4 * A compressor for the LZX compression format, as used in WIM files.
8 * Copyright (C) 2012, 2013, 2014, 2015 Eric Biggers
10 * This file is free software; you can redistribute it and/or modify it under
11 * the terms of the GNU Lesser General Public License as published by the Free
12 * Software Foundation; either version 3 of the License, or (at your option) any
15 * This file is distributed in the hope that it will be useful, but WITHOUT
16 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
17 * FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
20 * You should have received a copy of the GNU Lesser General Public License
21 * along with this file; if not, see http://www.gnu.org/licenses/.
26 * This file contains a compressor for the LZX ("Lempel-Ziv eXtended")
27 * compression format, as used in the WIM (Windows IMaging) file format.
29 * Two different parsing algorithms are implemented: "near-optimal" and "lazy".
30 * "Near-optimal" is significantly slower than "lazy", but results in a better
31 * compression ratio. The "near-optimal" algorithm is used at the default
34 * This file may need some slight modifications to be used outside of the WIM
35 * format. In particular, in other situations the LZX block header might be
36 * slightly different, and sliding window support might be required.
38 * Note: LZX is a compression format derived from DEFLATE, the format used by
39 * zlib and gzip. Both LZX and DEFLATE use LZ77 matching and Huffman coding.
40 * Certain details are quite similar, such as the method for storing Huffman
41 * codes. However, the main differences are:
43 * - LZX preprocesses the data to attempt to make x86 machine code slightly more
44 * compressible before attempting to compress it further.
46 * - LZX uses a "main" alphabet which combines literals and matches, with the
47 * match symbols containing a "length header" (giving all or part of the match
48 * length) and an "offset slot" (giving, roughly speaking, the order of
49 * magnitude of the match offset).
51 * - LZX does not have static Huffman blocks (that is, the kind with preset
52 * Huffman codes); however it does have two types of dynamic Huffman blocks
53 * ("verbatim" and "aligned").
55 * - LZX has a minimum match length of 2 rather than 3. Length 2 matches can be
56 * useful, but generally only if the parser is smart about choosing them.
58 * - In LZX, offset slots 0 through 2 actually represent entries in an LRU queue
59 * of match offsets. This is very useful for certain types of files, such as
60 * binary files that have repeating records.
68 * Start a new LZX block (with new Huffman codes) after this many bytes.
70 * Note: actual block sizes may slightly exceed this value.
72 * TODO: recursive splitting and cost evaluation might be good for an extremely
73 * high compression mode, but otherwise it is almost always far too slow for how
74 * much it helps. Perhaps some sort of heuristic would be useful?
76 #define LZX_DIV_BLOCK_SIZE 32768
79 * LZX_CACHE_PER_POS is the number of lz_match structures to reserve in the
80 * match cache for each byte position. This value should be high enough so that
81 * nearly the time, all matches found in a given block can fit in the match
82 * cache. However, fallback behavior (immediately terminating the block) on
83 * cache overflow is still required.
85 #define LZX_CACHE_PER_POS 7
88 * LZX_CACHE_LENGTH is the number of lz_match structures in the match cache,
89 * excluding the extra "overflow" entries. The per-position multiplier is '1 +
90 * LZX_CACHE_PER_POS' instead of 'LZX_CACHE_PER_POS' because there is an
91 * overhead of one lz_match per position, used to hold the match count at that
94 #define LZX_CACHE_LENGTH (LZX_DIV_BLOCK_SIZE * (1 + LZX_CACHE_PER_POS))
97 * LZX_MAX_MATCHES_PER_POS is an upper bound on the number of matches that can
98 * ever be saved in the match cache for a single position. Since each match we
99 * save for a single position has a distinct length, we can use the number of
100 * possible match lengths in LZX as this bound. This bound is guaranteed to be
101 * valid in all cases, although if 'nice_match_length < LZX_MAX_MATCH_LEN', then
102 * it will never actually be reached.
104 #define LZX_MAX_MATCHES_PER_POS LZX_NUM_LENS
107 * LZX_BIT_COST is a scaling factor that represents the cost to output one bit.
108 * This makes it possible to consider fractional bit costs.
110 * Note: this is only useful as a statistical trick for when the true costs are
111 * unknown. In reality, each token in LZX requires a whole number of bits to
114 #define LZX_BIT_COST 16
117 * Consideration of aligned offset costs is disabled for now, due to
118 * insufficient benefit gained from the time spent.
120 #define LZX_CONSIDER_ALIGNED_COSTS 0
123 * LZX_MAX_FAST_LEVEL is the maximum compression level at which we use the
126 #define LZX_MAX_FAST_LEVEL 34
129 * BT_MATCHFINDER_HASH2_ORDER is the log base 2 of the number of entries in the
130 * hash table for finding length 2 matches. This could be as high as 16, but
131 * using a smaller hash table speeds up compression due to reduced cache
134 #define BT_MATCHFINDER_HASH2_ORDER 12
137 * These are the compressor-side limits on the codeword lengths for each Huffman
138 * code. To make outputting bits slightly faster, some of these limits are
139 * lower than the limits defined by the LZX format. This does not significantly
140 * affect the compression ratio, at least for the block sizes we use.
142 #define MAIN_CODEWORD_LIMIT 12 /* 64-bit: can buffer 4 main symbols */
143 #define LENGTH_CODEWORD_LIMIT 12
144 #define ALIGNED_CODEWORD_LIMIT 7
145 #define PRE_CODEWORD_LIMIT 7
147 #include "wimlib/compress_common.h"
148 #include "wimlib/compressor_ops.h"
149 #include "wimlib/error.h"
150 #include "wimlib/lz_extend.h"
151 #include "wimlib/lzx_common.h"
152 #include "wimlib/unaligned.h"
153 #include "wimlib/util.h"
155 /* Matchfinders with 16-bit positions */
157 #define MF_SUFFIX _16
158 #include "wimlib/bt_matchfinder.h"
159 #include "wimlib/hc_matchfinder.h"
161 /* Matchfinders with 32-bit positions */
165 #define MF_SUFFIX _32
166 #include "wimlib/bt_matchfinder.h"
167 #include "wimlib/hc_matchfinder.h"
169 struct lzx_output_bitstream;
171 /* Codewords for the LZX Huffman codes. */
172 struct lzx_codewords {
173 u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
174 u32 len[LZX_LENCODE_NUM_SYMBOLS];
175 u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
178 /* Codeword lengths (in bits) for the LZX Huffman codes.
179 * A zero length means the corresponding codeword has zero frequency. */
181 u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS + 1];
182 u8 len[LZX_LENCODE_NUM_SYMBOLS + 1];
183 u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
186 /* Cost model for near-optimal parsing */
189 /* 'match_cost[offset_slot][len - LZX_MIN_MATCH_LEN]' is the cost for a
190 * length 'len' match that has an offset belonging to 'offset_slot'. */
191 u32 match_cost[LZX_MAX_OFFSET_SLOTS][LZX_NUM_LENS];
193 /* Cost for each symbol in the main code */
194 u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
196 /* Cost for each symbol in the length code */
197 u32 len[LZX_LENCODE_NUM_SYMBOLS];
199 #if LZX_CONSIDER_ALIGNED_COSTS
200 /* Cost for each symbol in the aligned code */
201 u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
205 /* Codewords and lengths for the LZX Huffman codes. */
207 struct lzx_codewords codewords;
208 struct lzx_lens lens;
211 /* Symbol frequency counters for the LZX Huffman codes. */
213 u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
214 u32 len[LZX_LENCODE_NUM_SYMBOLS];
215 u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
219 * Represents a run of literals followed by a match or end-of-block. This
220 * struct is needed to temporarily store items chosen by the parser, since items
221 * cannot be written until all items for the block have been chosen and the
222 * block's Huffman codes have been computed.
224 struct lzx_sequence {
226 /* The number of literals in the run. This may be 0. The literals are
227 * not stored explicitly in this structure; instead, they are read
228 * directly from the uncompressed data. */
231 /* If the next field doesn't indicate end-of-block, then this is the
232 * match length minus LZX_MIN_MATCH_LEN. */
235 /* If bit 31 is clear, then this field contains the match header in bits
236 * 0-8 and the match offset minus LZX_OFFSET_ADJUSTMENT in bits 9-30.
237 * Otherwise, this sequence's literal run was the last literal run in
238 * the block, so there is no match that follows it. */
239 u32 adjusted_offset_and_match_hdr;
243 * This structure represents a byte position in the input buffer and a node in
244 * the graph of possible match/literal choices.
246 * Logically, each incoming edge to this node is labeled with a literal or a
247 * match that can be taken to reach this position from an earlier position; and
248 * each outgoing edge from this node is labeled with a literal or a match that
249 * can be taken to advance from this position to a later position.
251 struct lzx_optimum_node {
253 /* The cost, in bits, of the lowest-cost path that has been found to
254 * reach this position. This can change as progressively lower cost
255 * paths are found to reach this position. */
259 * The match or literal that was taken to reach this position. This can
260 * change as progressively lower cost paths are found to reach this
263 * This variable is divided into two bitfields.
266 * Low bits are 0, high bits are the literal.
268 * Explicit offset matches:
269 * Low bits are the match length, high bits are the offset plus 2.
271 * Repeat offset matches:
272 * Low bits are the match length, high bits are the queue index.
275 #define OPTIMUM_OFFSET_SHIFT 9
276 #define OPTIMUM_LEN_MASK ((1 << OPTIMUM_OFFSET_SHIFT) - 1)
277 } _aligned_attribute(8);
280 * Least-recently-used queue for match offsets.
282 * This is represented as a 64-bit integer for efficiency. There are three
283 * offsets of 21 bits each. Bit 64 is garbage.
