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: offset slot => first match offset that uses that offset slot.
38 const u32 lzx_offset_slot_base[LZX_MAX_OFFSET_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: offset slot => how many extra bits must be read and added to the
53 * corresponding offset slot base to decode the match offset. */
54 const u8 lzx_extra_offset_bits[LZX_MAX_OFFSET_SLOTS] = {
68 /* Round the specified compression block size (not LZX block size) up to the
69 * next valid LZX window size, and return its order (log2). Or, if the block
70 * size is 0 or greater than the largest valid LZX window size, return 0. */
72 lzx_get_window_order(size_t max_block_size)
76 if (max_block_size == 0 || max_block_size > (1 << LZX_MAX_WINDOW_ORDER))
79 order = bsr32(max_block_size);
81 if ((1 << order) != max_block_size)
84 return max(order, LZX_MIN_WINDOW_ORDER);
87 /* Given a valid LZX window order, return the number of symbols that will exist
88 * in the main Huffman code. */
90 lzx_get_num_main_syms(unsigned window_order)
92 u32 window_size = 1 << window_order;
94 /* NOTE: the calculation *should* be as follows:
96 * u32 max_offset = window_size - LZX_MIN_MATCH_LEN;
97 * u32 max_adjusted_offset = max_offset + LZX_OFFSET_OFFSET;
98 * u32 num_offset_slots = 1 + lzx_get_offset_slot_raw(max_adjusted_offset);
100 * However since LZX_MIN_MATCH_LEN == LZX_OFFSET_OFFSET, we would get
101 * max_adjusted_offset == window_size, which would bump the number of
102 * offset slots up by 1 since every valid LZX window size is equal to a
103 * offset slot base value. The format doesn't do this, and instead
104 * disallows matches with minimum length and maximum offset. This sets
105 * max_adjusted_offset = window_size - 1, so instead we must calculate:
107 * num_offset_slots = 1 + lzx_get_offset_slot_raw(window_size - 1);
109 * ... which is the same as
111 * num_offset_slots = lzx_get_offset_slot_raw(window_size);
113 * ... since every valid window size is equal to an offset base value.
115 unsigned num_offset_slots = lzx_get_offset_slot_raw(window_size);
117 /* Now calculate the number of main symbols as LZX_NUM_CHARS literal
118 * symbols, plus 8 symbols per offset slot (since there are 8 possible
119 * length headers, and we need all (offset slot, length header)
121 return LZX_NUM_CHARS + (num_offset_slots << 3);
125 do_translate_target(sle32 *target, s32 input_pos)
127 s32 abs_offset, rel_offset;
129 /* XXX: This assumes unaligned memory accesses are okay. */
130 rel_offset = le32_to_cpu(*target);
131 if (rel_offset >= -input_pos && rel_offset < LZX_WIM_MAGIC_FILESIZE) {
132 if (rel_offset < LZX_WIM_MAGIC_FILESIZE - input_pos) {
133 /* "good translation" */
134 abs_offset = rel_offset + input_pos;
136 /* "compensating translation" */
137 abs_offset = rel_offset - LZX_WIM_MAGIC_FILESIZE;
139 *target = cpu_to_le32(abs_offset);
144 undo_translate_target(sle32 *target, s32 input_pos)
146 s32 abs_offset, rel_offset;
148 /* XXX: This assumes unaligned memory accesses are okay. */
149 abs_offset = le32_to_cpu(*target);
150 if (abs_offset >= 0) {
151 if (abs_offset < LZX_WIM_MAGIC_FILESIZE) {
152 /* "good translation" */
153 rel_offset = abs_offset - input_pos;
155 *target = cpu_to_le32(rel_offset);
158 if (abs_offset >= -input_pos) {
159 /* "compensating translation" */
160 rel_offset = abs_offset + LZX_WIM_MAGIC_FILESIZE;
162 *target = cpu_to_le32(rel_offset);
168 * Do or undo the 'E8' preprocessing used in LZX. Before compression, the
169 * uncompressed data is preprocessed by changing the targets of x86 CALL
170 * instructions from relative offsets to absolute offsets. After decompression,
171 * the translation is undone by changing the targets of x86 CALL instructions
172 * from absolute offsets to relative offsets.
