4 * A compressor for the LZX compression format, as used in WIM archives.
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 LZX-compatible algorithms are implemented: "near-optimal" and
30 * "lazy". "Near-optimal" is significantly slower than "lazy", but results in a
31 * better compression ratio. The "near-optimal" algorithm is used at the
32 * default compression level.
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 * LZX is a compression format derived from DEFLATE, the format used by zlib and
39 * gzip. Both LZX and DEFLATE use LZ77 matching and Huffman coding. Certain
40 * details are quite similar, such as the method for storing Huffman codes.
41 * 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 compressor 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.
63 /******************************************************************************/
64 /* General parameters */
65 /*----------------------------------------------------------------------------*/
68 * The compressor uses the faster algorithm at levels <= MAX_FAST_LEVEL. It
69 * uses the slower algorithm at levels > MAX_FAST_LEVEL.
71 #define MAX_FAST_LEVEL 34
74 * The compressor-side limits on the codeword lengths (in bits) for each Huffman
75 * code. To make outputting bits slightly faster, some of these limits are
76 * lower than the limits defined by the LZX format. This does not significantly
77 * affect the compression ratio.
79 #define MAIN_CODEWORD_LIMIT 16
80 #define LENGTH_CODEWORD_LIMIT 12
81 #define ALIGNED_CODEWORD_LIMIT 7
82 #define PRE_CODEWORD_LIMIT 7
85 /******************************************************************************/
86 /* Block splitting parameters */
87 /*----------------------------------------------------------------------------*/
90 * The compressor always outputs blocks of at least this size in bytes, except
91 * for the last block which may need to be smaller.
93 #define MIN_BLOCK_SIZE 6500
96 * The compressor attempts to end a block when it reaches this size in bytes.
97 * The final size might be slightly larger due to matches extending beyond the
98 * end of the block. Specifically:
100 * - The near-optimal compressor may choose a match of up to LZX_MAX_MATCH_LEN
101 * bytes starting at position 'SOFT_MAX_BLOCK_SIZE - 1'.
103 * - The lazy compressor may choose a sequence of literals starting at position
104 * 'SOFT_MAX_BLOCK_SIZE - 1' when it sees a sequence of increasingly better
105 * matches. The final match may be up to LZX_MAX_MATCH_LEN bytes. The
106 * length of the literal sequence is approximately limited by the "nice match
109 #define SOFT_MAX_BLOCK_SIZE 100000
112 * The number of observed items (matches and literals) that represents
113 * sufficient data for the compressor to decide whether the current block should
116 #define NUM_OBSERVATIONS_PER_BLOCK_CHECK 400
119 /******************************************************************************/
120 /* Parameters for slower algorithm */
121 /*----------------------------------------------------------------------------*/
124 * The log base 2 of the number of entries in the hash table for finding length
125 * 2 matches. This could be as high as 16, but using a smaller hash table
126 * speeds up compression due to reduced cache pressure.
128 #define BT_MATCHFINDER_HASH2_ORDER 12
131 * The number of lz_match structures in the match cache, excluding the extra
132 * "overflow" entries. This value should be high enough so that nearly the
133 * time, all matches found in a given block can fit in the match cache.
134 * However, fallback behavior (immediately terminating the block) on cache
135 * overflow is still required.
137 #define CACHE_LENGTH (SOFT_MAX_BLOCK_SIZE * 5)
140 * An upper bound on the number of matches that can ever be saved in the match
141 * cache for a single position. Since each match we save for a single position
142 * has a distinct length, we can use the number of possible match lengths in LZX
143 * as this bound. This bound is guaranteed to be valid in all cases, although
144 * if 'nice_match_length < LZX_MAX_MATCH_LEN', then it will never actually be
147 #define MAX_MATCHES_PER_POS LZX_NUM_LENS
150 * A scaling factor that makes it possible to consider fractional bit costs. A
151 * single bit has a cost of BIT_COST.
153 * Note: this is only useful as a statistical trick for when the true costs are
154 * unknown. Ultimately, each token in LZX requires a whole number of bits to
160 * Should the compressor take into account the costs of aligned offset symbols
161 * instead of assuming that all are equally likely?
163 #define CONSIDER_ALIGNED_COSTS 1
166 * Should the "minimum" cost path search algorithm consider "gap" matches, where
167 * a normal match is followed by a literal, then by a match with the same
168 * offset? This is one specific, somewhat common situation in which the true
169 * minimum cost path is often different from the path found by looking only one
172 #define CONSIDER_GAP_MATCHES 1
174 /******************************************************************************/
176 /*----------------------------------------------------------------------------*/
182 #include "wimlib/compress_common.h"
183 #include "wimlib/compressor_ops.h"
184 #include "wimlib/error.h"
185 #include "wimlib/lz_extend.h"
186 #include "wimlib/lzx_common.h"
187 #include "wimlib/unaligned.h"
188 #include "wimlib/util.h"
190 /* Note: BT_MATCHFINDER_HASH2_ORDER must be defined before including
191 * bt_matchfinder.h. */
193 /* Matchfinders with 16-bit positions */
195 #define MF_SUFFIX _16
196 #include "wimlib/bt_matchfinder.h"
197 #include "wimlib/hc_matchfinder.h"
199 /* Matchfinders with 32-bit positions */
203 #define MF_SUFFIX _32
204 #include "wimlib/bt_matchfinder.h"
205 #include "wimlib/hc_matchfinder.h"
207 /******************************************************************************/
208 /* Compressor structure */
209 /*----------------------------------------------------------------------------*/
211 /* Codewords for the Huffman codes */
212 struct lzx_codewords {
213 u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
214 u32 len[LZX_LENCODE_NUM_SYMBOLS];
215 u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
219 * Codeword lengths, in bits, for the Huffman codes.
221 * A codeword length of 0 means the corresponding codeword has zero frequency.
223 * The main and length codes each have one extra entry for use as a sentinel.
224 * See lzx_write_compressed_code().
227 u8 main[LZX_MAINCODE_MAX_NUM_SYMBOLS + 1];
228 u8 len[LZX_LENCODE_NUM_SYMBOLS + 1];
229 u8 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
232 /* Codewords and lengths for the Huffman codes */
234 struct lzx_codewords codewords;
235 struct lzx_lens lens;
238 /* Symbol frequency counters for the Huffman-encoded alphabets */
240 u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
241 u32 len[LZX_LENCODE_NUM_SYMBOLS];
242 u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
245 /* Block split statistics. See the "Block splitting algorithm" section later in
246 * this file for details. */
247 #define NUM_LITERAL_OBSERVATION_TYPES 8
248 #define NUM_MATCH_OBSERVATION_TYPES 2
249 #define NUM_OBSERVATION_TYPES (NUM_LITERAL_OBSERVATION_TYPES + \
250 NUM_MATCH_OBSERVATION_TYPES)
251 struct lzx_block_split_stats {
252 u32 new_observations[NUM_OBSERVATION_TYPES];
253 u32 observations[NUM_OBSERVATION_TYPES];
254 u32 num_new_observations;
255 u32 num_observations;
259 * Represents a run of literals followed by a match or end-of-block. This
260 * structure is needed to temporarily store items chosen by the compressor,
261 * since items cannot be written until all items for the block have been chosen
262 * and the block's Huffman codes have been computed.
264 struct lzx_sequence {
266 /* The number of literals in the run. This may be 0. The literals are
267 * not stored explicitly in this structure; instead, they are read
268 * directly from the uncompressed data. */
271 /* If the next field doesn't indicate end-of-block, then this is the
272 * match length minus LZX_MIN_MATCH_LEN. */
275 /* If bit 31 is clear, then this field contains the match header in bits
276 * 0-8, and either the match offset plus LZX_OFFSET_ADJUSTMENT or a
277 * recent offset code in bits 9-30. Otherwise (if bit 31 is set), this
278 * sequence's literal run was the last literal run in the block, so
279 * there is no match that follows it. */
280 u32 adjusted_offset_and_match_hdr;
284 * This structure represents a byte position in the input buffer and a node in
285 * the graph of possible match/literal choices.
287 * Logically, each incoming edge to this node is labeled with a literal or a
288 * match that can be taken to reach this position from an earlier position; and
289 * each outgoing edge from this node is labeled with a literal or a match that
290 * can be taken to advance from this position to a later position.
292 struct lzx_optimum_node {
294 /* The cost, in bits, of the lowest-cost path that has been found to
295 * reach this position. This can change as progressively lower cost
296 * paths are found to reach this position. */
300 * The best arrival to this node, i.e. the match or literal that was
301 * used to arrive to this position at the given 'cost'. This can change
302 * as progressively lower cost paths are found to reach this position.
304 * For non-gap matches, this variable is divided into two bitfields
305 * whose meanings depend on the item type:
308 * Low bits are 0, high bits are the literal.
310 * Explicit offset matches:
311 * Low bits are the match length, high bits are the offset plus
312 * LZX_OFFSET_ADJUSTMENT.
314 * Repeat offset matches:
315 * Low bits are the match length, high bits are the queue index.
317 * For gap matches, identified by OPTIMUM_GAP_MATCH set, special
318 * behavior applies --- see the code.
321 #define OPTIMUM_OFFSET_SHIFT 9
322 #define OPTIMUM_LEN_MASK ((1 << OPTIMUM_OFFSET_SHIFT) - 1)
323 #if CONSIDER_GAP_MATCHES
324 # define OPTIMUM_GAP_MATCH 0x80000000
327 } _aligned_attribute(8);
329 /* The cost model for near-optimal parsing */
333 * 'match_cost[offset_slot][len - LZX_MIN_MATCH_LEN]' is the cost of a
334 * length 'len' match which has an offset belonging to 'offset_slot'.
335 * The cost includes the main symbol, the length symbol if required, and
336 * the extra offset bits if any, excluding any entropy-coded bits
337 * (aligned offset bits). It does *not* include the cost of the aligned
338 * offset symbol which may be required.
340 u16 match_cost[LZX_MAX_OFFSET_SLOTS][LZX_NUM_LENS];
342 /* Cost of each symbol in the main code */
343 u32 main[LZX_MAINCODE_MAX_NUM_SYMBOLS];
345 /* Cost of each symbol in the length code */
346 u32 len[LZX_LENCODE_NUM_SYMBOLS];
348 #if CONSIDER_ALIGNED_COSTS
349 /* Cost of each symbol in the aligned offset code */
350 u32 aligned[LZX_ALIGNEDCODE_NUM_SYMBOLS];
354 struct lzx_output_bitstream;
356 /* The main LZX compressor structure */
357 struct lzx_compressor {
359 /* The buffer for preprocessed input data, if not using destructive
363 /* If true, then the compressor need not preserve the input buffer if it
364 * compresses the data successfully */
367 /* Pointer to the compress() implementation chosen at allocation time */
368 void (*impl)(struct lzx_compressor *, const u8 *, size_t,
369 struct lzx_output_bitstream *);
371 /* The log base 2 of the window size for match offset encoding purposes.
372 * This will be >= LZX_MIN_WINDOW_ORDER and <= LZX_MAX_WINDOW_ORDER. */
373 unsigned window_order;
375 /* The number of symbols in the main alphabet. This depends on the
376 * window order, since the window order determines the maximum possible
378 unsigned num_main_syms;
380 /* The "nice" match length: if a match of this length is found, then it
381 * is chosen immediately without further consideration. */
382 unsigned nice_match_length;
384 /* The maximum search depth: at most this many potential matches are
385 * considered at each position. */
386 unsigned max_search_depth;
388 /* The number of optimization passes per block */
389 unsigned num_optim_passes;
391 /* The symbol frequency counters for the current block */
392 struct lzx_freqs freqs;
394 /* Block split statistics for the current block */
395 struct lzx_block_split_stats split_stats;
397 /* The Huffman codes for the current and previous blocks. The one with
398 * index 'codes_index' is for the current block, and the other one is
399 * for the previous block. */
400 struct lzx_codes codes[2];
401 unsigned codes_index;
403 /* The matches and literals that the compressor has chosen for the
404 * current block. The required length of this array is limited by the
405 * maximum number of matches that can ever be chosen for a single block,
406 * plus one for the special entry at the end. */
407 struct lzx_sequence chosen_sequences[
408 DIV_ROUND_UP(SOFT_MAX_BLOCK_SIZE, LZX_MIN_MATCH_LEN) + 1];
410 /* Tables for mapping adjusted offsets to offset slots */
411 u8 offset_slot_tab_1[32768]; /* offset slots [0, 29] */
412 u8 offset_slot_tab_2[128]; /* offset slots [30, 49] */
415 /* Data for lzx_compress_lazy() */
417 /* Hash chains matchfinder (MUST BE LAST!!!) */
419 struct hc_matchfinder_16 hc_mf_16;
420 struct hc_matchfinder_32 hc_mf_32;
424 /* Data for lzx_compress_near_optimal() */
427 * Array of nodes, one per position, for running the
428 * minimum-cost path algorithm.
