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"
167 #ifdef ENABLE_LZX_DEBUG
168 # include "wimlib/decompress_common.h"
171 typedef u32 block_cost_t;
172 #define INFINITE_BLOCK_COST (~(block_cost_t)0)
174 #define LZX_OPTIM_ARRAY_SIZE 4096
176 #define LZX_DIV_BLOCK_SIZE 32768
178 #define LZX_MAX_CACHE_PER_POS 10
180 /* Codewords for the LZX main, length, and aligned offset Huffman codes */
181 struct lzx_codewords {
182 u16 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
183 u16 len[LZX_LENCODE_NUM_SYMBOLS];
184 u16 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
187 /* Codeword lengths (in bits) for the LZX main, length, and aligned offset
190 * A 0 length means the codeword has zero frequency.
193 u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
194 u8 len[LZX_LENCODE_NUM_SYMBOLS];
195 u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
198 /* Costs for the LZX main, length, and aligned offset Huffman symbols.
200 * If a codeword has zero frequency, it must still be assigned some nonzero cost
201 * --- generally a high cost, since even if it gets used in the next iteration,
202 * it probably will not be used very times. */
204 u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
205 u8 len[LZX_LENCODE_NUM_SYMBOLS];
206 u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
209 /* The LZX main, length, and aligned offset Huffman codes */
211 struct lzx_codewords codewords;
212 struct lzx_lens lens;
215 /* Tables for tallying symbol frequencies in the three LZX alphabets */
217 input_idx_t main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
218 input_idx_t len[LZX_LENCODE_NUM_SYMBOLS];
219 input_idx_t aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
222 /* LZX intermediate match/literal format */
226 * 31 1 if a match, 0 if a literal.
228 * 30-25 position slot. This can be at most 50, so it will fit in 6
231 * 8-24 position footer. This is the offset of the real formatted
232 * offset from the position base. This can be at most 17 bits
233 * (since lzx_extra_bits[LZX_MAX_POSITION_SLOTS - 1] is 17).
235 * 0-7 length of match, minus 2. This can be at most
236 * (LZX_MAX_MATCH_LEN - 2) == 255, so it will fit in 8 bits. */
240 /* Specification for an LZX block. */
241 struct lzx_block_spec {
243 /* One of the LZX_BLOCKTYPE_* constants indicating which type of this
247 /* 0-based position in the window at which this block starts. */
248 input_idx_t window_pos;
250 /* The number of bytes of uncompressed data this block represents. */
251 input_idx_t block_size;
253 /* The position in the 'chosen_matches' array in the `struct
254 * lzx_compressor' at which the match/literal specifications for
255 * this block begin. */
256 input_idx_t chosen_matches_start_pos;
258 /* The number of match/literal specifications for this block. */
259 input_idx_t num_chosen_matches;
261 /* Huffman codes for this block. */
262 struct lzx_codes codes;
265 /* Include template for the match-choosing algorithm. */
266 #define LZ_COMPRESSOR struct lzx_compressor
267 #define LZ_ADAPTIVE_STATE struct lzx_lru_queue
268 struct lzx_compressor;
269 #include "wimlib/lz_optimal.h"
271 /* State of the LZX compressor. */
272 struct lzx_compressor {
274 /* The parameters that were used to create the compressor. */
275 struct wimlib_lzx_compressor_params params;
277 /* The buffer of data to be compressed.
279 * 0xe8 byte preprocessing is done directly on the data here before
280 * further compression.
282 * Note that this compressor does *not* use a real sliding window!!!!
283 * It's not needed in the WIM format, since every chunk is compressed
284 * independently. This is by design, to allow random access to the
287 * We reserve a few extra bytes to potentially allow reading off the end
288 * of the array in the match-finding code for optimization purposes.
292 /* Number of bytes of data to be compressed, which is the number of
293 * bytes of data in @window that are actually valid. */
294 input_idx_t window_size;
296 /* Allocated size of the @window. */
297 input_idx_t max_window_size;
299 /* Number of symbols in the main alphabet (depends on the
300 * @max_window_size since it determines the maximum allowed offset). */
301 unsigned num_main_syms;
303 /* The current match offset LRU queue. */
304 struct lzx_lru_queue queue;
306 /* Space for the sequences of matches/literals that were chosen for each
308 struct lzx_match *chosen_matches;
310 /* Information about the LZX blocks the preprocessed input was divided
312 struct lzx_block_spec *block_specs;
314 /* Number of LZX blocks the input was divided into; a.k.a. the number of
315 * elements of @block_specs that are valid. */
318 /* This is simply filled in with zeroes and used to avoid special-casing
319 * the output of the first compressed Huffman code, which conceptually
320 * has a delta taken from a code with all symbols having zero-length
322 struct lzx_codes zero_codes;
324 /* The current cost model. */
325 struct lzx_costs costs;
327 /* Fast algorithm only: Array of hash table links. */
328 input_idx_t *prev_tab;
330 /* Slow algorithm only: Suffix array match-finder. */
331 struct lz_sarray lz_sarray;
333 /* Position in window of next match to return. */
334 input_idx_t match_window_pos;
336 /* The match-finder shall ensure the length of matches does not exceed
337 * this position in the input. */
338 input_idx_t match_window_end;
340 /* Matches found by the match-finder are cached in the following array
341 * to achieve a slight speedup when the same matches are needed on
342 * subsequent passes. This is suboptimal because different matches may
343 * be preferred with different cost models, but seems to be a worthwhile
345 struct raw_match *cached_matches;
346 unsigned cached_matches_pos;
350 struct lz_match_chooser mc;
353 /* Returns the LZX position slot that corresponds to a given match offset,
354 * taking into account the recent offset queue and updating it if the offset is
357 lzx_get_position_slot(unsigned offset, struct lzx_lru_queue *queue)
359 unsigned position_slot;
361 /* See if the offset was recently used. */
362 for (unsigned i = 0; i < LZX_NUM_RECENT_OFFSETS; i++) {
363 if (offset == queue->R[i]) {
366 /* Bring the repeat offset to the front of the
367 * queue. Note: this is, in fact, not a real
368 * LRU queue because repeat matches are simply
369 * swapped to the front. */
370 swap(queue->R[0], queue->R[i]);
372 /* The resulting position slot is simply the first index
373 * at which the offset was found in the queue. */
378 /* The offset was not recently used; look up its real position slot. */
379 position_slot = lzx_get_position_slot_raw(offset + LZX_OFFSET_OFFSET);
381 /* Bring the new offset to the front of the queue. */
382 for (unsigned i = LZX_NUM_RECENT_OFFSETS - 1; i > 0; i--)
383 queue->R[i] = queue->R[i - 1];
384 queue->R[0] = offset;
386 return position_slot;
389 /* Build the main, length, and aligned offset Huffman codes used in LZX.
