4 * A decompressor for the LZX compression format, as used in WIM files.
8 * Copyright (C) 2012-2016 Eric Biggers
10 * This file is free software; you can redistribute it and/or modify it under
11 * the terms of the GNU Lesser General Public License as published by the Free
12 * Software Foundation; either version 3 of the License, or (at your option) any
15 * This file is distributed in the hope that it will be useful, but WITHOUT
16 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
17 * FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
20 * You should have received a copy of the GNU Lesser General Public License
21 * along with this file; if not, see http://www.gnu.org/licenses/.
25 * LZX is an LZ77 and Huffman-code based compression format that has many
26 * similarities to DEFLATE (the format used by zlib/gzip). The compression
27 * ratio is as good or better than DEFLATE. See lzx_compress.c for a format
28 * overview, and see https://en.wikipedia.org/wiki/LZX_(algorithm) for a
29 * historical overview. Here I make some pragmatic notes.
31 * The old specification for LZX is the document "Microsoft LZX Data Compression
32 * Format" (1997). It defines the LZX format as used in cabinet files. Allowed
33 * window sizes are 2^n where 15 <= n <= 21. However, this document contains
34 * several errors, so don't read too much into it...
36 * The new specification for LZX is the document "[MS-PATCH]: LZX DELTA
37 * Compression and Decompression" (2014). It defines the LZX format as used by
38 * Microsoft's binary patcher. It corrects several errors in the 1997 document
39 * and extends the format in several ways --- namely, optional reference data,
40 * up to 2^25 byte windows, and longer match lengths.
42 * WIM files use a more restricted form of LZX. No LZX DELTA extensions are
43 * present, the window is not "sliding", E8 preprocessing is done
44 * unconditionally with a fixed file size, and the maximum window size is always
45 * 2^15 bytes (equal to the size of each "chunk" in a compressed WIM resource).
46 * This code is primarily intended to implement this form of LZX. But although
47 * not compatible with WIMGAPI, this code also supports maximum window sizes up
50 * TODO: Add support for window sizes up to 2^25 bytes.
59 #include "wimlib/decompressor_ops.h"
60 #include "wimlib/decompress_common.h"
61 #include "wimlib/error.h"
62 #include "wimlib/lzx_common.h"
63 #include "wimlib/util.h"
65 /* These values are chosen for fast decompression. */
66 #define LZX_MAINCODE_TABLEBITS 11
67 #define LZX_LENCODE_TABLEBITS 9
68 #define LZX_PRECODE_TABLEBITS 6
69 #define LZX_ALIGNEDCODE_TABLEBITS 7
71 #define LZX_READ_LENS_MAX_OVERRUN 50
73 struct lzx_decompressor {
75 DECODE_TABLE(maincode_decode_table, LZX_MAINCODE_MAX_NUM_SYMBOLS,
76 LZX_MAINCODE_TABLEBITS, LZX_MAX_MAIN_CODEWORD_LEN);
77 u8 maincode_lens[LZX_MAINCODE_MAX_NUM_SYMBOLS + LZX_READ_LENS_MAX_OVERRUN];
79 DECODE_TABLE(lencode_decode_table, LZX_LENCODE_NUM_SYMBOLS,
80 LZX_LENCODE_TABLEBITS, LZX_MAX_LEN_CODEWORD_LEN);
81 u8 lencode_lens[LZX_LENCODE_NUM_SYMBOLS + LZX_READ_LENS_MAX_OVERRUN];
84 DECODE_TABLE(alignedcode_decode_table, LZX_ALIGNEDCODE_NUM_SYMBOLS,
85 LZX_ALIGNEDCODE_TABLEBITS, LZX_MAX_ALIGNED_CODEWORD_LEN);
86 u8 alignedcode_lens[LZX_ALIGNEDCODE_NUM_SYMBOLS];
90 DECODE_TABLE(precode_decode_table, LZX_PRECODE_NUM_SYMBOLS,
