4 * A compressor that produces output compatible with 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/.
28 #include "wimlib/compress_common.h"
29 #include "wimlib/compressor_ops.h"
30 #include "wimlib/endianness.h"
31 #include "wimlib/error.h"
32 #include "wimlib/lz_mf.h"
33 #include "wimlib/lz_repsearch.h"
34 #include "wimlib/lzms.h"
35 #include "wimlib/util.h"
41 /* Stucture used for writing raw bits as a series of 16-bit little endian coding
42 * units. This starts at the *end* of the compressed data buffer and proceeds
44 struct lzms_output_bitstream {
46 /* Bits that haven't yet been written to the output buffer. */
49 /* Number of bits currently held in @bitbuf. */
52 /* Pointer to one past the next position in the compressed data buffer
53 * at which to output a 16-bit coding unit. */
56 /* Pointer to the beginning of the output buffer. (The "end" when
57 * writing backwards!) */
61 /* Stucture used for range encoding (raw version). This starts at the
62 * *beginning* of the compressed data buffer and proceeds forward. */
63 struct lzms_range_encoder_raw {
65 /* A 33-bit variable that holds the low boundary of the current range.
66 * The 33rd bit is needed to catch carries. */
69 /* Size of the current range. */
72 /* Next 16-bit coding unit to output. */
75 /* Number of 16-bit coding units whose output has been delayed due to
76 * possible carrying. The first such coding unit is @cache; all
77 * subsequent such coding units are 0xffff. */
80 /* Pointer to the beginning of the output buffer. */
83 /* Pointer to the position in the output buffer at which the next coding
84 * unit must be written. */
87 /* Pointer just past the end of the output buffer. */
91 /* Structure used for range encoding. This wraps around `struct
92 * lzms_range_encoder_raw' to use and maintain probability entries. */
93 struct lzms_range_encoder {
95 /* Pointer to the raw range encoder, which has no persistent knowledge
96 * of probabilities. Multiple lzms_range_encoder's share the same
97 * lzms_range_encoder_raw. */
98 struct lzms_range_encoder_raw *rc;
100 /* Bits recently encoded by this range encoder. This is used as an
101 * index into @prob_entries. */
104 /* Bitmask for @state to prevent its value from exceeding the number of
105 * probability entries. */
108 /* Probability entries being used for this range encoder. */
109 struct lzms_probability_entry prob_entries[LZMS_MAX_NUM_STATES];
112 /* Structure used for Huffman encoding. */
113 struct lzms_huffman_encoder {
115 /* Bitstream to write Huffman-encoded symbols and verbatim bits to.
116 * Multiple lzms_huffman_encoder's share the same lzms_output_bitstream.
118 struct lzms_output_bitstream *os;
120 /* Number of symbols that have been written using this code far. Reset
121 * to 0 whenever the code is rebuilt. */
122 u32 num_syms_written;
124 /* When @num_syms_written reaches this number, the Huffman code must be
128 /* Number of symbols in the represented Huffman code. */
131 /* Running totals of symbol frequencies. These are diluted slightly
132 * whenever the code is rebuilt. */
133 u32 sym_freqs[LZMS_MAX_NUM_SYMS];
135 /* The length, in bits, of each symbol in the Huffman code. */
136 u8 lens[LZMS_MAX_NUM_SYMS];
138 /* The codeword of each symbol in the Huffman code. */
139 u32 codewords[LZMS_MAX_NUM_SYMS];
142 /* Internal compression parameters */
143 struct lzms_compressor_params {
144 u32 min_match_length;
145 u32 nice_match_length;
146 u32 max_search_depth;
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 */
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];
209 * Match chooser position data:
211 * An array of these structures is used during the near-optimal match-choosing
212 * algorithm. They correspond to consecutive positions in the window and are
213 * used to keep track of the cost to reach each position, and the match/literal
214 * choices that need to be chosen to reach that position.
216 struct lzms_mc_pos_data {
218 /* The cost, in bits, of the lowest-cost path that has been found to
219 * reach this position. This can change as progressively lower cost
220 * paths are found to reach this position. */
222 #define MC_INFINITE_COST UINT32_MAX
224 /* The match or literal that was taken to reach this position. This can
225 * change as progressively lower cost paths are found to reach this
228 * This variable is divided into two bitfields.
231 * Low bits are 1, high bits are the literal.
233 * Explicit offset matches:
234 * Low bits are the match length, high bits are the offset plus 2.
236 * Repeat offset matches:
237 * Low bits are the match length, high bits are the queue index.
240 #define MC_OFFSET_SHIFT 32
241 #define MC_LEN_MASK (((u64)1 << MC_OFFSET_SHIFT) - 1)
243 /* The LZMS adaptive state that exists at this position. This is filled
244 * in lazily, only after the minimum-cost path to this position is
247 * Note: the way we handle this adaptive state in the "minimum-cost"
248 * parse is actually only an approximation. It's possible for the
249 * globally optimal, minimum cost path to contain a prefix, ending at a
250 * position, where that path prefix is *not* the minimum cost path to
251 * that position. This can happen if such a path prefix results in a
252 * different adaptive state which results in lower costs later. We do
253 * not solve this problem; we only consider the lowest cost to reach
254 * each position, which seems to be an acceptable approximation.
256 * Note: this adaptive state also does not include the probability
257 * entries or current Huffman codewords. Those aren't maintained
258 * per-position and are only updated occassionally. */
259 struct lzms_adaptive_state {
260 struct lzms_lz_lru_queues lru;
264 u8 lz_repeat_match_state[LZMS_NUM_RECENT_OFFSETS - 1];
269 lzms_init_fast_slots(struct lzms_compressor *c)
271 /* Create table mapping small lengths to length slots. */
272 for (unsigned slot = 0, i = 0; i < LZMS_NUM_FAST_LENGTHS; i++) {
273 while (i >= lzms_length_slot_base[slot + 1])
275 c->length_slot_fast[i] = slot;
278 /* Create table mapping small offsets to offset slots. */
279 for (unsigned slot = 0, i = 0; i < LZMS_NUM_FAST_OFFSETS; i++) {
280 while (i >= lzms_offset_slot_base[slot + 1])
282 c->offset_slot_fast[i] = slot;
286 static inline unsigned
287 lzms_get_length_slot_fast(const struct lzms_compressor *c, u32 length)
289 if (likely(length < LZMS_NUM_FAST_LENGTHS))
290 return c->length_slot_fast[length];
292 return lzms_get_length_slot(length);
295 static inline unsigned
296 lzms_get_offset_slot_fast(const struct lzms_compressor *c, u32 offset)
298 if (offset < LZMS_NUM_FAST_OFFSETS)
299 return c->offset_slot_fast[offset];
301 return lzms_get_offset_slot(offset);
304 /* Initialize the output bitstream @os to write backwards to the specified
305 * compressed data buffer @out that is @out_limit 16-bit integers long. */
307 lzms_output_bitstream_init(struct lzms_output_bitstream *os,
308 le16 *out, size_t out_limit)
312 os->next = out + out_limit;
317 * Write some bits, contained in the low @num_bits bits of @bits (ordered from
318 * high-order to low-order), to the output bitstream @os.
