4 * A decompressor for the LZMS compression format.
8 * Copyright (C) 2013, 2014, 2015 Eric Biggers
10 * This file is free software; you can redistribute it and/or modify it under
11 * the terms of the GNU Lesser General Public License as published by the Free
12 * Software Foundation; either version 3 of the License, or (at your option) any
15 * This file is distributed in the hope that it will be useful, but WITHOUT
16 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
17 * FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
20 * You should have received a copy of the GNU Lesser General Public License
21 * along with this file; if not, see http://www.gnu.org/licenses/.
25 * This is a decompressor for the LZMS compression format used by Microsoft.
26 * This format is not documented, but it is one of the formats supported by the
27 * compression API available in Windows 8, and as of Windows 8 it is one of the
28 * formats that can be used in WIM files.
30 * This decompressor only implements "raw" decompression, which decompresses a
31 * single LZMS-compressed block. This behavior is the same as that of
32 * Decompress() in the Windows 8 compression API when using a compression handle
33 * created with CreateDecompressor() with the Algorithm parameter specified as
34 * COMPRESS_ALGORITHM_LZMS | COMPRESS_RAW. Presumably, non-raw LZMS data is a
35 * container format from which the locations and sizes (both compressed and
36 * uncompressed) of the constituent blocks can be determined.
38 * An LZMS-compressed block must be read in 16-bit little endian units from both
39 * directions. One logical bitstream starts at the front of the block and
40 * proceeds forwards. Another logical bitstream starts at the end of the block
41 * and proceeds backwards. Bits read from the forwards bitstream constitute
42 * binary range-encoded data, whereas bits read from the backwards bitstream
43 * constitute Huffman-encoded symbols or verbatim bits. For both bitstreams,
44 * the ordering of the bits within the 16-bit coding units is such that the
45 * first bit is the high-order bit and the last bit is the low-order bit.
47 * From these two logical bitstreams, an LZMS decompressor can reconstitute the
48 * series of items that make up the LZMS data representation. Each such item
49 * may be a literal byte or a match. Matches may be either traditional LZ77
50 * matches or "delta" matches, either of which can have its offset encoded
51 * explicitly or encoded via a reference to a recently used (repeat) offset.
53 * A traditional LZ77 match consists of a length and offset. It asserts that
54 * the sequence of bytes beginning at the current position and extending for the
55 * length is equal to the same-length sequence of bytes at the offset back in
56 * the data buffer. This type of match can be visualized as follows, with the
57 * caveat that the sequences may overlap:
60 * --------------------
62 * B[1...len] A[1...len]
64 * Decoding proceeds as follows:
70 * On the other hand, a delta match consists of a "span" as well as a length and
71 * offset. A delta match can be visualized as follows, with the caveat that the
72 * various sequences may overlap:
75 * -----------------------------
78 * ------------- -------------
80 * D[1...len] C[1...len] B[1...len] A[1...len]
82 * Decoding proceeds as follows:
85 * *A++ = *B++ + *C++ - *D++;
88 * A delta match asserts that the bytewise differences of the A and B sequences
89 * are equal to the bytewise differences of the C and D sequences. The
90 * sequences within each pair are separated by the same number of bytes, the
91 * "span". The inter-pair distance is the "offset". In LZMS, spans are
92 * restricted to powers of 2 between 2**0 and 2**7 inclusively. Offsets are
93 * restricted to multiples of the span. The stored value for the offset is the
94 * "raw offset", which is the real offset divided by the span.
96 * Delta matches can cover data containing a series of power-of-2 sized integers
97 * that is linearly increasing or decreasing. Another way of thinking about it
98 * is that a delta match can match a longer sequence that is interrupted by a
99 * non-matching byte, provided that the non-matching byte is a continuation of a
100 * linearly changing pattern. Examples of files that may contain data like this
101 * are uncompressed bitmap images, uncompressed digital audio, and Unicode data
102 * tables. To some extent, this match type is a replacement for delta filters
103 * or multimedia filters that are sometimes used in other compression software
104 * (e.g. 'xz --delta --lzma2'). However, on most types of files, delta matches
105 * do not seem to be very useful.
107 * Both LZ and delta matches may use overlapping sequences. Therefore, they
108 * must be decoded as if only one byte is copied at a time.
110 * For both LZ and delta matches, any match length in [1, 1073809578] can be
111 * represented. Similarly, any match offset in [1, 1180427428] can be
112 * represented. For delta matches, this range applies to the raw offset, so the
113 * real offset may be larger.
115 * For LZ matches, up to 3 repeat offsets are allowed, similar to some other
116 * LZ-based formats such as LZX and LZMA. They must updated in an LRU fashion,
117 * except for a quirk: inserting anything to the front of the queue must be
118 * delayed by one LZMS item. The reason for this is presumably that there is
119 * almost no reason to code the same match offset twice in a row, since you
120 * might as well have coded a longer match at that offset. For this same
121 * reason, it also is a requirement that when an offset in the queue is used,
122 * that offset is removed from the queue immediately (and made pending for
123 * front-insertion after the following decoded item), and everything to the
124 * right is shifted left one queue slot. This creates a need for an "overflow"
125 * fourth entry in the queue, even though it is only possible to decode
126 * references to the first 3 entries at any given time. The queue must be
127 * initialized to the offsets {1, 2, 3, 4}.
