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 /* Mapping: position slot => first match offset that uses that position slot.
38 const u32 lzx_position_base[LZX_MAX_POSITION_SLOTS] = {
39 0 , 1 , 2 , 3 , 4 , /* 0 --- 4 */
40 6 , 8 , 12 , 16 , 24 , /* 5 --- 9 */
41 32 , 48 , 64 , 96 , 128 , /* 10 --- 14 */
42 192 , 256 , 384 , 512 , 768 , /* 15 --- 19 */
43 1024 , 1536 , 2048 , 3072 , 4096 , /* 20 --- 24 */
44 6144 , 8192 , 12288 , 16384 , 24576 , /* 25 --- 29 */
45 32768 , 49152 , 65536 , 98304 , 131072 , /* 30 --- 34 */
46 196608 , 262144 , 393216 , 524288 , 655360 , /* 35 --- 39 */
47 786432 , 917504 , 1048576, 1179648, 1310720, /* 40 --- 44 */
48 1441792, 1572864, 1703936, 1835008, 1966080, /* 45 --- 49 */
52 /* Mapping: position slot => how many extra bits must be read and added to the
53 * corresponding position base to decode the match offset. */
54 #ifdef USE_LZX_EXTRA_BITS_ARRAY
55 const u8 lzx_extra_bits[LZX_MAX_POSITION_SLOTS] = {
70 /* LZX window size must be a power of 2 between 2^15 and 2^21, inclusively. */
72 lzx_window_size_valid(size_t window_size)
74 if (window_size == 0 || (u32)window_size != window_size)
76 u32 order = bsr32(window_size);
77 if (window_size != 1U << order)
79 return (order >= LZX_MIN_WINDOW_ORDER && order <= LZX_MAX_WINDOW_ORDER);
82 /* Given a valid LZX window size, return the number of symbols that will exist
83 * in the main Huffman code. */
85 lzx_get_num_main_syms(u32 window_size)
87 /* NOTE: the calculation *should* be as follows:
89 * u32 max_offset = window_size - LZX_MIN_MATCH_LEN;
90 * u32 max_formatted_offset = max_offset + LZX_OFFSET_OFFSET;
91 * u32 num_position_slots = 1 + lzx_get_position_slot_raw(max_formatted_offset);
93 * However since LZX_MIN_MATCH_LEN == LZX_OFFSET_OFFSET, we would get
94 * max_formatted_offset == window_size, which would bump the number of
95 * position slots up by 1 since every valid LZX window size is equal to
96 * a position base value. The format doesn't do this, and instead
97 * disallows matches with minimum length and maximum offset. This sets
98 * max_formatted_offset = window_size - 1, so instead we must calculate:
100 * num_position_slots = 1 + lzx_get_position_slot_raw(window_size - 1);
102 * ... which is the same as
104 * num_position_slots = lzx_get_position_slot_raw(window_size);
106 * ... since every valid window size is equal to a position base value.
108 unsigned num_position_slots = lzx_get_position_slot_raw(window_size);
110 /* Now calculate the number of main symbols as LZX_NUM_CHARS literal
111 * symbols, plus 8 symbols per position slot (since there are 8 possible
112 * length headers, and we need all (position slot, length header)
114 return LZX_NUM_CHARS + (num_position_slots << 3);
118 do_translate_target(s32 *target, s32 input_pos)
120 s32 abs_offset, rel_offset;
122 /* XXX: This assumes unaligned memory accesses are okay. */
123 rel_offset = le32_to_cpu(*target);
124 if (rel_offset >= -input_pos && rel_offset < LZX_WIM_MAGIC_FILESIZE) {
125 if (rel_offset < LZX_WIM_MAGIC_FILESIZE - input_pos) {
126 /* "good translation" */
127 abs_offset = rel_offset + input_pos;
129 /* "compensating translation" */
130 abs_offset = rel_offset - LZX_WIM_MAGIC_FILESIZE;
132 *target = cpu_to_le32(abs_offset);
137 undo_translate_target(s32 *target, s32 input_pos)
139 s32 abs_offset, rel_offset;
141 /* XXX: This assumes unaligned memory accesses are okay. */
142 abs_offset = le32_to_cpu(*target);
143 if (abs_offset >= 0) {
144 if (abs_offset < LZX_WIM_MAGIC_FILESIZE) {
145 /* "good translation" */
146 rel_offset = abs_offset - input_pos;
148 *target = cpu_to_le32(rel_offset);
151 if (abs_offset >= -input_pos) {
152 /* "compensating translation" */
153 rel_offset = abs_offset + LZX_WIM_MAGIC_FILESIZE;
155 *target = cpu_to_le32(rel_offset);
161 * Do or undo the 'E8' preprocessing used in LZX. Before compression, the
162 * uncompressed data is preprocessed by changing the targets of x86 CALL
163 * instructions from relative offsets to absolute offsets. After decompression,
164 * the translation is undone by changing the targets of x86 CALL instructions
165 * from absolute offsets to relative offsets.