430 * This array must be large enough to accommodate the
431 * worst-case number of nodes, which occurs if the
432 * compressor finds a match of length LZX_MAX_MATCH_LEN
433 * at position 'SOFT_MAX_BLOCK_SIZE - 1', producing a
434 * block of size 'SOFT_MAX_BLOCK_SIZE - 1 +
435 * LZX_MAX_MATCH_LEN'. Add one for the end-of-block
438 struct lzx_optimum_node optimum_nodes[
439 SOFT_MAX_BLOCK_SIZE - 1 +
440 LZX_MAX_MATCH_LEN + 1];
442 /* The cost model for the current optimization pass */
443 struct lzx_costs costs;
446 * Cached matches for the current block. This array
447 * contains the matches that were found at each position
448 * in the block. Specifically, for each position, there
449 * is a special 'struct lz_match' whose 'length' field
450 * contains the number of matches that were found at
451 * that position; this is followed by the matches
452 * themselves, if any, sorted by strictly increasing
455 * Note: in rare cases, there will be a very high number
456 * of matches in the block and this array will overflow.
457 * If this happens, we force the end of the current
458 * block. CACHE_LENGTH is the length at which we
459 * actually check for overflow. The extra slots beyond
460 * this are enough to absorb the worst case overflow,
461 * which occurs if starting at &match_cache[CACHE_LENGTH
462 * - 1], we write the match count header, then write
463 * MAX_MATCHES_PER_POS matches, then skip searching for
464 * matches at 'LZX_MAX_MATCH_LEN - 1' positions and
465 * write the match count header for each.
467 struct lz_match match_cache[CACHE_LENGTH +
468 MAX_MATCHES_PER_POS +
469 LZX_MAX_MATCH_LEN - 1];
471 /* Binary trees matchfinder (MUST BE LAST!!!) */
473 struct bt_matchfinder_16 bt_mf_16;
474 struct bt_matchfinder_32 bt_mf_32;
480 /******************************************************************************/
481 /* Matchfinder utilities */
482 /*----------------------------------------------------------------------------*/
485 * Will a matchfinder using 16-bit positions be sufficient for compressing
486 * buffers of up to the specified size? The limit could be 65536 bytes, but we
487 * also want to optimize out the use of offset_slot_tab_2 in the 16-bit case.
488 * This requires that the limit be no more than the length of offset_slot_tab_1
492 lzx_is_16_bit(size_t max_bufsize)
494 STATIC_ASSERT(ARRAY_LEN(((struct lzx_compressor *)0)->offset_slot_tab_1) == 32768);
495 return max_bufsize <= 32768;
499 * Return the offset slot for the specified adjusted match offset.
501 static inline unsigned
502 lzx_get_offset_slot(struct lzx_compressor *c, u32 adjusted_offset,
505 if (is_16_bit || adjusted_offset < ARRAY_LEN(c->offset_slot_tab_1))
506 return c->offset_slot_tab_1[adjusted_offset];
507 return c->offset_slot_tab_2[adjusted_offset >> 14];
511 * The following macros call either the 16-bit or the 32-bit version of a
512 * matchfinder function based on the value of 'is_16_bit', which will be known
513 * at compilation time.
516 #define CALL_HC_MF(is_16_bit, c, funcname, ...) \
517 ((is_16_bit) ? CONCAT(funcname, _16)(&(c)->hc_mf_16, ##__VA_ARGS__) : \
518 CONCAT(funcname, _32)(&(c)->hc_mf_32, ##__VA_ARGS__));
520 #define CALL_BT_MF(is_16_bit, c, funcname, ...) \
521 ((is_16_bit) ? CONCAT(funcname, _16)(&(c)->bt_mf_16, ##__VA_ARGS__) : \
522 CONCAT(funcname, _32)(&(c)->bt_mf_32, ##__VA_ARGS__));
524 /******************************************************************************/
525 /* Output bitstream */
526 /*----------------------------------------------------------------------------*/
529 * The LZX bitstream is encoded as a sequence of little endian 16-bit coding
530 * units. Bits are ordered from most significant to least significant within
535 * Structure to keep track of the current state of sending bits to the
536 * compressed output buffer.
538 struct lzx_output_bitstream {
540 /* Bits that haven't yet been written to the output buffer */
541 machine_word_t bitbuf;
543 /* Number of bits currently held in @bitbuf */
544 machine_word_t bitcount;
546 /* Pointer to the start of the output buffer */
549 /* Pointer to the position in the output buffer at which the next coding
550 * unit should be written */
553 /* Pointer to just past the end of the output buffer, rounded down by
554 * one byte if needed to make 'end - start' a multiple of 2 */
558 /* Can the specified number of bits always be added to 'bitbuf' after all
559 * pending 16-bit coding units have been flushed? */
560 #define CAN_BUFFER(n) ((n) <= WORDBITS - 15)
562 /* Initialize the output bitstream to write to the specified buffer. */
564 lzx_init_output(struct lzx_output_bitstream *os, void *buffer, size_t size)
570 os->end = (u8 *)buffer + (size & ~1);
574 * Add some bits to the bitbuffer variable of the output bitstream. The caller
575 * 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;
585 * Flush bits from the bitbuffer variable to the output buffer. 'max_num_bits'
586 * 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 = WORDBITS - 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 /******************************************************************************/
638 /* Preparing Huffman codes */
639 /*----------------------------------------------------------------------------*/
642 * Build the Huffman codes. This takes as input the frequency tables for each
643 * code and produces as output a set of tables that map symbols to codewords and
647 lzx_build_huffman_codes(struct lzx_compressor *c)
649 const struct lzx_freqs *freqs = &c->freqs;
650 struct lzx_codes *codes = &c->codes[c->codes_index];
652 STATIC_ASSERT(MAIN_CODEWORD_LIMIT >= 9 &&
653 MAIN_CODEWORD_LIMIT <= LZX_MAX_MAIN_CODEWORD_LEN);
654 make_canonical_huffman_code(c->num_main_syms,
658 codes->codewords.main);
660 STATIC_ASSERT(LENGTH_CODEWORD_LIMIT >= 8 &&
661 LENGTH_CODEWORD_LIMIT <= LZX_MAX_LEN_CODEWORD_LEN);
662 make_canonical_huffman_code(LZX_LENCODE_NUM_SYMBOLS,
663 LENGTH_CODEWORD_LIMIT,
666 codes->codewords.len);
668 STATIC_ASSERT(ALIGNED_CODEWORD_LIMIT >= LZX_NUM_ALIGNED_OFFSET_BITS &&
669 ALIGNED_CODEWORD_LIMIT <= LZX_MAX_ALIGNED_CODEWORD_LEN);
670 make_canonical_huffman_code(LZX_ALIGNEDCODE_NUM_SYMBOLS,
671 ALIGNED_CODEWORD_LIMIT,
674 codes->codewords.aligned);
677 /* Reset the symbol frequencies for the current block. */
679 lzx_reset_symbol_frequencies(struct lzx_compressor *c)
681 memset(&c->freqs, 0, sizeof(c->freqs));
685 lzx_compute_precode_items(const u8 lens[restrict],
686 const u8 prev_lens[restrict],
687 u32 precode_freqs[restrict],
688 unsigned precode_items[restrict])
697 itemptr = precode_items;
700 while (!((len = lens[run_start]) & 0x80)) {
702 /* len = the length being repeated */
704 /* Find the next run of codeword lengths. */
706 run_end = run_start + 1;
708 /* Fast case for a single length. */
709 if (likely(len != lens[run_end])) {
710 delta = prev_lens[run_start] - len;
713 precode_freqs[delta]++;
719 /* Extend the run. */
722 } while (len == lens[run_end]);
727 /* Symbol 18: RLE 20 to 51 zeroes at a time. */
728 while ((run_end - run_start) >= 20) {
729 extra_bits = min((run_end - run_start) - 20, 0x1F);
731 *itemptr++ = 18 | (extra_bits << 5);
732 run_start += 20 + extra_bits;
735 /* Symbol 17: RLE 4 to 19 zeroes at a time. */
736 if ((run_end - run_start) >= 4) {
737 extra_bits = min((run_end - run_start) - 4, 0xF);
739 *itemptr++ = 17 | (extra_bits << 5);
740 run_start += 4 + extra_bits;
744 /* A run of nonzero lengths. */
746 /* Symbol 19: RLE 4 to 5 of any length at a time. */
747 while ((run_end - run_start) >= 4) {
748 extra_bits = (run_end - run_start) > 4;
749 delta = prev_lens[run_start] - len;
753 precode_freqs[delta]++;
754 *itemptr++ = 19 | (extra_bits << 5) | (delta << 6);
755 run_start += 4 + extra_bits;
759 /* Output any remaining lengths without RLE. */
760 while (run_start != run_end) {
761 delta = prev_lens[run_start] - len;
764 precode_freqs[delta]++;
770 return itemptr - precode_items;
773 /******************************************************************************/
774 /* Outputting compressed data */
775 /*----------------------------------------------------------------------------*/
778 * Output a Huffman code in the compressed form used in LZX.
780 * The Huffman code is represented in the output as a logical series of codeword
781 * lengths from which the Huffman code, which must be in canonical form, can be
784 * The codeword lengths are themselves compressed using a separate Huffman code,
785 * the "precode", which contains a symbol for each possible codeword length in
786 * the larger code as well as several special symbols to represent repeated
787 * codeword lengths (a form of run-length encoding). The precode is itself
788 * constructed in canonical form, and its codeword lengths are represented
789 * literally in 20 4-bit fields that immediately precede the compressed codeword
790 * lengths of the larger code.
792 * Furthermore, the codeword lengths of the larger code are actually represented
793 * as deltas from the codeword lengths of the corresponding code in the previous
797 * Bitstream to which to write the compressed Huffman code.
799 * The codeword lengths, indexed by symbol, in the Huffman code.
801 * The codeword lengths, indexed by symbol, in the corresponding Huffman
802 * code in the previous block, or all zeroes if this is the first block.
804 * The number of symbols in the Huffman code.
807 lzx_write_compressed_code(struct lzx_output_bitstream *os,
808 const u8 lens[restrict],
809 const u8 prev_lens[restrict],
812 u32 precode_freqs[LZX_PRECODE_NUM_SYMBOLS];
813 u8 precode_lens[LZX_PRECODE_NUM_SYMBOLS];
814 u32 precode_codewords[LZX_PRECODE_NUM_SYMBOLS];
815 unsigned precode_items[num_lens];
816 unsigned num_precode_items;
817 unsigned precode_item;
818 unsigned precode_sym;
820 u8 saved = lens[num_lens];
821 *(u8 *)(lens + num_lens) = 0x80;
823 for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++)
824 precode_freqs[i] = 0;
826 /* Compute the "items" (RLE / literal tokens and extra bits) with which
827 * the codeword lengths in the larger code will be output. */
828 num_precode_items = lzx_compute_precode_items(lens,
833 /* Build the precode. */
834 STATIC_ASSERT(PRE_CODEWORD_LIMIT >= 5 &&
835 PRE_CODEWORD_LIMIT <= LZX_MAX_PRE_CODEWORD_LEN);
836 make_canonical_huffman_code(LZX_PRECODE_NUM_SYMBOLS, PRE_CODEWORD_LIMIT,
837 precode_freqs, precode_lens,
840 /* Output the lengths of the codewords in the precode. */
841 for (i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++)
842 lzx_write_bits(os, precode_lens[i], LZX_PRECODE_ELEMENT_SIZE);
844 /* Output the encoded lengths of the codewords in the larger code. */
845 for (i = 0; i < num_precode_items; i++) {
846 precode_item = precode_items[i];
847 precode_sym = precode_item & 0x1F;
848 lzx_add_bits(os, precode_codewords[precode_sym],
849 precode_lens[precode_sym]);
850 if (precode_sym >= 17) {
851 if (precode_sym == 17) {
852 lzx_add_bits(os, precode_item >> 5, 4);
853 } else if (precode_sym == 18) {
854 lzx_add_bits(os, precode_item >> 5, 5);
856 lzx_add_bits(os, (precode_item >> 5) & 1, 1);
857 precode_sym = precode_item >> 6;
858 lzx_add_bits(os, precode_codewords[precode_sym],
859 precode_lens[precode_sym]);
862 STATIC_ASSERT(CAN_BUFFER(2 * PRE_CODEWORD_LIMIT + 1));
863 lzx_flush_bits(os, 2 * PRE_CODEWORD_LIMIT + 1);
866 *(u8 *)(lens + num_lens) = saved;
870 * Write all matches and literal bytes (which were precomputed) in an LZX
871 * compressed block to the output bitstream in the final compressed
875 * The output bitstream.
877 * The chosen type of the LZX compressed block (LZX_BLOCKTYPE_ALIGNED or
878 * LZX_BLOCKTYPE_VERBATIM).
880 * The uncompressed data of the block.
882 * The matches and literals to output, given as a series of sequences.
884 * The main, length, and aligned offset Huffman codes for the block.
