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/.
34 * Writes @match, which is a match given in the intermediate representation for
35 * XPRESS matches, to the output stream @ostream.
37 * @codewords and @lens provide the Huffman code that is being used.
40 xpress_write_match(struct output_bitstream *ostream, u32 match,
41 const u16 codewords[], const u8 lens[])
43 u32 adjusted_match_len = match & 0xffff;
44 u32 match_offset = match >> 16;
45 u32 len_hdr = min(adjusted_match_len, 0xf);
46 u32 offset_bsr = bsr32(match_offset);
47 u32 sym = len_hdr | (offset_bsr << 4) | XPRESS_NUM_CHARS;
50 ret = bitstream_put_bits(ostream, codewords[sym], lens[sym]);
54 if (adjusted_match_len >= 0xf) {
55 u8 byte1 = min(adjusted_match_len - 0xf, 0xff);
56 ret = bitstream_put_byte(ostream, byte1);
60 ret = bitstream_put_two_bytes(ostream, adjusted_match_len);
65 return bitstream_put_bits(ostream,
66 match_offset ^ (1 << offset_bsr), offset_bsr);
70 xpress_write_compressed_literals(struct output_bitstream *ostream,
71 const u32 match_tab[],
73 const u16 codewords[],
76 for (unsigned i = 0; i < num_matches; i++) {
78 u32 match = match_tab[i];
80 if (match >= XPRESS_NUM_CHARS) /* match */
81 ret = xpress_write_match(ostream, match, codewords,
83 else /* literal byte */
84 ret = bitstream_put_bits(ostream, codewords[match],
89 return bitstream_put_bits(ostream, codewords[XPRESS_END_OF_DATA],
90 lens[XPRESS_END_OF_DATA]);
94 xpress_record_literal(u8 literal, void *__freq_tab)
96 freq_t *freq_tab = __freq_tab;
102 xpress_record_match(unsigned match_offset, unsigned match_len,
103 void *freq_tab, void *ignore)
105 wimlib_assert(match_len >= XPRESS_MIN_MATCH &&
106 match_len <= XPRESS_MAX_MATCH);
107 wimlib_assert(match_offset >= XPRESS_MIN_OFFSET &&
108 match_offset <= XPRESS_MAX_OFFSET);
111 * The intermediate representation of XPRESS matches is as follows:
114 * ---- -----------------------------------------------------------
116 * 16-31 match offset (XPRESS_MIN_OFFSET < x < XPRESS_MAX_OFFSET)
118 * 0-15 adjusted match length (0 <= x <= XPRESS_MAX_MATCH - XPRESS_MIN_MATCH)
120 * Literals are simply represented as themselves and can be
121 * distinguished from matches by the fact that only literals will have
122 * the upper three bytes completely clear. */
124 u32 adjusted_match_len = match_len - XPRESS_MIN_MATCH;
125 u32 len_hdr = min(adjusted_match_len, 0xf);
126 u32 offset_bsr = bsr32(match_offset);
127 u32 sym = len_hdr | (offset_bsr << 4) | XPRESS_NUM_CHARS;
128 ((freq_t*)freq_tab)[sym]++;
129 return adjusted_match_len | (match_offset << 16);
132 static const struct lz_params xpress_lz_params = {
133 .min_match = XPRESS_MIN_MATCH,
134 .max_match = XPRESS_MAX_MATCH,
138 .max_lazy_match = 16,
143 * Performs XPRESS compression on a block of data.
145 * @__uncompressed_data: Pointer to the data to be compressed.
146 * @uncompressed_len: Length, in bytes, of the data to be compressed.
147 * @__compressed_data: Pointer to a location at least (@uncompressed_len - 1)
148 * bytes long into which the compressed data may be
150 * @compressed_len_ret: A pointer to an unsigned int into which the length of
151 * the compressed data may be returned.
