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 the queue contains
130 * (power, raw offset) pairs. This queue must be initialized to
131 * {(0, 1), (0, 2), (0, 3), (0, 4)}.
133 * Bits from the binary range decoder must be used to disambiguate item types.
134 * The range decoder must hold two state variables: the range, which must
135 * initially be set to 0xffffffff, and the current code, which must initially be
136 * set to the first 32 bits read from the forwards bitstream. The range must be
137 * maintained above 0xffff; when it falls below 0xffff, both the range and code
138 * must be left-shifted by 16 bits and the low 16 bits of the code must be
139 * filled in with the next 16 bits from the forwards bitstream.
141 * To decode each bit, the binary range decoder requires a probability that is
142 * logically a real number between 0 and 1. Multiplying this probability by the
143 * current range and taking the floor gives the bound between the 0-bit region of
144 * the range and the 1-bit region of the range. However, in LZMS, probabilities
145 * are restricted to values of n/64 where n is an integer is between 1 and 63
146 * inclusively, so the implementation may use integer operations instead.
147 * Following calculation of the bound, if the current code is in the 0-bit
148 * region, the new range becomes the current code and the decoded bit is 0;
149 * otherwise, the bound must be subtracted from both the range and the code, and
150 * the decoded bit is 1. More information about range coding can be found at
151 * https://en.wikipedia.org/wiki/Range_encoding. Furthermore, note that the
152 * LZMA format also uses range coding and has public domain code available for
155 * The probability used to range-decode each bit must be taken from a table, of
156 * which one instance must exist for each distinct context, or "binary decision
157 * class", in which a range-decoded bit is needed. At each call of the range
158 * decoder, the appropriate probability must be obtained by indexing the
159 * appropriate probability table with the last 4 (in the context disambiguating
160 * literals from matches), 5 (in the context disambiguating LZ matches from
161 * delta matches), or 6 (in all other contexts) bits recently range-decoded in
162 * that context, ordered such that the most recently decoded bit is the
163 * low-order bit of the index.
165 * Furthermore, each probability entry itself is variable, as its value must be
166 * maintained as n/64 where n is the number of 0 bits in the most recently
167 * decoded 64 bits with that same entry. This allows the compressed
168 * representation to adapt to the input and use fewer bits to represent the most
169 * likely data; note that LZMA uses a similar scheme. Initially, the most
170 * recently 64 decoded bits for each probability entry are assumed to be
171 * 0x0000000055555555 (high order to low order); therefore, all probabilities
172 * are initially 48/64. During the course of decoding, each probability may be
173 * updated to as low as 0/64 (as a result of reading many consecutive 1 bits
174 * with that entry) or as high as 64/64 (as a result of reading many consecutive
175 * 0 bits with that entry); however, probabilities of 0/64 and 64/64 cannot be
176 * used as-is but rather must be adjusted to 1/64 and 63/64, respectively,
177 * before being used for range decoding.
179 * Representations of the LZMS items themselves must be read from the backwards
180 * bitstream. For this, there are 5 different Huffman codes used:
182 * - The literal code, used for decoding literal bytes. Each of the 256
183 * symbols represents a literal byte. This code must be rebuilt whenever
184 * 1024 symbols have been decoded with it.
186 * - The LZ offset code, used for decoding the offsets of standard LZ77
187 * matches. Each symbol represents an offset slot, which corresponds to a
188 * base value and some number of extra bits which must be read and added to
189 * the base value to reconstitute the full offset. The number of symbols in
190 * this code is the number of offset slots needed to represent all possible
191 * offsets in the uncompressed block. This code must be rebuilt whenever
192 * 1024 symbols have been decoded with it.
194 * - The length code, used for decoding length symbols. Each of the 54 symbols
195 * represents a length slot, which corresponds to a base value and some
196 * number of extra bits which must be read and added to the base value to
197 * reconstitute the full length. This code must be rebuilt whenever 512
198 * symbols have been decoded with it.
200 * - The delta offset code, used for decoding the raw offsets of delta matches.
201 * Each symbol corresponds to an offset slot, which corresponds to a base
202 * value and some number of extra bits which must be read and added to the
203 * base value to reconstitute the full raw offset. The number of symbols in
204 * this code is equal to the number of symbols in the LZ offset code. This
205 * code must be rebuilt whenever 1024 symbols have been decoded with it.
