4 * A compressor for the LZX compression format, as used in WIM files.
8 * Copyright (C) 2012-2016 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 * The compressor always chooses a block of at least MIN_BLOCK_SIZE bytes,
69 * except if the last block has to be shorter.
71 #define MIN_BLOCK_SIZE 6500
74 * The compressor attempts to end blocks after SOFT_MAX_BLOCK_SIZE bytes, but
75 * the final size might be larger due to matches extending beyond the end of the
76 * block. Specifically:
78 * - The greedy parser may choose an arbitrarily long match starting at the
79 * SOFT_MAX_BLOCK_SIZE'th byte.
81 * - The lazy parser may choose a sequence of literals starting at the
82 * SOFT_MAX_BLOCK_SIZE'th byte when it sees a sequence of increasing good
83 * matches. The final match may be of arbitrary length. The length of the
84 * literal sequence is approximately limited by the "nice match length"
87 #define SOFT_MAX_BLOCK_SIZE 100000
90 * LZX_CACHE_LENGTH is the number of lz_match structures in the match cache,
91 * excluding the extra "overflow" entries. This value should be high enough so
92 * that nearly the time, all matches found in a given block can fit in the match
93 * cache. However, fallback behavior (immediately terminating the block) on
94 * cache overflow is still required.
96 #define LZX_CACHE_LENGTH (SOFT_MAX_BLOCK_SIZE * 5)
99 * LZX_MAX_MATCHES_PER_POS is an upper bound on the number of matches that can
100 * ever be saved in the match cache for a single position. Since each match we
101 * save for a single position has a distinct length, we can use the number of
102 * possible match lengths in LZX as this bound. This bound is guaranteed to be
103 * valid in all cases, although if 'nice_match_length < LZX_MAX_MATCH_LEN', then
104 * it will never actually be reached.
106 #define LZX_MAX_MATCHES_PER_POS LZX_NUM_LENS
109 * LZX_BIT_COST is a scaling factor that represents the cost to output one bit.
110 * This makes it possible to consider fractional bit costs.
112 * Note: this is only useful as a statistical trick for when the true costs are
113 * unknown. In reality, each token in LZX requires a whole number of bits to
116 #define LZX_BIT_COST 16
119 * Should the compressor take into account the costs of aligned offset symbols?
121 #define LZX_CONSIDER_ALIGNED_COSTS 1
124 * LZX_MAX_FAST_LEVEL is the maximum compression level at which we use the
127 #define LZX_MAX_FAST_LEVEL 34
130 * BT_MATCHFINDER_HASH2_ORDER is the log base 2 of the number of entries in the
131 * hash table for finding length 2 matches. This could be as high as 16, but
132 * using a smaller hash table speeds up compression due to reduced cache
135 #define BT_MATCHFINDER_HASH2_ORDER 12
138 * These are the compressor-side limits on the codeword lengths for each Huffman
139 * code. To make outputting bits slightly faster, some of these limits are
140 * lower than the limits defined by the LZX format. This does not significantly
141 * affect the compression ratio, at least for the block sizes we use.
143 #define MAIN_CODEWORD_LIMIT 12 /* 64-bit: can buffer 4 main symbols */
144 #define LENGTH_CODEWORD_LIMIT 12
145 #define ALIGNED_CODEWORD_LIMIT 7
146 #define PRE_CODEWORD_LIMIT 7
148 #include "wimlib/compress_common.h"
149 #include "wimlib/compressor_ops.h"
150 #include "wimlib/error.h"
151 #include "wimlib/lz_extend.h"
152 #include "wimlib/lzx_common.h"
153 #include "wimlib/unaligned.h"
154 #include "wimlib/util.h"
156 /* Matchfinders with 16-bit positions */
158 #define MF_SUFFIX _16
159 #include "wimlib/bt_matchfinder.h"
160 #include "wimlib/hc_matchfinder.h"
162 /* Matchfinders with 32-bit positions */
166 #define MF_SUFFIX _32
167 #include "wimlib/bt_matchfinder.h"
168 #include "wimlib/hc_matchfinder.h"
170 struct lzx_output_bitstream;
172 /* Codewords for the LZX Huffman codes. */
173 struct lzx_codewords {
174 u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
175 u32 len[LZX_LENCODE_NUM_SYMBOLS];
176 u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
179 /* Codeword lengths (in bits) for the LZX Huffman codes.
180 * A zero length means the corresponding codeword has zero frequency. */
182 u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS + 1];
183 u8 len[LZX_LENCODE_NUM_SYMBOLS + 1];
184 u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
187 /* Cost model for near-optimal parsing */
190 /* 'match_cost[offset_slot][len - LZX_MIN_MATCH_LEN]' is the cost for a
191 * length 'len' match that has an offset belonging to 'offset_slot'. */
192 u32 match_cost[LZX_MAX_OFFSET_SLOTS][LZX_NUM_LENS];
194 /* Cost for each symbol in the main code */
195 u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
197 /* Cost for each symbol in the length code */
198 u32 len[LZX_LENCODE_NUM_SYMBOLS];
200 #if LZX_CONSIDER_ALIGNED_COSTS
201 /* Cost for each symbol in the aligned code */
202 u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
206 /* Codewords and lengths for the LZX Huffman codes. */
208 struct lzx_codewords codewords;
209 struct lzx_lens lens;
212 /* Symbol frequency counters for the LZX Huffman codes. */
214 u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
215 u32 len[LZX_LENCODE_NUM_SYMBOLS];
216 u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
219 /* Block split statistics. See "Block splitting algorithm" below. */
220 #define NUM_LITERAL_OBSERVATION_TYPES 8
221 #define NUM_MATCH_OBSERVATION_TYPES 2
222 #define NUM_OBSERVATION_TYPES (NUM_LITERAL_OBSERVATION_TYPES + NUM_MATCH_OBSERVATION_TYPES)
223 struct block_split_stats {
224 u32 new_observations[NUM_OBSERVATION_TYPES];
225 u32 observations[NUM_OBSERVATION_TYPES];
226 u32 num_new_observations;
227 u32 num_observations;
231 * Represents a run of literals followed by a match or end-of-block. This
232 * struct is needed to temporarily store items chosen by the parser, since items
233 * cannot be written until all items for the block have been chosen and the
234 * block's Huffman codes have been computed.
236 struct lzx_sequence {
238 /* The number of literals in the run. This may be 0. The literals are
239 * not stored explicitly in this structure; instead, they are read
240 * directly from the uncompressed data. */
243 /* If the next field doesn't indicate end-of-block, then this is the
244 * match length minus LZX_MIN_MATCH_LEN. */
247 /* If bit 31 is clear, then this field contains the match header in bits
248 * 0-8, and either the match offset plus LZX_OFFSET_ADJUSTMENT or a
249 * recent offset code in bits 9-30. Otherwise (if bit 31 is set), this
250 * sequence's literal run was the last literal run in the block, so
251 * there is no match that follows it. */
252 u32 adjusted_offset_and_match_hdr;
256 * This structure represents a byte position in the input buffer and a node in
257 * the graph of possible match/literal choices.
259 * Logically, each incoming edge to this node is labeled with a literal or a
260 * match that can be taken to reach this position from an earlier position; and
261 * each outgoing edge from this node is labeled with a literal or a match that
262 * can be taken to advance from this position to a later position.
264 struct lzx_optimum_node {
266 /* The cost, in bits, of the lowest-cost path that has been found to
267 * reach this position. This can change as progressively lower cost
268 * paths are found to reach this position. */
272 * The match or literal that was taken to reach this position. This can
273 * change as progressively lower cost paths are found to reach this
276 * This variable is divided into two bitfields.
279 * Low bits are 0, high bits are the literal.
281 * Explicit offset matches:
282 * Low bits are the match length, high bits are the offset plus 2.
284 * Repeat offset matches:
285 * Low bits are the match length, high bits are the queue index.
288 #define OPTIMUM_OFFSET_SHIFT 9
289 #define OPTIMUM_LEN_MASK ((1 << OPTIMUM_OFFSET_SHIFT) - 1)
290 } _aligned_attribute(8);
293 * Least-recently-used queue for match offsets.
295 * This is represented as a 64-bit integer for efficiency. There are three
296 * offsets of 21 bits each. Bit 64 is garbage.