320 * @max_num_bits is a compile-time constant that specifies the maximum number of
321 * bits that can ever be written at this call site.
324 lzms_output_bitstream_put_varbits(struct lzms_output_bitstream *os,
325 u32 bits, unsigned num_bits,
326 unsigned max_num_bits)
328 LZMS_ASSERT(num_bits <= 48);
330 /* Add the bits to the bit buffer variable. */
331 os->bitcount += num_bits;
332 os->bitbuf = (os->bitbuf << num_bits) | bits;
334 /* Check whether any coding units need to be written. */
335 while (os->bitcount >= 16) {
339 /* Write a coding unit, unless it would underflow the buffer. */
340 if (os->next != os->begin)
341 *--os->next = cpu_to_le16(os->bitbuf >> os->bitcount);
343 /* Optimization for call sites that never write more than 16
345 if (max_num_bits <= 16)
350 /* Use when @num_bits is a compile-time constant. Otherwise use
351 * lzms_output_bitstream_put_bits(). */
353 lzms_output_bitstream_put_bits(struct lzms_output_bitstream *os,
354 u32 bits, unsigned num_bits)
356 lzms_output_bitstream_put_varbits(os, bits, num_bits, num_bits);
359 /* Flush the output bitstream, ensuring that all bits written to it have been
360 * written to memory. Returns %true if all bits have been output successfully,
361 * or %false if an overrun occurred. */
363 lzms_output_bitstream_flush(struct lzms_output_bitstream *os)
365 if (os->next == os->begin)
368 if (os->bitcount != 0)
369 *--os->next = cpu_to_le16(os->bitbuf << (16 - os->bitcount));
374 /* Initialize the range encoder @rc to write forwards to the specified
375 * compressed data buffer @out that is @out_limit 16-bit integers long. */
377 lzms_range_encoder_raw_init(struct lzms_range_encoder_raw *rc,
378 le16 *out, size_t out_limit)
381 rc->range = 0xffffffff;
386 rc->end = out + out_limit;
390 * Attempt to flush bits from the range encoder.
392 * Note: this is based on the public domain code for LZMA written by Igor
393 * Pavlov. The only differences in this function are that in LZMS the bits must
394 * be output in 16-bit coding units instead of 8-bit coding units, and that in
395 * LZMS the first coding unit is not ignored by the decompressor, so the encoder
396 * cannot output a dummy value to that position.
398 * The basic idea is that we're writing bits from @rc->low to the output.
399 * However, due to carrying, the writing of coding units with value 0xffff, as
400 * well as one prior coding unit, must be delayed until it is determined whether
404 lzms_range_encoder_raw_shift_low(struct lzms_range_encoder_raw *rc)
406 if ((u32)(rc->low) < 0xffff0000 ||
407 (u32)(rc->low >> 32) != 0)
409 /* Carry not needed (rc->low < 0xffff0000), or carry occurred
410 * ((rc->low >> 32) != 0, a.k.a. the carry bit is 1). */
412 if (likely(rc->next >= rc->begin)) {
413 if (rc->next != rc->end)
414 *rc->next++ = cpu_to_le16(rc->cache +
415 (u16)(rc->low >> 32));
420 } while (--rc->cache_size != 0);
422 rc->cache = (rc->low >> 16) & 0xffff;
425 rc->low = (rc->low & 0xffff) << 16;
429 lzms_range_encoder_raw_normalize(struct lzms_range_encoder_raw *rc)
431 if (rc->range <= 0xffff) {
433 lzms_range_encoder_raw_shift_low(rc);
438 lzms_range_encoder_raw_flush(struct lzms_range_encoder_raw *rc)
440 for (unsigned i = 0; i < 4; i++)
441 lzms_range_encoder_raw_shift_low(rc);
442 return rc->next != rc->end;
445 /* Encode the next bit using the range encoder (raw version).
