4 * Header for decompression code shared by multiple compression formats.
6 * The author dedicates this file to the public domain.
7 * You can do whatever you want with this file.
10 #ifndef _WIMLIB_DECOMPRESS_COMMON_H
11 #define _WIMLIB_DECOMPRESS_COMMON_H
13 #include "wimlib/compiler.h"
14 #include "wimlib/types.h"
15 #include "wimlib/unaligned.h"
17 /* Structure that encapsulates a block of in-memory data being interpreted as a
18 * stream of bits, optionally with interwoven literal bytes. Bits are assumed
19 * to be stored in little endian 16-bit coding units, with the bits ordered high
21 struct input_bitstream {
23 /* Bits that have been read from the input buffer. The bits are
24 * left-justified; the next bit is always bit 31. */
27 /* Number of bits currently held in @bitbuf. */
30 /* Pointer to the next byte to be retrieved from the input buffer. */
33 /* Pointer past the end of the input buffer. */
37 /* Initialize a bitstream to read from the specified input buffer. */
39 init_input_bitstream(struct input_bitstream *is, const void *buffer, u32 size)
44 is->end = is->next + size;
47 /* Note: for performance reasons, the following methods don't return error codes
48 * to the caller if the input buffer is overrun. Instead, they just assume that
49 * all overrun data is zeroes. This has no effect on well-formed compressed
50 * data. The only disadvantage is that bad compressed data may go undetected,
51 * but even this is irrelevant if higher level code checksums the uncompressed
54 /* Ensure the bit buffer variable for the bitstream contains at least @num_bits
55 * bits. Following this, bitstream_peek_bits() and/or bitstream_remove_bits()
56 * may be called on the bitstream to peek or remove up to @num_bits bits. */
58 bitstream_ensure_bits(struct input_bitstream *is, const unsigned num_bits)
60 /* This currently works for at most 17 bits. */
62 if (is->bitsleft >= num_bits)
65 if (unlikely(is->end - is->next < 2))
68 is->bitbuf |= (u32)get_unaligned_u16_le(is->next) << (16 - is->bitsleft);
72 if (unlikely(num_bits == 17 && is->bitsleft == 16)) {
73 if (unlikely(is->end - is->next < 2))
76 is->bitbuf |= (u32)get_unaligned_u16_le(is->next);
87 /* Return the next @num_bits bits from the bitstream, without removing them.
88 * There must be at least @num_bits remaining in the buffer variable, from a
89 * previous call to bitstream_ensure_bits(). */
91 bitstream_peek_bits(const struct input_bitstream *is, const unsigned num_bits)
93 return (is->bitbuf >> 1) >> (sizeof(is->bitbuf) * 8 - num_bits - 1);
96 /* Remove @num_bits from the bitstream. There must be at least @num_bits
97 * remaining in the buffer variable, from a previous call to
98 * bitstream_ensure_bits(). */
100 bitstream_remove_bits(struct input_bitstream *is, unsigned num_bits)
102 is->bitbuf <<= num_bits;
103 is->bitsleft -= num_bits;
106 /* Remove and return @num_bits bits from the bitstream. There must be at least
107 * @num_bits remaining in the buffer variable, from a previous call to
108 * bitstream_ensure_bits(). */
110 bitstream_pop_bits(struct input_bitstream *is, unsigned num_bits)
112 u32 bits = bitstream_peek_bits(is, num_bits);
113 bitstream_remove_bits(is, num_bits);
117 /* Read and return the next @num_bits bits from the bitstream. */
119 bitstream_read_bits(struct input_bitstream *is, unsigned num_bits)
121 bitstream_ensure_bits(is, num_bits);
122 return bitstream_pop_bits(is, num_bits);
125 /* Read and return the next literal byte embedded in the bitstream. */
127 bitstream_read_byte(struct input_bitstream *is)
129 if (unlikely(is->end == is->next))
134 /* Read and return the next 16-bit integer embedded in the bitstream. */
136 bitstream_read_u16(struct input_bitstream *is)
140 if (unlikely(is->end - is->next < 2))
142 v = get_unaligned_u16_le(is->next);
147 /* Read and return the next 32-bit integer embedded in the bitstream. */
149 bitstream_read_u32(struct input_bitstream *is)
153 if (unlikely(is->end - is->next < 4))
155 v = get_unaligned_u32_le(is->next);
160 /* Read an array of literal bytes embedded in the bitstream. Return a pointer
161 * to the resulting array, or NULL if the read overflows the input buffer. */
162 static inline const u8 *
163 bitstream_read_bytes(struct input_bitstream *is, size_t count)
167 if (unlikely(is->end - is->next < count))
174 /* Align the input bitstream on a coding-unit boundary. */
176 bitstream_align(struct input_bitstream *is)
182 /* Needed alignment of decode_table parameter to make_huffman_decode_table().
184 * Reason: We may fill the entries with SSE instructions without worrying
185 * about dealing with the unaligned case. */
186 #define DECODE_TABLE_ALIGNMENT 16
188 /* Maximum supported symbol count for make_huffman_decode_table().
190 * Reason: In direct mapping entries, we store the symbol in 11 bits. */
191 #define DECODE_TABLE_MAX_SYMBOLS 2048
193 /* Maximum supported table bits for make_huffman_decode_table().
