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 * Should the compressor take into account the costs of aligned offset symbols?
119 #define LZX_CONSIDER_ALIGNED_COSTS 1
122 * LZX_MAX_FAST_LEVEL is the maximum compression level at which we use the
125 #define LZX_MAX_FAST_LEVEL 34
128 * BT_MATCHFINDER_HASH2_ORDER is the log base 2 of the number of entries in the
129 * hash table for finding length 2 matches. This could be as high as 16, but
130 * using a smaller hash table speeds up compression due to reduced cache
133 #define BT_MATCHFINDER_HASH2_ORDER 12
136 * These are the compressor-side limits on the codeword lengths for each Huffman
137 * code. To make outputting bits slightly faster, some of these limits are
138 * lower than the limits defined by the LZX format. This does not significantly
139 * affect the compression ratio, at least for the block sizes we use.
141 #define MAIN_CODEWORD_LIMIT 12 /* 64-bit: can buffer 4 main symbols */
142 #define LENGTH_CODEWORD_LIMIT 12
143 #define ALIGNED_CODEWORD_LIMIT 7
144 #define PRE_CODEWORD_LIMIT 7
146 #include "wimlib/compress_common.h"
147 #include "wimlib/compressor_ops.h"
148 #include "wimlib/error.h"
149 #include "wimlib/lz_extend.h"
150 #include "wimlib/lzx_common.h"
151 #include "wimlib/unaligned.h"
152 #include "wimlib/util.h"
154 /* Matchfinders with 16-bit positions */
156 #define MF_SUFFIX _16
157 #include "wimlib/bt_matchfinder.h"
158 #include "wimlib/hc_matchfinder.h"
160 /* Matchfinders with 32-bit positions */
164 #define MF_SUFFIX _32
165 #include "wimlib/bt_matchfinder.h"
166 #include "wimlib/hc_matchfinder.h"
168 struct lzx_output_bitstream;
170 /* Codewords for the LZX Huffman codes. */
171 struct lzx_codewords {
172 u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
173 u32 len[LZX_LENCODE_NUM_SYMBOLS];
174 u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
177 /* Codeword lengths (in bits) for the LZX Huffman codes.
178 * A zero length means the corresponding codeword has zero frequency. */
180 u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS + 1];
181 u8 len[LZX_LENCODE_NUM_SYMBOLS + 1];
182 u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
185 /* Cost model for near-optimal parsing */
188 /* 'match_cost[offset_slot][len - LZX_MIN_MATCH_LEN]' is the cost for a
189 * length 'len' match that has an offset belonging to 'offset_slot'. */
190 u32 match_cost[LZX_MAX_OFFSET_SLOTS][LZX_NUM_LENS];
192 /* Cost for each symbol in the main code */
193 u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
195 /* Cost for each symbol in the length code */
196 u32 len[LZX_LENCODE_NUM_SYMBOLS];
198 #if LZX_CONSIDER_ALIGNED_COSTS
199 /* Cost for each symbol in the aligned code */
200 u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
204 /* Codewords and lengths for the LZX Huffman codes. */
206 struct lzx_codewords codewords;
207 struct lzx_lens lens;
210 /* Symbol frequency counters for the LZX Huffman codes. */
212 u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
213 u32 len[LZX_LENCODE_NUM_SYMBOLS];
214 u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
218 * Represents a run of literals followed by a match or end-of-block. This
219 * struct is needed to temporarily store items chosen by the parser, since items
220 * cannot be written until all items for the block have been chosen and the
221 * block's Huffman codes have been computed.
223 struct lzx_sequence {
225 /* The number of literals in the run. This may be 0. The literals are
226 * not stored explicitly in this structure; instead, they are read
227 * directly from the uncompressed data. */
230 /* If the next field doesn't indicate end-of-block, then this is the
231 * match length minus LZX_MIN_MATCH_LEN. */
234 /* If bit 31 is clear, then this field contains the match header in bits
235 * 0-8 and the match offset minus LZX_OFFSET_ADJUSTMENT in bits 9-30.
236 * Otherwise, this sequence's literal run was the last literal run in
237 * the block, so there is no match that follows it. */
238 u32 adjusted_offset_and_match_hdr;
242 * This structure represents a byte position in the input buffer and a node in
243 * the graph of possible match/literal choices.
245 * Logically, each incoming edge to this node is labeled with a literal or a
246 * match that can be taken to reach this position from an earlier position; and
247 * each outgoing edge from this node is labeled with a literal or a match that
248 * can be taken to advance from this position to a later position.
250 struct lzx_optimum_node {
252 /* The cost, in bits, of the lowest-cost path that has been found to
253 * reach this position. This can change as progressively lower cost
254 * paths are found to reach this position. */
258 * The match or literal that was taken to reach this position. This can
259 * change as progressively lower cost paths are found to reach this
262 * This variable is divided into two bitfields.
265 * Low bits are 0, high bits are the literal.
267 * Explicit offset matches:
268 * Low bits are the match length, high bits are the offset plus 2.
270 * Repeat offset matches:
271 * Low bits are the match length, high bits are the queue index.
274 #define OPTIMUM_OFFSET_SHIFT 9
275 #define OPTIMUM_LEN_MASK ((1 << OPTIMUM_OFFSET_SHIFT) - 1)
276 } _aligned_attribute(8);
279 * Least-recently-used queue for match offsets.
281 * This is represented as a 64-bit integer for efficiency. There are three
282 * offsets of 21 bits each. Bit 64 is garbage.
284 struct lzx_lru_queue {
288 #define LZX_QUEUE64_OFFSET_SHIFT 21
289 #define LZX_QUEUE64_OFFSET_MASK (((u64)1 << LZX_QUEUE64_OFFSET_SHIFT) - 1)
291 #define LZX_QUEUE64_R0_SHIFT (0 * LZX_QUEUE64_OFFSET_SHIFT)
292 #define LZX_QUEUE64_R1_SHIFT (1 * LZX_QUEUE64_OFFSET_SHIFT)
293 #define LZX_QUEUE64_R2_SHIFT (2 * LZX_QUEUE64_OFFSET_SHIFT)
295 #define LZX_QUEUE64_R0_MASK (LZX_QUEUE64_OFFSET_MASK << LZX_QUEUE64_R0_SHIFT)
296 #define LZX_QUEUE64_R1_MASK (LZX_QUEUE64_OFFSET_MASK << LZX_QUEUE64_R1_SHIFT)
297 #define LZX_QUEUE64_R2_MASK (LZX_QUEUE64_OFFSET_MASK << LZX_QUEUE64_R2_SHIFT)
300 lzx_lru_queue_init(struct lzx_lru_queue *queue)
302 queue->R = ((u64)1 << LZX_QUEUE64_R0_SHIFT) |
303 ((u64)1 << LZX_QUEUE64_R1_SHIFT) |
304 ((u64)1 << LZX_QUEUE64_R2_SHIFT);
308 lzx_lru_queue_R0(struct lzx_lru_queue queue)
310 return (queue.R >> LZX_QUEUE64_R0_SHIFT) & LZX_QUEUE64_OFFSET_MASK;
314 lzx_lru_queue_R1(struct lzx_lru_queue queue)
316 return (queue.R >> LZX_QUEUE64_R1_SHIFT) & LZX_QUEUE64_OFFSET_MASK;
320 lzx_lru_queue_R2(struct lzx_lru_queue queue)
322 return (queue.R >> LZX_QUEUE64_R2_SHIFT) & LZX_QUEUE64_OFFSET_MASK;
325 /* Push a match offset onto the front (most recently used) end of the queue. */
326 static inline struct lzx_lru_queue
327 lzx_lru_queue_push(struct lzx_lru_queue queue, u32 offset)
329 return (struct lzx_lru_queue) {
330 .R = (queue.R << LZX_QUEUE64_OFFSET_SHIFT) | offset,
334 /* Pop a match offset off the front (most recently used) end of the queue. */
336 lzx_lru_queue_pop(struct lzx_lru_queue *queue_p)
338 u32 offset = queue_p->R & LZX_QUEUE64_OFFSET_MASK;
339 queue_p->R >>= LZX_QUEUE64_OFFSET_SHIFT;
343 /* Swap a match offset to the front of the queue. */
344 static inline struct lzx_lru_queue
345 lzx_lru_queue_swap(struct lzx_lru_queue queue, unsigned idx)
351 return (struct lzx_lru_queue) {
352 .R = (lzx_lru_queue_R1(queue) << LZX_QUEUE64_R0_SHIFT) |
353 (lzx_lru_queue_R0(queue) << LZX_QUEUE64_R1_SHIFT) |
354 (queue.R & LZX_QUEUE64_R2_MASK),
357 return (struct lzx_lru_queue) {
358 .R = (lzx_lru_queue_R2(queue) << LZX_QUEUE64_R0_SHIFT) |
359 (queue.R & LZX_QUEUE64_R1_MASK) |
360 (lzx_lru_queue_R0(queue) << LZX_QUEUE64_R2_SHIFT),
364 /* The main LZX compressor structure */
365 struct lzx_compressor {
367 /* The "nice" match length: if a match of this length is found, then
368 * choose it immediately without further consideration. */
369 unsigned nice_match_length;
371 /* The maximum search depth: consider at most this many potential
372 * matches at each position. */
373 unsigned max_search_depth;
375 /* The log base 2 of the LZX window size for LZ match offset encoding
376 * purposes. This will be >= LZX_MIN_WINDOW_ORDER and <=
377 * LZX_MAX_WINDOW_ORDER. */
378 unsigned window_order;
380 /* The number of symbols in the main alphabet. This depends on
381 * @window_order, since @window_order determines the maximum possible
383 unsigned num_main_syms;
385 /* Number of optimization passes per block */
386 unsigned num_optim_passes;
388 /* The preprocessed buffer of data being compressed */
391 /* The number of bytes of data to be compressed, which is the number of
392 * bytes of data in @in_buffer that are actually valid. */
395 /* Pointer to the compress() implementation chosen at allocation time */
396 void (*impl)(struct lzx_compressor *, struct lzx_output_bitstream *);
398 /* If true, the compressor need not preserve the input buffer if it
399 * compresses the data successfully. */
402 /* The Huffman symbol frequency counters for the current block. */
403 struct lzx_freqs freqs;
405 /* The Huffman codes for the current and previous blocks. The one with
406 * index 'codes_index' is for the current block, and the other one is
407 * for the previous block. */
408 struct lzx_codes codes[2];
409 unsigned codes_index;
411 /* The matches and literals that the parser has chosen for the current
412 * block. The required length of this array is limited by the maximum
413 * number of matches that can ever be chosen for a single block, plus
414 * one for the special entry at the end. */
415 struct lzx_sequence chosen_sequences[
416 DIV_ROUND_UP(LZX_DIV_BLOCK_SIZE, LZX_MIN_MATCH_LEN) + 1];
418 /* Tables for mapping adjusted offsets to offset slots */
420 /* offset slots [0, 29] */
421 u8 offset_slot_tab_1[32768];
423 /* offset slots [30, 49] */
424 u8 offset_slot_tab_2[128];
427 /* Data for greedy or lazy parsing */
429 /* Hash chains matchfinder (MUST BE LAST!!!) */
431 struct hc_matchfinder_16 hc_mf_16;
432 struct hc_matchfinder_32 hc_mf_32;
436 /* Data for near-optimal parsing */
439 * The graph nodes for the current block.