174 * Note that despite its intent, E8 preprocessing can be done on any data even
175 * if it is not actually x86 machine code. In fact, E8 preprocessing appears to
176 * always be used in LZX-compressed resources in WIM files; there is no bit to
177 * indicate whether it is used or not, unlike in the LZX compressed format as
178 * used in cabinet files, where a bit is reserved for that purpose.
180 * E8 preprocessing is disabled in the last 6 bytes of the uncompressed data,
181 * which really means the 5-byte call instruction cannot start in the last 10
182 * bytes of the uncompressed data. This is one of the errors in the LZX
185 * E8 preprocessing does not appear to be disabled after the 32768th chunk of a
186 * WIM resource, which apparently is another difference from the LZX compression
187 * used in cabinet files.
189 * E8 processing is supposed to take the file size as a parameter, as it is used
190 * in calculating the translated jump targets. But in WIM files, this file size
191 * is always the same (LZX_WIM_MAGIC_FILESIZE == 12000000).
195 inline /* Although inlining the 'process_target' function still speeds up the
196 SSE2 case, it bloats the binary more. */
199 lzx_e8_filter(u8 *data, u32 size, void (*process_target)(sle32 *, s32))
202 /* SSE2 vectorized implementation for x86_64. This speeds up LZX
203 * decompression by about 5-8% overall. (Usually --- the performance
204 * actually regresses slightly in the degenerate case that the data
205 * consists entirely of 0xe8 bytes. Also, this optimization affects
206 * compression as well, but the percentage improvement is less because
207 * LZX compression is much slower than LZX decompression. ) */
208 __m128i *p128 = (__m128i *)data;
209 u32 valid_mask = 0xFFFFFFFF;
211 if (size >= 32 && (uintptr_t)data % 16 == 0) {
212 __m128i * const end128 = p128 + size / 16 - 1;
214 /* Create a vector of all 0xe8 bytes */
215 const __m128i e8_bytes = _mm_set1_epi8(0xe8);
217 /* Iterate through the 16-byte vectors in the input. */
219 /* Compare the current 16-byte vector with the vector of
220 * all 0xe8 bytes. This produces 0xff where the byte is
221 * 0xe8 and 0x00 where it is not. */
222 __m128i cmpresult = _mm_cmpeq_epi8(*p128, e8_bytes);
224 /* Map the comparison results into a single 16-bit
225 * number. It will contain a 1 bit when the
226 * corresponding byte in the current 16-byte vector is
227 * an e8 byte. Note: the low-order bit corresponds to
228 * the first (lowest address) byte. */
229 u32 e8_mask = _mm_movemask_epi8(cmpresult);
232 /* If e8_mask is 0, then none of these 16 bytes
233 * have value 0xe8. No e8 translation is
234 * needed, and there is no restriction that
235 * carries over to the next 16 bytes. */
236 valid_mask = 0xFFFFFFFF;
238 /* At least one byte has value 0xe8.
240 * The AND with valid_mask accounts for the fact
241 * that we can't start an e8 translation that
242 * overlaps the previous one. */
243 while ((e8_mask &= valid_mask)) {
245 /* Count the number of trailing zeroes
246 * in e8_mask. This will produce the
247 * index of the byte, within the 16, at
248 * which the next e8 translation should
250 u32 bit = __builtin_ctz(e8_mask);
252 /* Do (or undo) the e8 translation. */
253 u8 *p8 = (u8 *)p128 + bit;
254 (*process_target)((sle32 *)(p8 + 1),
257 /* Don't start an e8 translation in the
259 valid_mask &= ~((u32)0x1F << bit);
261 /* Moving on to the next vector. Shift and set
262 * valid_mask accordingly. */
264 valid_mask |= 0xFFFF0000;
266 } while (++p128 < end128);
270 while (!(valid_mask & 1)) {
276 #endif /* !__SSE2__ */
279 /* Finish any bytes that weren't processed by the vectorized
281 u8 *p8_end = data + size - 10;
284 (*process_target)((sle32 *)(p8 + 1), p8 - data);
289 } while (p8 < p8_end);
294 lzx_do_e8_preprocessing(u8 *data, u32 size)
296 lzx_e8_filter(data, size, do_translate_target);
300 lzx_undo_e8_preprocessing(u8 *data, u32 size)
302 lzx_e8_filter(data, size, undo_translate_target);