4 * A compressor for the LZMS compression format.
8 * Copyright (C) 2013, 2014 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/.
32 #include "wimlib/compress_common.h"
33 #include "wimlib/compressor_ops.h"
34 #include "wimlib/endianness.h"
35 #include "wimlib/error.h"
36 #include "wimlib/lcpit_matchfinder.h"
37 #include "wimlib/lz_repsearch.h"
38 #include "wimlib/lzms_common.h"
39 #include "wimlib/unaligned.h"
40 #include "wimlib/util.h"
42 /* Stucture used for writing raw bits as a series of 16-bit little endian coding
43 * units. This starts at the *end* of the compressed data buffer and proceeds
45 struct lzms_output_bitstream {
47 /* Bits that haven't yet been written to the output buffer. */
50 /* Number of bits currently held in @bitbuf. */
53 /* Pointer to one past the next position in the compressed data buffer
54 * at which to output a 16-bit coding unit. */
57 /* Pointer to the beginning of the output buffer. (The "end" when
58 * writing backwards!) */
62 /* Stucture used for range encoding (raw version). This starts at the
63 * *beginning* of the compressed data buffer and proceeds forward. */
64 struct lzms_range_encoder_raw {
66 /* A 33-bit variable that holds the low boundary of the current range.
67 * The 33rd bit is needed to catch carries. */
70 /* Size of the current range. */
73 /* Next 16-bit coding unit to output. */
76 /* Number of 16-bit coding units whose output has been delayed due to
77 * possible carrying. The first such coding unit is @cache; all
78 * subsequent such coding units are 0xffff. */
81 /* Pointer to the beginning of the output buffer. */
84 /* Pointer to the position in the output buffer at which the next coding
85 * unit must be written. */
88 /* Pointer just past the end of the output buffer. */
92 /* Structure used for range encoding. This wraps around `struct
93 * lzms_range_encoder_raw' to use and maintain probability entries. */
94 struct lzms_range_encoder {
96 /* Pointer to the raw range encoder, which has no persistent knowledge
97 * of probabilities. Multiple lzms_range_encoder's share the same
98 * lzms_range_encoder_raw. */
99 struct lzms_range_encoder_raw *rc;
101 /* Bits recently encoded by this range encoder. This is used as an
102 * index into @prob_entries. */
105 /* Bitmask for @state to prevent its value from exceeding the number of
106 * probability entries. */
109 /* Probability entries being used for this range encoder. */
110 struct lzms_probability_entry prob_entries[LZMS_MAX_NUM_STATES];
113 /* Structure used for Huffman encoding. */
114 struct lzms_huffman_encoder {
116 /* Bitstream to write Huffman-encoded symbols and verbatim bits to.
117 * Multiple lzms_huffman_encoder's share the same lzms_output_bitstream.
119 struct lzms_output_bitstream *os;
121 /* Number of symbols that have been written using this code far. Reset
122 * to 0 whenever the code is rebuilt. */
123 u32 num_syms_written;
125 /* When @num_syms_written reaches this number, the Huffman code must be
129 /* Number of symbols in the represented Huffman code. */
132 /* Running totals of symbol frequencies. These are diluted slightly
133 * whenever the code is rebuilt. */
134 u32 sym_freqs[LZMS_MAX_NUM_SYMS];
136 /* The length, in bits, of each symbol in the Huffman code. */
137 u8 lens[LZMS_MAX_NUM_SYMS];
139 /* The codeword of each symbol in the Huffman code. */
140 u32 codewords[LZMS_MAX_NUM_SYMS];
143 /* Internal compression parameters */
144 struct lzms_compressor_params {
145 u32 min_match_length;
146 u32 nice_match_length;
147 u32 optim_array_length;
150 /* State of the LZMS compressor */
151 struct lzms_compressor {
153 /* Internal compression parameters */
154 struct lzms_compressor_params params;
156 /* Data currently being compressed */
160 /* Lempel-Ziv match-finder */
161 struct lcpit_matchfinder mf;
163 /* Temporary space to store found matches */
164 struct lz_match *matches;
166 /* Per-position data for near-optimal parsing */
167 struct lzms_mc_pos_data *optimum;
168 struct lzms_mc_pos_data *optimum_end;
170 /* Raw range encoder which outputs to the beginning of the compressed
171 * data buffer, proceeding forwards */
172 struct lzms_range_encoder_raw rc;
174 /* Bitstream which outputs to the end of the compressed data buffer,
175 * proceeding backwards */
176 struct lzms_output_bitstream os;
179 struct lzms_range_encoder main_range_encoder;
180 struct lzms_range_encoder match_range_encoder;
181 struct lzms_range_encoder lz_match_range_encoder;
182 struct lzms_range_encoder lz_repeat_match_range_encoders[LZMS_NUM_RECENT_OFFSETS - 1];
183 struct lzms_range_encoder delta_match_range_encoder;
184 struct lzms_range_encoder delta_repeat_match_range_encoders[LZMS_NUM_RECENT_OFFSETS - 1];
186 /* Huffman encoders */
187 struct lzms_huffman_encoder literal_encoder;
188 struct lzms_huffman_encoder lz_offset_encoder;
189 struct lzms_huffman_encoder length_encoder;
190 struct lzms_huffman_encoder delta_power_encoder;
191 struct lzms_huffman_encoder delta_offset_encoder;
193 /* Used for preprocessing */
194 s32 last_target_usages[65536];
196 #define LZMS_NUM_FAST_LENGTHS 256
197 /* Table: length => length slot for small lengths */
198 u8 length_slot_fast[LZMS_NUM_FAST_LENGTHS];
200 /* Table: length => current cost for small match lengths */
201 u32 length_cost_fast[LZMS_NUM_FAST_LENGTHS];
203 #define LZMS_NUM_FAST_OFFSETS 32768
204 /* Table: offset => offset slot for small offsets */
205 u8 offset_slot_fast[LZMS_NUM_FAST_OFFSETS];
208 struct lzms_lz_lru_queue {
209 u32 recent_offsets[LZMS_NUM_RECENT_OFFSETS + 1];
215 lzms_init_lz_lru_queue(struct lzms_lz_lru_queue *queue)
217 for (int i = 0; i < LZMS_NUM_RECENT_OFFSETS + 1; i++)
218 queue->recent_offsets[i] = i + 1;
220 queue->prev_offset = 0;
221 queue->upcoming_offset = 0;
225 lzms_update_lz_lru_queue(struct lzms_lz_lru_queue *queue)
227 if (queue->prev_offset != 0) {
228 for (int i = LZMS_NUM_RECENT_OFFSETS - 1; i >= 0; i--)
229 queue->recent_offsets[i + 1] = queue->recent_offsets[i];
230 queue->recent_offsets[0] = queue->prev_offset;
232 queue->prev_offset = queue->upcoming_offset;
236 * Match chooser position data:
238 * An array of these structures is used during the near-optimal match-choosing
239 * algorithm. They correspond to consecutive positions in the window and are
240 * used to keep track of the cost to reach each position, and the match/literal
241 * choices that need to be chosen to reach that position.
