4 * LZX compression routines
8 * Copyright (C) 2012, 2013 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 compression format, as used in
29 * the WIM file format.
34 * First, the primary reference for the LZX compression format is the
35 * specification released by Microsoft.
37 * Second, the comments in lzx-decompress.c provide some more information about
38 * the LZX compression format, including errors in the Microsoft specification.
40 * Do note that LZX shares many similarities with DEFLATE, the algorithm used by
41 * zlib and gzip. Both LZX and DEFLATE use LZ77 matching and Huffman coding,
42 * and certain other details are quite similar, such as the method for storing
43 * Huffman codes. However, some of the main differences are:
45 * - LZX preprocesses the data to attempt to make x86 machine code slightly more
46 * compressible before attempting to compress it further.
47 * - LZX uses a "main" alphabet which combines literals and matches, with the
48 * match symbols containing a "length header" (giving all or part of the match
49 * length) and a "position slot" (giving, roughly speaking, the order of
50 * magnitude of the match offset).
51 * - LZX does not have static Huffman blocks; however it does have two types of
52 * dynamic Huffman blocks ("aligned offset" and "verbatim").
53 * - LZX has a minimum match length of 2 rather than 3.
54 * - In LZX, match offsets 0 through 2 actually represent entries in an LRU
55 * queue of match offsets. This is very useful for certain types of files,
56 * such as binary files that have repeating records.
61 * There are actually two distinct overall algorithms implemented here. We
62 * shall refer to them as the "slow" algorithm and the "fast" algorithm. The
63 * "slow" algorithm spends more time compressing to achieve a higher compression
64 * ratio compared to the "fast" algorithm. More details are presented below.
69 * The "slow" algorithm to generate LZX-compressed data is roughly as follows:
71 * 1. Preprocess the input data to translate the targets of x86 call
72 * instructions to absolute offsets.
74 * 2. Build the suffix array and inverse suffix array for the input data. The
75 * suffix array contains the indices of all suffixes of the input data,
76 * sorted lexcographically by the corresponding suffixes. The "position" of
77 * a suffix is the index of that suffix in the original string, whereas the
78 * "rank" of a suffix is the index at which that suffix's position is found
79 * in the suffix array.
81 * 3. Build the longest common prefix array corresponding to the suffix array.
83 * 4. For each suffix, find the highest lower ranked suffix that has a lower
84 * position, the lowest higher ranked suffix that has a lower position, and
85 * the length of the common prefix shared between each. This information is
86 * later used to link suffix ranks into a doubly-linked list for searching
89 * 5. Set a default cost model for matches/literals.
91 * 6. Determine the lowest cost sequence of LZ77 matches ((offset, length)
92 * pairs) and literal bytes to divide the input into. Raw match-finding is
93 * done by searching the suffix array using a linked list to avoid
94 * considering any suffixes that start after the current position. Each run
95 * of the match-finder returns the approximate lowest-cost longest match as
96 * well as any shorter matches that have even lower approximate costs. Each
97 * such run also adds the suffix rank of the current position into the linked
98 * list being used to search the suffix array. Parsing, or match-choosing,
99 * is solved as a minimum-cost path problem using a forward "optimal parsing"
100 * algorithm based on the Deflate encoder from 7-Zip. This algorithm moves
101 * forward calculating the minimum cost to reach each byte until either a
102 * very long match is found or until a position is found at which no matches
105 * 7. Build the Huffman codes needed to output the matches/literals.
107 * 8. Up to a certain number of iterations, use the resulting Huffman codes to
108 * refine a cost model and go back to Step #6 to determine an improved
109 * sequence of matches and literals.
111 * 9. Output the resulting block using the match/literal sequences and the
112 * Huffman codes that were computed for the block.
114 * Note: the algorithm does not yet attempt to split the input into multiple LZX
115 * blocks, instead using a series of blocks of LZX_DIV_BLOCK_SIZE bytes.
120 * The fast algorithm (and the only one available in wimlib v1.5.1 and earlier)
121 * spends much less time on the main bottlenecks of the compression process ---
122 * that is, the match finding and match choosing. Matches are found and chosen
123 * with hash chains using a greedy parse with one position of look-ahead. No
124 * block splitting is done; only compressing the full input into an aligned
125 * offset block is considered.
130 * Acknowledgments to several open-source projects and research papers that made
131 * it possible to implement this code:
133 * - divsufsort (author: Yuta Mori), for the suffix array construction code,
134 * located in a separate directory (divsufsort/).
136 * - "Linear-Time Longest-Common-Prefix Computation in Suffix Arrays and Its
137 * Applications" (Kasai et al. 2001), for the LCP array computation.
139 * - "LPF computation revisited" (Crochemore et al. 2009) for the prev and next
140 * array computations.
142 * - 7-Zip (author: Igor Pavlov) for the algorithm for forward optimal parsing
145 * - zlib (author: Jean-loup Gailly and Mark Adler), for the hash table
146 * match-finding algorithm (used in lz77.c).
148 * - lzx-compress (author: Matthew T. Russotto), on which some parts of this
149 * code were originally based.
157 #include "wimlib/compressor_ops.h"
158 #include "wimlib/compress_common.h"
159 #include "wimlib/endianness.h"
160 #include "wimlib/error.h"
161 #include "wimlib/lz_hash.h"
162 #include "wimlib/lz_sarray.h"
163 #include "wimlib/lzx.h"
164 #include "wimlib/util.h"
169 #ifdef ENABLE_LZX_DEBUG
170 # include "wimlib/decompress_common.h"
173 typedef u32 block_cost_t;
174 #define INFINITE_BLOCK_COST (~(block_cost_t)0)
176 #define LZX_OPTIM_ARRAY_SIZE 4096
178 #define LZX_DIV_BLOCK_SIZE 32768
180 #define LZX_MAX_CACHE_PER_POS 10
182 /* Codewords for the LZX main, length, and aligned offset Huffman codes */
183 struct lzx_codewords {
184 u16 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
185 u16 len[LZX_LENCODE_NUM_SYMBOLS];
186 u16 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
189 /* Codeword lengths (in bits) for the LZX main, length, and aligned offset
192 * A 0 length means the codeword has zero frequency.
195 u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
196 u8 len[LZX_LENCODE_NUM_SYMBOLS];
197 u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
200 /* Costs for the LZX main, length, and aligned offset Huffman symbols.
202 * If a codeword has zero frequency, it must still be assigned some nonzero cost
203 * --- generally a high cost, since even if it gets used in the next iteration,
204 * it probably will not be used very times. */
206 u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
207 u8 len[LZX_LENCODE_NUM_SYMBOLS];
208 u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
211 /* The LZX main, length, and aligned offset Huffman codes */
213 struct lzx_codewords codewords;
214 struct lzx_lens lens;
217 /* Tables for tallying symbol frequencies in the three LZX alphabets */
219 input_idx_t main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
220 input_idx_t len[LZX_LENCODE_NUM_SYMBOLS];
221 input_idx_t aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
224 /* LZX intermediate match/literal format */
228 * 31 1 if a match, 0 if a literal.
230 * 30-25 position slot. This can be at most 50, so it will fit in 6
233 * 8-24 position footer. This is the offset of the real formatted
234 * offset from the position base. This can be at most 17 bits
235 * (since lzx_extra_bits[LZX_MAX_POSITION_SLOTS - 1] is 17).
