2 * lzx-common.c - Common data for LZX compression and decompression.
6 * Copyright (C) 2012, 2013 Eric Biggers
8 * This file is part of wimlib, a library for working with WIM files.
10 * wimlib is free software; you can redistribute it and/or modify it under the
11 * terms of the GNU General Public License as published by the Free
12 * Software Foundation; either version 3 of the License, or (at your option)
15 * wimlib is distributed in the hope that it will be useful, but WITHOUT ANY
16 * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
17 * A PARTICULAR PURPOSE. See the GNU General Public License for more
20 * You should have received a copy of the GNU General Public License
21 * along with wimlib; if not, see http://www.gnu.org/licenses/.
28 #include "wimlib/endianness.h"
29 #include "wimlib/lzx.h"
30 #include "wimlib/util.h"
33 # include <emmintrin.h>
36 /* LZX uses what it calls 'position slots' to represent match offsets.
37 * What this means is that a small 'position slot' number and a small
38 * offset from that slot are encoded instead of one large offset for
40 * - lzx_position_base is an index to the position slot bases
41 * - lzx_extra_bits states how many bits of offset-from-base data is needed.
44 const u32 lzx_position_base[LZX_MAX_POSITION_SLOTS] = {
45 0 , 1 , 2 , 3 , 4 , /* 0 --- 4 */
46 6 , 8 , 12 , 16 , 24 , /* 5 --- 9 */
47 32 , 48 , 64 , 96 , 128 , /* 10 --- 14 */
48 192 , 256 , 384 , 512 , 768 , /* 15 --- 19 */
49 1024 , 1536 , 2048 , 3072 , 4096 , /* 20 --- 24 */
50 6144 , 8192 , 12288 , 16384 , 24576 , /* 25 --- 29 */
51 32768 , 49152 , 65536 , 98304 , 131072 , /* 30 --- 34 */
52 196608 , 262144 , 393216 , 524288 , 655360 , /* 35 --- 39 */
53 786432 , 917504 , 1048576, 1179648, 1310720, /* 40 --- 44 */
54 1441792, 1572864, 1703936, 1835008, 1966080, /* 45 --- 49 */
58 #ifdef USE_LZX_EXTRA_BITS_ARRAY
59 const u8 lzx_extra_bits[LZX_MAX_POSITION_SLOTS] = {
74 /* LZX window size must be a power of 2 between 2^15 and 2^21, inclusively. */
76 lzx_window_size_valid(size_t window_size)
78 if (window_size == 0 || (u32)window_size != window_size)
80 u32 order = bsr32(window_size);
81 if (window_size != 1U << order)
83 return (order >= LZX_MIN_WINDOW_ORDER && order <= LZX_MAX_WINDOW_ORDER);
86 /* Given a valid LZX window size, return the number of symbols that will exist
87 * in the main Huffman code. */
89 lzx_get_num_main_syms(u32 window_size)
91 /* NOTE: the calculation *should* be as follows:
93 * u32 max_offset = window_size - LZX_MIN_MATCH_LEN;
94 * u32 max_formatted_offset = max_offset + LZX_OFFSET_OFFSET;
95 * u32 num_position_slots = 1 + lzx_get_position_slot_raw(max_formatted_offset);
97 * However since LZX_MIN_MATCH_LEN == LZX_OFFSET_OFFSET, we would get
98 * max_formatted_offset == window_size, which would bump the number of
99 * position slots up by 1 since every valid LZX window size is equal to
100 * a position base value. The format doesn't do this, and instead
101 * disallows matches with minimum length and maximum offset. This sets
102 * max_formatted_offset = window_size - 1, so instead we must calculate:
104 * num_position_slots = 1 + lzx_get_position_slot_raw(window_size - 1);
106 * ... which is the same as
108 * num_position_slots = lzx_get_position_slot_raw(window_size);
110 * ... since every valid window size is equal to a position base value.
112 unsigned num_position_slots = lzx_get_position_slot_raw(window_size);
114 /* Now calculate the number of main symbols as LZX_NUM_CHARS literal
115 * symbols, plus 8 symbols per position slot (since there are 8 possible
116 * length headers, and we need all (position slot, length header)
118 return LZX_NUM_CHARS + (num_position_slots << 3);
122 do_translate_target(s32 *target, s32 input_pos)
124 s32 abs_offset, rel_offset;
126 /* XXX: This assumes unaligned memory accesses are okay. */
127 rel_offset = le32_to_cpu(*target);
128 if (rel_offset >= -input_pos && rel_offset < LZX_WIM_MAGIC_FILESIZE) {
129 if (rel_offset < LZX_WIM_MAGIC_FILESIZE - input_pos) {
130 /* "good translation" */
131 abs_offset = rel_offset + input_pos;
133 /* "compensating translation" */
134 abs_offset = rel_offset - LZX_WIM_MAGIC_FILESIZE;
136 *target = cpu_to_le32(abs_offset);
141 undo_translate_target(s32 *target, s32 input_pos)
143 s32 abs_offset, rel_offset;
145 /* XXX: This assumes unaligned memory accesses are okay. */
146 abs_offset = le32_to_cpu(*target);
147 if (abs_offset >= 0) {
148 if (abs_offset < LZX_WIM_MAGIC_FILESIZE) {
149 /* "good translation" */
150 rel_offset = abs_offset - input_pos;
152 *target = cpu_to_le32(rel_offset);
155 if (abs_offset >= -input_pos) {
156 /* "compensating translation" */
157 rel_offset = abs_offset + LZX_WIM_MAGIC_FILESIZE;
159 *target = cpu_to_le32(rel_offset);
165 * Do or undo the 'E8' preprocessing used in LZX. Before compression, the
166 * uncompressed data is preprocessed by changing the targets of x86 CALL
167 * instructions from relative offsets to absolute offsets. After decompression,
168 * the translation is undone by changing the targets of x86 CALL instructions
169 * from absolute offsets to relative offsets.