129 * Repeat delta matches are handled similarly, but for them there are two queues
130 * updated in lock-step: one for powers and one for raw offsets. The power
131 * queue must be initialized to {0, 0, 0, 0}, and the raw offset queue must be
132 * initialized to {1, 2, 3, 4}.
134 * Bits from the binary range decoder must be used to disambiguate item types.
135 * The range decoder must hold two state variables: the range, which must
136 * initially be set to 0xffffffff, and the current code, which must initially be
137 * set to the first 32 bits read from the forwards bitstream. The range must be
138 * maintained above 0xffff; when it falls below 0xffff, both the range and code
139 * must be left-shifted by 16 bits and the low 16 bits of the code must be
140 * filled in with the next 16 bits from the forwards bitstream.
142 * To decode each bit, the binary range decoder requires a probability that is
143 * logically a real number between 0 and 1. Multiplying this probability by the
144 * current range and taking the floor gives the bound between the 0-bit region of
145 * the range and the 1-bit region of the range. However, in LZMS, probabilities
146 * are restricted to values of n/64 where n is an integer is between 1 and 63
147 * inclusively, so the implementation may use integer operations instead.
148 * Following calculation of the bound, if the current code is in the 0-bit
149 * region, the new range becomes the current code and the decoded bit is 0;
150 * otherwise, the bound must be subtracted from both the range and the code, and
151 * the decoded bit is 1. More information about range coding can be found at
152 * https://en.wikipedia.org/wiki/Range_encoding. Furthermore, note that the
153 * LZMA format also uses range coding and has public domain code available for
156 * The probability used to range-decode each bit must be taken from a table, of
157 * which one instance must exist for each distinct context in which a
158 * range-decoded bit is needed. At each call of the range decoder, the
159 * appropriate probability must be obtained by indexing the appropriate
160 * probability table with the last 4 (in the context disambiguating literals
161 * from matches), 5 (in the context disambiguating LZ matches from delta
162 * matches), or 6 (in all other contexts) bits recently range-decoded in that
163 * context, ordered such that the most recently decoded bit is the low-order bit
166 * Furthermore, each probability entry itself is variable, as its value must be
167 * maintained as n/64 where n is the number of 0 bits in the most recently
168 * decoded 64 bits with that same entry. This allows the compressed
169 * representation to adapt to the input and use fewer bits to represent the most
170 * likely data; note that LZMA uses a similar scheme. Initially, the most
171 * recently 64 decoded bits for each probability entry are assumed to be
172 * 0x0000000055555555 (high order to low order); therefore, all probabilities
173 * are initially 48/64. During the course of decoding, each probability may be
174 * updated to as low as 0/64 (as a result of reading many consecutive 1 bits
175 * with that entry) or as high as 64/64 (as a result of reading many consecutive
176 * 0 bits with that entry); however, probabilities of 0/64 and 64/64 cannot be
177 * used as-is but rather must be adjusted to 1/64 and 63/64, respectively,
178 * before being used for range decoding.
180 * Representations of the LZMS items themselves must be read from the backwards
181 * bitstream. For this, there are 5 different Huffman codes used:
183 * - The literal code, used for decoding literal bytes. Each of the 256
184 * symbols represents a literal byte. This code must be rebuilt whenever
185 * 1024 symbols have been decoded with it.
187 * - The LZ offset code, used for decoding the offsets of standard LZ77
188 * matches. Each symbol represents an offset slot, which corresponds to a
189 * base value and some number of extra bits which must be read and added to
190 * the base value to reconstitute the full offset. The number of symbols in
191 * this code is the number of offset slots needed to represent all possible
192 * offsets in the uncompressed block. This code must be rebuilt whenever
193 * 1024 symbols have been decoded with it.
195 * - The length code, used for decoding length symbols. Each of the 54 symbols
196 * represents a length slot, which corresponds to a base value and some
197 * number of extra bits which must be read and added to the base value to
198 * reconstitute the full length. This code must be rebuilt whenever 512
199 * symbols have been decoded with it.
201 * - The delta offset code, used for decoding the offsets of delta matches.
202 * Each symbol corresponds to an offset slot, which corresponds to a base
203 * value and some number of extra bits which must be read and added to the
204 * base value to reconstitute the full offset. The number of symbols in this
205 * code is equal to the number of symbols in the LZ offset code. This code
206 * must be rebuilt whenever 1024 symbols have been decoded with it.
208 * - The delta power code, used for decoding the powers of delta matches. Each
209 * of the 8 symbols corresponds to a power. This code must be rebuilt
210 * whenever 512 symbols have been decoded with it.