441 * We need at least 'LZX_DIV_BLOCK_SIZE +
442 * LZX_MAX_MATCH_LEN - 1' nodes because that is the
443 * maximum block size that may be used. Add 1 because
444 * we need a node to represent end-of-block.
446 * It is possible that nodes past end-of-block are
447 * accessed during match consideration, but this can
448 * only occur if the block was truncated at
449 * LZX_DIV_BLOCK_SIZE. So the same bound still applies.
450 * Note that since nodes past the end of the block will
451 * never actually have an effect on the items that are
452 * chosen for the block, it makes no difference what
453 * their costs are initialized to (if anything).
455 struct lzx_optimum_node optimum_nodes[LZX_DIV_BLOCK_SIZE +
456 LZX_MAX_MATCH_LEN - 1 + 1];
458 /* The cost model for the current block */
459 struct lzx_costs costs;
462 * Cached matches for the current block. This array
463 * contains the matches that were found at each position
464 * in the block. Specifically, for each position, there
465 * is a special 'struct lz_match' whose 'length' field
466 * contains the number of matches that were found at
467 * that position; this is followed by the matches
468 * themselves, if any, sorted by strictly increasing
471 * Note: in rare cases, there will be a very high number
472 * of matches in the block and this array will overflow.
473 * If this happens, we force the end of the current
474 * block. LZX_CACHE_LENGTH is the length at which we
475 * actually check for overflow. The extra slots beyond
476 * this are enough to absorb the worst case overflow,
477 * which occurs if starting at
478 * &match_cache[LZX_CACHE_LENGTH - 1], we write the
479 * match count header, then write
480 * LZX_MAX_MATCHES_PER_POS matches, then skip searching
481 * for matches at 'LZX_MAX_MATCH_LEN - 1' positions and
482 * write the match count header for each.
484 struct lz_match match_cache[LZX_CACHE_LENGTH +
485 LZX_MAX_MATCHES_PER_POS +
486 LZX_MAX_MATCH_LEN - 1];
488 /* Binary trees matchfinder (MUST BE LAST!!!) */
490 struct bt_matchfinder_16 bt_mf_16;
491 struct bt_matchfinder_32 bt_mf_32;
498 * Will a matchfinder using 16-bit positions be sufficient for compressing
499 * buffers of up to the specified size? The limit could be 65536 bytes, but we
500 * also want to optimize out the use of offset_slot_tab_2 in the 16-bit case.
501 * This requires that the limit be no more than the length of offset_slot_tab_1
505 lzx_is_16_bit(size_t max_bufsize)
507 STATIC_ASSERT(ARRAY_LEN(((struct lzx_compressor *)0)->offset_slot_tab_1) == 32768);
508 return max_bufsize <= 32768;
512 * The following macros call either the 16-bit or the 32-bit version of a
513 * matchfinder function based on the value of 'is_16_bit', which will be known
514 * at compilation time.
517 #define CALL_HC_MF(is_16_bit, c, funcname, ...) \
518 ((is_16_bit) ? CONCAT(funcname, _16)(&(c)->hc_mf_16, ##__VA_ARGS__) : \
519 CONCAT(funcname, _32)(&(c)->hc_mf_32, ##__VA_ARGS__));
521 #define CALL_BT_MF(is_16_bit, c, funcname, ...) \
522 ((is_16_bit) ? CONCAT(funcname, _16)(&(c)->bt_mf_16, ##__VA_ARGS__) : \
523 CONCAT(funcname, _32)(&(c)->bt_mf_32, ##__VA_ARGS__));
526 * Structure to keep track of the current state of sending bits to the
527 * compressed output buffer.
529 * The LZX bitstream is encoded as a sequence of 16-bit coding units.
531 struct lzx_output_bitstream {
533 /* Bits that haven't yet been written to the output buffer. */
534 machine_word_t bitbuf;
536 /* Number of bits currently held in @bitbuf. */
539 /* Pointer to the start of the output buffer. */
542 /* Pointer to the position in the output buffer at which the next coding
543 * unit should be written. */
546 /* Pointer just past the end of the output buffer, rounded down to a
547 * 2-byte boundary. */
551 /* Can the specified number of bits always be added to 'bitbuf' after any
552 * pending 16-bit coding units have been flushed? */
553 #define CAN_BUFFER(n) ((n) <= (8 * sizeof(machine_word_t)) - 15)
556 * Initialize the output bitstream.
559 * The output bitstream structure to initialize.
561 * The buffer being written to.
563 * Size of @buffer, in bytes.
566 lzx_init_output(struct lzx_output_bitstream *os, void *buffer, size_t size)
571 os->next = os->start;
572 os->end = os->start + (size & ~1);
575 /* Add some bits to the bitbuffer variable of the output bitstream. The caller
576 * must make sure there is enough room. */
578 lzx_add_bits(struct lzx_output_bitstream *os, u32 bits, unsigned num_bits)
580 os->bitbuf = (os->bitbuf << num_bits) | bits;
581 os->bitcount += num_bits;
584 /* Flush bits from the bitbuffer variable to the output buffer. 'max_num_bits'
585 * specifies the maximum number of bits that may have been added since the last
588 lzx_flush_bits(struct lzx_output_bitstream *os, unsigned max_num_bits)
590 if (os->end - os->next < 6)
592 put_unaligned_u16_le(os->bitbuf >> (os->bitcount - 16), os->next + 0);
593 if (max_num_bits > 16)
594 put_unaligned_u16_le(os->bitbuf >> (os->bitcount - 32), os->next + 2);
595 if (max_num_bits > 32)
596 put_unaligned_u16_le(os->bitbuf >> (os->bitcount - 48), os->next + 4);
597 os->next += (os->bitcount >> 4) << 1;
601 /* Add at most 16 bits to the bitbuffer and flush it. */
603 lzx_write_bits(struct lzx_output_bitstream *os, u32 bits, unsigned num_bits)
605 lzx_add_bits(os, bits, num_bits);
606 lzx_flush_bits(os, 16);
610 * Flush the last coding unit to the output buffer if needed. Return the total
611 * number of bytes written to the output buffer, or 0 if an overflow occurred.
614 lzx_flush_output(struct lzx_output_bitstream *os)
616 if (os->end - os->next < 6)
619 if (os->bitcount != 0) {
620 put_unaligned_u16_le(os->bitbuf << (16 - os->bitcount), os->next);
624 return os->next - os->start;
627 /* Build the main, length, and aligned offset Huffman codes used in LZX.