207 * - The delta power code, used for decoding the powers of delta matches. Each
208 * of the 8 symbols corresponds to a power. This code must be rebuilt
209 * whenever 512 symbols have been decoded with it.
211 * Initially, each Huffman code must be built assuming that each symbol in that
212 * code has frequency 1. Following that, each code must be rebuilt each time a
213 * certain number of symbols, as noted above, has been decoded with it. The
214 * symbol frequencies for a code must be halved after each rebuild of that code;
215 * this makes the codes adapt to the more recent data.
217 * Like other compression formats such as XPRESS, LZX, and DEFLATE, the LZMS
218 * format requires that all Huffman codes be constructed in canonical form.
219 * This form requires that same-length codewords be lexicographically ordered
220 * the same way as the corresponding symbols and that all shorter codewords
221 * lexicographically precede longer codewords. Such a code can be constructed
222 * directly from codeword lengths.
224 * Even with the canonical code restriction, the same frequencies can be used to
225 * construct multiple valid Huffman codes. Therefore, the decompressor needs to
226 * construct the right one. Specifically, the LZMS format requires that the
227 * Huffman code be constructed as if the well-known priority queue algorithm is
228 * used and frequency ties are always broken in favor of leaf nodes.
230 * Codewords in LZMS are guaranteed to not exceed 15 bits. The format otherwise
231 * places no restrictions on codeword length. Therefore, the Huffman code
232 * construction algorithm that a correct LZMS decompressor uses need not
233 * implement length-limited code construction. But if it does (e.g. by virtue
234 * of being shared among multiple compression algorithms), the details of how it
235 * does so are unimportant, provided that the maximum codeword length parameter
236 * is set to at least 15 bits.
238 * After all LZMS items have been decoded, the data must be postprocessed to
239 * translate absolute address encoded in x86 instructions into their original
240 * relative addresses.
242 * Details omitted above can be found in the code. Note that in the absence of
243 * an official specification there is no guarantee that this decompressor
244 * handles all possible cases.
251 #include "wimlib/compress_common.h"
252 #include "wimlib/decompress_common.h"
253 #include "wimlib/decompressor_ops.h"
254 #include "wimlib/error.h"
255 #include "wimlib/lzms_common.h"
256 #include "wimlib/util.h"
258 /* The TABLEBITS values can be changed; they only affect decoding speed. */
259 #define LZMS_LITERAL_TABLEBITS 10
260 #define LZMS_LENGTH_TABLEBITS 10
261 #define LZMS_LZ_OFFSET_TABLEBITS 10
262 #define LZMS_DELTA_OFFSET_TABLEBITS 10
263 #define LZMS_DELTA_POWER_TABLEBITS 8
265 struct lzms_range_decoder {
267 /* The relevant part of the current range. Although the logical range
268 * for range decoding is a very large integer, only a small portion
269 * matters at any given time, and it can be normalized (shifted left)
270 * whenever it gets too small. */
273 /* The current position in the range encoded by the portion of the input
277 /* Pointer to the next little-endian 16-bit integer in the compressed
278 * input data (reading forwards). */
281 /* Pointer to the end of the compressed input data. */
285 typedef u64 bitbuf_t;
287 struct lzms_input_bitstream {
289 /* Holding variable for bits that have been read from the compressed
290 * data. The bit ordering is high to low. */
293 /* Number of bits currently held in @bitbuf. */
296 /* Pointer to the one past the next little-endian 16-bit integer in the
297 * compressed input data (reading backwards). */
300 /* Pointer to the beginning of the compressed input data. */
304 #define BITBUF_NBITS (8 * sizeof(bitbuf_t))
306 /* Bookkeeping information for an adaptive Huffman code */