243 struct lzms_mc_pos_data {
245 /* The cost, in bits, of the lowest-cost path that has been found to
246 * reach this position. This can change as progressively lower cost
247 * paths are found to reach this position. */
249 #define MC_INFINITE_COST UINT32_MAX
251 /* The match or literal that was taken to reach this position. This can
252 * change as progressively lower cost paths are found to reach this
255 * This variable is divided into two bitfields.
258 * Low bits are 1, high bits are the literal.
260 * Explicit offset matches:
261 * Low bits are the match length, high bits are the offset plus 2.
263 * Repeat offset matches:
264 * Low bits are the match length, high bits are the queue index.
267 #define MC_OFFSET_SHIFT 32
268 #define MC_LEN_MASK (((u64)1 << MC_OFFSET_SHIFT) - 1)
270 /* The LZMS adaptive state that exists at this position. This is filled
271 * in lazily, only after the minimum-cost path to this position is
274 * Note: the way we handle this adaptive state in the "minimum-cost"
275 * parse is actually only an approximation. It's possible for the
276 * globally optimal, minimum cost path to contain a prefix, ending at a
277 * position, where that path prefix is *not* the minimum cost path to
278 * that position. This can happen if such a path prefix results in a
279 * different adaptive state which results in lower costs later. We do
280 * not solve this problem; we only consider the lowest cost to reach
281 * each position, which seems to be an acceptable approximation.
283 * Note: this adaptive state also does not include the probability
284 * entries or current Huffman codewords. Those aren't maintained
285 * per-position and are only updated occassionally. */
286 struct lzms_adaptive_state {
287 struct lzms_lz_lru_queue lru;
291 u8 lz_repeat_match_state[LZMS_NUM_RECENT_OFFSETS - 1];
296 lzms_init_fast_slots(struct lzms_compressor *c)
298 /* Create table mapping small lengths to length slots. */
299 for (unsigned slot = 0, i = 0; i < LZMS_NUM_FAST_LENGTHS; i++) {
300 while (i >= lzms_length_slot_base[slot + 1])
302 c->length_slot_fast[i] = slot;
305 /* Create table mapping small offsets to offset slots. */
306 for (unsigned slot = 0, i = 0; i < LZMS_NUM_FAST_OFFSETS; i++) {
307 while (i >= lzms_offset_slot_base[slot + 1])
309 c->offset_slot_fast[i] = slot;
313 static inline unsigned
314 lzms_get_length_slot_fast(const struct lzms_compressor *c, u32 length)
316 if (likely(length < LZMS_NUM_FAST_LENGTHS))
317 return c->length_slot_fast[length];
319 return lzms_get_length_slot(length);
322 static inline unsigned
323 lzms_get_offset_slot_fast(const struct lzms_compressor *c, u32 offset)
325 if (offset < LZMS_NUM_FAST_OFFSETS)
326 return c->offset_slot_fast[offset];
328 return lzms_get_offset_slot(offset);
331 /* Initialize the output bitstream @os to write backwards to the specified
332 * compressed data buffer @out that is @out_limit 16-bit integers long. */
334 lzms_output_bitstream_init(struct lzms_output_bitstream *os,
335 le16 *out, size_t out_limit)
339 os->next = out + out_limit;
344 * Write some bits, contained in the low @num_bits bits of @bits (ordered from
345 * high-order to low-order), to the output bitstream @os.
347 * @max_num_bits is a compile-time constant that specifies the maximum number of
348 * bits that can ever be written at this call site.
351 lzms_output_bitstream_put_varbits(struct lzms_output_bitstream *os,
352 u32 bits, unsigned num_bits,
353 unsigned max_num_bits)
355 LZMS_ASSERT(num_bits <= 48);
357 /* Add the bits to the bit buffer variable. */
358 os->bitcount += num_bits;
359 os->bitbuf = (os->bitbuf << num_bits) | bits;
361 /* Check whether any coding units need to be written. */
362 while (os->bitcount >= 16) {
366 /* Write a coding unit, unless it would underflow the buffer. */
367 if (os->next != os->begin)
368 put_unaligned_u16_le(os->bitbuf >> os->bitcount, --os->next);
370 /* Optimization for call sites that never write more than 16
372 if (max_num_bits <= 16)
377 /* Flush the output bitstream, ensuring that all bits written to it have been
378 * written to memory. Returns %true if all bits have been output successfully,
379 * or %false if an overrun occurred. */
381 lzms_output_bitstream_flush(struct lzms_output_bitstream *os)
383 if (os->next == os->begin)
386 if (os->bitcount != 0)
387 put_unaligned_u16_le(os->bitbuf << (16 - os->bitcount), --os->next);
392 /* Initialize the range encoder @rc to write forwards to the specified
393 * compressed data buffer @out that is @out_limit 16-bit integers long. */
395 lzms_range_encoder_raw_init(struct lzms_range_encoder_raw *rc,
396 le16 *out, size_t out_limit)
399 rc->range = 0xffffffff;
404 rc->end = out + out_limit;
408 * Attempt to flush bits from the range encoder.
410 * Note: this is based on the public domain code for LZMA written by Igor
411 * Pavlov. The only differences in this function are that in LZMS the bits must
412 * be output in 16-bit coding units instead of 8-bit coding units, and that in
413 * LZMS the first coding unit is not ignored by the decompressor, so the encoder
414 * cannot output a dummy value to that position.
416 * The basic idea is that we're writing bits from @rc->low to the output.