212 * Initially, each Huffman code must be built assuming that each symbol in that
213 * code has frequency 1. Following that, each code must be rebuilt each time a
214 * certain number of symbols, as noted above, has been decoded with it. The
215 * symbol frequencies for a code must be halved after each rebuild of that code;
216 * this makes the codes adapt to the more recent data.
218 * Like other compression formats such as XPRESS, LZX, and DEFLATE, the LZMS
219 * format requires that all Huffman codes be constructed in canonical form.
220 * This form requires that same-length codewords be lexicographically ordered
221 * the same way as the corresponding symbols and that all shorter codewords
222 * lexicographically precede longer codewords. Such a code can be constructed
223 * directly from codeword lengths.
225 * Even with the canonical code restriction, the same frequencies can be used to
226 * construct multiple valid Huffman codes. Therefore, the decompressor needs to
227 * construct the right one. Specifically, the LZMS format requires that the
228 * Huffman code be constructed as if the well-known priority queue algorithm is
229 * used and frequency ties are always broken in favor of leaf nodes.
231 * Codewords in LZMS are guaranteed to not exceed 15 bits. The format otherwise
232 * places no restrictions on codeword length. Therefore, the Huffman code
233 * construction algorithm that a correct LZMS decompressor uses need not
234 * implement length-limited code construction. But if it does (e.g. by virtue
235 * of being shared among multiple compression algorithms), the details of how it
236 * does so are unimportant, provided that the maximum codeword length parameter
237 * is set to at least 15 bits.
239 * After all LZMS items have been decoded, the data must be postprocessed to
240 * translate absolute address encoded in x86 instructions into their original
241 * relative addresses.
243 * Details omitted above can be found in the code. Note that in the absence of
244 * an official specification there is no guarantee that this decompressor
245 * handles all possible cases.
252 #include "wimlib/compress_common.h"
253 #include "wimlib/decompress_common.h"
254 #include "wimlib/decompressor_ops.h"
255 #include "wimlib/error.h"
256 #include "wimlib/lzms_common.h"
257 #include "wimlib/util.h"
259 /* The TABLEBITS values can be changed; they only affect decoding speed. */
260 #define LZMS_LITERAL_TABLEBITS 10
261 #define LZMS_LENGTH_TABLEBITS 10
262 #define LZMS_LZ_OFFSET_TABLEBITS 10
263 #define LZMS_DELTA_OFFSET_TABLEBITS 10
264 #define LZMS_DELTA_POWER_TABLEBITS 8
266 struct lzms_range_decoder {
268 /* The relevant part of the current range. Although the logical range
269 * for range decoding is a very large integer, only a small portion
270 * matters at any given time, and it can be normalized (shifted left)
271 * whenever it gets too small. */
274 /* The current position in the range encoded by the portion of the input
278 /* Pointer to the next little-endian 16-bit integer in the compressed
279 * input data (reading forwards). */
282 /* Pointer to the end of the compressed input data. */
286 typedef u64 bitbuf_t;
288 struct lzms_input_bitstream {
290 /* Holding variable for bits that have been read from the compressed
291 * data. The bit ordering is high to low. */
294 /* Number of bits currently held in @bitbuf. */
297 /* Pointer to the one past the next little-endian 16-bit integer in the
298 * compressed input data (reading backwards). */
301 /* Pointer to the beginning of the compressed input data. */
305 /* Bookkeeping information for an adaptive Huffman code */
306 struct lzms_huffman_rebuild_info {
307 unsigned num_syms_until_rebuild;
308 unsigned rebuild_freq;
317 struct lzms_decompressor {
319 /* 'last_target_usages' is in union with everything else because it is
320 * only used for postprocessing. */
324 struct lzms_range_decoder rd;
325 struct lzms_input_bitstream is;
327 /* Match offset LRU queues */
328 u32 recent_lz_offsets[LZMS_NUM_RECENT_OFFSETS + 1];
329 u64 recent_delta_offsets[LZMS_NUM_RECENT_OFFSETS + 1];
330 u32 pending_lz_offset;
331 u64 pending_delta_offset;
332 const u8 *lz_offset_still_pending;
333 const u8 *delta_offset_still_pending;
335 /* States and probabilities for range decoding */