307 struct lzms_huffman_rebuild_info {
308 unsigned num_syms_until_rebuild;
310 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_probabilites probs;
326 u16 literal_decode_table[(1 << LZMS_LITERAL_TABLEBITS) +
327 (2 * LZMS_NUM_LITERAL_SYMS)]
328 _aligned_attribute(DECODE_TABLE_ALIGNMENT);
329 u32 literal_freqs[LZMS_NUM_LITERAL_SYMS];
330 struct lzms_huffman_rebuild_info literal_rebuild_info;
332 u16 lz_offset_decode_table[(1 << LZMS_LZ_OFFSET_TABLEBITS) +
333 ( 2 * LZMS_MAX_NUM_OFFSET_SYMS)]
334 _aligned_attribute(DECODE_TABLE_ALIGNMENT);
335 u32 lz_offset_freqs[LZMS_MAX_NUM_OFFSET_SYMS];
336 struct lzms_huffman_rebuild_info lz_offset_rebuild_info;
338 u16 length_decode_table[(1 << LZMS_LENGTH_TABLEBITS) +
339 (2 * LZMS_NUM_LENGTH_SYMS)]
340 _aligned_attribute(DECODE_TABLE_ALIGNMENT);
341 u32 length_freqs[LZMS_NUM_LENGTH_SYMS];
342 struct lzms_huffman_rebuild_info length_rebuild_info;
344 u16 delta_offset_decode_table[(1 << LZMS_DELTA_OFFSET_TABLEBITS) +
345 (2 * LZMS_MAX_NUM_OFFSET_SYMS)]
346 _aligned_attribute(DECODE_TABLE_ALIGNMENT);
347 u32 delta_offset_freqs[LZMS_MAX_NUM_OFFSET_SYMS];
348 struct lzms_huffman_rebuild_info delta_offset_rebuild_info;
350 u16 delta_power_decode_table[(1 << LZMS_DELTA_POWER_TABLEBITS) +
351 (2 * LZMS_NUM_DELTA_POWER_SYMS)]
352 _aligned_attribute(DECODE_TABLE_ALIGNMENT);
353 u32 delta_power_freqs[LZMS_NUM_DELTA_POWER_SYMS];
354 struct lzms_huffman_rebuild_info delta_power_rebuild_info;
356 u32 codewords[LZMS_MAX_NUM_SYMS];
360 s32 last_target_usages[65536];
365 /* Initialize the input bitstream @is to read backwards from the compressed data
366 * buffer @in that is @count 16-bit integers long. */
368 lzms_input_bitstream_init(struct lzms_input_bitstream *is,
369 const le16 *in, size_t count)
373 is->next = in + count;
377 /* Ensure that at least @num_bits bits are in the bitbuffer variable.
378 * @num_bits cannot be more than 32. */
380 lzms_ensure_bits(struct lzms_input_bitstream *is, unsigned num_bits)
384 if (is->bitsleft >= num_bits)
387 avail = BITBUF_NBITS - is->bitsleft;
389 if (UNALIGNED_ACCESS_IS_FAST && CPU_IS_LITTLE_ENDIAN &&
390 WORDSIZE == 8 && likely((u8 *)is->next - (u8 *)is->begin >= 8))
392 is->next -= avail >> 4;
393 is->bitbuf |= load_u64_unaligned(is->next) << (avail & 15);
394 is->bitsleft += avail & ~15;
396 if (likely(is->next != is->begin))
397 is->bitbuf |= (bitbuf_t)le16_to_cpu(*--is->next)
399 if (likely(is->next != is->begin))
400 is->bitbuf |=(bitbuf_t)le16_to_cpu(*--is->next)
406 /* Get @num_bits bits from the bitbuffer variable. */
407 static inline bitbuf_t
408 lzms_peek_bits(struct lzms_input_bitstream *is, unsigned num_bits)
410 return (is->bitbuf >> 1) >> (BITBUF_NBITS - num_bits - 1);
413 /* Remove @num_bits bits from the bitbuffer variable. */
415 lzms_remove_bits(struct lzms_input_bitstream *is, unsigned num_bits)
417 is->bitbuf <<= num_bits;
418 is->bitsleft -= num_bits;
421 /* Remove and return @num_bits bits from the bitbuffer variable. */
422 static inline bitbuf_t
423 lzms_pop_bits(struct lzms_input_bitstream *is, unsigned num_bits)
425 bitbuf_t bits = lzms_peek_bits(is, num_bits);
426 lzms_remove_bits(is, num_bits);
430 /* Read @num_bits bits from the input bitstream. */
431 static inline bitbuf_t
432 lzms_read_bits(struct lzms_input_bitstream *is, unsigned num_bits)
434 lzms_ensure_bits(is, num_bits);
435 return lzms_pop_bits(is, num_bits);
438 /* Initialize the range decoder @rd to read forwards from the compressed data
439 * buffer @in that is @count 16-bit integers long. */
441 lzms_range_decoder_init(struct lzms_range_decoder *rd,
442 const le16 *in, size_t count)
444 rd->range = 0xffffffff;
445 rd->code = ((u32)le16_to_cpu(in[0]) << 16) | le16_to_cpu(in[1]);
447 rd->end = in + count;
451 * Decode a bit using the range coder. The current state specifies the
452 * probability entry to use. The state and probability entry will be updated
453 * based on the decoded bit.
