4 * LZX compression routines, originally based on code written by Matthew T.
5 * Russotto (liblzxcomp), but heavily modified.
9 * Copyright (C) 2002 Matthew T. Russotto
10 * Copyright (C) 2012, 2013 Eric Biggers
12 * This file is part of wimlib, a library for working with WIM files.
14 * wimlib is free software; you can redistribute it and/or modify it under the
15 * terms of the GNU General Public License as published by the Free
16 * Software Foundation; either version 3 of the License, or (at your option)
19 * wimlib 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
21 * A PARTICULAR PURPOSE. See the GNU General Public License for more
24 * You should have received a copy of the GNU General Public License
25 * along with wimlib; if not, see http://www.gnu.org/licenses/.
30 * This file provides wimlib_lzx_compress(), a function to compress an in-memory
31 * buffer of data using LZX compression, as used in the WIM file format.
33 * Please see the comments in lzx-decompress.c for more information about this
36 * One thing to keep in mind is that there is no sliding window, since the
37 * window is always the entirety of a WIM chunk, which is at most WIM_CHUNK_SIZE
40 * The basic compression algorithm used here should be familiar if you are
41 * familiar with Huffman trees and with other LZ77 and Huffman-based formats
42 * such as DEFLATE. Otherwise it can be quite tricky to understand. Basically
43 * it is the following:
45 * - Preprocess the input data (LZX-specific)
46 * - Go through the input data and determine matches. This part is based on
47 * code from zlib, and a hash table of 3-character strings is used to
48 * accelerate the process of finding matches.
49 * - Build the Huffman trees based on the frequencies of symbols determined
50 * while recording matches.
51 * - Output the block header, including the Huffman trees; then output the
52 * compressed stream of matches and literal characters.
54 * It is possible for a WIM chunk to include multiple LZX blocks, since for some
55 * input data this will produce a better compression ratio (especially since
56 * each block can include new Huffman codes). However, producing multiple LZX
57 * blocks from one input chunk is not yet implemented.
65 #include "wimlib/compress.h"
66 #include "wimlib/lzx.h"
67 #include "wimlib/util.h"
73 /* Structure to contain the Huffman codes for the main, length, and aligned
76 u16 main_codewords[LZX_MAINTREE_NUM_SYMBOLS];
77 u8 main_lens[LZX_MAINTREE_NUM_SYMBOLS];
79 u16 len_codewords[LZX_LENTREE_NUM_SYMBOLS];
80 u8 len_lens[LZX_LENTREE_NUM_SYMBOLS];
82 u16 aligned_codewords[LZX_ALIGNEDTREE_NUM_SYMBOLS];
83 u8 aligned_lens[LZX_ALIGNEDTREE_NUM_SYMBOLS];
86 struct lzx_freq_tables {
87 freq_t main_freq_table[LZX_MAINTREE_NUM_SYMBOLS];
88 freq_t len_freq_table[LZX_LENTREE_NUM_SYMBOLS];
89 freq_t aligned_freq_table[LZX_ALIGNEDTREE_NUM_SYMBOLS];
92 /* Returns the LZX position slot that corresponds to a given formatted offset.
94 * Logically, this returns the smallest i such that
95 * formatted_offset >= lzx_position_base[i].
97 * The actual implementation below takes advantage of the regularity of the
98 * numbers in the lzx_position_base array to calculate the slot directly from
99 * the formatted offset without actually looking at the array.
101 static inline unsigned
102 lzx_get_position_slot(unsigned formatted_offset)
106 * Slots 36-49 (formatted_offset >= 262144) can be found by
107 * (formatted_offset/131072) + 34 == (formatted_offset >> 17) + 34;
108 * however, this check for formatted_offset >= 262144 is commented out
109 * because WIM chunks cannot be that large.
111 if (formatted_offset >= 262144) {
112 return (formatted_offset >> 17) + 34;
116 /* Note: this part here only works if:
118 * 2 <= formatted_offset < 655360
120 * It is < 655360 because the frequency of the position bases
121 * increases starting at the 655360 entry, and it is >= 2
122 * because the below calculation fails if the most significant
123 * bit is lower than the 2's place. */
124 wimlib_assert(formatted_offset >= 2 && formatted_offset < 655360);
125 unsigned mssb_idx = bsr32(formatted_offset);
126 return (mssb_idx << 1) |
127 ((formatted_offset >> (mssb_idx - 1)) & 1);
132 lzx_record_literal(u8 literal, void *__main_freq_tab)
134 freq_t *main_freq_tab = __main_freq_tab;
135 main_freq_tab[literal]++;
139 /* Constructs a match from an offset and a length, and updates the LRU queue and
140 * the frequency of symbols in the main, length, and aligned offset alphabets.