447 * @prob is the chance out of LZMS_PROBABILITY_MAX that the next bit is 0. */
449 lzms_range_encoder_raw_encode_bit(struct lzms_range_encoder_raw *rc,
452 lzms_range_encoder_raw_normalize(rc);
454 u32 bound = (rc->range >> LZMS_PROBABILITY_BITS) * prob;
463 /* Encode a bit using the specified range encoder. This wraps around
464 * lzms_range_encoder_raw_encode_bit() to handle using and updating the
465 * appropriate state and probability entry. */
467 lzms_range_encode_bit(struct lzms_range_encoder *enc, int bit)
469 struct lzms_probability_entry *prob_entry;
472 /* Load the probability entry corresponding to the current state. */
473 prob_entry = &enc->prob_entries[enc->state];
475 /* Update the state based on the next bit. */
476 enc->state = ((enc->state << 1) | bit) & enc->mask;
478 /* Get the probability that the bit is 0. */
479 prob = lzms_get_probability(prob_entry);
481 /* Update the probability entry. */
482 lzms_update_probability_entry(prob_entry, bit);
484 /* Encode the bit. */
485 lzms_range_encoder_raw_encode_bit(enc->rc, bit, prob);
488 /* Called when an adaptive Huffman code needs to be rebuilt. */
490 lzms_rebuild_huffman_code(struct lzms_huffman_encoder *enc)
492 make_canonical_huffman_code(enc->num_syms,
493 LZMS_MAX_CODEWORD_LEN,
498 /* Dilute the frequencies. */
499 for (unsigned i = 0; i < enc->num_syms; i++) {
500 enc->sym_freqs[i] >>= 1;
501 enc->sym_freqs[i] += 1;
503 enc->num_syms_written = 0;
506 /* Encode a symbol using the specified Huffman encoder. */
508 lzms_huffman_encode_symbol(struct lzms_huffman_encoder *enc, unsigned sym)
510 lzms_output_bitstream_put_varbits(enc->os,
513 LZMS_MAX_CODEWORD_LEN);
514 ++enc->sym_freqs[sym];
515 if (++enc->num_syms_written == enc->rebuild_freq)
516 lzms_rebuild_huffman_code(enc);
520 lzms_update_fast_length_costs(struct lzms_compressor *c);
522 /* Encode a match length. */
524 lzms_encode_length(struct lzms_compressor *c, u32 length)
527 unsigned num_extra_bits;
530 slot = lzms_get_length_slot_fast(c, length);
532 extra_bits = length - lzms_length_slot_base[slot];
533 num_extra_bits = lzms_extra_length_bits[slot];
535 lzms_huffman_encode_symbol(&c->length_encoder, slot);
536 if (c->length_encoder.num_syms_written == 0)
537 lzms_update_fast_length_costs(c);
539 lzms_output_bitstream_put_varbits(c->length_encoder.os,
540 extra_bits, num_extra_bits, 30);
543 /* Encode an LZ match offset. */
545 lzms_encode_lz_offset(struct lzms_compressor *c, u32 offset)
548 unsigned num_extra_bits;
551 slot = lzms_get_offset_slot_fast(c, offset);
553 extra_bits = offset - lzms_offset_slot_base[slot];
554 num_extra_bits = lzms_extra_offset_bits[slot];
556 lzms_huffman_encode_symbol(&c->lz_offset_encoder, slot);
557 lzms_output_bitstream_put_varbits(c->lz_offset_encoder.os,
558 extra_bits, num_extra_bits, 30);
561 /* Encode a literal byte. */
563 lzms_encode_literal(struct lzms_compressor *c, unsigned literal)
565 /* Main bit: 0 = a literal, not a match. */
566 lzms_range_encode_bit(&c->main_range_encoder, 0);
568 /* Encode the literal using the current literal Huffman code. */
569 lzms_huffman_encode_symbol(&c->literal_encoder, literal);
572 /* Encode an LZ repeat offset match. */
574 lzms_encode_lz_repeat_offset_match(struct lzms_compressor *c,
575 u32 length, unsigned rep_index)
579 /* Main bit: 1 = a match, not a literal. */
580 lzms_range_encode_bit(&c->main_range_encoder, 1);
582 /* Match bit: 0 = an LZ match, not a delta match. */
583 lzms_range_encode_bit(&c->match_range_encoder, 0);
585 /* LZ match bit: 1 = repeat offset, not an explicit offset. */
586 lzms_range_encode_bit(&c->lz_match_range_encoder, 1);
588 /* Encode the repeat offset index. A 1 bit is encoded for each index
589 * passed up. This sequence of 1 bits is terminated by a 0 bit, or
590 * automatically when (LZMS_NUM_RECENT_OFFSETS - 1) 1 bits have been
592 for (i = 0; i < rep_index; i++)
593 lzms_range_encode_bit(&c->lz_repeat_match_range_encoders[i], 1);
595 if (i < LZMS_NUM_RECENT_OFFSETS - 1)
596 lzms_range_encode_bit(&c->lz_repeat_match_range_encoders[i], 0);
598 /* Encode the match length. */
599 lzms_encode_length(c, length);
602 /* Encode an LZ explicit offset match. */
604 lzms_encode_lz_explicit_offset_match(struct lzms_compressor *c,
605 u32 length, u32 offset)
607 /* Main bit: 1 = a match, not a literal. */
608 lzms_range_encode_bit(&c->main_range_encoder, 1);
610 /* Match bit: 0 = an LZ match, not a delta match. */
611 lzms_range_encode_bit(&c->match_range_encoder, 0);
613 /* LZ match bit: 0 = explicit offset, not a repeat offset. */
614 lzms_range_encode_bit(&c->lz_match_range_encoder, 0);
616 /* Encode the match offset. */
617 lzms_encode_lz_offset(c, offset);
619 /* Encode the match length. */
620 lzms_encode_length(c, length);
624 lzms_encode_item(struct lzms_compressor *c, u64 mc_item_data)
626 u32 len = mc_item_data & MC_LEN_MASK;
627 u32 offset_data = mc_item_data >> MC_OFFSET_SHIFT;
630 lzms_encode_literal(c, offset_data);
631 else if (offset_data < LZMS_NUM_RECENT_OFFSETS)
632 lzms_encode_lz_repeat_offset_match(c, len, offset_data);
634 lzms_encode_lz_explicit_offset_match(c, len, offset_data - LZMS_OFFSET_OFFSET);
637 /* Encode a list of matches and literals chosen by the parsing algorithm. */
639 lzms_encode_item_list(struct lzms_compressor *c,
640 struct lzms_mc_pos_data *cur_optimum_ptr)
642 struct lzms_mc_pos_data *end_optimum_ptr;
646 /* The list is currently in reverse order (last item to first item).
648 end_optimum_ptr = cur_optimum_ptr;
649 saved_item = cur_optimum_ptr->mc_item_data;
652 cur_optimum_ptr -= item & MC_LEN_MASK;
653 saved_item = cur_optimum_ptr->mc_item_data;
654 cur_optimum_ptr->mc_item_data = item;
655 } while (cur_optimum_ptr != c->optimum);
657 /* Walk the list of items from beginning to end, encoding each item. */
659 lzms_encode_item(c, cur_optimum_ptr->mc_item_data);
660 cur_optimum_ptr += (cur_optimum_ptr->mc_item_data) & MC_LEN_MASK;
661 } while (cur_optimum_ptr != end_optimum_ptr);
664 /* Each bit costs 1 << LZMS_COST_SHIFT units. */
665 #define LZMS_COST_SHIFT 6
667 /*#define LZMS_RC_COSTS_USE_FLOATING_POINT*/
670 lzms_rc_costs[LZMS_PROBABILITY_MAX + 1];
672 #ifdef LZMS_RC_COSTS_USE_FLOATING_POINT
677 lzms_do_init_rc_costs(void)
679 /* Fill in a table that maps range coding probabilities needed to code a
680 * bit X (0 or 1) to the number of bits (scaled by a constant factor, to
681 * handle fractional costs) needed to code that bit X.