456 lzms_decode_bit(struct lzms_range_decoder *rd, u32 *state_p, u32 num_states,
457 struct lzms_probability_entry *probs)
459 struct lzms_probability_entry *prob_entry;
463 /* Load the probability entry corresponding to the current state. */
464 prob_entry = &probs[*state_p];
466 /* Normalize if needed. */
467 if (rd->range <= 0xffff) {
470 if (likely(rd->next != rd->end))
471 rd->code |= le16_to_cpu(*rd->next++);
474 /* Get the probability (out of LZMS_PROBABILITY_DENOMINATOR) that the
476 prob = lzms_get_probability(prob_entry);
478 /* Based on the probability, calculate the bound between the 0-bit
479 * region and the 1-bit region of the range. */
480 bound = (rd->range >> LZMS_PROBABILITY_BITS) * prob;
482 if (rd->code < bound) {
483 /* Current code is in the 0-bit region of the range. */
486 /* Update the state and probability entry based on the decoded bit. */
487 *state_p = ((*state_p << 1) | 0) & (num_states - 1);
488 lzms_update_probability_entry(prob_entry, 0);
491 /* Current code is in the 1-bit region of the range. */
495 /* Update the state and probability entry based on the decoded bit. */
496 *state_p = ((*state_p << 1) | 1) & (num_states - 1);
497 lzms_update_probability_entry(prob_entry, 1);
503 lzms_build_huffman_code(struct lzms_huffman_rebuild_info *rebuild_info)
505 make_canonical_huffman_code(rebuild_info->num_syms,
506 LZMS_MAX_CODEWORD_LENGTH,
508 (u8 *)rebuild_info->decode_table,
509 rebuild_info->codewords);
511 make_huffman_decode_table(rebuild_info->decode_table,
512 rebuild_info->num_syms,
513 rebuild_info->table_bits,
514 (u8 *)rebuild_info->decode_table,
515 LZMS_MAX_CODEWORD_LENGTH);
517 rebuild_info->num_syms_until_rebuild = rebuild_info->rebuild_freq;
521 lzms_init_huffman_code(struct lzms_huffman_rebuild_info *rebuild_info,
522 unsigned num_syms, unsigned rebuild_freq,
523 u32 *codewords, u32 *freqs,
524 u16 *decode_table, unsigned table_bits)
526 rebuild_info->num_syms = num_syms;
527 rebuild_info->rebuild_freq = rebuild_freq;
528 rebuild_info->codewords = codewords;
529 rebuild_info->freqs = freqs;
530 rebuild_info->decode_table = decode_table;
531 rebuild_info->table_bits = table_bits;
532 lzms_init_symbol_frequencies(freqs, num_syms);
533 lzms_build_huffman_code(rebuild_info);
537 lzms_init_huffman_codes(struct lzms_decompressor *d, unsigned num_offset_slots)
539 lzms_init_huffman_code(&d->literal_rebuild_info,
540 LZMS_NUM_LITERAL_SYMS,
541 LZMS_LITERAL_CODE_REBUILD_FREQ,
544 d->literal_decode_table,
545 LZMS_LITERAL_TABLEBITS);
547 lzms_init_huffman_code(&d->lz_offset_rebuild_info,
549 LZMS_LZ_OFFSET_CODE_REBUILD_FREQ,
552 d->lz_offset_decode_table,
553 LZMS_LZ_OFFSET_TABLEBITS);
555 lzms_init_huffman_code(&d->length_rebuild_info,
556 LZMS_NUM_LENGTH_SYMS,
557 LZMS_LENGTH_CODE_REBUILD_FREQ,
560 d->length_decode_table,
561 LZMS_LENGTH_TABLEBITS);
563 lzms_init_huffman_code(&d->delta_offset_rebuild_info,
565 LZMS_DELTA_OFFSET_CODE_REBUILD_FREQ,
567 d->delta_offset_freqs,
568 d->delta_offset_decode_table,
569 LZMS_DELTA_OFFSET_TABLEBITS);
571 lzms_init_huffman_code(&d->delta_power_rebuild_info,
572 LZMS_NUM_DELTA_POWER_SYMS,
573 LZMS_DELTA_POWER_CODE_REBUILD_FREQ,
575 d->delta_power_freqs,
576 d->delta_power_decode_table,
577 LZMS_DELTA_POWER_TABLEBITS);
581 lzms_rebuild_huffman_code(struct lzms_huffman_rebuild_info *rebuild_info)
583 lzms_build_huffman_code(rebuild_info);
584 lzms_dilute_symbol_frequencies(rebuild_info->freqs, rebuild_info->num_syms);