141 * The return value is a 32-bit number that provides the match in an
142 * intermediate representation documented below. */
144 lzx_record_match(unsigned match_offset, unsigned match_len,
145 void *__freq_tabs, void *__queue)
147 struct lzx_freq_tables *freq_tabs = __freq_tabs;
148 struct lru_queue *queue = __queue;
149 unsigned position_slot;
150 unsigned position_footer = 0;
155 unsigned adjusted_match_len;
157 wimlib_assert(match_len >= LZX_MIN_MATCH && match_len <= LZX_MAX_MATCH);
158 wimlib_assert(match_offset != 0);
160 /* If possible, encode this offset as a repeated offset. */
161 if (match_offset == queue->R0) {
163 } else if (match_offset == queue->R1) {
164 swap(queue->R0, queue->R1);
166 } else if (match_offset == queue->R2) {
167 swap(queue->R0, queue->R2);
170 /* Not a repeated offset. */
172 /* offsets of 0, 1, and 2 are reserved for the repeated offset
173 * codes, so non-repeated offsets must be encoded as 3+. The
174 * minimum offset is 1, so encode the offsets offset by 2. */
175 unsigned formatted_offset = match_offset + LZX_MIN_MATCH;
177 queue->R2 = queue->R1;
178 queue->R1 = queue->R0;
179 queue->R0 = match_offset;
181 /* The (now-formatted) offset will actually be encoded as a
182 * small position slot number that maps to a certain hard-coded
183 * offset (position base), followed by a number of extra bits---
184 * the position footer--- that are added to the position base to
185 * get the original formatted offset. */
187 position_slot = lzx_get_position_slot(formatted_offset);
188 position_footer = formatted_offset &
189 ((1 << lzx_get_num_extra_bits(position_slot)) - 1);
192 adjusted_match_len = match_len - LZX_MIN_MATCH;
194 /* Pack the position slot, position footer, and match length into an
195 * intermediate representation.
198 * ---- -----------------------------------------------------------
200 * 31 1 if a match, 0 if a literal.
202 * 30-25 position slot. This can be at most 50, so it will fit in 6
205 * 8-24 position footer. This is the offset of the real formatted
206 * offset from the position base. This can be at most 17 bits
207 * (since lzx_extra_bits[LZX_NUM_POSITION_SLOTS - 1] is 17).
209 * 0-7 length of match, offset by 2. This can be at most
210 * (LZX_MAX_MATCH - 2) == 255, so it will fit in 8 bits. */
212 (position_slot << 25) |
213 (position_footer << 8) |
214 (adjusted_match_len);
216 /* The match length must be at least 2, so let the adjusted match length
217 * be the match length minus 2.
219 * If it is less than 7, the adjusted match length is encoded as a 3-bit
220 * number offset by 2. Otherwise, the 3-bit length header is all 1's
221 * and the actual adjusted length is given as a symbol encoded with the
222 * length tree, offset by 7.
224 if (adjusted_match_len < LZX_NUM_PRIMARY_LENS) {
225 len_header = adjusted_match_len;
227 len_header = LZX_NUM_PRIMARY_LENS;
228 len_footer = adjusted_match_len - LZX_NUM_PRIMARY_LENS;
229 freq_tabs->len_freq_table[len_footer]++;
231 len_pos_header = (position_slot << 3) | len_header;
233 wimlib_assert(len_pos_header < LZX_MAINTREE_NUM_SYMBOLS - LZX_NUM_CHARS);
235 freq_tabs->main_freq_table[len_pos_header + LZX_NUM_CHARS]++;
238 * if (lzx_extra_bits[position_slot] >= 3) */
239 if (position_slot >= 8)
240 freq_tabs->aligned_freq_table[position_footer & 7]++;
246 * Writes a compressed literal match to the output.
248 * @out: The output bitstream.