298 struct lzx_lru_queue {
302 #define LZX_QUEUE64_OFFSET_SHIFT 21
303 #define LZX_QUEUE64_OFFSET_MASK (((u64)1 << LZX_QUEUE64_OFFSET_SHIFT) - 1)
305 #define LZX_QUEUE64_R0_SHIFT (0 * LZX_QUEUE64_OFFSET_SHIFT)
306 #define LZX_QUEUE64_R1_SHIFT (1 * LZX_QUEUE64_OFFSET_SHIFT)
307 #define LZX_QUEUE64_R2_SHIFT (2 * LZX_QUEUE64_OFFSET_SHIFT)
309 #define LZX_QUEUE64_R0_MASK (LZX_QUEUE64_OFFSET_MASK << LZX_QUEUE64_R0_SHIFT)
310 #define LZX_QUEUE64_R1_MASK (LZX_QUEUE64_OFFSET_MASK << LZX_QUEUE64_R1_SHIFT)
311 #define LZX_QUEUE64_R2_MASK (LZX_QUEUE64_OFFSET_MASK << LZX_QUEUE64_R2_SHIFT)
314 lzx_lru_queue_init(struct lzx_lru_queue *queue)
316 queue->R = ((u64)1 << LZX_QUEUE64_R0_SHIFT) |
317 ((u64)1 << LZX_QUEUE64_R1_SHIFT) |
318 ((u64)1 << LZX_QUEUE64_R2_SHIFT);
322 lzx_lru_queue_R0(struct lzx_lru_queue queue)
324 return (queue.R >> LZX_QUEUE64_R0_SHIFT) & LZX_QUEUE64_OFFSET_MASK;
328 lzx_lru_queue_R1(struct lzx_lru_queue queue)
330 return (queue.R >> LZX_QUEUE64_R1_SHIFT) & LZX_QUEUE64_OFFSET_MASK;
334 lzx_lru_queue_R2(struct lzx_lru_queue queue)
336 return (queue.R >> LZX_QUEUE64_R2_SHIFT) & LZX_QUEUE64_OFFSET_MASK;
339 /* Push a match offset onto the front (most recently used) end of the queue. */
340 static inline struct lzx_lru_queue
341 lzx_lru_queue_push(struct lzx_lru_queue queue, u32 offset)
343 return (struct lzx_lru_queue) {
344 .R = (queue.R << LZX_QUEUE64_OFFSET_SHIFT) | offset,
348 /* Swap a match offset to the front of the queue. */
349 static inline struct lzx_lru_queue
350 lzx_lru_queue_swap(struct lzx_lru_queue queue, unsigned idx)
356 return (struct lzx_lru_queue) {
357 .R = (lzx_lru_queue_R1(queue) << LZX_QUEUE64_R0_SHIFT) |
358 (lzx_lru_queue_R0(queue) << LZX_QUEUE64_R1_SHIFT) |
359 (queue.R & LZX_QUEUE64_R2_MASK),
362 return (struct lzx_lru_queue) {
363 .R = (lzx_lru_queue_R2(queue) << LZX_QUEUE64_R0_SHIFT) |
364 (queue.R & LZX_QUEUE64_R1_MASK) |
365 (lzx_lru_queue_R0(queue) << LZX_QUEUE64_R2_SHIFT),
369 /* The main LZX compressor structure */
370 struct lzx_compressor {
372 /* The "nice" match length: if a match of this length is found, then
373 * choose it immediately without further consideration. */
374 unsigned nice_match_length;
376 /* The maximum search depth: consider at most this many potential
377 * matches at each position. */
378 unsigned max_search_depth;
380 /* The log base 2 of the LZX window size for LZ match offset encoding
381 * purposes. This will be >= LZX_MIN_WINDOW_ORDER and <=
382 * LZX_MAX_WINDOW_ORDER. */
383 unsigned window_order;
385 /* The number of symbols in the main alphabet. This depends on
386 * @window_order, since @window_order determines the maximum possible
388 unsigned num_main_syms;
390 /* Number of optimization passes per block */
391 unsigned num_optim_passes;
393 /* The preprocessed buffer of data being compressed */
396 /* The number of bytes of data to be compressed, which is the number of
397 * bytes of data in @in_buffer that are actually valid. */
400 /* Pointer to the compress() implementation chosen at allocation time */
401 void (*impl)(struct lzx_compressor *, struct lzx_output_bitstream *);
403 /* If true, the compressor need not preserve the input buffer if it
404 * compresses the data successfully. */
407 /* The Huffman symbol frequency counters for the current block. */
408 struct lzx_freqs freqs;
410 /* Block split statistics. */
411 struct block_split_stats split_stats;
413 /* The Huffman codes for the current and previous blocks. The one with
414 * index 'codes_index' is for the current block, and the other one is
415 * for the previous block. */
416 struct lzx_codes codes[2];
417 unsigned codes_index;
419 /* The matches and literals that the parser has chosen for the current
420 * block. The required length of this array is limited by the maximum
421 * number of matches that can ever be chosen for a single block, plus
422 * one for the special entry at the end. */
423 struct lzx_sequence chosen_sequences[
424 DIV_ROUND_UP(SOFT_MAX_BLOCK_SIZE, LZX_MIN_MATCH_LEN) + 1];
426 /* Tables for mapping adjusted offsets to offset slots */
428 /* offset slots [0, 29] */
429 u8 offset_slot_tab_1[32768];
431 /* offset slots [30, 49] */
432 u8 offset_slot_tab_2[128];
435 /* Data for greedy or lazy parsing */
437 /* Hash chains matchfinder (MUST BE LAST!!!) */
439 struct hc_matchfinder_16 hc_mf_16;
440 struct hc_matchfinder_32 hc_mf_32;
444 /* Data for near-optimal parsing */
447 * Array of nodes, one per position, for running the
448 * minimum-cost path algorithm.
450 * This array must be large enough to accommodate the
451 * worst-case number of nodes, which occurs if we find a
452 * match of length LZX_MAX_MATCH_LEN at position
453 * SOFT_MAX_BLOCK_SIZE - 1, producing a block of length
454 * SOFT_MAX_BLOCK_SIZE - 1 + LZX_MAX_MATCH_LEN. Add one
455 * for the end-of-block node.
457 struct lzx_optimum_node optimum_nodes[SOFT_MAX_BLOCK_SIZE - 1 +
458 LZX_MAX_MATCH_LEN + 1];
460 /* The cost model for the current block */
461 struct lzx_costs costs;
464 * Cached matches for the current block. This array
465 * contains the matches that were found at each position
466 * in the block. Specifically, for each position, there
467 * is a special 'struct lz_match' whose 'length' field
468 * contains the number of matches that were found at
469 * that position; this is followed by the matches
470 * themselves, if any, sorted by strictly increasing
473 * Note: in rare cases, there will be a very high number
474 * of matches in the block and this array will overflow.
475 * If this happens, we force the end of the current
476 * block. LZX_CACHE_LENGTH is the length at which we
477 * actually check for overflow. The extra slots beyond
478 * this are enough to absorb the worst case overflow,
479 * which occurs if starting at
480 * &match_cache[LZX_CACHE_LENGTH - 1], we write the
481 * match count header, then write
482 * LZX_MAX_MATCHES_PER_POS matches, then skip searching
483 * for matches at 'LZX_MAX_MATCH_LEN - 1' positions and
484 * write the match count header for each.
486 struct lz_match match_cache[LZX_CACHE_LENGTH +
487 LZX_MAX_MATCHES_PER_POS +
488 LZX_MAX_MATCH_LEN - 1];
490 /* Binary trees matchfinder (MUST BE LAST!!!) */
492 struct bt_matchfinder_16 bt_mf_16;
493 struct bt_matchfinder_32 bt_mf_32;
500 * Will a matchfinder using 16-bit positions be sufficient for compressing
501 * buffers of up to the specified size? The limit could be 65536 bytes, but we
502 * also want to optimize out the use of offset_slot_tab_2 in the 16-bit case.
503 * This requires that the limit be no more than the length of offset_slot_tab_1
507 lzx_is_16_bit(size_t max_bufsize)
509 STATIC_ASSERT(ARRAY_LEN(((struct lzx_compressor *)0)->offset_slot_tab_1) == 32768);
510 return max_bufsize <= 32768;
514 * The following macros call either the 16-bit or the 32-bit version of a
515 * matchfinder function based on the value of 'is_16_bit', which will be known
516 * at compilation time.
519 #define CALL_HC_MF(is_16_bit, c, funcname, ...) \
520 ((is_16_bit) ? CONCAT(funcname, _16)(&(c)->hc_mf_16, ##__VA_ARGS__) : \
521 CONCAT(funcname, _32)(&(c)->hc_mf_32, ##__VA_ARGS__));
523 #define CALL_BT_MF(is_16_bit, c, funcname, ...) \
524 ((is_16_bit) ? CONCAT(funcname, _16)(&(c)->bt_mf_16, ##__VA_ARGS__) : \
525 CONCAT(funcname, _32)(&(c)->bt_mf_32, ##__VA_ARGS__));
528 * Structure to keep track of the current state of sending bits to the
529 * compressed output buffer.
531 * The LZX bitstream is encoded as a sequence of 16-bit coding units.
533 struct lzx_output_bitstream {
535 /* Bits that haven't yet been written to the output buffer. */
536 machine_word_t bitbuf;
538 /* Number of bits currently held in @bitbuf. */
541 /* Pointer to the start of the output buffer. */
544 /* Pointer to the position in the output buffer at which the next coding
545 * unit should be written. */
548 /* Pointer just past the end of the output buffer, rounded down to a
549 * 2-byte boundary. */
553 /* Can the specified number of bits always be added to 'bitbuf' after any
554 * pending 16-bit coding units have been flushed? */
555 #define CAN_BUFFER(n) ((n) <= (8 * sizeof(machine_word_t)) - 15)
558 * Initialize the output bitstream.
561 * The output bitstream structure to initialize.
563 * The buffer being written to.
565 * Size of @buffer, in bytes.
568 lzx_init_output(struct lzx_output_bitstream *os, void *buffer, size_t size)
573 os->next = os->start;
574 os->end = os->start + (size & ~1);
577 /* Add some bits to the bitbuffer variable of the output bitstream. The caller
578 * must make sure there is enough room. */
580 lzx_add_bits(struct lzx_output_bitstream *os, u32 bits, unsigned num_bits)
582 os->bitbuf = (os->bitbuf << num_bits) | bits;
583 os->bitcount += num_bits;
586 /* Flush bits from the bitbuffer variable to the output buffer. 'max_num_bits'
587 * specifies the maximum number of bits that may have been added since the last
590 lzx_flush_bits(struct lzx_output_bitstream *os, unsigned max_num_bits)
592 /* Masking the number of bits to shift is only needed to avoid undefined
593 * behavior; we don't actually care about the results of bad shifts. On
594 * x86, the explicit masking generates no extra code. */
595 const u32 shift_mask = 8 * sizeof(os->bitbuf) - 1;
597 if (os->end - os->next < 6)
599 put_unaligned_le16(os->bitbuf >> ((os->bitcount - 16) &
600 shift_mask), os->next + 0);
601 if (max_num_bits > 16)
602 put_unaligned_le16(os->bitbuf >> ((os->bitcount - 32) &
603 shift_mask), os->next + 2);
604 if (max_num_bits > 32)
605 put_unaligned_le16(os->bitbuf >> ((os->bitcount - 48) &
606 shift_mask), os->next + 4);
607 os->next += (os->bitcount >> 4) << 1;
611 /* Add at most 16 bits to the bitbuffer and flush it. */
613 lzx_write_bits(struct lzx_output_bitstream *os, u32 bits, unsigned num_bits)
615 lzx_add_bits(os, bits, num_bits);
616 lzx_flush_bits(os, 16);
620 * Flush the last coding unit to the output buffer if needed. Return the total
621 * number of bytes written to the output buffer, or 0 if an overflow occurred.
624 lzx_flush_output(struct lzx_output_bitstream *os)
626 if (os->end - os->next < 6)
629 if (os->bitcount != 0) {
630 put_unaligned_le16(os->bitbuf << (16 - os->bitcount), os->next);
634 return os->next - os->start;
637 /* Build the main, length, and aligned offset Huffman codes used in LZX.