683 * Consider the range of the range decoder. To eliminate exactly half
684 * the range (logical probability of 0.5), we need exactly 1 bit. For
685 * lower probabilities we need more bits and for higher probabilities we
686 * need fewer bits. In general, a logical probability of N will
687 * eliminate the proportion 1 - N of the range; this information takes
688 * log2(1 / N) bits to encode.
690 * The below loop is simply calculating this number of bits for each
691 * possible probability allowed by the LZMS compression format, but
692 * without using real numbers. To handle fractional probabilities, each
693 * cost is multiplied by (1 << LZMS_COST_SHIFT). These techniques are
694 * based on those used by LZMA.
696 * Note that in LZMS, a probability x really means x / 64, and 0 / 64 is
697 * really interpreted as 1 / 64 and 64 / 64 is really interpreted as
700 for (u32 i = 0; i <= LZMS_PROBABILITY_MAX; i++) {
705 else if (prob == LZMS_PROBABILITY_MAX)
706 prob = LZMS_PROBABILITY_MAX - 1;
708 #ifdef LZMS_RC_COSTS_USE_FLOATING_POINT
709 lzms_rc_costs[i] = log2((double)LZMS_PROBABILITY_MAX / prob) *
710 (1 << LZMS_COST_SHIFT);
714 for (u32 j = 0; j < LZMS_COST_SHIFT; j++) {
717 while (w >= ((u32)1 << 16)) {
722 lzms_rc_costs[i] = (LZMS_PROBABILITY_BITS << LZMS_COST_SHIFT) -
729 lzms_init_rc_costs(void)
731 static pthread_once_t once = PTHREAD_ONCE_INIT;
733 pthread_once(&once, lzms_do_init_rc_costs);
736 /* Return the cost to range-encode the specified bit from the specified state.*/
738 lzms_rc_bit_cost(const struct lzms_range_encoder *enc, u8 cur_state, int bit)
743 prob_zero = enc->prob_entries[cur_state].num_recent_zero_bits;
746 prob_correct = prob_zero;
748 prob_correct = LZMS_PROBABILITY_MAX - prob_zero;
750 return lzms_rc_costs[prob_correct];
753 /* Return the cost to Huffman-encode the specified symbol. */
755 lzms_huffman_symbol_cost(const struct lzms_huffman_encoder *enc, unsigned sym)
757 return (u32)enc->lens[sym] << LZMS_COST_SHIFT;
760 /* Return the cost to encode the specified literal byte. */
762 lzms_literal_cost(const struct lzms_compressor *c, unsigned literal,
763 const struct lzms_adaptive_state *state)
765 return lzms_rc_bit_cost(&c->main_range_encoder, state->main_state, 0) +
766 lzms_huffman_symbol_cost(&c->literal_encoder, literal);
769 /* Update the table that directly provides the costs for small lengths. */
771 lzms_update_fast_length_costs(struct lzms_compressor *c)
777 for (len = 1; len < LZMS_NUM_FAST_LENGTHS; len++) {
779 while (len >= lzms_length_slot_base[slot + 1]) {
781 cost = (u32)(c->length_encoder.lens[slot] +
782 lzms_extra_length_bits[slot]) << LZMS_COST_SHIFT;
785 c->length_cost_fast[len] = cost;
789 /* Return the cost to encode the specified match length, which must be less than
790 * LZMS_NUM_FAST_LENGTHS. */
792 lzms_fast_length_cost(const struct lzms_compressor *c, u32 length)
794 LZMS_ASSERT(length < LZMS_NUM_FAST_LENGTHS);
795 return c->length_cost_fast[length];
798 /* Return the cost to encode the specified LZ match offset. */
800 lzms_lz_offset_cost(const struct lzms_compressor *c, u32 offset)
802 unsigned slot = lzms_get_offset_slot_fast(c, offset);
804 return (u32)(c->lz_offset_encoder.lens[slot] +
805 lzms_extra_offset_bits[slot]) << LZMS_COST_SHIFT;
809 * Consider coding the match at repeat offset index @rep_idx. Consider each
810 * length from the minimum (2) to the full match length (@rep_len).
813 lzms_consider_lz_repeat_offset_match(const struct lzms_compressor *c,
814 struct lzms_mc_pos_data *cur_optimum_ptr,
815 u32 rep_len, unsigned rep_idx)
822 base_cost = cur_optimum_ptr->cost;
824 base_cost += lzms_rc_bit_cost(&c->main_range_encoder,
825 cur_optimum_ptr->state.main_state, 1);
827 base_cost += lzms_rc_bit_cost(&c->match_range_encoder,
828 cur_optimum_ptr->state.match_state, 0);
830 base_cost += lzms_rc_bit_cost(&c->lz_match_range_encoder,
831 cur_optimum_ptr->state.lz_match_state, 1);
833 for (i = 0; i < rep_idx; i++)
834 base_cost += lzms_rc_bit_cost(&c->lz_repeat_match_range_encoders[i],
835 cur_optimum_ptr->state.lz_repeat_match_state[i], 1);
837 if (i < LZMS_NUM_RECENT_OFFSETS - 1)
838 base_cost += lzms_rc_bit_cost(&c->lz_repeat_match_range_encoders[i],
839 cur_optimum_ptr->state.lz_repeat_match_state[i], 0);
843 cost = base_cost + lzms_fast_length_cost(c, len);
844 if (cost < (cur_optimum_ptr + len)->cost) {
845 (cur_optimum_ptr + len)->mc_item_data =
846 ((u64)rep_idx << MC_OFFSET_SHIFT) | len;
847 (cur_optimum_ptr + len)->cost = cost;
849 } while (++len <= rep_len);
853 * Consider coding each match in @matches as an explicit offset match.
855 * @matches must be sorted by strictly increasing length and strictly increasing
856 * offset. This is guaranteed by the match-finder.
858 * We consider each length from the minimum (2) to the longest
859 * (matches[num_matches - 1].len). For each length, we consider only the
860 * smallest offset for which that length is available. Although this is not
861 * guaranteed to be optimal due to the possibility of a larger offset costing
862 * less than a smaller offset to code, this is a very useful heuristic.