417 * However, due to carrying, the writing of coding units with value 0xffff, as
418 * well as one prior coding unit, must be delayed until it is determined whether
422 lzms_range_encoder_raw_shift_low(struct lzms_range_encoder_raw *rc)
424 if ((u32)(rc->low) < 0xffff0000 ||
425 (u32)(rc->low >> 32) != 0)
427 /* Carry not needed (rc->low < 0xffff0000), or carry occurred
428 * ((rc->low >> 32) != 0, a.k.a. the carry bit is 1). */
430 if (likely(rc->next >= rc->begin)) {
431 if (rc->next != rc->end) {
432 put_unaligned_u16_le(rc->cache +
433 (u16)(rc->low >> 32),
440 } while (--rc->cache_size != 0);
442 rc->cache = (rc->low >> 16) & 0xffff;
445 rc->low = (rc->low & 0xffff) << 16;
449 lzms_range_encoder_raw_normalize(struct lzms_range_encoder_raw *rc)
451 if (rc->range <= 0xffff) {
453 lzms_range_encoder_raw_shift_low(rc);
458 lzms_range_encoder_raw_flush(struct lzms_range_encoder_raw *rc)
460 for (unsigned i = 0; i < 4; i++)
461 lzms_range_encoder_raw_shift_low(rc);
462 return rc->next != rc->end;
465 /* Encode the next bit using the range encoder (raw version).
467 * @prob is the chance out of LZMS_PROBABILITY_MAX that the next bit is 0. */
469 lzms_range_encoder_raw_encode_bit(struct lzms_range_encoder_raw *rc,
472 lzms_range_encoder_raw_normalize(rc);
474 u32 bound = (rc->range >> LZMS_PROBABILITY_BITS) * prob;
483 /* Encode a bit using the specified range encoder. This wraps around
484 * lzms_range_encoder_raw_encode_bit() to handle using and updating the
485 * appropriate state and probability entry. */
487 lzms_range_encode_bit(struct lzms_range_encoder *enc, int bit)
489 struct lzms_probability_entry *prob_entry;
492 /* Load the probability entry corresponding to the current state. */
493 prob_entry = &enc->prob_entries[enc->state];
495 /* Update the state based on the next bit. */
496 enc->state = ((enc->state << 1) | bit) & enc->mask;
498 /* Get the probability that the bit is 0. */
499 prob = lzms_get_probability(prob_entry);
501 /* Update the probability entry. */
502 lzms_update_probability_entry(prob_entry, bit);
504 /* Encode the bit. */
505 lzms_range_encoder_raw_encode_bit(enc->rc, bit, prob);
508 /* Called when an adaptive Huffman code needs to be rebuilt. */
510 lzms_rebuild_huffman_code(struct lzms_huffman_encoder *enc)
512 make_canonical_huffman_code(enc->num_syms,
513 LZMS_MAX_CODEWORD_LEN,
518 /* Dilute the frequencies. */
519 for (unsigned i = 0; i < enc->num_syms; i++) {
520 enc->sym_freqs[i] >>= 1;
521 enc->sym_freqs[i] += 1;
523 enc->num_syms_written = 0;
526 /* Encode a symbol using the specified Huffman encoder. */
528 lzms_huffman_encode_symbol(struct lzms_huffman_encoder *enc, unsigned sym)
530 lzms_output_bitstream_put_varbits(enc->os,
533 LZMS_MAX_CODEWORD_LEN);
534 ++enc->sym_freqs[sym];
535 if (++enc->num_syms_written == enc->rebuild_freq)
536 lzms_rebuild_huffman_code(enc);
540 lzms_update_fast_length_costs(struct lzms_compressor *c);
542 /* Encode a match length. */
544 lzms_encode_length(struct lzms_compressor *c, u32 length)
547 unsigned num_extra_bits;
550 slot = lzms_get_length_slot_fast(c, length);
552 extra_bits = length - lzms_length_slot_base[slot];
553 num_extra_bits = lzms_extra_length_bits[slot];
555 lzms_huffman_encode_symbol(&c->length_encoder, slot);
556 if (c->length_encoder.num_syms_written == 0)
557 lzms_update_fast_length_costs(c);
559 lzms_output_bitstream_put_varbits(c->length_encoder.os,
560 extra_bits, num_extra_bits, 30);
563 /* Encode an LZ match offset. */
565 lzms_encode_lz_offset(struct lzms_compressor *c, u32 offset)
568 unsigned num_extra_bits;
571 slot = lzms_get_offset_slot_fast(c, offset);
573 extra_bits = offset - lzms_offset_slot_base[slot];
574 num_extra_bits = lzms_extra_offset_bits[slot];
576 lzms_huffman_encode_symbol(&c->lz_offset_encoder, slot);
577 lzms_output_bitstream_put_varbits(c->lz_offset_encoder.os,
578 extra_bits, num_extra_bits, 30);
581 /* Encode a literal byte. */
583 lzms_encode_literal(struct lzms_compressor *c, unsigned literal)
585 /* Main bit: 0 = a literal, not a match. */
586 lzms_range_encode_bit(&c->main_range_encoder, 0);
588 /* Encode the literal using the current literal Huffman code. */
589 lzms_huffman_encode_symbol(&c->literal_encoder, literal);
592 /* Encode an LZ repeat offset match. */
594 lzms_encode_lz_repeat_offset_match(struct lzms_compressor *c,
595 u32 length, unsigned rep_index)
599 /* Main bit: 1 = a match, not a literal. */
600 lzms_range_encode_bit(&c->main_range_encoder, 1);
602 /* Match bit: 0 = an LZ match, not a delta match. */
603 lzms_range_encode_bit(&c->match_range_encoder, 0);
605 /* LZ match bit: 1 = repeat offset, not an explicit offset. */
606 lzms_range_encode_bit(&c->lz_match_range_encoder, 1);
608 /* Encode the repeat offset index. A 1 bit is encoded for each index
609 * passed up. This sequence of 1 bits is terminated by a 0 bit, or
610 * automatically when (LZMS_NUM_RECENT_OFFSETS - 1) 1 bits have been
612 for (i = 0; i < rep_index; i++)
613 lzms_range_encode_bit(&c->lz_repeat_match_range_encoders[i], 1);
615 if (i < LZMS_NUM_RECENT_OFFSETS - 1)
616 lzms_range_encode_bit(&c->lz_repeat_match_range_encoders[i], 0);
618 /* Encode the match length. */
619 lzms_encode_length(c, length);
622 /* Encode an LZ explicit offset match. */
624 lzms_encode_lz_explicit_offset_match(struct lzms_compressor *c,
625 u32 length, u32 offset)
627 /* Main bit: 1 = a match, not a literal. */
628 lzms_range_encode_bit(&c->main_range_encoder, 1);
630 /* Match bit: 0 = an LZ match, not a delta match. */
631 lzms_range_encode_bit(&c->match_range_encoder, 0);
633 /* LZ match bit: 0 = explicit offset, not a repeat offset. */
634 lzms_range_encode_bit(&c->lz_match_range_encoder, 0);
636 /* Encode the match offset. */
637 lzms_encode_lz_offset(c, offset);
639 /* Encode the match length. */
640 lzms_encode_length(c, length);
644 lzms_encode_item(struct lzms_compressor *c, u64 mc_item_data)
646 u32 len = mc_item_data & MC_LEN_MASK;
647 u32 offset_data = mc_item_data >> MC_OFFSET_SHIFT;
650 lzms_encode_literal(c, offset_data);
651 else if (offset_data < LZMS_NUM_RECENT_OFFSETS)
652 lzms_encode_lz_repeat_offset_match(c, len, offset_data);
654 lzms_encode_lz_explicit_offset_match(c, len, offset_data - LZMS_OFFSET_OFFSET);
657 /* Encode a list of matches and literals chosen by the parsing algorithm. */
659 lzms_encode_item_list(struct lzms_compressor *c,
660 struct lzms_mc_pos_data *cur_optimum_ptr)
662 struct lzms_mc_pos_data *end_optimum_ptr;
666 /* The list is currently in reverse order (last item to first item).