639 * This takes as input the frequency tables for each code and produces as output
640 * a set of tables that map symbols to codewords and codeword lengths. */
642 lzx_make_huffman_codes(struct lzx_compressor *c)
644 const struct lzx_freqs *freqs = &c->freqs;
645 struct lzx_codes *codes = &c->codes[c->codes_index];
647 STATIC_ASSERT(MAIN_CODEWORD_LIMIT >= 9 &&
648 MAIN_CODEWORD_LIMIT <= LZX_MAX_MAIN_CODEWORD_LEN);
649 STATIC_ASSERT(LENGTH_CODEWORD_LIMIT >= 8 &&
650 LENGTH_CODEWORD_LIMIT <= LZX_MAX_LEN_CODEWORD_LEN);
651 STATIC_ASSERT(ALIGNED_CODEWORD_LIMIT >= LZX_NUM_ALIGNED_OFFSET_BITS &&
652 ALIGNED_CODEWORD_LIMIT <= LZX_MAX_ALIGNED_CODEWORD_LEN);
654 make_canonical_huffman_code(c->num_main_syms,
658 codes->codewords.main);
660 make_canonical_huffman_code(LZX_LENCODE_NUM_SYMBOLS,
661 LENGTH_CODEWORD_LIMIT,
664 codes->codewords.len);
666 make_canonical_huffman_code(LZX_ALIGNEDCODE_NUM_SYMBOLS,
667 ALIGNED_CODEWORD_LIMIT,
670 codes->codewords.aligned);
673 /* Reset the symbol frequencies for the LZX Huffman codes. */
675 lzx_reset_symbol_frequencies(struct lzx_compressor *c)
677 memset(&c->freqs, 0, sizeof(c->freqs));
681 lzx_compute_precode_items(const u8 lens[restrict],
682 const u8 prev_lens[restrict],
683 u32 precode_freqs[restrict],
684 unsigned precode_items[restrict])
693 itemptr = precode_items;
696 while (!((len = lens[run_start]) & 0x80)) {
698 /* len = the length being repeated */
700 /* Find the next run of codeword lengths. */
702 run_end = run_start + 1;
704 /* Fast case for a single length. */
705 if (likely(len != lens[run_end])) {
706 delta = prev_lens[run_start] - len;
709 precode_freqs[delta]++;
715 /* Extend the run. */
718 } while (len == lens[run_end]);
723 /* Symbol 18: RLE 20 to 51 zeroes at a time. */
724 while ((run_end - run_start) >= 20) {
725 extra_bits = min((run_end - run_start) - 20, 0x1f);
727 *itemptr++ = 18 | (extra_bits << 5);
728 run_start += 20 + extra_bits;
731 /* Symbol 17: RLE 4 to 19 zeroes at a time. */
732 if ((run_end - run_start) >= 4) {
733 extra_bits = min((run_end - run_start) - 4, 0xf);
735 *itemptr++ = 17 | (extra_bits << 5);
736 run_start += 4 + extra_bits;
740 /* A run of nonzero lengths. */
742 /* Symbol 19: RLE 4 to 5 of any length at a time. */
743 while ((run_end - run_start) >= 4) {
744 extra_bits = (run_end - run_start) > 4;
745 delta = prev_lens[run_start] - len;
749 precode_freqs[delta]++;
750 *itemptr++ = 19 | (extra_bits << 5) | (delta << 6);
751 run_start += 4 + extra_bits;
755 /* Output any remaining lengths without RLE. */
756 while (run_start != run_end) {
757 delta = prev_lens[run_start] - len;
760 precode_freqs[delta]++;
766 return itemptr - precode_items;
770 * Output a Huffman code in the compressed form used in LZX.
772 * The Huffman code is represented in the output as a logical series of codeword
773 * lengths from which the Huffman code, which must be in canonical form, can be
776 * The codeword lengths are themselves compressed using a separate Huffman code,
777 * the "precode", which contains a symbol for each possible codeword length in
778 * the larger code as well as several special symbols to represent repeated
779 * codeword lengths (a form of run-length encoding). The precode is itself
780 * constructed in canonical form, and its codeword lengths are represented
781 * literally in 20 4-bit fields that immediately precede the compressed codeword
782 * lengths of the larger code.
784 * Furthermore, the codeword lengths of the larger code are actually represented
785 * as deltas from the codeword lengths of the corresponding code in the previous
789 * Bitstream to which to write the compressed Huffman code.
791 * The codeword lengths, indexed by symbol, in the Huffman code.
793 * The codeword lengths, indexed by symbol, in the corresponding Huffman
794 * code in the previous block, or all zeroes if this is the first block.
796 * The number of symbols in the Huffman code.
799 lzx_write_compressed_code(struct lzx_output_bitstream *os,
800 const u8 lens[restrict],
801 const u8 prev_lens[restrict],
804 u32 precode_freqs[LZX_PRECODE_NUM_SYMBOLS];
805 u8 precode_lens[LZX_PRECODE_NUM_SYMBOLS];
806 u32 precode_codewords[LZX_PRECODE_NUM_SYMBOLS];
807 unsigned precode_items[num_lens];
808 unsigned num_precode_items;
809 unsigned precode_item;
810 unsigned precode_sym;
812 u8 saved = lens[num_lens];
813 *(u8 *)(lens + num_lens) = 0x80;
815 for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++)
816 precode_freqs[i] = 0;
818 /* Compute the "items" (RLE / literal tokens and extra bits) with which
819 * the codeword lengths in the larger code will be output. */
820 num_precode_items = lzx_compute_precode_items(lens,
825 /* Build the precode. */
826 STATIC_ASSERT(PRE_CODEWORD_LIMIT >= 5 &&
827 PRE_CODEWORD_LIMIT <= LZX_MAX_PRE_CODEWORD_LEN);
828 make_canonical_huffman_code(LZX_PRECODE_NUM_SYMBOLS,
830 precode_freqs, precode_lens,
833 /* Output the lengths of the codewords in the precode. */
834 for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++)
835 lzx_write_bits(os, precode_lens[i], LZX_PRECODE_ELEMENT_SIZE);
837 /* Output the encoded lengths of the codewords in the larger code. */
838 for (i = 0; i < num_precode_items; i++) {
839 precode_item = precode_items[i];
840 precode_sym = precode_item & 0x1F;
841 lzx_add_bits(os, precode_codewords[precode_sym],
842 precode_lens[precode_sym]);
843 if (precode_sym >= 17) {
844 if (precode_sym == 17) {
845 lzx_add_bits(os, precode_item >> 5, 4);
846 } else if (precode_sym == 18) {
847 lzx_add_bits(os, precode_item >> 5, 5);
849 lzx_add_bits(os, (precode_item >> 5) & 1, 1);
850 precode_sym = precode_item >> 6;
851 lzx_add_bits(os, precode_codewords[precode_sym],
852 precode_lens[precode_sym]);
855 STATIC_ASSERT(CAN_BUFFER(2 * PRE_CODEWORD_LIMIT + 1));
856 lzx_flush_bits(os, 2 * PRE_CODEWORD_LIMIT + 1);
859 *(u8 *)(lens + num_lens) = saved;
863 * Write all matches and literal bytes (which were precomputed) in an LZX
864 * compressed block to the output bitstream in the final compressed
868 * The output bitstream.
870 * The chosen type of the LZX compressed block (LZX_BLOCKTYPE_ALIGNED or
871 * LZX_BLOCKTYPE_VERBATIM).
873 * The uncompressed data of the block.
875 * The matches and literals to output, given as a series of sequences.
877 * The main, length, and aligned offset Huffman codes for the current
878 * LZX compressed block.
881 lzx_write_sequences(struct lzx_output_bitstream *os, int block_type,
882 const u8 *block_data, const struct lzx_sequence sequences[],
883 const struct lzx_codes *codes)
885 const struct lzx_sequence *seq = sequences;
886 u32 ones_if_aligned = 0 - (block_type == LZX_BLOCKTYPE_ALIGNED);
889 /* Output the next sequence. */
891 unsigned litrunlen = seq->litrunlen;
893 unsigned main_symbol;
894 unsigned adjusted_length;
896 unsigned offset_slot;
897 unsigned num_extra_bits;
900 /* Output the literal run of the sequence. */
902 if (litrunlen) { /* Is the literal run nonempty? */
904 /* Verify optimization is enabled on 64-bit */
905 STATIC_ASSERT(sizeof(machine_word_t) < 8 ||
906 CAN_BUFFER(4 * MAIN_CODEWORD_LIMIT));
908 if (CAN_BUFFER(4 * MAIN_CODEWORD_LIMIT)) {
910 /* 64-bit: write 4 literals at a time. */
911 while (litrunlen >= 4) {
912 unsigned lit0 = block_data[0];
913 unsigned lit1 = block_data[1];
914 unsigned lit2 = block_data[2];
915 unsigned lit3 = block_data[3];
916 lzx_add_bits(os, codes->codewords.main[lit0],
917 codes->lens.main[lit0]);
918 lzx_add_bits(os, codes->codewords.main[lit1],
919 codes->lens.main[lit1]);
920 lzx_add_bits(os, codes->codewords.main[lit2],
921 codes->lens.main[lit2]);
922 lzx_add_bits(os, codes->codewords.main[lit3],
923 codes->lens.main[lit3]);
924 lzx_flush_bits(os, 4 * MAIN_CODEWORD_LIMIT);
929 unsigned lit = *block_data++;
930 lzx_add_bits(os, codes->codewords.main[lit],
931 codes->lens.main[lit]);
933 unsigned lit = *block_data++;
934 lzx_add_bits(os, codes->codewords.main[lit],
935 codes->lens.main[lit]);
937 unsigned lit = *block_data++;
938 lzx_add_bits(os, codes->codewords.main[lit],
939 codes->lens.main[lit]);
940 lzx_flush_bits(os, 3 * MAIN_CODEWORD_LIMIT);
942 lzx_flush_bits(os, 2 * MAIN_CODEWORD_LIMIT);
945 lzx_flush_bits(os, 1 * MAIN_CODEWORD_LIMIT);
949 /* 32-bit: write 1 literal at a time. */
951 unsigned lit = *block_data++;
952 lzx_add_bits(os, codes->codewords.main[lit],
953 codes->lens.main[lit]);
954 lzx_flush_bits(os, MAIN_CODEWORD_LIMIT);
955 } while (--litrunlen);
959 /* Was this the last literal run? */
960 if (seq->adjusted_offset_and_match_hdr & 0x80000000)
963 /* Nope; output the match. */
965 match_hdr = seq->adjusted_offset_and_match_hdr & 0x1FF;
966 main_symbol = LZX_NUM_CHARS + match_hdr;
967 adjusted_length = seq->adjusted_length;
969 block_data += adjusted_length + LZX_MIN_MATCH_LEN;
971 offset_slot = match_hdr / LZX_NUM_LEN_HEADERS;
972 adjusted_offset = seq->adjusted_offset_and_match_hdr >> 9;
974 num_extra_bits = lzx_extra_offset_bits[offset_slot];
975 extra_bits = adjusted_offset - lzx_offset_slot_base[offset_slot];
977 #define MAX_MATCH_BITS (MAIN_CODEWORD_LIMIT + LENGTH_CODEWORD_LIMIT + \
978 14 + ALIGNED_CODEWORD_LIMIT)
980 /* Verify optimization is enabled on 64-bit */
981 STATIC_ASSERT(sizeof(machine_word_t) < 8 || CAN_BUFFER(MAX_MATCH_BITS));
983 /* Output the main symbol for the match. */
985 lzx_add_bits(os, codes->codewords.main[main_symbol],
986 codes->lens.main[main_symbol]);
987 if (!CAN_BUFFER(MAX_MATCH_BITS))
988 lzx_flush_bits(os, MAIN_CODEWORD_LIMIT);
990 /* If needed, output the length symbol for the match. */
992 if (adjusted_length >= LZX_NUM_PRIMARY_LENS) {
993 lzx_add_bits(os, codes->codewords.len[adjusted_length -
994 LZX_NUM_PRIMARY_LENS],
995 codes->lens.len[adjusted_length -
996 LZX_NUM_PRIMARY_LENS]);
997 if (!CAN_BUFFER(MAX_MATCH_BITS))
998 lzx_flush_bits(os, LENGTH_CODEWORD_LIMIT);
1001 /* Output the extra offset bits for the match. In aligned
1002 * offset blocks, the lowest 3 bits of the adjusted offset are
1003 * Huffman-encoded using the aligned offset code, provided that
1004 * there are at least extra 3 offset bits required. All other
1005 * extra offset bits are output verbatim. */
1007 if ((adjusted_offset & ones_if_aligned) >= 16) {
1009 lzx_add_bits(os, extra_bits >> LZX_NUM_ALIGNED_OFFSET_BITS,
1010 num_extra_bits - LZX_NUM_ALIGNED_OFFSET_BITS);
1011 if (!CAN_BUFFER(MAX_MATCH_BITS))
1012 lzx_flush_bits(os, 14);
1014 lzx_add_bits(os, codes->codewords.aligned[adjusted_offset &
1015 LZX_ALIGNED_OFFSET_BITMASK],
1016 codes->lens.aligned[adjusted_offset &
1017 LZX_ALIGNED_OFFSET_BITMASK]);
1018 if (!CAN_BUFFER(MAX_MATCH_BITS))
1019 lzx_flush_bits(os, ALIGNED_CODEWORD_LIMIT);
1021 STATIC_ASSERT(CAN_BUFFER(17));
1023 lzx_add_bits(os, extra_bits, num_extra_bits);
1024 if (!CAN_BUFFER(MAX_MATCH_BITS))
1025 lzx_flush_bits(os, 17);
1028 if (CAN_BUFFER(MAX_MATCH_BITS))
1029 lzx_flush_bits(os, MAX_MATCH_BITS);
1031 /* Advance to the next sequence. */
1037 lzx_write_compressed_block(const u8 *block_begin,
1040 unsigned window_order,
1041 unsigned num_main_syms,
1042 const struct lzx_sequence sequences[],
1043 const struct lzx_codes * codes,
1044 const struct lzx_lens * prev_lens,
1045 struct lzx_output_bitstream * os)
1047 /* The first three bits indicate the type of block and are one of the
1048 * LZX_BLOCKTYPE_* constants. */
1049 lzx_write_bits(os, block_type, 3);
1051 /* Output the block size.