587 static inline unsigned
588 lzms_decode_huffman_symbol(struct lzms_input_bitstream *is, u16 decode_table[],
589 unsigned table_bits, u32 freqs[],
590 struct lzms_huffman_rebuild_info *rebuild_info)
596 lzms_ensure_bits(is, LZMS_MAX_CODEWORD_LENGTH);
598 /* Index the decode table by the next table_bits bits of the input. */
599 key_bits = lzms_peek_bits(is, table_bits);
600 entry = decode_table[key_bits];
601 if (likely(entry < 0xC000)) {
602 /* Fast case: The decode table directly provided the symbol and
603 * codeword length. The low 11 bits are the symbol, and the
604 * high 5 bits are the codeword length. */
605 lzms_remove_bits(is, entry >> 11);
608 /* Slow case: The codeword for the symbol is longer than
609 * table_bits, so the symbol does not have an entry directly in
610 * the first (1 << table_bits) entries of the decode table.
611 * Traverse the appropriate binary tree bit-by-bit in order to
612 * decode the symbol. */
613 lzms_remove_bits(is, table_bits);
615 key_bits = (entry & 0x3FFF) + lzms_pop_bits(is, 1);
616 } while ((entry = decode_table[key_bits]) >= 0xC000);
621 if (--rebuild_info->num_syms_until_rebuild == 0)
622 lzms_rebuild_huffman_code(rebuild_info);
626 static inline unsigned
627 lzms_decode_literal(struct lzms_decompressor *d,
628 struct lzms_input_bitstream *is)
630 return lzms_decode_huffman_symbol(is,
631 d->literal_decode_table,
632 LZMS_LITERAL_TABLEBITS,
634 &d->literal_rebuild_info);
638 lzms_decode_lz_offset(struct lzms_decompressor *d,
639 struct lzms_input_bitstream *is)
641 unsigned slot = lzms_decode_huffman_symbol(is,
642 d->lz_offset_decode_table,
643 LZMS_LZ_OFFSET_TABLEBITS,
645 &d->lz_offset_rebuild_info);
646 return lzms_offset_slot_base[slot] +
647 lzms_read_bits(is, lzms_extra_offset_bits[slot]);
651 lzms_decode_length(struct lzms_decompressor *d,
652 struct lzms_input_bitstream *is)
654 unsigned slot = lzms_decode_huffman_symbol(is,
655 d->length_decode_table,
656 LZMS_LENGTH_TABLEBITS,
658 &d->length_rebuild_info);
659 u32 length = lzms_length_slot_base[slot];
660 unsigned num_extra_bits = lzms_extra_length_bits[slot];
661 /* Usually most lengths are short and have no extra bits. */
663 length += lzms_read_bits(is, num_extra_bits);
668 lzms_decode_delta_offset(struct lzms_decompressor *d,
669 struct lzms_input_bitstream *is)
671 unsigned slot = lzms_decode_huffman_symbol(is,
672 d->delta_offset_decode_table,
673 LZMS_DELTA_OFFSET_TABLEBITS,
674 d->delta_offset_freqs,
675 &d->delta_offset_rebuild_info);
676 return lzms_offset_slot_base[slot] +
677 lzms_read_bits(is, lzms_extra_offset_bits[slot]);
680 static inline unsigned
681 lzms_decode_delta_power(struct lzms_decompressor *d,
682 struct lzms_input_bitstream *is)
684 return lzms_decode_huffman_symbol(is,
685 d->delta_power_decode_table,
686 LZMS_DELTA_POWER_TABLEBITS,
687 d->delta_power_freqs,
688 &d->delta_power_rebuild_info);
692 lzms_create_decompressor(size_t max_bufsize, void **d_ret)
694 struct lzms_decompressor *d;
696 if (max_bufsize > LZMS_MAX_BUFFER_SIZE)
697 return WIMLIB_ERR_INVALID_PARAM;
699 d = ALIGNED_MALLOC(sizeof(struct lzms_decompressor),
700 DECODE_TABLE_ALIGNMENT);
702 return WIMLIB_ERR_NOMEM;
709 * Decompress @in_nbytes bytes of LZMS-compressed data at @in and write the
710 * uncompressed data, which had original size @out_nbytes, to @out. Return 0 if
711 * successful or -1 if the compressed data is invalid.