887 lzx_write_sequences(struct lzx_output_bitstream *os, int block_type,
888 const u8 *block_data, const struct lzx_sequence sequences[],
889 const struct lzx_codes *codes)
891 const struct lzx_sequence *seq = sequences;
892 u32 ones_if_aligned = 0 - (block_type == LZX_BLOCKTYPE_ALIGNED);
895 /* Output the next sequence. */
897 unsigned litrunlen = seq->litrunlen;
899 unsigned main_symbol;
900 unsigned adjusted_length;
902 unsigned offset_slot;
903 unsigned num_extra_bits;
906 /* Output the literal run of the sequence. */
908 if (litrunlen) { /* Is the literal run nonempty? */
910 /* Verify optimization is enabled on 64-bit */
911 STATIC_ASSERT(WORDBITS < 64 ||
912 CAN_BUFFER(3 * MAIN_CODEWORD_LIMIT));
914 if (CAN_BUFFER(3 * MAIN_CODEWORD_LIMIT)) {
916 /* 64-bit: write 3 literals at a time. */
917 while (litrunlen >= 3) {
918 unsigned lit0 = block_data[0];
919 unsigned lit1 = block_data[1];
920 unsigned lit2 = block_data[2];
921 lzx_add_bits(os, codes->codewords.main[lit0],
922 codes->lens.main[lit0]);
923 lzx_add_bits(os, codes->codewords.main[lit1],
924 codes->lens.main[lit1]);
925 lzx_add_bits(os, codes->codewords.main[lit2],
926 codes->lens.main[lit2]);
927 lzx_flush_bits(os, 3 * MAIN_CODEWORD_LIMIT);
932 unsigned lit = *block_data++;
933 lzx_add_bits(os, codes->codewords.main[lit],
934 codes->lens.main[lit]);
936 unsigned lit = *block_data++;
937 lzx_add_bits(os, codes->codewords.main[lit],
938 codes->lens.main[lit]);
939 lzx_flush_bits(os, 2 * MAIN_CODEWORD_LIMIT);
941 lzx_flush_bits(os, 1 * MAIN_CODEWORD_LIMIT);
945 /* 32-bit: write 1 literal at a time. */
947 unsigned lit = *block_data++;
948 lzx_add_bits(os, codes->codewords.main[lit],
949 codes->lens.main[lit]);
950 lzx_flush_bits(os, MAIN_CODEWORD_LIMIT);
951 } while (--litrunlen);
955 /* Was this the last literal run? */
956 if (seq->adjusted_offset_and_match_hdr & 0x80000000)
959 /* Nope; output the match. */
961 match_hdr = seq->adjusted_offset_and_match_hdr & 0x1FF;
962 main_symbol = LZX_NUM_CHARS + match_hdr;
963 adjusted_length = seq->adjusted_length;
965 block_data += adjusted_length + LZX_MIN_MATCH_LEN;
967 offset_slot = match_hdr / LZX_NUM_LEN_HEADERS;
968 adjusted_offset = seq->adjusted_offset_and_match_hdr >> 9;
970 num_extra_bits = lzx_extra_offset_bits[offset_slot];
971 extra_bits = adjusted_offset - (lzx_offset_slot_base[offset_slot] +
972 LZX_OFFSET_ADJUSTMENT);
974 #define MAX_MATCH_BITS (MAIN_CODEWORD_LIMIT + LENGTH_CODEWORD_LIMIT + \
975 14 + ALIGNED_CODEWORD_LIMIT)
977 /* Verify optimization is enabled on 64-bit */
978 STATIC_ASSERT(WORDBITS < 64 || CAN_BUFFER(MAX_MATCH_BITS));
980 /* Output the main symbol for the match. */
982 lzx_add_bits(os, codes->codewords.main[main_symbol],
983 codes->lens.main[main_symbol]);
984 if (!CAN_BUFFER(MAX_MATCH_BITS))
985 lzx_flush_bits(os, MAIN_CODEWORD_LIMIT);
987 /* If needed, output the length symbol for the match. */
989 if (adjusted_length >= LZX_NUM_PRIMARY_LENS) {
990 lzx_add_bits(os, codes->codewords.len[adjusted_length -
991 LZX_NUM_PRIMARY_LENS],
992 codes->lens.len[adjusted_length -
993 LZX_NUM_PRIMARY_LENS]);
994 if (!CAN_BUFFER(MAX_MATCH_BITS))
995 lzx_flush_bits(os, LENGTH_CODEWORD_LIMIT);
998 /* Output the extra offset bits for the match. In aligned
999 * offset blocks, the lowest 3 bits of the adjusted offset are
1000 * Huffman-encoded using the aligned offset code, provided that
1001 * there are at least extra 3 offset bits required. All other
1002 * extra offset bits are output verbatim. */
1004 if ((adjusted_offset & ones_if_aligned) >= 16) {
1006 lzx_add_bits(os, extra_bits >> LZX_NUM_ALIGNED_OFFSET_BITS,
1007 num_extra_bits - LZX_NUM_ALIGNED_OFFSET_BITS);
1008 if (!CAN_BUFFER(MAX_MATCH_BITS))
1009 lzx_flush_bits(os, 14);
1011 lzx_add_bits(os, codes->codewords.aligned[adjusted_offset &
1012 LZX_ALIGNED_OFFSET_BITMASK],
1013 codes->lens.aligned[adjusted_offset &
1014 LZX_ALIGNED_OFFSET_BITMASK]);
1015 if (!CAN_BUFFER(MAX_MATCH_BITS))
1016 lzx_flush_bits(os, ALIGNED_CODEWORD_LIMIT);
1018 STATIC_ASSERT(CAN_BUFFER(17));
1020 lzx_add_bits(os, extra_bits, num_extra_bits);
1021 if (!CAN_BUFFER(MAX_MATCH_BITS))
1022 lzx_flush_bits(os, 17);
1025 if (CAN_BUFFER(MAX_MATCH_BITS))
1026 lzx_flush_bits(os, MAX_MATCH_BITS);
1028 /* Advance to the next sequence. */
1034 lzx_write_compressed_block(const u8 *block_begin,
1037 unsigned window_order,
1038 unsigned num_main_syms,
1039 const struct lzx_sequence sequences[],
1040 const struct lzx_codes * codes,
1041 const struct lzx_lens * prev_lens,
1042 struct lzx_output_bitstream * os)
1044 /* The first three bits indicate the type of block and are one of the
1045 * LZX_BLOCKTYPE_* constants. */
1046 lzx_write_bits(os, block_type, 3);
1049 * Output the block size.
1051 * The original LZX format encoded the block size in 24 bits. However,
1052 * the LZX format used in WIM archives uses 1 bit to specify whether the
1053 * block has the default size of 32768 bytes, then optionally 16 bits to
1054 * specify a non-default size. This works fine for Microsoft's WIM
1055 * software (WIMGAPI), which never compresses more than 32768 bytes at a
1056 * time with LZX. However, as an extension, our LZX compressor supports
1057 * compressing up to 2097152 bytes, with a corresponding increase in
1058 * window size. It is possible for blocks in these larger buffers to
1059 * exceed 65535 bytes; such blocks cannot have their size represented in
1062 * The chosen solution was to use 24 bits for the block size when
1063 * possibly required --- specifically, when the compressor has been
1064 * allocated to be capable of compressing more than 32768 bytes at once
1065 * (which also causes the number of main symbols to be increased).
1067 if (block_size == LZX_DEFAULT_BLOCK_SIZE) {
1068 lzx_write_bits(os, 1, 1);
1070 lzx_write_bits(os, 0, 1);
1072 if (window_order >= 16)
1073 lzx_write_bits(os, block_size >> 16, 8);
1075 lzx_write_bits(os, block_size & 0xFFFF, 16);
1078 /* If it's an aligned offset block, output the aligned offset code. */
1079 if (block_type == LZX_BLOCKTYPE_ALIGNED) {
1080 for (int i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1081 lzx_write_bits(os, codes->lens.aligned[i],
1082 LZX_ALIGNEDCODE_ELEMENT_SIZE);
1086 /* Output the main code (two parts). */
1087 lzx_write_compressed_code(os, codes->lens.main,
1090 lzx_write_compressed_code(os, codes->lens.main + LZX_NUM_CHARS,
1091 prev_lens->main + LZX_NUM_CHARS,
1092 num_main_syms - LZX_NUM_CHARS);
1094 /* Output the length code. */
1095 lzx_write_compressed_code(os, codes->lens.len,
1097 LZX_LENCODE_NUM_SYMBOLS);
1099 /* Output the compressed matches and literals. */
1100 lzx_write_sequences(os, block_type, block_begin, sequences, codes);
1104 * Given the frequencies of symbols in an LZX-compressed block and the
1105 * corresponding Huffman codes, return LZX_BLOCKTYPE_ALIGNED or
1106 * LZX_BLOCKTYPE_VERBATIM if an aligned offset or verbatim block, respectively,
1107 * will take fewer bits to output.
1110 lzx_choose_verbatim_or_aligned(const struct lzx_freqs * freqs,
1111 const struct lzx_codes * codes)
1113 u32 verbatim_cost = 0;
1114 u32 aligned_cost = 0;
1116 /* A verbatim block requires 3 bits in each place that an aligned offset
1117 * symbol would be used in an aligned offset block. */
1118 for (unsigned i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
1119 verbatim_cost += LZX_NUM_ALIGNED_OFFSET_BITS * freqs->aligned[i];
1120 aligned_cost += codes->lens.aligned[i] * freqs->aligned[i];
1123 /* Account for the cost of sending the codeword lengths of the aligned
1125 aligned_cost += LZX_ALIGNEDCODE_ELEMENT_SIZE *
1126 LZX_ALIGNEDCODE_NUM_SYMBOLS;
1128 if (aligned_cost < verbatim_cost)
1129 return LZX_BLOCKTYPE_ALIGNED;
1131 return LZX_BLOCKTYPE_VERBATIM;
1135 * Flush an LZX block:
1137 * 1. Build the Huffman codes.
1138 * 2. Decide whether to output the block as VERBATIM or ALIGNED.
1139 * 3. Write the block.
1140 * 4. Swap the indices of the current and previous Huffman codes.
1142 * Note: we never output UNCOMPRESSED blocks. This probably should be
1143 * implemented sometime, but it doesn't make much difference.
1146 lzx_flush_block(struct lzx_compressor *c, struct lzx_output_bitstream *os,
1147 const u8 *block_begin, u32 block_size, u32 seq_idx)
1151 lzx_build_huffman_codes(c);
1153 block_type = lzx_choose_verbatim_or_aligned(&c->freqs,
1154 &c->codes[c->codes_index]);
1155 lzx_write_compressed_block(block_begin,
1160 &c->chosen_sequences[seq_idx],
1161 &c->codes[c->codes_index],
1162 &c->codes[c->codes_index ^ 1].lens,
1164 c->codes_index ^= 1;
1167 /******************************************************************************/
1168 /* Block splitting algorithm */
1169 /*----------------------------------------------------------------------------*/
1172 * The problem of block splitting is to decide when it is worthwhile to start a
1173 * new block with new entropy codes. There is a theoretically optimal solution:
1174 * recursively consider every possible block split, considering the exact cost
1175 * of each block, and choose the minimum cost approach. But this is far too
1176 * slow. Instead, as an approximation, we can count symbols and after every N
1177 * symbols, compare the expected distribution of symbols based on the previous
1178 * data with the actual distribution. If they differ "by enough", then start a
1181 * As an optimization and heuristic, we don't distinguish between every symbol
1182 * but rather we combine many symbols into a single "observation type". For
1183 * literals we only look at the high bits and low bits, and for matches we only
1184 * look at whether the match is long or not. The assumption is that for typical
1185 * "real" data, places that are good block boundaries will tend to be noticable
1186 * based only on changes in these aggregate frequencies, without looking for
1187 * subtle differences in individual symbols. For example, a change from ASCII
1188 * bytes to non-ASCII bytes, or from few matches (generally less compressible)
1189 * to many matches (generally more compressible), would be easily noticed based
1190 * on the aggregates.
1192 * For determining whether the frequency distributions are "different enough" to
1193 * start a new block, the simply heuristic of splitting when the sum of absolute
1194 * differences exceeds a constant seems to be good enough.
1196 * Finally, for an approximation, it is not strictly necessary that the exact
1197 * symbols being used are considered. With "near-optimal parsing", for example,
1198 * the actual symbols that will be used are unknown until after the block
1199 * boundary is chosen and the block has been optimized. Since the final choices
1200 * cannot be used, we can use preliminary "greedy" choices instead.