629 * This takes as input the frequency tables for each code and produces as output
630 * a set of tables that map symbols to codewords and codeword lengths. */
632 lzx_make_huffman_codes(struct lzx_compressor *c)
634 const struct lzx_freqs *freqs = &c->freqs;
635 struct lzx_codes *codes = &c->codes[c->codes_index];
637 STATIC_ASSERT(MAIN_CODEWORD_LIMIT >= 9 &&
638 MAIN_CODEWORD_LIMIT <= LZX_MAX_MAIN_CODEWORD_LEN);
639 STATIC_ASSERT(LENGTH_CODEWORD_LIMIT >= 8 &&
640 LENGTH_CODEWORD_LIMIT <= LZX_MAX_LEN_CODEWORD_LEN);
641 STATIC_ASSERT(ALIGNED_CODEWORD_LIMIT >= LZX_NUM_ALIGNED_OFFSET_BITS &&
642 ALIGNED_CODEWORD_LIMIT <= LZX_MAX_ALIGNED_CODEWORD_LEN);
644 make_canonical_huffman_code(c->num_main_syms,
648 codes->codewords.main);
650 make_canonical_huffman_code(LZX_LENCODE_NUM_SYMBOLS,
651 LENGTH_CODEWORD_LIMIT,
654 codes->codewords.len);
656 make_canonical_huffman_code(LZX_ALIGNEDCODE_NUM_SYMBOLS,
657 ALIGNED_CODEWORD_LIMIT,
660 codes->codewords.aligned);
663 /* Reset the symbol frequencies for the LZX Huffman codes. */
665 lzx_reset_symbol_frequencies(struct lzx_compressor *c)
667 memset(&c->freqs, 0, sizeof(c->freqs));
671 lzx_compute_precode_items(const u8 lens[restrict],
672 const u8 prev_lens[restrict],
673 u32 precode_freqs[restrict],
674 unsigned precode_items[restrict])
683 itemptr = precode_items;
686 while (!((len = lens[run_start]) & 0x80)) {
688 /* len = the length being repeated */
690 /* Find the next run of codeword lengths. */
692 run_end = run_start + 1;
694 /* Fast case for a single length. */
695 if (likely(len != lens[run_end])) {
696 delta = prev_lens[run_start] - len;
699 precode_freqs[delta]++;
705 /* Extend the run. */
708 } while (len == lens[run_end]);
713 /* Symbol 18: RLE 20 to 51 zeroes at a time. */
714 while ((run_end - run_start) >= 20) {
715 extra_bits = min((run_end - run_start) - 20, 0x1f);
717 *itemptr++ = 18 | (extra_bits << 5);
718 run_start += 20 + extra_bits;
721 /* Symbol 17: RLE 4 to 19 zeroes at a time. */
722 if ((run_end - run_start) >= 4) {
723 extra_bits = min((run_end - run_start) - 4, 0xf);
725 *itemptr++ = 17 | (extra_bits << 5);
726 run_start += 4 + extra_bits;
730 /* A run of nonzero lengths. */
732 /* Symbol 19: RLE 4 to 5 of any length at a time. */
733 while ((run_end - run_start) >= 4) {
734 extra_bits = (run_end - run_start) > 4;
735 delta = prev_lens[run_start] - len;
739 precode_freqs[delta]++;
740 *itemptr++ = 19 | (extra_bits << 5) | (delta << 6);
741 run_start += 4 + extra_bits;
745 /* Output any remaining lengths without RLE. */
746 while (run_start != run_end) {
747 delta = prev_lens[run_start] - len;
750 precode_freqs[delta]++;
756 return itemptr - precode_items;
760 * Output a Huffman code in the compressed form used in LZX.
762 * The Huffman code is represented in the output as a logical series of codeword
763 * lengths from which the Huffman code, which must be in canonical form, can be
766 * The codeword lengths are themselves compressed using a separate Huffman code,
767 * the "precode", which contains a symbol for each possible codeword length in
768 * the larger code as well as several special symbols to represent repeated
769 * codeword lengths (a form of run-length encoding). The precode is itself
770 * constructed in canonical form, and its codeword lengths are represented
771 * literally in 20 4-bit fields that immediately precede the compressed codeword
772 * lengths of the larger code.
774 * Furthermore, the codeword lengths of the larger code are actually represented
775 * as deltas from the codeword lengths of the corresponding code in the previous
779 * Bitstream to which to write the compressed Huffman code.
781 * The codeword lengths, indexed by symbol, in the Huffman code.
783 * The codeword lengths, indexed by symbol, in the corresponding Huffman
784 * code in the previous block, or all zeroes if this is the first block.
786 * The number of symbols in the Huffman code.
789 lzx_write_compressed_code(struct lzx_output_bitstream *os,
790 const u8 lens[restrict],
791 const u8 prev_lens[restrict],
794 u32 precode_freqs[LZX_PRECODE_NUM_SYMBOLS];
795 u8 precode_lens[LZX_PRECODE_NUM_SYMBOLS];
796 u32 precode_codewords[LZX_PRECODE_NUM_SYMBOLS];
797 unsigned precode_items[num_lens];
798 unsigned num_precode_items;
799 unsigned precode_item;
800 unsigned precode_sym;
802 u8 saved = lens[num_lens];
803 *(u8 *)(lens + num_lens) = 0x80;
805 for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++)
806 precode_freqs[i] = 0;
808 /* Compute the "items" (RLE / literal tokens and extra bits) with which
809 * the codeword lengths in the larger code will be output. */
810 num_precode_items = lzx_compute_precode_items(lens,
815 /* Build the precode. */
816 STATIC_ASSERT(PRE_CODEWORD_LIMIT >= 5 &&
817 PRE_CODEWORD_LIMIT <= LZX_MAX_PRE_CODEWORD_LEN);
818 make_canonical_huffman_code(LZX_PRECODE_NUM_SYMBOLS,
820 precode_freqs, precode_lens,
823 /* Output the lengths of the codewords in the precode. */
824 for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++)
825 lzx_write_bits(os, precode_lens[i], LZX_PRECODE_ELEMENT_SIZE);
827 /* Output the encoded lengths of the codewords in the larger code. */
828 for (i = 0; i < num_precode_items; i++) {
829 precode_item = precode_items[i];
830 precode_sym = precode_item & 0x1F;
831 lzx_add_bits(os, precode_codewords[precode_sym],
832 precode_lens[precode_sym]);
833 if (precode_sym >= 17) {
834 if (precode_sym == 17) {
835 lzx_add_bits(os, precode_item >> 5, 4);
836 } else if (precode_sym == 18) {
837 lzx_add_bits(os, precode_item >> 5, 5);
839 lzx_add_bits(os, (precode_item >> 5) & 1, 1);
840 precode_sym = precode_item >> 6;
841 lzx_add_bits(os, precode_codewords[precode_sym],
842 precode_lens[precode_sym]);
845 STATIC_ASSERT(CAN_BUFFER(2 * PRE_CODEWORD_LIMIT + 1));
846 lzx_flush_bits(os, 2 * PRE_CODEWORD_LIMIT + 1);
849 *(u8 *)(lens + num_lens) = saved;
853 * Write all matches and literal bytes (which were precomputed) in an LZX
854 * compressed block to the output bitstream in the final compressed
858 * The output bitstream.
860 * The chosen type of the LZX compressed block (LZX_BLOCKTYPE_ALIGNED or
861 * LZX_BLOCKTYPE_VERBATIM).
863 * The uncompressed data of the block.
865 * The matches and literals to output, given as a series of sequences.
867 * The main, length, and aligned offset Huffman codes for the current
868 * LZX compressed block.
871 lzx_write_sequences(struct lzx_output_bitstream *os, int block_type,
872 const u8 *block_data, const struct lzx_sequence sequences[],
873 const struct lzx_codes *codes)
875 const struct lzx_sequence *seq = sequences;
876 u32 ones_if_aligned = 0 - (block_type == LZX_BLOCKTYPE_ALIGNED);
879 /* Output the next sequence. */
881 unsigned litrunlen = seq->litrunlen;
883 unsigned main_symbol;
884 unsigned adjusted_length;
886 unsigned offset_slot;
887 unsigned num_extra_bits;
890 /* Output the literal run of the sequence. */
892 if (litrunlen) { /* Is the literal run nonempty? */
894 /* Verify optimization is enabled on 64-bit */
895 STATIC_ASSERT(sizeof(machine_word_t) < 8 ||
896 CAN_BUFFER(4 * MAIN_CODEWORD_LIMIT));
898 if (CAN_BUFFER(4 * MAIN_CODEWORD_LIMIT)) {
900 /* 64-bit: write 4 literals at a time. */
901 while (litrunlen >= 4) {
902 unsigned lit0 = block_data[0];
903 unsigned lit1 = block_data[1];
904 unsigned lit2 = block_data[2];
905 unsigned lit3 = block_data[3];
906 lzx_add_bits(os, codes->codewords.main[lit0], codes->lens.main[lit0]);
907 lzx_add_bits(os, codes->codewords.main[lit1], codes->lens.main[lit1]);
908 lzx_add_bits(os, codes->codewords.main[lit2], codes->lens.main[lit2]);
909 lzx_add_bits(os, codes->codewords.main[lit3], codes->lens.main[lit3]);
910 lzx_flush_bits(os, 4 * MAIN_CODEWORD_LIMIT);
915 unsigned lit = *block_data++;
916 lzx_add_bits(os, codes->codewords.main[lit], codes->lens.main[lit]);
918 unsigned lit = *block_data++;
919 lzx_add_bits(os, codes->codewords.main[lit], codes->lens.main[lit]);
921 unsigned lit = *block_data++;
922 lzx_add_bits(os, codes->codewords.main[lit], codes->lens.main[lit]);
923 lzx_flush_bits(os, 3 * MAIN_CODEWORD_LIMIT);
925 lzx_flush_bits(os, 2 * MAIN_CODEWORD_LIMIT);
928 lzx_flush_bits(os, 1 * MAIN_CODEWORD_LIMIT);
932 /* 32-bit: write 1 literal at a time. */
934 unsigned lit = *block_data++;
935 lzx_add_bits(os, codes->codewords.main[lit], codes->lens.main[lit]);
936 lzx_flush_bits(os, MAIN_CODEWORD_LIMIT);
937 } while (--litrunlen);
941 /* Was this the last literal run? */
942 if (seq->adjusted_offset_and_match_hdr & 0x80000000)
945 /* Nope; output the match. */
947 match_hdr = seq->adjusted_offset_and_match_hdr & 0x1FF;
948 main_symbol = LZX_NUM_CHARS + match_hdr;
949 adjusted_length = seq->adjusted_length;
951 block_data += adjusted_length + LZX_MIN_MATCH_LEN;
953 offset_slot = match_hdr / LZX_NUM_LEN_HEADERS;
954 adjusted_offset = seq->adjusted_offset_and_match_hdr >> 9;
956 num_extra_bits = lzx_extra_offset_bits[offset_slot];
957 extra_bits = adjusted_offset - lzx_offset_slot_base[offset_slot];
959 #define MAX_MATCH_BITS (MAIN_CODEWORD_LIMIT + LENGTH_CODEWORD_LIMIT + \
960 14 + ALIGNED_CODEWORD_LIMIT)