1053 * The original LZX format seemed to always encode the block size in 3
1054 * bytes. However, the implementation in WIMGAPI, as used in WIM files,
1055 * uses the first bit to indicate whether the block is the default size
1056 * (32768) or a different size given explicitly by the next 16 bits.
1058 * By default, this compressor uses a window size of 32768 and therefore
1059 * follows the WIMGAPI behavior. However, this compressor also supports
1060 * window sizes greater than 32768 bytes, which do not appear to be
1061 * supported by WIMGAPI. In such cases, we retain the default size bit
1062 * to mean a size of 32768 bytes but output non-default block size in 24
1063 * bits rather than 16. The compatibility of this behavior is unknown
1064 * because WIMs created with chunk size greater than 32768 can seemingly
1065 * only be opened by wimlib anyway. */
1066 if (block_size == LZX_DEFAULT_BLOCK_SIZE) {
1067 lzx_write_bits(os, 1, 1);
1069 lzx_write_bits(os, 0, 1);
1071 if (window_order >= 16)
1072 lzx_write_bits(os, block_size >> 16, 8);
1074 lzx_write_bits(os, block_size & 0xFFFF, 16);
1077 /* If it's an aligned offset block, output the aligned offset code. */
1078 if (block_type == LZX_BLOCKTYPE_ALIGNED) {
1079 for (int i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1080 lzx_write_bits(os, codes->lens.aligned[i],
1081 LZX_ALIGNEDCODE_ELEMENT_SIZE);
1085 /* Output the main code (two parts). */
1086 lzx_write_compressed_code(os, codes->lens.main,
1089 lzx_write_compressed_code(os, codes->lens.main + LZX_NUM_CHARS,
1090 prev_lens->main + LZX_NUM_CHARS,
1091 num_main_syms - LZX_NUM_CHARS);
1093 /* Output the length code. */
1094 lzx_write_compressed_code(os, codes->lens.len,
1096 LZX_LENCODE_NUM_SYMBOLS);
1098 /* Output the compressed matches and literals. */
1099 lzx_write_sequences(os, block_type, block_begin, sequences, codes);
1102 /* Given the frequencies of symbols in an LZX-compressed block and the
1103 * corresponding Huffman codes, return LZX_BLOCKTYPE_ALIGNED or
1104 * LZX_BLOCKTYPE_VERBATIM if an aligned offset or verbatim block, respectively,
1105 * will take fewer bits to output. */
1107 lzx_choose_verbatim_or_aligned(const struct lzx_freqs * freqs,
1108 const struct lzx_codes * codes)
1110 u32 aligned_cost = 0;
1111 u32 verbatim_cost = 0;
1113 /* A verbatim block requires 3 bits in each place that an aligned symbol
1114 * would be used in an aligned offset block. */
1115 for (unsigned i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1116 verbatim_cost += LZX_NUM_ALIGNED_OFFSET_BITS * freqs->aligned[i];
1117 aligned_cost += codes->lens.aligned[i] * freqs->aligned[i];
1120 /* Account for output of the aligned offset code. */
1121 aligned_cost += LZX_ALIGNEDCODE_ELEMENT_SIZE * LZX_ALIGNEDCODE_NUM_SYMBOLS;
1123 if (aligned_cost < verbatim_cost)
1124 return LZX_BLOCKTYPE_ALIGNED;
1126 return LZX_BLOCKTYPE_VERBATIM;
1130 * Return the offset slot for the specified adjusted match offset, using the
1131 * compressor's acceleration tables to speed up the mapping.
1133 static inline unsigned
1134 lzx_comp_get_offset_slot(struct lzx_compressor *c, u32 adjusted_offset,
1137 if (is_16_bit || adjusted_offset < ARRAY_LEN(c->offset_slot_tab_1))
1138 return c->offset_slot_tab_1[adjusted_offset];
1139 return c->offset_slot_tab_2[adjusted_offset >> 14];
1143 * Finish an LZX block:
1145 * - build the Huffman codes
1146 * - decide whether to output the block as VERBATIM or ALIGNED
1147 * - output the block
1148 * - swap the indices of the current and previous Huffman codes
1151 lzx_finish_block(struct lzx_compressor *c, struct lzx_output_bitstream *os,
1152 const u8 *block_begin, u32 block_size, u32 seq_idx)
1156 lzx_make_huffman_codes(c);
1158 block_type = lzx_choose_verbatim_or_aligned(&c->freqs,
1159 &c->codes[c->codes_index]);
1160 lzx_write_compressed_block(block_begin,
1165 &c->chosen_sequences[seq_idx],
1166 &c->codes[c->codes_index],
1167 &c->codes[c->codes_index ^ 1].lens,
1169 c->codes_index ^= 1;
1172 /* Tally the Huffman symbol for a literal and increment the literal run length.
1175 lzx_record_literal(struct lzx_compressor *c, unsigned literal, u32 *litrunlen_p)
1177 c->freqs.main[literal]++;
1181 /* Tally the Huffman symbol for a match, save the match data and the length of
1182 * the preceding literal run in the next lzx_sequence, and update the recent
1185 lzx_record_match(struct lzx_compressor *c, unsigned length, u32 offset_data,
1186 u32 recent_offsets[LZX_NUM_RECENT_OFFSETS], bool is_16_bit,
1187 u32 *litrunlen_p, struct lzx_sequence **next_seq_p)
1189 u32 litrunlen = *litrunlen_p;
1190 struct lzx_sequence *next_seq = *next_seq_p;
1191 unsigned offset_slot;
1194 v = length - LZX_MIN_MATCH_LEN;
1196 /* Save the literal run length and adjusted length. */
1197 next_seq->litrunlen = litrunlen;
1198 next_seq->adjusted_length = v;
1200 /* Compute the length header and tally the length symbol if needed */
1201 if (v >= LZX_NUM_PRIMARY_LENS) {
1202 c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++;
1203 v = LZX_NUM_PRIMARY_LENS;
1206 /* Compute the offset slot */
1207 offset_slot = lzx_comp_get_offset_slot(c, offset_data, is_16_bit);
1209 /* Compute the match header. */
1210 v += offset_slot * LZX_NUM_LEN_HEADERS;
1212 /* Save the adjusted offset and match header. */
1213 next_seq->adjusted_offset_and_match_hdr = (offset_data << 9) | v;
1215 /* Tally the main symbol. */
1216 c->freqs.main[LZX_NUM_CHARS + v]++;
1218 /* Update the recent offsets queue. */
1219 if (offset_data < LZX_NUM_RECENT_OFFSETS) {
1220 /* Repeat offset match */
1221 swap(recent_offsets[0], recent_offsets[offset_data]);
1223 /* Explicit offset match */
1225 /* Tally the aligned offset symbol if needed */
1226 if (offset_data >= 16)
1227 c->freqs.aligned[offset_data & LZX_ALIGNED_OFFSET_BITMASK]++;
1229 recent_offsets[2] = recent_offsets[1];
1230 recent_offsets[1] = recent_offsets[0];
1231 recent_offsets[0] = offset_data - LZX_OFFSET_ADJUSTMENT;
1234 /* Reset the literal run length and advance to the next sequence. */
1235 *next_seq_p = next_seq + 1;
1239 /* Finish the last lzx_sequence. The last lzx_sequence is just a literal run;
1240 * there is no match. This literal run may be empty. */
1242 lzx_finish_sequence(struct lzx_sequence *last_seq, u32 litrunlen)
1244 last_seq->litrunlen = litrunlen;
1246 /* Special value to mark last sequence */
1247 last_seq->adjusted_offset_and_match_hdr = 0x80000000;
1250 /******************************************************************************/
1253 * Block splitting algorithm. The problem is to decide when it is worthwhile to
1254 * start a new block with new entropy codes. There is a theoretically optimal
1255 * solution: recursively consider every possible block split, considering the
1256 * exact cost of each block, and choose the minimum cost approach. But this is
1257 * far too slow. Instead, as an approximation, we can count symbols and after
1258 * every N symbols, compare the expected distribution of symbols based on the
1259 * previous data with the actual distribution. If they differ "by enough", then
1260 * start a new block.