249 * @block_type: The type of the block (LZX_BLOCKTYPE_ALIGNED or LZX_BLOCKTYPE_VERBATIM)
250 * @match: The match, encoded as a 32-bit number.
251 * @codes: Pointer to a structure that contains the codewords for the
252 * main, length, and aligned offset Huffman codes.
255 lzx_write_match(struct output_bitstream *out, int block_type,
256 u32 match, const struct lzx_codes *codes)
258 /* low 8 bits are the match length minus 2 */
259 unsigned match_len_minus_2 = match & 0xff;
260 /* Next 17 bits are the position footer */
261 unsigned position_footer = (match >> 8) & 0x1ffff; /* 17 bits */
262 /* Next 6 bits are the position slot. */
263 unsigned position_slot = (match >> 25) & 0x3f; /* 6 bits */
266 unsigned len_pos_header;
267 unsigned main_symbol;
268 unsigned num_extra_bits;
269 unsigned verbatim_bits;
270 unsigned aligned_bits;
273 /* If the match length is less than MIN_MATCH (= 2) +
274 * NUM_PRIMARY_LENS (= 7), the length header contains
275 * the match length minus MIN_MATCH, and there is no
278 * Otherwise, the length header contains
279 * NUM_PRIMARY_LENS, and the length footer contains
280 * the match length minus NUM_PRIMARY_LENS minus
282 if (match_len_minus_2 < LZX_NUM_PRIMARY_LENS) {
283 len_header = match_len_minus_2;
284 /* No length footer-- mark it with a special
286 len_footer = (unsigned)(-1);
288 len_header = LZX_NUM_PRIMARY_LENS;
289 len_footer = match_len_minus_2 - LZX_NUM_PRIMARY_LENS;
292 /* Combine the position slot with the length header into
293 * a single symbol that will be encoded with the main
295 len_pos_header = (position_slot << 3) | len_header;
297 /* The actual main symbol is offset by LZX_NUM_CHARS because
298 * values under LZX_NUM_CHARS are used to indicate a literal
299 * byte rather than a match. */
300 main_symbol = len_pos_header + LZX_NUM_CHARS;
302 /* Output main symbol. */
303 ret = bitstream_put_bits(out, codes->main_codewords[main_symbol],
304 codes->main_lens[main_symbol]);
308 /* If there is a length footer, output it using the
309 * length Huffman code. */
310 if (len_footer != (unsigned)(-1)) {
311 ret = bitstream_put_bits(out, codes->len_codewords[len_footer],
312 codes->len_lens[len_footer]);
317 wimlib_assert(position_slot < LZX_NUM_POSITION_SLOTS);
319 num_extra_bits = lzx_get_num_extra_bits(position_slot);
321 /* For aligned offset blocks with at least 3 extra bits, output the
322 * verbatim bits literally, then the aligned bits encoded using the
323 * aligned offset tree. Otherwise, only the verbatim bits need to be
325 if ((block_type == LZX_BLOCKTYPE_ALIGNED) && (num_extra_bits >= 3)) {
327 verbatim_bits = position_footer >> 3;
328 ret = bitstream_put_bits(out, verbatim_bits,
333 aligned_bits = (position_footer & 7);
334 ret = bitstream_put_bits(out,
335 codes->aligned_codewords[aligned_bits],
336 codes->aligned_lens[aligned_bits]);
340 /* verbatim bits is the same as the position
341 * footer, in this case. */
342 ret = bitstream_put_bits(out, position_footer, num_extra_bits);
350 * Writes all compressed literals in a block, both matches and literal bytes, to
351 * the output bitstream.
353 * @out: The output bitstream.
354 * @block_type: The type of the block (LZX_BLOCKTYPE_ALIGNED or LZX_BLOCKTYPE_VERBATIM)
355 * @match_tab[]: The array of matches that will be output. It has length
356 * of @num_compressed_literals.
357 * @num_compressed_literals: Number of compressed literals to be output.
358 * @codes: Pointer to a structure that contains the codewords for the
359 * main, length, and aligned offset Huffman codes.