1203 /* Initialize the block split statistics when starting a new block. */
1205 lzx_init_block_split_stats(struct lzx_block_split_stats *stats)
1207 memset(stats, 0, sizeof(*stats));
1210 /* Literal observation. Heuristic: use the top 2 bits and low 1 bits of the
1211 * literal, for 8 possible literal observation types. */
1213 lzx_observe_literal(struct lzx_block_split_stats *stats, u8 lit)
1215 stats->new_observations[((lit >> 5) & 0x6) | (lit & 1)]++;
1216 stats->num_new_observations++;
1219 /* Match observation. Heuristic: use one observation type for "short match" and
1220 * one observation type for "long match". */
1222 lzx_observe_match(struct lzx_block_split_stats *stats, unsigned length)
1224 stats->new_observations[NUM_LITERAL_OBSERVATION_TYPES + (length >= 5)]++;
1225 stats->num_new_observations++;
1229 lzx_should_end_block(struct lzx_block_split_stats *stats)
1231 if (stats->num_observations > 0) {
1233 /* Note: to avoid slow divisions, we do not divide by
1234 * 'num_observations', but rather do all math with the numbers
1235 * multiplied by 'num_observations'. */
1236 u32 total_delta = 0;
1237 for (int i = 0; i < NUM_OBSERVATION_TYPES; i++) {
1238 u32 expected = stats->observations[i] *
1239 stats->num_new_observations;
1240 u32 actual = stats->new_observations[i] *
1241 stats->num_observations;
1242 u32 delta = (actual > expected) ? actual - expected :
1244 total_delta += delta;
1247 /* Ready to end the block? */
1249 stats->num_new_observations * 7 / 8 * stats->num_observations)
1253 for (int i = 0; i < NUM_OBSERVATION_TYPES; i++) {
1254 stats->num_observations += stats->new_observations[i];
1255 stats->observations[i] += stats->new_observations[i];
1256 stats->new_observations[i] = 0;
1258 stats->num_new_observations = 0;
1262 /******************************************************************************/
1263 /* Slower ("near-optimal") compression algorithm */
1264 /*----------------------------------------------------------------------------*/
1267 * Least-recently-used queue for match offsets.
1269 * This is represented as a 64-bit integer for efficiency. There are three
1270 * offsets of 21 bits each. Bit 64 is garbage.
1272 struct lzx_lru_queue {
1274 } _aligned_attribute(8);
1276 #define LZX_QUEUE_OFFSET_SHIFT 21
1277 #define LZX_QUEUE_OFFSET_MASK (((u64)1 << LZX_QUEUE_OFFSET_SHIFT) - 1)
1279 #define LZX_QUEUE_R0_SHIFT (0 * LZX_QUEUE_OFFSET_SHIFT)
1280 #define LZX_QUEUE_R1_SHIFT (1 * LZX_QUEUE_OFFSET_SHIFT)
1281 #define LZX_QUEUE_R2_SHIFT (2 * LZX_QUEUE_OFFSET_SHIFT)
1283 #define LZX_QUEUE_R0_MASK (LZX_QUEUE_OFFSET_MASK << LZX_QUEUE_R0_SHIFT)
1284 #define LZX_QUEUE_R1_MASK (LZX_QUEUE_OFFSET_MASK << LZX_QUEUE_R1_SHIFT)
1285 #define LZX_QUEUE_R2_MASK (LZX_QUEUE_OFFSET_MASK << LZX_QUEUE_R2_SHIFT)
1287 #define LZX_QUEUE_INITIALIZER { \
1288 ((u64)1 << LZX_QUEUE_R0_SHIFT) | \
1289 ((u64)1 << LZX_QUEUE_R1_SHIFT) | \
1290 ((u64)1 << LZX_QUEUE_R2_SHIFT) }
1293 lzx_lru_queue_R0(struct lzx_lru_queue queue)
1295 return (queue.R >> LZX_QUEUE_R0_SHIFT) & LZX_QUEUE_OFFSET_MASK;
1299 lzx_lru_queue_R1(struct lzx_lru_queue queue)
1301 return (queue.R >> LZX_QUEUE_R1_SHIFT) & LZX_QUEUE_OFFSET_MASK;
1305 lzx_lru_queue_R2(struct lzx_lru_queue queue)
1307 return (queue.R >> LZX_QUEUE_R2_SHIFT) & LZX_QUEUE_OFFSET_MASK;
1310 /* Push a match offset onto the front (most recently used) end of the queue. */
1311 static inline struct lzx_lru_queue
1312 lzx_lru_queue_push(struct lzx_lru_queue queue, u32 offset)
1314 return (struct lzx_lru_queue) {
1315 .R = (queue.R << LZX_QUEUE_OFFSET_SHIFT) | offset,
1319 /* Swap a match offset to the front of the queue. */
1320 static inline struct lzx_lru_queue
1321 lzx_lru_queue_swap(struct lzx_lru_queue queue, unsigned idx)
1323 unsigned shift = idx * 21;
1324 const u64 mask = LZX_QUEUE_R0_MASK;
1325 const u64 mask_high = mask << shift;
1327 return (struct lzx_lru_queue) {
1328 (queue.R & ~(mask | mask_high)) |
1329 ((queue.R & mask_high) >> shift) |
1330 ((queue.R & mask) << shift)
1335 lzx_walk_item_list(struct lzx_compressor *c, u32 block_size, bool is_16_bit,
1338 u32 node_idx = block_size;
1339 u32 seq_idx = ARRAY_LEN(c->chosen_sequences) - 1;
1343 /* Special value to mark last sequence */
1344 c->chosen_sequences[seq_idx].adjusted_offset_and_match_hdr = 0x80000000;
1345 lit_start_node = node_idx;
1351 u32 adjusted_offset;
1353 unsigned offset_slot;
1355 /* Tally literals until either a match or the beginning of the
1356 * block is reached. Note: the item in the node at the
1357 * beginning of the block has all bits set, causing this loop to
1358 * end when it is reached. */
1360 item = c->optimum_nodes[node_idx].item;
1361 if (item & OPTIMUM_LEN_MASK)
1363 c->freqs.main[item >> OPTIMUM_OFFSET_SHIFT]++;
1367 #if CONSIDER_GAP_MATCHES
1368 if (item & OPTIMUM_GAP_MATCH) {
1373 /* Record the literal run length for the next sequence
1374 * (the "previous sequence" when walking backwards). */
1375 len = item & OPTIMUM_LEN_MASK;
1377 c->chosen_sequences[seq_idx--].litrunlen =
1378 lit_start_node - node_idx;
1379 lit_start_node = node_idx - len;
1382 /* Tally the rep0 match after the gap. */
1383 v = len - LZX_MIN_MATCH_LEN;
1385 c->chosen_sequences[seq_idx].adjusted_length = v;
1386 if (v >= LZX_NUM_PRIMARY_LENS) {
1387 c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++;
1388 v = LZX_NUM_PRIMARY_LENS;
1390 c->freqs.main[LZX_NUM_CHARS + v]++;
1392 c->chosen_sequences[seq_idx].adjusted_offset_and_match_hdr = v;
1394 /* Tally the literal in the gap. */
1395 c->freqs.main[(u8)(item >> OPTIMUM_OFFSET_SHIFT)]++;
1397 /* Fall through and tally the match before the gap.
1398 * (It was temporarily saved in the 'cost' field of the
1399 * previous node, which was free to reuse.) */
1400 item = c->optimum_nodes[--node_idx].cost;
1403 #else /* CONSIDER_GAP_MATCHES */
1406 #endif /* !CONSIDER_GAP_MATCHES */
1408 len = item & OPTIMUM_LEN_MASK;
1409 adjusted_offset = item >> OPTIMUM_OFFSET_SHIFT;
1411 /* Record the literal run length for the next sequence (the
1412 * "previous sequence" when walking backwards). */
1414 c->chosen_sequences[seq_idx--].litrunlen =
1415 lit_start_node - node_idx;
1417 lit_start_node = node_idx;
1422 /* Record a match. */
1424 /* Tally the aligned offset symbol if needed. */
1425 if (adjusted_offset >= 16)
1426 c->freqs.aligned[adjusted_offset & LZX_ALIGNED_OFFSET_BITMASK]++;
1428 /* Record the adjusted length. */
1429 v = len - LZX_MIN_MATCH_LEN;
1431 c->chosen_sequences[seq_idx].adjusted_length = v;
1433 /* Tally the length symbol if needed. */
1434 if (v >= LZX_NUM_PRIMARY_LENS) {
1435 c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++;
1436 v = LZX_NUM_PRIMARY_LENS;
1439 /* Tally the main symbol. */
1440 offset_slot = lzx_get_offset_slot(c, adjusted_offset, is_16_bit);
1441 v += offset_slot * LZX_NUM_LEN_HEADERS;
1442 c->freqs.main[LZX_NUM_CHARS + v]++;
1444 /* Record the adjusted offset and match header. */
1446 c->chosen_sequences[seq_idx].adjusted_offset_and_match_hdr =
1447 (adjusted_offset << 9) | v;
1451 /* Record the literal run length for the first sequence. */
1453 c->chosen_sequences[seq_idx].litrunlen = lit_start_node - node_idx;
1455 /* Return the index in chosen_sequences at which the sequences begin. */
1460 * Given the minimum-cost path computed through the item graph for the current
1461 * block, walk the path and count how many of each symbol in each Huffman-coded
1462 * alphabet would be required to output the items (matches and literals) along
1465 * Note that the path will be walked backwards (from the end of the block to the
1466 * beginning of the block), but this doesn't matter because this function only
1467 * computes frequencies.
1470 lzx_tally_item_list(struct lzx_compressor *c, u32 block_size, bool is_16_bit)
1472 lzx_walk_item_list(c, block_size, is_16_bit, false);
1476 * Like lzx_tally_item_list(), but this function also generates the list of
1477 * lzx_sequences for the minimum-cost path and writes it to c->chosen_sequences,
1478 * ready to be output to the bitstream after the Huffman codes are computed.
1479 * The lzx_sequences will be written to decreasing memory addresses as the path
1480 * is walked backwards, which means they will end up in the expected
1481 * first-to-last order. The return value is the index in c->chosen_sequences at
1482 * which the lzx_sequences begin.
1485 lzx_record_item_list(struct lzx_compressor *c, u32 block_size, bool is_16_bit)
1487 return lzx_walk_item_list(c, block_size, is_16_bit, true);
1491 * Find an inexpensive path through the graph of possible match/literal choices
1492 * for the current block. The nodes of the graph are
1493 * c->optimum_nodes[0...block_size]. They correspond directly to the bytes in
1494 * the current block, plus one extra node for end-of-block. The edges of the
1495 * graph are matches and literals. The goal is to find the minimum cost path
1496 * from 'c->optimum_nodes[0]' to 'c->optimum_nodes[block_size]', given the cost
1499 * The algorithm works forwards, starting at 'c->optimum_nodes[0]' and
1500 * proceeding forwards one node at a time. At each node, a selection of matches
1501 * (len >= 2), as well as the literal byte (len = 1), is considered. An item of
1502 * length 'len' provides a new path to reach the node 'len' bytes later. If
1503 * such a path is the lowest cost found so far to reach that later node, then
1504 * that later node is updated with the new cost and the "arrival" which provided
1507 * Note that although this algorithm is based on minimum cost path search, due
1508 * to various simplifying assumptions the result is not guaranteed to be the
1509 * true minimum cost, or "optimal", path over the graph of all valid LZX
1510 * representations of this block.
1512 * Also, note that because of the presence of the recent offsets queue (which is
1513 * a type of adaptive state), the algorithm cannot work backwards and compute
1514 * "cost to end" instead of "cost to beginning". Furthermore, the way the
1515 * algorithm handles this adaptive state in the "minimum cost" parse is actually
1516 * only an approximation. It's possible for the globally optimal, minimum cost
1517 * path to contain a prefix, ending at a position, where that path prefix is
1518 * *not* the minimum cost path to that position. This can happen if such a path
1519 * prefix results in a different adaptive state which results in lower costs
1520 * later. The algorithm does not solve this problem in general; it only looks
1521 * one step ahead, with the exception of special consideration for "gap
1524 static inline struct lzx_lru_queue
1525 lzx_find_min_cost_path(struct lzx_compressor * const restrict c,
1526 const u8 * const restrict block_begin,
1527 const u32 block_size,
1528 const struct lzx_lru_queue initial_queue,
1531 struct lzx_optimum_node *cur_node = c->optimum_nodes;
1532 struct lzx_optimum_node * const end_node = cur_node + block_size;
1533 struct lz_match *cache_ptr = c->match_cache;
1534 const u8 *in_next = block_begin;
1535 const u8 * const block_end = block_begin + block_size;
1538 * Instead of storing the match offset LRU queues in the
1539 * 'lzx_optimum_node' structures, we save memory (and cache lines) by
1540 * storing them in a smaller array. This works because the algorithm
1541 * only requires a limited history of the adaptive state. Once a given
1542 * state is more than LZX_MAX_MATCH_LEN bytes behind the current node
1543 * (more if gap match consideration is enabled; we just round up to 512
1544 * so it's a power of 2), it is no longer needed.