338 struct lzms_probability_entry main_prob_entries[
339 LZMS_NUM_MAIN_STATES];
342 struct lzms_probability_entry match_prob_entries[
343 LZMS_NUM_MATCH_STATES];
346 struct lzms_probability_entry lz_match_prob_entries[
347 LZMS_NUM_LZ_MATCH_STATES];
349 u32 delta_match_state;
350 struct lzms_probability_entry delta_match_prob_entries[
351 LZMS_NUM_DELTA_MATCH_STATES];
353 u32 lz_repeat_match_states[LZMS_NUM_RECENT_OFFSETS - 1];
354 struct lzms_probability_entry lz_repeat_match_prob_entries[
355 LZMS_NUM_RECENT_OFFSETS - 1][LZMS_NUM_LZ_REPEAT_MATCH_STATES];
357 u32 delta_repeat_match_states[LZMS_NUM_RECENT_OFFSETS - 1];
358 struct lzms_probability_entry delta_repeat_match_prob_entries[
359 LZMS_NUM_RECENT_OFFSETS - 1][LZMS_NUM_DELTA_REPEAT_MATCH_STATES];
361 /* Huffman decoding */
363 u16 literal_decode_table[(1 << LZMS_LITERAL_TABLEBITS) +
364 (2 * LZMS_NUM_LITERAL_SYMS)]
365 _aligned_attribute(DECODE_TABLE_ALIGNMENT);
366 u32 literal_freqs[LZMS_NUM_LITERAL_SYMS];
367 struct lzms_huffman_rebuild_info literal_rebuild_info;
369 u16 length_decode_table[(1 << LZMS_LENGTH_TABLEBITS) +
370 (2 * LZMS_NUM_LENGTH_SYMS)]
371 _aligned_attribute(DECODE_TABLE_ALIGNMENT);
372 u32 length_freqs[LZMS_NUM_LENGTH_SYMS];
373 struct lzms_huffman_rebuild_info length_rebuild_info;
375 u16 lz_offset_decode_table[(1 << LZMS_LZ_OFFSET_TABLEBITS) +
376 ( 2 * LZMS_MAX_NUM_OFFSET_SYMS)]
377 _aligned_attribute(DECODE_TABLE_ALIGNMENT);
378 u32 lz_offset_freqs[LZMS_MAX_NUM_OFFSET_SYMS];
379 struct lzms_huffman_rebuild_info lz_offset_rebuild_info;
381 u16 delta_offset_decode_table[(1 << LZMS_DELTA_OFFSET_TABLEBITS) +
382 (2 * LZMS_MAX_NUM_OFFSET_SYMS)]
383 _aligned_attribute(DECODE_TABLE_ALIGNMENT);
384 u32 delta_offset_freqs[LZMS_MAX_NUM_OFFSET_SYMS];
385 struct lzms_huffman_rebuild_info delta_offset_rebuild_info;
387 u16 delta_power_decode_table[(1 << LZMS_DELTA_POWER_TABLEBITS) +
388 (2 * LZMS_NUM_DELTA_POWER_SYMS)]
389 _aligned_attribute(DECODE_TABLE_ALIGNMENT);
390 u32 delta_power_freqs[LZMS_NUM_DELTA_POWER_SYMS];
391 struct lzms_huffman_rebuild_info delta_power_rebuild_info;
393 u32 codewords[LZMS_MAX_NUM_SYMS];
394 u8 lens[LZMS_MAX_NUM_SYMS];
398 s32 last_target_usages[65536];
403 /* Initialize the input bitstream @is to read backwards from the compressed data
404 * buffer @in that is @count 16-bit integers long. */
406 lzms_input_bitstream_init(struct lzms_input_bitstream *is,
407 const le16 *in, size_t count)
411 is->next = in + count;
415 /* Ensure that at least @num_bits bits are in the bitbuffer variable.
416 * @num_bits cannot be more than 32. */
418 lzms_ensure_bits(struct lzms_input_bitstream *is, unsigned num_bits)
420 if (is->bitsleft >= num_bits)
423 if (likely(is->next != is->begin))
424 is->bitbuf |= (bitbuf_t)le16_to_cpu(*--is->next)
425 << (sizeof(is->bitbuf) * 8 - is->bitsleft - 16);
428 if (likely(is->next != is->begin))
429 is->bitbuf |= (bitbuf_t)le16_to_cpu(*--is->next)
430 << (sizeof(is->bitbuf) * 8 - is->bitsleft - 16);
434 /* Get @num_bits bits from the bitbuffer variable. */
435 static inline bitbuf_t
436 lzms_peek_bits(struct lzms_input_bitstream *is, unsigned num_bits)
438 if (unlikely(num_bits == 0))
440 return is->bitbuf >> (sizeof(is->bitbuf) * 8 - num_bits);
443 /* Remove @num_bits bits from the bitbuffer variable. */
445 lzms_remove_bits(struct lzms_input_bitstream *is, unsigned num_bits)
447 is->bitbuf <<= num_bits;
448 is->bitsleft -= num_bits;
451 /* Remove and return @num_bits bits from the bitbuffer variable. */
452 static inline bitbuf_t
453 lzms_pop_bits(struct lzms_input_bitstream *is, unsigned num_bits)
455 bitbuf_t bits = lzms_peek_bits(is, num_bits);
456 lzms_remove_bits(is, num_bits);
460 /* Read @num_bits bits from the input bitstream. */
461 static inline bitbuf_t
462 lzms_read_bits(struct lzms_input_bitstream *is, unsigned num_bits)
464 lzms_ensure_bits(is, num_bits);
465 return lzms_pop_bits(is, num_bits);
468 /* Initialize the range decoder @rd to read forwards from the compressed data
469 * buffer @in that is @count 16-bit integers long. */
471 lzms_range_decoder_init(struct lzms_range_decoder *rd,
472 const le16 *in, size_t count)
474 rd->range = 0xffffffff;
475 rd->code = ((u32)le16_to_cpu(in[0]) << 16) | le16_to_cpu(in[1]);
477 rd->end = in + count;
480 /* Decode and return the next bit from the range decoder.