285 struct lzx_lru_queue {
289 #define LZX_QUEUE64_OFFSET_SHIFT 21
290 #define LZX_QUEUE64_OFFSET_MASK (((u64)1 << LZX_QUEUE64_OFFSET_SHIFT) - 1)
292 #define LZX_QUEUE64_R0_SHIFT (0 * LZX_QUEUE64_OFFSET_SHIFT)
293 #define LZX_QUEUE64_R1_SHIFT (1 * LZX_QUEUE64_OFFSET_SHIFT)
294 #define LZX_QUEUE64_R2_SHIFT (2 * LZX_QUEUE64_OFFSET_SHIFT)
296 #define LZX_QUEUE64_R0_MASK (LZX_QUEUE64_OFFSET_MASK << LZX_QUEUE64_R0_SHIFT)
297 #define LZX_QUEUE64_R1_MASK (LZX_QUEUE64_OFFSET_MASK << LZX_QUEUE64_R1_SHIFT)
298 #define LZX_QUEUE64_R2_MASK (LZX_QUEUE64_OFFSET_MASK << LZX_QUEUE64_R2_SHIFT)
301 lzx_lru_queue_init(struct lzx_lru_queue *queue)
303 queue->R = ((u64)1 << LZX_QUEUE64_R0_SHIFT) |
304 ((u64)1 << LZX_QUEUE64_R1_SHIFT) |
305 ((u64)1 << LZX_QUEUE64_R2_SHIFT);
309 lzx_lru_queue_R0(struct lzx_lru_queue queue)
311 return (queue.R >> LZX_QUEUE64_R0_SHIFT) & LZX_QUEUE64_OFFSET_MASK;
315 lzx_lru_queue_R1(struct lzx_lru_queue queue)
317 return (queue.R >> LZX_QUEUE64_R1_SHIFT) & LZX_QUEUE64_OFFSET_MASK;
321 lzx_lru_queue_R2(struct lzx_lru_queue queue)
323 return (queue.R >> LZX_QUEUE64_R2_SHIFT) & LZX_QUEUE64_OFFSET_MASK;
326 /* Push a match offset onto the front (most recently used) end of the queue. */
327 static inline struct lzx_lru_queue
328 lzx_lru_queue_push(struct lzx_lru_queue queue, u32 offset)
330 return (struct lzx_lru_queue) {
331 .R = (queue.R << LZX_QUEUE64_OFFSET_SHIFT) | offset,
335 /* Pop a match offset off the front (most recently used) end of the queue. */
337 lzx_lru_queue_pop(struct lzx_lru_queue *queue_p)
339 u32 offset = queue_p->R & LZX_QUEUE64_OFFSET_MASK;
340 queue_p->R >>= LZX_QUEUE64_OFFSET_SHIFT;
344 /* Swap a match offset to the front of the queue. */
345 static inline struct lzx_lru_queue
346 lzx_lru_queue_swap(struct lzx_lru_queue queue, unsigned idx)
352 return (struct lzx_lru_queue) {
353 .R = (lzx_lru_queue_R1(queue) << LZX_QUEUE64_R0_SHIFT) |
354 (lzx_lru_queue_R0(queue) << LZX_QUEUE64_R1_SHIFT) |
355 (queue.R & LZX_QUEUE64_R2_MASK),
358 return (struct lzx_lru_queue) {
359 .R = (lzx_lru_queue_R2(queue) << LZX_QUEUE64_R0_SHIFT) |
360 (queue.R & LZX_QUEUE64_R1_MASK) |
361 (lzx_lru_queue_R0(queue) << LZX_QUEUE64_R2_SHIFT),
365 /* The main LZX compressor structure */
366 struct lzx_compressor {
368 /* The "nice" match length: if a match of this length is found, then
369 * choose it immediately without further consideration. */
370 unsigned nice_match_length;
372 /* The maximum search depth: consider at most this many potential
373 * matches at each position. */
374 unsigned max_search_depth;
376 /* The log base 2 of the LZX window size for LZ match offset encoding
377 * purposes. This will be >= LZX_MIN_WINDOW_ORDER and <=
378 * LZX_MAX_WINDOW_ORDER. */
379 unsigned window_order;
381 /* The number of symbols in the main alphabet. This depends on
382 * @window_order, since @window_order determines the maximum possible
384 unsigned num_main_syms;
386 /* Number of optimization passes per block */
387 unsigned num_optim_passes;
389 /* The preprocessed buffer of data being compressed */
392 /* The number of bytes of data to be compressed, which is the number of
393 * bytes of data in @in_buffer that are actually valid. */
396 /* Pointer to the compress() implementation chosen at allocation time */
397 void (*impl)(struct lzx_compressor *, struct lzx_output_bitstream *);
399 /* If true, the compressor need not preserve the input buffer if it
400 * compresses the data successfully. */
403 /* The Huffman symbol frequency counters for the current block. */
404 struct lzx_freqs freqs;
406 /* The Huffman codes for the current and previous blocks. The one with
407 * index 'codes_index' is for the current block, and the other one is
408 * for the previous block. */
409 struct lzx_codes codes[2];
410 unsigned codes_index;
412 /* The matches and literals that the parser has chosen for the current
413 * block. The required length of this array is limited by the maximum
414 * number of matches that can ever be chosen for a single block. */
415 struct lzx_sequence chosen_sequences[DIV_ROUND_UP(LZX_DIV_BLOCK_SIZE, LZX_MIN_MATCH_LEN)];
417 /* Tables for mapping adjusted offsets to offset slots */
419 /* offset slots [0, 29] */
420 u8 offset_slot_tab_1[32768];
422 /* offset slots [30, 49] */
423 u8 offset_slot_tab_2[128];
426 /* Data for greedy or lazy parsing */
428 /* Hash chains matchfinder (MUST BE LAST!!!) */
430 struct hc_matchfinder_16 hc_mf_16;
431 struct hc_matchfinder_32 hc_mf_32;
435 /* Data for near-optimal parsing */
438 * The graph nodes for the current block.
440 * We need at least 'LZX_DIV_BLOCK_SIZE +
441 * LZX_MAX_MATCH_LEN - 1' nodes because that is the
442 * maximum block size that may be used. Add 1 because
443 * we need a node to represent end-of-block.
445 * It is possible that nodes past end-of-block are
446 * accessed during match consideration, but this can
447 * only occur if the block was truncated at
448 * LZX_DIV_BLOCK_SIZE. So the same bound still applies.
449 * Note that since nodes past the end of the block will
450 * never actually have an effect on the items that are
451 * chosen for the block, it makes no difference what
452 * their costs are initialized to (if anything).
454 struct lzx_optimum_node optimum_nodes[LZX_DIV_BLOCK_SIZE +
455 LZX_MAX_MATCH_LEN - 1 + 1];
457 /* The cost model for the current block */
458 struct lzx_costs costs;
461 * Cached matches for the current block. This array
462 * contains the matches that were found at each position
463 * in the block. Specifically, for each position, there
464 * is a special 'struct lz_match' whose 'length' field
465 * contains the number of matches that were found at
466 * that position; this is followed by the matches
467 * themselves, if any, sorted by strictly increasing
470 * Note: in rare cases, there will be a very high number
471 * of matches in the block and this array will overflow.
472 * If this happens, we force the end of the current
473 * block. LZX_CACHE_LENGTH is the length at which we
474 * actually check for overflow. The extra slots beyond
475 * this are enough to absorb the worst case overflow,
476 * which occurs if starting at
477 * &match_cache[LZX_CACHE_LENGTH - 1], we write the
478 * match count header, then write
479 * LZX_MAX_MATCHES_PER_POS matches, then skip searching
480 * for matches at 'LZX_MAX_MATCH_LEN - 1' positions and
481 * write the match count header for each.
483 struct lz_match match_cache[LZX_CACHE_LENGTH +
484 LZX_MAX_MATCHES_PER_POS +
485 LZX_MAX_MATCH_LEN - 1];
487 /* Binary trees matchfinder (MUST BE LAST!!!) */
489 struct bt_matchfinder_16 bt_mf_16;
490 struct bt_matchfinder_32 bt_mf_32;
497 * Will a matchfinder using 16-bit positions be sufficient for compressing
498 * buffers of up to the specified size? The limit could be 65536 bytes, but we
499 * also want to optimize out the use of offset_slot_tab_2 in the 16-bit case.
500 * This requires that the limit be no more than the length of offset_slot_tab_1
504 lzx_is_16_bit(size_t max_bufsize)
506 STATIC_ASSERT(ARRAY_LEN(((struct lzx_compressor *)0)->offset_slot_tab_1) == 32768);
507 return max_bufsize <= 32768;
511 * The following macros call either the 16-bit or the 32-bit version of a
512 * matchfinder function based on the value of 'is_16_bit', which will be known
513 * at compilation time.
516 #define CALL_HC_MF(is_16_bit, c, funcname, ...) \
517 ((is_16_bit) ? CONCAT(funcname, _16)(&(c)->hc_mf_16, ##__VA_ARGS__) : \
518 CONCAT(funcname, _32)(&(c)->hc_mf_32, ##__VA_ARGS__));
520 #define CALL_BT_MF(is_16_bit, c, funcname, ...) \
521 ((is_16_bit) ? CONCAT(funcname, _16)(&(c)->bt_mf_16, ##__VA_ARGS__) : \
522 CONCAT(funcname, _32)(&(c)->bt_mf_32, ##__VA_ARGS__));
525 * Structure to keep track of the current state of sending bits to the
526 * compressed output buffer.
528 * The LZX bitstream is encoded as a sequence of 16-bit coding units.
530 struct lzx_output_bitstream {
532 /* Bits that haven't yet been written to the output buffer. */
533 machine_word_t bitbuf;
535 /* Number of bits currently held in @bitbuf. */
538 /* Pointer to the start of the output buffer. */
541 /* Pointer to the position in the output buffer at which the next coding
542 * unit should be written. */
545 /* Pointer just past the end of the output buffer, rounded down to a
546 * 2-byte boundary. */
550 /* Can the specified number of bits always be added to 'bitbuf' after any
551 * pending 16-bit coding units have been flushed? */
552 #define CAN_BUFFER(n) ((n) <= (8 * sizeof(machine_word_t)) - 16)
555 * Initialize the output bitstream.
558 * The output bitstream structure to initialize.
560 * The buffer being written to.
562 * Size of @buffer, in bytes.