171 * Note that despite its intent, E8 preprocessing can be done on any data even
172 * if it is not actually x86 machine code. In fact, E8 preprocessing appears to
173 * always be used in LZX-compressed resources in WIM files; there is no bit to
174 * indicate whether it is used or not, unlike in the LZX compressed format as
175 * used in cabinet files, where a bit is reserved for that purpose.
177 * E8 preprocessing is disabled in the last 6 bytes of the uncompressed data,
178 * which really means the 5-byte call instruction cannot start in the last 10
179 * bytes of the uncompressed data. This is one of the errors in the LZX
182 * E8 preprocessing does not appear to be disabled after the 32768th chunk of a
183 * WIM resource, which apparently is another difference from the LZX compression
184 * used in cabinet files.
186 * E8 processing is supposed to take the file size as a parameter, as it is used
187 * in calculating the translated jump targets. But in WIM files, this file size
188 * is always the same (LZX_WIM_MAGIC_FILESIZE == 12000000).
192 inline /* Although inlining the 'process_target' function still speeds up the
193 SSE2 case, it bloats the binary more. */
196 lzx_e8_filter(u8 *data, s32 size, void (*process_target)(s32 *, s32))
199 /* SSE2 vectorized implementation for x86_64. This speeds up LZX
200 * decompression by about 5-8% overall. (Usually --- the performance
201 * actually regresses slightly in the degenerate case that the data
202 * consists entirely of 0xe8 bytes. Also, this optimization affects
203 * compression as well, but the percentage improvement is less because
204 * LZX compression is much slower than LZX decompression. ) */
205 __m128i *p128 = (__m128i *)data;
206 u32 valid_mask = 0xFFFFFFFF;
208 if (size >= 32 && (uintptr_t)data % 16 == 0) {
209 __m128i * const end128 = p128 + size / 16 - 1;
211 /* Create a vector of all 0xe8 bytes */
212 const __m128i e8_bytes = _mm_set1_epi8(0xe8);
214 /* Iterate through the 16-byte vectors in the input. */
216 /* Compare the current 16-byte vector with the vector of
217 * all 0xe8 bytes. This produces 0xff where the byte is
218 * 0xe8 and 0x00 where it is not. */
219 __m128i cmpresult = _mm_cmpeq_epi8(*p128, e8_bytes);
221 /* Map the comparison results into a single 16-bit
222 * number. It will contain a 1 bit when the
223 * corresponding byte in the current 16-byte vector is
224 * an e8 byte. Note: the low-order bit corresponds to
225 * the first (lowest address) byte. */
226 u32 e8_mask = _mm_movemask_epi8(cmpresult);
229 /* If e8_mask is 0, then none of these 16 bytes
230 * have value 0xe8. No e8 translation is
231 * needed, and there is no restriction that
232 * carries over to the next 16 bytes. */
233 valid_mask = 0xFFFFFFFF;
235 /* At least one byte has value 0xe8.
237 * The AND with valid_mask accounts for the fact
238 * that we can't start an e8 translation that
239 * overlaps the previous one. */
240 while ((e8_mask &= valid_mask)) {
242 /* Count the number of trailing zeroes
243 * in e8_mask. This will produce the
244 * index of the byte, within the 16, at
245 * which the next e8 translation should
247 u32 bit = __builtin_ctz(e8_mask);
249 /* Do (or undo) the e8 translation. */
250 u8 *p8 = (u8 *)p128 + bit;
251 (*process_target)((s32 *)(p8 + 1),
254 /* Don't start an e8 translation in the
256 valid_mask &= ~((u32)0x1F << bit);
258 /* Moving on to the next vector. Shift and set
259 * valid_mask accordingly. */
261 valid_mask |= 0xFFFF0000;
263 } while (++p128 < end128);
267 while (!(valid_mask & 1)) {
273 #endif /* !__SSE2__ */
276 /* Finish any bytes that weren't processed by the vectorized
278 u8 *p8_end = data + size - 10;
281 (*process_target)((s32 *)(p8 + 1), p8 - data);
286 } while (p8 < p8_end);
291 lzx_do_e8_preprocessing(u8 *data, s32 size)
293 lzx_e8_filter(data, size, do_translate_target);
297 lzx_undo_e8_preprocessing(u8 *data, s32 size)
299 lzx_e8_filter(data, size, undo_translate_target);