668 end_optimum_ptr = cur_optimum_ptr;
669 saved_item = cur_optimum_ptr->mc_item_data;
672 cur_optimum_ptr -= item & MC_LEN_MASK;
673 saved_item = cur_optimum_ptr->mc_item_data;
674 cur_optimum_ptr->mc_item_data = item;
675 } while (cur_optimum_ptr != c->optimum);
677 /* Walk the list of items from beginning to end, encoding each item. */
679 lzms_encode_item(c, cur_optimum_ptr->mc_item_data);
680 cur_optimum_ptr += (cur_optimum_ptr->mc_item_data) & MC_LEN_MASK;
681 } while (cur_optimum_ptr != end_optimum_ptr);
684 /* Each bit costs 1 << LZMS_COST_SHIFT units. */
685 #define LZMS_COST_SHIFT 6
687 /*#define LZMS_RC_COSTS_USE_FLOATING_POINT*/
690 lzms_rc_costs[LZMS_PROBABILITY_MAX + 1];
692 #ifdef LZMS_RC_COSTS_USE_FLOATING_POINT
697 lzms_do_init_rc_costs(void)
699 /* Fill in a table that maps range coding probabilities needed to code a
700 * bit X (0 or 1) to the number of bits (scaled by a constant factor, to
701 * handle fractional costs) needed to code that bit X.
703 * Consider the range of the range decoder. To eliminate exactly half
704 * the range (logical probability of 0.5), we need exactly 1 bit. For
705 * lower probabilities we need more bits and for higher probabilities we
706 * need fewer bits. In general, a logical probability of N will
707 * eliminate the proportion 1 - N of the range; this information takes
708 * log2(1 / N) bits to encode.
710 * The below loop is simply calculating this number of bits for each
711 * possible probability allowed by the LZMS compression format, but
712 * without using real numbers. To handle fractional probabilities, each
713 * cost is multiplied by (1 << LZMS_COST_SHIFT). These techniques are
714 * based on those used by LZMA.
716 * Note that in LZMS, a probability x really means x / 64, and 0 / 64 is
717 * really interpreted as 1 / 64 and 64 / 64 is really interpreted as
720 for (u32 i = 0; i <= LZMS_PROBABILITY_MAX; i++) {
725 else if (prob == LZMS_PROBABILITY_MAX)
726 prob = LZMS_PROBABILITY_MAX - 1;
728 #ifdef LZMS_RC_COSTS_USE_FLOATING_POINT
729 lzms_rc_costs[i] = log2((double)LZMS_PROBABILITY_MAX / prob) *
730 (1 << LZMS_COST_SHIFT);
734 for (u32 j = 0; j < LZMS_COST_SHIFT; j++) {
737 while (w >= ((u32)1 << 16)) {
742 lzms_rc_costs[i] = (LZMS_PROBABILITY_BITS << LZMS_COST_SHIFT) -
749 lzms_init_rc_costs(void)
751 static pthread_once_t once = PTHREAD_ONCE_INIT;
753 pthread_once(&once, lzms_do_init_rc_costs);
756 /* Return the cost to range-encode the specified bit from the specified state.*/
758 lzms_rc_bit_cost(const struct lzms_range_encoder *enc, u8 cur_state, int bit)
763 prob_zero = enc->prob_entries[cur_state].num_recent_zero_bits;
766 prob_correct = prob_zero;
768 prob_correct = LZMS_PROBABILITY_MAX - prob_zero;
770 return lzms_rc_costs[prob_correct];
773 /* Return the cost to Huffman-encode the specified symbol. */
775 lzms_huffman_symbol_cost(const struct lzms_huffman_encoder *enc, unsigned sym)
777 return (u32)enc->lens[sym] << LZMS_COST_SHIFT;
780 /* Return the cost to encode the specified literal byte. */
782 lzms_literal_cost(const struct lzms_compressor *c, unsigned literal,
783 const struct lzms_adaptive_state *state)
785 return lzms_rc_bit_cost(&c->main_range_encoder, state->main_state, 0) +
786 lzms_huffman_symbol_cost(&c->literal_encoder, literal);
789 /* Update the table that directly provides the costs for small lengths. */
791 lzms_update_fast_length_costs(struct lzms_compressor *c)
797 for (len = 1; len < LZMS_NUM_FAST_LENGTHS; len++) {
799 while (len >= lzms_length_slot_base[slot + 1]) {
801 cost = (u32)(c->length_encoder.lens[slot] +
802 lzms_extra_length_bits[slot]) << LZMS_COST_SHIFT;
805 c->length_cost_fast[len] = cost;
809 /* Return the cost to encode the specified match length, which must be less than
810 * LZMS_NUM_FAST_LENGTHS. */
812 lzms_fast_length_cost(const struct lzms_compressor *c, u32 length)
814 LZMS_ASSERT(length < LZMS_NUM_FAST_LENGTHS);
815 return c->length_cost_fast[length];
818 /* Return the cost to encode the specified LZ match offset. */
820 lzms_lz_offset_cost(const struct lzms_compressor *c, u32 offset)
822 unsigned slot = lzms_get_offset_slot_fast(c, offset);
824 return (u32)(c->lz_offset_encoder.lens[slot] +
825 lzms_extra_offset_bits[slot]) << LZMS_COST_SHIFT;
829 * Consider coding the match at repeat offset index @rep_idx. Consider each
830 * length from the minimum (2) to the full match length (@rep_len).