714 lzms_decompress(const void * const restrict in, const size_t in_nbytes,
715 void * const restrict out, const size_t out_nbytes,
716 void * const restrict _d)
718 struct lzms_decompressor *d = _d;
720 u8 * const out_end = out + out_nbytes;
721 struct lzms_range_decoder rd;
722 struct lzms_input_bitstream is;
724 /* LRU queues for match sources */
725 u32 recent_lz_offsets[LZMS_NUM_LZ_REPS + 1];
726 u64 recent_delta_pairs[LZMS_NUM_DELTA_REPS + 1];
727 u32 pending_lz_offset = 0;
728 u64 pending_delta_pair = 0;
729 const u8 *lz_offset_still_pending;
730 const u8 *delta_pair_still_pending;
732 /* States and probability entries for item type disambiguation */
737 u32 lz_rep_states[LZMS_NUM_LZ_REP_DECISIONS] = {};
738 u32 delta_rep_states[LZMS_NUM_DELTA_REP_DECISIONS] = {};
741 * Requirements on the compressed data:
743 * 1. LZMS-compressed data is a series of 16-bit integers, so the
744 * compressed data buffer cannot take up an odd number of bytes.
745 * 2. To prevent poor performance on some architectures, we require that
746 * the compressed data buffer is 2-byte aligned.
747 * 3. There must be at least 4 bytes of compressed data, since otherwise
748 * we cannot even initialize the range decoder.
750 if ((in_nbytes & 1) || ((uintptr_t)in & 1) || (in_nbytes < 4))
753 lzms_range_decoder_init(&rd, in, in_nbytes / sizeof(le16));
755 lzms_input_bitstream_init(&is, in, in_nbytes / sizeof(le16));
757 lzms_init_probabilities(&d->probs);
759 lzms_init_huffman_codes(d, lzms_get_num_offset_slots(out_nbytes));
761 for (int i = 0; i < LZMS_NUM_LZ_REPS + 1; i++)
762 recent_lz_offsets[i] = i + 1;
764 for (int i = 0; i < LZMS_NUM_DELTA_REPS + 1; i++)
765 recent_delta_pairs[i] = i + 1;
767 /* Main decode loop */
768 while (out_next != out_end) {
770 if (!lzms_decode_bit(&rd, &main_state,
771 LZMS_NUM_MAIN_PROBS, d->probs.main))
774 *out_next++ = lzms_decode_literal(d, &is);
776 } else if (!lzms_decode_bit(&rd, &match_state,
777 LZMS_NUM_MATCH_PROBS,
785 if (!lzms_decode_bit(&rd, &lz_state,
786 LZMS_NUM_LZ_PROBS, d->probs.lz))
788 /* Explicit offset */
789 offset = lzms_decode_lz_offset(d, &is);
793 if (pending_lz_offset != 0 &&
794 out_next != lz_offset_still_pending)
796 BUILD_BUG_ON(LZMS_NUM_LZ_REPS != 3);
797 recent_lz_offsets[3] = recent_lz_offsets[2];
798 recent_lz_offsets[2] = recent_lz_offsets[1];
799 recent_lz_offsets[1] = recent_lz_offsets[0];
800 recent_lz_offsets[0] = pending_lz_offset;
801 pending_lz_offset = 0;
804 BUILD_BUG_ON(LZMS_NUM_LZ_REPS != 3);
805 if (!lzms_decode_bit(&rd, &lz_rep_states[0],
806 LZMS_NUM_LZ_REP_PROBS,
809 offset = recent_lz_offsets[0];
810 recent_lz_offsets[0] = recent_lz_offsets[1];
811 recent_lz_offsets[1] = recent_lz_offsets[2];
812 recent_lz_offsets[2] = recent_lz_offsets[3];
813 } else if (!lzms_decode_bit(&rd, &lz_rep_states[1],
814 LZMS_NUM_LZ_REP_PROBS,
817 offset = recent_lz_offsets[1];
818 recent_lz_offsets[1] = recent_lz_offsets[2];
819 recent_lz_offsets[2] = recent_lz_offsets[3];
821 offset = recent_lz_offsets[2];
822 recent_lz_offsets[2] = recent_lz_offsets[3];