237 * 0-7 length of match, minus 2. This can be at most
238 * (LZX_MAX_MATCH_LEN - 2) == 255, so it will fit in 8 bits. */
242 /* Specification for an LZX block. */
243 struct lzx_block_spec {
245 /* One of the LZX_BLOCKTYPE_* constants indicating which type of this
249 /* 0-based position in the window at which this block starts. */
250 input_idx_t window_pos;
252 /* The number of bytes of uncompressed data this block represents. */
253 input_idx_t block_size;
255 /* The position in the 'chosen_matches' array in the `struct
256 * lzx_compressor' at which the match/literal specifications for
257 * this block begin. */
258 input_idx_t chosen_matches_start_pos;
260 /* The number of match/literal specifications for this block. */
261 input_idx_t num_chosen_matches;
263 /* Huffman codes for this block. */
264 struct lzx_codes codes;
267 /* Include template for the match-choosing algorithm. */
268 #define LZ_COMPRESSOR struct lzx_compressor
269 #define LZ_ADAPTIVE_STATE struct lzx_lru_queue
270 struct lzx_compressor;
271 #include "wimlib/lz_optimal.h"
273 /* State of the LZX compressor. */
274 struct lzx_compressor {
276 /* The parameters that were used to create the compressor. */
277 struct wimlib_lzx_compressor_params params;
279 /* The buffer of data to be compressed.
281 * 0xe8 byte preprocessing is done directly on the data here before
282 * further compression.
284 * Note that this compressor does *not* use a real sliding window!!!!
285 * It's not needed in the WIM format, since every chunk is compressed
286 * independently. This is by design, to allow random access to the
289 * We reserve a few extra bytes to potentially allow reading off the end
290 * of the array in the match-finding code for optimization purposes.
294 /* Number of bytes of data to be compressed, which is the number of
295 * bytes of data in @window that are actually valid. */
296 input_idx_t window_size;
298 /* Allocated size of the @window. */
299 input_idx_t max_window_size;
301 /* Number of symbols in the main alphabet (depends on the
302 * @max_window_size since it determines the maximum allowed offset). */
303 unsigned num_main_syms;
305 /* The current match offset LRU queue. */
306 struct lzx_lru_queue queue;
308 /* Space for the sequences of matches/literals that were chosen for each
310 struct lzx_match *chosen_matches;
312 /* Information about the LZX blocks the preprocessed input was divided
314 struct lzx_block_spec *block_specs;
316 /* Number of LZX blocks the input was divided into; a.k.a. the number of
317 * elements of @block_specs that are valid. */
320 /* This is simply filled in with zeroes and used to avoid special-casing
321 * the output of the first compressed Huffman code, which conceptually
322 * has a delta taken from a code with all symbols having zero-length
324 struct lzx_codes zero_codes;
326 /* The current cost model. */
327 struct lzx_costs costs;
329 /* Fast algorithm only: Array of hash table links. */
330 input_idx_t *prev_tab;
332 /* Slow algorithm only: Suffix array match-finder. */
333 struct lz_sarray lz_sarray;
335 /* Position in window of next match to return. */
336 input_idx_t match_window_pos;
338 /* The match-finder shall ensure the length of matches does not exceed
339 * this position in the input. */
340 input_idx_t match_window_end;
342 /* Matches found by the match-finder are cached in the following array
343 * to achieve a slight speedup when the same matches are needed on
344 * subsequent passes. This is suboptimal because different matches may
345 * be preferred with different cost models, but seems to be a worthwhile
347 struct raw_match *cached_matches;
348 unsigned cached_matches_pos;
352 struct lz_match_chooser mc;
355 /* Returns the LZX position slot that corresponds to a given match offset,
356 * taking into account the recent offset queue and updating it if the offset is
359 lzx_get_position_slot(unsigned offset, struct lzx_lru_queue *queue)
361 unsigned position_slot;
363 /* See if the offset was recently used. */
364 for (unsigned i = 0; i < LZX_NUM_RECENT_OFFSETS; i++) {
365 if (offset == queue->R[i]) {
368 /* Bring the repeat offset to the front of the
369 * queue. Note: this is, in fact, not a real
370 * LRU queue because repeat matches are simply
371 * swapped to the front. */
372 swap(queue->R[0], queue->R[i]);
374 /* The resulting position slot is simply the first index
375 * at which the offset was found in the queue. */
380 /* The offset was not recently used; look up its real position slot. */
381 position_slot = lzx_get_position_slot_raw(offset + LZX_OFFSET_OFFSET);
383 /* Bring the new offset to the front of the queue. */
384 for (unsigned i = LZX_NUM_RECENT_OFFSETS - 1; i > 0; i--)
385 queue->R[i] = queue->R[i - 1];
386 queue->R[0] = offset;
388 return position_slot;
391 /* Build the main, length, and aligned offset Huffman codes used in LZX.
393 * This takes as input the frequency tables for each code and produces as output
394 * a set of tables that map symbols to codewords and codeword lengths. */
396 lzx_make_huffman_codes(const struct lzx_freqs *freqs,
397 struct lzx_codes *codes,
398 unsigned num_main_syms)
400 make_canonical_huffman_code(num_main_syms,
401 LZX_MAX_MAIN_CODEWORD_LEN,
404 codes->codewords.main);
406 make_canonical_huffman_code(LZX_LENCODE_NUM_SYMBOLS,
407 LZX_MAX_LEN_CODEWORD_LEN,
410 codes->codewords.len);
412 make_canonical_huffman_code(LZX_ALIGNEDCODE_NUM_SYMBOLS,
413 LZX_MAX_ALIGNED_CODEWORD_LEN,
416 codes->codewords.aligned);
420 * Output an LZX match.
422 * @out: The bitstream to write the match to.
423 * @block_type: The type of the LZX block (LZX_BLOCKTYPE_ALIGNED or LZX_BLOCKTYPE_VERBATIM)
425 * @codes: Pointer to a structure that contains the codewords for the
426 * main, length, and aligned offset Huffman codes.
429 lzx_write_match(struct output_bitstream *out, int block_type,
430 struct lzx_match match, const struct lzx_codes *codes)
432 /* low 8 bits are the match length minus 2 */
433 unsigned match_len_minus_2 = match.data & 0xff;
434 /* Next 17 bits are the position footer */
435 unsigned position_footer = (match.data >> 8) & 0x1ffff; /* 17 bits */
436 /* Next 6 bits are the position slot. */
437 unsigned position_slot = (match.data >> 25) & 0x3f; /* 6 bits */
440 unsigned main_symbol;
441 unsigned num_extra_bits;
442 unsigned verbatim_bits;
443 unsigned aligned_bits;
445 /* If the match length is less than MIN_MATCH_LEN (= 2) +
446 * NUM_PRIMARY_LENS (= 7), the length header contains
447 * the match length minus MIN_MATCH_LEN, and there is no
450 * Otherwise, the length header contains
451 * NUM_PRIMARY_LENS, and the length footer contains
452 * the match length minus NUM_PRIMARY_LENS minus
454 if (match_len_minus_2 < LZX_NUM_PRIMARY_LENS) {
455 len_header = match_len_minus_2;
456 /* No length footer-- mark it with a special
458 len_footer = (unsigned)(-1);
460 len_header = LZX_NUM_PRIMARY_LENS;
461 len_footer = match_len_minus_2 - LZX_NUM_PRIMARY_LENS;
464 /* Combine the position slot with the length header into a single symbol
465 * that will be encoded with the main code.