962 /* Verify optimization is enabled on 64-bit */
963 STATIC_ASSERT(sizeof(machine_word_t) < 8 || CAN_BUFFER(MAX_MATCH_BITS));
965 /* Output the main symbol for the match. */
967 lzx_add_bits(os, codes->codewords.main[main_symbol],
968 codes->lens.main[main_symbol]);
969 if (!CAN_BUFFER(MAX_MATCH_BITS))
970 lzx_flush_bits(os, MAIN_CODEWORD_LIMIT);
972 /* If needed, output the length symbol for the match. */
974 if (adjusted_length >= LZX_NUM_PRIMARY_LENS) {
975 lzx_add_bits(os, codes->codewords.len[adjusted_length - LZX_NUM_PRIMARY_LENS],
976 codes->lens.len[adjusted_length - LZX_NUM_PRIMARY_LENS]);
977 if (!CAN_BUFFER(MAX_MATCH_BITS))
978 lzx_flush_bits(os, LENGTH_CODEWORD_LIMIT);
981 /* Output the extra offset bits for the match. In aligned
982 * offset blocks, the lowest 3 bits of the adjusted offset are
983 * Huffman-encoded using the aligned offset code, provided that
984 * there are at least extra 3 offset bits required. All other
985 * extra offset bits are output verbatim. */
987 if ((adjusted_offset & ones_if_aligned) >= 16) {
989 lzx_add_bits(os, extra_bits >> LZX_NUM_ALIGNED_OFFSET_BITS,
990 num_extra_bits - LZX_NUM_ALIGNED_OFFSET_BITS);
991 if (!CAN_BUFFER(MAX_MATCH_BITS))
992 lzx_flush_bits(os, 14);
994 lzx_add_bits(os, codes->codewords.aligned[adjusted_offset & LZX_ALIGNED_OFFSET_BITMASK],
995 codes->lens.aligned[adjusted_offset & LZX_ALIGNED_OFFSET_BITMASK]);
996 if (!CAN_BUFFER(MAX_MATCH_BITS))
997 lzx_flush_bits(os, ALIGNED_CODEWORD_LIMIT);
999 STATIC_ASSERT(CAN_BUFFER(17));
1001 lzx_add_bits(os, extra_bits, num_extra_bits);
1002 if (!CAN_BUFFER(MAX_MATCH_BITS))
1003 lzx_flush_bits(os, 17);
1006 if (CAN_BUFFER(MAX_MATCH_BITS))
1007 lzx_flush_bits(os, MAX_MATCH_BITS);
1009 /* Advance to the next sequence. */
1015 lzx_write_compressed_block(const u8 *block_begin,
1018 unsigned window_order,
1019 unsigned num_main_syms,
1020 const struct lzx_sequence sequences[],
1021 const struct lzx_codes * codes,
1022 const struct lzx_lens * prev_lens,
1023 struct lzx_output_bitstream * os)
1025 /* The first three bits indicate the type of block and are one of the
1026 * LZX_BLOCKTYPE_* constants. */
1027 lzx_write_bits(os, block_type, 3);
1029 /* Output the block size.
1031 * The original LZX format seemed to always encode the block size in 3
1032 * bytes. However, the implementation in WIMGAPI, as used in WIM files,
1033 * uses the first bit to indicate whether the block is the default size
1034 * (32768) or a different size given explicitly by the next 16 bits.
1036 * By default, this compressor uses a window size of 32768 and therefore
1037 * follows the WIMGAPI behavior. However, this compressor also supports
1038 * window sizes greater than 32768 bytes, which do not appear to be
1039 * supported by WIMGAPI. In such cases, we retain the default size bit
1040 * to mean a size of 32768 bytes but output non-default block size in 24
1041 * bits rather than 16. The compatibility of this behavior is unknown
1042 * because WIMs created with chunk size greater than 32768 can seemingly
1043 * only be opened by wimlib anyway. */
1044 if (block_size == LZX_DEFAULT_BLOCK_SIZE) {
1045 lzx_write_bits(os, 1, 1);
1047 lzx_write_bits(os, 0, 1);
1049 if (window_order >= 16)
1050 lzx_write_bits(os, block_size >> 16, 8);
1052 lzx_write_bits(os, block_size & 0xFFFF, 16);
1055 /* If it's an aligned offset block, output the aligned offset code. */
1056 if (block_type == LZX_BLOCKTYPE_ALIGNED) {
1057 for (int i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1058 lzx_write_bits(os, codes->lens.aligned[i],
1059 LZX_ALIGNEDCODE_ELEMENT_SIZE);
1063 /* Output the main code (two parts). */
1064 lzx_write_compressed_code(os, codes->lens.main,
1067 lzx_write_compressed_code(os, codes->lens.main + LZX_NUM_CHARS,
1068 prev_lens->main + LZX_NUM_CHARS,
1069 num_main_syms - LZX_NUM_CHARS);
1071 /* Output the length code. */
1072 lzx_write_compressed_code(os, codes->lens.len,
1074 LZX_LENCODE_NUM_SYMBOLS);
1076 /* Output the compressed matches and literals. */
1077 lzx_write_sequences(os, block_type, block_begin, sequences, codes);
1080 /* Given the frequencies of symbols in an LZX-compressed block and the
1081 * corresponding Huffman codes, return LZX_BLOCKTYPE_ALIGNED or
1082 * LZX_BLOCKTYPE_VERBATIM if an aligned offset or verbatim block, respectively,
1083 * will take fewer bits to output. */
1085 lzx_choose_verbatim_or_aligned(const struct lzx_freqs * freqs,
1086 const struct lzx_codes * codes)
1088 u32 aligned_cost = 0;
1089 u32 verbatim_cost = 0;
1091 /* A verbatim block requires 3 bits in each place that an aligned symbol
1092 * would be used in an aligned offset block. */
1093 for (unsigned i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1094 verbatim_cost += LZX_NUM_ALIGNED_OFFSET_BITS * freqs->aligned[i];
1095 aligned_cost += codes->lens.aligned[i] * freqs->aligned[i];
1098 /* Account for output of the aligned offset code. */
1099 aligned_cost += LZX_ALIGNEDCODE_ELEMENT_SIZE * LZX_ALIGNEDCODE_NUM_SYMBOLS;
1101 if (aligned_cost < verbatim_cost)
1102 return LZX_BLOCKTYPE_ALIGNED;
1104 return LZX_BLOCKTYPE_VERBATIM;
1108 * Return the offset slot for the specified adjusted match offset, using the
1109 * compressor's acceleration tables to speed up the mapping.
1111 static inline unsigned
1112 lzx_comp_get_offset_slot(struct lzx_compressor *c, u32 adjusted_offset,
1115 if (is_16_bit || adjusted_offset < ARRAY_LEN(c->offset_slot_tab_1))
1116 return c->offset_slot_tab_1[adjusted_offset];
1117 return c->offset_slot_tab_2[adjusted_offset >> 14];
1121 * Finish an LZX block:
1123 * - build the Huffman codes
1124 * - decide whether to output the block as VERBATIM or ALIGNED
1125 * - output the block
1126 * - swap the indices of the current and previous Huffman codes
1129 lzx_finish_block(struct lzx_compressor *c, struct lzx_output_bitstream *os,
1130 const u8 *block_begin, u32 block_size, u32 seq_idx)
1134 lzx_make_huffman_codes(c);
1136 block_type = lzx_choose_verbatim_or_aligned(&c->freqs,
1137 &c->codes[c->codes_index]);
1138 lzx_write_compressed_block(block_begin,
1143 &c->chosen_sequences[seq_idx],
1144 &c->codes[c->codes_index],
1145 &c->codes[c->codes_index ^ 1].lens,
1147 c->codes_index ^= 1;
1150 /* Tally the Huffman symbol for a literal and increment the literal run length.
1153 lzx_record_literal(struct lzx_compressor *c, unsigned literal, u32 *litrunlen_p)
1155 c->freqs.main[literal]++;
1159 /* Tally the Huffman symbol for a match, save the match data and the length of
1160 * the preceding literal run in the next lzx_sequence, and update the recent
1163 lzx_record_match(struct lzx_compressor *c, unsigned length, u32 offset_data,
1164 u32 recent_offsets[LZX_NUM_RECENT_OFFSETS], bool is_16_bit,
1165 u32 *litrunlen_p, struct lzx_sequence **next_seq_p)
1167 u32 litrunlen = *litrunlen_p;
1168 struct lzx_sequence *next_seq = *next_seq_p;
1169 unsigned offset_slot;
1172 v = length - LZX_MIN_MATCH_LEN;
1174 /* Save the literal run length and adjusted length. */
1175 next_seq->litrunlen = litrunlen;
1176 next_seq->adjusted_length = v;
1178 /* Compute the length header and tally the length symbol if needed */
1179 if (v >= LZX_NUM_PRIMARY_LENS) {
1180 c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++;
1181 v = LZX_NUM_PRIMARY_LENS;
1184 /* Compute the offset slot */
1185 offset_slot = lzx_comp_get_offset_slot(c, offset_data, is_16_bit);
1187 /* Compute the match header. */
1188 v += offset_slot * LZX_NUM_LEN_HEADERS;
1190 /* Save the adjusted offset and match header. */
1191 next_seq->adjusted_offset_and_match_hdr = (offset_data << 9) | v;
1193 /* Tally the main symbol. */
1194 c->freqs.main[LZX_NUM_CHARS + v]++;
1196 /* Update the recent offsets queue. */
1197 if (offset_data < LZX_NUM_RECENT_OFFSETS) {
1198 /* Repeat offset match */
1199 swap(recent_offsets[0], recent_offsets[offset_data]);
1201 /* Explicit offset match */
1203 /* Tally the aligned offset symbol if needed */
1204 if (offset_data >= 16)
1205 c->freqs.aligned[offset_data & LZX_ALIGNED_OFFSET_BITMASK]++;
1207 recent_offsets[2] = recent_offsets[1];
1208 recent_offsets[1] = recent_offsets[0];
1209 recent_offsets[0] = offset_data - LZX_OFFSET_ADJUSTMENT;
1212 /* Reset the literal run length and advance to the next sequence. */
1213 *next_seq_p = next_seq + 1;
1217 /* Finish the last lzx_sequence. The last lzx_sequence is just a literal run;
1218 * there is no match. This literal run may be empty. */
1220 lzx_finish_sequence(struct lzx_sequence *last_seq, u32 litrunlen)
1222 last_seq->litrunlen = litrunlen;
1224 /* Special value to mark last sequence */
1225 last_seq->adjusted_offset_and_match_hdr = 0x80000000;
1229 * Given the minimum-cost path computed through the item graph for the current
1230 * block, walk the path and count how many of each symbol in each Huffman-coded
1231 * alphabet would be required to output the items (matches and literals) along
1234 * Note that the path will be walked backwards (from the end of the block to the
1235 * beginning of the block), but this doesn't matter because this function only
1236 * computes frequencies.