482 * @prob is the chance out of LZMS_PROBABILITY_MAX that the next bit is 0.
485 lzms_range_decoder_decode_bit(struct lzms_range_decoder *rd, u32 prob)
489 /* Normalize if needed. */
490 if (rd->range <= 0xffff) {
493 if (likely(rd->next != rd->end))
494 rd->code |= le16_to_cpu(*rd->next++);
497 /* Based on the probability, calculate the bound between the 0-bit
498 * region and the 1-bit region of the range. */
499 bound = (rd->range >> LZMS_PROBABILITY_BITS) * prob;
501 if (rd->code < bound) {
502 /* Current code is in the 0-bit region of the range. */
506 /* Current code is in the 1-bit region of the range. */
513 /* Decode and return the next bit from the range decoder. This wraps around
514 * lzms_range_decoder_decode_bit() to handle using and updating the appropriate
515 * state and probability entry. */
517 lzms_range_decode_bit(struct lzms_range_decoder *rd,
518 u32 *state_p, u32 num_states,
519 struct lzms_probability_entry prob_entries[])
521 struct lzms_probability_entry *prob_entry;
525 /* Load the probability entry corresponding to the current state. */
526 prob_entry = &prob_entries[*state_p];
528 /* Get the probability that the next bit is 0. */
529 prob = lzms_get_probability(prob_entry);
531 /* Decode the next bit. */
532 bit = lzms_range_decoder_decode_bit(rd, prob);
534 /* Update the state and probability entry based on the decoded bit. */
535 *state_p = ((*state_p << 1) | bit) & (num_states - 1);
536 lzms_update_probability_entry(prob_entry, bit);
538 /* Return the decoded bit. */
543 lzms_decode_main_bit(struct lzms_decompressor *d)
545 return lzms_range_decode_bit(&d->rd, &d->main_state,
546 LZMS_NUM_MAIN_STATES,
547 d->main_prob_entries);
551 lzms_decode_match_bit(struct lzms_decompressor *d)
553 return lzms_range_decode_bit(&d->rd, &d->match_state,
554 LZMS_NUM_MATCH_STATES,
555 d->match_prob_entries);
559 lzms_decode_lz_match_bit(struct lzms_decompressor *d)
561 return lzms_range_decode_bit(&d->rd, &d->lz_match_state,
562 LZMS_NUM_LZ_MATCH_STATES,
563 d->lz_match_prob_entries);
567 lzms_decode_delta_match_bit(struct lzms_decompressor *d)
569 return lzms_range_decode_bit(&d->rd, &d->delta_match_state,
570 LZMS_NUM_DELTA_MATCH_STATES,
571 d->delta_match_prob_entries);
575 lzms_decode_lz_repeat_match_bit(struct lzms_decompressor *d, int idx)
577 return lzms_range_decode_bit(&d->rd, &d->lz_repeat_match_states[idx],
578 LZMS_NUM_LZ_REPEAT_MATCH_STATES,
579 d->lz_repeat_match_prob_entries[idx]);
583 lzms_decode_delta_repeat_match_bit(struct lzms_decompressor *d, int idx)
585 return lzms_range_decode_bit(&d->rd, &d->delta_repeat_match_states[idx],
586 LZMS_NUM_DELTA_REPEAT_MATCH_STATES,
587 d->delta_repeat_match_prob_entries[idx]);
591 lzms_init_huffman_rebuild_info(struct lzms_huffman_rebuild_info *info,
592 unsigned rebuild_freq,
593 u16 *decode_table, unsigned table_bits,
594 u32 *freqs, u32 *codewords, u8 *lens,
597 info->num_syms_until_rebuild = 1;
598 info->rebuild_freq = rebuild_freq;
599 info->decode_table = decode_table;
600 info->table_bits = table_bits;
602 info->codewords = codewords;
604 info->num_syms = num_syms;
605 lzms_init_symbol_frequencies(freqs, num_syms);
609 lzms_rebuild_huffman_code(struct lzms_huffman_rebuild_info *info)
611 make_canonical_huffman_code(info->num_syms, LZMS_MAX_CODEWORD_LEN,
612 info->freqs, info->lens, info->codewords);
613 make_huffman_decode_table(info->decode_table, info->num_syms,
614 info->table_bits, info->lens,
615 LZMS_MAX_CODEWORD_LEN);
616 for (unsigned i = 0; i < info->num_syms; i++)
617 info->freqs[i] = (info->freqs[i] >> 1) + 1;
618 info->num_syms_until_rebuild = info->rebuild_freq;
621 static inline unsigned
622 lzms_decode_huffman_symbol(struct lzms_input_bitstream *is,
623 u16 decode_table[], unsigned table_bits,
624 struct lzms_huffman_rebuild_info *rebuild_info)
630 if (unlikely(--rebuild_info->num_syms_until_rebuild == 0))
631 lzms_rebuild_huffman_code(rebuild_info);
633 lzms_ensure_bits(is, LZMS_MAX_CODEWORD_LEN);
635 /* Index the decode table by the next table_bits bits of the input. */
636 key_bits = lzms_peek_bits(is, table_bits);
637 entry = decode_table[key_bits];
638 if (likely(entry < 0xC000)) {
639 /* Fast case: The decode table directly provided the symbol and
640 * codeword length. The low 11 bits are the symbol, and the
641 * high 5 bits are the codeword length. */
642 lzms_remove_bits(is, entry >> 11);
645 /* Slow case: The codeword for the symbol is longer than
646 * table_bits, so the symbol does not have an entry directly in
647 * the first (1 << table_bits) entries of the decode table.