91 LZX_PRECODE_TABLEBITS, LZX_MAX_PRE_CODEWORD_LEN);
92 u8 precode_lens[LZX_PRECODE_NUM_SYMBOLS];
93 u8 extra_offset_bits[LZX_MAX_OFFSET_SLOTS];
97 DECODE_TABLE_WORKING_SPACE(maincode_working_space,
98 LZX_MAINCODE_MAX_NUM_SYMBOLS,
99 LZX_MAX_MAIN_CODEWORD_LEN);
100 DECODE_TABLE_WORKING_SPACE(lencode_working_space,
101 LZX_LENCODE_NUM_SYMBOLS,
102 LZX_MAX_LEN_CODEWORD_LEN);
103 DECODE_TABLE_WORKING_SPACE(alignedcode_working_space,
104 LZX_ALIGNEDCODE_NUM_SYMBOLS,
105 LZX_MAX_ALIGNED_CODEWORD_LEN);
106 DECODE_TABLE_WORKING_SPACE(precode_working_space,
107 LZX_PRECODE_NUM_SYMBOLS,
108 LZX_MAX_PRE_CODEWORD_LEN);
111 unsigned window_order;
112 unsigned num_main_syms;
114 /* Like lzx_extra_offset_bits[], but does not include the entropy-coded
115 * bits of aligned offset blocks */
116 u8 extra_offset_bits_minus_aligned[LZX_MAX_OFFSET_SLOTS];
118 } _aligned_attribute(DECODE_TABLE_ALIGNMENT);
120 /* Read a Huffman-encoded symbol using the precode. */
121 static forceinline unsigned
122 read_presym(const struct lzx_decompressor *d, struct input_bitstream *is)
124 return read_huffsym(is, d->precode_decode_table,
125 LZX_PRECODE_TABLEBITS, LZX_MAX_PRE_CODEWORD_LEN);
128 /* Read a Huffman-encoded symbol using the main code. */
129 static forceinline unsigned
130 read_mainsym(const struct lzx_decompressor *d, struct input_bitstream *is)
132 return read_huffsym(is, d->maincode_decode_table,
133 LZX_MAINCODE_TABLEBITS, LZX_MAX_MAIN_CODEWORD_LEN);
136 /* Read a Huffman-encoded symbol using the length code. */
137 static forceinline unsigned
138 read_lensym(const struct lzx_decompressor *d, struct input_bitstream *is)
140 return read_huffsym(is, d->lencode_decode_table,
141 LZX_LENCODE_TABLEBITS, LZX_MAX_LEN_CODEWORD_LEN);
144 /* Read a Huffman-encoded symbol using the aligned offset code. */
145 static forceinline unsigned
146 read_alignedsym(const struct lzx_decompressor *d, struct input_bitstream *is)
148 return read_huffsym(is, d->alignedcode_decode_table,
149 LZX_ALIGNEDCODE_TABLEBITS, LZX_MAX_ALIGNED_CODEWORD_LEN);
153 * Read a precode from the compressed input bitstream, then use it to decode
154 * @num_lens codeword length values and write them to @lens.
157 lzx_read_codeword_lens(struct lzx_decompressor *d, struct input_bitstream *is,
158 u8 *lens, unsigned num_lens)
161 u8 *lens_end = lens + num_lens;
163 /* Read the lengths of the precode codewords. These are stored
165 for (int i = 0; i < LZX_PRECODE_NUM_SYMBOLS; i++) {
167 bitstream_read_bits(is, LZX_PRECODE_ELEMENT_SIZE);
170 /* Build the decoding table for the precode. */
171 if (make_huffman_decode_table(d->precode_decode_table,
172 LZX_PRECODE_NUM_SYMBOLS,
173 LZX_PRECODE_TABLEBITS,
175 LZX_MAX_PRE_CODEWORD_LEN,
176 d->precode_working_space))
179 /* Decode the codeword lengths. */
184 /* Read the next precode symbol. */
185 presym = read_presym(d, is);
187 /* Difference from old length */
188 len = *len_ptr - presym;
193 /* Special RLE values */
199 run_len = 4 + bitstream_read_bits(is, 4);
201 } else if (presym == 18) {
202 /* Longer run of 0's */
203 run_len = 20 + bitstream_read_bits(is, 5);
206 /* Run of identical lengths */
207 run_len = 4 + bitstream_read_bits(is, 1);
208 presym = read_presym(d, is);
209 if (unlikely(presym > 17))
211 len = *len_ptr - presym;
220 * The worst case overrun is when presym == 18,
221 * run_len == 20 + 31, and only 1 length was remaining.