833 lzms_consider_lz_repeat_offset_match(const struct lzms_compressor *c,
834 struct lzms_mc_pos_data *cur_optimum_ptr,
835 u32 rep_len, unsigned rep_idx)
842 base_cost = cur_optimum_ptr->cost;
844 base_cost += lzms_rc_bit_cost(&c->main_range_encoder,
845 cur_optimum_ptr->state.main_state, 1);
847 base_cost += lzms_rc_bit_cost(&c->match_range_encoder,
848 cur_optimum_ptr->state.match_state, 0);
850 base_cost += lzms_rc_bit_cost(&c->lz_match_range_encoder,
851 cur_optimum_ptr->state.lz_match_state, 1);
853 for (i = 0; i < rep_idx; i++)
854 base_cost += lzms_rc_bit_cost(&c->lz_repeat_match_range_encoders[i],
855 cur_optimum_ptr->state.lz_repeat_match_state[i], 1);
857 if (i < LZMS_NUM_RECENT_OFFSETS - 1)
858 base_cost += lzms_rc_bit_cost(&c->lz_repeat_match_range_encoders[i],
859 cur_optimum_ptr->state.lz_repeat_match_state[i], 0);
863 cost = base_cost + lzms_fast_length_cost(c, len);
864 if (cost < (cur_optimum_ptr + len)->cost) {
865 (cur_optimum_ptr + len)->mc_item_data =
866 ((u64)rep_idx << MC_OFFSET_SHIFT) | len;
867 (cur_optimum_ptr + len)->cost = cost;
869 } while (++len <= rep_len);
873 * Consider coding each match in @matches as an explicit offset match.
875 * @matches must be sorted by strictly decreasing length. This is guaranteed by
878 * We consider each length from the minimum (2) to the longest
879 * (matches[num_matches - 1].len). For each length, we consider only the
880 * smallest offset for which that length is available. Although this is not
881 * guaranteed to be optimal due to the possibility of a larger offset costing
882 * less than a smaller offset to code, this is a very useful heuristic.
885 lzms_consider_lz_explicit_offset_matches(const struct lzms_compressor *c,
886 struct lzms_mc_pos_data *cur_optimum_ptr,
887 const struct lz_match matches[],
896 base_cost = cur_optimum_ptr->cost;
898 base_cost += lzms_rc_bit_cost(&c->main_range_encoder,
899 cur_optimum_ptr->state.main_state, 1);
901 base_cost += lzms_rc_bit_cost(&c->match_range_encoder,
902 cur_optimum_ptr->state.match_state, 0);
904 base_cost += lzms_rc_bit_cost(&c->lz_match_range_encoder,
905 cur_optimum_ptr->state.lz_match_state, 0);
909 position_cost = base_cost + lzms_lz_offset_cost(c, matches[i].offset);
911 cost = position_cost + lzms_fast_length_cost(c, len);
912 if (cost < (cur_optimum_ptr + len)->cost) {
913 (cur_optimum_ptr + len)->mc_item_data =
914 ((u64)(matches[i].offset + LZMS_OFFSET_OFFSET)
915 << MC_OFFSET_SHIFT) | len;
916 (cur_optimum_ptr + len)->cost = cost;
918 } while (++len <= matches[i].length);
923 lzms_init_adaptive_state(struct lzms_adaptive_state *state)
927 lzms_init_lz_lru_queue(&state->lru);
928 state->main_state = 0;
929 state->match_state = 0;
930 state->lz_match_state = 0;
931 for (i = 0; i < LZMS_NUM_RECENT_OFFSETS - 1; i++)
932 state->lz_repeat_match_state[i] = 0;
936 lzms_update_main_state(struct lzms_adaptive_state *state, int is_match)
938 state->main_state = ((state->main_state << 1) | is_match) % LZMS_NUM_MAIN_STATES;
942 lzms_update_match_state(struct lzms_adaptive_state *state, int is_delta)
944 state->match_state = ((state->match_state << 1) | is_delta) % LZMS_NUM_MATCH_STATES;
948 lzms_update_lz_match_state(struct lzms_adaptive_state *state, int is_repeat_offset)
950 state->lz_match_state = ((state->lz_match_state << 1) | is_repeat_offset) % LZMS_NUM_LZ_MATCH_STATES;
954 lzms_update_lz_repeat_match_state(struct lzms_adaptive_state *state, int rep_idx)
958 for (i = 0; i < rep_idx; i++)
959 state->lz_repeat_match_state[i] =
960 ((state->lz_repeat_match_state[i] << 1) | 1) %
961 LZMS_NUM_LZ_REPEAT_MATCH_STATES;
963 if (i < LZMS_NUM_RECENT_OFFSETS - 1)
964 state->lz_repeat_match_state[i] =
965 ((state->lz_repeat_match_state[i] << 1) | 0) %
966 LZMS_NUM_LZ_REPEAT_MATCH_STATES;
970 * The main near-optimal parsing routine.
972 * Briefly, the algorithm does an approximate minimum-cost path search to find a
973 * "near-optimal" sequence of matches and literals to output, based on the
974 * current cost model. The algorithm steps forward, position by position (byte
975 * by byte), and updates the minimum cost path to reach each later position that
976 * can be reached using a match or literal from the current position. This is
977 * essentially Dijkstra's algorithm in disguise: the graph nodes are positions,
978 * the graph edges are possible matches/literals to code, and the cost of each
979 * edge is the estimated number of bits that will be required to output the
980 * corresponding match or literal. But one difference is that we actually
981 * compute the lowest-cost path in pieces, where each piece is terminated when
982 * there are no choices to be made.
986 * - This does not output any delta matches.
988 * - The costs of literals and matches are estimated using the range encoder
989 * states and the semi-adaptive Huffman codes. Except for range encoding
990 * states, costs are assumed to be constant throughout a single run of the
991 * parsing algorithm, which can parse up to @optim_array_length bytes of data.
992 * This introduces a source of inaccuracy because the probabilities and
993 * Huffman codes can change over this part of the data.