1239 lzx_tally_item_list(struct lzx_compressor *c, u32 block_size, bool is_16_bit)
1241 u32 node_idx = block_size;
1246 unsigned offset_slot;
1248 /* Tally literals until either a match or the beginning of the
1249 * block is reached. */
1251 u32 item = c->optimum_nodes[node_idx].item;
1253 len = item & OPTIMUM_LEN_MASK;
1254 offset_data = item >> OPTIMUM_OFFSET_SHIFT;
1256 if (len != 0) /* Not a literal? */
1259 /* Tally the main symbol for the literal. */
1260 c->freqs.main[offset_data]++;
1262 if (--node_idx == 0) /* Beginning of block was reached? */
1268 /* Tally a match. */
1270 /* Tally the aligned offset symbol if needed. */
1271 if (offset_data >= 16)
1272 c->freqs.aligned[offset_data & LZX_ALIGNED_OFFSET_BITMASK]++;
1274 /* Tally the length symbol if needed. */
1275 v = len - LZX_MIN_MATCH_LEN;;
1276 if (v >= LZX_NUM_PRIMARY_LENS) {
1277 c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++;
1278 v = LZX_NUM_PRIMARY_LENS;
1281 /* Tally the main symbol. */
1282 offset_slot = lzx_comp_get_offset_slot(c, offset_data, is_16_bit);
1283 v += offset_slot * LZX_NUM_LEN_HEADERS;
1284 c->freqs.main[LZX_NUM_CHARS + v]++;
1286 if (node_idx == 0) /* Beginning of block was reached? */
1292 * Like lzx_tally_item_list(), but this function also generates the list of
1293 * lzx_sequences for the minimum-cost path and writes it to c->chosen_sequences,
1294 * ready to be output to the bitstream after the Huffman codes are computed.
1295 * The lzx_sequences will be written to decreasing memory addresses as the path
1296 * is walked backwards, which means they will end up in the expected
1297 * first-to-last order. The return value is the index in c->chosen_sequences at
1298 * which the lzx_sequences begin.
1301 lzx_record_item_list(struct lzx_compressor *c, u32 block_size, bool is_16_bit)
1303 u32 node_idx = block_size;
1304 u32 seq_idx = ARRAY_LEN(c->chosen_sequences) - 1;
1307 /* Special value to mark last sequence */
1308 c->chosen_sequences[seq_idx].adjusted_offset_and_match_hdr = 0x80000000;
1310 lit_start_node = node_idx;
1315 unsigned offset_slot;
1317 /* Record literals until either a match or the beginning of the
1318 * block is reached. */
1320 u32 item = c->optimum_nodes[node_idx].item;
1322 len = item & OPTIMUM_LEN_MASK;
1323 offset_data = item >> OPTIMUM_OFFSET_SHIFT;
1325 if (len != 0) /* Not a literal? */
1328 /* Tally the main symbol for the literal. */
1329 c->freqs.main[offset_data]++;
1331 if (--node_idx == 0) /* Beginning of block was reached? */
1335 /* Save the literal run length for the next sequence (the
1336 * "previous sequence" when walking backwards). */
1337 c->chosen_sequences[seq_idx--].litrunlen = lit_start_node - node_idx;
1339 lit_start_node = node_idx;
1341 /* Record a match. */
1343 /* Tally the aligned offset symbol if needed. */
1344 if (offset_data >= 16)
1345 c->freqs.aligned[offset_data & LZX_ALIGNED_OFFSET_BITMASK]++;
1347 /* Save the adjusted length. */
1348 v = len - LZX_MIN_MATCH_LEN;
1349 c->chosen_sequences[seq_idx].adjusted_length = v;
1351 /* Tally the length symbol if needed. */
1352 if (v >= LZX_NUM_PRIMARY_LENS) {
1353 c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++;
1354 v = LZX_NUM_PRIMARY_LENS;
1357 /* Tally the main symbol. */
1358 offset_slot = lzx_comp_get_offset_slot(c, offset_data, is_16_bit);
1359 v += offset_slot * LZX_NUM_LEN_HEADERS;
1360 c->freqs.main[LZX_NUM_CHARS + v]++;
1362 /* Save the adjusted offset and match header. */
1363 c->chosen_sequences[seq_idx].adjusted_offset_and_match_hdr =
1364 (offset_data << 9) | v;
1366 if (node_idx == 0) /* Beginning of block was reached? */
1371 /* Save the literal run length for the first sequence. */
1372 c->chosen_sequences[seq_idx].litrunlen = lit_start_node - node_idx;
1374 /* Return the index in c->chosen_sequences at which the lzx_sequences
1380 * Find an inexpensive path through the graph of possible match/literal choices
1381 * for the current block. The nodes of the graph are
1382 * c->optimum_nodes[0...block_size]. They correspond directly to the bytes in
1383 * the current block, plus one extra node for end-of-block. The edges of the
1384 * graph are matches and literals. The goal is to find the minimum cost path
1385 * from 'c->optimum_nodes[0]' to 'c->optimum_nodes[block_size]'.
1387 * The algorithm works forwards, starting at 'c->optimum_nodes[0]' and
1388 * proceeding forwards one node at a time. At each node, a selection of matches
1389 * (len >= 2), as well as the literal byte (len = 1), is considered. An item of
1390 * length 'len' provides a new path to reach the node 'len' bytes later. If
1391 * such a path is the lowest cost found so far to reach that later node, then
1392 * that later node is updated with the new path.
1394 * Note that although this algorithm is based on minimum cost path search, due
1395 * to various simplifying assumptions the result is not guaranteed to be the
1396 * true minimum cost, or "optimal", path over the graph of all valid LZX
1397 * representations of this block.
1399 * Also, note that because of the presence of the recent offsets queue (which is
1400 * a type of adaptive state), the algorithm cannot work backwards and compute
1401 * "cost to end" instead of "cost to beginning". Furthermore, the way the
1402 * algorithm handles this adaptive state in the "minimum cost" parse is actually
1403 * only an approximation. It's possible for the globally optimal, minimum cost
1404 * path to contain a prefix, ending at a position, where that path prefix is
1405 * *not* the minimum cost path to that position. This can happen if such a path
1406 * prefix results in a different adaptive state which results in lower costs
1407 * later. The algorithm does not solve this problem; it only considers the
1408 * lowest cost to reach each individual position.
1410 static inline struct lzx_lru_queue
1411 lzx_find_min_cost_path(struct lzx_compressor * const restrict c,
1412 const u8 * const restrict block_begin,
1413 const u32 block_size,
1414 const struct lzx_lru_queue initial_queue,
1417 struct lzx_optimum_node *cur_node = c->optimum_nodes;
1418 struct lzx_optimum_node * const end_node = &c->optimum_nodes[block_size];
1419 struct lz_match *cache_ptr = c->match_cache;
1420 const u8 *in_next = block_begin;
1421 const u8 * const block_end = block_begin + block_size;
1423 /* Instead of storing the match offset LRU queues in the
1424 * 'lzx_optimum_node' structures, we save memory (and cache lines) by
1425 * storing them in a smaller array. This works because the algorithm
1426 * only requires a limited history of the adaptive state. Once a given
1427 * state is more than LZX_MAX_MATCH_LEN bytes behind the current node,
1428 * it is no longer needed. */
1429 struct lzx_lru_queue queues[512];
1431 STATIC_ASSERT(ARRAY_LEN(queues) >= LZX_MAX_MATCH_LEN + 1);
1432 #define QUEUE(in) (queues[(uintptr_t)(in) % ARRAY_LEN(queues)])
1434 /* Initially, the cost to reach each node is "infinity". */
1435 memset(c->optimum_nodes, 0xFF,
1436 (block_size + 1) * sizeof(c->optimum_nodes[0]));
1438 QUEUE(block_begin) = initial_queue;
1440 /* The following loop runs 'block_size' iterations, one per node. */
1442 unsigned num_matches;
1447 * A selection of matches for the block was already saved in
1448 * memory so that we don't have to run the uncompressed data
1449 * through the matchfinder on every optimization pass. However,
1450 * we still search for repeat offset matches during each
1451 * optimization pass because we cannot predict the state of the
1452 * recent offsets queue. But as a heuristic, we don't bother
1453 * searching for repeat offset matches if the general-purpose
1454 * matchfinder failed to find any matches.
