaining bytes of the long match. */
1798 lzx_skip_bytes(c, longest_len - 1);
1800 /* Return first match in the list. */
1807 /* If we are reaching any positions for the first time, we need
1808 * to initialize their costs to infinity. */
1809 while (end_pos < cur_pos + longest_len)
1810 optimum[++end_pos].cost = MC_INFINITE_COST;
1812 /* Consider coding a literal. */
1813 cost = optimum[cur_pos].cost +
1814 lzx_literal_cost(c->cur_window[c->match_window_pos - 1],
1816 if (cost < optimum[cur_pos + 1].cost) {
1817 optimum[cur_pos + 1].queue = optimum[cur_pos].queue;
1818 optimum[cur_pos + 1].cost = cost;
1819 optimum[cur_pos + 1].prev.link = cur_pos;
1822 /* Consider coding a match.
1824 * The hard-coded cost calculation is done for the same reason
1825 * stated in the comment for the similar loop earlier.
1826 * Actually, it is *this* one that has the biggest effect on
1827 * performance; overall LZX compression is > 10% faster with
1828 * this code compared to calling lzx_match_cost() with each
1830 for (unsigned i = 0, len = 2; i < num_matches; i++) {
1833 unsigned position_slot;
1834 unsigned num_extra_bits;
1836 offset = matches[i].offset;
1837 position_cost = optimum[cur_pos].cost;
1839 /* Yet another optimization: instead of calling
1840 * lzx_get_position_slot(), hand-inline the search of
1841 * the repeat offset queue. Then we can omit the
1842 * extra_bits calculation for repeat offset matches, and
1843 * also only compute the updated queue if we actually do
1844 * find a new lowest cost path. */
1845 for (position_slot = 0; position_slot < LZX_NUM_RECENT_OFFSETS; position_slot++)
1846 if (offset == optimum[cur_pos].queue.R[position_slot])
1847 goto have_position_cost;
1849 position_slot = lzx_get_position_slot_raw(offset + LZX_OFFSET_OFFSET);
1851 num_extra_bits = lzx_get_num_extra_bits(position_slot);
1852 if (num_extra_bits >= 3) {
1853 position_cost += num_extra_bits - 3;
1854 position_cost += c->costs.aligned[
1855 (offset + LZX_OFFSET_OFFSET) & 7];
1857 position_cost += num_extra_bits;
1863 unsigned len_header;
1864 unsigned main_symbol;
1867 cost = position_cost;
1869 len_header = min(len - LZX_MIN_MATCH_LEN,
1870 LZX_NUM_PRIMARY_LENS);
1871 main_symbol = ((position_slot << 3) | len_header) +
1873 cost += c->costs.main[main_symbol];
1874 if (len_header == LZX_NUM_PRIMARY_LENS) {
1875 cost += c->costs.len[len -
1877 LZX_NUM_PRIMARY_LENS];
1879 if (cost < optimum[cur_pos + len].cost) {
1880 if (position_slot < LZX_NUM_RECENT_OFFSETS) {
1881 optimum[cur_pos + len].queue = optimum[cur_pos].queue;
1882 swap(optimum[cur_pos + len].queue.R[0],
1883 optimum[cur_pos + len].queue.R[position_slot]);
1885 optimum[cur_pos + len].queue.R[0] = offset;
1886 optimum[cur_pos + len].queue.R[1] = optimum[cur_pos].queue.R[0];
1887 optimum[cur_pos + len].queue.R[2] = optimum[cur_pos].queue.R[1];
1889 optimum[cur_pos + len].prev.link = cur_pos;
1890 optimum[cur_pos + len].prev.match_offset = offset;
1891 optimum[cur_pos + len].cost = cost;
1893 } while (++len <= matches[i].len);
1896 /* Consider coding a repeat offset match.
1898 * As a heuristic, we only consider the longest length of the
1899 * longest repeat offset match. This does not, however,
1900 * necessarily mean that we will never consider any other repeat
1901 * offsets, because above we detect repeat offset matches that
1902 * were found by the regular match-finder. Therefore, this
1903 * special handling of the longest repeat-offset match is only
1904 * helpful for coding a repeat offset match that was *not* found
1905 * by the match-finder, e.g. due to being obscured by a less
1906 * distant match that is at least as long.