996 lzms_near_optimal_parse(struct lzms_compressor *c)
998 const u8 *window_ptr;
999 const u8 *window_end;
1000 struct lzms_mc_pos_data *cur_optimum_ptr;
1001 struct lzms_mc_pos_data *end_optimum_ptr;
1005 unsigned rep_max_idx;
1012 window_ptr = c->cur_window;
1013 window_end = window_ptr + c->cur_window_size;
1015 lzms_init_adaptive_state(&c->optimum[0].state);
1018 /* Start building a new list of items, which will correspond to the next
1019 * piece of the overall minimum-cost path. */
1021 cur_optimum_ptr = c->optimum;
1022 cur_optimum_ptr->cost = 0;
1023 end_optimum_ptr = cur_optimum_ptr;
1025 /* States should currently be consistent with the encoders. */
1026 LZMS_ASSERT(cur_optimum_ptr->state.main_state == c->main_range_encoder.state);
1027 LZMS_ASSERT(cur_optimum_ptr->state.match_state == c->match_range_encoder.state);
1028 LZMS_ASSERT(cur_optimum_ptr->state.lz_match_state == c->lz_match_range_encoder.state);
1029 for (i = 0; i < LZMS_NUM_RECENT_OFFSETS - 1; i++)
1030 LZMS_ASSERT(cur_optimum_ptr->state.lz_repeat_match_state[i] ==
1031 c->lz_repeat_match_range_encoders[i].state);
1033 if (window_ptr == window_end)
1036 /* The following loop runs once for each per byte in the window, except
1037 * in a couple shortcut cases. */
1040 /* Find explicit offset matches with the current position. */
1041 num_matches = lcpit_matchfinder_get_matches(&c->mf, c->matches);
1044 * Find the longest repeat offset match with the current
1049 * - Only search for repeat offset matches if the
1050 * match-finder already found at least one match.
1052 * - Only consider the longest repeat offset match. It
1053 * seems to be rare for the optimal parse to include a
1054 * repeat offset match that doesn't have the longest
1055 * length (allowing for the possibility that not all
1056 * of that length is actually used).
1058 if (likely(window_ptr - c->cur_window >= LZMS_MAX_INIT_RECENT_OFFSET)) {
1059 BUILD_BUG_ON(LZMS_NUM_RECENT_OFFSETS != 3);
1060 rep_max_len = lz_repsearch3(window_ptr,
1061 window_end - window_ptr,
1062 cur_optimum_ptr->state.lru.recent_offsets,
1069 /* If there's a very long repeat offset match,
1070 * choose it immediately. */
1071 if (rep_max_len >= c->params.nice_match_length) {
1073 lcpit_matchfinder_skip_bytes(&c->mf, rep_max_len - 1);
1074 window_ptr += rep_max_len;
1076 if (cur_optimum_ptr != c->optimum)
1077 lzms_encode_item_list(c, cur_optimum_ptr);
1079 lzms_encode_lz_repeat_offset_match(c, rep_max_len,
1082 c->optimum[0].state = cur_optimum_ptr->state;
1084 lzms_update_main_state(&c->optimum[0].state, 1);
1085 lzms_update_match_state(&c->optimum[0].state, 0);
1086 lzms_update_lz_match_state(&c->optimum[0].state, 1);
1087 lzms_update_lz_repeat_match_state(&c->optimum[0].state,
1090 c->optimum[0].state.lru.upcoming_offset =
1091 c->optimum[0].state.lru.recent_offsets[rep_max_idx];
1093 for (i = rep_max_idx; i < LZMS_NUM_RECENT_OFFSETS; i++)
1094 c->optimum[0].state.lru.recent_offsets[i] =
1095 c->optimum[0].state.lru.recent_offsets[i + 1];
1097 lzms_update_lz_lru_queue(&c->optimum[0].state.lru);
1101 /* If reaching any positions for the first time,
1102 * initialize their costs to "infinity". */
1103 while (end_optimum_ptr < cur_optimum_ptr + rep_max_len)
1104 (++end_optimum_ptr)->cost = MC_INFINITE_COST;
1106 /* Consider coding a repeat offset match. */
1107 lzms_consider_lz_repeat_offset_match(c, cur_optimum_ptr,
1108 rep_max_len, rep_max_idx);
1111 longest_len = c->matches[0].length;
1113 /* If there's a very long explicit offset match, choose
1114 * it immediately. */
1115 if (longest_len >= c->params.nice_match_length) {
1117 u32 offset = c->matches[0].offset;
1119 /* Extend the match as far as possible. (The
1120 * LCP-interval tree matchfinder only reports up
1121 * to the "nice" length.) */
1122 longest_len = lz_extend(window_ptr,
1123 window_ptr - offset,
1125 window_end - window_ptr);
1127 lcpit_matchfinder_skip_bytes(&c->mf, longest_len - 1);
1128 window_ptr += longest_len;
1130 if (cur_optimum_ptr != c->optimum)
1131 lzms_encode_item_list(c, cur_optimum_ptr);
1133 lzms_encode_lz_explicit_offset_match(c, longest_len, offset);
1135 c->optimum[0].state = cur_optimum_ptr->state;
1137 lzms_update_main_state(&c->optimum[0].state, 1);
1138 lzms_update_match_state(&c->optimum[0].state, 0);
1139 lzms_update_lz_match_state(&c->optimum[0].state, 0);
1141 c->optimum[0].state.lru.upcoming_offset = offset;
1143 lzms_update_lz_lru_queue(&c->optimum[0].state.lru);
1147 /* If reaching any positions for the first time,
1148 * initialize their costs to "infinity". */
1149 while (end_optimum_ptr < cur_optimum_ptr + longest_len)
1150 (++end_optimum_ptr)->cost = MC_INFINITE_COST;
1152 /* Consider coding an explicit offset match. */
1153 lzms_consider_lz_explicit_offset_matches(c, cur_optimum_ptr,
1154 c->matches, num_matches);
1156 /* No matches found. The only choice at this position
1157 * is to code a literal. */
1159 if (end_optimum_ptr == cur_optimum_ptr)
1160 (++end_optimum_ptr)->cost = MC_INFINITE_COST;
1163 /* Consider coding a literal.
1165 * To avoid an extra unpredictable brench, actually checking the
1166 * preferability of coding a literal is integrated into the
1167 * adaptive state update code below. */
1168 literal = *window_ptr++;
1169 cost = cur_optimum_ptr->cost +
1170 lzms_literal_cost(c, literal, &cur_optimum_ptr->state);
1172 /* Advance to the next position. */
1175 /* The lowest-cost path to the current position is now known.