565 lzx_init_output(struct lzx_output_bitstream *os, void *buffer, size_t size)
570 os->next = os->start;
571 os->end = os->start + (size & ~1);
574 /* Add some bits to the bitbuffer variable of the output bitstream. The caller
575 * must make sure there is enough room. */
577 lzx_add_bits(struct lzx_output_bitstream *os, u32 bits, unsigned num_bits)
579 os->bitbuf = (os->bitbuf << num_bits) | bits;
580 os->bitcount += num_bits;
583 /* Flush bits from the bitbuffer variable to the output buffer. 'max_num_bits'
584 * specifies the maximum number of bits that may have been added since the last
587 lzx_flush_bits(struct lzx_output_bitstream *os, unsigned max_num_bits)
589 if (os->end - os->next < 6)
591 put_unaligned_u16_le(os->bitbuf >> (os->bitcount - 16), os->next + 0);
592 if (max_num_bits > 16)
593 put_unaligned_u16_le(os->bitbuf >> (os->bitcount - 32), os->next + 2);
594 if (max_num_bits > 32)
595 put_unaligned_u16_le(os->bitbuf >> (os->bitcount - 48), os->next + 4);
596 os->next += (os->bitcount >> 4) << 1;
600 /* Add at most 16 bits to the bitbuffer and flush it. */
602 lzx_write_bits(struct lzx_output_bitstream *os, u32 bits, unsigned num_bits)
604 lzx_add_bits(os, bits, num_bits);
605 lzx_flush_bits(os, 16);
609 * Flush the last coding unit to the output buffer if needed. Return the total
610 * number of bytes written to the output buffer, or 0 if an overflow occurred.
613 lzx_flush_output(struct lzx_output_bitstream *os)
615 if (os->end - os->next < 6)
618 if (os->bitcount != 0) {
619 put_unaligned_u16_le(os->bitbuf << (16 - os->bitcount), os->next);
623 return os->next - os->start;
626 /* Build the main, length, and aligned offset Huffman codes used in LZX.
628 * This takes as input the frequency tables for each code and produces as output
629 * a set of tables that map symbols to codewords and codeword lengths. */
631 lzx_make_huffman_codes(struct lzx_compressor *c)
633 const struct lzx_freqs *freqs = &c->freqs;
634 struct lzx_codes *codes = &c->codes[c->codes_index];
636 STATIC_ASSERT(MAIN_CODEWORD_LIMIT >= 9 &&
637 MAIN_CODEWORD_LIMIT <= LZX_MAX_MAIN_CODEWORD_LEN);
638 STATIC_ASSERT(LENGTH_CODEWORD_LIMIT >= 9 &&
639 LENGTH_CODEWORD_LIMIT <= LZX_MAX_LEN_CODEWORD_LEN);
640 STATIC_ASSERT(ALIGNED_CODEWORD_LIMIT >= LZX_NUM_ALIGNED_OFFSET_BITS &&
641 ALIGNED_CODEWORD_LIMIT <= LZX_MAX_ALIGNED_CODEWORD_LEN);
643 make_canonical_huffman_code(c->num_main_syms,
647 codes->codewords.main);
649 make_canonical_huffman_code(LZX_LENCODE_NUM_SYMBOLS,
650 LENGTH_CODEWORD_LIMIT,
653 codes->codewords.len);
655 make_canonical_huffman_code(LZX_ALIGNEDCODE_NUM_SYMBOLS,
656 ALIGNED_CODEWORD_LIMIT,
659 codes->codewords.aligned);
662 /* Reset the symbol frequencies for the LZX Huffman codes. */
664 lzx_reset_symbol_frequencies(struct lzx_compressor *c)
666 memset(&c->freqs, 0, sizeof(c->freqs));
670 lzx_compute_precode_items(const u8 lens[restrict],
671 const u8 prev_lens[restrict],
672 u32 precode_freqs[restrict],
673 unsigned precode_items[restrict])
682 itemptr = precode_items;
685 while (!((len = lens[run_start]) & 0x80)) {
687 /* len = the length being repeated */
689 /* Find the next run of codeword lengths. */
691 run_end = run_start + 1;
693 /* Fast case for a single length. */
694 if (likely(len != lens[run_end])) {
695 delta = prev_lens[run_start] - len;
698 precode_freqs[delta]++;
704 /* Extend the run. */
707 } while (len == lens[run_end]);
712 /* Symbol 18: RLE 20 to 51 zeroes at a time. */
713 while ((run_end - run_start) >= 20) {
714 extra_bits = min((run_end - run_start) - 20, 0x1f);
716 *itemptr++ = 18 | (extra_bits << 5);
717 run_start += 20 + extra_bits;
720 /* Symbol 17: RLE 4 to 19 zeroes at a time. */
721 if ((run_end - run_start) >= 4) {
722 extra_bits = min((run_end - run_start) - 4, 0xf);
724 *itemptr++ = 17 | (extra_bits << 5);
725 run_start += 4 + extra_bits;
729 /* A run of nonzero lengths. */
731 /* Symbol 19: RLE 4 to 5 of any length at a time. */
732 while ((run_end - run_start) >= 4) {
733 extra_bits = (run_end - run_start) > 4;
734 delta = prev_lens[run_start] - len;
738 precode_freqs[delta]++;
739 *itemptr++ = 19 | (extra_bits << 5) | (delta << 6);
740 run_start += 4 + extra_bits;
744 /* Output any remaining lengths without RLE. */
745 while (run_start != run_end) {
746 delta = prev_lens[run_start] - len;
749 precode_freqs[delta]++;
755 return itemptr - precode_items;
759 * Output a Huffman code in the compressed form used in LZX.
761 * The Huffman code is represented in the output as a logical series of codeword
762 * lengths from which the Huffman code, which must be in canonical form, can be
765 * The codeword lengths are themselves compressed using a separate Huffman code,
766 * the "precode", which contains a symbol for each possible codeword length in
767 * the larger code as well as several special symbols to represent repeated
768 * codeword lengths (a form of run-length encoding). The precode is itself
769 * constructed in canonical form, and its codeword lengths are represented
770 * literally in 20 4-bit fields that immediately precede the compressed codeword
771 * lengths of the larger code.
773 * Furthermore, the codeword lengths of the larger code are actually represented
774 * as deltas from the codeword lengths of the corresponding code in the previous
778 * Bitstream to which to write the compressed Huffman code.
780 * The codeword lengths, indexed by symbol, in the Huffman code.
782 * The codeword lengths, indexed by symbol, in the corresponding Huffman
783 * code in the previous block, or all zeroes if this is the first block.
785 * The number of symbols in the Huffman code.
788 lzx_write_compressed_code(struct lzx_output_bitstream *os,
789 const u8 lens[restrict],
790 const u8 prev_lens[restrict],
793 u32 precode_freqs[LZX_PRECODE_NUM_SYMBOLS];
794 u8 precode_lens[LZX_PRECODE_NUM_SYMBOLS];
795 u32 precode_codewords[LZX_PRECODE_NUM_SYMBOLS];
796 unsigned precode_items[num_lens];
797 unsigned num_precode_items;
798 unsigned precode_item;
799 unsigned precode_sym;
801 u8 saved = lens[num_lens];
802 *(u8 *)(lens + num_lens) = 0x80;
804 for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++)
805 precode_freqs[i] = 0;
807 /* Compute the "items" (RLE / literal tokens and extra bits) with which
808 * the codeword lengths in the larger code will be output. */
809 num_precode_items = lzx_compute_precode_items(lens,
814 /* Build the precode. */
815 STATIC_ASSERT(PRE_CODEWORD_LIMIT >= 5 &&
816 PRE_CODEWORD_LIMIT <= LZX_MAX_PRE_CODEWORD_LEN);
817 make_canonical_huffman_code(LZX_PRECODE_NUM_SYMBOLS,
819 precode_freqs, precode_lens,
822 /* Output the lengths of the codewords in the precode. */
823 for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++)
824 lzx_write_bits(os, precode_lens[i], LZX_PRECODE_ELEMENT_SIZE);
826 /* Output the encoded lengths of the codewords in the larger code. */
827 for (i = 0; i < num_precode_items; i++) {
828 precode_item = precode_items[i];
829 precode_sym = precode_item & 0x1F;
830 lzx_add_bits(os, precode_codewords[precode_sym],
831 precode_lens[precode_sym]);
832 if (precode_sym >= 17) {
833 if (precode_sym == 17) {
834 lzx_add_bits(os, precode_item >> 5, 4);
835 } else if (precode_sym == 18) {
836 lzx_add_bits(os, precode_item >> 5, 5);
838 lzx_add_bits(os, (precode_item >> 5) & 1, 1);
839 precode_sym = precode_item >> 6;
840 lzx_add_bits(os, precode_codewords[precode_sym],
841 precode_lens[precode_sym]);
844 STATIC_ASSERT(CAN_BUFFER(2 * PRE_CODEWORD_LIMIT + 1));
845 lzx_flush_bits(os, 2 * PRE_CODEWORD_LIMIT + 1);
848 *(u8 *)(lens + num_lens) = saved;
852 * Write all matches and literal bytes (which were precomputed) in an LZX
853 * compressed block to the output bitstream in the final compressed
857 * The output bitstream.
859 * The chosen type of the LZX compressed block (LZX_BLOCKTYPE_ALIGNED or
860 * LZX_BLOCKTYPE_VERBATIM).
862 * The uncompressed data of the block.
864 * The matches and literals to output, given as a series of sequences.
866 * The main, length, and aligned offset Huffman codes for the current
867 * LZX compressed block.