222 * So LZX_READ_LENS_MAX_OVERRUN == 50.
224 * Overrun while reading the first half of maincode_lens
225 * can corrupt the previous values in the second half.
226 * This doesn't really matter because the resulting
227 * lengths will still be in range, and data that
228 * generates overruns is invalid anyway.
231 } while (len_ptr < lens_end);
237 * Read the header of an LZX block. For all block types, the block type and
238 * size is saved in *block_type_ret and *block_size_ret, respectively. For
239 * compressed blocks, the codeword lengths are also saved. For uncompressed
240 * blocks, the recent offsets queue is also updated.
243 lzx_read_block_header(struct lzx_decompressor *d, struct input_bitstream *is,
244 u32 recent_offsets[], int *block_type_ret,
250 bitstream_ensure_bits(is, 4);
252 /* Read the block type. */
253 block_type = bitstream_pop_bits(is, 3);
255 /* Read the block size. */
256 if (bitstream_pop_bits(is, 1)) {
257 block_size = LZX_DEFAULT_BLOCK_SIZE;
259 block_size = bitstream_read_bits(is, 16);
260 if (d->window_order >= 16) {
262 block_size |= bitstream_read_bits(is, 8);
266 switch (block_type) {
268 case LZX_BLOCKTYPE_ALIGNED:
270 /* Read the aligned offset codeword lengths. */
272 for (int i = 0; i < LZX_ALIGNEDCODE_NUM_SYMBOLS; i++) {
273 d->alignedcode_lens[i] =
274 bitstream_read_bits(is,
275 LZX_ALIGNEDCODE_ELEMENT_SIZE);
278 /* Fall though, since the rest of the header for aligned offset
279 * blocks is the same as that for verbatim blocks. */
281 case LZX_BLOCKTYPE_VERBATIM:
283 /* Read the main codeword lengths, which are divided into two
284 * parts: literal symbols and match headers. */
286 if (lzx_read_codeword_lens(d, is, d->maincode_lens,
290 if (lzx_read_codeword_lens(d, is, d->maincode_lens + LZX_NUM_CHARS,
291 d->num_main_syms - LZX_NUM_CHARS))
295 /* Read the length codeword lengths. */
297 if (lzx_read_codeword_lens(d, is, d->lencode_lens,
298 LZX_LENCODE_NUM_SYMBOLS))
303 case LZX_BLOCKTYPE_UNCOMPRESSED:
305 * The header of an uncompressed block contains new values for
306 * the recent offsets queue, starting on the next 16-bit
307 * boundary in the bitstream. Careful: if the stream is
308 * *already* aligned, the correct thing to do is to throw away
309 * the next 16 bits (this is probably a mistake in the format).
311 bitstream_ensure_bits(is, 1);
313 recent_offsets[0] = bitstream_read_u32(is);
314 recent_offsets[1] = bitstream_read_u32(is);
315 recent_offsets[2] = bitstream_read_u32(is);
317 /* Offsets of 0 are invalid. */
318 if (recent_offsets[0] == 0 || recent_offsets[1] == 0 ||
319 recent_offsets[2] == 0)
324 /* Unrecognized block type. */
328 *block_type_ret = block_type;
329 *block_size_ret = block_size;
333 /* Decompress a block of LZX-compressed data. */
335 lzx_decompress_block(struct lzx_decompressor *d, struct input_bitstream *_is,
336 int block_type, u32 block_size,
337 u8 * const out_begin, u8 *out_next, u32 recent_offsets[])
340 * Redeclare the input bitstream on the stack. This shouldn't be
341 * needed, but it can improve the main loop's performance significantly
342 * with both gcc and clang, apparently because the compiler otherwise
343 * gets confused and doesn't properly allocate registers for
344 * 'is->bitbuf' et al. and/or thinks 'is->next' may point into 'is'.
