2 * lzx_common.c - Common code for LZX compression and decompression.
6 * Copyright (C) 2012-2016 Eric Biggers
8 * This file is free software; you can redistribute it and/or modify it under
9 * the terms of the GNU Lesser General Public License as published by the Free
10 * Software Foundation; either version 3 of the License, or (at your option) any
13 * This file is distributed in the hope that it will be useful, but WITHOUT
14 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
15 * FOR A PARTICULAR PURPOSE. See the GNU Lesser General Public License for more
18 * You should have received a copy of the GNU Lesser General Public License
19 * along with this file; if not, see http://www.gnu.org/licenses/.
29 # include <emmintrin.h>
33 # include <immintrin.h>
36 #include "wimlib/bitops.h"
37 #include "wimlib/endianness.h"
38 #include "wimlib/lzx_common.h"
39 #include "wimlib/unaligned.h"
40 #include "wimlib/util.h"
42 /* Mapping: offset slot => first match offset that uses that offset slot. */
43 const u32 lzx_offset_slot_base[LZX_MAX_OFFSET_SLOTS + 1] = {
44 -2 , -1 , 0 , 1 , 2 , /* 0 --- 2 */
45 4 , 6 , 10 , 14 , 22 , /* 5 --- 9 */
46 30 , 46 , 62 , 94 , 126 , /* 10 --- 14 */
47 190 , 254 , 382 , 510 , 766 , /* 15 --- 19 */
48 1022 , 1534 , 2046 , 3070 , 4094 , /* 20 --- 24 */
49 6142 , 8190 , 12286 , 16382 , 24574 , /* 25 --- 29 */
50 32766 , 49150 , 65534 , 98302 , 131070 , /* 30 --- 34 */
51 196606 , 262142 , 393214 , 524286 , 655358 , /* 35 --- 39 */
52 786430 , 917502 , 1048574, 1179646, 1310718, /* 40 --- 44 */
53 1441790, 1572862, 1703934, 1835006, 1966078, /* 45 --- 49 */
57 /* Mapping: offset slot => how many extra bits must be read and added to the
58 * corresponding offset slot base to decode the match offset. */
59 const u8 lzx_extra_offset_bits[LZX_MAX_OFFSET_SLOTS] = {
72 /* Round the specified buffer size up to the next valid LZX window size, and
73 * return its order (log2). Or, if the buffer size is 0 or greater than the
74 * largest valid LZX window size, return 0. */
76 lzx_get_window_order(size_t max_bufsize)
78 if (max_bufsize == 0 || max_bufsize > LZX_MAX_WINDOW_SIZE)
81 return max(ilog2_ceil(max_bufsize), LZX_MIN_WINDOW_ORDER);
84 /* Given a valid LZX window order, return the number of symbols that will exist
85 * in the main Huffman code. */
87 lzx_get_num_main_syms(unsigned window_order)
89 /* Note: one would expect that the maximum match offset would be
90 * 'window_size - LZX_MIN_MATCH_LEN', which would occur if the first two
91 * bytes were to match the last two bytes. However, the format
92 * disallows this case. This reduces the number of needed offset slots
94 u32 window_size = (u32)1 << window_order;
95 u32 max_offset = window_size - LZX_MIN_MATCH_LEN - 1;
96 unsigned num_offset_slots = 30;
97 while (max_offset >= lzx_offset_slot_base[num_offset_slots])
100 return LZX_NUM_CHARS + (num_offset_slots * LZX_NUM_LEN_HEADERS);
104 do_translate_target(void *target, s32 input_pos)
106 s32 abs_offset, rel_offset;
108 rel_offset = get_unaligned_le32(target);
109 if (rel_offset >= -input_pos && rel_offset < LZX_WIM_MAGIC_FILESIZE) {
110 if (rel_offset < LZX_WIM_MAGIC_FILESIZE - input_pos) {
111 /* "good translation" */
112 abs_offset = rel_offset + input_pos;
114 /* "compensating translation" */
115 abs_offset = rel_offset - LZX_WIM_MAGIC_FILESIZE;
117 put_unaligned_le32(abs_offset, target);
122 undo_translate_target(void *target, s32 input_pos)
124 s32 abs_offset, rel_offset;
126 abs_offset = get_unaligned_le32(target);
127 if (abs_offset >= 0) {
128 if (abs_offset < LZX_WIM_MAGIC_FILESIZE) {
129 /* "good translation" */
130 rel_offset = abs_offset - input_pos;
131 put_unaligned_le32(rel_offset, target);
134 if (abs_offset >= -input_pos) {
135 /* "compensating translation" */
136 rel_offset = abs_offset + LZX_WIM_MAGIC_FILESIZE;
137 put_unaligned_le32(rel_offset, target);
143 * Do or undo the 'E8' preprocessing used in LZX. Before compression, the
144 * uncompressed data is preprocessed by changing the targets of x86 CALL
145 * instructions from relative offsets to absolute offsets. After decompression,
146 * the translation is undone by changing the targets of x86 CALL instructions
147 * from absolute offsets to relative offsets.