1546 * The QUEUE() macro finds the queue for the given node. This macro has
1547 * been optimized by taking advantage of 'struct lzx_lru_queue' and
1548 * 'struct lzx_optimum_node' both being 8 bytes in size and alignment.
1550 struct lzx_lru_queue queues[512];
1551 STATIC_ASSERT(ARRAY_LEN(queues) >= LZX_MAX_MATCH_LEN + 1);
1552 STATIC_ASSERT(sizeof(c->optimum_nodes[0]) == sizeof(queues[0]));
1553 #define QUEUE(node) \
1554 (*(struct lzx_lru_queue *)((char *)queues + \
1555 ((uintptr_t)(node) % (ARRAY_LEN(queues) * sizeof(queues[0])))))
1556 /*(queues[(uintptr_t)(node) / sizeof(*(node)) % ARRAY_LEN(queues)])*/
1558 #if CONSIDER_GAP_MATCHES
1559 u32 matches_before_gap[ARRAY_LEN(queues)];
1560 #define MATCH_BEFORE_GAP(node) \
1561 (matches_before_gap[(uintptr_t)(node) / sizeof(*(node)) % \
1562 ARRAY_LEN(matches_before_gap)])
1566 * Initially, the cost to reach each node is "infinity".
1568 * The first node actually should have cost 0, but "infinity"
1569 * (0xFFFFFFFF) works just as well because it immediately overflows.
1571 * The following statement also intentionally sets the 'item' of the
1572 * first node, which would otherwise have no meaning, to 0xFFFFFFFF for
1573 * use as a sentinel. See lzx_walk_item_list().
1575 memset(c->optimum_nodes, 0xFF,
1576 (block_size + 1) * sizeof(c->optimum_nodes[0]));
1578 /* Initialize the recent offsets queue for the first node. */
1579 QUEUE(cur_node) = initial_queue;
1581 do { /* For each node in the block in position order... */
1583 unsigned num_matches;
1588 * A selection of matches for the block was already saved in
1589 * memory so that we don't have to run the uncompressed data
1590 * through the matchfinder on every optimization pass. However,
1591 * we still search for repeat offset matches during each
1592 * optimization pass because we cannot predict the state of the
1593 * recent offsets queue. But as a heuristic, we don't bother
1594 * searching for repeat offset matches if the general-purpose
1595 * matchfinder failed to find any matches.
1597 * Note that a match of length n at some offset implies there is
1598 * also a match of length l for LZX_MIN_MATCH_LEN <= l <= n at
1599 * that same offset. In other words, we don't necessarily need
1600 * to use the full length of a match. The key heuristic that
1601 * saves a significicant amount of time is that for each
1602 * distinct length, we only consider the smallest offset for
1603 * which that length is available. This heuristic also applies
1604 * to repeat offsets, which we order specially: R0 < R1 < R2 <
1605 * any explicit offset. Of course, this heuristic may be
1606 * produce suboptimal results because offset slots in LZX are
1607 * subject to entropy encoding, but in practice this is a useful
1611 num_matches = cache_ptr->length;
1615 struct lz_match *end_matches = cache_ptr + num_matches;
1616 unsigned next_len = LZX_MIN_MATCH_LEN;
1617 unsigned max_len = min(block_end - in_next, LZX_MAX_MATCH_LEN);
1620 /* Consider rep0 matches. */
1621 matchptr = in_next - lzx_lru_queue_R0(QUEUE(cur_node));
1622 if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1624 STATIC_ASSERT(LZX_MIN_MATCH_LEN == 2);
1626 u32 cost = cur_node->cost +
1627 c->costs.match_cost[0][
1628 next_len - LZX_MIN_MATCH_LEN];
1629 if (cost <= (cur_node + next_len)->cost) {
1630 (cur_node + next_len)->cost = cost;
1631 (cur_node + next_len)->item =
1632 (0 << OPTIMUM_OFFSET_SHIFT) | next_len;
1634 if (unlikely(++next_len > max_len)) {
1635 cache_ptr = end_matches;
1638 } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1642 /* Consider rep1 matches. */
1643 matchptr = in_next - lzx_lru_queue_R1(QUEUE(cur_node));
1644 if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1646 if (matchptr[next_len - 1] != in_next[next_len - 1])
1648 for (unsigned len = 2; len < next_len - 1; len++)
1649 if (matchptr[len] != in_next[len])
1652 u32 cost = cur_node->cost +
1653 c->costs.match_cost[1][
1654 next_len - LZX_MIN_MATCH_LEN];
1655 if (cost <= (cur_node + next_len)->cost) {
1656 (cur_node + next_len)->cost = cost;
1657 (cur_node + next_len)->item =
1658 (1 << OPTIMUM_OFFSET_SHIFT) | next_len;
1660 if (unlikely(++next_len > max_len)) {
1661 cache_ptr = end_matches;
1664 } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1668 /* Consider rep2 matches. */
1669 matchptr = in_next - lzx_lru_queue_R2(QUEUE(cur_node));
1670 if (load_u16_unaligned(matchptr) != load_u16_unaligned(in_next))
1672 if (matchptr[next_len - 1] != in_next[next_len - 1])
1674 for (unsigned len = 2; len < next_len - 1; len++)
1675 if (matchptr[len] != in_next[len])
1678 u32 cost = cur_node->cost +
1679 c->costs.match_cost[2][
1680 next_len - LZX_MIN_MATCH_LEN];
1681 if (cost <= (cur_node + next_len)->cost) {
1682 (cur_node + next_len)->cost = cost;
1683 (cur_node + next_len)->item =
1684 (2 << OPTIMUM_OFFSET_SHIFT) | next_len;
1686 if (unlikely(++next_len > max_len)) {
1687 cache_ptr = end_matches;
1690 } while (in_next[next_len - 1] == matchptr[next_len - 1]);
1694 while (next_len > cache_ptr->length)
1695 if (++cache_ptr == end_matches)
1698 /* Consider explicit offset matches. */
1700 u32 offset = cache_ptr->offset;
1701 u32 adjusted_offset = offset + LZX_OFFSET_ADJUSTMENT;
1702 unsigned offset_slot = lzx_get_offset_slot(c, adjusted_offset, is_16_bit);
1703 u32 base_cost = cur_node->cost;
1706 #if CONSIDER_ALIGNED_COSTS
1707 if (offset >= 16 - LZX_OFFSET_ADJUSTMENT)
1708 base_cost += c->costs.aligned[adjusted_offset &
1709 LZX_ALIGNED_OFFSET_BITMASK];
1713 c->costs.match_cost[offset_slot][
1714 next_len - LZX_MIN_MATCH_LEN];
1715 if (cost < (cur_node + next_len)->cost) {
1716 (cur_node + next_len)->cost = cost;
1717 (cur_node + next_len)->item =
1718 (adjusted_offset << OPTIMUM_OFFSET_SHIFT) | next_len;
1720 } while (++next_len <= cache_ptr->length);
1722 if (++cache_ptr == end_matches) {
1723 #if CONSIDER_GAP_MATCHES
1724 /* Also consider the longest explicit
1725 * offset match as a "gap match": match
1727 s32 remaining = (block_end - in_next) - (s32)next_len;
1728 if (likely(remaining >= 2)) {
1729 const u8 *strptr = in_next + next_len;
1730 const u8 *matchptr = strptr - offset;
1731 if (load_u16_unaligned(strptr) == load_u16_unaligned(matchptr)) {
1732 STATIC_ASSERT(ARRAY_LEN(queues) - LZX_MAX_MATCH_LEN - 2 >= 250);
1733 STATIC_ASSERT(ARRAY_LEN(queues) == ARRAY_LEN(matches_before_gap));
1734 unsigned limit = min(remaining,
1735 min(ARRAY_LEN(queues) - LZX_MAX_MATCH_LEN - 2,
1736 LZX_MAX_MATCH_LEN));
1737 unsigned rep0_len = lz_extend(strptr, matchptr, 2, limit);
1738 u8 lit = strptr[-1];
1739 cost += c->costs.main[lit] +
1740 c->costs.match_cost[0][rep0_len - LZX_MIN_MATCH_LEN];
1741 unsigned total_len = next_len + rep0_len;
1742 if (cost < (cur_node + total_len)->cost) {
1743 (cur_node + total_len)->cost = cost;
1744 (cur_node + total_len)->item =
1746 ((u32)lit << OPTIMUM_OFFSET_SHIFT) |
1748 MATCH_BEFORE_GAP(cur_node + total_len) =
1749 (adjusted_offset << OPTIMUM_OFFSET_SHIFT) |
1754 #endif /* CONSIDER_GAP_MATCHES */
1762 /* Consider coding a literal.
1764 * To avoid an extra branch, actually checking the preferability
1765 * of coding the literal is integrated into the queue update
1767 literal = *in_next++;
1768 cost = cur_node->cost + c->costs.main[literal];
1770 /* Advance to the next position. */
1773 /* The lowest-cost path to the current position is now known.
1774 * Finalize the recent offsets queue that results from taking
1775 * this lowest-cost path. */
1777 if (cost <= cur_node->cost) {
1778 /* Literal: queue remains unchanged. */
1779 cur_node->cost = cost;
1780 cur_node->item = (u32)literal << OPTIMUM_OFFSET_SHIFT;
1781 QUEUE(cur_node) = QUEUE(cur_node - 1);
1783 /* Match: queue update is needed. */
1784 unsigned len = cur_node->item & OPTIMUM_LEN_MASK;
1785 #if CONSIDER_GAP_MATCHES
1786 s32 adjusted_offset = (s32)cur_node->item >> OPTIMUM_OFFSET_SHIFT;
1787 STATIC_ASSERT(OPTIMUM_GAP_MATCH == 0x80000000); /* assuming sign extension */
1789 u32 adjusted_offset = cur_node->item >> OPTIMUM_OFFSET_SHIFT;
1792 if (adjusted_offset >= LZX_NUM_RECENT_OFFSETS) {
1793 /* Explicit offset match: insert offset at front. */
1795 lzx_lru_queue_push(QUEUE(cur_node - len),
1796 adjusted_offset - LZX_OFFSET_ADJUSTMENT);
1798 #if CONSIDER_GAP_MATCHES
1799 else if (adjusted_offset < 0) {
1800 /* "Gap match": Explicit offset match, then a
1801 * literal, then rep0 match. Save the explicit
1802 * offset match information in the cost field of
1803 * the previous node, which isn't needed
1804 * anymore. Then insert the offset at the front
1806 u32 match_before_gap = MATCH_BEFORE_GAP(cur_node);
1807 (cur_node - 1)->cost = match_before_gap;
1809 lzx_lru_queue_push(QUEUE(cur_node - len - 1 -
1810 (match_before_gap & OPTIMUM_LEN_MASK)),
1811 (match_before_gap >> OPTIMUM_OFFSET_SHIFT) -
1812 LZX_OFFSET_ADJUSTMENT);
1816 /* Repeat offset match: swap offset to front. */
1818 lzx_lru_queue_swap(QUEUE(cur_node - len),
1822 } while (cur_node != end_node);
1824 /* Return the recent offsets queue at the end of the path. */
1825 return QUEUE(cur_node);
1829 * Given the costs for the main and length codewords (c->costs.main and
1830 * c->costs.len), initialize the match cost array (c->costs.match_cost) which
1831 * directly provides the cost of every possible (length, offset slot) pair.
1834 lzx_compute_match_costs(struct lzx_compressor *c)
1836 unsigned num_offset_slots = (c->num_main_syms - LZX_NUM_CHARS) /
1837 LZX_NUM_LEN_HEADERS;
1838 struct lzx_costs *costs = &c->costs;
1839 unsigned main_symbol = LZX_NUM_CHARS;
1841 for (unsigned offset_slot = 0; offset_slot < num_offset_slots;
1844 u32 extra_cost = lzx_extra_offset_bits[offset_slot] * BIT_COST;
1847 #if CONSIDER_ALIGNED_COSTS
1848 if (offset_slot >= 8)
1849 extra_cost -= LZX_NUM_ALIGNED_OFFSET_BITS * BIT_COST;
1852 for (i = 0; i < LZX_NUM_PRIMARY_LENS; i++) {
1853 costs->match_cost[offset_slot][i] =
1854 costs->main[main_symbol++] + extra_cost;
1857 extra_cost += costs->main[main_symbol++];
1859 for (; i < LZX_NUM_LENS; i++) {
1860 costs->match_cost[offset_slot][i] =
1861 costs->len[i - LZX_NUM_PRIMARY_LENS] +
1868 * Fast approximation for log2f(x). This is not as accurate as the standard C
1869 * version. It does not need to be perfectly accurate because it is only used
1870 * for estimating symbol costs, which is very approximate anyway.