467 * The actual main symbol is offset by LZX_NUM_CHARS because values
468 * under LZX_NUM_CHARS are used to indicate a literal byte rather than a
470 main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS;
472 /* Output main symbol. */
473 bitstream_put_bits(out, codes->codewords.main[main_symbol],
474 codes->lens.main[main_symbol]);
476 /* If there is a length footer, output it using the
477 * length Huffman code. */
478 if (len_footer != (unsigned)(-1)) {
479 bitstream_put_bits(out, codes->codewords.len[len_footer],
480 codes->lens.len[len_footer]);
483 num_extra_bits = lzx_get_num_extra_bits(position_slot);
485 /* For aligned offset blocks with at least 3 extra bits, output the
486 * verbatim bits literally, then the aligned bits encoded using the
487 * aligned offset code. Otherwise, only the verbatim bits need to be
489 if ((block_type == LZX_BLOCKTYPE_ALIGNED) && (num_extra_bits >= 3)) {
491 verbatim_bits = position_footer >> 3;
492 bitstream_put_bits(out, verbatim_bits,
495 aligned_bits = (position_footer & 7);
496 bitstream_put_bits(out,
497 codes->codewords.aligned[aligned_bits],
498 codes->lens.aligned[aligned_bits]);
500 /* verbatim bits is the same as the position
501 * footer, in this case. */
502 bitstream_put_bits(out, position_footer, num_extra_bits);
507 lzx_build_precode(const u8 lens[restrict],
508 const u8 prev_lens[restrict],
509 const unsigned num_syms,
510 input_idx_t precode_freqs[restrict LZX_PRECODE_NUM_SYMBOLS],
511 u8 output_syms[restrict num_syms],
512 u8 precode_lens[restrict LZX_PRECODE_NUM_SYMBOLS],
513 u16 precode_codewords[restrict LZX_PRECODE_NUM_SYMBOLS],
514 unsigned *num_additional_bits_ret)
516 memset(precode_freqs, 0,
517 LZX_PRECODE_NUM_SYMBOLS * sizeof(precode_freqs[0]));
519 /* Since the code word lengths use a form of RLE encoding, the goal here
520 * is to find each run of identical lengths when going through them in
521 * symbol order (including runs of length 1). For each run, as many
522 * lengths are encoded using RLE as possible, and the rest are output
525 * output_syms[] will be filled in with the length symbols that will be
526 * output, including RLE codes, not yet encoded using the precode.
528 * cur_run_len keeps track of how many code word lengths are in the
529 * current run of identical lengths. */
530 unsigned output_syms_idx = 0;
531 unsigned cur_run_len = 1;
532 unsigned num_additional_bits = 0;
533 for (unsigned i = 1; i <= num_syms; i++) {
535 if (i != num_syms && lens[i] == lens[i - 1]) {
536 /* Still in a run--- keep going. */
541 /* Run ended! Check if it is a run of zeroes or a run of
544 /* The symbol that was repeated in the run--- not to be confused
545 * with the length *of* the run (cur_run_len) */
546 unsigned len_in_run = lens[i - 1];
548 if (len_in_run == 0) {
549 /* A run of 0's. Encode it in as few length
550 * codes as we can. */
552 /* The magic length 18 indicates a run of 20 + n zeroes,
553 * where n is an uncompressed literal 5-bit integer that
554 * follows the magic length. */
555 while (cur_run_len >= 20) {
556 unsigned additional_bits;
558 additional_bits = min(cur_run_len - 20, 0x1f);
559 num_additional_bits += 5;
561 output_syms[output_syms_idx++] = 18;
562 output_syms[output_syms_idx++] = additional_bits;
563 cur_run_len -= 20 + additional_bits;
566 /* The magic length 17 indicates a run of 4 + n zeroes,
567 * where n is an uncompressed literal 4-bit integer that
568 * follows the magic length. */
569 while (cur_run_len >= 4) {
570 unsigned additional_bits;
572 additional_bits = min(cur_run_len - 4, 0xf);
573 num_additional_bits += 4;
575 output_syms[output_syms_idx++] = 17;
576 output_syms[output_syms_idx++] = additional_bits;
577 cur_run_len -= 4 + additional_bits;
582 /* A run of nonzero lengths. */
584 /* The magic length 19 indicates a run of 4 + n
585 * nonzeroes, where n is a literal bit that follows the
586 * magic length, and where the value of the lengths in
587 * the run is given by an extra length symbol, encoded
588 * with the precode, that follows the literal bit.
590 * The extra length symbol is encoded as a difference
591 * from the length of the codeword for the first symbol
592 * in the run in the previous code.
594 while (cur_run_len >= 4) {
595 unsigned additional_bits;
598 additional_bits = (cur_run_len > 4);
599 num_additional_bits += 1;
600 delta = (signed char)prev_lens[i - cur_run_len] -
601 (signed char)len_in_run;
605 precode_freqs[(unsigned char)delta]++;
606 output_syms[output_syms_idx++] = 19;
607 output_syms[output_syms_idx++] = additional_bits;
608 output_syms[output_syms_idx++] = delta;
609 cur_run_len -= 4 + additional_bits;
613 /* Any remaining lengths in the run are outputted without RLE,
614 * as a difference from the length of that codeword in the
616 while (cur_run_len > 0) {
619 delta = (signed char)prev_lens[i - cur_run_len] -
620 (signed char)len_in_run;
624 precode_freqs[(unsigned char)delta]++;
625 output_syms[output_syms_idx++] = delta;
632 /* Build the precode from the frequencies of the length symbols. */
634 make_canonical_huffman_code(LZX_PRECODE_NUM_SYMBOLS,
635 LZX_MAX_PRE_CODEWORD_LEN,
636 precode_freqs, precode_lens,
639 *num_additional_bits_ret = num_additional_bits;
641 return output_syms_idx;
645 * Writes a compressed Huffman code to the output, preceded by the precode for
648 * The Huffman code is represented in the output as a series of path lengths
649 * from which the canonical Huffman code can be reconstructed. The path lengths
650 * themselves are compressed using a separate Huffman code, the precode, which
651 * consists of LZX_PRECODE_NUM_SYMBOLS (= 20) symbols that cover all possible
652 * code lengths, plus extra codes for repeated lengths. The path lengths of the
653 * precode precede the path lengths of the larger code and are uncompressed,
654 * consisting of 20 entries of 4 bits each.
656 * @out: Bitstream to write the code to.
657 * @lens: The code lengths for the Huffman code, indexed by symbol.
658 * @prev_lens: Code lengths for this Huffman code, indexed by symbol,
659 * in the *previous block*, or all zeroes if this is the
661 * @num_syms: The number of symbols in the code.
664 lzx_write_compressed_code(struct output_bitstream *out,
665 const u8 lens[restrict],
666 const u8 prev_lens[restrict],
669 input_idx_t precode_freqs[LZX_PRECODE_NUM_SYMBOLS];
670 u8 output_syms[num_syms];
671 u8 precode_lens[LZX_PRECODE_NUM_SYMBOLS];
672 u16 precode_codewords[LZX_PRECODE_NUM_SYMBOLS];
674 unsigned num_output_syms;
678 num_output_syms = lzx_build_precode(lens,
687 /* Write the lengths of the precode codes to the output. */
688 for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++)
689 bitstream_put_bits(out, precode_lens[i],
690 LZX_PRECODE_ELEMENT_SIZE);
692 /* Write the length symbols, encoded with the precode, to the output. */
694 for (i = 0; i < num_output_syms; ) {
695 precode_sym = output_syms[i++];
697 bitstream_put_bits(out, precode_codewords[precode_sym],
698 precode_lens[precode_sym]);
699 switch (precode_sym) {
701 bitstream_put_bits(out, output_syms[i++], 4);
704 bitstream_put_bits(out, output_syms[i++], 5);
707 bitstream_put_bits(out, output_syms[i++], 1);
708 bitstream_put_bits(out,
709 precode_codewords[output_syms[i]],
710 precode_lens[output_syms[i]]);
720 * Writes all compressed matches and literal bytes in an LZX block to the the
724 * The output bitstream.