1262 * As an optimization and heuristic, we don't distinguish between every symbol
1263 * but rather we combine many symbols into a single "observation type". For
1264 * literals we only look at the high bits and low bits, and for matches we only
1265 * look at whether the match is long or not. The assumption is that for typical
1266 * "real" data, places that are good block boundaries will tend to be noticable
1267 * based only on changes in these aggregate frequencies, without looking for
1268 * subtle differences in individual symbols. For example, a change from ASCII
1269 * bytes to non-ASCII bytes, or from few matches (generally less compressible)
1270 * to many matches (generally more compressible), would be easily noticed based
1271 * on the aggregates.
1273 * For determining whether the frequency distributions are "different enough" to
1274 * start a new block, the simply heuristic of splitting when the sum of absolute
1275 * differences exceeds a constant seems to be good enough. We also add a number
1276 * proportional to the block size so that the algorithm is more likely to end
1277 * large blocks than small blocks. This reflects the general expectation that
1278 * it will become increasingly beneficial to start a new block as the current
1279 * blocks grows larger.
1281 * Finally, for an approximation, it is not strictly necessary that the exact
1282 * symbols being used are considered. With "near-optimal parsing", for example,
1283 * the actual symbols that will be used are unknown until after the block
1284 * boundary is chosen and the block has been optimized. Since the final choices
1285 * cannot be used, we can use preliminary "greedy" choices instead.
1288 /* Initialize the block split statistics when starting a new block. */
1290 init_block_split_stats(struct block_split_stats *stats)
1292 for (int i = 0; i < NUM_OBSERVATION_TYPES; i++) {
1293 stats->new_observations[i] = 0;
1294 stats->observations[i] = 0;
1296 stats->num_new_observations = 0;
1297 stats->num_observations = 0;
1300 /* Literal observation. Heuristic: use the top 2 bits and low 1 bits of the
1301 * literal, for 8 possible literal observation types. */
1303 observe_literal(struct block_split_stats *stats, u8 lit)
1305 stats->new_observations[((lit >> 5) & 0x6) | (lit & 1)]++;
1306 stats->num_new_observations++;
1309 /* Match observation. Heuristic: use one observation type for "short match" and
1310 * one observation type for "long match". */
1312 observe_match(struct block_split_stats *stats, unsigned length)
1314 stats->new_observations[NUM_LITERAL_OBSERVATION_TYPES + (length >= 9)]++;
1315 stats->num_new_observations++;
1319 do_end_block_check(struct block_split_stats *stats, u32 block_size)
1321 if (stats->num_observations > 0) {
1323 /* Note: to avoid slow divisions, we do not divide by
1324 * 'num_observations', but rather do all math with the numbers
1325 * multiplied by 'num_observations'. */
1326 u32 total_delta = 0;
1327 for (int i = 0; i < NUM_OBSERVATION_TYPES; i++) {
1328 u32 expected = stats->observations[i] * stats->num_new_observations;
1329 u32 actual = stats->new_observations[i] * stats->num_observations;
1330 u32 delta = (actual > expected) ? actual - expected :
1332 total_delta += delta;
1335 /* Ready to end the block? */
1336 if (total_delta + (block_size >> 12) * stats->num_observations >=
1337 200 * stats->num_observations)
1341 for (int i = 0; i < NUM_OBSERVATION_TYPES; i++) {
1342 stats->num_observations += stats->new_observations[i];
1343 stats->observations[i] += stats->new_observations[i];
1344 stats->new_observations[i] = 0;
1346 stats->num_new_observations = 0;
1351 should_end_block(struct block_split_stats *stats,
1352 const u8 *in_block_begin, const u8 *in_next, const u8 *in_end)
1354 /* Ready to check block split statistics? */
1355 if (stats->num_new_observations < 250 ||
1356 in_next - in_block_begin < MIN_BLOCK_SIZE ||
1357 in_end - in_next < MIN_BLOCK_SIZE)
1360 return do_end_block_check(stats, in_next - in_block_begin);
1363 /******************************************************************************/
1366 * Given the minimum-cost path computed through the item graph for the current
1367 * block, walk the path and count how many of each symbol in each Huffman-coded
1368 * alphabet would be required to output the items (matches and literals) along
1371 * Note that the path will be walked backwards (from the end of the block to the
1372 * beginning of the block), but this doesn't matter because this function only
1373 * computes frequencies.
1376 lzx_tally_item_list(struct lzx_compressor *c, u32 block_size, bool is_16_bit)
1378 u32 node_idx = block_size;
1383 unsigned offset_slot;
1385 /* Tally literals until either a match or the beginning of the
1386 * block is reached. */
1388 u32 item = c->optimum_nodes[node_idx].item;
1390 len = item & OPTIMUM_LEN_MASK;
1391 offset_data = item >> OPTIMUM_OFFSET_SHIFT;
1393 if (len != 0) /* Not a literal? */
1396 /* Tally the main symbol for the literal. */
1397 c->freqs.main[offset_data]++;
1399 if (--node_idx == 0) /* Beginning of block was reached? */
1405 /* Tally a match. */
1407 /* Tally the aligned offset symbol if needed. */
1408 if (offset_data >= 16)
1409 c->freqs.aligned[offset_data & LZX_ALIGNED_OFFSET_BITMASK]++;
1411 /* Tally the length symbol if needed. */
1412 v = len - LZX_MIN_MATCH_LEN;;
1413 if (v >= LZX_NUM_PRIMARY_LENS) {
1414 c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++;
1415 v = LZX_NUM_PRIMARY_LENS;
1418 /* Tally the main symbol. */
1419 offset_slot = lzx_comp_get_offset_slot(c, offset_data, is_16_bit);
1420 v += offset_slot * LZX_NUM_LEN_HEADERS;
1421 c->freqs.main[LZX_NUM_CHARS + v]++;
1423 if (node_idx == 0) /* Beginning of block was reached? */
1429 * Like lzx_tally_item_list(), but this function also generates the list of
1430 * lzx_sequences for the minimum-cost path and writes it to c->chosen_sequences,
1431 * ready to be output to the bitstream after the Huffman codes are computed.
1432 * The lzx_sequences will be written to decreasing memory addresses as the path
1433 * is walked backwards, which means they will end up in the expected
1434 * first-to-last order. The return value is the index in c->chosen_sequences at
1435 * which the lzx_sequences begin.
1438 lzx_record_item_list(struct lzx_compressor *c, u32 block_size, bool is_16_bit)
1440 u32 node_idx = block_size;
1441 u32 seq_idx = ARRAY_LEN(c->chosen_sequences) - 1;
1444 /* Special value to mark last sequence */
1445 c->chosen_sequences[seq_idx].adjusted_offset_and_match_hdr = 0x80000000;
1447 lit_start_node = node_idx;
1452 unsigned offset_slot;
1454 /* Record literals until either a match or the beginning of the
1455 * block is reached. */
1457 u32 item = c->optimum_nodes[node_idx].item;
1459 len = item & OPTIMUM_LEN_MASK;
1460 offset_data = item >> OPTIMUM_OFFSET_SHIFT;
1462 if (len != 0) /* Not a literal? */
1465 /* Tally the main symbol for the literal. */
1466 c->freqs.main[offset_data]++;
1468 if (--node_idx == 0) /* Beginning of block was reached? */
1472 /* Save the literal run length for the next sequence (the
1473 * "previous sequence" when walking backwards). */
1474 c->chosen_sequences[seq_idx--].litrunlen = lit_start_node - node_idx;
1476 lit_start_node = node_idx;
1478 /* Record a match. */
1480 /* Tally the aligned offset symbol if needed. */
1481 if (offset_data >= 16)
1482 c->freqs.aligned[offset_data & LZX_ALIGNED_OFFSET_BITMASK]++;
1484 /* Save the adjusted length. */
1485 v = len - LZX_MIN_MATCH_LEN;
1486 c->chosen_sequences[seq_idx].adjusted_length = v;
1488 /* Tally the length symbol if needed. */
1489 if (v >= LZX_NUM_PRIMARY_LENS) {
1490 c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++;
1491 v = LZX_NUM_PRIMARY_LENS;
1494 /* Tally the main symbol. */
1495 offset_slot = lzx_comp_get_offset_slot(c, offset_data, is_16_bit);
1496 v += offset_slot * LZX_NUM_LEN_HEADERS;
1497 c->freqs.main[LZX_NUM_CHARS + v]++;
1499 /* Save the adjusted offset and match header. */
1500 c->chosen_sequences[seq_idx].adjusted_offset_and_match_hdr =
1501 (offset_data << 9) | v;
1503 if (node_idx == 0) /* Beginning of block was reached? */
1508 /* Save the literal run length for the first sequence. */
1509 c->chosen_sequences[seq_idx].litrunlen = lit_start_node - node_idx;
1511 /* Return the index in c->chosen_sequences at which the lzx_sequences
1517 * Find an inexpensive path through the graph of possible match/literal choices
1518 * for the current block. The nodes of the graph are
1519 * c->optimum_nodes[0...block_size]. They correspond directly to the bytes in
1520 * the current block, plus one extra node for end-of-block. The edges of the
1521 * graph are matches and literals. The goal is to find the minimum cost path
1522 * from 'c->optimum_nodes[0]' to 'c->optimum_nodes[block_size]'.
1524 * The algorithm works forwards, starting at 'c->optimum_nodes[0]' and
1525 * proceeding forwards one node at a time. At each node, a selection of matches
1526 * (len >= 2), as well as the literal byte (len = 1), is considered. An item of
1527 * length 'len' provides a new path to reach the node 'len' bytes later. If
1528 * such a path is the lowest cost found so far to reach that later node, then
1529 * that later node is updated with the new path.
1531 * Note that although this algorithm is based on minimum cost path search, due
1532 * to various simplifying assumptions the result is not guaranteed to be the
1533 * true minimum cost, or "optimal", path over the graph of all valid LZX
1534 * representations of this block.
1536 * Also, note that because of the presence of the recent offsets queue (which is
1537 * a type of adaptive state), the algorithm cannot work backwards and compute
1538 * "cost to end" instead of "cost to beginning". Furthermore, the way the
1539 * algorithm handles this adaptive state in the "minimum cost" parse is actually
1540 * only an approximation. It's possible for the globally optimal, minimum cost
1541 * path to contain a prefix, ending at a position, where that path prefix is
1542 * *not* the minimum cost path to that position. This can happen if such a path
1543 * prefix results in a different adaptive state which results in lower costs
1544 * later. The algorithm does not solve this problem; it only considers the
1545 * lowest cost to reach each individual position.
