4 * XPRESS compression routines.
6 * See the comments in xpress-decompress.c about the XPRESS format.
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/.
36 * Writes @match, which is a match given in the intermediate representation for
37 * XPRESS matches, to the output stream @ostream.
39 * @codewords and @lens provide the Huffman code that is being used.
42 xpress_write_match(struct output_bitstream *ostream, u32 match,
43 const u16 codewords[], const u8 lens[])
45 u32 adjusted_match_len = match & 0xffff;
46 u32 match_offset = match >> 16;
47 u32 len_hdr = min(adjusted_match_len, 0xf);
48 u32 offset_bsr = bsr32(match_offset);
49 u32 sym = len_hdr | (offset_bsr << 4) | XPRESS_NUM_CHARS;
52 ret = bitstream_put_bits(ostream, codewords[sym], lens[sym]);
56 if (adjusted_match_len >= 0xf) {
57 u8 byte1 = min(adjusted_match_len - 0xf, 0xff);
58 ret = bitstream_put_byte(ostream, byte1);
62 ret = bitstream_put_two_bytes(ostream, adjusted_match_len);
67 return bitstream_put_bits(ostream,
68 match_offset ^ (1 << offset_bsr), offset_bsr);
72 xpress_write_compressed_literals(struct output_bitstream *ostream,
73 const u32 match_tab[],
75 const u16 codewords[],
78 for (unsigned i = 0; i < num_matches; i++) {
80 u32 match = match_tab[i];
82 if (match >= XPRESS_NUM_CHARS) /* match */
83 ret = xpress_write_match(ostream, match, codewords,
85 else /* literal byte */
86 ret = bitstream_put_bits(ostream, codewords[match],
91 return bitstream_put_bits(ostream, codewords[XPRESS_END_OF_DATA],
92 lens[XPRESS_END_OF_DATA]);
96 xpress_record_literal(u8 literal, void *__freq_tab)
98 freq_t *freq_tab = __freq_tab;
104 xpress_record_match(unsigned match_offset, unsigned match_len,
105 void *freq_tab, void *ignore)
107 wimlib_assert(match_len >= XPRESS_MIN_MATCH &&
108 match_len <= XPRESS_MAX_MATCH);
109 wimlib_assert(match_offset >= XPRESS_MIN_OFFSET &&
110 match_offset <= XPRESS_MAX_OFFSET);
113 * The intermediate representation of XPRESS matches is as follows:
116 * ---- -----------------------------------------------------------
118 * 16-31 match offset (XPRESS_MIN_OFFSET < x < XPRESS_MAX_OFFSET)
120 * 0-15 adjusted match length (0 <= x <= XPRESS_MAX_MATCH - XPRESS_MIN_MATCH)
122 * Literals are simply represented as themselves and can be
123 * distinguished from matches by the fact that only literals will have
124 * the upper three bytes completely clear. */
126 u32 adjusted_match_len = match_len - XPRESS_MIN_MATCH;
127 u32 len_hdr = min(adjusted_match_len, 0xf);
128 u32 offset_bsr = bsr32(match_offset);
129 u32 sym = len_hdr | (offset_bsr << 4) | XPRESS_NUM_CHARS;
130 ((freq_t*)freq_tab)[sym]++;
131 return adjusted_match_len | (match_offset << 16);
134 static const struct lz_params xpress_lz_params = {
135 .min_match = XPRESS_MIN_MATCH,
136 .max_match = XPRESS_MAX_MATCH,
140 .max_lazy_match = 16,
144 /* Documented in wimlib.h */
146 wimlib_xpress_compress(const void *__uncompressed_data,
147 unsigned uncompressed_len, void *__compressed_data)
149 const u8 *uncompressed_data = __uncompressed_data;
150 u8 *compressed_data = __compressed_data;
151 struct output_bitstream ostream;
152 u32 match_tab[uncompressed_len];
153 freq_t freq_tab[XPRESS_NUM_SYMBOLS];
154 u16 codewords[XPRESS_NUM_SYMBOLS];
155 u8 lens[XPRESS_NUM_SYMBOLS];
156 unsigned num_matches;
157 unsigned compressed_len;
161 wimlib_assert(uncompressed_len <= 32768);
163 /* XPRESS requires 256 bytes of overhead for the Huffman tables, so it's
164 * impossible cannot compress 256 bytes or less of data to less than the
167 * +1 to take into account that the buffer for compressed data is 1 byte
168 * smaller than the buffer for uncompressed data.