149 * Note that despite its intent, E8 preprocessing can be done on any data even
150 * if it is not actually x86 machine code. In fact, E8 preprocessing appears to
151 * always be used in LZX-compressed resources in WIM files; there is no bit to
152 * indicate whether it is used or not, unlike in the LZX compressed format as
153 * used in cabinet files, where a bit is reserved for that purpose.
155 * E8 preprocessing is disabled in the last 6 bytes of the uncompressed data,
156 * which really means the 5-byte call instruction cannot start in the last 10
157 * bytes of the uncompressed data. This is one of the errors in the LZX
160 * E8 preprocessing does not appear to be disabled after the 32768th chunk of a
161 * WIM resource, which apparently is another difference from the LZX compression
162 * used in cabinet files.
164 * E8 processing is supposed to take the file size as a parameter, as it is used
165 * in calculating the translated jump targets. But in WIM files, this file size
166 * is always the same (LZX_WIM_MAGIC_FILESIZE == 12000000).
169 lzx_e8_filter(u8 *data, u32 size, void (*process_target)(void *, s32))
172 #if !defined(__SSE2__) && !defined(__AVX2__)
174 * A worthwhile optimization is to push the end-of-buffer check into the
175 * relatively rare E8 case. This is possible if we replace the last six
176 * bytes of data with E8 bytes; then we are guaranteed to hit an E8 byte
177 * before reaching end-of-buffer. In addition, this scheme guarantees
178 * that no translation can begin following an E8 byte in the last 10
179 * bytes because a 4-byte offset containing E8 as its high byte is a
180 * large negative number that is not valid for translation. That is
181 * exactly what we need.
190 tail = &data[size - 6];
191 memcpy(saved_bytes, tail, 6);
192 memset(tail, 0xE8, 6);
199 (*process_target)(p + 1, p - data);
202 memcpy(tail, saved_bytes, 6);
204 /* SSE2 or AVX-2 optimized version for x86_64 */
212 # define ALIGNMENT_REQUIRED 32
214 # define ALIGNMENT_REQUIRED 16
217 /* Process one byte at a time until the pointer is properly aligned. */
218 while ((uintptr_t)p % ALIGNMENT_REQUIRED != 0) {
219 if (p >= data + size - 10)
221 if (*p == 0xE8 && (valid_mask & 1)) {
222 (*process_target)(p + 1, p - data);
227 valid_mask |= (u64)1 << 63;
230 if (data + size - p >= 64) {
232 /* Vectorized processing */
234 /* Note: we use a "trap" E8 byte to eliminate the need to check
235 * for end-of-buffer in the inner loop. This byte is carefully
236 * positioned so that it will never be changed by a previous
237 * translation before it is detected. */
239 u8 *trap = p + ((data + size - p) & ~31) - 32 + 4;
240 u8 saved_byte = *trap;
247 const __m256i e8_bytes = _mm256_set1_epi8(0xE8);
249 __m256i bytes = *(const __m256i *)p;
250 __m256i cmpresult = _mm256_cmpeq_epi8(bytes, e8_bytes);
251 e8_mask = _mm256_movemask_epi8(cmpresult);
257 const __m128i e8_bytes = _mm_set1_epi8(0xE8);
259 /* Read the next 32 bytes of data and test them
261 __m128i bytes1 = *(const __m128i *)p;
262 __m128i bytes2 = *(const __m128i *)(p + 16);
263 __m128i cmpresult1 = _mm_cmpeq_epi8(bytes1, e8_bytes);
264 __m128i cmpresult2 = _mm_cmpeq_epi8(bytes2, e8_bytes);
265 u32 mask1 = _mm_movemask_epi8(cmpresult1);
266 u32 mask2 = _mm_movemask_epi8(cmpresult2);
267 /* The masks have a bit set for each E8 byte.
268 * We stay in this fast inner loop as long as
269 * there are no E8 bytes. */
271 e8_mask = mask1 | (mask2 << 16);
278 /* Did we pass over data with no E8 bytes? */
282 /* Are we nearing end-of-buffer? */
286 /* Process the E8 bytes. However, the AND with
287 * 'valid_mask' ensures we never process an E8 byte that
288 * was itself part of a translation target. */
289 while ((e8_mask &= valid_mask)) {
290 unsigned bit = ffs32(e8_mask);
291 (*process_target)(p + bit + 1, p + bit - data);
292 valid_mask &= ~((u64)0x1F << bit);
296 valid_mask |= 0xFFFFFFFF00000000;
303 /* Approaching the end of the buffer; process one byte a time. */
304 while (p < data + size - 10) {
305 if (*p == 0xE8 && (valid_mask & 1)) {
306 (*process_target)(p + 1, p - data);
311 valid_mask |= (u64)1 << 63;
313 #endif /* __SSE2__ || __AVX2__ */
317 lzx_preprocess(u8 *data, u32 size)
319 lzx_e8_filter(data, size, do_translate_target);
323 lzx_postprocess(u8 *data, u32 size)
325 lzx_e8_filter(data, size, undo_translate_target);