195 * Reason: In internal binary tree nodes, offsets are encoded in 14 bits.
196 * But the real limit is 13, because we allocate entries past the end of
197 * the direct lookup part of the table for binary tree nodes. (Note: if
198 * needed this limit could be removed by encoding the offsets relative to
199 * &decode_table[1 << table_bits].) */
200 #define DECODE_TABLE_MAX_TABLE_BITS 13
202 /* Maximum supported codeword length for make_huffman_decode_table().
204 * Reason: In direct mapping entries, we encode the codeword length in 5
205 * bits, and the top 2 bits can't both be set because that has special
207 #define DECODE_TABLE_MAX_CODEWORD_LEN 23
209 /* Reads and returns the next Huffman-encoded symbol from a bitstream. If the
210 * input data is exhausted, the Huffman symbol is decoded as if the missing bits
213 * XXX: This is mostly duplicated in lzms_decode_huffman_symbol() in
214 * lzms_decompress.c. */
215 static inline unsigned
216 read_huffsym(struct input_bitstream *istream, const u16 decode_table[],
217 unsigned table_bits, unsigned max_codeword_len)
222 bitstream_ensure_bits(istream, max_codeword_len);
224 /* Index the decode table by the next table_bits bits of the input. */
225 key_bits = bitstream_peek_bits(istream, table_bits);
226 entry = decode_table[key_bits];
227 if (likely(entry < 0xC000)) {
228 /* Fast case: The decode table directly provided the
229 * symbol and codeword length. The low 11 bits are the
230 * symbol, and the high 5 bits are the codeword length. */
231 bitstream_remove_bits(istream, entry >> 11);
232 return entry & 0x7FF;
234 /* Slow case: The codeword for the symbol is longer than
235 * table_bits, so the symbol does not have an entry
236 * directly in the first (1 << table_bits) entries of the
237 * decode table. Traverse the appropriate binary tree
238 * bit-by-bit to decode the symbol. */
239 bitstream_remove_bits(istream, table_bits);
241 key_bits = (entry & 0x3FFF) + bitstream_pop_bits(istream, 1);
242 } while ((entry = decode_table[key_bits]) >= 0xC000);
248 make_huffman_decode_table(u16 decode_table[], unsigned num_syms,
249 unsigned num_bits, const u8 lens[],
250 unsigned max_codeword_len);
253 copy_word_unaligned(const void *src, void *dst)
255 store_word_unaligned(load_word_unaligned(src), dst);
258 static inline machine_word_t
263 BUILD_BUG_ON(WORDSIZE != 4 && WORDSIZE != 8);
268 v |= v << ((WORDSIZE == 8) ? 32 : 0);
273 * Copy an LZ77 match at (dst - offset) to dst.
275 * The length and offset must be already validated --- that is, (dst - offset)
276 * can't underrun the output buffer, and (dst + length) can't overrun the output
277 * buffer. Also, the length cannot be 0.
279 * @winend points to the byte past the end of the output buffer.
280 * This function won't write any data beyond this position.
283 lz_copy(u8 *dst, u32 length, u32 offset, const u8 *winend, u32 min_length)
285 const u8 *src = dst - offset;
286 const u8 * const end = dst + length;
289 * Try to copy one machine word at a time. On i386 and x86_64 this is
290 * faster than copying one byte at a time, unless the data is
291 * near-random and all the matches have very short lengths. Note that
292 * since this requires unaligned memory accesses, it won't necessarily
293 * be faster on every architecture.
295 * Also note that we might copy more than the length of the match. For
296 * example, if a word is 8 bytes and the match is of length 5, then
297 * we'll simply copy 8 bytes. This is okay as long as we don't write
298 * beyond the end of the output buffer, hence the check for (winend -
299 * end >= WORDSIZE - 1).
301 if (UNALIGNED_ACCESS_IS_VERY_FAST &&
302 likely(winend - end >= WORDSIZE - 1))
305 if (offset >= WORDSIZE) {
306 /* The source and destination words don't overlap. */
308 /* To improve branch prediction, one iteration of this
309 * loop is unrolled. Most matches are short and will
310 * fail the first check. But if that check passes, then
311 * it becomes increasing likely that the match is long
312 * and we'll need to continue copying. */
314 copy_word_unaligned(src, dst);
320 copy_word_unaligned(src, dst);
326 } else if (offset == 1) {
328 /* Offset 1 matches are equivalent to run-length
329 * encoding of the previous byte. This case is common
330 * if the data contains many repeated bytes. */
332 machine_word_t v = repeat_byte(*(dst - 1));
334 store_word_unaligned(v, dst);
341 * We don't bother with special cases for other 'offset <
342 * WORDSIZE', which are usually rarer than 'offset == 1'. Extra
343 * checks will just slow things down. Actually, it's possible
344 * to handle all the 'offset < WORDSIZE' cases using the same
345 * code, but it still becomes more complicated doesn't seem any
346 * faster overall; it definitely slows down the more common
347 * 'offset == 1' case.
351 /* Fall back to a bytewise copy. */
353 if (min_length >= 2) {
357 if (min_length >= 3) {
361 if (min_length >= 4) {
370 #endif /* _WIMLIB_DECOMPRESS_COMMON_H */