4 * LZX compression routines.
6 * This code was originally based on code written by Matthew T. Russotto
9 * Copyright (C) 2002 Matthew T. Russotto
10 * Copyright (C) 2012 Eric Biggers
12 * wimlib - Library for working with WIM files
14 * This library is free software; you can redistribute it and/or modify it under
15 * the terms of the GNU Lesser General Public License as published by the Free
16 * Software Foundation; either version 2.1 of the License, or (at your option) any
19 * This library is distributed in the hope that it will be useful, but WITHOUT ANY
20 * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
21 * PARTICULAR PURPOSE. See the GNU Lesser General Public License for more details.
23 * You should have received a copy of the GNU Lesser General Public License along
24 * with this library; if not, write to the Free Software Foundation, Inc., 59
25 * Temple Place, Suite 330, Boston, MA 02111-1307 USA
31 * This file provides lzx_compress(), a function to compress an in-memory buffer
32 * of data using LZX compression, as used in the WIM file format.
34 * There is no sliding window, as for the compressed chunks in WIM resources,
35 * the window is always the length of the input.
37 * The basic algorithm should be familiar if you are familiar with Huffman trees
38 * and with other LZ77-based formats such as DEFLATE. Otherwise it can be quite
39 * tricky to understand. Basically it is the following:
41 * - Preprocess the input data (LZX-specific)
42 * - Go through the input data and determine matches. This part is based on
43 * code from zlib, and a hash table of 3-character strings is used to
44 * accelerate the process of finding matches.
45 * - Build the Huffman trees based on the frequencies of symbols determined
46 * while recording matches.
47 * - Output the block header, including the Huffman trees; then output the
48 * compressed stream of matches and literal characters.
60 /* Structure to contain the Huffman codes for the main, length, and aligned
63 u16 main_codewords[LZX_MAINTREE_NUM_SYMBOLS];
64 u8 main_lens[LZX_MAINTREE_NUM_SYMBOLS];
66 u16 len_codewords[LZX_LENTREE_NUM_SYMBOLS];
67 u8 len_lens[LZX_LENTREE_NUM_SYMBOLS];
69 u16 aligned_codewords[LZX_ALIGNEDTREE_NUM_SYMBOLS];
70 u8 aligned_lens[LZX_ALIGNEDTREE_NUM_SYMBOLS];
73 struct lzx_freq_tables {
74 u32 main_freq_table[LZX_MAINTREE_NUM_SYMBOLS];
75 u32 len_freq_table[LZX_LENTREE_NUM_SYMBOLS];
76 u32 aligned_freq_table[LZX_ALIGNEDTREE_NUM_SYMBOLS];
82 /* Returns the position slot that corresponds to a given formatted offset. This
83 * means searching the lzx_position_base array to find what slot contains a
84 * position base that is less than or equal to formatted_offset, where the next
85 * slot contains a position base that is greater than or equal to
86 * formatted_offset. */
87 static uint lzx_get_position_slot(uint formatted_offset)
93 /* Calculate position base using binary search of table; if log2 can be
94 * done in hardware, approximation might work;
95 * trunc(log2(formatted_offset*formatted_offset)) gets either the proper
96 * position slot or the next one, except for slots 0, 1, and 39-49
98 * Slots 0-1 are handled by the R0-R1 procedures
100 * Slots 36-49 (formatted_offset >= 262144) can be found by