726 * The type of the block (LZX_BLOCKTYPE_ALIGNED or LZX_BLOCKTYPE_VERBATIM).
728 * The array of matches/literals that will be output (length @match_count).
730 * Number of matches/literals to be output.
732 * Pointer to a structure that contains the codewords for the main, length,
733 * and aligned offset Huffman codes.
736 lzx_write_matches_and_literals(struct output_bitstream *ostream,
738 const struct lzx_match match_tab[],
739 unsigned match_count,
740 const struct lzx_codes *codes)
742 for (unsigned i = 0; i < match_count; i++) {
743 struct lzx_match match = match_tab[i];
745 /* High bit of the match indicates whether the match is an
746 * actual match (1) or a literal uncompressed byte (0) */
747 if (match.data & 0x80000000) {
749 lzx_write_match(ostream, block_type,
753 bitstream_put_bits(ostream,
754 codes->codewords.main[match.data],
755 codes->lens.main[match.data]);
761 lzx_assert_codes_valid(const struct lzx_codes * codes, unsigned num_main_syms)
763 #ifdef ENABLE_LZX_DEBUG
766 for (i = 0; i < num_main_syms; i++)
767 LZX_ASSERT(codes->lens.main[i] <= LZX_MAX_MAIN_CODEWORD_LEN);
769 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
770 LZX_ASSERT(codes->lens.len[i] <= LZX_MAX_LEN_CODEWORD_LEN);
772 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
773 LZX_ASSERT(codes->lens.aligned[i] <= LZX_MAX_ALIGNED_CODEWORD_LEN);
775 const unsigned tablebits = 10;
776 u16 decode_table[(1 << tablebits) +
777 (2 * max(num_main_syms, LZX_LENCODE_NUM_SYMBOLS))]
778 _aligned_attribute(DECODE_TABLE_ALIGNMENT);
779 LZX_ASSERT(0 == make_huffman_decode_table(decode_table,
781 min(tablebits, LZX_MAINCODE_TABLEBITS),
783 LZX_MAX_MAIN_CODEWORD_LEN));
784 LZX_ASSERT(0 == make_huffman_decode_table(decode_table,
785 LZX_LENCODE_NUM_SYMBOLS,
786 min(tablebits, LZX_LENCODE_TABLEBITS),
788 LZX_MAX_LEN_CODEWORD_LEN));
789 LZX_ASSERT(0 == make_huffman_decode_table(decode_table,
790 LZX_ALIGNEDCODE_NUM_SYMBOLS,
791 min(tablebits, LZX_ALIGNEDCODE_TABLEBITS),
793 LZX_MAX_ALIGNED_CODEWORD_LEN));
794 #endif /* ENABLE_LZX_DEBUG */
797 /* Write an LZX aligned offset or verbatim block to the output. */
799 lzx_write_compressed_block(int block_type,
801 unsigned max_window_size,
802 unsigned num_main_syms,
803 struct lzx_match * chosen_matches,
804 unsigned num_chosen_matches,
805 const struct lzx_codes * codes,
806 const struct lzx_codes * prev_codes,
807 struct output_bitstream * ostream)
811 LZX_ASSERT(block_type == LZX_BLOCKTYPE_ALIGNED ||
812 block_type == LZX_BLOCKTYPE_VERBATIM);
813 lzx_assert_codes_valid(codes, num_main_syms);
815 /* The first three bits indicate the type of block and are one of the
816 * LZX_BLOCKTYPE_* constants. */
817 bitstream_put_bits(ostream, block_type, 3);
819 /* Output the block size.
821 * The original LZX format seemed to always encode the block size in 3
822 * bytes. However, the implementation in WIMGAPI, as used in WIM files,
823 * uses the first bit to indicate whether the block is the default size
824 * (32768) or a different size given explicitly by the next 16 bits.
826 * By default, this compressor uses a window size of 32768 and therefore
827 * follows the WIMGAPI behavior. However, this compressor also supports
828 * window sizes greater than 32768 bytes, which do not appear to be
829 * supported by WIMGAPI. In such cases, we retain the default size bit
830 * to mean a size of 32768 bytes but output non-default block size in 24
831 * bits rather than 16. The compatibility of this behavior is unknown
832 * because WIMs created with chunk size greater than 32768 can seemingly
833 * only be opened by wimlib anyway. */
834 if (block_size == LZX_DEFAULT_BLOCK_SIZE) {
835 bitstream_put_bits(ostream, 1, 1);
837 bitstream_put_bits(ostream, 0, 1);
839 if (max_window_size >= 65536)
840 bitstream_put_bits(ostream, block_size >> 16, 8);
842 bitstream_put_bits(ostream, block_size, 16);
845 /* Write out lengths of the main code. Note that the LZX specification
846 * incorrectly states that the aligned offset code comes after the
847 * length code, but in fact it is the very first code to be written
848 * (before the main code). */
849 if (block_type == LZX_BLOCKTYPE_ALIGNED)
850 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
851 bitstream_put_bits(ostream, codes->lens.aligned[i],
852 LZX_ALIGNEDCODE_ELEMENT_SIZE);
854 LZX_DEBUG("Writing main code...");
856 /* Write the precode and lengths for the first LZX_NUM_CHARS symbols in
857 * the main code, which are the codewords for literal bytes. */
858 lzx_write_compressed_code(ostream,
860 prev_codes->lens.main,
863 /* Write the precode and lengths for the rest of the main code, which
864 * are the codewords for match headers. */
865 lzx_write_compressed_code(ostream,
866 codes->lens.main + LZX_NUM_CHARS,
867 prev_codes->lens.main + LZX_NUM_CHARS,
868 num_main_syms - LZX_NUM_CHARS);
870 LZX_DEBUG("Writing length code...");
872 /* Write the precode and lengths for the length code. */
873 lzx_write_compressed_code(ostream,
875 prev_codes->lens.len,
876 LZX_LENCODE_NUM_SYMBOLS);
878 LZX_DEBUG("Writing matches and literals...");
880 /* Write the actual matches and literals. */
881 lzx_write_matches_and_literals(ostream, block_type,
882 chosen_matches, num_chosen_matches,
885 LZX_DEBUG("Done writing block.");
888 /* Write out the LZX blocks that were computed. */
890 lzx_write_all_blocks(struct lzx_compressor *ctx, struct output_bitstream *ostream)
893 const struct lzx_codes *prev_codes = &ctx->zero_codes;
894 for (unsigned i = 0; i < ctx->num_blocks; i++) {
895 const struct lzx_block_spec *spec = &ctx->block_specs[i];
897 LZX_DEBUG("Writing block %u/%u (type=%d, size=%u, num_chosen_matches=%u)...",
898 i + 1, ctx->num_blocks,
899 spec->block_type, spec->block_size,
900 spec->num_chosen_matches);
902 lzx_write_compressed_block(spec->block_type,
904 ctx->max_window_size,
906 &ctx->chosen_matches[spec->chosen_matches_start_pos],
907 spec->num_chosen_matches,
912 prev_codes = &spec->codes;
916 /* Constructs an LZX match from a literal byte and updates the main code symbol
919 lzx_tally_literal(u8 lit, struct lzx_freqs *freqs)
925 /* Constructs an LZX match from an offset and a length, and updates the LRU
926 * queue and the frequency of symbols in the main, length, and aligned offset
927 * alphabets. The return value is a 32-bit number that provides the match in an
928 * intermediate representation documented below. */
930 lzx_tally_match(unsigned match_len, unsigned match_offset,
931 struct lzx_freqs *freqs, struct lzx_lru_queue *queue)
933 unsigned position_slot;