870 lzx_write_sequences(struct lzx_output_bitstream *os, int block_type,
871 const u8 *block_data, const struct lzx_sequence sequences[],
872 const struct lzx_codes *codes)
874 const struct lzx_sequence *seq = sequences;
875 u32 ones_if_aligned = 0 - (block_type == LZX_BLOCKTYPE_ALIGNED);
878 /* Output the next sequence. */
880 unsigned litrunlen = seq->litrunlen;
882 unsigned main_symbol;
883 unsigned adjusted_length;
885 unsigned offset_slot;
886 unsigned num_extra_bits;
889 /* Output the literal run of the sequence. */
891 if (litrunlen) { /* Is the literal run nonempty? */
893 /* Verify optimization is enabled on 64-bit */
894 STATIC_ASSERT(sizeof(machine_word_t) < 8 ||
895 CAN_BUFFER(4 * MAIN_CODEWORD_LIMIT));
897 if (CAN_BUFFER(4 * MAIN_CODEWORD_LIMIT)) {
899 /* 64-bit: write 4 literals at a time. */
900 while (litrunlen >= 4) {
901 unsigned lit0 = block_data[0];
902 unsigned lit1 = block_data[1];
903 unsigned lit2 = block_data[2];
904 unsigned lit3 = block_data[3];
905 lzx_add_bits(os, codes->codewords.main[lit0], codes->lens.main[lit0]);
906 lzx_add_bits(os, codes->codewords.main[lit1], codes->lens.main[lit1]);
907 lzx_add_bits(os, codes->codewords.main[lit2], codes->lens.main[lit2]);
908 lzx_add_bits(os, codes->codewords.main[lit3], codes->lens.main[lit3]);
909 lzx_flush_bits(os, 4 * MAIN_CODEWORD_LIMIT);
914 unsigned lit = *block_data++;
915 lzx_add_bits(os, codes->codewords.main[lit], codes->lens.main[lit]);
917 unsigned lit = *block_data++;
918 lzx_add_bits(os, codes->codewords.main[lit], codes->lens.main[lit]);
920 unsigned lit = *block_data++;
921 lzx_add_bits(os, codes->codewords.main[lit], codes->lens.main[lit]);
922 lzx_flush_bits(os, 3 * MAIN_CODEWORD_LIMIT);
924 lzx_flush_bits(os, 2 * MAIN_CODEWORD_LIMIT);
927 lzx_flush_bits(os, 1 * MAIN_CODEWORD_LIMIT);
931 /* 32-bit: write 1 literal at a time. */
933 unsigned lit = *block_data++;
934 lzx_add_bits(os, codes->codewords.main[lit], codes->lens.main[lit]);
935 lzx_flush_bits(os, MAIN_CODEWORD_LIMIT);
936 } while (--litrunlen);
940 /* Was this the last literal run? */
941 if (seq->adjusted_offset_and_match_hdr & 0x80000000)
944 /* Nope; output the match. */
946 match_hdr = seq->adjusted_offset_and_match_hdr & 0x1FF;
947 main_symbol = LZX_NUM_CHARS + match_hdr;
948 adjusted_length = seq->adjusted_length;
950 block_data += adjusted_length + LZX_MIN_MATCH_LEN;
952 offset_slot = match_hdr / LZX_NUM_LEN_HEADERS;
953 adjusted_offset = seq->adjusted_offset_and_match_hdr >> 9;
955 num_extra_bits = lzx_extra_offset_bits[offset_slot];
956 extra_bits = adjusted_offset - lzx_offset_slot_base[offset_slot];
958 #define MAX_MATCH_BITS (MAIN_CODEWORD_LIMIT + LENGTH_CODEWORD_LIMIT + \
959 14 + ALIGNED_CODEWORD_LIMIT)
961 /* Verify optimization is enabled on 64-bit */
962 STATIC_ASSERT(sizeof(machine_word_t) < 8 || CAN_BUFFER(MAX_MATCH_BITS));
964 /* Output the main symbol for the match. */
966 lzx_add_bits(os, codes->codewords.main[main_symbol],
967 codes->lens.main[main_symbol]);
968 if (!CAN_BUFFER(MAX_MATCH_BITS))
969 lzx_flush_bits(os, MAIN_CODEWORD_LIMIT);
971 /* If needed, output the length symbol for the match. */
973 if (adjusted_length >= LZX_NUM_PRIMARY_LENS) {
974 lzx_add_bits(os, codes->codewords.len[adjusted_length - LZX_NUM_PRIMARY_LENS],
975 codes->lens.len[adjusted_length - LZX_NUM_PRIMARY_LENS]);
976 if (!CAN_BUFFER(MAX_MATCH_BITS))
977 lzx_flush_bits(os, LENGTH_CODEWORD_LIMIT);
980 /* Output the extra offset bits for the match. In aligned
981 * offset blocks, the lowest 3 bits of the adjusted offset are
982 * Huffman-encoded using the aligned offset code, provided that
983 * there are at least extra 3 offset bits required. All other
984 * extra offset bits are output verbatim. */
986 if ((adjusted_offset & ones_if_aligned) >= 16) {
988 lzx_add_bits(os, extra_bits >> LZX_NUM_ALIGNED_OFFSET_BITS,
989 num_extra_bits - LZX_NUM_ALIGNED_OFFSET_BITS);
990 if (!CAN_BUFFER(MAX_MATCH_BITS))
991 lzx_flush_bits(os, 14);
993 lzx_add_bits(os, codes->codewords.aligned[adjusted_offset & LZX_ALIGNED_OFFSET_BITMASK],
994 codes->lens.aligned[adjusted_offset & LZX_ALIGNED_OFFSET_BITMASK]);
995 if (!CAN_BUFFER(MAX_MATCH_BITS))
996 lzx_flush_bits(os, ALIGNED_CODEWORD_LIMIT);
998 lzx_add_bits(os, extra_bits, num_extra_bits);
999 if (!CAN_BUFFER(MAX_MATCH_BITS))
1000 lzx_flush_bits(os, 17);
1003 if (CAN_BUFFER(MAX_MATCH_BITS))
1004 lzx_flush_bits(os, MAX_MATCH_BITS);
1006 /* Advance to the next sequence. */
1012 lzx_write_compressed_block(const u8 *block_begin,
1015 unsigned window_order,
1016 unsigned num_main_syms,
1017 const struct lzx_sequence sequences[],
1018 const struct lzx_codes * codes,
1019 const struct lzx_lens * prev_lens,
1020 struct lzx_output_bitstream * os)
1022 /* The first three bits indicate the type of block and are one of the
1023 * LZX_BLOCKTYPE_* constants. */
1024 lzx_write_bits(os, block_type, 3);
1026 /* Output the block size.
1028 * The original LZX format seemed to always encode the block size in 3
1029 * bytes. However, the implementation in WIMGAPI, as used in WIM files,
1030 * uses the first bit to indicate whether the block is the default size
1031 * (32768) or a different size given explicitly by the next 16 bits.
1033 * By default, this compressor uses a window size of 32768 and therefore
1034 * follows the WIMGAPI behavior. However, this compressor also supports
1035 * window sizes greater than 32768 bytes, which do not appear to be
1036 * supported by WIMGAPI. In such cases, we retain the default size bit
1037 * to mean a size of 32768 bytes but output non-default block size in 24
1038 * bits rather than 16. The compatibility of this behavior is unknown
1039 * because WIMs created with chunk size greater than 32768 can seemingly
1040 * only be opened by wimlib anyway. */
1041 if (block_size == LZX_DEFAULT_BLOCK_SIZE) {
1042 lzx_write_bits(os, 1, 1);
1044 lzx_write_bits(os, 0, 1);
1046 if (window_order >= 16)
1047 lzx_write_bits(os, block_size >> 16, 8);
1049 lzx_write_bits(os, block_size & 0xFFFF, 16);
1052 /* If it's an aligned offset block, output the aligned offset code. */
1053 if (block_type == LZX_BLOCKTYPE_ALIGNED) {
1054 for (int i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1055 lzx_write_bits(os, codes->lens.aligned[i],
1056 LZX_ALIGNEDCODE_ELEMENT_SIZE);
1060 /* Output the main code (two parts). */
1061 lzx_write_compressed_code(os, codes->lens.main,
1064 lzx_write_compressed_code(os, codes->lens.main + LZX_NUM_CHARS,
1065 prev_lens->main + LZX_NUM_CHARS,
1066 num_main_syms - LZX_NUM_CHARS);
1068 /* Output the length code. */
1069 lzx_write_compressed_code(os, codes->lens.len,
1071 LZX_LENCODE_NUM_SYMBOLS);
1073 /* Output the compressed matches and literals. */
1074 lzx_write_sequences(os, block_type, block_begin, sequences, codes);
1077 /* Given the frequencies of symbols in an LZX-compressed block and the
1078 * corresponding Huffman codes, return LZX_BLOCKTYPE_ALIGNED or
1079 * LZX_BLOCKTYPE_VERBATIM if an aligned offset or verbatim block, respectively,
1080 * will take fewer bits to output. */
1082 lzx_choose_verbatim_or_aligned(const struct lzx_freqs * freqs,
1083 const struct lzx_codes * codes)
1085 u32 aligned_cost = 0;
1086 u32 verbatim_cost = 0;
1088 /* A verbatim block requires 3 bits in each place that an aligned symbol
1089 * would be used in an aligned offset block. */
1090 for (unsigned i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1091 verbatim_cost += LZX_NUM_ALIGNED_OFFSET_BITS * freqs->aligned[i];
1092 aligned_cost += codes->lens.aligned[i] * freqs->aligned[i];
1095 /* Account for output of the aligned offset code. */
1096 aligned_cost += LZX_ALIGNEDCODE_ELEMENT_SIZE * LZX_ALIGNEDCODE_NUM_SYMBOLS;
1098 if (aligned_cost < verbatim_cost)
1099 return LZX_BLOCKTYPE_ALIGNED;
1101 return LZX_BLOCKTYPE_VERBATIM;
1105 * Return the offset slot for the specified adjusted match offset, using the
1106 * compressor's acceleration tables to speed up the mapping.
1108 static inline unsigned
1109 lzx_comp_get_offset_slot(struct lzx_compressor *c, u32 adjusted_offset,
1112 if (is_16_bit || adjusted_offset < ARRAY_LEN(c->offset_slot_tab_1))
1113 return c->offset_slot_tab_1[adjusted_offset];
1114 return c->offset_slot_tab_2[adjusted_offset >> 14];
1118 * Finish an LZX block:
1120 * - build the Huffman codes
1121 * - decide whether to output the block as VERBATIM or ALIGNED
1122 * - output the block
1123 * - swap the indices of the current and previous Huffman codes
1126 lzx_finish_block(struct lzx_compressor *c, struct lzx_output_bitstream *os,
1127 const u8 *block_begin, u32 block_size, u32 seq_idx)
1131 lzx_make_huffman_codes(c);
1133 block_type = lzx_choose_verbatim_or_aligned(&c->freqs,
1134 &c->codes[c->codes_index]);
1135 lzx_write_compressed_block(block_begin,
1140 &c->chosen_sequences[seq_idx],
1141 &c->codes[c->codes_index],
1142 &c->codes[c->codes_index ^ 1].lens,
1144 c->codes_index ^= 1;
1147 /* Tally the Huffman symbol for a literal and increment the literal run length.