391 * This takes as input the frequency tables for each code and produces as output
392 * a set of tables that map symbols to codewords and codeword lengths. */
394 lzx_make_huffman_codes(const struct lzx_freqs *freqs,
395 struct lzx_codes *codes,
396 unsigned num_main_syms)
398 make_canonical_huffman_code(num_main_syms,
399 LZX_MAX_MAIN_CODEWORD_LEN,
402 codes->codewords.main);
404 make_canonical_huffman_code(LZX_LENCODE_NUM_SYMBOLS,
405 LZX_MAX_LEN_CODEWORD_LEN,
408 codes->codewords.len);
410 make_canonical_huffman_code(LZX_ALIGNEDCODE_NUM_SYMBOLS,
411 LZX_MAX_ALIGNED_CODEWORD_LEN,
414 codes->codewords.aligned);
418 * Output an LZX match.
420 * @out: The bitstream to write the match to.
421 * @block_type: The type of the LZX block (LZX_BLOCKTYPE_ALIGNED or LZX_BLOCKTYPE_VERBATIM)
423 * @codes: Pointer to a structure that contains the codewords for the
424 * main, length, and aligned offset Huffman codes.
427 lzx_write_match(struct output_bitstream *out, int block_type,
428 struct lzx_match match, const struct lzx_codes *codes)
430 /* low 8 bits are the match length minus 2 */
431 unsigned match_len_minus_2 = match.data & 0xff;
432 /* Next 17 bits are the position footer */
433 unsigned position_footer = (match.data >> 8) & 0x1ffff; /* 17 bits */
434 /* Next 6 bits are the position slot. */
435 unsigned position_slot = (match.data >> 25) & 0x3f; /* 6 bits */
438 unsigned main_symbol;
439 unsigned num_extra_bits;
440 unsigned verbatim_bits;
441 unsigned aligned_bits;
443 /* If the match length is less than MIN_MATCH_LEN (= 2) +
444 * NUM_PRIMARY_LENS (= 7), the length header contains
445 * the match length minus MIN_MATCH_LEN, and there is no
448 * Otherwise, the length header contains
449 * NUM_PRIMARY_LENS, and the length footer contains
450 * the match length minus NUM_PRIMARY_LENS minus
452 if (match_len_minus_2 < LZX_NUM_PRIMARY_LENS) {
453 len_header = match_len_minus_2;
454 /* No length footer-- mark it with a special
456 len_footer = (unsigned)(-1);
458 len_header = LZX_NUM_PRIMARY_LENS;
459 len_footer = match_len_minus_2 - LZX_NUM_PRIMARY_LENS;
462 /* Combine the position slot with the length header into a single symbol
463 * that will be encoded with the main code.
465 * The actual main symbol is offset by LZX_NUM_CHARS because values
466 * under LZX_NUM_CHARS are used to indicate a literal byte rather than a
468 main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS;
470 /* Output main symbol. */
471 bitstream_put_bits(out, codes->codewords.main[main_symbol],
472 codes->lens.main[main_symbol]);
474 /* If there is a length footer, output it using the
475 * length Huffman code. */
476 if (len_footer != (unsigned)(-1)) {
477 bitstream_put_bits(out, codes->codewords.len[len_footer],
478 codes->lens.len[len_footer]);
481 num_extra_bits = lzx_get_num_extra_bits(position_slot);
483 /* For aligned offset blocks with at least 3 extra bits, output the
484 * verbatim bits literally, then the aligned bits encoded using the
485 * aligned offset code. Otherwise, only the verbatim bits need to be
487 if ((block_type == LZX_BLOCKTYPE_ALIGNED) && (num_extra_bits >= 3)) {
489 verbatim_bits = position_footer >> 3;
490 bitstream_put_bits(out, verbatim_bits,
493 aligned_bits = (position_footer & 7);
494 bitstream_put_bits(out,
495 codes->codewords.aligned[aligned_bits],
496 codes->lens.aligned[aligned_bits]);
498 /* verbatim bits is the same as the position
499 * footer, in this case. */
500 bitstream_put_bits(out, position_footer, num_extra_bits);
505 lzx_build_precode(const u8 lens[restrict],
506 const u8 prev_lens[restrict],
507 const unsigned num_syms,
508 input_idx_t precode_freqs[restrict LZX_PRECODE_NUM_SYMBOLS],
509 u8 output_syms[restrict num_syms],
510 u8 precode_lens[restrict LZX_PRECODE_NUM_SYMBOLS],
511 u16 precode_codewords[restrict LZX_PRECODE_NUM_SYMBOLS],
512 unsigned *num_additional_bits_ret)
514 memset(precode_freqs, 0,
515 LZX_PRECODE_NUM_SYMBOLS * sizeof(precode_freqs[0]));
517 /* Since the code word lengths use a form of RLE encoding, the goal here
518 * is to find each run of identical lengths when going through them in
519 * symbol order (including runs of length 1). For each run, as many
520 * lengths are encoded using RLE as possible, and the rest are output
523 * output_syms[] will be filled in with the length symbols that will be
524 * output, including RLE codes, not yet encoded using the precode.
526 * cur_run_len keeps track of how many code word lengths are in the
527 * current run of identical lengths. */
528 unsigned output_syms_idx = 0;
529 unsigned cur_run_len = 1;
530 unsigned num_additional_bits = 0;
531 for (unsigned i = 1; i <= num_syms; i++) {
533 if (i != num_syms && lens[i] == lens[i - 1]) {
534 /* Still in a run--- keep going. */
539 /* Run ended! Check if it is a run of zeroes or a run of
542 /* The symbol that was repeated in the run--- not to be confused
543 * with the length *of* the run (cur_run_len) */
544 unsigned len_in_run = lens[i - 1];
546 if (len_in_run == 0) {
547 /* A run of 0's. Encode it in as few length
548 * codes as we can. */
550 /* The magic length 18 indicates a run of 20 + n zeroes,
551 * where n is an uncompressed literal 5-bit integer that
552 * follows the magic length. */
553 while (cur_run_len >= 20) {
554 unsigned additional_bits;
556 additional_bits = min(cur_run_len - 20, 0x1f);
557 num_additional_bits += 5;
559 output_syms[output_syms_idx++] = 18;
560 output_syms[output_syms_idx++] = additional_bits;
561 cur_run_len -= 20 + additional_bits;
564 /* The magic length 17 indicates a run of 4 + n zeroes,
565 * where n is an uncompressed literal 4-bit integer that
566 * follows the magic length. */
567 while (cur_run_len >= 4) {
568 unsigned additional_bits;
570 additional_bits = min(cur_run_len - 4, 0xf);
571 num_additional_bits += 4;
573 output_syms[output_syms_idx++] = 17;
574 output_syms[output_syms_idx++] = additional_bits;
575 cur_run_len -= 4 + additional_bits;
580 /* A run of nonzero lengths. */
582 /* The magic length 19 indicates a run of 4 + n
583 * nonzeroes, where n is a literal bit that follows the
584 * magic length, and where the value of the lengths in
585 * the run is given by an extra length symbol, encoded
586 * with the precode, that follows the literal bit.
588 * The extra length symbol is encoded as a difference
589 * from the length of the codeword for the first symbol
590 * in the run in the previous code.