1456 * Note that a match of length n at some offset implies there is
1457 * also a match of length l for LZX_MIN_MATCH_LEN <= l <= n at
1458 * that same offset. In other words, we don't necessarily need
1459 * to use the full length of a match. The key heuristic that
1460 * saves a significicant amount of time is that for each
1461 * distinct length, we only consider the smallest offset for
1462 * which that length is available. This heuristic also applies
1463 * to repeat offsets, which we order specially: R0 < R1 < R2 <
1464 * any explicit offset. Of course, this heuristic may be
1465 * produce suboptimal results because offset slots in LZX are
1466 * subject to entropy encoding, but in practice this is a useful
1470 num_matches = cache_ptr->length;
1474 struct lz_match *end_matches = cache_ptr + num_matches;
1475 unsigned next_len = LZX_MIN_MATCH_LEN;
1476 unsigned max_len = min(block_end - in_next, LZX_MAX_MATCH_LEN);
1479 /* Consider R0 match */
1480 matchptr = in_next - lzx_lru_queue_R0(QUEUE(in_next));
1481 if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1483 STATIC_ASSERT(LZX_MIN_MATCH_LEN == 2);
1485 u32 cost = cur_node->cost +
1486 c->costs.match_cost[0][
1487 next_len - LZX_MIN_MATCH_LEN];
1488 if (cost <= (cur_node + next_len)->cost) {
1489 (cur_node + next_len)->cost = cost;
1490 (cur_node + next_len)->item =
1491 (0 << OPTIMUM_OFFSET_SHIFT) | next_len;
1493 if (unlikely(++next_len > max_len)) {
1494 cache_ptr = end_matches;
1497 } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1501 /* Consider R1 match */
1502 matchptr = in_next - lzx_lru_queue_R1(QUEUE(in_next));
1503 if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1505 if (matchptr[next_len - 1] != in_next[next_len - 1])
1507 for (unsigned len = 2; len < next_len - 1; len++)
1508 if (matchptr[len] != in_next[len])
1511 u32 cost = cur_node->cost +
1512 c->costs.match_cost[1][
1513 next_len - LZX_MIN_MATCH_LEN];
1514 if (cost <= (cur_node + next_len)->cost) {
1515 (cur_node + next_len)->cost = cost;
1516 (cur_node + next_len)->item =
1517 (1 << OPTIMUM_OFFSET_SHIFT) | next_len;
1519 if (unlikely(++next_len > max_len)) {
1520 cache_ptr = end_matches;
1523 } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1527 /* Consider R2 match */
1528 matchptr = in_next - lzx_lru_queue_R2(QUEUE(in_next));
1529 if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1531 if (matchptr[next_len - 1] != in_next[next_len - 1])
1533 for (unsigned len = 2; len < next_len - 1; len++)
1534 if (matchptr[len] != in_next[len])
1537 u32 cost = cur_node->cost +
1538 c->costs.match_cost[2][
1539 next_len - LZX_MIN_MATCH_LEN];
1540 if (cost <= (cur_node + next_len)->cost) {
1541 (cur_node + next_len)->cost = cost;
1542 (cur_node + next_len)->item =
1543 (2 << OPTIMUM_OFFSET_SHIFT) | next_len;
1545 if (unlikely(++next_len > max_len)) {
1546 cache_ptr = end_matches;
1549 } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1553 while (next_len > cache_ptr->length)
1554 if (++cache_ptr == end_matches)
1557 /* Consider explicit offset matches */
1559 u32 offset = cache_ptr->offset;
1560 u32 offset_data = offset + LZX_OFFSET_ADJUSTMENT;
1561 unsigned offset_slot = lzx_comp_get_offset_slot(c, offset_data,
1563 u32 base_cost = cur_node->cost;
1565 #if LZX_CONSIDER_ALIGNED_COSTS
1566 if (offset_data >= 16)
1567 base_cost += c->costs.aligned[offset_data &
1568 LZX_ALIGNED_OFFSET_BITMASK];
1572 u32 cost = base_cost +
1573 c->costs.match_cost[offset_slot][
1574 next_len - LZX_MIN_MATCH_LEN];
1575 if (cost < (cur_node + next_len)->cost) {
1576 (cur_node + next_len)->cost = cost;
1577 (cur_node + next_len)->item =
1578 (offset_data << OPTIMUM_OFFSET_SHIFT) | next_len;
1580 } while (++next_len <= cache_ptr->length);
1581 } while (++cache_ptr != end_matches);
1586 /* Consider coding a literal.
1588 * To avoid an extra branch, actually checking the preferability
1589 * of coding the literal is integrated into the queue update
1591 literal = *in_next++;
1592 cost = cur_node->cost + c->costs.main[literal];
1594 /* Advance to the next position. */
1597 /* The lowest-cost path to the current position is now known.
1598 * Finalize the recent offsets queue that results from taking
1599 * this lowest-cost path. */
1601 if (cost <= cur_node->cost) {
1602 /* Literal: queue remains unchanged. */
1603 cur_node->cost = cost;
1604 cur_node->item = (u32)literal << OPTIMUM_OFFSET_SHIFT;
1605 QUEUE(in_next) = QUEUE(in_next - 1);
1607 /* Match: queue update is needed. */
1608 unsigned len = cur_node->item & OPTIMUM_LEN_MASK;
1609 u32 offset_data = cur_node->item >> OPTIMUM_OFFSET_SHIFT;
1610 if (offset_data >= LZX_NUM_RECENT_OFFSETS) {
1611 /* Explicit offset match: insert offset at front */
1613 lzx_lru_queue_push(QUEUE(in_next - len),
1614 offset_data - LZX_OFFSET_ADJUSTMENT);
1616 /* Repeat offset match: swap offset to front */
1618 lzx_lru_queue_swap(QUEUE(in_next - len),
1622 } while (cur_node != end_node);
1624 /* Return the match offset queue at the end of the minimum cost path. */
1625 return QUEUE(block_end);
1628 /* Given the costs for the main and length codewords, compute 'match_costs'. */
1630 lzx_compute_match_costs(struct lzx_compressor *c)
1632 unsigned num_offset_slots = (c->num_main_syms - LZX_NUM_CHARS) / LZX_NUM_LEN_HEADERS;
1633 struct lzx_costs *costs = &c->costs;
1635 for (unsigned offset_slot = 0; offset_slot < num_offset_slots; offset_slot++) {
1637 u32 extra_cost = (u32)lzx_extra_offset_bits[offset_slot] * LZX_BIT_COST;
1638 unsigned main_symbol = LZX_NUM_CHARS + (offset_slot * LZX_NUM_LEN_HEADERS);
1641 #if LZX_CONSIDER_ALIGNED_COSTS
1642 if (offset_slot >= 8)
1643 extra_cost -= LZX_NUM_ALIGNED_OFFSET_BITS * LZX_BIT_COST;
1646 for (i = 0; i < LZX_NUM_PRIMARY_LENS; i++)
1647 costs->match_cost[offset_slot][i] =
1648 costs->main[main_symbol++] + extra_cost;
1650 extra_cost += costs->main[main_symbol];
1652 for (; i < LZX_NUM_LENS; i++)
1653 costs->match_cost[offset_slot][i] =
1654 costs->len[i - LZX_NUM_PRIMARY_LENS] + extra_cost;
1658 /* Set default LZX Huffman symbol costs to bootstrap the iterative optimization
1661 lzx_set_default_costs(struct lzx_compressor *c, const u8 *block, u32 block_size)
1664 bool have_byte[256];
1665 unsigned num_used_bytes;
1667 /* The costs below are hard coded to use a scaling factor of 16. */
1668 STATIC_ASSERT(LZX_BIT_COST == 16);
1673 * - Use smaller initial costs for literal symbols when the input buffer
1674 * contains fewer distinct bytes.
1676 * - Assume that match symbols are more costly than literal symbols.
1678 * - Assume that length symbols for shorter lengths are less costly than
1679 * length symbols for longer lengths.
1682 for (i = 0; i < 256; i++)
1683 have_byte[i] = false;
1685 for (i = 0; i < block_size; i++)
1686 have_byte[block[i]] = true;
1689 for (i = 0; i < 256; i++)
1690 num_used_bytes += have_byte[i];
1692 for (i = 0; i < 256; i++)
1693 c->costs.main[i] = 140 - (256 - num_used_bytes) / 4;
1695 for (; i < c->num_main_syms; i++)
1696 c->costs.main[i] = 170;
1698 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
1699 c->costs.len[i] = 103 + (i / 4);
1701 #if LZX_CONSIDER_ALIGNED_COSTS
1702 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
1703 c->costs.aligned[i] = LZX_NUM_ALIGNED_OFFSET_BITS * LZX_BIT_COST;
1706 lzx_compute_match_costs(c);
1709 /* Update the current cost model to reflect the computed Huffman codes. */
1711 lzx_update_costs(struct lzx_compressor *c)
1714 const struct lzx_lens *lens = &c->codes[c->codes_index].lens;
1716 for (i = 0; i < c->num_main_syms; i++) {
1717 c->costs.main[i] = (lens->main[i] ? lens->main[i] :
1718 MAIN_CODEWORD_LIMIT) * LZX_BIT_COST;
1721 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) {
1722 c->costs.len[i] = (lens->len[i] ? lens->len[i] :
1723 LENGTH_CODEWORD_LIMIT) * LZX_BIT_COST;
1726 #if LZX_CONSIDER_ALIGNED_COSTS
1727 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1728 c->costs.aligned[i] = (lens->aligned[i] ? lens->aligned[i] :
1729 ALIGNED_CODEWORD_LIMIT) * LZX_BIT_COST;
1733 lzx_compute_match_costs(c);
1736 static inline struct lzx_lru_queue
1737 lzx_optimize_and_write_block(struct lzx_compressor * const restrict c,
1738 struct lzx_output_bitstream * const restrict os,
1739 const u8 * const restrict block_begin,
1740 const u32 block_size,
1741 const struct lzx_lru_queue initial_queue,
1744 unsigned num_passes_remaining = c->num_optim_passes;
1745 struct lzx_lru_queue new_queue;
1748 /* The first optimization pass uses a default cost model. Each
1749 * additional optimization pass uses a cost model derived from the
1750 * Huffman code computed in the previous pass. */
1752 lzx_set_default_costs(c, block_begin, block_size);
1753 lzx_reset_symbol_frequencies(c);
1755 new_queue = lzx_find_min_cost_path(c, block_begin, block_size,
1756 initial_queue, is_16_bit);
1757 if (num_passes_remaining > 1) {
1758 lzx_tally_item_list(c, block_size, is_16_bit);
1759 lzx_make_huffman_codes(c);
1760 lzx_update_costs(c);
1761 lzx_reset_symbol_frequencies(c);
1763 } while (--num_passes_remaining);
1765 seq_idx = lzx_record_item_list(c, block_size, is_16_bit);
1766 lzx_finish_block(c, os, block_begin, block_size, seq_idx);
1771 * This is the "near-optimal" LZX compressor.