934 unsigned position_footer;
936 unsigned main_symbol;
938 unsigned adjusted_match_len;
940 LZX_ASSERT(match_len >= LZX_MIN_MATCH_LEN && match_len <= LZX_MAX_MATCH_LEN);
942 /* The match offset shall be encoded as a position slot (itself encoded
943 * as part of the main symbol) and a position footer. */
944 position_slot = lzx_get_position_slot(match_offset, queue);
945 position_footer = (match_offset + LZX_OFFSET_OFFSET) &
946 ((1U << lzx_get_num_extra_bits(position_slot)) - 1);
948 /* The match length shall be encoded as a length header (itself encoded
949 * as part of the main symbol) and an optional length footer. */
950 adjusted_match_len = match_len - LZX_MIN_MATCH_LEN;
951 if (adjusted_match_len < LZX_NUM_PRIMARY_LENS) {
952 /* No length footer needed. */
953 len_header = adjusted_match_len;
955 /* Length footer needed. It will be encoded using the length
957 len_header = LZX_NUM_PRIMARY_LENS;
958 len_footer = adjusted_match_len - LZX_NUM_PRIMARY_LENS;
959 freqs->len[len_footer]++;
962 /* Account for the main symbol. */
963 main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS;
965 freqs->main[main_symbol]++;
967 /* In an aligned offset block, 3 bits of the position footer are output
968 * as an aligned offset symbol. Account for this, although we may
969 * ultimately decide to output the block as verbatim. */
971 /* The following check is equivalent to:
973 * if (lzx_extra_bits[position_slot] >= 3)
975 * Note that this correctly excludes position slots that correspond to
977 if (position_slot >= 8)
978 freqs->aligned[position_footer & 7]++;
980 /* Pack the position slot, position footer, and match length into an
981 * intermediate representation. See `struct lzx_match' for details.
983 LZX_ASSERT(LZX_MAX_POSITION_SLOTS <= 64);
984 LZX_ASSERT(lzx_get_num_extra_bits(LZX_MAX_POSITION_SLOTS - 1) <= 17);
985 LZX_ASSERT(LZX_MAX_MATCH_LEN - LZX_MIN_MATCH_LEN + 1 <= 256);
987 LZX_ASSERT(position_slot <= (1U << (31 - 25)) - 1);
988 LZX_ASSERT(position_footer <= (1U << (25 - 8)) - 1);
989 LZX_ASSERT(adjusted_match_len <= (1U << (8 - 0)) - 1);
991 (position_slot << 25) |
992 (position_footer << 8) |
993 (adjusted_match_len);
996 struct lzx_record_ctx {
997 struct lzx_freqs freqs;
998 struct lzx_lru_queue queue;
999 struct lzx_match *matches;
1003 lzx_record_match(unsigned len, unsigned offset, void *_ctx)
1005 struct lzx_record_ctx *ctx = _ctx;
1007 (ctx->matches++)->data = lzx_tally_match(len, offset, &ctx->freqs, &ctx->queue);
1011 lzx_record_literal(u8 lit, void *_ctx)
1013 struct lzx_record_ctx *ctx = _ctx;
1015 (ctx->matches++)->data = lzx_tally_literal(lit, &ctx->freqs);
1018 /* Returns the cost, in bits, to output a literal byte using the specified cost
1021 lzx_literal_cost(u8 c, const struct lzx_costs * costs)
1023 return costs->main[c];
1026 /* Given a (length, offset) pair that could be turned into a valid LZX match as
1027 * well as costs for the codewords in the main, length, and aligned Huffman
1028 * codes, return the approximate number of bits it will take to represent this
1029 * match in the compressed output. Take into account the match offset LRU
1030 * queue and optionally update it. */
1032 lzx_match_cost(unsigned length, unsigned offset, const struct lzx_costs *costs,
1033 struct lzx_lru_queue *queue)
1035 unsigned position_slot;
1036 unsigned len_header, main_symbol;
1039 position_slot = lzx_get_position_slot(offset, queue);
1041 len_header = min(length - LZX_MIN_MATCH_LEN, LZX_NUM_PRIMARY_LENS);
1042 main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS;
1044 /* Account for main symbol. */
1045 cost += costs->main[main_symbol];
1047 /* Account for extra position information. */
1048 unsigned num_extra_bits = lzx_get_num_extra_bits(position_slot);
1049 if (num_extra_bits >= 3) {
1050 cost += num_extra_bits - 3;
1051 cost += costs->aligned[(offset + LZX_OFFSET_OFFSET) & 7];
1053 cost += num_extra_bits;
1056 /* Account for extra length information. */
1057 if (len_header == LZX_NUM_PRIMARY_LENS)
1058 cost += costs->len[length - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS];
1064 /* Fast heuristic cost evaluation to use in the inner loop of the match-finder.
1065 * Unlike lzx_match_cost() which does a true cost evaluation, this simply
1066 * prioritize matches based on their offset. */
1068 lzx_match_cost_fast(input_idx_t length, input_idx_t offset, const void *_queue)
1070 const struct lzx_lru_queue *queue = _queue;
1072 /* It seems well worth it to take the time to give priority to recently
1074 for (input_idx_t i = 0; i < LZX_NUM_RECENT_OFFSETS; i++)
1075 if (offset == queue->R[i])
1081 /* Set the cost model @ctx->costs from the Huffman codeword lengths specified in
1084 * The cost model and codeword lengths are almost the same thing, but the
1085 * Huffman codewords with length 0 correspond to symbols with zero frequency
1086 * that still need to be assigned actual costs. The specific values assigned
1087 * are arbitrary, but they should be fairly high (near the maximum codeword
1088 * length) to take into account the fact that uses of these symbols are expected
1091 lzx_set_costs(struct lzx_compressor * ctx, const struct lzx_lens * lens)
1094 unsigned num_main_syms = ctx->num_main_syms;
1097 for (i = 0; i < num_main_syms; i++) {
1098 ctx->costs.main[i] = lens->main[i];
1099 if (ctx->costs.main[i] == 0)
1100 ctx->costs.main[i] = ctx->params.alg_params.slow.main_nostat_cost;
1104 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) {
1105 ctx->costs.len[i] = lens->len[i];
1106 if (ctx->costs.len[i] == 0)
1107 ctx->costs.len[i] = ctx->params.alg_params.slow.len_nostat_cost;
1110 /* Aligned offset code */
1111 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1112 ctx->costs.aligned[i] = lens->aligned[i];
1113 if (ctx->costs.aligned[i] == 0)
1114 ctx->costs.aligned[i] = ctx->params.alg_params.slow.aligned_nostat_cost;
1118 /* Tell the match-finder to skip the specified number of bytes (@n) in the
1121 lzx_lz_skip_bytes(struct lzx_compressor *ctx, input_idx_t n)
1123 LZX_ASSERT(n <= ctx->match_window_end - ctx->match_window_pos);
1124 if (ctx->matches_cached) {
1125 ctx->match_window_pos += n;
1127 ctx->cached_matches_pos +=
1128 ctx->cached_matches[ctx->cached_matches_pos].len + 1;
1132 ctx->cached_matches[ctx->cached_matches_pos++].len = 0;
1133 lz_sarray_skip_position(&ctx->lz_sarray);
1134 ctx->match_window_pos++;
1136 LZX_ASSERT(lz_sarray_get_pos(&ctx->lz_sarray) == ctx->match_window_pos);
1140 /* Retrieve a list of matches available at the next position in the input.