1150 lzx_record_literal(struct lzx_compressor *c, unsigned literal, u32 *litrunlen_p)
1152 c->freqs.main[literal]++;
1156 /* Tally the Huffman symbol for a match, save the match data and the length of
1157 * the preceding literal run in the next lzx_sequence, and update the recent
1160 lzx_record_match(struct lzx_compressor *c, unsigned length, u32 offset_data,
1161 u32 recent_offsets[LZX_NUM_RECENT_OFFSETS], bool is_16_bit,
1162 u32 *litrunlen_p, struct lzx_sequence **next_seq_p)
1164 u32 litrunlen = *litrunlen_p;
1165 struct lzx_sequence *next_seq = *next_seq_p;
1166 unsigned offset_slot;
1169 v = length - LZX_MIN_MATCH_LEN;
1171 /* Save the literal run length and adjusted length. */
1172 next_seq->litrunlen = litrunlen;
1173 next_seq->adjusted_length = v;
1175 /* Compute the length header and tally the length symbol if needed */
1176 if (v >= LZX_NUM_PRIMARY_LENS) {
1177 c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++;
1178 v = LZX_NUM_PRIMARY_LENS;
1181 /* Compute the offset slot */
1182 offset_slot = lzx_comp_get_offset_slot(c, offset_data, is_16_bit);
1184 /* Compute the match header. */
1185 v += offset_slot * LZX_NUM_LEN_HEADERS;
1187 /* Save the adjusted offset and match header. */
1188 next_seq->adjusted_offset_and_match_hdr = (offset_data << 9) | v;
1190 /* Tally the main symbol. */
1191 c->freqs.main[LZX_NUM_CHARS + v]++;
1193 /* Update the recent offsets queue. */
1194 if (offset_data < LZX_NUM_RECENT_OFFSETS) {
1195 /* Repeat offset match */
1196 swap(recent_offsets[0], recent_offsets[offset_data]);
1198 /* Explicit offset match */
1200 /* Tally the aligned offset symbol if needed */
1201 if (offset_data >= 16)
1202 c->freqs.aligned[offset_data & LZX_ALIGNED_OFFSET_BITMASK]++;
1204 recent_offsets[2] = recent_offsets[1];
1205 recent_offsets[1] = recent_offsets[0];
1206 recent_offsets[0] = offset_data - LZX_OFFSET_ADJUSTMENT;
1209 /* Reset the literal run length and advance to the next sequence. */
1210 *next_seq_p = next_seq + 1;
1214 /* Finish the last lzx_sequence. The last lzx_sequence is just a literal run;
1215 * there is no match. This literal run may be empty. */
1217 lzx_finish_sequence(struct lzx_sequence *last_seq, u32 litrunlen)
1219 last_seq->litrunlen = litrunlen;
1221 /* Special value to mark last sequence */
1222 last_seq->adjusted_offset_and_match_hdr = 0x80000000;
1226 * Given the minimum-cost path computed through the item graph for the current
1227 * block, walk the path and count how many of each symbol in each Huffman-coded
1228 * alphabet would be required to output the items (matches and literals) along
1231 * Note that the path will be walked backwards (from the end of the block to the
1232 * beginning of the block), but this doesn't matter because this function only
1233 * computes frequencies.
1236 lzx_tally_item_list(struct lzx_compressor *c, u32 block_size, bool is_16_bit)
1238 u32 node_idx = block_size;
1243 unsigned offset_slot;
1245 /* Tally literals until either a match or the beginning of the
1246 * block is reached. */
1248 u32 item = c->optimum_nodes[node_idx].item;
1250 len = item & OPTIMUM_LEN_MASK;
1251 offset_data = item >> OPTIMUM_OFFSET_SHIFT;
1253 if (len != 0) /* Not a literal? */
1256 /* Tally the main symbol for the literal. */
1257 c->freqs.main[offset_data]++;
1259 if (--node_idx == 0) /* Beginning of block was reached? */
1265 /* Tally a match. */
1267 /* Tally the aligned offset symbol if needed. */
1268 if (offset_data >= 16)
1269 c->freqs.aligned[offset_data & LZX_ALIGNED_OFFSET_BITMASK]++;
1271 /* Tally the length symbol if needed. */
1272 v = len - LZX_MIN_MATCH_LEN;;
1273 if (v >= LZX_NUM_PRIMARY_LENS) {
1274 c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++;
1275 v = LZX_NUM_PRIMARY_LENS;
1278 /* Tally the main symbol. */
1279 offset_slot = lzx_comp_get_offset_slot(c, offset_data, is_16_bit);
1280 v += offset_slot * LZX_NUM_LEN_HEADERS;
1281 c->freqs.main[LZX_NUM_CHARS + v]++;
1283 if (node_idx == 0) /* Beginning of block was reached? */
1289 * Like lzx_tally_item_list(), but this function also generates the list of
1290 * lzx_sequences for the minimum-cost path and writes it to c->chosen_sequences,
1291 * ready to be output to the bitstream after the Huffman codes are computed.
1292 * The lzx_sequences will be written to decreasing memory addresses as the path
1293 * is walked backwards, which means they will end up in the expected
1294 * first-to-last order. The return value is the index in c->chosen_sequences at
1295 * which the lzx_sequences begin.
1298 lzx_record_item_list(struct lzx_compressor *c, u32 block_size, bool is_16_bit)
1300 u32 node_idx = block_size;
1301 u32 seq_idx = ARRAY_LEN(c->chosen_sequences) - 1;
1304 /* Special value to mark last sequence */
1305 c->chosen_sequences[seq_idx].adjusted_offset_and_match_hdr = 0x80000000;
1307 lit_start_node = node_idx;
1312 unsigned offset_slot;
1314 /* Record literals until either a match or the beginning of the
1315 * block is reached. */
1317 u32 item = c->optimum_nodes[node_idx].item;
1319 len = item & OPTIMUM_LEN_MASK;
1320 offset_data = item >> OPTIMUM_OFFSET_SHIFT;
1322 if (len != 0) /* Not a literal? */
1325 /* Tally the main symbol for the literal. */
1326 c->freqs.main[offset_data]++;
1328 if (--node_idx == 0) /* Beginning of block was reached? */
1332 /* Save the literal run length for the next sequence (the
1333 * "previous sequence" when walking backwards). */
1334 c->chosen_sequences[seq_idx--].litrunlen = lit_start_node - node_idx;
1336 lit_start_node = node_idx;
1338 /* Record a match. */
1340 /* Tally the aligned offset symbol if needed. */
1341 if (offset_data >= 16)
1342 c->freqs.aligned[offset_data & LZX_ALIGNED_OFFSET_BITMASK]++;
1344 /* Save the adjusted length. */
1345 v = len - LZX_MIN_MATCH_LEN;
1346 c->chosen_sequences[seq_idx].adjusted_length = v;
1348 /* Tally the length symbol if needed. */
1349 if (v >= LZX_NUM_PRIMARY_LENS) {
1350 c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++;
1351 v = LZX_NUM_PRIMARY_LENS;
1354 /* Tally the main symbol. */
1355 offset_slot = lzx_comp_get_offset_slot(c, offset_data, is_16_bit);
1356 v += offset_slot * LZX_NUM_LEN_HEADERS;
1357 c->freqs.main[LZX_NUM_CHARS + v]++;
1359 /* Save the adjusted offset and match header. */
1360 c->chosen_sequences[seq_idx].adjusted_offset_and_match_hdr =
1361 (offset_data << 9) | v;
1363 if (node_idx == 0) /* Beginning of block was reached? */
1368 /* Save the literal run length for the first sequence. */
1369 c->chosen_sequences[seq_idx].litrunlen = lit_start_node - node_idx;
1371 /* Return the index in c->chosen_sequences at which the lzx_sequences
1377 * Find an inexpensive path through the graph of possible match/literal choices
1378 * for the current block. The nodes of the graph are
1379 * c->optimum_nodes[0...block_size]. They correspond directly to the bytes in
1380 * the current block, plus one extra node for end-of-block. The edges of the
1381 * graph are matches and literals. The goal is to find the minimum cost path
1382 * from 'c->optimum_nodes[0]' to 'c->optimum_nodes[block_size]'.
1384 * The algorithm works forwards, starting at 'c->optimum_nodes[0]' and
1385 * proceeding forwards one node at a time. At each node, a selection of matches
1386 * (len >= 2), as well as the literal byte (len = 1), is considered. An item of
1387 * length 'len' provides a new path to reach the node 'len' bytes later. If
1388 * such a path is the lowest cost found so far to reach that later node, then
1389 * that later node is updated with the new path.
1391 * Note that although this algorithm is based on minimum cost path search, due
1392 * to various simplifying assumptions the result is not guaranteed to be the
1393 * true minimum cost, or "optimal", path over the graph of all valid LZX
1394 * representations of this block.
1396 * Also, note that because of the presence of the recent offsets queue (which is
1397 * a type of adaptive state), the algorithm cannot work backwards and compute
1398 * "cost to end" instead of "cost to beginning". Furthermore, the way the
1399 * algorithm handles this adaptive state in the "minimum cost" parse is actually
1400 * only an approximation. It's possible for the globally optimal, minimum cost
1401 * path to contain a prefix, ending at a position, where that path prefix is
1402 * *not* the minimum cost path to that position. This can happen if such a path
1403 * prefix results in a different adaptive state which results in lower costs
1404 * later. The algorithm does not solve this problem; it only considers the
1405 * lowest cost to reach each individual position.