1908 * Note: an alternative, used in LZMA, is to consider every
1909 * length of every repeat offset match. This is a more thorough
1910 * search, and it makes it unnecessary to detect repeat offset
1911 * matches that were found by the regular match-finder. But by
1912 * my tests, for LZX the LZMA method slows down the compressor
1913 * by ~10% and doesn't actually help the compression ratio too
1916 * Also tested a compromise approach: consider every 3rd length
1917 * of the longest repeat offset match. Still didn't seem quite
1920 if (longest_rep_len >= LZX_MIN_MATCH_LEN) {
1922 while (end_pos < cur_pos + longest_rep_len)
1923 optimum[++end_pos].cost = MC_INFINITE_COST;
1925 cost = optimum[cur_pos].cost +
1926 lzx_repmatch_cost(longest_rep_len, longest_rep_slot,
1928 if (cost <= optimum[cur_pos + longest_rep_len].cost) {
1929 optimum[cur_pos + longest_rep_len].queue =
1930 optimum[cur_pos].queue;
1931 swap(optimum[cur_pos + longest_rep_len].queue.R[0],
1932 optimum[cur_pos + longest_rep_len].queue.R[longest_rep_slot]);
1933 optimum[cur_pos + longest_rep_len].prev.link =
1935 optimum[cur_pos + longest_rep_len].prev.match_offset =
1936 optimum[cur_pos + longest_rep_len].queue.R[0];
1937 optimum[cur_pos + longest_rep_len].cost =
1944 static struct lz_match
1945 lzx_choose_lazy_item(struct lzx_compressor *c)
1947 const struct lz_match *matches;
1948 struct lz_match cur_match;
1949 struct lz_match next_match;
1952 if (c->prev_match.len) {
1953 cur_match = c->prev_match;
1954 c->prev_match.len = 0;
1956 num_matches = lzx_get_matches(c, &matches);
1957 if (num_matches == 0 ||
1958 (matches[num_matches - 1].len <= 3 &&
1959 (matches[num_matches - 1].len <= 2 ||
1960 matches[num_matches - 1].offset > 4096)))
1962 return (struct lz_match) { };
1965 cur_match = matches[num_matches - 1];
1968 if (cur_match.len >= c->params.nice_match_length) {
1969 lzx_skip_bytes(c, cur_match.len - 1);
1973 num_matches = lzx_get_matches(c, &matches);
1974 if (num_matches == 0 ||
1975 (matches[num_matches - 1].len <= 3 &&
1976 (matches[num_matches - 1].len <= 2 ||
1977 matches[num_matches - 1].offset > 4096)))
1979 lzx_skip_bytes(c, cur_match.len - 2);
1983 next_match = matches[num_matches - 1];
1985 if (next_match.len <= cur_match.len) {
1986 lzx_skip_bytes(c, cur_match.len - 2);
1989 c->prev_match = next_match;
1990 return (struct lz_match) { };
1995 * Return the next match or literal to use, delegating to the currently selected
1996 * match-choosing algorithm.
1998 * If the length of the returned 'struct lz_match' is less than
1999 * LZX_MIN_MATCH_LEN, then it is really a literal.
2001 static inline struct lz_match
2002 lzx_choose_item(struct lzx_compressor *c)
2004 return (*c->params.choose_item_func)(c);
2007 /* Set default symbol costs for the LZX Huffman codes. */
2009 lzx_set_default_costs(struct lzx_costs * costs, unsigned num_main_syms)
2013 /* Main code (part 1): Literal symbols */
2014 for (i = 0; i < LZX_NUM_CHARS; i++)
2017 /* Main code (part 2): Match header symbols */
2018 for (; i < num_main_syms; i++)
2019 costs->main[i] = 10;
2022 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
2025 /* Aligned offset code */
2026 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
2027 costs->aligned[i] = 3;
2030 /* Given the frequencies of symbols in an LZX-compressed block and the
2031 * corresponding Huffman codes, return LZX_BLOCKTYPE_ALIGNED or
2032 * LZX_BLOCKTYPE_VERBATIM if an aligned offset or verbatim block, respectively,
2033 * will take fewer bits to output. */
2035 lzx_choose_verbatim_or_aligned(const struct lzx_freqs * freqs,
2036 const struct lzx_codes * codes)
2038 unsigned aligned_cost = 0;
2039 unsigned verbatim_cost = 0;
2041 /* Verbatim blocks have a constant 3 bits per position footer. Aligned
2042 * offset blocks have an aligned offset symbol per position footer, plus
2043 * an extra 24 bits per block to output the lengths necessary to
2044 * reconstruct the aligned offset code itself. */
2045 for (unsigned i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
2046 verbatim_cost += 3 * freqs->aligned[i];
2047 aligned_cost += codes->lens.aligned[i] * freqs->aligned[i];
2049 aligned_cost += LZX_ALIGNEDCODE_ELEMENT_SIZE * LZX_ALIGNEDCODE_NUM_SYMBOLS;
2050 if (aligned_cost < verbatim_cost)
2051 return LZX_BLOCKTYPE_ALIGNED;
2053 return LZX_BLOCKTYPE_VERBATIM;