1773 * For each block, it performs a relatively thorough graph search to find an
1774 * inexpensive (in terms of compressed size) way to output that block.
1776 * Note: there are actually many things this algorithm leaves on the table in
1777 * terms of compression ratio. So although it may be "near-optimal", it is
1778 * certainly not "optimal". The goal is not to produce the optimal compression
1779 * ratio, which for LZX is probably impossible within any practical amount of
1780 * time, but rather to produce a compression ratio significantly better than a
1781 * simpler "greedy" or "lazy" parse while still being relatively fast.
1784 lzx_compress_near_optimal(struct lzx_compressor *c,
1785 struct lzx_output_bitstream *os,
1788 const u8 * const in_begin = c->in_buffer;
1789 const u8 * in_next = in_begin;
1790 const u8 * const in_end = in_begin + c->in_nbytes;
1791 u32 max_len = LZX_MAX_MATCH_LEN;
1792 u32 nice_len = min(c->nice_match_length, max_len);
1793 u32 next_hashes[2] = {};
1794 struct lzx_lru_queue queue;
1796 CALL_BT_MF(is_16_bit, c, bt_matchfinder_init);
1797 lzx_lru_queue_init(&queue);
1800 /* Starting a new block */
1801 const u8 * const in_block_begin = in_next;
1802 const u8 * const in_block_end =
1803 in_next + min(LZX_DIV_BLOCK_SIZE, in_end - in_next);
1805 /* Run the block through the matchfinder and cache the matches. */
1806 struct lz_match *cache_ptr = c->match_cache;
1808 struct lz_match *lz_matchptr;
1811 /* If approaching the end of the input buffer, adjust
1812 * 'max_len' and 'nice_len' accordingly. */
1813 if (unlikely(max_len > in_end - in_next)) {
1814 max_len = in_end - in_next;
1815 nice_len = min(max_len, nice_len);
1816 if (unlikely(max_len < BT_MATCHFINDER_REQUIRED_NBYTES)) {
1818 cache_ptr->length = 0;
1824 /* Check for matches. */
1825 lz_matchptr = CALL_BT_MF(is_16_bit, c, bt_matchfinder_get_matches,
1830 c->max_search_depth,
1835 cache_ptr->length = lz_matchptr - (cache_ptr + 1);
1836 cache_ptr = lz_matchptr;
1839 * If there was a very long match found, then don't
1840 * cache any matches for the bytes covered by that
1841 * match. This avoids degenerate behavior when
1842 * compressing highly redundant data, where the number
1843 * of matches can be very large.
1845 * This heuristic doesn't actually hurt the compression
1846 * ratio very much. If there's a long match, then the
1847 * data must be highly compressible, so it doesn't
1848 * matter as much what we do.
1850 if (best_len >= nice_len) {
1853 if (unlikely(max_len > in_end - in_next)) {
1854 max_len = in_end - in_next;
1855 nice_len = min(max_len, nice_len);
1856 if (unlikely(max_len < BT_MATCHFINDER_REQUIRED_NBYTES)) {
1858 cache_ptr->length = 0;
1863 CALL_BT_MF(is_16_bit, c, bt_matchfinder_skip_position,
1868 c->max_search_depth,
1871 cache_ptr->length = 0;
1873 } while (--best_len);
1875 } while (in_next < in_block_end &&
1876 likely(cache_ptr < &c->match_cache[LZX_CACHE_LENGTH]));
1878 /* We've finished running the block through the matchfinder.
1879 * Now choose a match/literal sequence and write the block. */
1881 queue = lzx_optimize_and_write_block(c, os, in_block_begin,
1882 in_next - in_block_begin,
1884 } while (in_next != in_end);
1888 lzx_compress_near_optimal_16(struct lzx_compressor *c,
1889 struct lzx_output_bitstream *os)
1891 lzx_compress_near_optimal(c, os, true);
1895 lzx_compress_near_optimal_32(struct lzx_compressor *c,
1896 struct lzx_output_bitstream *os)
1898 lzx_compress_near_optimal(c, os, false);
1902 * Given a pointer to the current byte sequence and the current list of recent
1903 * match offsets, find the longest repeat offset match.
1905 * If no match of at least 2 bytes is found, then return 0.
1907 * If a match of at least 2 bytes is found, then return its length and set
1908 * *rep_max_idx_ret to the index of its offset in @queue.
1911 lzx_find_longest_repeat_offset_match(const u8 * const in_next,
1912 const u32 bytes_remaining,
1913 const u32 recent_offsets[LZX_NUM_RECENT_OFFSETS],
1914 unsigned *rep_max_idx_ret)
1916 STATIC_ASSERT(LZX_NUM_RECENT_OFFSETS == 3);
1918 const unsigned max_len = min(bytes_remaining, LZX_MAX_MATCH_LEN);
1919 const u16 next_2_bytes = load_u16_unaligned(in_next);
1921 unsigned rep_max_len;
1922 unsigned rep_max_idx;
1925 matchptr = in_next - recent_offsets[0];
1926 if (load_u16_unaligned(matchptr) == next_2_bytes)
1927 rep_max_len = lz_extend(in_next, matchptr, 2, max_len);
1932 matchptr = in_next - recent_offsets[1];
1933 if (load_u16_unaligned(matchptr) == next_2_bytes) {
1934 rep_len = lz_extend(in_next, matchptr, 2, max_len);
1935 if (rep_len > rep_max_len) {
1936 rep_max_len = rep_len;
1941 matchptr = in_next - recent_offsets[2];
1942 if (load_u16_unaligned(matchptr) == next_2_bytes) {
1943 rep_len = lz_extend(in_next, matchptr, 2, max_len);
1944 if (rep_len > rep_max_len) {
1945 rep_max_len = rep_len;
1950 *rep_max_idx_ret = rep_max_idx;
1954 /* Fast heuristic scoring for lazy parsing: how "good" is this match? */
1955 static inline unsigned
1956 lzx_explicit_offset_match_score(unsigned len, u32 adjusted_offset)
1958 unsigned score = len;
1960 if (adjusted_offset < 4096)
1963 if (adjusted_offset < 256)
1969 static inline unsigned
1970 lzx_repeat_offset_match_score(unsigned rep_len, unsigned rep_idx)
1975 /* This is the "lazy" LZX compressor. */
1977 lzx_compress_lazy(struct lzx_compressor *c, struct lzx_output_bitstream *os,
1980 const u8 * const in_begin = c->in_buffer;
1981 const u8 * in_next = in_begin;
1982 const u8 * const in_end = in_begin + c->in_nbytes;
1983 unsigned max_len = LZX_MAX_MATCH_LEN;
1984 unsigned nice_len = min(c->nice_match_length, max_len);
1985 STATIC_ASSERT(LZX_NUM_RECENT_OFFSETS == 3);
1986 u32 recent_offsets[3] = {1, 1, 1};
1987 u32 next_hashes[2] = {};
1989 CALL_HC_MF(is_16_bit, c, hc_matchfinder_init);
1992 /* Starting a new block */
1994 const u8 * const in_block_begin = in_next;
1995 const u8 * const in_block_end =
1996 in_next + min(LZX_DIV_BLOCK_SIZE, in_end - in_next);
1997 struct lzx_sequence *next_seq = c->chosen_sequences;
2000 u32 cur_offset_data;
2004 u32 next_offset_data;
2005 unsigned next_score;
2006 unsigned rep_max_len;
2007 unsigned rep_max_idx;
2012 lzx_reset_symbol_frequencies(c);
2015 if (unlikely(max_len > in_end - in_next)) {
2016 max_len = in_end - in_next;
2017 nice_len = min(max_len, nice_len);
2020 /* Find the longest match at the current position. */
2022 cur_len = CALL_HC_MF(is_16_bit, c, hc_matchfinder_longest_match,
2028 c->max_search_depth,
2033 cur_offset >= 8192 - LZX_OFFSET_ADJUSTMENT &&
2034 cur_offset != recent_offsets[0] &&
2035 cur_offset != recent_offsets[1] &&
2036 cur_offset != recent_offsets[2]))
2038 /* There was no match found, or the only match found
2039 * was a distant length 3 match. Output a literal. */
2040 lzx_record_literal(c, *in_next++, &litrunlen);
2044 if (cur_offset == recent_offsets[0]) {
2046 cur_offset_data = 0;
2047 skip_len = cur_len - 1;
2048 goto choose_cur_match;
2051 cur_offset_data = cur_offset + LZX_OFFSET_ADJUSTMENT;
2052 cur_score = lzx_explicit_offset_match_score(cur_len, cur_offset_data);
2054 /* Consider a repeat offset match */
2055 rep_max_len = lzx_find_longest_repeat_offset_match(in_next,
2061 if (rep_max_len >= 3 &&
2062 (rep_score = lzx_repeat_offset_match_score(rep_max_len,
2063 rep_max_idx)) >= cur_score)
2065 cur_len = rep_max_len;
2066 cur_offset_data = rep_max_idx;