1142 * The matches are written to ctx->matches in decreasing order of length, and
1143 * the return value is the number of matches found. */
1145 lzx_lz_get_matches_caching(struct lzx_compressor *ctx,
1146 const struct lzx_lru_queue *queue,
1147 struct raw_match **matches_ret)
1150 struct raw_match *matches;
1152 LZX_ASSERT(ctx->match_window_pos <= ctx->match_window_end);
1154 matches = &ctx->cached_matches[ctx->cached_matches_pos + 1];
1156 if (ctx->matches_cached) {
1157 num_matches = matches[-1].len;
1159 LZX_ASSERT(lz_sarray_get_pos(&ctx->lz_sarray) == ctx->match_window_pos);
1160 num_matches = lz_sarray_get_matches(&ctx->lz_sarray,
1162 lzx_match_cost_fast,
1164 matches[-1].len = num_matches;
1166 ctx->cached_matches_pos += num_matches + 1;
1167 *matches_ret = matches;
1169 /* Cap the length of returned matches to the number of bytes remaining,
1170 * if it is not the whole window. */
1171 if (ctx->match_window_end < ctx->window_size) {
1172 unsigned maxlen = ctx->match_window_end - ctx->match_window_pos;
1173 for (u32 i = 0; i < num_matches; i++)
1174 if (matches[i].len > maxlen)
1175 matches[i].len = maxlen;
1178 fprintf(stderr, "Pos %u/%u: %u matches\n",
1179 ctx->match_window_pos, ctx->match_window_end, num_matches);
1180 for (unsigned i = 0; i < num_matches; i++)
1181 fprintf(stderr, "\tLen %u Offset %u\n", matches[i].len, matches[i].offset);
1184 #ifdef ENABLE_LZX_DEBUG
1185 for (u32 i = 0; i < num_matches; i++) {
1186 LZX_ASSERT(matches[i].len >= LZX_MIN_MATCH_LEN);
1187 LZX_ASSERT(matches[i].len <= LZX_MAX_MATCH_LEN);
1188 LZX_ASSERT(matches[i].len <= ctx->match_window_end - ctx->match_window_pos);
1189 LZX_ASSERT(matches[i].offset > 0);
1190 LZX_ASSERT(matches[i].offset <= ctx->match_window_pos);
1191 LZX_ASSERT(!memcmp(&ctx->window[ctx->match_window_pos],
1192 &ctx->window[ctx->match_window_pos - matches[i].offset],
1197 ctx->match_window_pos++;
1202 lzx_get_prev_literal_cost(struct lzx_compressor *ctx,
1203 struct lzx_lru_queue *queue)
1205 return lzx_literal_cost(ctx->window[ctx->match_window_pos - 1],
1210 lzx_get_match_cost(struct lzx_compressor *ctx,
1211 struct lzx_lru_queue *queue,
1212 input_idx_t length, input_idx_t offset)
1214 return lzx_match_cost(length, offset, &ctx->costs, queue);
1217 static struct raw_match
1218 lzx_lz_get_near_optimal_match(struct lzx_compressor *ctx)
1220 return lz_get_near_optimal_match(&ctx->mc,
1221 lzx_lz_get_matches_caching,
1223 lzx_get_prev_literal_cost,
1230 * Set default symbol costs.
1233 lzx_set_default_costs(struct lzx_costs * costs, unsigned num_main_syms)
1237 /* Literal symbols */
1238 for (i = 0; i < LZX_NUM_CHARS; i++)
1241 /* Match header symbols */
1242 for (; i < num_main_syms; i++)
1243 costs->main[i] = 10;
1245 /* Length symbols */
1246 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
1249 /* Aligned offset symbols */
1250 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
1251 costs->aligned[i] = 3;
1254 /* Given the frequencies of symbols in a compressed block and the corresponding
1255 * Huffman codes, return LZX_BLOCKTYPE_ALIGNED or LZX_BLOCKTYPE_VERBATIM if an
1256 * aligned offset or verbatim block, respectively, will take fewer bits to
1259 lzx_choose_verbatim_or_aligned(const struct lzx_freqs * freqs,
1260 const struct lzx_codes * codes)
1262 unsigned aligned_cost = 0;
1263 unsigned verbatim_cost = 0;
1265 /* Verbatim blocks have a constant 3 bits per position footer. Aligned
1266 * offset blocks have an aligned offset symbol per position footer, plus
1267 * an extra 24 bits to output the lengths necessary to reconstruct the
1268 * aligned offset code itself. */
1269 for (unsigned i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1270 verbatim_cost += 3 * freqs->aligned[i];
1271 aligned_cost += codes->lens.aligned[i] * freqs->aligned[i];
1273 aligned_cost += LZX_ALIGNEDCODE_ELEMENT_SIZE * LZX_ALIGNEDCODE_NUM_SYMBOLS;
1274 if (aligned_cost < verbatim_cost)
1275 return LZX_BLOCKTYPE_ALIGNED;
1277 return LZX_BLOCKTYPE_VERBATIM;
1280 /* Find a near-optimal sequence of matches/literals with which to output the
1281 * specified LZX block, then set its type to that which has the minimum cost to
1284 lzx_optimize_block(struct lzx_compressor *ctx, struct lzx_block_spec *spec,
1285 unsigned num_passes)
1287 const struct lzx_lru_queue orig_queue = ctx->queue;
1288 struct lzx_freqs freqs;
1290 unsigned orig_window_pos = spec->window_pos;
1291 unsigned orig_cached_pos = ctx->cached_matches_pos;
1293 LZX_ASSERT(ctx->match_window_pos == spec->window_pos);
1295 ctx->match_window_end = spec->window_pos + spec->block_size;
1296 spec->chosen_matches_start_pos = spec->window_pos;
1298 LZX_ASSERT(num_passes >= 1);
1300 /* The first optimal parsing pass is done using the cost model already
1301 * set in ctx->costs. Each later pass is done using a cost model
1302 * computed from the previous pass. */
1303 for (unsigned pass = 0; pass < num_passes; pass++) {
1305 ctx->match_window_pos = orig_window_pos;
1306 ctx->cached_matches_pos = orig_cached_pos;
1307 ctx->queue = orig_queue;
1308 spec->num_chosen_matches = 0;
1309 memset(&freqs, 0, sizeof(freqs));
1311 for (unsigned i = spec->window_pos; i < spec->window_pos + spec->block_size; ) {
1312 struct raw_match raw_match;
1313 struct lzx_match lzx_match;
1315 raw_match = lzx_lz_get_near_optimal_match(ctx);
1316 if (raw_match.len >= LZX_MIN_MATCH_LEN) {
1317 lzx_match.data = lzx_tally_match(raw_match.len, raw_match.offset,
1318 &freqs, &ctx->queue);
1321 lzx_match.data = lzx_tally_literal(ctx->window[i], &freqs);
1324 ctx->chosen_matches[spec->chosen_matches_start_pos +
1325 spec->num_chosen_matches++] = lzx_match;
1328 lzx_make_huffman_codes(&freqs, &spec->codes,
1329 ctx->num_main_syms);
1330 if (pass < num_passes - 1)
1331 lzx_set_costs(ctx, &spec->codes.lens);
1332 ctx->matches_cached = true;
1334 spec->block_type = lzx_choose_verbatim_or_aligned(&freqs, &spec->codes);