101 * (formatted_offset/131072) + 34 == (formatted_offset >> 17) + 34;
103 if (formatted_offset >= 262144) {
104 return (formatted_offset >> 17) + 34;
107 right = LZX_NUM_POSITION_SLOTS - 1;
109 mid = (left + right) >> 1;
110 if ((lzx_position_base[mid] <= formatted_offset) &&
111 lzx_position_base[mid + 1] > formatted_offset) {
114 if (formatted_offset > lzx_position_base[mid])
123 static u32 lzx_record_literal(u8 literal, void *__main_freq_tab)
125 u32 *main_freq_tab = __main_freq_tab;
126 main_freq_tab[literal]++;
130 /* Constructs a match from an offset and a length, and updates the LRU queue
131 * and the frequency of symbols in the main, length, and aligned offset
132 * alphabets. The return value is a 32-bit integer that, if the high bit is
133 * set, contains the match length, the position slot, and the position footer
135 static u32 lzx_record_match(uint match_offset, uint match_len,
136 void *__freq_tabs, void *__queue)
138 struct lzx_freq_tables *freq_tabs = __freq_tabs;
139 struct lru_queue *queue = __queue;
140 uint formatted_offset;
142 uint position_footer = 0;
148 wimlib_assert(match_len >= LZX_MIN_MATCH && match_len <= LZX_MAX_MATCH);
151 if (match_offset == queue->R0) {
152 formatted_offset = 0;
154 } else if (match_offset == queue->R1) {
155 swap(queue->R0, queue->R1);
156 formatted_offset = 1;
158 } else if (match_offset == queue->R2) {
159 swap(queue->R0, queue->R2);
160 formatted_offset = 2;
163 /* Not a repeated offset. */
165 formatted_offset = match_offset + LZX_MIN_MATCH;
167 queue->R2 = queue->R1;
168 queue->R1 = queue->R0;
169 queue->R0 = match_offset;
171 position_slot = lzx_get_position_slot(formatted_offset);
173 /* Just the extra bits of the formatted offset. */
174 position_footer = ((1UL << lzx_extra_bits[position_slot]) - 1) &
178 /* (match length - 2) = 8 bits */
179 /* position_slot = 6 bits */
180 /* position_footer = 17 bits */
181 /* total = 31 bits */
182 /* plus one to say whether it's a literal or not */
184 match = 0x80000000 | /* bit 31 in intelligent bit ordering */
185 (position_slot << 25) | /* bits 30-25 */
186 (position_footer << 8) | /* bits 8-24 */
187 (match_len - LZX_MIN_MATCH); /* bits 0-7 */
189 /* Update the frequency for the main tree, the length tree (only if a
190 * length symbol is to be output), and the aligned tree (only if an
191 * aligned symbol is to be output.) */
192 if (match_len < (LZX_NUM_PRIMARY_LENS + LZX_MIN_MATCH)) {
193 len_header = match_len - LZX_MIN_MATCH;
195 len_header = LZX_NUM_PRIMARY_LENS;
196 len_footer = match_len - (LZX_NUM_PRIMARY_LENS + LZX_MIN_MATCH);
197 freq_tabs->len_freq_table[len_footer]++;
199 len_pos_header = (position_slot << 3) | len_header;
201 wimlib_assert(len_pos_header < LZX_MAINTREE_NUM_SYMBOLS - LZX_NUM_CHARS);
203 freq_tabs->main_freq_table[len_pos_header + LZX_NUM_CHARS]++;
205 if (lzx_extra_bits[position_slot] >= 3)
206 freq_tabs->aligned_freq_table[position_footer & 7]++;
212 * Writes a compressed literal match to the output.
214 * @out: The output bitstream.
215 * @block_type: The type of the block (LZX_BLOCKTYPE_ALIGNED or LZX_BLOCKTYPE_VERBATIM)
216 * @match: The match, encoded as a 32-bit number.
217 * @codes: Pointer to a structure that contains the codewords for the
218 * main, length, and aligned offset Huffman codes.