1880 /* Extract the exponent and subtract 127 to remove the bias. This gives
1881 * the integer part of the result. */
1882 float res = ((u.i >> 23) & 0xFF) - 127;
1884 /* Set the exponent to 0 (plus bias of 127). This transforms the number
1885 * to the range [1, 2) while retaining the same mantissa. */
1886 u.i = (u.i & ~(0xFF << 23)) | (127 << 23);
1889 * Approximate the log2 of the transformed number using a degree 2
1890 * interpolating polynomial for log2(x) over the interval [1, 2). Then
1891 * add this to the extracted exponent to produce the final approximation
1894 * The coefficients of the interpolating polynomial used here were found
1895 * using the script tools/log2_interpolation.r.
1897 return res - 1.653124006f + u.f * (1.9941812f - u.f * 0.3347490189f);
1902 * Return the estimated cost of a symbol which has been estimated to have the
1903 * given probability.
1906 lzx_cost_for_probability(float prob)
1909 * The basic formula is:
1911 * entropy = -log2(probability)
1913 * Use this to get the cost in fractional bits. Then multiply by our
1914 * scaling factor of BIT_COST and truncate to a u32.
1916 * In addition, the minimum cost is BIT_COST (one bit) because the
1917 * entropy coding method will be Huffman codes.
1919 u32 cost = -log2f_fast(prob) * BIT_COST;
1920 return max(cost, BIT_COST);
1924 * Mapping: number of used literals => heuristic probability of a literal times
1925 * 6870. Generated by running this R command:
1927 * cat(paste(round(6870*2^-((304+(0:256))/64)), collapse=", "))
1929 static const u8 literal_scaled_probs[257] = {
1930 255, 253, 250, 247, 244, 242, 239, 237, 234, 232, 229, 227, 224, 222,
1931 219, 217, 215, 212, 210, 208, 206, 203, 201, 199, 197, 195, 193, 191,
1932 189, 186, 184, 182, 181, 179, 177, 175, 173, 171, 169, 167, 166, 164,
1933 162, 160, 159, 157, 155, 153, 152, 150, 149, 147, 145, 144, 142, 141,
1934 139, 138, 136, 135, 133, 132, 130, 129, 128, 126, 125, 124, 122, 121,
1935 120, 118, 117, 116, 115, 113, 112, 111, 110, 109, 107, 106, 105, 104,
1936 103, 102, 101, 100, 98, 97, 96, 95, 94, 93, 92, 91, 90, 89, 88, 87, 86,
1937 86, 85, 84, 83, 82, 81, 80, 79, 78, 78, 77, 76, 75, 74, 73, 73, 72, 71,
1938 70, 70, 69, 68, 67, 67, 66, 65, 65, 64, 63, 62, 62, 61, 60, 60, 59, 59,
1939 58, 57, 57, 56, 55, 55, 54, 54, 53, 53, 52, 51, 51, 50, 50, 49, 49, 48,
1940 48, 47, 47, 46, 46, 45, 45, 44, 44, 43, 43, 42, 42, 41, 41, 40, 40, 40,
1941 39, 39, 38, 38, 38, 37, 37, 36, 36, 36, 35, 35, 34, 34, 34, 33, 33, 33,
1942 32, 32, 32, 31, 31, 31, 30, 30, 30, 29, 29, 29, 28, 28, 28, 27, 27, 27,
1943 27, 26, 26, 26, 25, 25, 25, 25, 24, 24, 24, 24, 23, 23, 23, 23, 22, 22,
1944 22, 22, 21, 21, 21, 21, 20, 20, 20, 20, 20, 19, 19, 19, 19, 19, 18, 18,
1945 18, 18, 18, 17, 17, 17, 17, 17, 16, 16, 16, 16
1949 * Mapping: length symbol => default cost of that symbol. This is derived from
1950 * sample data but has been slightly edited to add more bias towards the
1951 * shortest lengths, which are the most common.
1953 static const u16 lzx_default_len_costs[LZX_LENCODE_NUM_SYMBOLS] = {
1954 300, 310, 320, 330, 360, 396, 399, 416, 451, 448, 463, 466, 505, 492,
1955 503, 514, 547, 531, 566, 561, 589, 563, 592, 586, 623, 602, 639, 627,
1956 659, 643, 657, 650, 685, 662, 661, 672, 685, 686, 696, 680, 657, 682,
1957 666, 699, 674, 699, 679, 709, 688, 712, 692, 714, 694, 716, 698, 712,
1958 706, 727, 714, 727, 713, 723, 712, 718, 719, 719, 720, 735, 725, 735,
1959 728, 740, 727, 739, 727, 742, 716, 733, 733, 740, 738, 746, 737, 747,
1960 738, 745, 736, 748, 742, 749, 745, 749, 743, 748, 741, 752, 745, 752,
1961 747, 750, 747, 752, 748, 753, 750, 752, 753, 753, 749, 744, 752, 755,
1962 753, 756, 745, 748, 746, 745, 723, 757, 755, 758, 755, 758, 752, 757,
1963 754, 757, 755, 759, 755, 758, 753, 755, 755, 758, 757, 761, 755, 750,
1964 758, 759, 759, 760, 758, 751, 757, 757, 759, 759, 758, 759, 758, 761,
1965 750, 761, 758, 760, 759, 761, 758, 761, 760, 752, 759, 760, 759, 759,
1966 757, 762, 760, 761, 761, 748, 761, 760, 762, 763, 752, 762, 762, 763,
1967 762, 762, 763, 763, 762, 763, 762, 763, 762, 763, 763, 764, 763, 762,
1968 763, 762, 762, 762, 764, 764, 763, 764, 763, 763, 763, 762, 763, 763,
1969 762, 764, 764, 763, 762, 763, 763, 763, 763, 762, 764, 763, 762, 764,
1970 764, 763, 763, 765, 764, 764, 762, 763, 764, 765, 763, 764, 763, 764,
1971 762, 764, 764, 754, 763, 764, 763, 763, 762, 763, 584,
1974 /* Set default costs to bootstrap the iterative optimization algorithm. */
1976 lzx_set_default_costs(struct lzx_compressor *c)
1979 u32 num_literals = 0;
1980 u32 num_used_literals = 0;
1981 float inv_num_matches = 1.0f / c->freqs.main[LZX_NUM_CHARS];
1982 float inv_num_items;
1983 float prob_match = 1.0f;
1985 float base_literal_prob;
1987 /* Some numbers here have been hardcoded to assume a bit cost of 64. */
1988 STATIC_ASSERT(BIT_COST == 64);
1990 /* Estimate the number of literals that will used. 'num_literals' is
1991 * the total number, whereas 'num_used_literals' is the number of
1992 * distinct symbols. */
1993 for (i = 0; i < LZX_NUM_CHARS; i++) {
1994 num_literals += c->freqs.main[i];
1995 num_used_literals += (c->freqs.main[i] != 0);
1998 /* Note: all match headers were tallied as symbol 'LZX_NUM_CHARS'. We
1999 * don't attempt to estimate which ones will be used. */
2001 inv_num_items = 1.0f / (num_literals + c->freqs.main[LZX_NUM_CHARS]);
2002 base_literal_prob = literal_scaled_probs[num_used_literals] *
2005 /* Literal costs. We use two different methods to compute the
2006 * probability of each literal and mix together their results. */
2007 for (i = 0; i < LZX_NUM_CHARS; i++) {
2008 u32 freq = c->freqs.main[i];
2010 float prob = 0.5f * ((freq * inv_num_items) +
2012 c->costs.main[i] = lzx_cost_for_probability(prob);
2015 c->costs.main[i] = 11 * BIT_COST;
2019 /* Match header costs. We just assume that all match headers are
2020 * equally probable, but we do take into account the relative cost of a
2021 * match header vs. a literal depending on how common matches are
2022 * expected to be vs. literals. */
2023 prob_match = max(prob_match, 0.15f);
2024 match_cost = lzx_cost_for_probability(prob_match / (c->num_main_syms -
2026 for (; i < c->num_main_syms; i++)
2027 c->costs.main[i] = match_cost;
2029 /* Length symbol costs. These are just set to fixed values which
2030 * reflect the fact the smallest lengths are typically the most common,
2031 * and therefore are typically the cheapest. */
2032 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++)
2033 c->costs.len[i] = lzx_default_len_costs[i];
2035 #if CONSIDER_ALIGNED_COSTS
2036 /* Aligned offset symbol costs. These are derived from the estimated
2037 * probability of each aligned offset symbol. */
2038 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
2039 /* We intentionally tallied the frequencies in the wrong slots,
2040 * not accounting for LZX_OFFSET_ADJUSTMENT, since doing the
2041 * fixup here is faster: a constant 8 subtractions here vs. one
2042 * addition for every match. */
2043 unsigned j = (i - LZX_OFFSET_ADJUSTMENT) & LZX_ALIGNED_OFFSET_BITMASK;
2044 if (c->freqs.aligned[j] != 0) {
2045 float prob = c->freqs.aligned[j] * inv_num_matches;
2046 c->costs.aligned[i] = lzx_cost_for_probability(prob);
2048 c->costs.aligned[i] =
2049 (2 * LZX_NUM_ALIGNED_OFFSET_BITS) * BIT_COST;
2055 /* Update the current cost model to reflect the computed Huffman codes. */
2057 lzx_set_costs_from_codes(struct lzx_compressor *c)
2060 const struct lzx_lens *lens = &c->codes[c->codes_index].lens;
2062 for (i = 0; i < c->num_main_syms; i++) {
2063 c->costs.main[i] = (lens->main[i] ? lens->main[i] :
2064 MAIN_CODEWORD_LIMIT) * BIT_COST;
2067 for (i = 0; i < LZX_LENCODE_NUM_SYMBOLS; i++) {
2068 c->costs.len[i] = (lens->len[i] ? lens->len[i] :
2069 LENGTH_CODEWORD_LIMIT) * BIT_COST;
2072 #if CONSIDER_ALIGNED_COSTS
2073 for (i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
2074 c->costs.aligned[i] = (lens->aligned[i] ? lens->aligned[i] :
2075 ALIGNED_CODEWORD_LIMIT) * BIT_COST;
2081 * Choose a "near-optimal" literal/match sequence to use for the current block,
2082 * then flush the block. Because the cost of each Huffman symbol is unknown
2083 * until the Huffman codes have been built and the Huffman codes themselves
2084 * depend on the symbol frequencies, this uses an iterative optimization
2085 * algorithm to approximate an optimal solution. The first optimization pass
2086 * for the block uses default costs; additional passes use costs derived from
2087 * the Huffman codes computed in the previous pass.
2089 static inline struct lzx_lru_queue
2090 lzx_optimize_and_flush_block(struct lzx_compressor * const restrict c,
2091 struct lzx_output_bitstream * const restrict os,
2092 const u8 * const restrict block_begin,
2093 const u32 block_size,
2094 const struct lzx_lru_queue initial_queue,
2097 unsigned num_passes_remaining = c->num_optim_passes;
2098 struct lzx_lru_queue new_queue;
2101 lzx_set_default_costs(c);
2104 lzx_compute_match_costs(c);
2105 new_queue = lzx_find_min_cost_path(c, block_begin, block_size,
2106 initial_queue, is_16_bit);
2108 if (--num_passes_remaining == 0)
2111 /* At least one optimization pass remains. Update the costs. */
2112 lzx_reset_symbol_frequencies(c);
2113 lzx_tally_item_list(c, block_size, is_16_bit);
2114 lzx_build_huffman_codes(c);
2115 lzx_set_costs_from_codes(c);
2118 /* Done optimizing. Generate the sequence list and flush the block. */
2119 lzx_reset_symbol_frequencies(c);
2120 seq_idx = lzx_record_item_list(c, block_size, is_16_bit);
2121 lzx_flush_block(c, os, block_begin, block_size, seq_idx);
2126 * This is the "near-optimal" LZX compressor.
2128 * For each block, it performs a relatively thorough graph search to find an
2129 * inexpensive (in terms of compressed size) way to output the block.
2131 * Note: there are actually many things this algorithm leaves on the table in
2132 * terms of compression ratio. So although it may be "near-optimal", it is
2133 * certainly not "optimal". The goal is not to produce the optimal compression
2134 * ratio, which for LZX is probably impossible within any practical amount of
2135 * time, but rather to produce a compression ratio significantly better than a
2136 * simpler "greedy" or "lazy" parse while still being relatively fast.
2139 lzx_compress_near_optimal(struct lzx_compressor * restrict c,
2140 const u8 * const restrict in_begin, size_t in_nbytes,
2141 struct lzx_output_bitstream * restrict os,
2144 const u8 * in_next = in_begin;
2145 const u8 * const in_end = in_begin + in_nbytes;
2146 u32 max_len = LZX_MAX_MATCH_LEN;
2147 u32 nice_len = min(c->nice_match_length, max_len);
2148 u32 next_hashes[2] = {0, 0};
2149 struct lzx_lru_queue queue = LZX_QUEUE_INITIALIZER;
2151 /* Initialize the matchfinder. */
2152 CALL_BT_MF(is_16_bit, c, bt_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 lz_match *cache_ptr = c->match_cache;
2161 const u8 *next_search_pos = in_next;
2162 const u8 *next_observation = in_next;
2163 const u8 *next_pause_point =
2164 min(in_next + min(MIN_BLOCK_SIZE,
2165 in_max_block_end - in_next),
2166 in_max_block_end - min(LZX_MAX_MATCH_LEN - 1,
2167 in_max_block_end - in_next));
2169 lzx_init_block_split_stats(&c->split_stats);
2170 lzx_reset_symbol_frequencies(c);
2172 if (in_next >= next_pause_point)
2176 * Run the input buffer through the matchfinder, caching the
2177 * matches, until we decide to end the block.