362 lzx_write_compressed_literals(struct output_bitstream *ostream,
364 const u32 match_tab[],
365 unsigned num_compressed_literals,
366 const struct lzx_codes *codes)
372 for (i = 0; i < num_compressed_literals; i++) {
373 match = match_tab[i];
375 /* High bit of the match indicates whether the match is an
376 * actual match (1) or a literal uncompressed byte (0) */
377 if (match & 0x80000000) {
379 ret = lzx_write_match(ostream, block_type, match,
385 wimlib_assert(match < LZX_NUM_CHARS);
386 ret = bitstream_put_bits(ostream,
387 codes->main_codewords[match],
388 codes->main_lens[match]);
397 * Writes a compressed Huffman tree to the output, preceded by the pretree for
400 * The Huffman tree is represented in the output as a series of path lengths
401 * from which the canonical Huffman code can be reconstructed. The path lengths
402 * themselves are compressed using a separate Huffman code, the pretree, which
403 * consists of LZX_PRETREE_NUM_SYMBOLS (= 20) symbols that cover all possible code
404 * lengths, plus extra codes for repeated lengths. The path lengths of the
405 * pretree precede the path lengths of the larger code and are uncompressed,
406 * consisting of 20 entries of 4 bits each.
408 * @out: The bitstream for the compressed output.
409 * @lens: The code lengths for the Huffman tree, indexed by symbol.
410 * @num_symbols: The number of symbols in the code.
413 lzx_write_compressed_tree(struct output_bitstream *out,
414 const u8 lens[], unsigned num_symbols)
416 /* Frequencies of the length symbols, including the RLE symbols (NOT the
417 * actual lengths themselves). */
418 freq_t pretree_freqs[LZX_PRETREE_NUM_SYMBOLS];
419 u8 pretree_lens[LZX_PRETREE_NUM_SYMBOLS];
420 u16 pretree_codewords[LZX_PRETREE_NUM_SYMBOLS];
421 u8 output_syms[num_symbols * 2];
422 unsigned output_syms_idx;
423 unsigned cur_run_len;
426 unsigned additional_bits;
430 ZERO_ARRAY(pretree_freqs);
432 /* Since the code word lengths use a form of RLE encoding, the goal here
433 * is to find each run of identical lengths when going through them in
434 * symbol order (including runs of length 1). For each run, as many
435 * lengths are encoded using RLE as possible, and the rest are output
438 * output_syms[] will be filled in with the length symbols that will be
439 * output, including RLE codes, not yet encoded using the pre-tree.
441 * cur_run_len keeps track of how many code word lengths are in the
442 * current run of identical lengths.
446 for (i = 1; i <= num_symbols; i++) {
448 if (i != num_symbols && lens[i] == lens[i - 1]) {
449 /* Still in a run--- keep going. */
454 /* Run ended! Check if it is a run of zeroes or a run of
457 /* The symbol that was repeated in the run--- not to be confused
458 * with the length *of* the run (cur_run_len) */
459 len_in_run = lens[i - 1];
461 if (len_in_run == 0) {
462 /* A run of 0's. Encode it in as few length
463 * codes as we can. */
465 /* The magic length 18 indicates a run of 20 + n zeroes,
466 * where n is an uncompressed literal 5-bit integer that
467 * follows the magic length. */
468 while (cur_run_len >= 20) {
470 additional_bits = min(cur_run_len - 20, 0x1f);
472 output_syms[output_syms_idx++] = 18;
473 output_syms[output_syms_idx++] = additional_bits;
474 cur_run_len -= 20 + additional_bits;
477 /* The magic length 17 indicates a run of 4 + n zeroes,
478 * where n is an uncompressed literal 4-bit integer that
479 * follows the magic length. */
480 while (cur_run_len >= 4) {
481 additional_bits = min(cur_run_len - 4, 0xf);
483 output_syms[output_syms_idx++] = 17;
484 output_syms[output_syms_idx++] = additional_bits;
485 cur_run_len -= 4 + additional_bits;
490 /* A run of nonzero lengths. */
492 /* The magic length 19 indicates a run of 4 + n
493 * nonzeroes, where n is a literal bit that follows the
494 * magic length, and where the value of the lengths in
495 * the run is given by an extra length symbol, encoded
496 * with the pretree, that follows the literal bit.
498 * The extra length symbol is encoded as a difference
499 * from the length of the codeword for the first symbol
500 * in the run in the previous tree.