826 if (pending_lz_offset != 0) {
827 BUILD_BUG_ON(LZMS_NUM_LZ_REPS != 3);
828 recent_lz_offsets[3] = recent_lz_offsets[2];
829 recent_lz_offsets[2] = recent_lz_offsets[1];
830 recent_lz_offsets[1] = recent_lz_offsets[0];
831 recent_lz_offsets[0] = pending_lz_offset;
833 pending_lz_offset = offset;
835 length = lzms_decode_length(d, &is);
837 if (unlikely(length > out_end - out_next))
839 if (unlikely(offset > out_next - (u8 *)out))
842 lz_copy(out_next, length, offset, out_end, LZMS_MIN_MATCH_LENGTH);
845 lz_offset_still_pending = out_next;
849 /* (See beginning of file for more information.) */
858 if (!lzms_decode_bit(&rd, &delta_state,
859 LZMS_NUM_DELTA_PROBS,
862 /* Explicit offset */
863 power = lzms_decode_delta_power(d, &is);
864 raw_offset = lzms_decode_delta_offset(d, &is);
869 if (pending_delta_pair != 0 &&
870 out_next != delta_pair_still_pending)
872 BUILD_BUG_ON(LZMS_NUM_DELTA_REPS != 3);
873 recent_delta_pairs[3] = recent_delta_pairs[2];
874 recent_delta_pairs[2] = recent_delta_pairs[1];
875 recent_delta_pairs[1] = recent_delta_pairs[0];
876 recent_delta_pairs[0] = pending_delta_pair;
877 pending_delta_pair = 0;
880 BUILD_BUG_ON(LZMS_NUM_DELTA_REPS != 3);
881 if (!lzms_decode_bit(&rd, &delta_rep_states[0],
882 LZMS_NUM_DELTA_REP_PROBS,
883 d->probs.delta_rep[0]))
885 val = recent_delta_pairs[0];
886 recent_delta_pairs[0] = recent_delta_pairs[1];
887 recent_delta_pairs[1] = recent_delta_pairs[2];
888 recent_delta_pairs[2] = recent_delta_pairs[3];
889 } else if (!lzms_decode_bit(&rd, &delta_rep_states[1],
890 LZMS_NUM_DELTA_REP_PROBS,
891 d->probs.delta_rep[1]))
893 val = recent_delta_pairs[1];
894 recent_delta_pairs[1] = recent_delta_pairs[2];
895 recent_delta_pairs[2] = recent_delta_pairs[3];
897 val = recent_delta_pairs[2];
898 recent_delta_pairs[2] = recent_delta_pairs[3];
901 raw_offset = (u32)val;
904 if (pending_delta_pair != 0) {
905 BUILD_BUG_ON(LZMS_NUM_DELTA_REPS != 3);
906 recent_delta_pairs[3] = recent_delta_pairs[2];
907 recent_delta_pairs[2] = recent_delta_pairs[1];
908 recent_delta_pairs[1] = recent_delta_pairs[0];
909 recent_delta_pairs[0] = pending_delta_pair;
911 pending_delta_pair = raw_offset | ((u64)power << 32);
913 length = lzms_decode_length(d, &is);
915 span = (u32)1 << power;
916 offset = raw_offset << power;
918 /* raw_offset<<power overflows? */
919 if (unlikely(offset >> power != raw_offset))
922 /* offset+span overflows? */
923 if (unlikely(offset + span < offset))
926 /* buffer underrun? */
927 if (unlikely(offset + span > out_next - (u8 *)out))
930 /* buffer overrun? */
931 if (unlikely(length > out_end - out_next))
934 matchptr = out_next - offset;
936 *out_next = *matchptr + *(out_next - span) -
942 delta_pair_still_pending = out_next;
946 lzms_x86_filter(out, out_nbytes, d->last_target_usages, true);
951 lzms_free_decompressor(void *_d)
953 struct lzms_decompressor *d = _d;
958 const struct decompressor_ops lzms_decompressor_ops = {
959 .create_decompressor = lzms_create_decompressor,
960 .decompress = lzms_decompress,
961 .free_decompressor = lzms_free_decompressor,