170 * +4 to take into account that init_output_bitstream() requires at
171 * least 4 bytes of data. */
172 if (uncompressed_len < XPRESS_NUM_SYMBOLS / 2 + 1 + 4)
175 ZERO_ARRAY(freq_tab);
176 num_matches = lz_analyze_block(uncompressed_data, uncompressed_len,
177 match_tab, xpress_record_match,
178 xpress_record_literal, freq_tab,
182 freq_tab[XPRESS_END_OF_DATA]++;
184 make_canonical_huffman_code(XPRESS_NUM_SYMBOLS, XPRESS_MAX_CODEWORD_LEN,
185 freq_tab, lens, codewords);
189 * It's tempting to output the 512 Huffman codeword lengths using the
190 * bitstream_put_bits() function. However, this is NOT correct because
191 * bitstream_put_bits() will output 2 bytes at a time in little-endian
192 * order, which is the order that is needed for the compressed literals.
193 * However, the bytes in the lengths table are in order, so they need to
194 * be written one at a time without using bitstream_put_bits().
196 * Because of this, init_output_bitstream() is not called until after
197 * the lengths table is output.
199 for (i = 0; i < XPRESS_NUM_SYMBOLS; i += 2)
200 *compressed_data++ = (lens[i] & 0xf) | (lens[i + 1] << 4);
202 init_output_bitstream(&ostream, compressed_data,
203 uncompressed_len - XPRESS_NUM_SYMBOLS / 2 - 1);
205 ret = xpress_write_compressed_literals(&ostream, match_tab,
206 num_matches, codewords, lens);
210 /* Flush any bits that are buffered. */
211 ret = flush_output_bitstream(&ostream);
215 /* Assert that there are no output bytes between the ostream.output
216 * pointer and the ostream.next_bit_output pointer. This can only
217 * happen if bytes had been written at the ostream.output pointer before
218 * the last bit word was written to the stream. But, this does not
219 * occur since xpress_write_match() always finishes by writing some bits
220 * (a Huffman symbol), and the bitstream was just flushed. */
221 wimlib_assert(ostream.output - ostream.next_bit_output == 2);
223 /* The length of the compressed data is supposed to be the value of the
224 * ostream.output pointer before flushing, which is now the
225 * output.next_bit_output pointer after flushing.
227 * There will be an extra 2 bytes at the ostream.bit_output pointer,
228 * which is zeroed out. (These 2 bytes may be either the last bytes in
229 * the compressed data, in which case they are actually unnecessary, or
230 * they may precede a number of bytes embedded into the bitstream.) */
231 if (ostream.bit_output >
232 (const u8*)__compressed_data + uncompressed_len - 3)
234 *(u16*)ostream.bit_output = cpu_to_le16(0);
235 compressed_len = ostream.next_bit_output - (const u8*)__compressed_data;
237 wimlib_assert(compressed_len <= uncompressed_len - 1);
239 #ifdef ENABLE_VERIFY_COMPRESSION
240 /* Verify that we really get the same thing back when decompressing. */
241 u8 buf[uncompressed_len];
242 ret = wimlib_xpress_decompress(__compressed_data, compressed_len,
243 buf, uncompressed_len);
245 ERROR("xpress_compress(): Failed to decompress data we "
249 for (i = 0; i < uncompressed_len; i++) {
250 if (buf[i] != uncompressed_data[i]) {
251 ERROR("xpress_compress(): Data we compressed didn't "
252 "decompress to the original data (difference at "
253 "byte %u of %u)", i + 1, uncompressed_len);
258 return compressed_len;