865 lzms_consider_lz_explicit_offset_matches(const struct lzms_compressor *c,
866 struct lzms_mc_pos_data *cur_optimum_ptr,
867 const struct lz_match matches[],
876 base_cost = cur_optimum_ptr->cost;
878 base_cost += lzms_rc_bit_cost(&c->main_range_encoder,
879 cur_optimum_ptr->state.main_state, 1);
881 base_cost += lzms_rc_bit_cost(&c->match_range_encoder,
882 cur_optimum_ptr->state.match_state, 0);
884 base_cost += lzms_rc_bit_cost(&c->lz_match_range_encoder,
885 cur_optimum_ptr->state.lz_match_state, 0);
889 position_cost = base_cost + lzms_lz_offset_cost(c, matches[i].offset);
891 cost = position_cost + lzms_fast_length_cost(c, len);
892 if (cost < (cur_optimum_ptr + len)->cost) {
893 (cur_optimum_ptr + len)->mc_item_data =
894 ((u64)(matches[i].offset + LZMS_OFFSET_OFFSET)
895 << MC_OFFSET_SHIFT) | len;
896 (cur_optimum_ptr + len)->cost = cost;
898 } while (++len <= matches[i].len);
899 } while (++i != num_matches);
903 lzms_init_adaptive_state(struct lzms_adaptive_state *state)
907 lzms_init_lz_lru_queues(&state->lru);
908 state->main_state = 0;
909 state->match_state = 0;
910 state->lz_match_state = 0;
911 for (i = 0; i < LZMS_NUM_RECENT_OFFSETS - 1; i++)
912 state->lz_repeat_match_state[i] = 0;
916 lzms_update_main_state(struct lzms_adaptive_state *state, int is_match)
918 state->main_state = ((state->main_state << 1) | is_match) % LZMS_NUM_MAIN_STATES;
922 lzms_update_match_state(struct lzms_adaptive_state *state, int is_delta)
924 state->match_state = ((state->match_state << 1) | is_delta) % LZMS_NUM_MATCH_STATES;
928 lzms_update_lz_match_state(struct lzms_adaptive_state *state, int is_repeat_offset)
930 state->lz_match_state = ((state->lz_match_state << 1) | is_repeat_offset) % LZMS_NUM_LZ_MATCH_STATES;
934 lzms_update_lz_repeat_match_state(struct lzms_adaptive_state *state, int rep_idx)
938 for (i = 0; i < rep_idx; i++)
939 state->lz_repeat_match_state[i] =
940 ((state->lz_repeat_match_state[i] << 1) | 1) %
941 LZMS_NUM_LZ_REPEAT_MATCH_STATES;
943 if (i < LZMS_NUM_RECENT_OFFSETS - 1)
944 state->lz_repeat_match_state[i] =
945 ((state->lz_repeat_match_state[i] << 1) | 0) %
946 LZMS_NUM_LZ_REPEAT_MATCH_STATES;
950 * The main near-optimal parsing routine.
952 * Briefly, the algorithm does an approximate minimum-cost path search to find a
953 * "near-optimal" sequence of matches and literals to output, based on the
954 * current cost model. The algorithm steps forward, position by position (byte
955 * by byte), and updates the minimum cost path to reach each later position that
956 * can be reached using a match or literal from the current position. This is
957 * essentially Dijkstra's algorithm in disguise: the graph nodes are positions,
958 * the graph edges are possible matches/literals to code, and the cost of each
959 * edge is the estimated number of bits that will be required to output the
960 * corresponding match or literal. But one difference is that we actually
961 * compute the lowest-cost path in pieces, where each piece is terminated when
962 * there are no choices to be made.
966 * - This does not output any delta matches.
968 * - The costs of literals and matches are estimated using the range encoder
969 * states and the semi-adaptive Huffman codes. Except for range encoding
970 * states, costs are assumed to be constant throughout a single run of the
971 * parsing algorithm, which can parse up to @optim_array_length bytes of data.
972 * This introduces a source of inaccuracy because the probabilities and
973 * Huffman codes can change over this part of the data.
976 lzms_near_optimal_parse(struct lzms_compressor *c)
978 const u8 *window_ptr;
979 const u8 *window_end;
980 struct lzms_mc_pos_data *cur_optimum_ptr;
981 struct lzms_mc_pos_data *end_optimum_ptr;
985 unsigned rep_max_idx;
992 window_ptr = c->cur_window;
993 window_end = window_ptr + c->cur_window_size;
995 lzms_init_adaptive_state(&c->optimum[0].state);
998 /* Start building a new list of items, which will correspond to the next
999 * piece of the overall minimum-cost path. */
1001 cur_optimum_ptr = c->optimum;
1002 cur_optimum_ptr->cost = 0;
1003 end_optimum_ptr = cur_optimum_ptr;
1005 /* States should currently be consistent with the encoders. */
1006 LZMS_ASSERT(cur_optimum_ptr->state.main_state == c->main_range_encoder.state);
1007 LZMS_ASSERT(cur_optimum_ptr->state.match_state == c->match_range_encoder.state);
1008 LZMS_ASSERT(cur_optimum_ptr->state.lz_match_state == c->lz_match_range_encoder.state);
1009 for (i = 0; i < LZMS_NUM_RECENT_OFFSETS - 1; i++)
1010 LZMS_ASSERT(cur_optimum_ptr->state.lz_repeat_match_state[i] ==
1011 c->lz_repeat_match_range_encoders[i].state);
1013 if (window_ptr == window_end)
1016 /* The following loop runs once for each per byte in the window, except
1017 * in a couple shortcut cases. */
1020 /* Find explicit offset matches with the current position. */
1021 num_matches = lz_mf_get_matches(c->mf, c->matches);
1025 * Find the longest repeat offset match with the current
1030 * - Only search for repeat offset matches if the
1031 * match-finder already found at least one match.
1033 * - Only consider the longest repeat offset match. It
1034 * seems to be rare for the optimal parse to include a
1035 * repeat offset match that doesn't have the longest
1036 * length (allowing for the possibility that not all
1037 * of that length is actually used).