220 static int lzx_write_match(struct output_bitstream *out, int block_type,
221 u32 match, const struct lzx_codes *codes)
223 /* low 8 bits are the match length minus 2 */
224 uint match_len_minus_2 = match & 0xff;
225 /* Next 17 bits are the position footer */
226 uint position_footer = (match >> 8) & 0x1ffff; /* 17 bits */
227 /* Next 6 bits are the position slot. */
228 uint position_slot = (match >> 25) & 0x3f; /* 6 bits */
238 /* If the match length is less than MIN_MATCH (= 2) +
239 * NUM_PRIMARY_LENS (= 7), the length header contains
240 * the match length minus MIN_MATCH, and there is no
243 * Otherwise, the length header contains
244 * NUM_PRIMARY_LENS, and the length footer contains
245 * the match length minus NUM_PRIMARY_LENS minus
247 if (match_len_minus_2 < LZX_NUM_PRIMARY_LENS) {
248 len_header = match_len_minus_2;
249 /* No length footer-- mark it with a special
251 len_footer = (uint)(-1);
253 len_header = LZX_NUM_PRIMARY_LENS;
254 len_footer = match_len_minus_2 - LZX_NUM_PRIMARY_LENS;
257 /* Combine the position slot with the length header into
258 * a single symbol that will be encoded with the main
260 len_pos_header = (position_slot << 3) | len_header;
262 /* The actual main symbol is offset by LZX_NUM_CHARS because
263 * values under LZX_NUM_CHARS are used to indicate a literal
264 * byte rather than a match. */
265 main_symbol = len_pos_header + LZX_NUM_CHARS;
267 /* Output main symbol. */
268 ret = bitstream_put_bits(out, codes->main_codewords[main_symbol],
269 codes->main_lens[main_symbol]);
273 /* If there is a length footer, output it using the
274 * length Huffman code. */
275 if (len_footer != (uint)(-1)) {
276 ret = bitstream_put_bits(out, codes->len_codewords[len_footer],
277 codes->len_lens[len_footer]);
282 wimlib_assert(position_slot < LZX_NUM_POSITION_SLOTS);
284 num_extra_bits = lzx_extra_bits[position_slot];
286 /* For aligned offset blocks with at least 3 extra bits, output the
287 * verbatim bits literally, then the aligned bits encoded using the
288 * aligned offset tree. Otherwise, only the verbatim bits need to be
290 if ((block_type == LZX_BLOCKTYPE_ALIGNED) && (num_extra_bits >= 3)) {
292 verbatim_bits = position_footer >> 3;
293 ret = bitstream_put_bits(out, verbatim_bits,
298 aligned_bits = (position_footer & 7);
299 ret = bitstream_put_bits(out,
300 codes->aligned_codewords[aligned_bits],
301 codes->aligned_lens[aligned_bits]);
305 /* verbatim bits is the same as the position
306 * footer, in this case. */
307 ret = bitstream_put_bits(out, position_footer, num_extra_bits);
315 * Writes all compressed literals in a block, both matches and literal bytes, to
316 * the output bitstream.
318 * @out: The output bitstream.
319 * @block_type: The type of the block (LZX_BLOCKTYPE_ALIGNED or LZX_BLOCKTYPE_VERBATIM)
320 * @match_tab[]: The array of matches that will be output. It has length
321 * of @num_compressed_literals.
322 * @num_compressed_literals: Number of compressed literals to be output.
323 * @codes: Pointer to a structure that contains the codewords for the
324 * main, length, and aligned offset Huffman codes.
326 static int lzx_write_compressed_literals(struct output_bitstream *ostream,
328 const u32 match_tab[],
329 uint num_compressed_literals,
330 const struct lzx_codes *codes)
336 for (i = 0; i < num_compressed_literals; i++) {
337 match = match_tab[i];
339 /* High bit of the match indicates whether the match is an
340 * actual match (1) or a literal uncompressed byte (0) */
341 if (match & 0x80000000) {
343 ret = lzx_write_match(ostream, block_type, match,
349 wimlib_assert(match < LZX_NUM_CHARS);
350 ret = bitstream_put_bits(ostream,
351 codes->main_codewords[match],
352 codes->main_lens[match]);
361 * Writes a compressed Huffman tree to the output, preceded by the pretree for
364 * The Huffman tree is represented in the output as a series of path lengths
365 * from which the canonical Huffman code can be reconstructed. The path lengths
366 * themselves are compressed using a separate Huffman code, the pretree, which
367 * consists of LZX_PRETREE_NUM_SYMBOLS (= 20) symbols that cover all possible code
368 * lengths, plus extra codes for repeated lengths. The path lengths of the
369 * pretree precede the path lengths of the larger code and are uncompressed,
370 * consisting of 20 entries of 4 bits each.
372 * @out: The bitstream for the compressed output.
373 * @lens: The code lengths for the Huffman tree, indexed by symbol.
374 * @num_symbols: The number of symbols in the code.