592 while (cur_run_len >= 4) {
593 unsigned additional_bits;
596 additional_bits = (cur_run_len > 4);
597 num_additional_bits += 1;
598 delta = (signed char)prev_lens[i - cur_run_len] -
599 (signed char)len_in_run;
603 precode_freqs[(unsigned char)delta]++;
604 output_syms[output_syms_idx++] = 19;
605 output_syms[output_syms_idx++] = additional_bits;
606 output_syms[output_syms_idx++] = delta;
607 cur_run_len -= 4 + additional_bits;
611 /* Any remaining lengths in the run are outputted without RLE,
612 * as a difference from the length of that codeword in the
614 while (cur_run_len > 0) {
617 delta = (signed char)prev_lens[i - cur_run_len] -
618 (signed char)len_in_run;
622 precode_freqs[(unsigned char)delta]++;
623 output_syms[output_syms_idx++] = delta;
630 /* Build the precode from the frequencies of the length symbols. */
632 make_canonical_huffman_code(LZX_PRECODE_NUM_SYMBOLS,
633 LZX_MAX_PRE_CODEWORD_LEN,
634 precode_freqs, precode_lens,
637 *num_additional_bits_ret = num_additional_bits;
639 return output_syms_idx;
643 * Writes a compressed Huffman code to the output, preceded by the precode for
646 * The Huffman code is represented in the output as a series of path lengths
647 * from which the canonical Huffman code can be reconstructed. The path lengths
648 * themselves are compressed using a separate Huffman code, the precode, which
649 * consists of LZX_PRECODE_NUM_SYMBOLS (= 20) symbols that cover all possible
650 * code lengths, plus extra codes for repeated lengths. The path lengths of the
651 * precode precede the path lengths of the larger code and are uncompressed,
652 * consisting of 20 entries of 4 bits each.
654 * @out: Bitstream to write the code to.
655 * @lens: The code lengths for the Huffman code, indexed by symbol.
656 * @prev_lens: Code lengths for this Huffman code, indexed by symbol,
657 * in the *previous block*, or all zeroes if this is the
659 * @num_syms: The number of symbols in the code.
662 lzx_write_compressed_code(struct output_bitstream *out,
663 const u8 lens[restrict],
664 const u8 prev_lens[restrict],
667 input_idx_t precode_freqs[LZX_PRECODE_NUM_SYMBOLS];
668 u8 output_syms[num_syms];
669 u8 precode_lens[LZX_PRECODE_NUM_SYMBOLS];
670 u16 precode_codewords[LZX_PRECODE_NUM_SYMBOLS];
672 unsigned num_output_syms;
676 num_output_syms = lzx_build_precode(lens,
685 /* Write the lengths of the precode codes to the output. */
686 for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++)
687 bitstream_put_bits(out, precode_lens[i],
688 LZX_PRECODE_ELEMENT_SIZE);
690 /* Write the length symbols, encoded with the precode, to the output. */
692 for (i = 0; i < num_output_syms; ) {
693 precode_sym = output_syms[i++];
695 bitstream_put_bits(out, precode_codewords[precode_sym],
696 precode_lens[precode_sym]);
697 switch (precode_sym) {
699 bitstream_put_bits(out, output_syms[i++], 4);
702 bitstream_put_bits(out, output_syms[i++], 5);
705 bitstream_put_bits(out, output_syms[i++], 1);
706 bitstream_put_bits(out,
707 precode_codewords[output_syms[i]],
708 precode_lens[output_syms[i]]);
718 * Writes all compressed matches and literal bytes in an LZX block to the the
722 * The output bitstream.
724 * The type of the block (LZX_BLOCKTYPE_ALIGNED or LZX_BLOCKTYPE_VERBATIM).
726 * The array of matches/literals that will be output (length @match_count).
728 * Number of matches/literals to be output.
730 * Pointer to a structure that contains the codewords for the main, length,
731 * and aligned offset Huffman codes.
734 lzx_write_matches_and_literals(struct output_bitstream *ostream,
736 const struct lzx_match match_tab[],
737 unsigned match_count,
738 const struct lzx_codes *codes)
740 for (unsigned i = 0; i < match_count; i++) {
741 struct lzx_match match = match_tab[i];
743 /* High bit of the match indicates whether the match is an
744 * actual match (1) or a literal uncompressed byte (0) */
745 if (match.data & 0x80000000) {
747 lzx_write_match(ostream, block_type,
751 bitstream_put_bits(ostream,
752 codes->codewords.main[match.data],
753 codes->lens.main[match.data]);
759 lzx_assert_codes_valid(const struct lzx_codes * codes, unsigned num_main_syms)
761 #ifdef ENABLE_LZX_DEBUG
764 for (i = 0; i < num_main_syms; i++)
765 LZX_ASSERT(codes->lens.main[i] <= LZX_MAX_MAIN_CODEWORD_LEN);
767 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
768 LZX_ASSERT(codes->lens.len[i] <= LZX_MAX_LEN_CODEWORD_LEN);
770 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
771 LZX_ASSERT(codes->lens.aligned[i] <= LZX_MAX_ALIGNED_CODEWORD_LEN);
773 const unsigned tablebits = 10;
774 u16 decode_table[(1 << tablebits) +
775 (2 * max(num_main_syms, LZX_LENCODE_NUM_SYMBOLS))]
776 _aligned_attribute(DECODE_TABLE_ALIGNMENT);
777 LZX_ASSERT(0 == make_huffman_decode_table(decode_table,
779 min(tablebits, LZX_MAINCODE_TABLEBITS),
781 LZX_MAX_MAIN_CODEWORD_LEN));
782 LZX_ASSERT(0 == make_huffman_decode_table(decode_table,
783 LZX_LENCODE_NUM_SYMBOLS,
784 min(tablebits, LZX_LENCODE_TABLEBITS),
786 LZX_MAX_LEN_CODEWORD_LEN));
787 LZX_ASSERT(0 == make_huffman_decode_table(decode_table,
788 LZX_ALIGNEDCODE_NUM_SYMBOLS,
789 min(tablebits, LZX_ALIGNEDCODE_TABLEBITS),
791 LZX_MAX_ALIGNED_CODEWORD_LEN));
792 #endif /* ENABLE_LZX_DEBUG */
795 /* Write an LZX aligned offset or verbatim block to the output. */
797 lzx_write_compressed_block(int block_type,
799 unsigned max_window_size,
800 unsigned num_main_syms,
801 struct lzx_match * chosen_matches,
802 unsigned num_chosen_matches,
803 const struct lzx_codes * codes,
804 const struct lzx_codes * prev_codes,
805 struct output_bitstream * ostream)
809 LZX_ASSERT(block_type == LZX_BLOCKTYPE_ALIGNED ||
810 block_type == LZX_BLOCKTYPE_VERBATIM);
811 lzx_assert_codes_valid(codes, num_main_syms);
813 /* The first three bits indicate the type of block and are one of the
814 * LZX_BLOCKTYPE_* constants. */
815 bitstream_put_bits(ostream, block_type, 3);
817 /* Output the block size.
819 * The original LZX format seemed to always encode the block size in 3
820 * bytes. However, the implementation in WIMGAPI, as used in WIM files,
821 * uses the first bit to indicate whether the block is the default size
822 * (32768) or a different size given explicitly by the next 16 bits.