648 * Traverse the appropriate binary tree bit-by-bit in order to
649 * decode the symbol. */
650 lzms_remove_bits(is, table_bits);
652 key_bits = (entry & 0x3FFF) + lzms_pop_bits(is, 1);
653 } while ((entry = decode_table[key_bits]) >= 0xC000);
657 /* Tally and return the decoded symbol. */
658 rebuild_info->freqs[sym]++;
663 lzms_decode_literal(struct lzms_decompressor *d)
665 return lzms_decode_huffman_symbol(&d->is,
666 d->literal_decode_table,
667 LZMS_LITERAL_TABLEBITS,
668 &d->literal_rebuild_info);
672 lzms_decode_length(struct lzms_decompressor *d)
674 unsigned slot = lzms_decode_huffman_symbol(&d->is,
675 d->length_decode_table,
676 LZMS_LENGTH_TABLEBITS,
677 &d->length_rebuild_info);
678 u32 length = lzms_length_slot_base[slot];
679 unsigned num_extra_bits = lzms_extra_length_bits[slot];
680 /* Usually most lengths are short and have no extra bits. */
682 length += lzms_read_bits(&d->is, num_extra_bits);
687 lzms_decode_lz_offset(struct lzms_decompressor *d)
689 unsigned slot = lzms_decode_huffman_symbol(&d->is,
690 d->lz_offset_decode_table,
691 LZMS_LZ_OFFSET_TABLEBITS,
692 &d->lz_offset_rebuild_info);
693 return lzms_offset_slot_base[slot] +
694 lzms_read_bits(&d->is, lzms_extra_offset_bits[slot]);
698 lzms_decode_delta_offset(struct lzms_decompressor *d)
700 unsigned slot = lzms_decode_huffman_symbol(&d->is,
701 d->delta_offset_decode_table,
702 LZMS_DELTA_OFFSET_TABLEBITS,
703 &d->delta_offset_rebuild_info);
704 return lzms_offset_slot_base[slot] +
705 lzms_read_bits(&d->is, lzms_extra_offset_bits[slot]);
709 lzms_decode_delta_power(struct lzms_decompressor *d)
711 return lzms_decode_huffman_symbol(&d->is,
712 d->delta_power_decode_table,
713 LZMS_DELTA_POWER_TABLEBITS,
714 &d->delta_power_rebuild_info);
717 /* Decode the series of literals and matches from the LZMS-compressed data.