346 struct input_bitstream is_onstack = *_is;
347 struct input_bitstream *is = &is_onstack;
348 u8 * const block_end = out_next + block_size;
349 unsigned min_aligned_offset_slot;
352 * Build the Huffman decode tables. We always need to build the main
353 * and length decode tables. For aligned blocks we additionally need to
354 * build the aligned offset decode table.
357 if (make_huffman_decode_table(d->maincode_decode_table,
359 LZX_MAINCODE_TABLEBITS,
361 LZX_MAX_MAIN_CODEWORD_LEN,
362 d->maincode_working_space))
365 if (make_huffman_decode_table(d->lencode_decode_table,
366 LZX_LENCODE_NUM_SYMBOLS,
367 LZX_LENCODE_TABLEBITS,
369 LZX_MAX_LEN_CODEWORD_LEN,
370 d->lencode_working_space))
373 if (block_type == LZX_BLOCKTYPE_ALIGNED) {
374 if (make_huffman_decode_table(d->alignedcode_decode_table,
375 LZX_ALIGNEDCODE_NUM_SYMBOLS,
376 LZX_ALIGNEDCODE_TABLEBITS,
378 LZX_MAX_ALIGNED_CODEWORD_LEN,
379 d->alignedcode_working_space))
381 min_aligned_offset_slot = LZX_MIN_ALIGNED_OFFSET_SLOT;
382 memcpy(d->extra_offset_bits, d->extra_offset_bits_minus_aligned,
383 sizeof(lzx_extra_offset_bits));
385 min_aligned_offset_slot = LZX_MAX_OFFSET_SLOTS;
386 memcpy(d->extra_offset_bits, lzx_extra_offset_bits,
387 sizeof(lzx_extra_offset_bits));
390 /* Decode the literals and matches. */
396 unsigned offset_slot;
398 mainsym = read_mainsym(d, is);
399 if (mainsym < LZX_NUM_CHARS) {
401 *out_next++ = mainsym;
407 /* Decode the length header and offset slot. */
408 STATIC_ASSERT(LZX_NUM_CHARS % LZX_NUM_LEN_HEADERS == 0);
409 length = mainsym % LZX_NUM_LEN_HEADERS;
410 offset_slot = (mainsym - LZX_NUM_CHARS) / LZX_NUM_LEN_HEADERS;
412 /* If needed, read a length symbol to decode the full length. */
413 if (length == LZX_NUM_PRIMARY_LENS)
414 length += read_lensym(d, is);
415 length += LZX_MIN_MATCH_LEN;
417 if (offset_slot < LZX_NUM_RECENT_OFFSETS) {
420 /* Note: This isn't a real LRU queue, since using the R2
421 * offset doesn't bump the R1 offset down to R2. */
422 offset = recent_offsets[offset_slot];
423 recent_offsets[offset_slot] = recent_offsets[0];
425 /* Explicit offset */
426 offset = bitstream_read_bits(is, d->extra_offset_bits[offset_slot]);
427 if (offset_slot >= min_aligned_offset_slot) {
428 offset = (offset << LZX_NUM_ALIGNED_OFFSET_BITS) |
429 read_alignedsym(d, is);
431 offset += lzx_offset_slot_base[offset_slot];
433 /* Update the match offset LRU queue. */
434 STATIC_ASSERT(LZX_NUM_RECENT_OFFSETS == 3);
435 recent_offsets[2] = recent_offsets[1];
436 recent_offsets[1] = recent_offsets[0];
438 recent_offsets[0] = offset;
440 /* Validate the match and copy it to the current position. */
441 if (unlikely(lz_copy(length, offset, out_begin,
442 out_next, block_end, LZX_MIN_MATCH_LEN)))
445 } while (out_next != block_end);
452 lzx_decompress(const void *restrict compressed_data, size_t compressed_size,