2179 * For a tighter matchfinding loop, we compute a "pause point",
2180 * which is the next position at which we may need to check
2181 * whether to end the block or to decrease max_len. We then
2182 * only do these extra checks upon reaching the pause point.
2184 resume_matchfinding:
2186 if (in_next >= next_search_pos) {
2187 /* Search for matches at this position. */
2188 struct lz_match *lz_matchptr;
2191 lz_matchptr = CALL_BT_MF(is_16_bit, c,
2192 bt_matchfinder_get_matches,
2197 c->max_search_depth,
2201 cache_ptr->length = lz_matchptr - (cache_ptr + 1);
2202 cache_ptr = lz_matchptr;
2204 /* Accumulate literal/match statistics for block
2205 * splitting and for generating the initial cost
2207 if (in_next >= next_observation) {
2208 best_len = cache_ptr[-1].length;
2209 if (best_len >= 3) {
2210 /* Match (len >= 3) */
2213 * Note: for performance reasons this has
2214 * been simplified significantly:
2216 * - We wait until later to account for
2217 * LZX_OFFSET_ADJUSTMENT.
2218 * - We don't account for repeat offsets.
2219 * - We don't account for different match headers.
2221 c->freqs.aligned[cache_ptr[-1].offset &
2222 LZX_ALIGNED_OFFSET_BITMASK]++;
2223 c->freqs.main[LZX_NUM_CHARS]++;
2225 lzx_observe_match(&c->split_stats, best_len);
2226 next_observation = in_next + best_len;
2229 c->freqs.main[*in_next]++;
2230 lzx_observe_literal(&c->split_stats, *in_next);
2231 next_observation = in_next + 1;
2236 * If there was a very long match found, then
2237 * don't cache any matches for the bytes covered
2238 * by that match. This avoids degenerate
2239 * behavior when compressing highly redundant
2240 * data, where the number of matches can be very
2243 * This heuristic doesn't actually hurt the
2244 * compression ratio *too* much. If there's a
2245 * long match, then the data must be highly
2246 * compressible, so it doesn't matter as much
2249 if (best_len >= nice_len)
2250 next_search_pos = in_next + best_len;
2252 /* Don't search for matches at this position. */
2253 CALL_BT_MF(is_16_bit, c,
2254 bt_matchfinder_skip_position,
2258 c->max_search_depth,
2260 cache_ptr->length = 0;
2263 } while (++in_next < next_pause_point &&
2264 likely(cache_ptr < &c->match_cache[CACHE_LENGTH]));
2268 /* Adjust max_len and nice_len if we're nearing the end of the
2269 * input buffer. In addition, if we are so close to the end of
2270 * the input buffer that there cannot be any more matches, then
2271 * just advance through the last few positions and record no
2273 if (unlikely(max_len > in_end - in_next)) {
2274 max_len = in_end - in_next;
2275 nice_len = min(max_len, nice_len);
2276 if (max_len < BT_MATCHFINDER_REQUIRED_NBYTES) {
2277 while (in_next != in_end) {
2278 cache_ptr->length = 0;
2285 /* End the block if the match cache may overflow. */
2286 if (unlikely(cache_ptr >= &c->match_cache[CACHE_LENGTH]))
2289 /* End the block if the soft maximum size has been reached. */
2290 if (in_next >= in_max_block_end)
2293 /* End the block if the block splitting algorithm thinks this is
2294 * a good place to do so. */
2295 if (c->split_stats.num_new_observations >=
2296 NUM_OBSERVATIONS_PER_BLOCK_CHECK &&
2297 in_max_block_end - in_next >= MIN_BLOCK_SIZE &&
2298 lzx_should_end_block(&c->split_stats))
2301 /* It's not time to end the block yet. Compute the next pause
2302 * point and resume matchfinding. */
2304 min(in_next + min(NUM_OBSERVATIONS_PER_BLOCK_CHECK * 2 -
2305 c->split_stats.num_new_observations,
2306 in_max_block_end - in_next),
2307 in_max_block_end - min(LZX_MAX_MATCH_LEN - 1,
2308 in_max_block_end - in_next));
2309 goto resume_matchfinding;
2312 /* We've decided on a block boundary and cached matches. Now
2313 * choose a match/literal sequence and flush the block. */
2314 queue = lzx_optimize_and_flush_block(c, os, in_block_begin,
2315 in_next - in_block_begin,
2317 } while (in_next != in_end);
2321 lzx_compress_near_optimal_16(struct lzx_compressor *c, const u8 *in,
2322 size_t in_nbytes, struct lzx_output_bitstream *os)
2324 lzx_compress_near_optimal(c, in, in_nbytes, os, true);
2328 lzx_compress_near_optimal_32(struct lzx_compressor *c, const u8 *in,
2329 size_t in_nbytes, struct lzx_output_bitstream *os)
2331 lzx_compress_near_optimal(c, in, in_nbytes, os, false);
2334 /******************************************************************************/
2335 /* Faster ("lazy") compression algorithm */
2336 /*----------------------------------------------------------------------------*/
2339 * Called when the compressor chooses to use a literal. This tallies the
2340 * Huffman symbol for the literal, increments the current literal run length,
2341 * and "observes" the literal for the block split statistics.
2344 lzx_choose_literal(struct lzx_compressor *c, unsigned literal, u32 *litrunlen_p)
2346 lzx_observe_literal(&c->split_stats, literal);
2347 c->freqs.main[literal]++;
2352 * Called when the compressor chooses to use a match. This tallies the Huffman
2353 * symbol(s) for a match, saves the match data and the length of the preceding
2354 * literal run, updates the recent offsets queue, and "observes" the match for
2355 * the block split statistics.
2358 lzx_choose_match(struct lzx_compressor *c, unsigned length, u32 adjusted_offset,
2359 u32 recent_offsets[LZX_NUM_RECENT_OFFSETS], bool is_16_bit,
2360 u32 *litrunlen_p, struct lzx_sequence **next_seq_p)
2362 u32 litrunlen = *litrunlen_p;
2363 struct lzx_sequence *next_seq = *next_seq_p;
2364 unsigned offset_slot;
2367 lzx_observe_match(&c->split_stats, length);
2369 v = length - LZX_MIN_MATCH_LEN;
2371 /* Save the literal run length and adjusted length. */
2372 next_seq->litrunlen = litrunlen;
2373 next_seq->adjusted_length = v;
2375 /* Compute the length header, then tally the length symbol if needed. */
2376 if (v >= LZX_NUM_PRIMARY_LENS) {
2377 c->freqs.len[v - LZX_NUM_PRIMARY_LENS]++;
2378 v = LZX_NUM_PRIMARY_LENS;
2381 /* Compute the offset slot. */
2382 offset_slot = lzx_get_offset_slot(c, adjusted_offset, is_16_bit);
2384 /* Compute the match header. */
2385 v += offset_slot * LZX_NUM_LEN_HEADERS;
2387 /* Save the adjusted offset and match header. */
2388 next_seq->adjusted_offset_and_match_hdr = (adjusted_offset << 9) | v;
2390 /* Tally the main symbol. */
2391 c->freqs.main[LZX_NUM_CHARS + v]++;
2393 /* Update the recent offsets queue. */
2394 if (adjusted_offset < LZX_NUM_RECENT_OFFSETS) {
2395 /* Repeat offset match. */
2396 swap(recent_offsets[0], recent_offsets[adjusted_offset]);
2398 /* Explicit offset match. */
2400 /* Tally the aligned offset symbol if needed. */
2401 if (adjusted_offset >= 16)
2402 c->freqs.aligned[adjusted_offset & LZX_ALIGNED_OFFSET_BITMASK]++;
2404 recent_offsets[2] = recent_offsets[1];
2405 recent_offsets[1] = recent_offsets[0];
2406 recent_offsets[0] = adjusted_offset - LZX_OFFSET_ADJUSTMENT;
2409 /* Reset the literal run length and advance to the next sequence. */
2410 *next_seq_p = next_seq + 1;
2415 * Called when the compressor ends a block. This finshes the last lzx_sequence,
2416 * which is just a literal run with no following match. This literal run might
2420 lzx_finish_sequence(struct lzx_sequence *last_seq, u32 litrunlen)
2422 last_seq->litrunlen = litrunlen;
2424 /* Special value to mark last sequence */
2425 last_seq->adjusted_offset_and_match_hdr = 0x80000000;
2429 * Find the longest repeat offset match with the current position. If a match
2430 * is found, return its length and set *best_rep_idx_ret to the index of its
2431 * offset in @recent_offsets. Otherwise, return 0.
2433 * Don't bother with length 2 matches; consider matches of length >= 3 only.
2434 * Also assume that max_len >= 3.
2437 lzx_find_longest_repeat_offset_match(const u8 * const in_next,
2438 const u32 recent_offsets[],
2439 const unsigned max_len,
2440 unsigned *best_rep_idx_ret)
2442 STATIC_ASSERT(LZX_NUM_RECENT_OFFSETS == 3); /* loop is unrolled */
2444 const u32 seq3 = load_u24_unaligned(in_next);
2446 unsigned best_rep_len = 0;
2447 unsigned best_rep_idx = 0;
2450 /* Check for rep0 match (most recent offset) */
2451 matchptr = in_next - recent_offsets[0];
2452 if (load_u24_unaligned(matchptr) == seq3)
2453 best_rep_len = lz_extend(in_next, matchptr, 3, max_len);
2455 /* Check for rep1 match (second most recent offset) */
2456 matchptr = in_next - recent_offsets[1];
2457 if (load_u24_unaligned(matchptr) == seq3) {
2458 rep_len = lz_extend(in_next, matchptr, 3, max_len);
2459 if (rep_len > best_rep_len) {
2460 best_rep_len = rep_len;
2465 /* Check for rep2 match (third most recent offset) */
2466 matchptr = in_next - recent_offsets[2];
2467 if (load_u24_unaligned(matchptr) == seq3) {
2468 rep_len = lz_extend(in_next, matchptr, 3, max_len);
2469 if (rep_len > best_rep_len) {
2470 best_rep_len = rep_len;
2475 *best_rep_idx_ret = best_rep_idx;
2476 return best_rep_len;
2480 * Fast heuristic scoring for lazy parsing: how "good" is this match?
2481 * This is mainly determined by the length: longer matches are better.
2482 * However, we also give a bonus to close (small offset) matches and to repeat
2483 * offset matches, since those require fewer bits to encode.
2486 static inline unsigned
2487 lzx_explicit_offset_match_score(unsigned len, u32 adjusted_offset)
2489 unsigned score = len;
2491 if (adjusted_offset < 4096)
2493 if (adjusted_offset < 256)
2499 static inline unsigned
2500 lzx_repeat_offset_match_score(unsigned rep_len, unsigned rep_idx)
2506 * This is the "lazy" LZX compressor. The basic idea is that before it chooses
2507 * a match, it checks to see if there's a longer match at the next position. If
2508 * yes, it chooses a literal and continues to the next position. If no, it
2509 * chooses the match.
2511 * Some additional heuristics are used as well. Repeat offset matches are
2512 * considered favorably and sometimes are chosen immediately. In addition, long
2513 * matches (at least "nice_len" bytes) are chosen immediately as well. Finally,
2514 * when we decide whether a match is "better" than another, we take the offset
2515 * into consideration as well as the length.