718 * Return 0 if successful or -1 if the compressed data is invalid. */
720 lzms_decode_items(struct lzms_decompressor * const restrict d,
721 u8 * const restrict out, const size_t out_nbytes)
724 u8 * const out_end = out + out_nbytes;
726 while (out_next != out_end) {
728 if (!lzms_decode_main_bit(d)) {
731 *out_next++ = lzms_decode_literal(d);
733 } else if (!lzms_decode_match_bit(d)) {
740 if (d->pending_lz_offset != 0 &&
741 out_next != d->lz_offset_still_pending)
743 BUILD_BUG_ON(LZMS_NUM_RECENT_OFFSETS != 3);
744 d->recent_lz_offsets[3] = d->recent_lz_offsets[2];
745 d->recent_lz_offsets[2] = d->recent_lz_offsets[1];
746 d->recent_lz_offsets[1] = d->recent_lz_offsets[0];
747 d->recent_lz_offsets[0] = d->pending_lz_offset;
748 d->pending_lz_offset = 0;
751 if (!lzms_decode_lz_match_bit(d)) {
752 /* Explicit offset */
753 offset = lzms_decode_lz_offset(d);
757 BUILD_BUG_ON(LZMS_NUM_RECENT_OFFSETS != 3);
758 if (!lzms_decode_lz_repeat_match_bit(d, 0)) {
759 offset = d->recent_lz_offsets[0];
760 d->recent_lz_offsets[0] = d->recent_lz_offsets[1];
761 d->recent_lz_offsets[1] = d->recent_lz_offsets[2];
762 d->recent_lz_offsets[2] = d->recent_lz_offsets[3];
763 } else if (!lzms_decode_lz_repeat_match_bit(d, 1)) {
764 offset = d->recent_lz_offsets[1];
765 d->recent_lz_offsets[1] = d->recent_lz_offsets[2];
766 d->recent_lz_offsets[2] = d->recent_lz_offsets[3];
768 offset = d->recent_lz_offsets[2];
769 d->recent_lz_offsets[2] = d->recent_lz_offsets[3];
773 if (d->pending_lz_offset != 0) {
774 BUILD_BUG_ON(LZMS_NUM_RECENT_OFFSETS != 3);
775 d->recent_lz_offsets[3] = d->recent_lz_offsets[2];
776 d->recent_lz_offsets[2] = d->recent_lz_offsets[1];
777 d->recent_lz_offsets[1] = d->recent_lz_offsets[0];
778 d->recent_lz_offsets[0] = d->pending_lz_offset;
780 d->pending_lz_offset = offset;
782 length = lzms_decode_length(d);
784 if (unlikely(length > out_end - out_next))
786 if (unlikely(offset > out_next - out))
789 lz_copy(out_next, length, offset, out_end, LZMS_MIN_MATCH_LEN);
792 d->lz_offset_still_pending = out_next;
797 u32 raw_offset, offset1, offset2, offset;
798 const u8 *matchptr1, *matchptr2, *matchptr;
801 if (d->pending_delta_offset != 0 &&
802 out_next != d->delta_offset_still_pending)
804 BUILD_BUG_ON(LZMS_NUM_RECENT_OFFSETS != 3);
805 d->recent_delta_offsets[3] = d->recent_delta_offsets[2];
806 d->recent_delta_offsets[2] = d->recent_delta_offsets[1];
807 d->recent_delta_offsets[1] = d->recent_delta_offsets[0];
808 d->recent_delta_offsets[0] = d->pending_delta_offset;
809 d->pending_delta_offset = 0;
812 if (!lzms_decode_delta_match_bit(d)) {
813 /* Explicit offset */
814 power = lzms_decode_delta_power(d);
815 raw_offset = lzms_decode_delta_offset(d);
820 BUILD_BUG_ON(LZMS_NUM_RECENT_OFFSETS != 3);
821 if (!lzms_decode_delta_repeat_match_bit(d, 0)) {
822 val = d->recent_delta_offsets[0];
823 d->recent_delta_offsets[0] = d->recent_delta_offsets[1];
824 d->recent_delta_offsets[1] = d->recent_delta_offsets[2];
825 d->recent_delta_offsets[2] = d->recent_delta_offsets[3];
826 } else if (!lzms_decode_delta_repeat_match_bit(d, 1)) {
827 val = d->recent_delta_offsets[1];
828 d->recent_delta_offsets[1] = d->recent_delta_offsets[2];
829 d->recent_delta_offsets[2] = d->recent_delta_offsets[3];
831 val = d->recent_delta_offsets[2];
832 d->recent_delta_offsets[2] = d->recent_delta_offsets[3];
835 raw_offset = (u32)val;
838 if (d->pending_delta_offset != 0) {
839 BUILD_BUG_ON(LZMS_NUM_RECENT_OFFSETS != 3);
840 d->recent_delta_offsets[3] = d->recent_delta_offsets[2];
841 d->recent_delta_offsets[2] = d->recent_delta_offsets[1];
842 d->recent_delta_offsets[1] = d->recent_delta_offsets[0];
843 d->recent_delta_offsets[0] = d->pending_delta_offset;
845 d->pending_delta_offset = raw_offset | ((u64)power << 32);
847 length = lzms_decode_length(d);
849 offset1 = (u32)1 << power;
850 offset2 = raw_offset << power;
851 offset = offset1 + offset2;
853 /* raw_offset<<power overflowed? */
854 if (unlikely((offset2 >> power) != raw_offset))
857 /* offset1+offset2 overflowed? */
858 if (unlikely(offset < offset2))
861 if (unlikely(length > out_end - out_next))