1547 static inline struct lzx_lru_queue
1548 lzx_find_min_cost_path(struct lzx_compressor * const restrict c,
1549 const u8 * const restrict block_begin,
1550 const u32 block_size,
1551 const struct lzx_lru_queue initial_queue,
1554 struct lzx_optimum_node *cur_node = c->optimum_nodes;
1555 struct lzx_optimum_node * const end_node = &c->optimum_nodes[block_size];
1556 struct lz_match *cache_ptr = c->match_cache;
1557 const u8 *in_next = block_begin;
1558 const u8 * const block_end = block_begin + block_size;
1560 /* Instead of storing the match offset LRU queues in the
1561 * 'lzx_optimum_node' structures, we save memory (and cache lines) by
1562 * storing them in a smaller array. This works because the algorithm
1563 * only requires a limited history of the adaptive state. Once a given
1564 * state is more than LZX_MAX_MATCH_LEN bytes behind the current node,
1565 * it is no longer needed. */
1566 struct lzx_lru_queue queues[512];
1568 STATIC_ASSERT(ARRAY_LEN(queues) >= LZX_MAX_MATCH_LEN + 1);
1569 #define QUEUE(in) (queues[(uintptr_t)(in) % ARRAY_LEN(queues)])
1571 /* Initially, the cost to reach each node is "infinity". */
1572 memset(c->optimum_nodes, 0xFF,
1573 (block_size + 1) * sizeof(c->optimum_nodes[0]));
1575 QUEUE(block_begin) = initial_queue;
1577 /* The following loop runs 'block_size' iterations, one per node. */
1579 unsigned num_matches;
1584 * A selection of matches for the block was already saved in
1585 * memory so that we don't have to run the uncompressed data
1586 * through the matchfinder on every optimization pass. However,
1587 * we still search for repeat offset matches during each
1588 * optimization pass because we cannot predict the state of the
1589 * recent offsets queue. But as a heuristic, we don't bother
1590 * searching for repeat offset matches if the general-purpose
1591 * matchfinder failed to find any matches.
1593 * Note that a match of length n at some offset implies there is
1594 * also a match of length l for LZX_MIN_MATCH_LEN <= l <= n at
1595 * that same offset. In other words, we don't necessarily need
1596 * to use the full length of a match. The key heuristic that
1597 * saves a significicant amount of time is that for each
1598 * distinct length, we only consider the smallest offset for
1599 * which that length is available. This heuristic also applies
1600 * to repeat offsets, which we order specially: R0 < R1 < R2 <
1601 * any explicit offset. Of course, this heuristic may be
1602 * produce suboptimal results because offset slots in LZX are
1603 * subject to entropy encoding, but in practice this is a useful
1607 num_matches = cache_ptr->length;
1611 struct lz_match *end_matches = cache_ptr + num_matches;
1612 unsigned next_len = LZX_MIN_MATCH_LEN;
1613 unsigned max_len = min(block_end - in_next, LZX_MAX_MATCH_LEN);
1616 /* Consider R0 match */
1617 matchptr = in_next - lzx_lru_queue_R0(QUEUE(in_next));
1618 if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1620 STATIC_ASSERT(LZX_MIN_MATCH_LEN == 2);
1622 u32 cost = cur_node->cost +
1623 c->costs.match_cost[0][
1624 next_len - LZX_MIN_MATCH_LEN];
1625 if (cost <= (cur_node + next_len)->cost) {
1626 (cur_node + next_len)->cost = cost;
1627 (cur_node + next_len)->item =
1628 (0 << OPTIMUM_OFFSET_SHIFT) | next_len;
1630 if (unlikely(++next_len > max_len)) {
1631 cache_ptr = end_matches;
1634 } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1638 /* Consider R1 match */
1639 matchptr = in_next - lzx_lru_queue_R1(QUEUE(in_next));
1640 if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1642 if (matchptr[next_len - 1] != in_next[next_len - 1])
1644 for (unsigned len = 2; len < next_len - 1; len++)
1645 if (matchptr[len] != in_next[len])
1648 u32 cost = cur_node->cost +
1649 c->costs.match_cost[1][
1650 next_len - LZX_MIN_MATCH_LEN];
1651 if (cost <= (cur_node + next_len)->cost) {
1652 (cur_node + next_len)->cost = cost;
1653 (cur_node + next_len)->item =
1654 (1 << OPTIMUM_OFFSET_SHIFT) | next_len;
1656 if (unlikely(++next_len > max_len)) {
1657 cache_ptr = end_matches;
1660 } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1664 /* Consider R2 match */
1665 matchptr = in_next - lzx_lru_queue_R2(QUEUE(in_next));
1666 if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1668 if (matchptr[next_len - 1] != in_next[next_len - 1])
1670 for (unsigned len = 2; len < next_len - 1; len++)
1671 if (matchptr[len] != in_next[len])
1674 u32 cost = cur_node->cost +
1675 c->costs.match_cost[2][
1676 next_len - LZX_MIN_MATCH_LEN];
1677 if (cost <= (cur_node + next_len)->cost) {
1678 (cur_node + next_len)->cost = cost;
1679 (cur_node + next_len)->item =
1680 (2 << OPTIMUM_OFFSET_SHIFT) | next_len;
1682 if (unlikely(++next_len > max_len)) {
1683 cache_ptr = end_matches;
1686 } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1690 while (next_len > cache_ptr->length)
1691 if (++cache_ptr == end_matches)
1694 /* Consider explicit offset matches */
1696 u32 offset = cache_ptr->offset;
1697 u32 offset_data = offset + LZX_OFFSET_ADJUSTMENT;
1698 unsigned offset_slot = lzx_comp_get_offset_slot(c, offset_data,
1700 u32 base_cost = cur_node->cost;
1702 #if LZX_CONSIDER_ALIGNED_COSTS
1703 if (offset_data >= 16)
1704 base_cost += c->costs.aligned[offset_data &
1705 LZX_ALIGNED_OFFSET_BITMASK];
1709 u32 cost = base_cost +
1710 c->costs.match_cost[offset_slot][
1711 next_len - LZX_MIN_MATCH_LEN];
1712 if (cost < (cur_node + next_len)->cost) {
1713 (cur_node + next_len)->cost = cost;
1714 (cur_node + next_len)->item =
1715 (offset_data << OPTIMUM_OFFSET_SHIFT) | next_len;
1717 } while (++next_len <= cache_ptr->length);
1718 } while (++cache_ptr != end_matches);
1723 /* Consider coding a literal.
1725 * To avoid an extra branch, actually checking the preferability
1726 * of coding the literal is integrated into the queue update
1728 literal = *in_next++;
1729 cost = cur_node->cost + c->costs.main[literal];
1731 /* Advance to the next position. */
1734 /* The lowest-cost path to the current position is now known.
1735 * Finalize the recent offsets queue that results from taking
1736 * this lowest-cost path. */
1738 if (cost <= cur_node->cost) {
1739 /* Literal: queue remains unchanged. */
1740 cur_node->cost = cost;
1741 cur_node->item = (u32)literal << OPTIMUM_OFFSET_SHIFT;
1742 QUEUE(in_next) = QUEUE(in_next - 1);
1744 /* Match: queue update is needed. */
1745 unsigned len = cur_node->item & OPTIMUM_LEN_MASK;
1746 u32 offset_data = cur_node->item >> OPTIMUM_OFFSET_SHIFT;
1747 if (offset_data >= LZX_NUM_RECENT_OFFSETS) {
1748 /* Explicit offset match: insert offset at front */
1750 lzx_lru_queue_push(QUEUE(in_next - len),
1751 offset_data - LZX_OFFSET_ADJUSTMENT);
1753 /* Repeat offset match: swap offset to front */
1755 lzx_lru_queue_swap(QUEUE(in_next - len),
1759 } while (cur_node != end_node);
1761 /* Return the match offset queue at the end of the minimum cost path. */
1762 return QUEUE(block_end);
1765 /* Given the costs for the main and length codewords, compute 'match_costs'. */
1767 lzx_compute_match_costs(struct lzx_compressor *c)
1769 unsigned num_offset_slots = (c->num_main_syms - LZX_NUM_CHARS) /
1770 LZX_NUM_LEN_HEADERS;
1771 struct lzx_costs *costs = &c->costs;
1773 for (unsigned offset_slot = 0; offset_slot < num_offset_slots; offset_slot++) {
1775 u32 extra_cost = (u32)lzx_extra_offset_bits[offset_slot] * LZX_BIT_COST;
1776 unsigned main_symbol = LZX_NUM_CHARS + (offset_slot *
1777 LZX_NUM_LEN_HEADERS);
1780 #if LZX_CONSIDER_ALIGNED_COSTS
1781 if (offset_slot >= 8)
1782 extra_cost -= LZX_NUM_ALIGNED_OFFSET_BITS * LZX_BIT_COST;
1785 for (i = 0; i < LZX_NUM_PRIMARY_LENS; i++)
1786 costs->match_cost[offset_slot][i] =
1787 costs->main[main_symbol++] + extra_cost;
1789 extra_cost += costs->main[main_symbol];
1791 for (; i < LZX_NUM_LENS; i++)
1792 costs->match_cost[offset_slot][i] =
1793 costs->len[i - LZX_NUM_PRIMARY_LENS] + extra_cost;
1797 /* Set default LZX Huffman symbol costs to bootstrap the iterative optimization
1800 lzx_set_default_costs(struct lzx_compressor *c, const u8 *block, u32 block_size)
1803 bool have_byte[256];
1804 unsigned num_used_bytes;
1806 /* The costs below are hard coded to use a scaling factor of 16. */
1807 STATIC_ASSERT(LZX_BIT_COST == 16);
1812 * - Use smaller initial costs for literal symbols when the input buffer
1813 * contains fewer distinct bytes.
1815 * - Assume that match symbols are more costly than literal symbols.
1817 * - Assume that length symbols for shorter lengths are less costly than
1818 * length symbols for longer lengths.