824 * By default, this compressor uses a window size of 32768 and therefore
825 * follows the WIMGAPI behavior. However, this compressor also supports
826 * window sizes greater than 32768 bytes, which do not appear to be
827 * supported by WIMGAPI. In such cases, we retain the default size bit
828 * to mean a size of 32768 bytes but output non-default block size in 24
829 * bits rather than 16. The compatibility of this behavior is unknown
830 * because WIMs created with chunk size greater than 32768 can seemingly
831 * only be opened by wimlib anyway. */
832 if (block_size == LZX_DEFAULT_BLOCK_SIZE) {
833 bitstream_put_bits(ostream, 1, 1);
835 bitstream_put_bits(ostream, 0, 1);
837 if (max_window_size >= 65536)
838 bitstream_put_bits(ostream, block_size >> 16, 8);
840 bitstream_put_bits(ostream, block_size, 16);
843 /* Write out lengths of the main code. Note that the LZX specification
844 * incorrectly states that the aligned offset code comes after the
845 * length code, but in fact it is the very first code to be written
846 * (before the main code). */
847 if (block_type == LZX_BLOCKTYPE_ALIGNED)
848 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
849 bitstream_put_bits(ostream, codes->lens.aligned[i],
850 LZX_ALIGNEDCODE_ELEMENT_SIZE);
852 LZX_DEBUG("Writing main code...");
854 /* Write the precode and lengths for the first LZX_NUM_CHARS symbols in
855 * the main code, which are the codewords for literal bytes. */
856 lzx_write_compressed_code(ostream,
858 prev_codes->lens.main,
861 /* Write the precode and lengths for the rest of the main code, which
862 * are the codewords for match headers. */
863 lzx_write_compressed_code(ostream,
864 codes->lens.main + LZX_NUM_CHARS,
865 prev_codes->lens.main + LZX_NUM_CHARS,
866 num_main_syms - LZX_NUM_CHARS);
868 LZX_DEBUG("Writing length code...");
870 /* Write the precode and lengths for the length code. */
871 lzx_write_compressed_code(ostream,
873 prev_codes->lens.len,
874 LZX_LENCODE_NUM_SYMBOLS);
876 LZX_DEBUG("Writing matches and literals...");
878 /* Write the actual matches and literals. */
879 lzx_write_matches_and_literals(ostream, block_type,
880 chosen_matches, num_chosen_matches,
883 LZX_DEBUG("Done writing block.");
886 /* Write out the LZX blocks that were computed. */
888 lzx_write_all_blocks(struct lzx_compressor *ctx, struct output_bitstream *ostream)
891 const struct lzx_codes *prev_codes = &ctx->zero_codes;
892 for (unsigned i = 0; i < ctx->num_blocks; i++) {
893 const struct lzx_block_spec *spec = &ctx->block_specs[i];
895 LZX_DEBUG("Writing block %u/%u (type=%d, size=%u, num_chosen_matches=%u)...",
896 i + 1, ctx->num_blocks,
897 spec->block_type, spec->block_size,
898 spec->num_chosen_matches);
900 lzx_write_compressed_block(spec->block_type,
902 ctx->max_window_size,
904 &ctx->chosen_matches[spec->chosen_matches_start_pos],
905 spec->num_chosen_matches,
910 prev_codes = &spec->codes;
914 /* Constructs an LZX match from a literal byte and updates the main code symbol
917 lzx_tally_literal(u8 lit, struct lzx_freqs *freqs)
923 /* Constructs an LZX match from an offset and a length, and updates the LRU
924 * queue and the frequency of symbols in the main, length, and aligned offset
925 * alphabets. The return value is a 32-bit number that provides the match in an
926 * intermediate representation documented below. */
928 lzx_tally_match(unsigned match_len, unsigned match_offset,
929 struct lzx_freqs *freqs, struct lzx_lru_queue *queue)
931 unsigned position_slot;
932 unsigned position_footer;
934 unsigned main_symbol;
936 unsigned adjusted_match_len;
938 LZX_ASSERT(match_len >= LZX_MIN_MATCH_LEN && match_len <= LZX_MAX_MATCH_LEN);
940 /* The match offset shall be encoded as a position slot (itself encoded
941 * as part of the main symbol) and a position footer. */
942 position_slot = lzx_get_position_slot(match_offset, queue);
943 position_footer = (match_offset + LZX_OFFSET_OFFSET) &
944 ((1U << lzx_get_num_extra_bits(position_slot)) - 1);
946 /* The match length shall be encoded as a length header (itself encoded
947 * as part of the main symbol) and an optional length footer. */
948 adjusted_match_len = match_len - LZX_MIN_MATCH_LEN;
949 if (adjusted_match_len < LZX_NUM_PRIMARY_LENS) {
950 /* No length footer needed. */
951 len_header = adjusted_match_len;
953 /* Length footer needed. It will be encoded using the length
955 len_header = LZX_NUM_PRIMARY_LENS;
956 len_footer = adjusted_match_len - LZX_NUM_PRIMARY_LENS;
957 freqs->len[len_footer]++;
960 /* Account for the main symbol. */
961 main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS;
963 freqs->main[main_symbol]++;
965 /* In an aligned offset block, 3 bits of the position footer are output
966 * as an aligned offset symbol. Account for this, although we may
967 * ultimately decide to output the block as verbatim. */
969 /* The following check is equivalent to:
971 * if (lzx_extra_bits[position_slot] >= 3)
973 * Note that this correctly excludes position slots that correspond to
975 if (position_slot >= 8)
976 freqs->aligned[position_footer & 7]++;
978 /* Pack the position slot, position footer, and match length into an
979 * intermediate representation. See `struct lzx_match' for details.
981 LZX_ASSERT(LZX_MAX_POSITION_SLOTS <= 64);
982 LZX_ASSERT(lzx_get_num_extra_bits(LZX_MAX_POSITION_SLOTS - 1) <= 17);
983 LZX_ASSERT(LZX_MAX_MATCH_LEN - LZX_MIN_MATCH_LEN + 1 <= 256);
985 LZX_ASSERT(position_slot <= (1U << (31 - 25)) - 1);
986 LZX_ASSERT(position_footer <= (1U << (25 - 8)) - 1);
987 LZX_ASSERT(adjusted_match_len <= (1U << (8 - 0)) - 1);
989 (position_slot << 25) |
990 (position_footer << 8) |
991 (adjusted_match_len);
994 struct lzx_record_ctx {
995 struct lzx_freqs freqs;
996 struct lzx_lru_queue queue;
997 struct lzx_match *matches;
1001 lzx_record_match(unsigned len, unsigned offset, void *_ctx)
1003 struct lzx_record_ctx *ctx = _ctx;
1005 (ctx->matches++)->data = lzx_tally_match(len, offset, &ctx->freqs, &ctx->queue);
1009 lzx_record_literal(u8 lit, void *_ctx)
1011 struct lzx_record_ctx *ctx = _ctx;
1013 (ctx->matches++)->data = lzx_tally_literal(lit, &ctx->freqs);
1016 /* Returns the cost, in bits, to output a literal byte using the specified cost
1019 lzx_literal_cost(u8 c, const struct lzx_costs * costs)
1021 return costs->main[c];
1024 /* Given a (length, offset) pair that could be turned into a valid LZX match as
1025 * well as costs for the codewords in the main, length, and aligned Huffman
1026 * codes, return the approximate number of bits it will take to represent this
1027 * match in the compressed output. Take into account the match offset LRU
1028 * queue and optionally update it. */
1030 lzx_match_cost(unsigned length, unsigned offset, const struct lzx_costs *costs,
1031 struct lzx_lru_queue *queue)
1033 unsigned position_slot;
1034 unsigned len_header, main_symbol;
1037 position_slot = lzx_get_position_slot(offset, queue);
1039 len_header = min(length - LZX_MIN_MATCH_LEN, LZX_NUM_PRIMARY_LENS);
1040 main_symbol = ((position_slot << 3) | len_header) + LZX_NUM_CHARS;
1042 /* Account for main symbol. */
1043 cost += costs->main[main_symbol];
1045 /* Account for extra position information. */
1046 unsigned num_extra_bits = lzx_get_num_extra_bits(position_slot);
1047 if (num_extra_bits >= 3) {
1048 cost += num_extra_bits - 3;
1049 cost += costs->aligned[(offset + LZX_OFFSET_OFFSET) & 7];
1051 cost += num_extra_bits;
1054 /* Account for extra length information. */
1055 if (len_header == LZX_NUM_PRIMARY_LENS)
1056 cost += costs->len[length - LZX_MIN_MATCH_LEN - LZX_NUM_PRIMARY_LENS];
1062 /* Fast heuristic cost evaluation to use in the inner loop of the match-finder.