167 * Note that despite its intent, E8 preprocessing can be done on any data even
168 * if it is not actually x86 machine code. In fact, E8 preprocessing appears to
169 * always be used in LZX-compressed resources in WIM files; there is no bit to
170 * indicate whether it is used or not, unlike in the LZX compressed format as
171 * used in cabinet files, where a bit is reserved for that purpose.
173 * E8 preprocessing is disabled in the last 6 bytes of the uncompressed data,
174 * which really means the 5-byte call instruction cannot start in the last 10
175 * bytes of the uncompressed data. This is one of the errors in the LZX
178 * E8 preprocessing does not appear to be disabled after the 32768th chunk of a
179 * WIM resource, which apparently is another difference from the LZX compression
180 * used in cabinet files.
182 * E8 processing is supposed to take the file size as a parameter, as it is used
183 * in calculating the translated jump targets. But in WIM files, this file size
184 * is always the same (LZX_WIM_MAGIC_FILESIZE == 12000000).
188 inline /* Although inlining the 'process_target' function still speeds up the
189 SSE2 case, it bloats the binary more. */
192 lzx_e8_filter(u8 *data, u32 size, void (*process_target)(s32 *, s32))
195 /* SSE2 vectorized implementation for x86_64. This speeds up LZX
196 * decompression by about 5-8% overall. (Usually --- the performance
197 * actually regresses slightly in the degenerate case that the data
198 * consists entirely of 0xe8 bytes. Also, this optimization affects
199 * compression as well, but the percentage improvement is less because
200 * LZX compression is much slower than LZX decompression. ) */
201 __m128i *p128 = (__m128i *)data;
202 u32 valid_mask = 0xFFFFFFFF;
204 if (size >= 32 && (uintptr_t)data % 16 == 0) {
205 __m128i * const end128 = p128 + size / 16 - 1;
207 /* Create a vector of all 0xe8 bytes */
208 const __m128i e8_bytes = _mm_set1_epi8(0xe8);
210 /* Iterate through the 16-byte vectors in the input. */
212 /* Compare the current 16-byte vector with the vector of
213 * all 0xe8 bytes. This produces 0xff where the byte is
214 * 0xe8 and 0x00 where it is not. */
215 __m128i cmpresult = _mm_cmpeq_epi8(*p128, e8_bytes);
217 /* Map the comparison results into a single 16-bit
218 * number. It will contain a 1 bit when the
219 * corresponding byte in the current 16-byte vector is
220 * an e8 byte. Note: the low-order bit corresponds to
221 * the first (lowest address) byte. */
222 u32 e8_mask = _mm_movemask_epi8(cmpresult);
225 /* If e8_mask is 0, then none of these 16 bytes
226 * have value 0xe8. No e8 translation is
227 * needed, and there is no restriction that
228 * carries over to the next 16 bytes. */
229 valid_mask = 0xFFFFFFFF;
231 /* At least one byte has value 0xe8.
233 * The AND with valid_mask accounts for the fact
234 * that we can't start an e8 translation that
235 * overlaps the previous one. */
236 while ((e8_mask &= valid_mask)) {
238 /* Count the number of trailing zeroes
239 * in e8_mask. This will produce the
240 * index of the byte, within the 16, at
241 * which the next e8 translation should
243 u32 bit = __builtin_ctz(e8_mask);
245 /* Do (or undo) the e8 translation. */
246 u8 *p8 = (u8 *)p128 + bit;
247 (*process_target)((s32 *)(p8 + 1),
250 /* Don't start an e8 translation in the
252 valid_mask &= ~((u32)0x1F << bit);
254 /* Moving on to the next vector. Shift and set
255 * valid_mask accordingly. */
257 valid_mask |= 0xFFFF0000;
259 } while (++p128 < end128);
263 while (!(valid_mask & 1)) {
269 #endif /* !__SSE2__ */
272 /* Finish any bytes that weren't processed by the vectorized
274 u8 *p8_end = data + size - 10;
277 (*process_target)((s32 *)(p8 + 1), p8 - data);
282 } while (p8 < p8_end);
287 lzx_do_e8_preprocessing(u8 *data, u32 size)
289 lzx_e8_filter(data, size, do_translate_target);
293 lzx_undo_e8_preprocessing(u8 *data, u32 size)
295 lzx_e8_filter(data, size, undo_translate_target);