2518 lzx_compress_lazy(struct lzx_compressor * restrict c,
2519 const u8 * const restrict in_begin, size_t in_nbytes,
2520 struct lzx_output_bitstream * restrict os, bool is_16_bit)
2522 const u8 * in_next = in_begin;
2523 const u8 * const in_end = in_begin + in_nbytes;
2524 unsigned max_len = LZX_MAX_MATCH_LEN;
2525 unsigned nice_len = min(c->nice_match_length, max_len);
2526 STATIC_ASSERT(LZX_NUM_RECENT_OFFSETS == 3);
2527 u32 recent_offsets[LZX_NUM_RECENT_OFFSETS] = {1, 1, 1};
2528 u32 next_hashes[2] = {0, 0};
2530 /* Initialize the matchfinder. */
2531 CALL_HC_MF(is_16_bit, c, hc_matchfinder_init);
2534 /* Starting a new block */
2536 const u8 * const in_block_begin = in_next;
2537 const u8 * const in_max_block_end =
2538 in_next + min(SOFT_MAX_BLOCK_SIZE, in_end - in_next);
2539 struct lzx_sequence *next_seq = c->chosen_sequences;
2543 u32 cur_adjusted_offset;
2547 u32 next_adjusted_offset;
2548 unsigned next_score;
2549 unsigned best_rep_len;
2550 unsigned best_rep_idx;
2554 lzx_reset_symbol_frequencies(c);
2555 lzx_init_block_split_stats(&c->split_stats);
2558 /* Adjust max_len and nice_len if we're nearing the end
2559 * of the input buffer. */
2560 if (unlikely(max_len > in_end - in_next)) {
2561 max_len = in_end - in_next;
2562 nice_len = min(max_len, nice_len);
2565 /* Find the longest match (subject to the
2566 * max_search_depth cutoff parameter) with the current
2567 * position. Don't bother with length 2 matches; only
2568 * look for matches of length >= 3. */
2569 cur_len = CALL_HC_MF(is_16_bit, c,
2570 hc_matchfinder_longest_match,
2576 c->max_search_depth,
2580 /* If there was no match found, or the only match found
2581 * was a distant short match, then choose a literal. */
2584 cur_offset >= 8192 - LZX_OFFSET_ADJUSTMENT &&
2585 cur_offset != recent_offsets[0] &&
2586 cur_offset != recent_offsets[1] &&
2587 cur_offset != recent_offsets[2]))
2589 lzx_choose_literal(c, *in_next, &litrunlen);
2594 /* Heuristic: if this match has the most recent offset,
2595 * then go ahead and choose it as a rep0 match. */
2596 if (cur_offset == recent_offsets[0]) {
2598 skip_len = cur_len - 1;
2599 cur_adjusted_offset = 0;
2600 goto choose_cur_match;
2603 /* Compute the longest match's score as an explicit
2605 cur_adjusted_offset = cur_offset + LZX_OFFSET_ADJUSTMENT;
2606 cur_score = lzx_explicit_offset_match_score(cur_len, cur_adjusted_offset);
2608 /* Find the longest repeat offset match at this
2609 * position. If we find one and it's "better" than the
2610 * explicit offset match we found, then go ahead and
2611 * choose the repeat offset match immediately. */
2612 best_rep_len = lzx_find_longest_repeat_offset_match(in_next,
2618 if (best_rep_len != 0 &&
2619 (rep_score = lzx_repeat_offset_match_score(best_rep_len,
2620 best_rep_idx)) >= cur_score)
2622 cur_len = best_rep_len;
2623 cur_adjusted_offset = best_rep_idx;
2624 skip_len = best_rep_len - 1;
2625 goto choose_cur_match;
2630 * We have a match at the current position. If the
2631 * match is very long, then choose it immediately.
2632 * Otherwise, see if there's a better match at the next
2636 if (cur_len >= nice_len) {
2637 skip_len = cur_len - 1;
2638 goto choose_cur_match;
2641 if (unlikely(max_len > in_end - in_next)) {
2642 max_len = in_end - in_next;
2643 nice_len = min(max_len, nice_len);
2646 next_len = CALL_HC_MF(is_16_bit, c,
2647 hc_matchfinder_longest_match,
2653 c->max_search_depth / 2,
2657 if (next_len <= cur_len - 2) {
2658 /* No potentially better match was found. */
2660 skip_len = cur_len - 2;
2661 goto choose_cur_match;
2664 next_adjusted_offset = next_offset + LZX_OFFSET_ADJUSTMENT;
2665 next_score = lzx_explicit_offset_match_score(next_len, next_adjusted_offset);
2667 best_rep_len = lzx_find_longest_repeat_offset_match(in_next,
2673 if (best_rep_len != 0 &&
2674 (rep_score = lzx_repeat_offset_match_score(best_rep_len,
2675 best_rep_idx)) >= next_score)
2678 if (rep_score > cur_score) {
2679 /* The next match is better, and it's a
2680 * repeat offset match. */
2681 lzx_choose_literal(c, *(in_next - 2),
2683 cur_len = best_rep_len;
2684 cur_adjusted_offset = best_rep_idx;
2685 skip_len = cur_len - 1;
2686 goto choose_cur_match;
2689 if (next_score > cur_score) {
2690 /* The next match is better, and it's an
2691 * explicit offset match. */
2692 lzx_choose_literal(c, *(in_next - 2),
2695 cur_adjusted_offset = next_adjusted_offset;
2696 cur_score = next_score;
2697 goto have_cur_match;
2701 /* The original match was better; choose it. */
2702 skip_len = cur_len - 2;
2705 /* Choose a match and have the matchfinder skip over its
2706 * remaining bytes. */
2707 lzx_choose_match(c, cur_len, cur_adjusted_offset,
2708 recent_offsets, is_16_bit,
2709 &litrunlen, &next_seq);
2710 in_next = CALL_HC_MF(is_16_bit, c,
2711 hc_matchfinder_skip_positions,
2718 /* Keep going until it's time to end the block. */
2719 } while (in_next < in_max_block_end &&
2720 !(c->split_stats.num_new_observations >=
2721 NUM_OBSERVATIONS_PER_BLOCK_CHECK &&
2722 in_next - in_block_begin >= MIN_BLOCK_SIZE &&
2723 in_end - in_next >= MIN_BLOCK_SIZE &&
2724 lzx_should_end_block(&c->split_stats)));
2726 /* Flush the block. */
2727 lzx_finish_sequence(next_seq, litrunlen);
2728 lzx_flush_block(c, os, in_block_begin, in_next - in_block_begin, 0);
2730 /* Keep going until we've reached the end of the input buffer. */
2731 } while (in_next != in_end);
2735 lzx_compress_lazy_16(struct lzx_compressor *c, const u8 *in, size_t in_nbytes,
2736 struct lzx_output_bitstream *os)
2738 lzx_compress_lazy(c, in, in_nbytes, os, true);
2742 lzx_compress_lazy_32(struct lzx_compressor *c, const u8 *in, size_t in_nbytes,
2743 struct lzx_output_bitstream *os)
2745 lzx_compress_lazy(c, in, in_nbytes, os, false);
2748 /******************************************************************************/
2749 /* Compressor operations */
2750 /*----------------------------------------------------------------------------*/
2753 * Generate tables for mapping match offsets (actually, "adjusted" match
2754 * offsets) to offset slots.
2757 lzx_init_offset_slot_tabs(struct lzx_compressor *c)
2759 u32 adjusted_offset = 0;
2763 for (; adjusted_offset < ARRAY_LEN(c->offset_slot_tab_1);
2766 if (adjusted_offset >= lzx_offset_slot_base[slot + 1] +
2767 LZX_OFFSET_ADJUSTMENT)
2769 c->offset_slot_tab_1[adjusted_offset] = slot;
2772 /* slots [30, 49] */
2773 for (; adjusted_offset < LZX_MAX_WINDOW_SIZE;
2774 adjusted_offset += (u32)1 << 14)
2776 if (adjusted_offset >= lzx_offset_slot_base[slot + 1] +
2777 LZX_OFFSET_ADJUSTMENT)
2779 c->offset_slot_tab_2[adjusted_offset >> 14] = slot;
2784 lzx_get_compressor_size(size_t max_bufsize, unsigned compression_level)
2786 if (compression_level <= MAX_FAST_LEVEL) {
2787 if (lzx_is_16_bit(max_bufsize))
2788 return offsetof(struct lzx_compressor, hc_mf_16) +
2789 hc_matchfinder_size_16(max_bufsize);
2791 return offsetof(struct lzx_compressor, hc_mf_32) +
2792 hc_matchfinder_size_32(max_bufsize);
2794 if (lzx_is_16_bit(max_bufsize))
2795 return offsetof(struct lzx_compressor, bt_mf_16) +
2796 bt_matchfinder_size_16(max_bufsize);
2798 return offsetof(struct lzx_compressor, bt_mf_32) +
2799 bt_matchfinder_size_32(max_bufsize);
2803 /* Compute the amount of memory needed to allocate an LZX compressor. */
2805 lzx_get_needed_memory(size_t max_bufsize, unsigned compression_level,
2810 if (max_bufsize > LZX_MAX_WINDOW_SIZE)
2813 size += lzx_get_compressor_size(max_bufsize, compression_level);
2815 size += max_bufsize; /* account for in_buffer */
2819 /* Allocate an LZX compressor. */
2821 lzx_create_compressor(size_t max_bufsize, unsigned compression_level,
2822 bool destructive, void **c_ret)
2824 unsigned window_order;
2825 struct lzx_compressor *c;
2827 /* Validate the maximum buffer size and get the window order from it. */
2828 window_order = lzx_get_window_order(max_bufsize);
2829 if (window_order == 0)
2830 return WIMLIB_ERR_INVALID_PARAM;
2832 /* Allocate the compressor. */
2833 c = MALLOC(lzx_get_compressor_size(max_bufsize, compression_level));
2837 c->window_order = window_order;
2838 c->num_main_syms = lzx_get_num_main_syms(window_order);
2839 c->destructive = destructive;
2841 /* Allocate the buffer for preprocessed data if needed. */
2842 if (!c->destructive) {
2843 c->in_buffer = MALLOC(max_bufsize);
2848 if (compression_level <= MAX_FAST_LEVEL) {
2850 /* Fast compression: Use lazy parsing. */
2851 if (lzx_is_16_bit(max_bufsize))
2852 c->impl = lzx_compress_lazy_16;
2854 c->impl = lzx_compress_lazy_32;
2856 /* Scale max_search_depth and nice_match_length with the
2857 * compression level. */
2858 c->max_search_depth = (60 * compression_level) / 20;
2859 c->nice_match_length = (80 * compression_level) / 20;
2861 /* lzx_compress_lazy() needs max_search_depth >= 2 because it
2862 * halves the max_search_depth when attempting a lazy match, and
2863 * max_search_depth must be at least 1. */
2864 c->max_search_depth = max(c->max_search_depth, 2);
2867 /* Normal / high compression: Use near-optimal parsing. */
2868 if (lzx_is_16_bit(max_bufsize))
2869 c->impl = lzx_compress_near_optimal_16;
2871 c->impl = lzx_compress_near_optimal_32;
2873 /* Scale max_search_depth and nice_match_length with the
2874 * compression level. */
2875 c->max_search_depth = (24 * compression_level) / 50;
2876 c->nice_match_length = (48 * compression_level) / 50;
2878 /* Also scale num_optim_passes with the compression level. But
2879 * the more passes there are, the less they help --- so don't
2880 * add them linearly. */
2881 c->num_optim_passes = 1;
2882 c->num_optim_passes += (compression_level >= 45);
2883 c->num_optim_passes += (compression_level >= 70);
2884 c->num_optim_passes += (compression_level >= 100);
2885 c->num_optim_passes += (compression_level >= 150);
2886 c->num_optim_passes += (compression_level >= 200);
2887 c->num_optim_passes += (compression_level >= 300);
2889 /* max_search_depth must be at least 1. */
2890 c->max_search_depth = max(c->max_search_depth, 1);
2893 /* Prepare the offset => offset slot mapping. */
2894 lzx_init_offset_slot_tabs(c);
2902 return WIMLIB_ERR_NOMEM;
2905 /* Compress a buffer of data. */
2907 lzx_compress(const void *restrict in, size_t in_nbytes,
2908 void *restrict out, size_t out_nbytes_avail, void *restrict _c)
2910 struct lzx_compressor *c = _c;
2911 struct lzx_output_bitstream os;
2914 /* Don't bother trying to compress very small inputs. */
2918 /* If the compressor is in "destructive" mode, then we can directly
2919 * preprocess the input data. Otherwise, we need to copy it into an
2920 * internal buffer first. */
2921 if (!c->destructive) {
2922 memcpy(c->in_buffer, in, in_nbytes);
2926 /* Preprocess the input data. */
2927 lzx_preprocess((void *)in, in_nbytes);
2929 /* Initially, the previous Huffman codeword lengths are all zeroes. */
2931 memset(&c->codes[1].lens, 0, sizeof(struct lzx_lens));
2933 /* Initialize the output bitstream. */
2934 lzx_init_output(&os, out, out_nbytes_avail);
2936 /* Call the compression level-specific compress() function. */
2937 (*c->impl)(c, in, in_nbytes, &os);
2939 /* Flush the output bitstream. */
2940 result = lzx_flush_output(&os);
2942 /* If the data did not compress to less than its original size and we
2943 * preprocessed the original buffer, then postprocess it to restore it
2944 * to its original state. */
2945 if (result == 0 && c->destructive)
2946 lzx_postprocess((void *)in, in_nbytes);
2948 /* Return the number of compressed bytes, or 0 if the input did not
2949 * compress to less than its original size. */
2953 /* Free an LZX compressor. */
2955 lzx_free_compressor(void *_c)
2957 struct lzx_compressor *c = _c;
2959 if (!c->destructive)
2964 const struct compressor_ops lzx_compressor_ops = {
2965 .get_needed_memory = lzx_get_needed_memory,
2966 .create_compressor = lzx_create_compressor,
2967 .compress = lzx_compress,
2968 .free_compressor = lzx_free_compressor,