2056 /* Find a sequence of matches/literals with which to output the specified LZX
2057 * block, then set the block's type to that which has the minimum cost to output
2058 * (either verbatim or aligned). */
2060 lzx_choose_items_for_block(struct lzx_compressor *c, struct lzx_block_spec *spec)
2062 const struct lzx_lru_queue orig_queue = c->queue;
2063 u32 num_passes_remaining = c->params.num_optim_passes;
2064 struct lzx_freqs freqs;
2065 const u8 *window_ptr;
2066 const u8 *window_end;
2067 struct lzx_item *next_chosen_item;
2068 struct lz_match lz_match;
2069 struct lzx_item lzx_item;
2071 LZX_ASSERT(num_passes_remaining >= 1);
2072 LZX_ASSERT(lz_mf_get_position(c->mf) == spec->window_pos);
2074 c->match_window_end = spec->window_pos + spec->block_size;
2076 if (c->params.num_optim_passes > 1) {
2077 if (spec->block_size == c->cur_window_size)
2078 c->get_matches_func = lzx_get_matches_fillcache_singleblock;
2080 c->get_matches_func = lzx_get_matches_fillcache_multiblock;
2081 c->skip_bytes_func = lzx_skip_bytes_fillcache;
2083 if (spec->block_size == c->cur_window_size)
2084 c->get_matches_func = lzx_get_matches_nocache_singleblock;
2086 c->get_matches_func = lzx_get_matches_nocache_multiblock;
2087 c->skip_bytes_func = lzx_skip_bytes_nocache;
2090 /* The first optimal parsing pass is done using the cost model already
2091 * set in c->costs. Each later pass is done using a cost model
2092 * computed from the previous pass.
2094 * To improve performance we only generate the array containing the
2095 * matches and literals in intermediate form on the final pass. */
2097 while (--num_passes_remaining) {
2098 c->match_window_pos = spec->window_pos;
2099 c->cache_ptr = c->cached_matches;
2100 memset(&freqs, 0, sizeof(freqs));
2101 window_ptr = &c->cur_window[spec->window_pos];
2102 window_end = window_ptr + spec->block_size;
2104 while (window_ptr != window_end) {
2106 lz_match = lzx_choose_item(c);
2108 LZX_ASSERT(!(lz_match.len == LZX_MIN_MATCH_LEN &&
2109 lz_match.offset == c->max_window_size -
2110 LZX_MIN_MATCH_LEN));
2111 if (lz_match.len >= LZX_MIN_MATCH_LEN) {
2112 lzx_tally_match(lz_match.len, lz_match.offset,
2114 window_ptr += lz_match.len;
2116 lzx_tally_literal(*window_ptr, &freqs);
2120 lzx_make_huffman_codes(&freqs, &spec->codes, c->num_main_syms);
2121 lzx_set_costs(c, &spec->codes.lens, 15);
2122 c->queue = orig_queue;
2123 if (c->cache_ptr <= c->cache_limit) {
2124 c->get_matches_func = lzx_get_matches_usecache_nocheck;
2125 c->skip_bytes_func = lzx_skip_bytes_usecache_nocheck;
2127 c->get_matches_func = lzx_get_matches_usecache;
2128 c->skip_bytes_func = lzx_skip_bytes_usecache;
2132 c->match_window_pos = spec->window_pos;
2133 c->cache_ptr = c->cached_matches;
2134 memset(&freqs, 0, sizeof(freqs));
2135 window_ptr = &c->cur_window[spec->window_pos];
2136 window_end = window_ptr + spec->block_size;
2138 spec->chosen_items = &c->chosen_items[spec->window_pos];
2139 next_chosen_item = spec->chosen_items;
2141 unsigned unseen_cost = 9;
2142 while (window_ptr != window_end) {
2144 lz_match = lzx_choose_item(c);
2146 LZX_ASSERT(!(lz_match.len == LZX_MIN_MATCH_LEN &&
2147 lz_match.offset == c->max_window_size -
2148 LZX_MIN_MATCH_LEN));
2149 if (lz_match.len >= LZX_MIN_MATCH_LEN) {
2150 lzx_item.data = lzx_tally_match(lz_match.len,
2153 window_ptr += lz_match.len;
2155 lzx_item.data = lzx_tally_literal(*window_ptr, &freqs);
2158 *next_chosen_item++ = lzx_item;
2160 /* When doing one-pass "near-optimal" parsing, update the cost
2161 * model occassionally. */
2162 if (unlikely((next_chosen_item - spec->chosen_items) % 2048 == 0) &&
2163 c->params.choose_item_func == lzx_choose_near_optimal_item &&
2164 c->params.num_optim_passes == 1)
2166 lzx_make_huffman_codes(&freqs, &spec->codes, c->num_main_syms);
2167 lzx_set_costs(c, &spec->codes.lens, unseen_cost);
2168 if (unseen_cost < 15)
2172 spec->num_chosen_items = next_chosen_item - spec->chosen_items;
2173 lzx_make_huffman_codes(&freqs, &spec->codes, c->num_main_syms);
2174 spec->block_type = lzx_choose_verbatim_or_aligned(&freqs, &spec->codes);
2177 /* Prepare the input window into one or more LZX blocks ready to be output. */
2179 lzx_prepare_blocks(struct lzx_compressor *c)
2181 /* Set up a default cost model. */
2182 if (c->params.choose_item_func == lzx_choose_near_optimal_item)
2183 lzx_set_default_costs(&c->costs, c->num_main_syms);
2185 /* Set up the block specifications.