453 void *restrict uncompressed_data, size_t uncompressed_size,
456 struct lzx_decompressor *d = _d;
457 u8 * const out_begin = uncompressed_data;
458 u8 *out_next = out_begin;
459 u8 * const out_end = out_begin + uncompressed_size;
460 struct input_bitstream is;
461 STATIC_ASSERT(LZX_NUM_RECENT_OFFSETS == 3);
462 u32 recent_offsets[LZX_NUM_RECENT_OFFSETS] = {1, 1, 1};
463 unsigned may_have_e8_byte = 0;
465 init_input_bitstream(&is, compressed_data, compressed_size);
467 /* Codeword lengths begin as all 0's for delta encoding purposes. */
468 memset(d->maincode_lens, 0, d->num_main_syms);
469 memset(d->lencode_lens, 0, LZX_LENCODE_NUM_SYMBOLS);
471 /* Decompress blocks until we have all the uncompressed data. */
473 while (out_next != out_end) {
477 if (lzx_read_block_header(d, &is, recent_offsets,
478 &block_type, &block_size))
481 if (block_size < 1 || block_size > out_end - out_next)
484 if (likely(block_type != LZX_BLOCKTYPE_UNCOMPRESSED)) {
486 /* Compressed block */
487 if (lzx_decompress_block(d, &is, block_type, block_size,
492 /* If the first E8 byte was in this block, then it must
493 * have been encoded as a literal using mainsym E8. */
494 may_have_e8_byte |= d->maincode_lens[0xE8];
497 /* Uncompressed block */
498 if (bitstream_read_bytes(&is, out_next, block_size))
501 /* Re-align the bitstream if needed. */
503 bitstream_read_byte(&is);
505 /* There may have been an E8 byte in the block. */
506 may_have_e8_byte = 1;
508 out_next += block_size;
511 /* Postprocess the data unless it cannot possibly contain E8 bytes. */
512 if (may_have_e8_byte)
513 lzx_postprocess(uncompressed_data, uncompressed_size);
519 lzx_create_decompressor(size_t max_block_size, void **d_ret)
521 unsigned window_order;
522 struct lzx_decompressor *d;
524 window_order = lzx_get_window_order(max_block_size);
525 if (window_order == 0)
526 return WIMLIB_ERR_INVALID_PARAM;
528 d = ALIGNED_MALLOC(sizeof(*d), DECODE_TABLE_ALIGNMENT);
530 return WIMLIB_ERR_NOMEM;
532 d->window_order = window_order;
533 d->num_main_syms = lzx_get_num_main_syms(window_order);
535 /* Initialize 'd->extra_offset_bits_minus_aligned'. */
536 STATIC_ASSERT(sizeof(d->extra_offset_bits_minus_aligned) ==
537 sizeof(lzx_extra_offset_bits));
538 STATIC_ASSERT(sizeof(d->extra_offset_bits) ==
539 sizeof(lzx_extra_offset_bits));
540 memcpy(d->extra_offset_bits_minus_aligned, lzx_extra_offset_bits,
541 sizeof(lzx_extra_offset_bits));
542 for (unsigned offset_slot = LZX_MIN_ALIGNED_OFFSET_SLOT;
543 offset_slot < LZX_MAX_OFFSET_SLOTS; offset_slot++)
545 d->extra_offset_bits_minus_aligned[offset_slot] -=
546 LZX_NUM_ALIGNED_OFFSET_BITS;
554 lzx_free_decompressor(void *_d)
559 const struct decompressor_ops lzx_decompressor_ops = {
560 .create_decompressor = lzx_create_decompressor,
561 .decompress = lzx_decompress,
562 .free_decompressor = lzx_free_decompressor,