502 while (cur_run_len >= 4) {
503 additional_bits = (cur_run_len > 4);
504 delta = -(char)len_in_run;
508 pretree_freqs[(unsigned char)delta]++;
509 output_syms[output_syms_idx++] = 19;
510 output_syms[output_syms_idx++] = additional_bits;
511 output_syms[output_syms_idx++] = delta;
512 cur_run_len -= 4 + additional_bits;
516 /* Any remaining lengths in the run are outputted without RLE,
517 * as a difference from the length of that codeword in the
519 while (cur_run_len--) {
520 delta = -(char)len_in_run;
524 pretree_freqs[(unsigned char)delta]++;
525 output_syms[output_syms_idx++] = delta;
531 wimlib_assert(output_syms_idx < ARRAY_LEN(output_syms));
533 /* Build the pretree from the frequencies of the length symbols. */
535 make_canonical_huffman_code(LZX_PRETREE_NUM_SYMBOLS,
536 LZX_MAX_CODEWORD_LEN,
537 pretree_freqs, pretree_lens,
540 /* Write the lengths of the pretree codes to the output. */
541 for (i = 0; i < LZX_PRETREE_NUM_SYMBOLS; i++)
542 bitstream_put_bits(out, pretree_lens[i],
543 LZX_PRETREE_ELEMENT_SIZE);
545 /* Write the length symbols, encoded with the pretree, to the output. */
548 while (i < output_syms_idx) {
549 pretree_sym = output_syms[i++];
551 bitstream_put_bits(out, pretree_codewords[pretree_sym],
552 pretree_lens[pretree_sym]);
553 switch (pretree_sym) {
555 bitstream_put_bits(out, output_syms[i++], 4);
558 bitstream_put_bits(out, output_syms[i++], 5);
561 bitstream_put_bits(out, output_syms[i++], 1);
562 bitstream_put_bits(out,
563 pretree_codewords[output_syms[i]],
564 pretree_lens[output_syms[i]]);
574 /* Builds the canonical Huffman code for the main tree, the length tree, and the
575 * aligned offset tree. */
577 lzx_make_huffman_codes(const struct lzx_freq_tables *freq_tabs,
578 struct lzx_codes *codes)
580 make_canonical_huffman_code(LZX_MAINTREE_NUM_SYMBOLS,
581 LZX_MAX_CODEWORD_LEN,
582 freq_tabs->main_freq_table,
584 codes->main_codewords);
586 make_canonical_huffman_code(LZX_LENTREE_NUM_SYMBOLS,
587 LZX_MAX_CODEWORD_LEN,
588 freq_tabs->len_freq_table,
590 codes->len_codewords);
592 make_canonical_huffman_code(LZX_ALIGNEDTREE_NUM_SYMBOLS, 8,
593 freq_tabs->aligned_freq_table,
595 codes->aligned_codewords);
599 do_call_insn_translation(u32 *call_insn_target, int input_pos,
605 rel_offset = le32_to_cpu(*call_insn_target);
606 if (rel_offset >= -input_pos && rel_offset < file_size) {
607 if (rel_offset < file_size - input_pos) {
608 /* "good translation" */
609 abs_offset = rel_offset + input_pos;
611 /* "compensating translation" */
612 abs_offset = rel_offset - file_size;
614 *call_insn_target = cpu_to_le32(abs_offset);
618 /* This is the reverse of undo_call_insn_preprocessing() in lzx-decompress.c.