864 if (unlikely(offset > out_next - out))
867 matchptr1 = out_next - offset1;
868 matchptr2 = out_next - offset2;
869 matchptr = out_next - offset;
872 *out_next++ = *matchptr1++ + *matchptr2++ - *matchptr++;
875 d->delta_offset_still_pending = out_next;
882 lzms_init_decompressor(struct lzms_decompressor *d, const void *in,
883 size_t in_nbytes, unsigned num_offset_slots)
885 /* Match offset LRU queues */
886 for (int i = 0; i < LZMS_NUM_RECENT_OFFSETS + 1; i++) {
887 d->recent_lz_offsets[i] = i + 1;
888 d->recent_delta_offsets[i] = i + 1;
890 d->pending_lz_offset = 0;
891 d->pending_delta_offset = 0;
895 lzms_range_decoder_init(&d->rd, in, in_nbytes / sizeof(le16));
898 lzms_init_probability_entries(d->main_prob_entries, LZMS_NUM_MAIN_STATES);
901 lzms_init_probability_entries(d->match_prob_entries, LZMS_NUM_MATCH_STATES);
903 d->lz_match_state = 0;
904 lzms_init_probability_entries(d->lz_match_prob_entries, LZMS_NUM_LZ_MATCH_STATES);
906 d->delta_match_state = 0;
907 lzms_init_probability_entries(d->delta_match_prob_entries, LZMS_NUM_DELTA_MATCH_STATES);
909 for (int i = 0; i < LZMS_NUM_RECENT_OFFSETS - 1; i++) {
910 d->lz_repeat_match_states[i] = 0;
911 lzms_init_probability_entries(d->lz_repeat_match_prob_entries[i],
912 LZMS_NUM_LZ_REPEAT_MATCH_STATES);
914 d->delta_repeat_match_states[i] = 0;
915 lzms_init_probability_entries(d->delta_repeat_match_prob_entries[i],
916 LZMS_NUM_DELTA_REPEAT_MATCH_STATES);
919 /* Huffman decoding */
921 lzms_input_bitstream_init(&d->is, in, in_nbytes / sizeof(le16));
923 lzms_init_huffman_rebuild_info(&d->literal_rebuild_info,
924 LZMS_LITERAL_CODE_REBUILD_FREQ,
925 d->literal_decode_table,
926 LZMS_LITERAL_TABLEBITS,
930 LZMS_NUM_LITERAL_SYMS);
932 lzms_init_huffman_rebuild_info(&d->length_rebuild_info,
933 LZMS_LENGTH_CODE_REBUILD_FREQ,
934 d->length_decode_table,
935 LZMS_LENGTH_TABLEBITS,
939 LZMS_NUM_LENGTH_SYMS);
941 lzms_init_huffman_rebuild_info(&d->lz_offset_rebuild_info,
942 LZMS_LZ_OFFSET_CODE_REBUILD_FREQ,
943 d->lz_offset_decode_table,
944 LZMS_LZ_OFFSET_TABLEBITS,
950 lzms_init_huffman_rebuild_info(&d->delta_offset_rebuild_info,
951 LZMS_DELTA_OFFSET_CODE_REBUILD_FREQ,
952 d->delta_offset_decode_table,
953 LZMS_DELTA_OFFSET_TABLEBITS,
954 d->delta_offset_freqs,
959 lzms_init_huffman_rebuild_info(&d->delta_power_rebuild_info,
960 LZMS_DELTA_POWER_CODE_REBUILD_FREQ,
961 d->delta_power_decode_table,
962 LZMS_DELTA_POWER_TABLEBITS,
963 d->delta_power_freqs,
966 LZMS_NUM_DELTA_POWER_SYMS);
970 lzms_create_decompressor(size_t max_bufsize, void **d_ret)
972 struct lzms_decompressor *d;
974 if (max_bufsize > LZMS_MAX_BUFFER_SIZE)
975 return WIMLIB_ERR_INVALID_PARAM;
977 d = ALIGNED_MALLOC(sizeof(struct lzms_decompressor),
978 DECODE_TABLE_ALIGNMENT);
980 return WIMLIB_ERR_NOMEM;
986 /* Decompress @in_nbytes bytes of LZMS-compressed data at @in and write the
987 * uncompressed data, which had original size @out_nbytes, to @out. Return 0 if
988 * successful or -1 if the compressed data is invalid. */
990 lzms_decompress(const void *in, size_t in_nbytes, void *out, size_t out_nbytes,
993 struct lzms_decompressor *d = _d;
996 * Requirements on the compressed data:
998 * 1. LZMS-compressed data is a series of 16-bit integers, so the
999 * compressed data buffer cannot take up an odd number of bytes.
1000 * 2. To prevent poor performance on some architectures, we require that
1001 * the compressed data buffer is 2-byte aligned.
1002 * 3. There must be at least 4 bytes of compressed data, since otherwise
1003 * we cannot even initialize the range decoder.
1005 if ((in_nbytes & 1) || ((uintptr_t)in & 1) || (in_nbytes < 4))
1008 lzms_init_decompressor(d, in, in_nbytes,
1009 lzms_get_num_offset_slots(out_nbytes));
1011 if (lzms_decode_items(d, out, out_nbytes))
1014 lzms_x86_filter(out, out_nbytes, d->last_target_usages, true);
1019 lzms_free_decompressor(void *_d)
1021 struct lzms_decompressor *d = _d;
1026 const struct decompressor_ops lzms_decompressor_ops = {
1027 .create_decompressor = lzms_create_decompressor,
1028 .decompress = lzms_decompress,
1029 .free_decompressor = lzms_free_decompressor,