1039 if (likely(window_ptr - c->cur_window >= LZMS_MAX_INIT_RECENT_OFFSET)) {
1040 BUILD_BUG_ON(LZMS_NUM_RECENT_OFFSETS != 3);
1041 rep_max_len = lz_repsearch3(window_ptr,
1042 window_end - window_ptr,
1043 cur_optimum_ptr->state.lru.recent_offsets,
1050 /* If there's a very long repeat offset match,
1051 * choose it immediately. */
1052 if (rep_max_len >= c->params.nice_match_length) {
1054 lz_mf_skip_positions(c->mf, rep_max_len - 1);
1055 window_ptr += rep_max_len;
1057 if (cur_optimum_ptr != c->optimum)
1058 lzms_encode_item_list(c, cur_optimum_ptr);
1060 lzms_encode_lz_repeat_offset_match(c, rep_max_len,
1063 c->optimum[0].state = cur_optimum_ptr->state;
1065 lzms_update_main_state(&c->optimum[0].state, 1);
1066 lzms_update_match_state(&c->optimum[0].state, 0);
1067 lzms_update_lz_match_state(&c->optimum[0].state, 1);
1068 lzms_update_lz_repeat_match_state(&c->optimum[0].state,
1071 c->optimum[0].state.lru.upcoming_offset =
1072 c->optimum[0].state.lru.recent_offsets[rep_max_idx];
1074 for (i = rep_max_idx; i < LZMS_NUM_RECENT_OFFSETS; i++)
1075 c->optimum[0].state.lru.recent_offsets[i] =
1076 c->optimum[0].state.lru.recent_offsets[i + 1];
1078 lzms_update_lz_lru_queue(&c->optimum[0].state.lru);
1082 /* If reaching any positions for the first time,
1083 * initialize their costs to "infinity". */
1084 while (end_optimum_ptr < cur_optimum_ptr + rep_max_len)
1085 (++end_optimum_ptr)->cost = MC_INFINITE_COST;
1087 /* Consider coding a repeat offset match. */
1088 lzms_consider_lz_repeat_offset_match(c, cur_optimum_ptr,
1089 rep_max_len, rep_max_idx);
1092 longest_len = c->matches[num_matches - 1].len;
1094 /* If there's a very long explicit offset match, choose
1095 * it immediately. */
1096 if (longest_len >= c->params.nice_match_length) {
1098 lz_mf_skip_positions(c->mf, longest_len - 1);
1099 window_ptr += longest_len;
1101 if (cur_optimum_ptr != c->optimum)
1102 lzms_encode_item_list(c, cur_optimum_ptr);
1104 lzms_encode_lz_explicit_offset_match(c, longest_len,
1105 c->matches[num_matches - 1].offset);
1107 c->optimum[0].state = cur_optimum_ptr->state;
1109 lzms_update_main_state(&c->optimum[0].state, 1);
1110 lzms_update_match_state(&c->optimum[0].state, 0);
1111 lzms_update_lz_match_state(&c->optimum[0].state, 0);
1113 c->optimum[0].state.lru.upcoming_offset =
1114 c->matches[num_matches - 1].offset;
1116 lzms_update_lz_lru_queue(&c->optimum[0].state.lru);
1120 /* If reaching any positions for the first time,
1121 * initialize their costs to "infinity". */
1122 while (end_optimum_ptr < cur_optimum_ptr + longest_len)
1123 (++end_optimum_ptr)->cost = MC_INFINITE_COST;
1125 /* Consider coding an explicit offset match. */
1126 lzms_consider_lz_explicit_offset_matches(c, cur_optimum_ptr,
1127 c->matches, num_matches);
1129 /* No matches found. The only choice at this position
1130 * is to code a literal. */
1132 if (end_optimum_ptr == cur_optimum_ptr)
1133 (++end_optimum_ptr)->cost = MC_INFINITE_COST;
1136 /* Consider coding a literal.
1138 * To avoid an extra unpredictable brench, actually checking the
1139 * preferability of coding a literal is integrated into the
1140 * adaptive state update code below. */
1141 literal = *window_ptr++;
1142 cost = cur_optimum_ptr->cost +
1143 lzms_literal_cost(c, literal, &cur_optimum_ptr->state);
1145 /* Advance to the next position. */
1148 /* The lowest-cost path to the current position is now known.
1149 * Finalize the adaptive state that results from taking this
1150 * lowest-cost path. */
1152 if (cost < cur_optimum_ptr->cost) {
1154 cur_optimum_ptr->cost = cost;
1155 cur_optimum_ptr->mc_item_data = ((u64)literal << MC_OFFSET_SHIFT) | 1;
1157 cur_optimum_ptr->state = (cur_optimum_ptr - 1)->state;
1159 lzms_update_main_state(&cur_optimum_ptr->state, 0);
1161 cur_optimum_ptr->state.lru.upcoming_offset = 0;
1164 len = cur_optimum_ptr->mc_item_data & MC_LEN_MASK;
1165 offset_data = cur_optimum_ptr->mc_item_data >> MC_OFFSET_SHIFT;
1167 cur_optimum_ptr->state = (cur_optimum_ptr - len)->state;
1169 lzms_update_main_state(&cur_optimum_ptr->state, 1);
1170 lzms_update_match_state(&cur_optimum_ptr->state, 0);
1172 if (offset_data >= LZMS_NUM_RECENT_OFFSETS) {
1174 /* Explicit offset LZ match */
1176 lzms_update_lz_match_state(&cur_optimum_ptr->state, 0);
1178 cur_optimum_ptr->state.lru.upcoming_offset =
1179 offset_data - LZMS_OFFSET_OFFSET;
1181 /* Repeat offset LZ match */
1183 lzms_update_lz_match_state(&cur_optimum_ptr->state, 1);
1184 lzms_update_lz_repeat_match_state(&cur_optimum_ptr->state,
1187 cur_optimum_ptr->state.lru.upcoming_offset =
1188 cur_optimum_ptr->state.lru.recent_offsets[offset_data];
1190 for (i = offset_data; i < LZMS_NUM_RECENT_OFFSETS; i++)
1191 cur_optimum_ptr->state.lru.recent_offsets[i] =
1192 cur_optimum_ptr->state.lru.recent_offsets[i + 1];
1196 lzms_update_lz_lru_queue(&cur_optimum_ptr->state.lru);
1199 * This loop will terminate when either of the following
1200 * conditions is true:
1202 * (1) cur_optimum_ptr == end_optimum_ptr
1204 * There are no paths that extend beyond the current
1205 * position. In this case, any path to a later position
1206 * must pass through the current position, so we can go
1207 * ahead and choose the list of items that led to this
1210 * (2) cur_optimum_ptr == c->optimum_end
1212 * This bounds the number of times the algorithm can step
1213 * forward before it is guaranteed to start choosing items.