376 static int lzx_write_compressed_tree(struct output_bitstream *out,
380 /* Frequencies of the length symbols, including the RLE symbols (NOT the
381 * actual lengths themselves). */
382 uint pretree_freqs[LZX_PRETREE_NUM_SYMBOLS];
383 u8 pretree_lens[LZX_PRETREE_NUM_SYMBOLS];
384 u16 pretree_codewords[LZX_PRETREE_NUM_SYMBOLS];
385 u8 output_syms[num_symbols * 2];
386 uint output_syms_idx;
390 uint additional_bits;
394 ZERO_ARRAY(pretree_freqs);
396 /* Since the code word lengths use a form of RLE encoding, the goal here
397 * is to find each run of identical lengths when going through them in
398 * symbol order (including runs of length 1). For each run, as many
399 * lengths are encoded using RLE as possible, and the rest are output
402 * output_syms[] will be filled in with the length symbols that will be
403 * output, including RLE codes, not yet encoded using the pre-tree.
405 * cur_run_len keeps track of how many code word lengths are in the
406 * current run of identical lengths.
410 for (i = 1; i <= num_symbols; i++) {
412 if (i != num_symbols && lens[i] == lens[i - 1]) {
413 /* Still in a run--- keep going. */
418 /* Run ended! Check if it is a run of zeroes or a run of
421 /* The symbol that was repeated in the run--- not to be confused
422 * with the length *of* the run (cur_run_len) */
423 len_in_run = lens[i - 1];
425 if (len_in_run == 0) {
426 /* A run of 0's. Encode it in as few length
427 * codes as we can. */
429 /* The magic length 18 indicates a run of 20 + n zeroes,
430 * where n is an uncompressed literal 5-bit integer that
431 * follows the magic length. */
432 while (cur_run_len >= 20) {
434 additional_bits = min(cur_run_len - 20, 0x1f);
436 output_syms[output_syms_idx++] = 18;
437 output_syms[output_syms_idx++] = additional_bits;
438 cur_run_len -= 20 + additional_bits;
441 /* The magic length 17 indicates a run of 4 + n zeroes,
442 * where n is an uncompressed literal 4-bit integer that
443 * follows the magic length. */
444 while (cur_run_len >= 4) {
445 additional_bits = min(cur_run_len - 4, 0xf);
447 output_syms[output_syms_idx++] = 17;
448 output_syms[output_syms_idx++] = additional_bits;
449 cur_run_len -= 4 + additional_bits;
454 /* A run of nonzero lengths. */
456 /* The magic length 19 indicates a run of 4 + n
457 * nonzeroes, where n is a literal bit that follows the
458 * magic length, and where the value of the lengths in
459 * the run is given by an extra length symbol, encoded
460 * with the pretree, that follows the literal bit.
462 * The extra length symbol is encoded as a difference
463 * from the length of the codeword for the first symbol
464 * in the run in the previous tree.
466 while (cur_run_len >= 4) {
467 additional_bits = (cur_run_len > 4);
468 delta = -(char)len_in_run;
472 pretree_freqs[delta]++;
473 output_syms[output_syms_idx++] = 19;
474 output_syms[output_syms_idx++] = additional_bits;
475 output_syms[output_syms_idx++] = delta;
476 cur_run_len -= 4 + additional_bits;
480 /* Any remaining lengths in the run are outputted without RLE,
481 * as a difference from the length of that codeword in the
483 while (cur_run_len--) {
484 delta = -(char)len_in_run;
488 pretree_freqs[delta]++;
489 output_syms[output_syms_idx++] = delta;
495 wimlib_assert(output_syms_idx < ARRAY_LEN(output_syms));
497 /* Build the pretree from the frequencies of the length symbols. */