1063 * Unlike lzx_match_cost() which does a true cost evaluation, this simply
1064 * prioritize matches based on their offset. */
1066 lzx_match_cost_fast(input_idx_t length, input_idx_t offset, const void *_queue)
1068 const struct lzx_lru_queue *queue = _queue;
1070 /* It seems well worth it to take the time to give priority to recently
1072 for (input_idx_t i = 0; i < LZX_NUM_RECENT_OFFSETS; i++)
1073 if (offset == queue->R[i])
1079 /* Set the cost model @ctx->costs from the Huffman codeword lengths specified in
1082 * The cost model and codeword lengths are almost the same thing, but the
1083 * Huffman codewords with length 0 correspond to symbols with zero frequency
1084 * that still need to be assigned actual costs. The specific values assigned
1085 * are arbitrary, but they should be fairly high (near the maximum codeword
1086 * length) to take into account the fact that uses of these symbols are expected
1089 lzx_set_costs(struct lzx_compressor * ctx, const struct lzx_lens * lens)
1092 unsigned num_main_syms = ctx->num_main_syms;
1095 for (i = 0; i < num_main_syms; i++) {
1096 ctx->costs.main[i] = lens->main[i];
1097 if (ctx->costs.main[i] == 0)
1098 ctx->costs.main[i] = ctx->params.alg_params.slow.main_nostat_cost;
1102 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) {
1103 ctx->costs.len[i] = lens->len[i];
1104 if (ctx->costs.len[i] == 0)
1105 ctx->costs.len[i] = ctx->params.alg_params.slow.len_nostat_cost;
1108 /* Aligned offset code */
1109 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1110 ctx->costs.aligned[i] = lens->aligned[i];
1111 if (ctx->costs.aligned[i] == 0)
1112 ctx->costs.aligned[i] = ctx->params.alg_params.slow.aligned_nostat_cost;
1116 /* Tell the match-finder to skip the specified number of bytes (@n) in the
1119 lzx_lz_skip_bytes(struct lzx_compressor *ctx, input_idx_t n)
1121 LZX_ASSERT(n <= ctx->match_window_end - ctx->match_window_pos);
1122 if (ctx->matches_cached) {
1123 ctx->match_window_pos += n;
1125 ctx->cached_matches_pos +=
1126 ctx->cached_matches[ctx->cached_matches_pos].len + 1;
1130 ctx->cached_matches[ctx->cached_matches_pos++].len = 0;
1131 lz_sarray_skip_position(&ctx->lz_sarray);
1132 ctx->match_window_pos++;
1134 LZX_ASSERT(lz_sarray_get_pos(&ctx->lz_sarray) == ctx->match_window_pos);
1138 /* Retrieve a list of matches available at the next position in the input.
1140 * The matches are written to ctx->matches in decreasing order of length, and
1141 * the return value is the number of matches found. */
1143 lzx_lz_get_matches_caching(struct lzx_compressor *ctx,
1144 const struct lzx_lru_queue *queue,
1145 struct raw_match **matches_ret)
1148 struct raw_match *matches;
1150 LZX_ASSERT(ctx->match_window_pos <= ctx->match_window_end);
1152 matches = &ctx->cached_matches[ctx->cached_matches_pos + 1];
1154 if (ctx->matches_cached) {
1155 num_matches = matches[-1].len;
1157 LZX_ASSERT(lz_sarray_get_pos(&ctx->lz_sarray) == ctx->match_window_pos);
1158 num_matches = lz_sarray_get_matches(&ctx->lz_sarray,
1160 lzx_match_cost_fast,
1162 matches[-1].len = num_matches;
1164 ctx->cached_matches_pos += num_matches + 1;
1165 *matches_ret = matches;
1167 /* Cap the length of returned matches to the number of bytes remaining,
1168 * if it is not the whole window. */
1169 if (ctx->match_window_end < ctx->window_size) {
1170 unsigned maxlen = ctx->match_window_end - ctx->match_window_pos;
1171 for (u32 i = 0; i < num_matches; i++)
1172 if (matches[i].len > maxlen)
1173 matches[i].len = maxlen;
1176 fprintf(stderr, "Pos %u/%u: %u matches\n",
1177 ctx->match_window_pos, ctx->match_window_end, num_matches);
1178 for (unsigned i = 0; i < num_matches; i++)
1179 fprintf(stderr, "\tLen %u Offset %u\n", matches[i].len, matches[i].offset);
1182 #ifdef ENABLE_LZX_DEBUG
1183 for (u32 i = 0; i < num_matches; i++) {
1184 LZX_ASSERT(matches[i].len >= LZX_MIN_MATCH_LEN);
1185 LZX_ASSERT(matches[i].len <= LZX_MAX_MATCH_LEN);
1186 LZX_ASSERT(matches[i].len <= ctx->match_window_end - ctx->match_window_pos);
1187 LZX_ASSERT(matches[i].offset > 0);
1188 LZX_ASSERT(matches[i].offset <= ctx->match_window_pos);
1189 LZX_ASSERT(!memcmp(&ctx->window[ctx->match_window_pos],
1190 &ctx->window[ctx->match_window_pos - matches[i].offset],
1195 ctx->match_window_pos++;
1200 lzx_get_prev_literal_cost(struct lzx_compressor *ctx,
1201 struct lzx_lru_queue *queue)
1203 return lzx_literal_cost(ctx->window[ctx->match_window_pos - 1],
1208 lzx_get_match_cost(struct lzx_compressor *ctx,
1209 struct lzx_lru_queue *queue,
1210 input_idx_t length, input_idx_t offset)
1212 return lzx_match_cost(length, offset, &ctx->costs, queue);
1215 static struct raw_match
1216 lzx_lz_get_near_optimal_match(struct lzx_compressor *ctx)
1218 return lz_get_near_optimal_match(&ctx->mc,
1219 lzx_lz_get_matches_caching,
1221 lzx_get_prev_literal_cost,
1228 * Set default symbol costs.