1214 * This limits the memory usage. It also guarantees that
1215 * the parser will not go too long without updating the
1216 * probability tables.
1218 * Note: no check for end-of-window is needed because
1219 * end-of-window will trigger condition (1).
1221 if (cur_optimum_ptr == end_optimum_ptr ||
1222 cur_optimum_ptr == c->optimum_end)
1224 c->optimum[0].state = cur_optimum_ptr->state;
1229 /* Output the current list of items that constitute the minimum-cost
1230 * path to the current position. */
1231 lzms_encode_item_list(c, cur_optimum_ptr);
1236 lzms_init_range_encoder(struct lzms_range_encoder *enc,
1237 struct lzms_range_encoder_raw *rc, u32 num_states)
1241 LZMS_ASSERT(is_power_of_2(num_states));
1242 enc->mask = num_states - 1;
1243 for (u32 i = 0; i < num_states; i++) {
1244 enc->prob_entries[i].num_recent_zero_bits = LZMS_INITIAL_PROBABILITY;
1245 enc->prob_entries[i].recent_bits = LZMS_INITIAL_RECENT_BITS;
1250 lzms_init_huffman_encoder(struct lzms_huffman_encoder *enc,
1251 struct lzms_output_bitstream *os,
1253 unsigned rebuild_freq)
1256 enc->num_syms_written = 0;
1257 enc->rebuild_freq = rebuild_freq;
1258 enc->num_syms = num_syms;
1259 for (unsigned i = 0; i < num_syms; i++)
1260 enc->sym_freqs[i] = 1;
1262 make_canonical_huffman_code(enc->num_syms,
1263 LZMS_MAX_CODEWORD_LEN,
1269 /* Prepare the LZMS compressor for compressing a block of data. */
1271 lzms_prepare_compressor(struct lzms_compressor *c, const u8 *udata, u32 ulen,
1272 le16 *cdata, u32 clen16)
1274 unsigned num_offset_slots;
1276 /* Copy the uncompressed data into the @c->cur_window buffer. */
1277 memcpy(c->cur_window, udata, ulen);
1278 c->cur_window_size = ulen;
1280 /* Initialize the raw range encoder (writing forwards). */
1281 lzms_range_encoder_raw_init(&c->rc, cdata, clen16);
1283 /* Initialize the output bitstream for Huffman symbols and verbatim bits
1284 * (writing backwards). */
1285 lzms_output_bitstream_init(&c->os, cdata, clen16);
1287 /* Calculate the number of offset slots required. */
1288 num_offset_slots = lzms_get_offset_slot(ulen - 1) + 1;
1290 /* Initialize a Huffman encoder for each alphabet. */
1291 lzms_init_huffman_encoder(&c->literal_encoder, &c->os,
1292 LZMS_NUM_LITERAL_SYMS,
1293 LZMS_LITERAL_CODE_REBUILD_FREQ);
1295 lzms_init_huffman_encoder(&c->lz_offset_encoder, &c->os,
1297 LZMS_LZ_OFFSET_CODE_REBUILD_FREQ);
1299 lzms_init_huffman_encoder(&c->length_encoder, &c->os,
1301 LZMS_LENGTH_CODE_REBUILD_FREQ);
1303 lzms_init_huffman_encoder(&c->delta_offset_encoder, &c->os,
1305 LZMS_DELTA_OFFSET_CODE_REBUILD_FREQ);
1307 lzms_init_huffman_encoder(&c->delta_power_encoder, &c->os,
1308 LZMS_NUM_DELTA_POWER_SYMS,
1309 LZMS_DELTA_POWER_CODE_REBUILD_FREQ);
1311 /* Initialize range encoders, all of which wrap around the same
1312 * lzms_range_encoder_raw. */
1313 lzms_init_range_encoder(&c->main_range_encoder,
1314 &c->rc, LZMS_NUM_MAIN_STATES);
1316 lzms_init_range_encoder(&c->match_range_encoder,
1317 &c->rc, LZMS_NUM_MATCH_STATES);
1319 lzms_init_range_encoder(&c->lz_match_range_encoder,
1320 &c->rc, LZMS_NUM_LZ_MATCH_STATES);
1322 for (unsigned i = 0; i < ARRAY_LEN(c->lz_repeat_match_range_encoders); i++)
1323 lzms_init_range_encoder(&c->lz_repeat_match_range_encoders[i],
1324 &c->rc, LZMS_NUM_LZ_REPEAT_MATCH_STATES);
1326 lzms_init_range_encoder(&c->delta_match_range_encoder,
1327 &c->rc, LZMS_NUM_DELTA_MATCH_STATES);
1329 for (unsigned i = 0; i < ARRAY_LEN(c->delta_repeat_match_range_encoders); i++)
1330 lzms_init_range_encoder(&c->delta_repeat_match_range_encoders[i],
1331 &c->rc, LZMS_NUM_DELTA_REPEAT_MATCH_STATES);
1333 /* Set initial length costs for lengths < LZMS_NUM_FAST_LENGTHS. */
1334 lzms_update_fast_length_costs(c);
1337 /* Flush the output streams, prepare the final compressed data, and return its
1340 * A return value of 0 indicates that the data could not be compressed to fit in
1341 * the available space. */
1343 lzms_finalize(struct lzms_compressor *c, u8 *cdata, size_t csize_avail)
1345 size_t num_forwards_bytes;
1346 size_t num_backwards_bytes;
1348 /* Flush both the forwards and backwards streams, and make sure they
1349 * didn't cross each other and start overwriting each other's data. */
1350 if (!lzms_output_bitstream_flush(&c->os))
1353 if (!lzms_range_encoder_raw_flush(&c->rc))
1356 if (c->rc.next > c->os.next)
1359 /* Now the compressed buffer contains the data output by the forwards
1360 * bitstream, then empty space, then data output by the backwards
1361 * bitstream. Move the data output by the backwards bitstream to be
1362 * adjacent to the data output by the forward bitstream, and calculate
1363 * the compressed size that this results in. */
1364 num_forwards_bytes = (u8*)c->rc.next - (u8*)cdata;
1365 num_backwards_bytes = ((u8*)cdata + csize_avail) - (u8*)c->os.next;
1367 memmove(cdata + num_forwards_bytes, c->os.next, num_backwards_bytes);
1369 return num_forwards_bytes + num_backwards_bytes;
1372 /* Set internal compression parameters for the specified compression level and
1373 * maximum window size. */
1375 lzms_build_params(unsigned int compression_level,
1376 struct lzms_compressor_params *params)
1378 /* Allow length 2 matches if the compression level is sufficiently high.