499 make_canonical_huffman_code(LZX_PRETREE_NUM_SYMBOLS,
500 LZX_MAX_CODEWORD_LEN,
501 pretree_freqs, pretree_lens,
504 /* Write the lengths of the pretree codes to the output. */
505 for (i = 0; i < LZX_PRETREE_NUM_SYMBOLS; i++)
506 bitstream_put_bits(out, pretree_lens[i],
507 LZX_PRETREE_ELEMENT_SIZE);
509 /* Write the length symbols, encoded with the pretree, to the output. */
512 while (i < output_syms_idx) {
513 pretree_sym = output_syms[i++];
515 bitstream_put_bits(out, pretree_codewords[pretree_sym],
516 pretree_lens[pretree_sym]);
517 switch (pretree_sym) {
519 bitstream_put_bits(out, output_syms[i++], 4);
522 bitstream_put_bits(out, output_syms[i++], 5);
525 bitstream_put_bits(out, output_syms[i++], 1);
526 bitstream_put_bits(out,
527 pretree_codewords[output_syms[i]],
528 pretree_lens[output_syms[i]]);
538 /* Builds the canonical Huffman code for the main tree, the length tree, and the
539 * aligned offset tree. */
540 static void lzx_make_huffman_codes(const struct lzx_freq_tables *freq_tabs,
541 struct lzx_codes *codes)
543 make_canonical_huffman_code(LZX_MAINTREE_NUM_SYMBOLS,
544 LZX_MAX_CODEWORD_LEN,
545 freq_tabs->main_freq_table,
547 codes->main_codewords);
549 make_canonical_huffman_code(LZX_LENTREE_NUM_SYMBOLS,
550 LZX_MAX_CODEWORD_LEN,
551 freq_tabs->len_freq_table,
553 codes->len_codewords);
555 make_canonical_huffman_code(LZX_ALIGNEDTREE_NUM_SYMBOLS, 8,
556 freq_tabs->aligned_freq_table,
558 codes->aligned_codewords);
561 /* Do the 'E8' preprocessing, where the targets of x86 CALL instructions were
562 * changed from relative offsets to absolute offsets. This type of
563 * preprocessing can be used on any binary data even if it is not actually
564 * machine code. It seems to always be used in WIM files, even though there is
565 * no bit to indicate that it actually is used, unlike in the LZX compressed
566 * format as used in other file formats such as the cabinet format, where a bit
567 * is reserved for that purpose. */
568 static void do_call_insn_preprocessing(u8 uncompressed_data[],
569 uint uncompressed_data_len)
572 int file_size = LZX_MAGIC_FILESIZE;
576 /* Not enabled in the last 6 bytes, which means the 5-byte call
577 * instruction cannot start in the last *10* bytes. */
578 while (i < uncompressed_data_len - 10) {
579 if (uncompressed_data[i] != 0xe8) {
583 rel_offset = to_le32(*(int32_t*)(uncompressed_data + i + 1));
585 if (rel_offset >= -i && rel_offset < file_size) {
586 if (rel_offset < file_size - i) {
587 /* "good translation" */
588 abs_offset = rel_offset + i;
590 /* "compensating translation" */
591 abs_offset = rel_offset - file_size;
593 *(int32_t*)(uncompressed_data + i + 1) = to_le32(abs_offset);
600 static const struct lz_params lzx_lz_params = {
602 .max_match = LZX_MAX_MATCH,
603 .good_match = LZX_MAX_MATCH,
604 .nice_match = LZX_MAX_MATCH,
605 .max_chain_len = LZX_MAX_MATCH,
606 .max_lazy_match = LZX_MAX_MATCH,
611 * Performs LZX compression on a block of data.
613 * @__uncompressed_data: Pointer to the data to be compressed.
614 * @uncompressed_len: Length, in bytes, of the data to be compressed.
615 * @compressed_data: Pointer to a location at least (@uncompressed_len - 1)
616 * bytes long into which the compressed data may be
618 * @compressed_len_ret: A pointer to an unsigned int into which the length of
619 * the compressed data may be returned.