1407 static inline struct lzx_lru_queue
1408 lzx_find_min_cost_path(struct lzx_compressor * const restrict c,
1409 const u8 * const restrict block_begin,
1410 const u32 block_size,
1411 const struct lzx_lru_queue initial_queue,
1414 struct lzx_optimum_node *cur_node = c->optimum_nodes;
1415 struct lzx_optimum_node * const end_node = &c->optimum_nodes[block_size];
1416 struct lz_match *cache_ptr = c->match_cache;
1417 const u8 *in_next = block_begin;
1418 const u8 * const block_end = block_begin + block_size;
1420 /* Instead of storing the match offset LRU queues in the
1421 * 'lzx_optimum_node' structures, we save memory (and cache lines) by
1422 * storing them in a smaller array. This works because the algorithm
1423 * only requires a limited history of the adaptive state. Once a given
1424 * state is more than LZX_MAX_MATCH_LEN bytes behind the current node,
1425 * it is no longer needed. */
1426 struct lzx_lru_queue queues[512];
1428 STATIC_ASSERT(ARRAY_LEN(queues) >= LZX_MAX_MATCH_LEN + 1);
1429 #define QUEUE(in) (queues[(uintptr_t)(in) % ARRAY_LEN(queues)])
1431 /* Initially, the cost to reach each node is "infinity". */
1432 memset(c->optimum_nodes, 0xFF,
1433 (block_size + 1) * sizeof(c->optimum_nodes[0]));
1435 QUEUE(block_begin) = initial_queue;
1437 /* The following loop runs 'block_size' iterations, one per node. */
1439 unsigned num_matches;
1444 * A selection of matches for the block was already saved in
1445 * memory so that we don't have to run the uncompressed data
1446 * through the matchfinder on every optimization pass. However,
1447 * we still search for repeat offset matches during each
1448 * optimization pass because we cannot predict the state of the
1449 * recent offsets queue. But as a heuristic, we don't bother
1450 * searching for repeat offset matches if the general-purpose
1451 * matchfinder failed to find any matches.
1453 * Note that a match of length n at some offset implies there is
1454 * also a match of length l for LZX_MIN_MATCH_LEN <= l <= n at
1455 * that same offset. In other words, we don't necessarily need
1456 * to use the full length of a match. The key heuristic that
1457 * saves a significicant amount of time is that for each
1458 * distinct length, we only consider the smallest offset for
1459 * which that length is available. This heuristic also applies
1460 * to repeat offsets, which we order specially: R0 < R1 < R2 <
1461 * any explicit offset. Of course, this heuristic may be
1462 * produce suboptimal results because offset slots in LZX are
1463 * subject to entropy encoding, but in practice this is a useful
1467 num_matches = cache_ptr->length;
1471 struct lz_match *end_matches = cache_ptr + num_matches;
1472 unsigned next_len = LZX_MIN_MATCH_LEN;
1473 unsigned max_len = min(block_end - in_next, LZX_MAX_MATCH_LEN);
1476 /* Consider R0 match */
1477 matchptr = in_next - lzx_lru_queue_R0(QUEUE(in_next));
1478 if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1480 STATIC_ASSERT(LZX_MIN_MATCH_LEN == 2);
1482 u32 cost = cur_node->cost +
1483 c->costs.match_cost[0][
1484 next_len - LZX_MIN_MATCH_LEN];
1485 if (cost <= (cur_node + next_len)->cost) {
1486 (cur_node + next_len)->cost = cost;
1487 (cur_node + next_len)->item =
1488 (0 << OPTIMUM_OFFSET_SHIFT) | next_len;
1490 if (unlikely(++next_len > max_len)) {
1491 cache_ptr = end_matches;
1494 } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1498 /* Consider R1 match */
1499 matchptr = in_next - lzx_lru_queue_R1(QUEUE(in_next));
1500 if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1502 if (matchptr[next_len - 1] != in_next[next_len - 1])
1504 for (unsigned len = 2; len < next_len - 1; len++)
1505 if (matchptr[len] != in_next[len])
1508 u32 cost = cur_node->cost +
1509 c->costs.match_cost[1][
1510 next_len - LZX_MIN_MATCH_LEN];
1511 if (cost <= (cur_node + next_len)->cost) {
1512 (cur_node + next_len)->cost = cost;
1513 (cur_node + next_len)->item =
1514 (1 << OPTIMUM_OFFSET_SHIFT) | next_len;
1516 if (unlikely(++next_len > max_len)) {
1517 cache_ptr = end_matches;
1520 } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1524 /* Consider R2 match */
1525 matchptr = in_next - lzx_lru_queue_R2(QUEUE(in_next));
1526 if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1528 if (matchptr[next_len - 1] != in_next[next_len - 1])
1530 for (unsigned len = 2; len < next_len - 1; len++)
1531 if (matchptr[len] != in_next[len])
1534 u32 cost = cur_node->cost +
1535 c->costs.match_cost[2][
1536 next_len - LZX_MIN_MATCH_LEN];
1537 if (cost <= (cur_node + next_len)->cost) {
1538 (cur_node + next_len)->cost = cost;
1539 (cur_node + next_len)->item =
1540 (2 << OPTIMUM_OFFSET_SHIFT) | next_len;
1542 if (unlikely(++next_len > max_len)) {
1543 cache_ptr = end_matches;
1546 } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1550 while (next_len > cache_ptr->length)
1551 if (++cache_ptr == end_matches)
1554 /* Consider explicit offset matches */
1556 u32 offset = cache_ptr->offset;
1557 u32 offset_data = offset + LZX_OFFSET_ADJUSTMENT;
1558 unsigned offset_slot = lzx_comp_get_offset_slot(c, offset_data,
1561 u32 cost = cur_node->cost +
1562 c->costs.match_cost[offset_slot][
1563 next_len - LZX_MIN_MATCH_LEN];
1564 #if LZX_CONSIDER_ALIGNED_COSTS
1565 if (lzx_extra_offset_bits[offset_slot] >=
1566 LZX_NUM_ALIGNED_OFFSET_BITS)
1567 cost += c->costs.aligned[offset_data &
1568 LZX_ALIGNED_OFFSET_BITMASK];
1570 if (cost < (cur_node + next_len)->cost) {
1571 (cur_node + next_len)->cost = cost;
1572 (cur_node + next_len)->item =
1573 (offset_data << OPTIMUM_OFFSET_SHIFT) | next_len;
1575 } while (++next_len <= cache_ptr->length);
1576 } while (++cache_ptr != end_matches);
1581 /* Consider coding a literal.
1583 * To avoid an extra branch, actually checking the preferability
1584 * of coding the literal is integrated into the queue update
1586 literal = *in_next++;
1587 cost = cur_node->cost +
1588 c->costs.main[lzx_main_symbol_for_literal(literal)];
1590 /* Advance to the next position. */
1593 /* The lowest-cost path to the current position is now known.
1594 * Finalize the recent offsets queue that results from taking
1595 * this lowest-cost path. */
1597 if (cost <= cur_node->cost) {
1598 /* Literal: queue remains unchanged. */
1599 cur_node->cost = cost;
1600 cur_node->item = (u32)literal << OPTIMUM_OFFSET_SHIFT;
1601 QUEUE(in_next) = QUEUE(in_next - 1);
1603 /* Match: queue update is needed. */
1604 unsigned len = cur_node->item & OPTIMUM_LEN_MASK;
1605 u32 offset_data = cur_node->item >> OPTIMUM_OFFSET_SHIFT;
1606 if (offset_data >= LZX_NUM_RECENT_OFFSETS) {
1607 /* Explicit offset match: insert offset at front */
1609 lzx_lru_queue_push(QUEUE(in_next - len),
1610 offset_data - LZX_OFFSET_ADJUSTMENT);
1612 /* Repeat offset match: swap offset to front */
1614 lzx_lru_queue_swap(QUEUE(in_next - len),
1618 } while (cur_node != end_node);
1620 /* Return the match offset queue at the end of the minimum cost path. */
1621 return QUEUE(block_end);
1624 /* Given the costs for the main and length codewords, compute 'match_costs'. */
1626 lzx_compute_match_costs(struct lzx_compressor *c)
1628 unsigned num_offset_slots = lzx_get_num_offset_slots(c->window_order);
1629 struct lzx_costs *costs = &c->costs;
1631 for (unsigned offset_slot = 0; offset_slot < num_offset_slots; offset_slot++) {
1633 u32 extra_cost = (u32)lzx_extra_offset_bits[offset_slot] * LZX_BIT_COST;
1634 unsigned main_symbol = lzx_main_symbol_for_match(offset_slot, 0);
1637 #if LZX_CONSIDER_ALIGNED_COSTS
1638 if (lzx_extra_offset_bits[offset_slot] >= LZX_NUM_ALIGNED_OFFSET_BITS)
1639 extra_cost -= LZX_NUM_ALIGNED_OFFSET_BITS * LZX_BIT_COST;
1642 for (i = 0; i < LZX_NUM_PRIMARY_LENS; i++)
1643 costs->match_cost[offset_slot][i] =
1644 costs->main[main_symbol++] + extra_cost;
1646 extra_cost += costs->main[main_symbol];
1648 for (; i < LZX_NUM_LENS; i++)
1649 costs->match_cost[offset_slot][i] =
1650 costs->len[i - LZX_NUM_PRIMARY_LENS] + extra_cost;
1654 /* Set default LZX Huffman symbol costs to bootstrap the iterative optimization
1657 lzx_set_default_costs(struct lzx_compressor *c, const u8 *block, u32 block_size)
1660 bool have_byte[256];
1661 unsigned num_used_bytes;
1663 /* The costs below are hard coded to use a scaling factor of 16. */
1664 STATIC_ASSERT(LZX_BIT_COST == 16);
1669 * - Use smaller initial costs for literal symbols when the input buffer
1670 * contains fewer distinct bytes.
1672 * - Assume that match symbols are more costly than literal symbols.
1674 * - Assume that length symbols for shorter lengths are less costly than
1675 * length symbols for longer lengths.