2067 skip_len = rep_max_len - 1;
2068 goto choose_cur_match;
2073 /* We have a match at the current position. */
2075 /* If we have a very long match, choose it immediately. */
2076 if (cur_len >= nice_len) {
2077 skip_len = cur_len - 1;
2078 goto choose_cur_match;
2081 /* See if there's a better match at the next position. */
2083 if (unlikely(max_len > in_end - in_next)) {
2084 max_len = in_end - in_next;
2085 nice_len = min(max_len, nice_len);
2088 next_len = CALL_HC_MF(is_16_bit, c, hc_matchfinder_longest_match,
2094 c->max_search_depth / 2,
2098 if (next_len <= cur_len - 2) {
2100 skip_len = cur_len - 2;
2101 goto choose_cur_match;
2104 next_offset_data = next_offset + LZX_OFFSET_ADJUSTMENT;
2105 next_score = lzx_explicit_offset_match_score(next_len, next_offset_data);
2107 rep_max_len = lzx_find_longest_repeat_offset_match(in_next,
2113 if (rep_max_len >= 3 &&
2114 (rep_score = lzx_repeat_offset_match_score(rep_max_len,
2115 rep_max_idx)) >= next_score)
2118 if (rep_score > cur_score) {
2119 /* The next match is better, and it's a
2120 * repeat offset match. */
2121 lzx_record_literal(c, *(in_next - 2),
2123 cur_len = rep_max_len;
2124 cur_offset_data = rep_max_idx;
2125 skip_len = cur_len - 1;
2126 goto choose_cur_match;
2129 if (next_score > cur_score) {
2130 /* The next match is better, and it's an
2131 * explicit offset match. */
2132 lzx_record_literal(c, *(in_next - 2),
2135 cur_offset_data = next_offset_data;
2136 cur_score = next_score;
2137 goto have_cur_match;
2141 /* The original match was better. */
2142 skip_len = cur_len - 2;
2145 lzx_record_match(c, cur_len, cur_offset_data,
2146 recent_offsets, is_16_bit,
2147 &litrunlen, &next_seq);
2148 in_next = CALL_HC_MF(is_16_bit, c, hc_matchfinder_skip_positions,
2154 } while (in_next < in_block_end);
2156 lzx_finish_sequence(next_seq, litrunlen);
2158 lzx_finish_block(c, os, in_block_begin, in_next - in_block_begin, 0);
2160 } while (in_next != in_end);
2164 lzx_compress_lazy_16(struct lzx_compressor *c, struct lzx_output_bitstream *os)
2166 lzx_compress_lazy(c, os, true);
2170 lzx_compress_lazy_32(struct lzx_compressor *c, struct lzx_output_bitstream *os)
2172 lzx_compress_lazy(c, os, false);
2175 /* Generate the acceleration tables for offset slots. */
2177 lzx_init_offset_slot_tabs(struct lzx_compressor *c)
2179 u32 adjusted_offset = 0;
2183 for (; adjusted_offset < ARRAY_LEN(c->offset_slot_tab_1);
2186 if (adjusted_offset >= lzx_offset_slot_base[slot + 1])
2188 c->offset_slot_tab_1[adjusted_offset] = slot;
2191 /* slots [30, 49] */
2192 for (; adjusted_offset < LZX_MAX_WINDOW_SIZE;
2193 adjusted_offset += (u32)1 << 14)
2195 if (adjusted_offset >= lzx_offset_slot_base[slot + 1])
2197 c->offset_slot_tab_2[adjusted_offset >> 14] = slot;
2202 lzx_get_compressor_size(size_t max_bufsize, unsigned compression_level)
2204 if (compression_level <= LZX_MAX_FAST_LEVEL) {
2205 if (lzx_is_16_bit(max_bufsize))
2206 return offsetof(struct lzx_compressor, hc_mf_16) +
2207 hc_matchfinder_size_16(max_bufsize);
2209 return offsetof(struct lzx_compressor, hc_mf_32) +
2210 hc_matchfinder_size_32(max_bufsize);
2212 if (lzx_is_16_bit(max_bufsize))
2213 return offsetof(struct lzx_compressor, bt_mf_16) +
2214 bt_matchfinder_size_16(max_bufsize);
2216 return offsetof(struct lzx_compressor, bt_mf_32) +
2217 bt_matchfinder_size_32(max_bufsize);
2222 lzx_get_needed_memory(size_t max_bufsize, unsigned compression_level,
2227 if (max_bufsize > LZX_MAX_WINDOW_SIZE)
2230 size += lzx_get_compressor_size(max_bufsize, compression_level);
2232 size += max_bufsize; /* in_buffer */
2237 lzx_create_compressor(size_t max_bufsize, unsigned compression_level,
2238 bool destructive, void **c_ret)
2240 unsigned window_order;
2241 struct lzx_compressor *c;
2243 window_order = lzx_get_window_order(max_bufsize);
2244 if (window_order == 0)
2245 return WIMLIB_ERR_INVALID_PARAM;
2247 c = MALLOC(lzx_get_compressor_size(max_bufsize, compression_level));
2251 c->destructive = destructive;
2253 c->num_main_syms = lzx_get_num_main_syms(window_order);
2254 c->window_order = window_order;
2256 if (!c->destructive) {
2257 c->in_buffer = MALLOC(max_bufsize);
2262 if (compression_level <= LZX_MAX_FAST_LEVEL) {
2264 /* Fast compression: Use lazy parsing. */
2266 if (lzx_is_16_bit(max_bufsize))
2267 c->impl = lzx_compress_lazy_16;
2269 c->impl = lzx_compress_lazy_32;
2270 c->max_search_depth = (60 * compression_level) / 20;
2271 c->nice_match_length = (80 * compression_level) / 20;
2273 /* lzx_compress_lazy() needs max_search_depth >= 2 because it
2274 * halves the max_search_depth when attempting a lazy match, and
2275 * max_search_depth cannot be 0. */
2276 if (c->max_search_depth < 2)
2277 c->max_search_depth = 2;
2280 /* Normal / high compression: Use near-optimal parsing. */
2282 if (lzx_is_16_bit(max_bufsize))
2283 c->impl = lzx_compress_near_optimal_16;
2285 c->impl = lzx_compress_near_optimal_32;
2287 /* Scale nice_match_length and max_search_depth with the
2288 * compression level. */
2289 c->max_search_depth = (24 * compression_level) / 50;
2290 c->nice_match_length = (48 * compression_level) / 50;
2292 /* Set a number of optimization passes appropriate for the
2293 * compression level. */
2295 c->num_optim_passes = 1;
2297 if (compression_level >= 45)
2298 c->num_optim_passes++;
2300 /* Use more optimization passes for higher compression levels.
2301 * But the more passes there are, the less they help --- so
2302 * don't add them linearly. */
2303 if (compression_level >= 70) {
2304 c->num_optim_passes++;
2305 if (compression_level >= 100)
2306 c->num_optim_passes++;
2307 if (compression_level >= 150)
2308 c->num_optim_passes++;
2309 if (compression_level >= 200)
2310 c->num_optim_passes++;
2311 if (compression_level >= 300)
2312 c->num_optim_passes++;
2316 /* max_search_depth == 0 is invalid. */
2317 if (c->max_search_depth < 1)
2318 c->max_search_depth = 1;
2320 if (c->nice_match_length > LZX_MAX_MATCH_LEN)
2321 c->nice_match_length = LZX_MAX_MATCH_LEN;
2323 lzx_init_offset_slot_tabs(c);
2330 return WIMLIB_ERR_NOMEM;
2334 lzx_compress(const void *restrict in, size_t in_nbytes,
2335 void *restrict out, size_t out_nbytes_avail, void *restrict _c)
2337 struct lzx_compressor *c = _c;
2338 struct lzx_output_bitstream os;
2341 /* Don't bother trying to compress very small inputs. */
2342 if (in_nbytes < 100)
2345 /* Copy the input data into the internal buffer and preprocess it. */
2347 c->in_buffer = (void *)in;
2349 memcpy(c->in_buffer, in, in_nbytes);
2350 c->in_nbytes = in_nbytes;
2351 lzx_preprocess(c->in_buffer, in_nbytes);
2353 /* Initially, the previous Huffman codeword lengths are all zeroes. */
2355 memset(&c->codes[1].lens, 0, sizeof(struct lzx_lens));
2357 /* Initialize the output bitstream. */
2358 lzx_init_output(&os, out, out_nbytes_avail);
2360 /* Call the compression level-specific compress() function. */
2363 /* Flush the output bitstream and return the compressed size or 0. */
2364 result = lzx_flush_output(&os);
2365 if (!result && c->destructive)
2366 lzx_postprocess(c->in_buffer, c->in_nbytes);
2371 lzx_free_compressor(void *_c)
2373 struct lzx_compressor *c = _c;
2375 if (!c->destructive)
2380 const struct compressor_ops lzx_compressor_ops = {
2381 .get_needed_memory = lzx_get_needed_memory,
2382 .create_compressor = lzx_create_compressor,
2383 .compress = lzx_compress,
2384 .free_compressor = lzx_free_compressor,