621 * Returns zero if compression was successfully performed. In that case
622 * @compressed_data and @compressed_len_ret will contain the compressed data and
623 * its length. A return value of nonzero means that compressing the data did
624 * not reduce its size, and @compressed_data will not contain the full
627 int lzx_compress(const void *__uncompressed_data, uint uncompressed_len,
628 void *compressed_data, uint *compressed_len_ret)
630 struct output_bitstream ostream;
631 u8 uncompressed_data[uncompressed_len + LZX_MAX_MATCH];
632 struct lzx_freq_tables freq_tabs;
633 struct lzx_codes codes;
634 u32 match_tab[uncompressed_len];
635 struct lru_queue queue = {.R0 = 1, .R1 = 1, .R2 = 1};
641 LZX_DEBUG("uncompressed_len = %u\n", uncompressed_len);
643 if (uncompressed_len < 100)
647 memset(&freq_tabs, 0, sizeof(freq_tabs));
649 /* The input data must be preprocessed. To avoid changing the original
650 * input, copy it to a temporary buffer. */
651 memcpy(uncompressed_data, __uncompressed_data, uncompressed_len);
654 /* Before doing any actual compression, do the call instruction (0xe8
655 * byte) translation on the uncompressed data. */
656 do_call_insn_preprocessing(uncompressed_data, uncompressed_len);
659 /* Determine the sequence of matches and literals that will be output,
660 * and in the process, keep counts of the number of times each symbol
661 * will be output, so that the Huffman trees can be made. */
663 num_matches = lz_analyze_block(uncompressed_data, uncompressed_len,
664 match_tab, lzx_record_match,
665 lzx_record_literal, &freq_tabs,
666 &queue, freq_tabs.main_freq_table,
669 LZX_DEBUG("using %u matches\n", num_matches);
672 lzx_make_huffman_codes(&freq_tabs, &codes);
674 /* Initialize the output bitstream. */
675 init_output_bitstream(&ostream, compressed_data, uncompressed_len - 1);
677 /* The first three bits tell us what kind of block it is, and are one
678 * of the LZX_BLOCKTYPE_* values. */
679 bitstream_put_bits(&ostream, LZX_BLOCKTYPE_ALIGNED, 3);
681 /* The next bit indicates whether the block size is the default (32768),
682 * indicated by a 1 bit, or whether the block size is given by the next
683 * 16 bits, indicated by a 0 bit. */
684 if (uncompressed_len == 32768) {
685 bitstream_put_bits(&ostream, 1, 1);
687 bitstream_put_bits(&ostream, 0, 1);
688 bitstream_put_bits(&ostream, uncompressed_len, 16);
691 /* Write out the aligned offset tree. Note that M$ lies and says that
692 * the aligned offset tree comes after the length tree, but that is
693 * wrong; it actually is before the main tree. */
694 for (i = 0; i < LZX_ALIGNEDTREE_NUM_SYMBOLS; i++)
695 bitstream_put_bits(&ostream, codes.aligned_lens[i],
696 LZX_ALIGNEDTREE_ELEMENT_SIZE);
698 /* Write the pre-tree and lengths for the first LZX_NUM_CHARS symbols in the
700 ret = lzx_write_compressed_tree(&ostream, codes.main_lens,
705 /* Write the pre-tree and symbols for the rest of the main tree. */
706 ret = lzx_write_compressed_tree(&ostream, codes.main_lens +
708 LZX_MAINTREE_NUM_SYMBOLS -
713 /* Write the pre-tree and symbols for the length tree. */
714 ret = lzx_write_compressed_tree(&ostream, codes.len_lens,
715 LZX_LENTREE_NUM_SYMBOLS);
719 /* Write the compressed literals. */
720 ret = lzx_write_compressed_literals(&ostream, LZX_BLOCKTYPE_ALIGNED,
721 match_tab, num_matches, &codes);
725 ret = flush_output_bitstream(&ostream);
729 compressed_len = ostream.bit_output - (u8*)compressed_data;
731 LZX_DEBUG("Compressed %u => %u bytes\n",
732 uncompressed_len, compressed_len);
734 *compressed_len_ret = compressed_len;
736 #ifdef ENABLE_VERIFY_COMPRESSION
737 /* Verify that we really get the same thing back when decompressing. */
738 u8 buf[uncompressed_len];
739 ret = lzx_decompress(compressed_data, compressed_len, buf,
742 ERROR("ERROR: Failed to decompress data we compressed!\n");
747 for (i = 0; i < uncompressed_len; i++) {
748 if (buf[i] != *((u8*)__uncompressed_data + i)) {
749 ERROR("Data we compressed didn't decompress to "
750 "the original data (difference at byte %u of "
751 "%u)\n", i + 1, uncompressed_len);
755 LZX_DEBUG("Compression verified to be correct.\n");