1335 ctx->matches_cached = false;
1339 lzx_optimize_blocks(struct lzx_compressor *ctx)
1341 lzx_lru_queue_init(&ctx->queue);
1342 lz_match_chooser_begin(&ctx->mc);
1344 const unsigned num_passes = ctx->params.alg_params.slow.num_optim_passes;
1346 for (unsigned i = 0; i < ctx->num_blocks; i++)
1347 lzx_optimize_block(ctx, &ctx->block_specs[i], num_passes);
1350 /* Prepare the input window into one or more LZX blocks ready to be output. */
1352 lzx_prepare_blocks(struct lzx_compressor * ctx)
1354 /* Initialize the match-finder. */
1355 lz_sarray_load_window(&ctx->lz_sarray, ctx->window, ctx->window_size);
1356 ctx->cached_matches_pos = 0;
1357 ctx->matches_cached = false;
1358 ctx->match_window_pos = 0;
1360 /* Set up a default cost model. */
1361 lzx_set_default_costs(&ctx->costs, ctx->num_main_syms);
1363 ctx->num_blocks = DIV_ROUND_UP(ctx->window_size, LZX_DIV_BLOCK_SIZE);
1364 for (unsigned i = 0; i < ctx->num_blocks; i++) {
1365 unsigned pos = LZX_DIV_BLOCK_SIZE * i;
1366 ctx->block_specs[i].window_pos = pos;
1367 ctx->block_specs[i].block_size = min(ctx->window_size - pos, LZX_DIV_BLOCK_SIZE);
1370 /* Determine sequence of matches/literals to output for each block. */
1371 lzx_optimize_blocks(ctx);
1375 * This is the fast version of lzx_prepare_blocks(). This version "quickly"
1376 * prepares a single compressed block containing the entire input. See the
1377 * description of the "Fast algorithm" at the beginning of this file for more
1380 * Input --- the preprocessed data:
1385 * Output --- the block specification and the corresponding match/literal data:
1387 * ctx->block_specs[]
1389 * ctx->chosen_matches[]
1392 lzx_prepare_block_fast(struct lzx_compressor * ctx)
1394 struct lzx_record_ctx record_ctx;
1395 struct lzx_block_spec *spec;
1397 /* Parameters to hash chain LZ match finder
1398 * (lazy with 1 match lookahead) */
1399 static const struct lz_params lzx_lz_params = {
1400 /* Although LZX_MIN_MATCH_LEN == 2, length 2 matches typically
1401 * aren't worth choosing when using greedy or lazy parsing. */
1403 .max_match = LZX_MAX_MATCH_LEN,
1404 .max_offset = LZX_MAX_WINDOW_SIZE,
1405 .good_match = LZX_MAX_MATCH_LEN,
1406 .nice_match = LZX_MAX_MATCH_LEN,
1407 .max_chain_len = LZX_MAX_MATCH_LEN,
1408 .max_lazy_match = LZX_MAX_MATCH_LEN,
1412 /* Initialize symbol frequencies and match offset LRU queue. */
1413 memset(&record_ctx.freqs, 0, sizeof(struct lzx_freqs));
1414 lzx_lru_queue_init(&record_ctx.queue);
1415 record_ctx.matches = ctx->chosen_matches;
1417 /* Determine series of matches/literals to output. */
1418 lz_analyze_block(ctx->window,
1426 /* Set up block specification. */
1427 spec = &ctx->block_specs[0];
1428 spec->block_type = LZX_BLOCKTYPE_ALIGNED;
1429 spec->window_pos = 0;
1430 spec->block_size = ctx->window_size;
1431 spec->num_chosen_matches = (record_ctx.matches - ctx->chosen_matches);
1432 spec->chosen_matches_start_pos = 0;
1433 lzx_make_huffman_codes(&record_ctx.freqs, &spec->codes,
1434 ctx->num_main_syms);
1435 ctx->num_blocks = 1;
1439 do_call_insn_translation(u32 *call_insn_target, int input_pos,
1445 rel_offset = le32_to_cpu(*call_insn_target);
1446 if (rel_offset >= -input_pos && rel_offset < file_size) {
1447 if (rel_offset < file_size - input_pos) {
1448 /* "good translation" */
1449 abs_offset = rel_offset + input_pos;
1451 /* "compensating translation" */
1452 abs_offset = rel_offset - file_size;
1454 *call_insn_target = cpu_to_le32(abs_offset);
1458 /* This is the reverse of undo_call_insn_preprocessing() in lzx-decompress.c.
1459 * See the comment above that function for more information. */
1461 do_call_insn_preprocessing(u8 data[], int size)
1463 for (int i = 0; i < size - 10; i++) {
1464 if (data[i] == 0xe8) {
1465 do_call_insn_translation((u32*)&data[i + 1], i,
1466 LZX_WIM_MAGIC_FILESIZE);
1473 lzx_compress(const void *uncompressed_data, size_t uncompressed_size,
1474 void *compressed_data, size_t compressed_size_avail, void *_ctx)
1476 struct lzx_compressor *ctx = _ctx;
1477 struct output_bitstream ostream;
1478 size_t compressed_size;
1480 if (uncompressed_size < 100) {
1481 LZX_DEBUG("Too small to bother compressing.");
1485 if (uncompressed_size > ctx->max_window_size) {
1486 LZX_DEBUG("Can't compress %zu bytes using window of %u bytes!",
1487 uncompressed_size, ctx->max_window_size);
1491 LZX_DEBUG("Attempting to compress %zu bytes...",
1494 /* The input data must be preprocessed. To avoid changing the original
1495 * input, copy it to a temporary buffer. */
1496 memcpy(ctx->window, uncompressed_data, uncompressed_size);
1497 ctx->window_size = uncompressed_size;
1499 /* This line is unnecessary; it just avoids inconsequential accesses of
1500 * uninitialized memory that would show up in memory-checking tools such
1502 memset(&ctx->window[ctx->window_size], 0, 12);
1504 LZX_DEBUG("Preprocessing data...");
1506 /* Before doing any actual compression, do the call instruction (0xe8
1507 * byte) translation on the uncompressed data. */
1508 do_call_insn_preprocessing(ctx->window, ctx->window_size);
1510 LZX_DEBUG("Preparing blocks...");
1512 /* Prepare the compressed data. */
1513 if (ctx->params.algorithm == WIMLIB_LZX_ALGORITHM_FAST)
1514 lzx_prepare_block_fast(ctx);
1516 lzx_prepare_blocks(ctx);
1518 LZX_DEBUG("Writing compressed blocks...");
1520 /* Generate the compressed data. */
1521 init_output_bitstream(&ostream, compressed_data, compressed_size_avail);
1522 lzx_write_all_blocks(ctx, &ostream);
1524 LZX_DEBUG("Flushing bitstream...");
1525 compressed_size = flush_output_bitstream(&ostream);
1526 if (compressed_size == ~(input_idx_t)0) {
1527 LZX_DEBUG("Data did not compress to %zu bytes or less!",
1528 compressed_size_avail);
1532 LZX_DEBUG("Done: compressed %zu => %zu bytes.",
1533 uncompressed_size, compressed_size);
1535 /* Verify that we really get the same thing back when decompressing.