619 * See the comment above that function for more information. */
621 do_call_insn_preprocessing(u8 uncompressed_data[], int uncompressed_data_len)
623 for (int i = 0; i < uncompressed_data_len - 10; i++) {
624 if (uncompressed_data[i] == 0xe8) {
625 do_call_insn_translation((u32*)&uncompressed_data[i + 1],
627 LZX_WIM_MAGIC_FILESIZE);
634 static const struct lz_params lzx_lz_params = {
636 /* LZX_MIN_MATCH == 2, but 2-character matches are rarely useful; the
637 * minimum match for compression is set to 3 instead. */
640 .max_match = LZX_MAX_MATCH,
641 .good_match = LZX_MAX_MATCH,
642 .nice_match = LZX_MAX_MATCH,
643 .max_chain_len = LZX_MAX_MATCH,
644 .max_lazy_match = LZX_MAX_MATCH,
648 /* Documented in wimlib.h */
650 wimlib_lzx_compress(const void *__uncompressed_data, unsigned uncompressed_len,
651 void *compressed_data)
653 struct output_bitstream ostream;
654 u8 uncompressed_data[uncompressed_len + 8];
655 struct lzx_freq_tables freq_tabs;
656 struct lzx_codes codes;
657 u32 match_tab[uncompressed_len];
658 struct lru_queue queue;
659 unsigned num_matches;
660 unsigned compressed_len;
663 int block_type = LZX_BLOCKTYPE_ALIGNED;
665 wimlib_assert(uncompressed_len <= 32768);
667 if (uncompressed_len < 100)
670 memset(&freq_tabs, 0, sizeof(freq_tabs));
675 /* The input data must be preprocessed. To avoid changing the original
676 * input, copy it to a temporary buffer. */
677 memcpy(uncompressed_data, __uncompressed_data, uncompressed_len);
678 memset(uncompressed_data + uncompressed_len, 0, 8);
680 /* Before doing any actual compression, do the call instruction (0xe8
681 * byte) translation on the uncompressed data. */
682 do_call_insn_preprocessing(uncompressed_data, uncompressed_len);
684 /* Determine the sequence of matches and literals that will be output,
685 * and in the process, keep counts of the number of times each symbol
686 * will be output, so that the Huffman trees can be made. */
688 num_matches = lz_analyze_block(uncompressed_data, uncompressed_len,
689 match_tab, lzx_record_match,
690 lzx_record_literal, &freq_tabs,
691 &queue, freq_tabs.main_freq_table,
694 lzx_make_huffman_codes(&freq_tabs, &codes);
696 /* Initialize the output bitstream. */
697 init_output_bitstream(&ostream, compressed_data, uncompressed_len - 1);
699 /* The first three bits tell us what kind of block it is, and are one
700 * of the LZX_BLOCKTYPE_* values. */
701 bitstream_put_bits(&ostream, block_type, 3);
703 /* The next bit indicates whether the block size is the default (32768),
704 * indicated by a 1 bit, or whether the block size is given by the next
705 * 16 bits, indicated by a 0 bit. */
706 if (uncompressed_len == 32768) {
707 bitstream_put_bits(&ostream, 1, 1);
709 bitstream_put_bits(&ostream, 0, 1);
710 bitstream_put_bits(&ostream, uncompressed_len, 16);
713 /* Write out the aligned offset tree. Note that M$ lies and says that
714 * the aligned offset tree comes after the length tree, but that is
715 * wrong; it actually is before the main tree. */
716 if (block_type == LZX_BLOCKTYPE_ALIGNED)
717 for (i = 0; i < LZX_ALIGNEDTREE_NUM_SYMBOLS; i++)
718 bitstream_put_bits(&ostream, codes.aligned_lens[i],
719 LZX_ALIGNEDTREE_ELEMENT_SIZE);
721 /* Write the pre-tree and lengths for the first LZX_NUM_CHARS symbols in the
723 ret = lzx_write_compressed_tree(&ostream, codes.main_lens,
728 /* Write the pre-tree and symbols for the rest of the main tree. */
729 ret = lzx_write_compressed_tree(&ostream, codes.main_lens +
731 LZX_MAINTREE_NUM_SYMBOLS -
736 /* Write the pre-tree and symbols for the length tree. */
737 ret = lzx_write_compressed_tree(&ostream, codes.len_lens,
738 LZX_LENTREE_NUM_SYMBOLS);
742 /* Write the compressed literals. */
743 ret = lzx_write_compressed_literals(&ostream, block_type,
744 match_tab, num_matches, &codes);
748 ret = flush_output_bitstream(&ostream);
752 compressed_len = ostream.bit_output - (u8*)compressed_data;
754 #ifdef ENABLE_VERIFY_COMPRESSION
755 /* Verify that we really get the same thing back when decompressing. */
757 u8 buf[uncompressed_len];
758 ret = wimlib_lzx_decompress(compressed_data, compressed_len,
759 buf, uncompressed_len);
761 ERROR("lzx_compress(): Failed to decompress data we compressed");
765 for (i = 0; i < uncompressed_len; i++) {
766 if (buf[i] != *((u8*)__uncompressed_data + i)) {
767 ERROR("lzx_compress(): Data we compressed didn't "
768 "decompress to the original data (difference at "
769 "byte %u of %u)", i + 1, uncompressed_len);
775 return compressed_len;