1821 for (i = 0; i < 256; i++)
1822 have_byte[i] = false;
1824 for (i = 0; i < block_size; i++)
1825 have_byte[block[i]] = true;
1828 for (i = 0; i < 256; i++)
1829 num_used_bytes += have_byte[i];
1831 for (i = 0; i < 256; i++)
1832 c->costs.main[i] = 140 - (256 - num_used_bytes) / 4;
1834 for (; i < c->num_main_syms; i++)
1835 c->costs.main[i] = 170;
1837 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
1838 c->costs.len[i] = 103 + (i / 4);
1840 #if LZX_CONSIDER_ALIGNED_COSTS
1841 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++)
1842 c->costs.aligned[i] = LZX_NUM_ALIGNED_OFFSET_BITS * LZX_BIT_COST;
1845 lzx_compute_match_costs(c);
1848 /* Update the current cost model to reflect the computed Huffman codes. */
1850 lzx_update_costs(struct lzx_compressor *c)
1853 const struct lzx_lens *lens = &c->codes[c->codes_index].lens;
1855 for (i = 0; i < c->num_main_syms; i++) {
1856 c->costs.main[i] = (lens->main[i] ? lens->main[i] :
1857 MAIN_CODEWORD_LIMIT) * LZX_BIT_COST;
1860 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) {
1861 c->costs.len[i] = (lens->len[i] ? lens->len[i] :
1862 LENGTH_CODEWORD_LIMIT) * LZX_BIT_COST;
1865 #if LZX_CONSIDER_ALIGNED_COSTS
1866 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1867 c->costs.aligned[i] = (lens->aligned[i] ? lens->aligned[i] :
1868 ALIGNED_CODEWORD_LIMIT) * LZX_BIT_COST;
1872 lzx_compute_match_costs(c);
1875 static inline struct lzx_lru_queue
1876 lzx_optimize_and_write_block(struct lzx_compressor * const restrict c,
1877 struct lzx_output_bitstream * const restrict os,
1878 const u8 * const restrict block_begin,
1879 const u32 block_size,
1880 const struct lzx_lru_queue initial_queue,
1883 unsigned num_passes_remaining = c->num_optim_passes;
1884 struct lzx_lru_queue new_queue;
1887 /* The first optimization pass uses a default cost model. Each
1888 * additional optimization pass uses a cost model derived from the
1889 * Huffman code computed in the previous pass. */
1891 lzx_set_default_costs(c, block_begin, block_size);
1892 lzx_reset_symbol_frequencies(c);
1894 new_queue = lzx_find_min_cost_path(c, block_begin, block_size,
1895 initial_queue, is_16_bit);
1896 if (num_passes_remaining > 1) {
1897 lzx_tally_item_list(c, block_size, is_16_bit);
1898 lzx_make_huffman_codes(c);
1899 lzx_update_costs(c);
1900 lzx_reset_symbol_frequencies(c);
1902 } while (--num_passes_remaining);
1904 seq_idx = lzx_record_item_list(c, block_size, is_16_bit);
1905 lzx_finish_block(c, os, block_begin, block_size, seq_idx);
1910 * This is the "near-optimal" LZX compressor.
1912 * For each block, it performs a relatively thorough graph search to find an
1913 * inexpensive (in terms of compressed size) way to output that block.
1915 * Note: there are actually many things this algorithm leaves on the table in
1916 * terms of compression ratio. So although it may be "near-optimal", it is
1917 * certainly not "optimal". The goal is not to produce the optimal compression
1918 * ratio, which for LZX is probably impossible within any practical amount of
1919 * time, but rather to produce a compression ratio significantly better than a
1920 * simpler "greedy" or "lazy" parse while still being relatively fast.
1923 lzx_compress_near_optimal(struct lzx_compressor *c,
1924 struct lzx_output_bitstream *os,
1927 const u8 * const in_begin = c->in_buffer;
1928 const u8 * in_next = in_begin;
1929 const u8 * const in_end = in_begin + c->in_nbytes;
1930 u32 max_len = LZX_MAX_MATCH_LEN;
1931 u32 nice_len = min(c->nice_match_length, max_len);
1932 u32 next_hashes[2] = {};
1933 struct lzx_lru_queue queue;
1935 CALL_BT_MF(is_16_bit, c, bt_matchfinder_init);
1936 lzx_lru_queue_init(&queue);
1939 /* Starting a new block */
1940 const u8 * const in_block_begin = in_next;
1941 const u8 * const in_max_block_end =
1942 in_next + min(SOFT_MAX_BLOCK_SIZE, in_end - in_next);
1943 const u8 *next_observation = in_next;
1945 init_block_split_stats(&c->split_stats);
1947 /* Run the block through the matchfinder and cache the matches. */
1948 struct lz_match *cache_ptr = c->match_cache;
1950 struct lz_match *lz_matchptr;
1953 /* If approaching the end of the input buffer, adjust
1954 * 'max_len' and 'nice_len' accordingly. */
1955 if (unlikely(max_len > in_end - in_next)) {
1956 max_len = in_end - in_next;
1957 nice_len = min(max_len, nice_len);
1958 if (unlikely(max_len <
1959 BT_MATCHFINDER_REQUIRED_NBYTES))
1962 cache_ptr->length = 0;
1968 /* Check for matches. */
1969 lz_matchptr = CALL_BT_MF(is_16_bit, c,
1970 bt_matchfinder_get_matches,
1975 c->max_search_depth,
1980 if (in_next >= next_observation) {
1982 if (lz_matchptr > cache_ptr + 1)
1983 best_len = (lz_matchptr - 1)->length;
1984 if (best_len >= 2) {
1985 observe_match(&c->split_stats, best_len);
1986 next_observation = in_next + best_len;
1988 observe_literal(&c->split_stats, *in_next);
1989 next_observation = in_next + 1;
1994 cache_ptr->length = lz_matchptr - (cache_ptr + 1);
1995 cache_ptr = lz_matchptr;
1998 * If there was a very long match found, then don't
1999 * cache any matches for the bytes covered by that
2000 * match. This avoids degenerate behavior when
2001 * compressing highly redundant data, where the number
2002 * of matches can be very large.
2004 * This heuristic doesn't actually hurt the compression
2005 * ratio very much. If there's a long match, then the
2006 * data must be highly compressible, so it doesn't
2007 * matter as much what we do.
2009 if (best_len >= nice_len) {
2012 if (unlikely(max_len > in_end - in_next)) {
2013 max_len = in_end - in_next;
2014 nice_len = min(max_len, nice_len);
2015 if (unlikely(max_len <
2016 BT_MATCHFINDER_REQUIRED_NBYTES))
2019 cache_ptr->length = 0;
2024 CALL_BT_MF(is_16_bit, c,
2025 bt_matchfinder_skip_position,
2030 c->max_search_depth,
2033 cache_ptr->length = 0;
2035 } while (--best_len);
2037 } while (in_next < in_max_block_end &&
2038 likely(cache_ptr < &c->match_cache[LZX_CACHE_LENGTH]) &&
2039 !should_end_block(&c->split_stats, in_block_begin, in_next, in_end));
2041 /* We've finished running the block through the matchfinder.
2042 * Now choose a match/literal sequence and write the block. */
2044 queue = lzx_optimize_and_write_block(c, os, in_block_begin,
2045 in_next - in_block_begin,
2047 } while (in_next != in_end);
2051 lzx_compress_near_optimal_16(struct lzx_compressor *c,
2052 struct lzx_output_bitstream *os)
2054 lzx_compress_near_optimal(c, os, true);
2058 lzx_compress_near_optimal_32(struct lzx_compressor *c,
2059 struct lzx_output_bitstream *os)
2061 lzx_compress_near_optimal(c, os, false);
2065 * Given a pointer to the current byte sequence and the current list of recent
2066 * match offsets, find the longest repeat offset match.
2068 * If no match of at least 2 bytes is found, then return 0.
2070 * If a match of at least 2 bytes is found, then return its length and set
2071 * *rep_max_idx_ret to the index of its offset in @queue.
2074 lzx_find_longest_repeat_offset_match(const u8 * const in_next,
2075 const u32 bytes_remaining,
2076 const u32 recent_offsets[LZX_NUM_RECENT_OFFSETS],
2077 unsigned *rep_max_idx_ret)
2079 STATIC_ASSERT(LZX_NUM_RECENT_OFFSETS == 3);
2081 const unsigned max_len = min(bytes_remaining, LZX_MAX_MATCH_LEN);
2082 const u16 next_2_bytes = load_u16_unaligned(in_next);
2084 unsigned rep_max_len;
2085 unsigned rep_max_idx;
2088 matchptr = in_next - recent_offsets[0];
2089 if (load_u16_unaligned(matchptr) == next_2_bytes)
2090 rep_max_len = lz_extend(in_next, matchptr, 2, max_len);
2095 matchptr = in_next - recent_offsets[1];
2096 if (load_u16_unaligned(matchptr) == next_2_bytes) {
2097 rep_len = lz_extend(in_next, matchptr, 2, max_len);
2098 if (rep_len > rep_max_len) {
2099 rep_max_len = rep_len;
2104 matchptr = in_next - recent_offsets[2];
2105 if (load_u16_unaligned(matchptr) == next_2_bytes) {
2106 rep_len = lz_extend(in_next, matchptr, 2, max_len);
2107 if (rep_len > rep_max_len) {
2108 rep_max_len = rep_len;
2113 *rep_max_idx_ret = rep_max_idx;
2117 /* Fast heuristic scoring for lazy parsing: how "good" is this match? */
2118 static inline unsigned
2119 lzx_explicit_offset_match_score(unsigned len, u32 adjusted_offset)
2121 unsigned score = len;
2123 if (adjusted_offset < 4096)
2126 if (adjusted_offset < 256)
2132 static inline unsigned
2133 lzx_repeat_offset_match_score(unsigned rep_len, unsigned rep_idx)
2138 /* This is the "lazy" LZX compressor. */
2140 lzx_compress_lazy(struct lzx_compressor *c, struct lzx_output_bitstream *os,
2143 const u8 * const in_begin = c->in_buffer;
2144 const u8 * in_next = in_begin;
2145 const u8 * const in_end = in_begin + c->in_nbytes;
2146 unsigned max_len = LZX_MAX_MATCH_LEN;
2147 unsigned nice_len = min(c->nice_match_length, max_len);
2148 STATIC_ASSERT(LZX_NUM_RECENT_OFFSETS == 3);
2149 u32 recent_offsets[3] = {1, 1, 1};
2150 u32 next_hashes[2] = {};
2152 CALL_HC_MF(is_16_bit, c, hc_matchfinder_init);
2155 /* Starting a new block */
2157 const u8 * const in_block_begin = in_next;
2158 const u8 * const in_max_block_end =
2159 in_next + min(SOFT_MAX_BLOCK_SIZE, in_end - in_next);