1536 * Although this could be disabled by default in all cases, it only
1537 * takes around 2-3% of the running time of the slow algorithm to do the
1539 if (ctx->params.algorithm == WIMLIB_LZX_ALGORITHM_SLOW
1540 #if defined(ENABLE_LZX_DEBUG) || defined(ENABLE_VERIFY_COMPRESSION)
1545 struct wimlib_decompressor *decompressor;
1547 if (0 == wimlib_create_decompressor(WIMLIB_COMPRESSION_TYPE_LZX,
1548 ctx->max_window_size,
1553 ret = wimlib_decompress(compressed_data,
1558 wimlib_free_decompressor(decompressor);
1561 ERROR("Failed to decompress data we "
1562 "compressed using LZX algorithm");
1566 if (memcmp(uncompressed_data, ctx->window, uncompressed_size)) {
1567 ERROR("Data we compressed using LZX algorithm "
1568 "didn't decompress to original");
1573 WARNING("Failed to create decompressor for "
1574 "data verification!");
1577 return compressed_size;
1581 lzx_params_valid(const struct wimlib_lzx_compressor_params *params)
1583 /* Validate parameters. */
1584 if (params->hdr.size != sizeof(struct wimlib_lzx_compressor_params)) {
1585 LZX_DEBUG("Invalid parameter structure size!");
1589 if (params->algorithm != WIMLIB_LZX_ALGORITHM_SLOW &&
1590 params->algorithm != WIMLIB_LZX_ALGORITHM_FAST)
1592 LZX_DEBUG("Invalid algorithm.");
1596 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
1597 if (params->alg_params.slow.num_optim_passes < 1)
1599 LZX_DEBUG("Invalid number of optimization passes!");
1603 if (params->alg_params.slow.main_nostat_cost < 1 ||
1604 params->alg_params.slow.main_nostat_cost > 16)
1606 LZX_DEBUG("Invalid main_nostat_cost!");
1610 if (params->alg_params.slow.len_nostat_cost < 1 ||
1611 params->alg_params.slow.len_nostat_cost > 16)
1613 LZX_DEBUG("Invalid len_nostat_cost!");
1617 if (params->alg_params.slow.aligned_nostat_cost < 1 ||
1618 params->alg_params.slow.aligned_nostat_cost > 8)
1620 LZX_DEBUG("Invalid aligned_nostat_cost!");
1628 lzx_free_compressor(void *_ctx)
1630 struct lzx_compressor *ctx = _ctx;
1633 FREE(ctx->chosen_matches);
1634 FREE(ctx->cached_matches);
1635 lz_match_chooser_destroy(&ctx->mc);
1636 lz_sarray_destroy(&ctx->lz_sarray);
1637 FREE(ctx->block_specs);
1638 FREE(ctx->prev_tab);
1645 lzx_create_compressor(size_t window_size,
1646 const struct wimlib_compressor_params_header *_params,
1649 const struct wimlib_lzx_compressor_params *params =
1650 (const struct wimlib_lzx_compressor_params*)_params;
1651 struct lzx_compressor *ctx;
1653 LZX_DEBUG("Allocating LZX context...");
1655 if (!lzx_window_size_valid(window_size))
1656 return WIMLIB_ERR_INVALID_PARAM;
1658 static const struct wimlib_lzx_compressor_params fast_default = {
1660 .size = sizeof(struct wimlib_lzx_compressor_params),
1662 .algorithm = WIMLIB_LZX_ALGORITHM_FAST,
1669 static const struct wimlib_lzx_compressor_params slow_default = {
1671 .size = sizeof(struct wimlib_lzx_compressor_params),
1673 .algorithm = WIMLIB_LZX_ALGORITHM_SLOW,
1677 .use_len2_matches = 1,
1678 .num_fast_bytes = 32,
1679 .num_optim_passes = 2,
1680 .max_search_depth = 50,
1681 .max_matches_per_pos = 3,
1682 .main_nostat_cost = 15,
1683 .len_nostat_cost = 15,
1684 .aligned_nostat_cost = 7,
1690 if (!lzx_params_valid(params))
1691 return WIMLIB_ERR_INVALID_PARAM;
1693 LZX_DEBUG("Using default algorithm and parameters.");
1694 params = &slow_default;
1697 if (params->use_defaults) {
1698 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW)
1699 params = &slow_default;
1701 params = &fast_default;
1704 LZX_DEBUG("Allocating memory.");
1706 ctx = CALLOC(1, sizeof(struct lzx_compressor));
1710 ctx->num_main_syms = lzx_get_num_main_syms(window_size);
1711 ctx->max_window_size = window_size;
1712 ctx->window = MALLOC(window_size + 12);
1713 if (ctx->window == NULL)
1716 if (params->algorithm == WIMLIB_LZX_ALGORITHM_FAST) {
1717 ctx->prev_tab = MALLOC(window_size * sizeof(ctx->prev_tab[0]));
1718 if (ctx->prev_tab == NULL)
1722 size_t block_specs_length = DIV_ROUND_UP(window_size, LZX_DIV_BLOCK_SIZE);
1723 ctx->block_specs = MALLOC(block_specs_length * sizeof(ctx->block_specs[0]));
1724 if (ctx->block_specs == NULL)
1727 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
1728 unsigned min_match_len = LZX_MIN_MATCH_LEN;
1729 if (!params->alg_params.slow.use_len2_matches)
1730 min_match_len = max(min_match_len, 3);
1732 if (!lz_sarray_init(&ctx->lz_sarray,
1736 params->alg_params.slow.max_search_depth,
1737 params->alg_params.slow.max_matches_per_pos))
1741 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
1742 if (!lz_match_chooser_init(&ctx->mc,
1743 LZX_OPTIM_ARRAY_SIZE,
1744 params->alg_params.slow.num_fast_bytes,
1749 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
1752 cache_per_pos = params->alg_params.slow.max_matches_per_pos;
1753 if (cache_per_pos > LZX_MAX_CACHE_PER_POS)
1754 cache_per_pos = LZX_MAX_CACHE_PER_POS;
1756 ctx->cached_matches = MALLOC(window_size * (cache_per_pos + 1) *
1757 sizeof(ctx->cached_matches[0]));
1758 if (ctx->cached_matches == NULL)
1762 ctx->chosen_matches = MALLOC(window_size * sizeof(ctx->chosen_matches[0]));
1763 if (ctx->chosen_matches == NULL)
1766 memcpy(&ctx->params, params, sizeof(struct wimlib_lzx_compressor_params));
1767 memset(&ctx->zero_codes, 0, sizeof(ctx->zero_codes));
1769 LZX_DEBUG("Successfully allocated new LZX context.");
1775 lzx_free_compressor(ctx);
1776 return WIMLIB_ERR_NOMEM;
1779 const struct compressor_ops lzx_compressor_ops = {
1780 .create_compressor = lzx_create_compressor,
1781 .compress = lzx_compress,
1782 .free_compressor = lzx_free_compressor,