1231 lzx_set_default_costs(struct lzx_costs * costs, unsigned num_main_syms)
1235 /* Literal symbols */
1236 for (i = 0; i < LZX_NUM_CHARS; i++)
1239 /* Match header symbols */
1240 for (; i < num_main_syms; i++)
1241 costs->main[i] = 10;
1243 /* Length symbols */
1244 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
1247 /* Aligned offset symbols */
1248 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
1249 costs->aligned[i] = 3;
1252 /* Given the frequencies of symbols in a compressed block and the corresponding
1253 * Huffman codes, return LZX_BLOCKTYPE_ALIGNED or LZX_BLOCKTYPE_VERBATIM if an
1254 * aligned offset or verbatim block, respectively, will take fewer bits to
1257 lzx_choose_verbatim_or_aligned(const struct lzx_freqs * freqs,
1258 const struct lzx_codes * codes)
1260 unsigned aligned_cost = 0;
1261 unsigned verbatim_cost = 0;
1263 /* Verbatim blocks have a constant 3 bits per position footer. Aligned
1264 * offset blocks have an aligned offset symbol per position footer, plus
1265 * an extra 24 bits to output the lengths necessary to reconstruct the
1266 * aligned offset code itself. */
1267 for (unsigned i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1268 verbatim_cost += 3 * freqs->aligned[i];
1269 aligned_cost += codes->lens.aligned[i] * freqs->aligned[i];
1271 aligned_cost += LZX_ALIGNEDCODE_ELEMENT_SIZE * LZX_ALIGNEDCODE_NUM_SYMBOLS;
1272 if (aligned_cost < verbatim_cost)
1273 return LZX_BLOCKTYPE_ALIGNED;
1275 return LZX_BLOCKTYPE_VERBATIM;
1278 /* Find a near-optimal sequence of matches/literals with which to output the
1279 * specified LZX block, then set its type to that which has the minimum cost to
1282 lzx_optimize_block(struct lzx_compressor *ctx, struct lzx_block_spec *spec,
1283 unsigned num_passes)
1285 const struct lzx_lru_queue orig_queue = ctx->queue;
1286 struct lzx_freqs freqs;
1288 unsigned orig_window_pos = spec->window_pos;
1289 unsigned orig_cached_pos = ctx->cached_matches_pos;
1291 LZX_ASSERT(ctx->match_window_pos == spec->window_pos);
1293 ctx->match_window_end = spec->window_pos + spec->block_size;
1294 spec->chosen_matches_start_pos = spec->window_pos;
1296 LZX_ASSERT(num_passes >= 1);
1298 /* The first optimal parsing pass is done using the cost model already
1299 * set in ctx->costs. Each later pass is done using a cost model
1300 * computed from the previous pass. */
1301 for (unsigned pass = 0; pass < num_passes; pass++) {
1303 ctx->match_window_pos = orig_window_pos;
1304 ctx->cached_matches_pos = orig_cached_pos;
1305 ctx->queue = orig_queue;
1306 spec->num_chosen_matches = 0;
1307 memset(&freqs, 0, sizeof(freqs));
1309 for (unsigned i = spec->window_pos; i < spec->window_pos + spec->block_size; ) {
1310 struct raw_match raw_match;
1311 struct lzx_match lzx_match;
1313 raw_match = lzx_lz_get_near_optimal_match(ctx);
1314 if (raw_match.len >= LZX_MIN_MATCH_LEN) {
1315 if (unlikely(raw_match.len == LZX_MIN_MATCH_LEN &&
1316 raw_match.offset == ctx->max_window_size -
1319 /* Degenerate case where the parser
1320 * generated the minimum match length
1321 * with the maximum offset. There
1322 * aren't actually enough position slots
1323 * to represent this offset, as noted in
1325 * lzx_get_num_main_syms(), so we cannot
1326 * allow it. Use literals instead.
1328 * Note that this case only occurs if
1329 * the match-finder can generate matches
1330 * to the very start of the window. The
1331 * suffix array match-finder can,
1332 * although typical hash chain and
1333 * binary tree match-finders use 0 as a
1334 * null value and therefore cannot
1335 * generate such matches. */
1336 BUILD_BUG_ON(LZX_MIN_MATCH_LEN != 2);
1337 lzx_match.data = lzx_tally_literal(ctx->window[i],
1340 ctx->chosen_matches[spec->chosen_matches_start_pos +
1341 spec->num_chosen_matches++]
1343 lzx_match.data = lzx_tally_literal(ctx->window[i],
1347 lzx_match.data = lzx_tally_match(raw_match.len,
1354 lzx_match.data = lzx_tally_literal(ctx->window[i], &freqs);
1357 ctx->chosen_matches[spec->chosen_matches_start_pos +
1358 spec->num_chosen_matches++] = lzx_match;
1361 lzx_make_huffman_codes(&freqs, &spec->codes,
1362 ctx->num_main_syms);
1363 if (pass < num_passes - 1)
1364 lzx_set_costs(ctx, &spec->codes.lens);
1365 ctx->matches_cached = true;
1367 spec->block_type = lzx_choose_verbatim_or_aligned(&freqs, &spec->codes);
1368 ctx->matches_cached = false;
1372 lzx_optimize_blocks(struct lzx_compressor *ctx)
1374 lzx_lru_queue_init(&ctx->queue);
1375 lz_match_chooser_begin(&ctx->mc);
1377 const unsigned num_passes = ctx->params.alg_params.slow.num_optim_passes;
1379 for (unsigned i = 0; i < ctx->num_blocks; i++)
1380 lzx_optimize_block(ctx, &ctx->block_specs[i], num_passes);
1383 /* Prepare the input window into one or more LZX blocks ready to be output. */
1385 lzx_prepare_blocks(struct lzx_compressor * ctx)
1387 /* Initialize the match-finder. */
1388 lz_sarray_load_window(&ctx->lz_sarray, ctx->window, ctx->window_size);
1389 ctx->cached_matches_pos = 0;
1390 ctx->matches_cached = false;
1391 ctx->match_window_pos = 0;
1393 /* Set up a default cost model. */