2160 struct lzx_sequence *next_seq = c->chosen_sequences;
2163 u32 cur_offset_data;
2167 u32 next_offset_data;
2168 unsigned next_score;
2169 unsigned rep_max_len;
2170 unsigned rep_max_idx;
2175 lzx_reset_symbol_frequencies(c);
2176 init_block_split_stats(&c->split_stats);
2179 if (unlikely(max_len > in_end - in_next)) {
2180 max_len = in_end - in_next;
2181 nice_len = min(max_len, nice_len);
2184 /* Find the longest match at the current position. */
2186 cur_len = CALL_HC_MF(is_16_bit, c,
2187 hc_matchfinder_longest_match,
2193 c->max_search_depth,
2198 cur_offset >= 8192 - LZX_OFFSET_ADJUSTMENT &&
2199 cur_offset != recent_offsets[0] &&
2200 cur_offset != recent_offsets[1] &&
2201 cur_offset != recent_offsets[2]))
2203 /* There was no match found, or the only match found
2204 * was a distant length 3 match. Output a literal. */
2205 lzx_record_literal(c, *in_next, &litrunlen);
2206 observe_literal(&c->split_stats, *in_next);
2211 observe_match(&c->split_stats, cur_len);
2213 if (cur_offset == recent_offsets[0]) {
2215 cur_offset_data = 0;
2216 skip_len = cur_len - 1;
2217 goto choose_cur_match;
2220 cur_offset_data = cur_offset + LZX_OFFSET_ADJUSTMENT;
2221 cur_score = lzx_explicit_offset_match_score(cur_len, cur_offset_data);
2223 /* Consider a repeat offset match */
2224 rep_max_len = lzx_find_longest_repeat_offset_match(in_next,
2230 if (rep_max_len >= 3 &&
2231 (rep_score = lzx_repeat_offset_match_score(rep_max_len,
2232 rep_max_idx)) >= cur_score)
2234 cur_len = rep_max_len;
2235 cur_offset_data = rep_max_idx;
2236 skip_len = rep_max_len - 1;
2237 goto choose_cur_match;
2242 /* We have a match at the current position. */
2244 /* If we have a very long match, choose it immediately. */
2245 if (cur_len >= nice_len) {
2246 skip_len = cur_len - 1;
2247 goto choose_cur_match;
2250 /* See if there's a better match at the next position. */
2252 if (unlikely(max_len > in_end - in_next)) {
2253 max_len = in_end - in_next;
2254 nice_len = min(max_len, nice_len);
2257 next_len = CALL_HC_MF(is_16_bit, c,
2258 hc_matchfinder_longest_match,
2264 c->max_search_depth / 2,
2268 if (next_len <= cur_len - 2) {
2270 skip_len = cur_len - 2;
2271 goto choose_cur_match;
2274 next_offset_data = next_offset + LZX_OFFSET_ADJUSTMENT;
2275 next_score = lzx_explicit_offset_match_score(next_len, next_offset_data);
2277 rep_max_len = lzx_find_longest_repeat_offset_match(in_next,
2283 if (rep_max_len >= 3 &&
2284 (rep_score = lzx_repeat_offset_match_score(rep_max_len,
2285 rep_max_idx)) >= next_score)
2288 if (rep_score > cur_score) {
2289 /* The next match is better, and it's a
2290 * repeat offset match. */
2291 lzx_record_literal(c, *(in_next - 2),
2293 cur_len = rep_max_len;
2294 cur_offset_data = rep_max_idx;
2295 skip_len = cur_len - 1;
2296 goto choose_cur_match;
2299 if (next_score > cur_score) {
2300 /* The next match is better, and it's an
2301 * explicit offset match. */
2302 lzx_record_literal(c, *(in_next - 2),
2305 cur_offset_data = next_offset_data;
2306 cur_score = next_score;
2307 goto have_cur_match;
2311 /* The original match was better. */
2312 skip_len = cur_len - 2;
2315 lzx_record_match(c, cur_len, cur_offset_data,
2316 recent_offsets, is_16_bit,
2317 &litrunlen, &next_seq);
2318 in_next = CALL_HC_MF(is_16_bit, c,
2319 hc_matchfinder_skip_positions,
2325 } while (in_next < in_max_block_end &&
2326 !should_end_block(&c->split_stats, in_block_begin, in_next, in_end));
2328 lzx_finish_sequence(next_seq, litrunlen);
2330 lzx_finish_block(c, os, in_block_begin, in_next - in_block_begin, 0);
2332 } while (in_next != in_end);
2336 lzx_compress_lazy_16(struct lzx_compressor *c, struct lzx_output_bitstream *os)
2338 lzx_compress_lazy(c, os, true);
2342 lzx_compress_lazy_32(struct lzx_compressor *c, struct lzx_output_bitstream *os)
2344 lzx_compress_lazy(c, os, false);
2347 /* Generate the acceleration tables for offset slots. */
2349 lzx_init_offset_slot_tabs(struct lzx_compressor *c)
2351 u32 adjusted_offset = 0;
2355 for (; adjusted_offset < ARRAY_LEN(c->offset_slot_tab_1);
2358 if (adjusted_offset >= lzx_offset_slot_base[slot + 1])
2360 c->offset_slot_tab_1[adjusted_offset] = slot;
2363 /* slots [30, 49] */
2364 for (; adjusted_offset < LZX_MAX_WINDOW_SIZE;
2365 adjusted_offset += (u32)1 << 14)
2367 if (adjusted_offset >= lzx_offset_slot_base[slot + 1])
2369 c->offset_slot_tab_2[adjusted_offset >> 14] = slot;
2374 lzx_get_compressor_size(size_t max_bufsize, unsigned compression_level)
2376 if (compression_level <= LZX_MAX_FAST_LEVEL) {
2377 if (lzx_is_16_bit(max_bufsize))
2378 return offsetof(struct lzx_compressor, hc_mf_16) +
2379 hc_matchfinder_size_16(max_bufsize);
2381 return offsetof(struct lzx_compressor, hc_mf_32) +
2382 hc_matchfinder_size_32(max_bufsize);
2384 if (lzx_is_16_bit(max_bufsize))
2385 return offsetof(struct lzx_compressor, bt_mf_16) +
2386 bt_matchfinder_size_16(max_bufsize);
2388 return offsetof(struct lzx_compressor, bt_mf_32) +
2389 bt_matchfinder_size_32(max_bufsize);
2394 lzx_get_needed_memory(size_t max_bufsize, unsigned compression_level,
2399 if (max_bufsize > LZX_MAX_WINDOW_SIZE)
2402 size += lzx_get_compressor_size(max_bufsize, compression_level);
2404 size += max_bufsize; /* in_buffer */
2409 lzx_create_compressor(size_t max_bufsize, unsigned compression_level,
2410 bool destructive, void **c_ret)
2412 unsigned window_order;
2413 struct lzx_compressor *c;
2415 window_order = lzx_get_window_order(max_bufsize);
2416 if (window_order == 0)
2417 return WIMLIB_ERR_INVALID_PARAM;
2419 c = MALLOC(lzx_get_compressor_size(max_bufsize, compression_level));
2423 c->destructive = destructive;
2425 c->num_main_syms = lzx_get_num_main_syms(window_order);
2426 c->window_order = window_order;
2428 if (!c->destructive) {
2429 c->in_buffer = MALLOC(max_bufsize);
2434 if (compression_level <= LZX_MAX_FAST_LEVEL) {
2436 /* Fast compression: Use lazy parsing. */
2438 if (lzx_is_16_bit(max_bufsize))
2439 c->impl = lzx_compress_lazy_16;
2441 c->impl = lzx_compress_lazy_32;
2442 c->max_search_depth = (60 * compression_level) / 20;
2443 c->nice_match_length = (80 * compression_level) / 20;
2445 /* lzx_compress_lazy() needs max_search_depth >= 2 because it
2446 * halves the max_search_depth when attempting a lazy match, and
2447 * max_search_depth cannot be 0. */
2448 if (c->max_search_depth < 2)
2449 c->max_search_depth = 2;
2452 /* Normal / high compression: Use near-optimal parsing. */
2454 if (lzx_is_16_bit(max_bufsize))
2455 c->impl = lzx_compress_near_optimal_16;
2457 c->impl = lzx_compress_near_optimal_32;
2459 /* Scale nice_match_length and max_search_depth with the
2460 * compression level. */
2461 c->max_search_depth = (24 * compression_level) / 50;
2462 c->nice_match_length = (48 * compression_level) / 50;
2464 /* Set a number of optimization passes appropriate for the
2465 * compression level. */
2467 c->num_optim_passes = 1;
2469 if (compression_level >= 45)
2470 c->num_optim_passes++;
2472 /* Use more optimization passes for higher compression levels.
2473 * But the more passes there are, the less they help --- so
2474 * don't add them linearly. */
2475 if (compression_level >= 70) {
2476 c->num_optim_passes++;
2477 if (compression_level >= 100)
2478 c->num_optim_passes++;
2479 if (compression_level >= 150)
2480 c->num_optim_passes++;
2481 if (compression_level >= 200)
2482 c->num_optim_passes++;
2483 if (compression_level >= 300)
2484 c->num_optim_passes++;
2488 /* max_search_depth == 0 is invalid. */
2489 if (c->max_search_depth < 1)
2490 c->max_search_depth = 1;
2492 if (c->nice_match_length > LZX_MAX_MATCH_LEN)
2493 c->nice_match_length = LZX_MAX_MATCH_LEN;
2495 lzx_init_offset_slot_tabs(c);
2502 return WIMLIB_ERR_NOMEM;
2506 lzx_compress(const void *restrict in, size_t in_nbytes,
2507 void *restrict out, size_t out_nbytes_avail, void *restrict _c)
2509 struct lzx_compressor *c = _c;
2510 struct lzx_output_bitstream os;
2513 /* Don't bother trying to compress very small inputs. */
2514 if (in_nbytes < 100)
2517 /* Copy the input data into the internal buffer and preprocess it. */
2519 c->in_buffer = (void *)in;
2521 memcpy(c->in_buffer, in, in_nbytes);
2522 c->in_nbytes = in_nbytes;
2523 lzx_preprocess(c->in_buffer, in_nbytes);
2525 /* Initially, the previous Huffman codeword lengths are all zeroes. */
2527 memset(&c->codes[1].lens, 0, sizeof(struct lzx_lens));
2529 /* Initialize the output bitstream. */
2530 lzx_init_output(&os, out, out_nbytes_avail);
2532 /* Call the compression level-specific compress() function. */
2535 /* Flush the output bitstream and return the compressed size or 0. */
2536 result = lzx_flush_output(&os);
2537 if (!result && c->destructive)
2538 lzx_postprocess(c->in_buffer, c->in_nbytes);
2543 lzx_free_compressor(void *_c)
2545 struct lzx_compressor *c = _c;
2547 if (!c->destructive)
2552 const struct compressor_ops lzx_compressor_ops = {
2553 .get_needed_memory = lzx_get_needed_memory,
2554 .create_compressor = lzx_create_compressor,
2555 .compress = lzx_compress,
2556 .free_compressor = lzx_free_compressor,