1176 * Finalize the adaptive state that results from taking this
1177 * lowest-cost path. */
1179 if (cost < cur_optimum_ptr->cost) {
1181 cur_optimum_ptr->cost = cost;
1182 cur_optimum_ptr->mc_item_data = ((u64)literal << MC_OFFSET_SHIFT) | 1;
1184 cur_optimum_ptr->state = (cur_optimum_ptr - 1)->state;
1186 lzms_update_main_state(&cur_optimum_ptr->state, 0);
1188 cur_optimum_ptr->state.lru.upcoming_offset = 0;
1191 len = cur_optimum_ptr->mc_item_data & MC_LEN_MASK;
1192 offset_data = cur_optimum_ptr->mc_item_data >> MC_OFFSET_SHIFT;
1194 cur_optimum_ptr->state = (cur_optimum_ptr - len)->state;
1196 lzms_update_main_state(&cur_optimum_ptr->state, 1);
1197 lzms_update_match_state(&cur_optimum_ptr->state, 0);
1199 if (offset_data >= LZMS_NUM_RECENT_OFFSETS) {
1201 /* Explicit offset LZ match */
1203 lzms_update_lz_match_state(&cur_optimum_ptr->state, 0);
1205 cur_optimum_ptr->state.lru.upcoming_offset =
1206 offset_data - LZMS_OFFSET_OFFSET;
1208 /* Repeat offset LZ match */
1210 lzms_update_lz_match_state(&cur_optimum_ptr->state, 1);
1211 lzms_update_lz_repeat_match_state(&cur_optimum_ptr->state,
1214 cur_optimum_ptr->state.lru.upcoming_offset =
1215 cur_optimum_ptr->state.lru.recent_offsets[offset_data];
1217 for (i = offset_data; i < LZMS_NUM_RECENT_OFFSETS; i++)
1218 cur_optimum_ptr->state.lru.recent_offsets[i] =
1219 cur_optimum_ptr->state.lru.recent_offsets[i + 1];
1223 lzms_update_lz_lru_queue(&cur_optimum_ptr->state.lru);
1226 * This loop will terminate when either of the following
1227 * conditions is true:
1229 * (1) cur_optimum_ptr == end_optimum_ptr
1231 * There are no paths that extend beyond the current
1232 * position. In this case, any path to a later position
1233 * must pass through the current position, so we can go
1234 * ahead and choose the list of items that led to this
1237 * (2) cur_optimum_ptr == c->optimum_end
1239 * This bounds the number of times the algorithm can step
1240 * forward before it is guaranteed to start choosing items.
1241 * This limits the memory usage. It also guarantees that
1242 * the parser will not go too long without updating the
1243 * probability tables.
1245 * Note: no check for end-of-window is needed because
1246 * end-of-window will trigger condition (1).
1248 if (cur_optimum_ptr == end_optimum_ptr ||
1249 cur_optimum_ptr == c->optimum_end)
1251 c->optimum[0].state = cur_optimum_ptr->state;
1256 /* Output the current list of items that constitute the minimum-cost
1257 * path to the current position. */
1258 lzms_encode_item_list(c, cur_optimum_ptr);
1263 lzms_init_range_encoder(struct lzms_range_encoder *enc,
1264 struct lzms_range_encoder_raw *rc, u32 num_states)
1268 LZMS_ASSERT(is_power_of_2(num_states));
1269 enc->mask = num_states - 1;
1270 lzms_init_probability_entries(enc->prob_entries, num_states);
1274 lzms_init_huffman_encoder(struct lzms_huffman_encoder *enc,
1275 struct lzms_output_bitstream *os,
1277 unsigned rebuild_freq)
1280 enc->num_syms_written = 0;
1281 enc->rebuild_freq = rebuild_freq;
1282 enc->num_syms = num_syms;
1283 for (unsigned i = 0; i < num_syms; i++)
1284 enc->sym_freqs[i] = 1;
1286 make_canonical_huffman_code(enc->num_syms,
1287 LZMS_MAX_CODEWORD_LEN,
1293 /* Prepare the LZMS compressor for compressing a block of data. */
1295 lzms_prepare_compressor(struct lzms_compressor *c, const u8 *udata, u32 ulen,
1296 le16 *cdata, u32 clen16)
1298 unsigned num_offset_slots;
1300 /* Copy the uncompressed data into the @c->cur_window buffer. */
1301 memcpy(c->cur_window, udata, ulen);
1302 c->cur_window_size = ulen;
1304 /* Initialize the raw range encoder (writing forwards). */
1305 lzms_range_encoder_raw_init(&c->rc, cdata, clen16);
1307 /* Initialize the output bitstream for Huffman symbols and verbatim bits
1308 * (writing backwards). */
1309 lzms_output_bitstream_init(&c->os, cdata, clen16);
1311 /* Calculate the number of offset slots required. */
1312 num_offset_slots = lzms_get_offset_slot(ulen - 1) + 1;
1314 /* Initialize a Huffman encoder for each alphabet. */
1315 lzms_init_huffman_encoder(&c->literal_encoder, &c->os,
1316 LZMS_NUM_LITERAL_SYMS,
1317 LZMS_LITERAL_CODE_REBUILD_FREQ);
1319 lzms_init_huffman_encoder(&c->lz_offset_encoder, &c->os,
1321 LZMS_LZ_OFFSET_CODE_REBUILD_FREQ);
1323 lzms_init_huffman_encoder(&c->length_encoder, &c->os,
1324 LZMS_NUM_LENGTH_SYMS,
1325 LZMS_LENGTH_CODE_REBUILD_FREQ);
1327 lzms_init_huffman_encoder(&c->delta_offset_encoder, &c->os,
1329 LZMS_DELTA_OFFSET_CODE_REBUILD_FREQ);
1331 lzms_init_huffman_encoder(&c->delta_power_encoder, &c->os,
1332 LZMS_NUM_DELTA_POWER_SYMS,
1333 LZMS_DELTA_POWER_CODE_REBUILD_FREQ);
1335 /* Initialize range encoders, all of which wrap around the same
1336 * lzms_range_encoder_raw. */
1337 lzms_init_range_encoder(&c->main_range_encoder,
1338 &c->rc, LZMS_NUM_MAIN_STATES);
1340 lzms_init_range_encoder(&c->match_range_encoder,
1341 &c->rc, LZMS_NUM_MATCH_STATES);
1343 lzms_init_range_encoder(&c->lz_match_range_encoder,
1344 &c->rc, LZMS_NUM_LZ_MATCH_STATES);
1346 for (unsigned i = 0; i < ARRAY_LEN(c->lz_repeat_match_range_encoders); i++)
1347 lzms_init_range_encoder(&c->lz_repeat_match_range_encoders[i],
1348 &c->rc, LZMS_NUM_LZ_REPEAT_MATCH_STATES);
1350 lzms_init_range_encoder(&c->delta_match_range_encoder,
1351 &c->rc, LZMS_NUM_DELTA_MATCH_STATES);
1353 for (unsigned i = 0; i < ARRAY_LEN(c->delta_repeat_match_range_encoders); i++)
1354 lzms_init_range_encoder(&c->delta_repeat_match_range_encoders[i],