153 * Returns zero if compression was successfully performed. In that case
154 * @compressed_data and @compressed_len_ret will contain the compressed data and
155 * its length. A return value of nonzero means that compressing the data did
156 * not reduce its size, and @compressed_data will not contain the full
160 xpress_compress(const void *__uncompressed_data, unsigned uncompressed_len,
161 void *__compressed_data, unsigned *compressed_len_ret)
163 const u8 *uncompressed_data = __uncompressed_data;
164 u8 *compressed_data = __compressed_data;
165 struct output_bitstream ostream;
166 u32 match_tab[uncompressed_len];
167 freq_t freq_tab[XPRESS_NUM_SYMBOLS];
168 u16 codewords[XPRESS_NUM_SYMBOLS];
169 u8 lens[XPRESS_NUM_SYMBOLS];
170 unsigned num_matches;
171 unsigned compressed_len;
175 /* XPRESS requires 256 bytes of overhead for the Huffman tables, so it's
176 * impossible cannot compress 256 bytes or less of data to less than the
179 * +1 to take into account that the buffer for compressed data is 1 byte
180 * smaller than the buffer for uncompressed data.
182 * +4 to take into account that init_output_bitstream() requires at
183 * least 4 bytes of data. */
184 if (uncompressed_len < XPRESS_NUM_SYMBOLS / 2 + 1 + 4)
187 ZERO_ARRAY(freq_tab);
188 num_matches = lz_analyze_block(uncompressed_data, uncompressed_len,
189 match_tab, xpress_record_match,
190 xpress_record_literal, freq_tab,
194 freq_tab[XPRESS_END_OF_DATA]++;
196 make_canonical_huffman_code(XPRESS_NUM_SYMBOLS, XPRESS_MAX_CODEWORD_LEN,
197 freq_tab, lens, codewords);
201 * It's tempting to output the 512 Huffman codeword lengths using the
202 * bitstream_put_bits() function. However, this is NOT correct because
203 * bitstream_put_bits() will output 2 bytes at a time in little-endian
204 * order, which is the order that is needed for the compressed literals.
205 * However, the bytes in the lengths table are in order, so they need to
206 * be written one at a time without using bitstream_put_bits().
208 * Because of this, init_output_bitstream() is not called until after
209 * the lengths table is output.
211 for (i = 0; i < XPRESS_NUM_SYMBOLS; i += 2)
212 *compressed_data++ = (lens[i] & 0xf) | (lens[i + 1] << 4);
214 init_output_bitstream(&ostream, compressed_data,
215 uncompressed_len - XPRESS_NUM_SYMBOLS / 2 - 1);
217 ret = xpress_write_compressed_literals(&ostream, match_tab,
218 num_matches, codewords, lens);
222 /* Flush any bits that are buffered. */
223 ret = flush_output_bitstream(&ostream);
227 /* Assert that there are no output bytes between the ostream.output
228 * pointer and the ostream.next_bit_output pointer. This can only
229 * happen if bytes had been written at the ostream.output pointer before
230 * the last bit word was written to the stream. But, this does not
231 * occur since xpress_write_match() always finishes by writing some bits
232 * (a Huffman symbol), and the bitstream was just flushed. */
233 wimlib_assert(ostream.output - ostream.next_bit_output == 2);
235 /* The length of the compressed data is supposed to be the value of the
236 * ostream.output pointer before flushing, which is now the
237 * output.next_bit_output pointer after flushing.
239 * There will be an extra 2 bytes at the ostream.bit_output pointer,
240 * which is zeroed out. (These 2 bytes may be either the last bytes in
241 * the compressed data, in which case they are actually unnecessary, or
242 * they may precede a number of bytes embedded into the bitstream.) */
243 if (ostream.bit_output >
244 (const u8*)__compressed_data + uncompressed_len - 3)
246 *(u16*)ostream.bit_output = cpu_to_le16(0);
247 compressed_len = ostream.next_bit_output - (const u8*)__compressed_data;
249 wimlib_assert(compressed_len <= uncompressed_len - 1);
251 *compressed_len_ret = compressed_len;
253 #ifdef ENABLE_VERIFY_COMPRESSION
254 /* Verify that we really get the same thing back when decompressing. */
255 u8 buf[uncompressed_len];
256 ret = xpress_decompress(__compressed_data, compressed_len, buf,
259 ERROR("xpress_compress(): Failed to decompress data we "
263 for (i = 0; i < uncompressed_len; i++) {
264 if (buf[i] != uncompressed_data[i]) {
265 ERROR("xpress_compress(): Data we compressed didn't "
266 "decompress to the original data (difference at "
267 "byte %u of %u)", i + 1, uncompressed_len);