2186 * TODO: The compression ratio could be slightly improved by performing
2187 * data-dependent block splitting instead of using fixed-size blocks.
2188 * Doing so well is a computationally hard problem, however. */
2189 c->num_blocks = DIV_ROUND_UP(c->cur_window_size, LZX_DIV_BLOCK_SIZE);
2190 for (unsigned i = 0; i < c->num_blocks; i++) {
2191 u32 pos = LZX_DIV_BLOCK_SIZE * i;
2192 c->block_specs[i].window_pos = pos;
2193 c->block_specs[i].block_size = min(c->cur_window_size - pos,
2194 LZX_DIV_BLOCK_SIZE);
2197 /* Load the window into the match-finder. */
2198 lz_mf_load_window(c->mf, c->cur_window, c->cur_window_size);
2200 /* Determine sequence of matches/literals to output for each block. */
2201 lzx_lru_queue_init(&c->queue);
2202 c->optimum_cur_idx = 0;
2203 c->optimum_end_idx = 0;
2204 c->prev_match.len = 0;
2205 for (unsigned i = 0; i < c->num_blocks; i++)
2206 lzx_choose_items_for_block(c, &c->block_specs[i]);
2210 lzx_build_params(unsigned int compression_level,
2211 u32 max_window_size,
2212 struct lzx_compressor_params *lzx_params)
2214 if (compression_level < 25) {
2215 lzx_params->choose_item_func = lzx_choose_lazy_item;
2216 lzx_params->num_optim_passes = 1;
2217 if (max_window_size <= 262144)
2218 lzx_params->mf_algo = LZ_MF_HASH_CHAINS;
2220 lzx_params->mf_algo = LZ_MF_BINARY_TREES;
2221 lzx_params->min_match_length = 3;
2222 lzx_params->nice_match_length = 25 + compression_level * 2;
2223 lzx_params->max_search_depth = 25 + compression_level;
2225 lzx_params->choose_item_func = lzx_choose_near_optimal_item;
2226 lzx_params->num_optim_passes = compression_level / 20;
2227 if (max_window_size <= 32768 && lzx_params->num_optim_passes == 1)
2228 lzx_params->mf_algo = LZ_MF_HASH_CHAINS;
2230 lzx_params->mf_algo = LZ_MF_BINARY_TREES;
2231 lzx_params->min_match_length = (compression_level >= 45) ? 2 : 3;
2232 lzx_params->nice_match_length = min(((u64)compression_level * 32) / 50,
2234 lzx_params->max_search_depth = min(((u64)compression_level * 50) / 50,
2240 lzx_build_mf_params(const struct lzx_compressor_params *lzx_params,
2241 u32 max_window_size, struct lz_mf_params *mf_params)
2243 memset(mf_params, 0, sizeof(*mf_params));
2245 mf_params->algorithm = lzx_params->mf_algo;
2246 mf_params->max_window_size = max_window_size;
2247 mf_params->min_match_len = lzx_params->min_match_length;
2248 mf_params->max_match_len = LZX_MAX_MATCH_LEN;
2249 mf_params->max_search_depth = lzx_params->max_search_depth;
2250 mf_params->nice_match_len = lzx_params->nice_match_length;
2254 lzx_free_compressor(void *_c);
2257 lzx_get_needed_memory(size_t max_window_size, unsigned int compression_level)
2259 struct lzx_compressor_params params;
2262 if (!lzx_window_size_valid(max_window_size))
2265 lzx_build_params(compression_level, max_window_size, ¶ms);
2267 size += sizeof(struct lzx_compressor);
2269 size += max_window_size;
2271 size += DIV_ROUND_UP(max_window_size, LZX_DIV_BLOCK_SIZE) *
2272 sizeof(struct lzx_block_spec);
2274 size += max_window_size * sizeof(struct lzx_item);
2276 size += lz_mf_get_needed_memory(params.mf_algo, max_window_size);
2277 if (params.choose_item_func == lzx_choose_near_optimal_item) {