1380 if (compression_level >= 45)
1381 params->min_match_length = 2;
1383 params->min_match_length = 3;
1385 /* Scale nice_match_length and max_search_depth with the compression
1386 * level. But to allow an optimization on length cost calculations,
1387 * don't allow nice_match_length to exceed LZMS_NUM_FAST_LENGTH. */
1388 params->nice_match_length = ((u64)compression_level * 32) / 50;
1389 if (params->nice_match_length < params->min_match_length)
1390 params->nice_match_length = params->min_match_length;
1391 if (params->nice_match_length > LZMS_NUM_FAST_LENGTHS)
1392 params->nice_match_length = LZMS_NUM_FAST_LENGTHS;
1393 params->max_search_depth = compression_level;
1395 params->optim_array_length = 1024;
1398 /* Given the internal compression parameters and maximum window size, build the
1399 * Lempel-Ziv match-finder parameters. */
1401 lzms_build_mf_params(const struct lzms_compressor_params *lzms_params,
1402 u32 max_window_size, struct lz_mf_params *mf_params)
1404 memset(mf_params, 0, sizeof(*mf_params));
1406 /* Choose an appropriate match-finding algorithm. */
1407 if (max_window_size <= 2097152)
1408 mf_params->algorithm = LZ_MF_BINARY_TREES;
1409 else if (max_window_size <= 33554432)
1410 mf_params->algorithm = LZ_MF_LCP_INTERVAL_TREE;
1412 mf_params->algorithm = LZ_MF_LINKED_SUFFIX_ARRAY;
1414 mf_params->max_window_size = max_window_size;
1415 mf_params->min_match_len = lzms_params->min_match_length;
1416 mf_params->max_search_depth = lzms_params->max_search_depth;
1417 mf_params->nice_match_len = lzms_params->nice_match_length;
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;
1427 struct lz_mf_params mf_params;
1430 if (max_block_size >= INT32_MAX)
1433 lzms_build_params(compression_level, ¶ms);
1434 lzms_build_mf_params(¶ms, max_block_size, &mf_params);
1436 size += sizeof(struct lzms_compressor);
1439 size += max_block_size;
1442 size += lz_mf_get_needed_memory(mf_params.algorithm, max_block_size);
1445 size += min(params.max_search_depth, params.nice_match_length) *
1446 sizeof(struct lz_match);
1449 size += (params.optim_array_length + params.nice_match_length) *
1450 sizeof(struct lzms_mc_pos_data);
1456 lzms_create_compressor(size_t max_block_size, unsigned int compression_level,
1459 struct lzms_compressor *c;
1460 struct lzms_compressor_params params;
1461 struct lz_mf_params mf_params;
1463 if (max_block_size >= INT32_MAX)
1464 return WIMLIB_ERR_INVALID_PARAM;
1466 lzms_build_params(compression_level, ¶ms);
1467 lzms_build_mf_params(¶ms, max_block_size, &mf_params);
1468 if (!lz_mf_params_valid(&mf_params))
1469 return WIMLIB_ERR_INVALID_PARAM;
1471 c = CALLOC(1, sizeof(struct lzms_compressor));
1477 c->cur_window = MALLOC(max_block_size);
1481 c->mf = lz_mf_alloc(&mf_params);
1485 c->matches = MALLOC(min(params.max_search_depth,
1486 params.nice_match_length) *
1487 sizeof(struct lz_match));
1491 c->optimum = MALLOC((params.optim_array_length +
1492 params.nice_match_length) *
1493 sizeof(struct lzms_mc_pos_data));
1496 c->optimum_end = &c->optimum[params.optim_array_length];
1500 lzms_init_rc_costs();
1502 lzms_init_fast_slots(c);
1508 lzms_free_compressor(c);
1509 return WIMLIB_ERR_NOMEM;
1513 lzms_compress(const void *uncompressed_data, size_t uncompressed_size,
1514 void *compressed_data, size_t compressed_size_avail, void *_c)
1516 struct lzms_compressor *c = _c;
1518 /* Don't bother compressing extremely small inputs. */
1519 if (uncompressed_size < 4)
1522 /* Cap the available compressed size to a 32-bit integer and round it
1523 * down to the nearest multiple of 2. */
1524 if (compressed_size_avail > UINT32_MAX)
1525 compressed_size_avail = UINT32_MAX;
1526 if (compressed_size_avail & 1)
1527 compressed_size_avail--;
1529 /* Initialize the compressor structures. */
1530 lzms_prepare_compressor(c, uncompressed_data, uncompressed_size,
1531 compressed_data, compressed_size_avail / 2);
1533 /* Preprocess the uncompressed data. */
1534 lzms_x86_filter(c->cur_window, c->cur_window_size,
1535 c->last_target_usages, false);
1537 /* Load the window into the match-finder. */
1538 lz_mf_load_window(c->mf, c->cur_window, c->cur_window_size);
1540 /* Compute and encode a literal/match sequence that decompresses to the
1541 * preprocessed data. */
1542 lzms_near_optimal_parse(c);
1544 /* Return the compressed data size or 0. */
1545 return lzms_finalize(c, compressed_data, compressed_size_avail);
1549 lzms_free_compressor(void *_c)
1551 struct lzms_compressor *c = _c;
1554 FREE(c->cur_window);
1562 const struct compressor_ops lzms_compressor_ops = {
1563 .get_needed_memory = lzms_get_needed_memory,
1564 .create_compressor = lzms_create_compressor,
1565 .compress = lzms_compress,
1566 .free_compressor = lzms_free_compressor,