1678 for (i = 0; i < 256; i++)
1679 have_byte[i] = false;
1681 for (i = 0; i < block_size; i++)
1682 have_byte[block[i]] = true;
1685 for (i = 0; i < 256; i++)
1686 num_used_bytes += have_byte[i];
1688 for (i = 0; i < 256; i++)
1689 c->costs.main[i] = 140 - (256 - num_used_bytes) / 4;
1691 for (; i < c->num_main_syms; i++)
1692 c->costs.main[i] = 170;
1694 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
1695 c->costs.len[i] = 103 + (i / 4);
1697 #if LZX_CONSIDER_ALIGNED_COSTS
1698 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
1699 c->costs.aligned[i] = LZX_NUM_ALIGNED_OFFSET_BITS * LZX_BIT_COST;
1702 lzx_compute_match_costs(c);
1705 /* Update the current cost model to reflect the computed Huffman codes. */
1707 lzx_update_costs(struct lzx_compressor *c)
1710 const struct lzx_lens *lens = &c->codes[c->codes_index].lens;
1712 for (i = 0; i < c->num_main_syms; i++)
1713 c->costs.main[i] = (lens->main[i] ? lens->main[i] : 15) * LZX_BIT_COST;
1715 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
1716 c->costs.len[i] = (lens->len[i] ? lens->len[i] : 15) * LZX_BIT_COST;
1718 #if LZX_CONSIDER_ALIGNED_COSTS
1719 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
1720 c->costs.aligned[i] = (lens->aligned[i] ? lens->aligned[i] : 7) * LZX_BIT_COST;
1723 lzx_compute_match_costs(c);
1726 static inline struct lzx_lru_queue
1727 lzx_optimize_and_write_block(struct lzx_compressor * const restrict c,
1728 struct lzx_output_bitstream * const restrict os,
1729 const u8 * const restrict block_begin,
1730 const u32 block_size,
1731 const struct lzx_lru_queue initial_queue,
1734 unsigned num_passes_remaining = c->num_optim_passes;
1735 struct lzx_lru_queue new_queue;
1738 /* The first optimization pass uses a default cost model. Each
1739 * additional optimization pass uses a cost model derived from the
1740 * Huffman code computed in the previous pass. */
1742 lzx_set_default_costs(c, block_begin, block_size);
1743 lzx_reset_symbol_frequencies(c);
1745 new_queue = lzx_find_min_cost_path(c, block_begin, block_size,
1746 initial_queue, is_16_bit);
1747 if (num_passes_remaining > 1) {
1748 lzx_tally_item_list(c, block_size, is_16_bit);
1749 lzx_make_huffman_codes(c);
1750 lzx_update_costs(c);
1751 lzx_reset_symbol_frequencies(c);
1753 } while (--num_passes_remaining);
1755 seq_idx = lzx_record_item_list(c, block_size, is_16_bit);
1756 lzx_finish_block(c, os, block_begin, block_size, seq_idx);
1761 * This is the "near-optimal" LZX compressor.
1763 * For each block, it performs a relatively thorough graph search to find an
1764 * inexpensive (in terms of compressed size) way to output that block.
1766 * Note: there are actually many things this algorithm leaves on the table in
1767 * terms of compression ratio. So although it may be "near-optimal", it is
1768 * certainly not "optimal". The goal is not to produce the optimal compression
1769 * ratio, which for LZX is probably impossible within any practical amount of
1770 * time, but rather to produce a compression ratio significantly better than a
1771 * simpler "greedy" or "lazy" parse while still being relatively fast.
1774 lzx_compress_near_optimal(struct lzx_compressor *c,
1775 struct lzx_output_bitstream *os,
1778 const u8 * const in_begin = c->in_buffer;
1779 const u8 * in_next = in_begin;
1780 const u8 * const in_end = in_begin + c->in_nbytes;
1781 unsigned max_len = LZX_MAX_MATCH_LEN;
1782 unsigned nice_len = min(c->nice_match_length, max_len);
1784 struct lzx_lru_queue queue;
1786 CALL_BT_MF(is_16_bit, c, bt_matchfinder_init);
1787 lzx_lru_queue_init(&queue);
1790 /* Starting a new block */
1791 const u8 * const in_block_begin = in_next;
1792 const u8 * const in_block_end =
1793 in_next + min(LZX_DIV_BLOCK_SIZE, in_end - in_next);
1795 /* Run the block through the matchfinder and cache the matches. */
1796 struct lz_match *cache_ptr = c->match_cache;
1798 struct lz_match *lz_matchptr;
1801 /* If approaching the end of the input buffer, adjust
1802 * 'max_len' and 'nice_len' accordingly. */
1803 if (unlikely(max_len > in_end - in_next)) {
1804 max_len = in_end - in_next;
1805 nice_len = min(max_len, nice_len);
1807 /* This extra check is needed to ensure that we
1808 * never output a length 2 match of the very
1809 * last two bytes with the very first two bytes,
1810 * since such a match has an offset too large to
1811 * be represented. */
1812 if (unlikely(max_len < 3)) {
1814 cache_ptr->length = 0;
1820 /* Check for matches. */
1821 lz_matchptr = CALL_BT_MF(is_16_bit, c, bt_matchfinder_get_matches,
1826 c->max_search_depth,
1831 cache_ptr->length = lz_matchptr - (cache_ptr + 1);
1832 cache_ptr = lz_matchptr;
1835 * If there was a very long match found, then don't
1836 * cache any matches for the bytes covered by that
1837 * match. This avoids degenerate behavior when
1838 * compressing highly redundant data, where the number
1839 * of matches can be very large.
1841 * This heuristic doesn't actually hurt the compression
1842 * ratio very much. If there's a long match, then the
1843 * data must be highly compressible, so it doesn't
1844 * matter as much what we do.
1846 if (best_len >= nice_len) {
1849 if (unlikely(max_len > in_end - in_next)) {
1850 max_len = in_end - in_next;
1851 nice_len = min(max_len, nice_len);
1852 if (unlikely(max_len < 3)) {
1854 cache_ptr->length = 0;
1859 CALL_BT_MF(is_16_bit, c, bt_matchfinder_skip_position,
1864 c->max_search_depth,
1867 cache_ptr->length = 0;
1869 } while (--best_len);
1871 } while (in_next < in_block_end &&
1872 likely(cache_ptr < &c->match_cache[LZX_CACHE_LENGTH]));
1874 /* We've finished running the block through the matchfinder.
1875 * Now choose a match/literal sequence and write the block. */
1877 queue = lzx_optimize_and_write_block(c, os, in_block_begin,
1878 in_next - in_block_begin,
1880 } while (in_next != in_end);
1884 lzx_compress_near_optimal_16(struct lzx_compressor *c,
1885 struct lzx_output_bitstream *os)
1887 lzx_compress_near_optimal(c, os, true);
1891 lzx_compress_near_optimal_32(struct lzx_compressor *c,
1892 struct lzx_output_bitstream *os)
1894 lzx_compress_near_optimal(c, os, false);
1898 * Given a pointer to the current byte sequence and the current list of recent
1899 * match offsets, find the longest repeat offset match.
1901 * If no match of at least 2 bytes is found, then return 0.
1903 * If a match of at least 2 bytes is found, then return its length and set
1904 * *rep_max_idx_ret to the index of its offset in @queue.
1907 lzx_find_longest_repeat_offset_match(const u8 * const in_next,
1908 const u32 bytes_remaining,
1909 const u32 recent_offsets[LZX_NUM_RECENT_OFFSETS],
1910 unsigned *rep_max_idx_ret)
1912 STATIC_ASSERT(LZX_NUM_RECENT_OFFSETS == 3);
1914 const unsigned max_len = min(bytes_remaining, LZX_MAX_MATCH_LEN);
1915 const u16 next_2_bytes = load_u16_unaligned(in_next);
1917 unsigned rep_max_len;
1918 unsigned rep_max_idx;
1921 matchptr = in_next - recent_offsets[0];
1922 if (load_u16_unaligned(matchptr) == next_2_bytes)
1923 rep_max_len = lz_extend(in_next, matchptr, 2, max_len);
1928 matchptr = in_next - recent_offsets[1];
1929 if (load_u16_unaligned(matchptr) == next_2_bytes) {
1930 rep_len = lz_extend(in_next, matchptr, 2, max_len);
1931 if (rep_len > rep_max_len) {
1932 rep_max_len = rep_len;
1937 matchptr = in_next - recent_offsets[2];
1938 if (load_u16_unaligned(matchptr) == next_2_bytes) {
1939 rep_len = lz_extend(in_next, matchptr, 2, max_len);
1940 if (rep_len > rep_max_len) {
1941 rep_max_len = rep_len;
1946 *rep_max_idx_ret = rep_max_idx;
1950 /* Fast heuristic scoring for lazy parsing: how "good" is this match? */
1951 static inline unsigned
1952 lzx_explicit_offset_match_score(unsigned len, u32 adjusted_offset)
1954 unsigned score = len;
1956 if (adjusted_offset < 4096)
1959 if (adjusted_offset < 256)
1965 static inline unsigned
1966 lzx_repeat_offset_match_score(unsigned rep_len, unsigned rep_idx)
1971 /* This is the "lazy" LZX compressor. */
1973 lzx_compress_lazy(struct lzx_compressor *c, struct lzx_output_bitstream *os,
1976 const u8 * const in_begin = c->in_buffer;
1977 const u8 * in_next = in_begin;
1978 const u8 * const in_end = in_begin + c->in_nbytes;
1979 unsigned max_len = LZX_MAX_MATCH_LEN;
1980 unsigned nice_len = min(c->nice_match_length, max_len);
1981 STATIC_ASSERT(LZX_NUM_RECENT_OFFSETS == 3);
1982 u32 recent_offsets[3] = {1, 1, 1};
1983 u32 next_hashes[2] = {};
1985 CALL_HC_MF(is_16_bit, c, hc_matchfinder_init);
1988 /* Starting a new block */
1990 const u8 * const in_block_begin = in_next;
1991 const u8 * const in_block_end =
1992 in_next + min(LZX_DIV_BLOCK_SIZE, in_end - in_next);
1993 struct lzx_sequence *next_seq = c->chosen_sequences;
1996 u32 cur_offset_data;