1394 lzx_set_default_costs(&ctx->costs, ctx->num_main_syms);
1396 ctx->num_blocks = DIV_ROUND_UP(ctx->window_size, LZX_DIV_BLOCK_SIZE);
1397 for (unsigned i = 0; i < ctx->num_blocks; i++) {
1398 unsigned pos = LZX_DIV_BLOCK_SIZE * i;
1399 ctx->block_specs[i].window_pos = pos;
1400 ctx->block_specs[i].block_size = min(ctx->window_size - pos, LZX_DIV_BLOCK_SIZE);
1403 /* Determine sequence of matches/literals to output for each block. */
1404 lzx_optimize_blocks(ctx);
1408 * This is the fast version of lzx_prepare_blocks(). This version "quickly"
1409 * prepares a single compressed block containing the entire input. See the
1410 * description of the "Fast algorithm" at the beginning of this file for more
1413 * Input --- the preprocessed data:
1418 * Output --- the block specification and the corresponding match/literal data:
1420 * ctx->block_specs[]
1422 * ctx->chosen_matches[]
1425 lzx_prepare_block_fast(struct lzx_compressor * ctx)
1427 struct lzx_record_ctx record_ctx;
1428 struct lzx_block_spec *spec;
1430 /* Parameters to hash chain LZ match finder
1431 * (lazy with 1 match lookahead) */
1432 static const struct lz_params lzx_lz_params = {
1433 /* Although LZX_MIN_MATCH_LEN == 2, length 2 matches typically
1434 * aren't worth choosing when using greedy or lazy parsing. */
1436 .max_match = LZX_MAX_MATCH_LEN,
1437 .max_offset = LZX_MAX_WINDOW_SIZE,
1438 .good_match = LZX_MAX_MATCH_LEN,
1439 .nice_match = LZX_MAX_MATCH_LEN,
1440 .max_chain_len = LZX_MAX_MATCH_LEN,
1441 .max_lazy_match = LZX_MAX_MATCH_LEN,
1445 /* Initialize symbol frequencies and match offset LRU queue. */
1446 memset(&record_ctx.freqs, 0, sizeof(struct lzx_freqs));
1447 lzx_lru_queue_init(&record_ctx.queue);
1448 record_ctx.matches = ctx->chosen_matches;
1450 /* Determine series of matches/literals to output. */
1451 lz_analyze_block(ctx->window,
1459 /* Set up block specification. */
1460 spec = &ctx->block_specs[0];
1461 spec->block_type = LZX_BLOCKTYPE_ALIGNED;
1462 spec->window_pos = 0;
1463 spec->block_size = ctx->window_size;
1464 spec->num_chosen_matches = (record_ctx.matches - ctx->chosen_matches);
1465 spec->chosen_matches_start_pos = 0;
1466 lzx_make_huffman_codes(&record_ctx.freqs, &spec->codes,
1467 ctx->num_main_syms);
1468 ctx->num_blocks = 1;
1472 do_call_insn_translation(u32 *call_insn_target, int input_pos,
1478 rel_offset = le32_to_cpu(*call_insn_target);
1479 if (rel_offset >= -input_pos && rel_offset < file_size) {
1480 if (rel_offset < file_size - input_pos) {
1481 /* "good translation" */
1482 abs_offset = rel_offset + input_pos;
1484 /* "compensating translation" */
1485 abs_offset = rel_offset - file_size;
1487 *call_insn_target = cpu_to_le32(abs_offset);
1491 /* This is the reverse of undo_call_insn_preprocessing() in lzx-decompress.c.
1492 * See the comment above that function for more information. */
1494 do_call_insn_preprocessing(u8 data[], int size)
1496 for (int i = 0; i < size - 10; i++) {
1497 if (data[i] == 0xe8) {
1498 do_call_insn_translation((u32*)&data[i + 1], i,
1499 LZX_WIM_MAGIC_FILESIZE);
1506 lzx_compress(const void *uncompressed_data, size_t uncompressed_size,
1507 void *compressed_data, size_t compressed_size_avail, void *_ctx)
1509 struct lzx_compressor *ctx = _ctx;
1510 struct output_bitstream ostream;
1511 size_t compressed_size;
1513 if (uncompressed_size < 100) {
1514 LZX_DEBUG("Too small to bother compressing.");
1518 if (uncompressed_size > ctx->max_window_size) {
1519 LZX_DEBUG("Can't compress %zu bytes using window of %u bytes!",
1520 uncompressed_size, ctx->max_window_size);
1524 LZX_DEBUG("Attempting to compress %zu bytes...",
1527 /* The input data must be preprocessed. To avoid changing the original
1528 * input, copy it to a temporary buffer. */
1529 memcpy(ctx->window, uncompressed_data, uncompressed_size);
1530 ctx->window_size = uncompressed_size;
1532 /* This line is unnecessary; it just avoids inconsequential accesses of
1533 * uninitialized memory that would show up in memory-checking tools such
1535 memset(&ctx->window[ctx->window_size], 0, 12);
1537 LZX_DEBUG("Preprocessing data...");
1539 /* Before doing any actual compression, do the call instruction (0xe8
1540 * byte) translation on the uncompressed data. */
1541 do_call_insn_preprocessing(ctx->window, ctx->window_size);
1543 LZX_DEBUG("Preparing blocks...");
1545 /* Prepare the compressed data. */
1546 if (ctx->params.algorithm == WIMLIB_LZX_ALGORITHM_FAST)
1547 lzx_prepare_block_fast(ctx);
1549 lzx_prepare_blocks(ctx);
1551 LZX_DEBUG("Writing compressed blocks...");
1553 /* Generate the compressed data. */
1554 init_output_bitstream(&ostream, compressed_data, compressed_size_avail);
1555 lzx_write_all_blocks(ctx, &ostream);
1557 LZX_DEBUG("Flushing bitstream...");
1558 compressed_size = flush_output_bitstream(&ostream);
1559 if (compressed_size == ~(input_idx_t)0) {
1560 LZX_DEBUG("Data did not compress to %zu bytes or less!",
1561 compressed_size_avail);
1565 LZX_DEBUG("Done: compressed %zu => %zu bytes.",
1566 uncompressed_size, compressed_size);
1568 /* Verify that we really get the same thing back when decompressing.