2278 size += (LZX_OPTIM_ARRAY_LENGTH + params.nice_match_length) *
2279 sizeof(struct lzx_mc_pos_data);
2281 if (params.num_optim_passes > 1)
2282 size += LZX_CACHE_LEN * sizeof(struct lz_match);
2284 size += LZX_MAX_MATCHES_PER_POS * sizeof(struct lz_match);
2289 lzx_create_compressor(size_t max_window_size, unsigned int compression_level,
2292 struct lzx_compressor *c;
2293 struct lzx_compressor_params params;
2294 struct lz_mf_params mf_params;
2296 if (!lzx_window_size_valid(max_window_size))
2297 return WIMLIB_ERR_INVALID_PARAM;
2299 lzx_build_params(compression_level, max_window_size, ¶ms);
2300 lzx_build_mf_params(¶ms, max_window_size, &mf_params);
2301 if (!lz_mf_params_valid(&mf_params))
2302 return WIMLIB_ERR_INVALID_PARAM;
2304 c = CALLOC(1, sizeof(struct lzx_compressor));
2309 c->num_main_syms = lzx_get_num_main_syms(max_window_size);
2310 c->max_window_size = max_window_size;
2312 c->cur_window = ALIGNED_MALLOC(max_window_size, 16);
2316 c->block_specs = MALLOC(DIV_ROUND_UP(max_window_size,
2317 LZX_DIV_BLOCK_SIZE) *
2318 sizeof(struct lzx_block_spec));
2319 if (!c->block_specs)
2322 c->chosen_items = MALLOC(max_window_size * sizeof(struct lzx_item));
2323 if (!c->chosen_items)
2326 c->mf = lz_mf_alloc(&mf_params);
2330 if (params.choose_item_func == lzx_choose_near_optimal_item) {
2331 c->optimum = MALLOC((LZX_OPTIM_ARRAY_LENGTH +
2332 params.nice_match_length) *
2333 sizeof(struct lzx_mc_pos_data));
2338 if (params.num_optim_passes > 1) {
2339 c->cached_matches = MALLOC(LZX_CACHE_LEN *
2340 sizeof(struct lz_match));
2341 if (!c->cached_matches)
2343 c->cache_limit = c->cached_matches + LZX_CACHE_LEN -
2344 (LZX_MAX_MATCHES_PER_POS + 1);
2346 c->cached_matches = MALLOC(LZX_MAX_MATCHES_PER_POS *
2347 sizeof(struct lz_match));
2348 if (!c->cached_matches)
2356 lzx_free_compressor(c);
2357 return WIMLIB_ERR_NOMEM;
2361 lzx_compress(const void *uncompressed_data, size_t uncompressed_size,
2362 void *compressed_data, size_t compressed_size_avail, void *_c)
2364 struct lzx_compressor *c = _c;
2365 struct lzx_output_bitstream os;
2367 /* Don't bother compressing very small inputs. */
2368 if (uncompressed_size < 100)
2371 /* The input data must be preprocessed. To avoid changing the original
2372 * input, copy it to a temporary buffer. */
2373 memcpy(c->cur_window, uncompressed_data, uncompressed_size);
2374 c->cur_window_size = uncompressed_size;
2376 /* Preprocess the data. */
2377 lzx_do_e8_preprocessing(c->cur_window, c->cur_window_size);
2379 /* Prepare the compressed data. */
2380 lzx_prepare_blocks(c);
2382 /* Generate the compressed data and return its size, or 0 if an overflow
2384 lzx_init_output(&os, compressed_data, compressed_size_avail);
2385 lzx_write_all_blocks(c, &os);
2386 return lzx_flush_output(&os);
2390 lzx_free_compressor(void *_c)
2392 struct lzx_compressor *c = _c;
2395 ALIGNED_FREE(c->cur_window);
2396 FREE(c->block_specs);
2397 FREE(c->chosen_items);
2400 FREE(c->cached_matches);
2405 const struct compressor_ops lzx_compressor_ops = {
2406 .get_needed_memory = lzx_get_needed_memory,
2407 .create_compressor = lzx_create_compressor,
2408 .compress = lzx_compress,
2409 .free_compressor = lzx_free_compressor,