1355 &c->rc, LZMS_NUM_DELTA_REPEAT_MATCH_STATES);
1357 /* Set initial length costs for lengths < LZMS_NUM_FAST_LENGTHS. */
1358 lzms_update_fast_length_costs(c);
1361 /* Flush the output streams, prepare the final compressed data, and return its
1364 * A return value of 0 indicates that the data could not be compressed to fit in
1365 * the available space. */
1367 lzms_finalize(struct lzms_compressor *c, u8 *cdata, size_t csize_avail)
1369 size_t num_forwards_bytes;
1370 size_t num_backwards_bytes;
1372 /* Flush both the forwards and backwards streams, and make sure they
1373 * didn't cross each other and start overwriting each other's data. */
1374 if (!lzms_output_bitstream_flush(&c->os))
1377 if (!lzms_range_encoder_raw_flush(&c->rc))
1380 if (c->rc.next > c->os.next)
1383 /* Now the compressed buffer contains the data output by the forwards
1384 * bitstream, then empty space, then data output by the backwards
1385 * bitstream. Move the data output by the backwards bitstream to be
1386 * adjacent to the data output by the forward bitstream, and calculate
1387 * the compressed size that this results in. */
1388 num_forwards_bytes = (u8*)c->rc.next - (u8*)cdata;
1389 num_backwards_bytes = ((u8*)cdata + csize_avail) - (u8*)c->os.next;
1391 memmove(cdata + num_forwards_bytes, c->os.next, num_backwards_bytes);
1393 return num_forwards_bytes + num_backwards_bytes;
1396 /* Set internal compression parameters for the specified compression level and
1397 * maximum window size. */
1399 lzms_build_params(unsigned int compression_level,
1400 struct lzms_compressor_params *params)
1402 /* Allow length 2 matches if the compression level is sufficiently high.
1404 if (compression_level >= 45)
1405 params->min_match_length = 2;
1407 params->min_match_length = 3;
1409 /* Scale nice_match_length with the compression level. But to allow an
1410 * optimization on length cost calculations, don't allow
1411 * nice_match_length to exceed LZMS_NUM_FAST_LENGTH. */
1412 params->nice_match_length = ((u64)compression_level * 32) / 50;
1413 if (params->nice_match_length < params->min_match_length)
1414 params->nice_match_length = params->min_match_length;
1415 if (params->nice_match_length > LZMS_NUM_FAST_LENGTHS)
1416 params->nice_match_length = LZMS_NUM_FAST_LENGTHS;
1417 params->optim_array_length = 1024;
1421 lzms_free_compressor(void *_c);
1424 lzms_get_needed_memory(size_t max_block_size, unsigned int compression_level)
1426 struct lzms_compressor_params params;
1429 if (max_block_size > LZMS_MAX_BUFFER_SIZE)
1432 lzms_build_params(compression_level, ¶ms);
1434 size += sizeof(struct lzms_compressor);
1437 size += max_block_size;
1440 size += lcpit_matchfinder_get_needed_memory(max_block_size);
1443 size += (params.nice_match_length - params.min_match_length + 1) *
1444 sizeof(struct lz_match);
1447 size += (params.optim_array_length + params.nice_match_length) *
1448 sizeof(struct lzms_mc_pos_data);
1454 lzms_create_compressor(size_t max_block_size, unsigned int compression_level,
1457 struct lzms_compressor *c;
1458 struct lzms_compressor_params params;
1460 if (max_block_size > LZMS_MAX_BUFFER_SIZE)
1461 return WIMLIB_ERR_INVALID_PARAM;
1463 lzms_build_params(compression_level, ¶ms);
1465 c = CALLOC(1, sizeof(struct lzms_compressor));
1471 c->cur_window = MALLOC(max_block_size);
1475 if (!lcpit_matchfinder_init(&c->mf, max_block_size,
1476 c->params.min_match_length,
1477 c->params.nice_match_length))
1480 c->matches = MALLOC((params.nice_match_length - params.min_match_length + 1) *
1481 sizeof(struct lz_match));
1485 c->optimum = MALLOC((params.optim_array_length +
1486 params.nice_match_length) *
1487 sizeof(struct lzms_mc_pos_data));
1490 c->optimum_end = &c->optimum[params.optim_array_length];
1492 lzms_init_rc_costs();
1494 lzms_init_fast_slots(c);
1500 lzms_free_compressor(c);
1501 return WIMLIB_ERR_NOMEM;
1505 lzms_compress(const void *uncompressed_data, size_t uncompressed_size,
1506 void *compressed_data, size_t compressed_size_avail, void *_c)
1508 struct lzms_compressor *c = _c;
1510 /* Don't bother compressing extremely small inputs. */
1511 if (uncompressed_size < 4)
1514 /* Cap the available compressed size to a 32-bit integer and round it
1515 * down to the nearest multiple of 2. */
1516 if (compressed_size_avail > UINT32_MAX)
1517 compressed_size_avail = UINT32_MAX;
1518 if (compressed_size_avail & 1)
1519 compressed_size_avail--;
1521 /* Initialize the compressor structures. */
1522 lzms_prepare_compressor(c, uncompressed_data, uncompressed_size,
1523 compressed_data, compressed_size_avail / 2);
1525 /* Preprocess the uncompressed data. */
1526 lzms_x86_filter(c->cur_window, c->cur_window_size,
1527 c->last_target_usages, false);
1529 /* Load the window into the match-finder. */
1530 lcpit_matchfinder_load_buffer(&c->mf, c->cur_window, c->cur_window_size);
1532 /* Compute and encode a literal/match sequence that decompresses to the
1533 * preprocessed data. */
1534 lzms_near_optimal_parse(c);
1536 /* Return the compressed data size or 0. */
1537 return lzms_finalize(c, compressed_data, compressed_size_avail);
1541 lzms_free_compressor(void *_c)
1543 struct lzms_compressor *c = _c;
1546 FREE(c->cur_window);
1547 lcpit_matchfinder_destroy(&c->mf);
1554 const struct compressor_ops lzms_compressor_ops = {
1555 .get_needed_memory = lzms_get_needed_memory,
1556 .create_compressor = lzms_create_compressor,
1557 .compress = lzms_compress,
1558 .free_compressor = lzms_free_compressor,