1569 * Although this could be disabled by default in all cases, it only
1570 * takes around 2-3% of the running time of the slow algorithm to do the
1572 if (ctx->params.algorithm == WIMLIB_LZX_ALGORITHM_SLOW
1573 #if defined(ENABLE_LZX_DEBUG) || defined(ENABLE_VERIFY_COMPRESSION)
1578 struct wimlib_decompressor *decompressor;
1580 if (0 == wimlib_create_decompressor(WIMLIB_COMPRESSION_TYPE_LZX,
1581 ctx->max_window_size,
1586 ret = wimlib_decompress(compressed_data,
1591 wimlib_free_decompressor(decompressor);
1594 ERROR("Failed to decompress data we "
1595 "compressed using LZX algorithm");
1599 if (memcmp(uncompressed_data, ctx->window, uncompressed_size)) {
1600 ERROR("Data we compressed using LZX algorithm "
1601 "didn't decompress to original");
1606 WARNING("Failed to create decompressor for "
1607 "data verification!");
1610 return compressed_size;
1614 lzx_free_compressor(void *_ctx)
1616 struct lzx_compressor *ctx = _ctx;
1619 FREE(ctx->chosen_matches);
1620 FREE(ctx->cached_matches);
1621 lz_match_chooser_destroy(&ctx->mc);
1622 lz_sarray_destroy(&ctx->lz_sarray);
1623 FREE(ctx->block_specs);
1624 FREE(ctx->prev_tab);
1630 static const struct wimlib_lzx_compressor_params lzx_fast_default = {
1632 .size = sizeof(struct wimlib_lzx_compressor_params),
1634 .algorithm = WIMLIB_LZX_ALGORITHM_FAST,
1641 static const struct wimlib_lzx_compressor_params lzx_slow_default = {
1643 .size = sizeof(struct wimlib_lzx_compressor_params),
1645 .algorithm = WIMLIB_LZX_ALGORITHM_SLOW,
1649 .use_len2_matches = 1,
1650 .nice_match_length = 32,
1651 .num_optim_passes = 2,
1652 .max_search_depth = 50,
1653 .max_matches_per_pos = 3,
1654 .main_nostat_cost = 15,
1655 .len_nostat_cost = 15,
1656 .aligned_nostat_cost = 7,
1661 static const struct wimlib_lzx_compressor_params *
1662 lzx_get_params(const struct wimlib_compressor_params_header *_params)
1664 const struct wimlib_lzx_compressor_params *params =
1665 (const struct wimlib_lzx_compressor_params*)_params;
1667 if (params == NULL) {
1668 LZX_DEBUG("Using default algorithm and parameters.");
1669 params = &lzx_slow_default;
1671 if (params->use_defaults) {
1672 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW)
1673 params = &lzx_slow_default;
1675 params = &lzx_fast_default;
1682 lzx_create_compressor(size_t window_size,
1683 const struct wimlib_compressor_params_header *_params,
1686 const struct wimlib_lzx_compressor_params *params = lzx_get_params(_params);
1687 struct lzx_compressor *ctx;
1689 LZX_DEBUG("Allocating LZX context...");
1691 if (!lzx_window_size_valid(window_size))
1692 return WIMLIB_ERR_INVALID_PARAM;
1694 LZX_DEBUG("Allocating memory.");
1696 ctx = CALLOC(1, sizeof(struct lzx_compressor));
1700 ctx->num_main_syms = lzx_get_num_main_syms(window_size);
1701 ctx->max_window_size = window_size;
1702 ctx->window = MALLOC(window_size + 12);
1703 if (ctx->window == NULL)
1706 if (params->algorithm == WIMLIB_LZX_ALGORITHM_FAST) {
1707 ctx->prev_tab = MALLOC(window_size * sizeof(ctx->prev_tab[0]));
1708 if (ctx->prev_tab == NULL)
1712 size_t block_specs_length = DIV_ROUND_UP(window_size, LZX_DIV_BLOCK_SIZE);
1713 ctx->block_specs = MALLOC(block_specs_length * sizeof(ctx->block_specs[0]));
1714 if (ctx->block_specs == NULL)
1717 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
1718 unsigned min_match_len = LZX_MIN_MATCH_LEN;
1719 if (!params->alg_params.slow.use_len2_matches)
1720 min_match_len = max(min_match_len, 3);
1722 if (!lz_sarray_init(&ctx->lz_sarray,
1726 params->alg_params.slow.max_search_depth,
1727 params->alg_params.slow.max_matches_per_pos))
1731 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
1732 if (!lz_match_chooser_init(&ctx->mc,
1733 LZX_OPTIM_ARRAY_SIZE,
1734 params->alg_params.slow.nice_match_length,
1739 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
1742 cache_per_pos = params->alg_params.slow.max_matches_per_pos;
1743 if (cache_per_pos > LZX_MAX_CACHE_PER_POS)
1744 cache_per_pos = LZX_MAX_CACHE_PER_POS;
1746 ctx->cached_matches = MALLOC(window_size * (cache_per_pos + 1) *
1747 sizeof(ctx->cached_matches[0]));
1748 if (ctx->cached_matches == NULL)
1752 ctx->chosen_matches = MALLOC(window_size * sizeof(ctx->chosen_matches[0]));
1753 if (ctx->chosen_matches == NULL)
1756 memcpy(&ctx->params, params, sizeof(struct wimlib_lzx_compressor_params));
1757 memset(&ctx->zero_codes, 0, sizeof(ctx->zero_codes));
1759 LZX_DEBUG("Successfully allocated new LZX context.");
1765 lzx_free_compressor(ctx);
1766 return WIMLIB_ERR_NOMEM;
1770 lzx_get_needed_memory(size_t max_block_size,
1771 const struct wimlib_compressor_params_header *_params)
1773 const struct wimlib_lzx_compressor_params *params = lzx_get_params(_params);
1777 size += sizeof(struct lzx_compressor);
1779 size += max_block_size + 12;
1781 size += DIV_ROUND_UP(max_block_size, LZX_DIV_BLOCK_SIZE) *
1782 sizeof(((struct lzx_compressor*)0)->block_specs[0]);
1784 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW) {
1785 size += max_block_size * sizeof(((struct lzx_compressor*)0)->chosen_matches[0]);
1786 size += lz_sarray_get_needed_memory(max_block_size);
1787 size += lz_match_chooser_get_needed_memory(LZX_OPTIM_ARRAY_SIZE,
1788 params->alg_params.slow.nice_match_length,
1792 cache_per_pos = params->alg_params.slow.max_matches_per_pos;
1793 if (cache_per_pos > LZX_MAX_CACHE_PER_POS)
1794 cache_per_pos = LZX_MAX_CACHE_PER_POS;
1796 size += max_block_size * (cache_per_pos + 1) *
1797 sizeof(((struct lzx_compressor*)0)->cached_matches[0]);
1799 size += max_block_size * sizeof(((struct lzx_compressor*)0)->prev_tab[0]);
1805 lzx_params_valid(const struct wimlib_compressor_params_header *_params)
1807 const struct wimlib_lzx_compressor_params *params =
1808 (const struct wimlib_lzx_compressor_params*)_params;
1810 if (params->hdr.size != sizeof(struct wimlib_lzx_compressor_params)) {
1811 LZX_DEBUG("Invalid parameter structure size!");
1815 if (params->algorithm != WIMLIB_LZX_ALGORITHM_SLOW &&
1816 params->algorithm != WIMLIB_LZX_ALGORITHM_FAST)
1818 LZX_DEBUG("Invalid algorithm.");
1822 if (params->algorithm == WIMLIB_LZX_ALGORITHM_SLOW &&
1823 !params->use_defaults)
1825 if (params->alg_params.slow.num_optim_passes < 1)
1827 LZX_DEBUG("Invalid number of optimization passes!");
1831 if (params->alg_params.slow.main_nostat_cost < 1 ||
1832 params->alg_params.slow.main_nostat_cost > 16)
1834 LZX_DEBUG("Invalid main_nostat_cost!");
1838 if (params->alg_params.slow.len_nostat_cost < 1 ||
1839 params->alg_params.slow.len_nostat_cost > 16)
1841 LZX_DEBUG("Invalid len_nostat_cost!");
1845 if (params->alg_params.slow.aligned_nostat_cost < 1 ||
1846 params->alg_params.slow.aligned_nostat_cost > 8)
1848 LZX_DEBUG("Invalid aligned_nostat_cost!");
1856 const struct compressor_ops lzx_compressor_ops = {
1857 .params_valid = lzx_params_valid,
1858 .get_needed_memory = lzx_get_needed_memory,
1859 .create_compressor = lzx_create_compressor,
1860 .compress = lzx_compress,
1861 .free_compressor = lzx_free_compressor,