2000 u32 next_offset_data;
2001 unsigned next_score;
2002 unsigned rep_max_len;
2003 unsigned rep_max_idx;
2008 lzx_reset_symbol_frequencies(c);
2011 if (unlikely(max_len > in_end - in_next)) {
2012 max_len = in_end - in_next;
2013 nice_len = min(max_len, nice_len);
2016 /* Find the longest match at the current position. */
2018 cur_len = CALL_HC_MF(is_16_bit, c, hc_matchfinder_longest_match,
2024 c->max_search_depth,
2029 cur_offset >= 8192 - LZX_OFFSET_ADJUSTMENT &&
2030 cur_offset != recent_offsets[0] &&
2031 cur_offset != recent_offsets[1] &&
2032 cur_offset != recent_offsets[2]))
2034 /* There was no match found, or the only match found
2035 * was a distant length 3 match. Output a literal. */
2036 lzx_record_literal(c, *in_next++, &litrunlen);
2040 if (cur_offset == recent_offsets[0]) {
2042 cur_offset_data = 0;
2043 skip_len = cur_len - 1;
2044 goto choose_cur_match;
2047 cur_offset_data = cur_offset + LZX_OFFSET_ADJUSTMENT;
2048 cur_score = lzx_explicit_offset_match_score(cur_len, cur_offset_data);
2050 /* Consider a repeat offset match */
2051 rep_max_len = lzx_find_longest_repeat_offset_match(in_next,
2057 if (rep_max_len >= 3 &&
2058 (rep_score = lzx_repeat_offset_match_score(rep_max_len,
2059 rep_max_idx)) >= cur_score)
2061 cur_len = rep_max_len;
2062 cur_offset_data = rep_max_idx;
2063 skip_len = rep_max_len - 1;
2064 goto choose_cur_match;
2069 /* We have a match at the current position. */
2071 /* If we have a very long match, choose it immediately. */
2072 if (cur_len >= nice_len) {
2073 skip_len = cur_len - 1;
2074 goto choose_cur_match;
2077 /* See if there's a better match at the next position. */
2079 if (unlikely(max_len > in_end - in_next)) {
2080 max_len = in_end - in_next;
2081 nice_len = min(max_len, nice_len);
2084 next_len = CALL_HC_MF(is_16_bit, c, hc_matchfinder_longest_match,
2090 c->max_search_depth / 2,
2094 if (next_len <= cur_len - 2) {
2096 skip_len = cur_len - 2;
2097 goto choose_cur_match;
2100 next_offset_data = next_offset + LZX_OFFSET_ADJUSTMENT;
2101 next_score = lzx_explicit_offset_match_score(next_len, next_offset_data);
2103 rep_max_len = lzx_find_longest_repeat_offset_match(in_next,
2109 if (rep_max_len >= 3 &&
2110 (rep_score = lzx_repeat_offset_match_score(rep_max_len,
2111 rep_max_idx)) >= next_score)
2114 if (rep_score > cur_score) {
2115 /* The next match is better, and it's a
2116 * repeat offset match. */
2117 lzx_record_literal(c, *(in_next - 2),
2119 cur_len = rep_max_len;
2120 cur_offset_data = rep_max_idx;
2121 skip_len = cur_len - 1;
2122 goto choose_cur_match;
2125 if (next_score > cur_score) {
2126 /* The next match is better, and it's an
2127 * explicit offset match. */
2128 lzx_record_literal(c, *(in_next - 2),
2131 cur_offset_data = next_offset_data;
2132 cur_score = next_score;
2133 goto have_cur_match;
2137 /* The original match was better. */
2138 skip_len = cur_len - 2;
2141 lzx_record_match(c, cur_len, cur_offset_data,
2142 recent_offsets, is_16_bit,
2143 &litrunlen, &next_seq);
2144 in_next = CALL_HC_MF(is_16_bit, c, hc_matchfinder_skip_positions,
2150 } while (in_next < in_block_end);
2152 lzx_finish_sequence(next_seq, litrunlen);
2154 lzx_finish_block(c, os, in_block_begin, in_next - in_block_begin, 0);
2156 } while (in_next != in_end);
2160 lzx_compress_lazy_16(struct lzx_compressor *c, struct lzx_output_bitstream *os)
2162 lzx_compress_lazy(c, os, true);
2166 lzx_compress_lazy_32(struct lzx_compressor *c, struct lzx_output_bitstream *os)
2168 lzx_compress_lazy(c, os, false);
2171 /* Generate the acceleration tables for offset slots. */
2173 lzx_init_offset_slot_tabs(struct lzx_compressor *c)
2175 u32 adjusted_offset = 0;
2179 for (; adjusted_offset < ARRAY_LEN(c->offset_slot_tab_1);
2182 if (adjusted_offset >= lzx_offset_slot_base[slot + 1])
2184 c->offset_slot_tab_1[adjusted_offset] = slot;
2187 /* slots [30, 49] */
2188 for (; adjusted_offset < LZX_MAX_WINDOW_SIZE;
2189 adjusted_offset += (u32)1 << 14)
2191 if (adjusted_offset >= lzx_offset_slot_base[slot + 1])
2193 c->offset_slot_tab_2[adjusted_offset >> 14] = slot;
2198 lzx_get_compressor_size(size_t max_bufsize, unsigned compression_level)
2200 if (compression_level <= LZX_MAX_FAST_LEVEL) {
2201 if (lzx_is_16_bit(max_bufsize))
2202 return offsetof(struct lzx_compressor, hc_mf_16) +
2203 hc_matchfinder_size_16(max_bufsize);
2205 return offsetof(struct lzx_compressor, hc_mf_32) +
2206 hc_matchfinder_size_32(max_bufsize);
2208 if (lzx_is_16_bit(max_bufsize))
2209 return offsetof(struct lzx_compressor, bt_mf_16) +
2210 bt_matchfinder_size_16(max_bufsize);
2212 return offsetof(struct lzx_compressor, bt_mf_32) +
2213 bt_matchfinder_size_32(max_bufsize);
2218 lzx_get_needed_memory(size_t max_bufsize, unsigned compression_level,
2223 if (max_bufsize > LZX_MAX_WINDOW_SIZE)
2226 size += lzx_get_compressor_size(max_bufsize, compression_level);
2228 size += max_bufsize; /* in_buffer */
2233 lzx_create_compressor(size_t max_bufsize, unsigned compression_level,
2234 bool destructive, void **c_ret)
2236 unsigned window_order;
2237 struct lzx_compressor *c;
2239 window_order = lzx_get_window_order(max_bufsize);
2240 if (window_order == 0)
2241 return WIMLIB_ERR_INVALID_PARAM;
2243 c = MALLOC(lzx_get_compressor_size(max_bufsize, compression_level));
2247 c->destructive = destructive;
2249 c->num_main_syms = lzx_get_num_main_syms(window_order);
2250 c->window_order = window_order;
2252 if (!c->destructive) {
2253 c->in_buffer = MALLOC(max_bufsize);
2258 if (compression_level <= LZX_MAX_FAST_LEVEL) {
2260 /* Fast compression: Use lazy parsing. */
2262 if (lzx_is_16_bit(max_bufsize))
2263 c->impl = lzx_compress_lazy_16;
2265 c->impl = lzx_compress_lazy_32;
2266 c->max_search_depth = (36 * compression_level) / 20;
2267 c->nice_match_length = (72 * compression_level) / 20;
2269 /* lzx_compress_lazy() needs max_search_depth >= 2 because it
2270 * halves the max_search_depth when attempting a lazy match, and
2271 * max_search_depth cannot be 0. */
2272 if (c->max_search_depth < 2)
2273 c->max_search_depth = 2;
2276 /* Normal / high compression: Use near-optimal parsing. */
2278 if (lzx_is_16_bit(max_bufsize))
2279 c->impl = lzx_compress_near_optimal_16;
2281 c->impl = lzx_compress_near_optimal_32;
2283 /* Scale nice_match_length and max_search_depth with the
2284 * compression level. */
2285 c->max_search_depth = (24 * compression_level) / 50;
2286 c->nice_match_length = (32 * compression_level) / 50;
2288 /* Set a number of optimization passes appropriate for the
2289 * compression level. */
2291 c->num_optim_passes = 1;
2293 if (compression_level >= 45)
2294 c->num_optim_passes++;
2296 /* Use more optimization passes for higher compression levels.
2297 * But the more passes there are, the less they help --- so
2298 * don't add them linearly. */
2299 if (compression_level >= 70) {
2300 c->num_optim_passes++;
2301 if (compression_level >= 100)
2302 c->num_optim_passes++;
2303 if (compression_level >= 150)
2304 c->num_optim_passes++;
2305 if (compression_level >= 200)
2306 c->num_optim_passes++;
2307 if (compression_level >= 300)
2308 c->num_optim_passes++;
2312 /* max_search_depth == 0 is invalid. */
2313 if (c->max_search_depth < 1)
2314 c->max_search_depth = 1;
2316 if (c->nice_match_length > LZX_MAX_MATCH_LEN)
2317 c->nice_match_length = LZX_MAX_MATCH_LEN;
2319 lzx_init_offset_slot_tabs(c);
2326 return WIMLIB_ERR_NOMEM;
2330 lzx_compress(const void *restrict in, size_t in_nbytes,
2331 void *restrict out, size_t out_nbytes_avail, void *restrict _c)
2333 struct lzx_compressor *c = _c;
2334 struct lzx_output_bitstream os;
2337 /* Don't bother trying to compress very small inputs. */
2338 if (in_nbytes < 100)
2341 /* Copy the input data into the internal buffer and preprocess it. */
2343 c->in_buffer = (void *)in;
2345 memcpy(c->in_buffer, in, in_nbytes);
2346 c->in_nbytes = in_nbytes;
2347 lzx_do_e8_preprocessing(c->in_buffer, in_nbytes);
2349 /* Initially, the previous Huffman codeword lengths are all zeroes. */
2351 memset(&c->codes[1].lens, 0, sizeof(struct lzx_lens));
2353 /* Initialize the output bitstream. */
2354 lzx_init_output(&os, out, out_nbytes_avail);
2356 /* Call the compression level-specific compress() function. */
2359 /* Flush the output bitstream and return the compressed size or 0. */
2360 result = lzx_flush_output(&os);
2361 if (!result && c->destructive)
2362 lzx_undo_e8_preprocessing(c->in_buffer, c->in_nbytes);
2367 lzx_free_compressor(void *_c)
2369 struct lzx_compressor *c = _c;
2371 if (!c->destructive)
2376 const struct compressor_ops lzx_compressor_ops = {
2377 .get_needed_memory = lzx_get_needed_memory,
2378 .create_compressor = lzx_create_compressor,
2379 .compress = lzx_compress,
2380 .free_compressor = lzx_free_compressor,