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/endianness.h"
15 #include "wimlib/types.h"
16 #include "wimlib/unaligned.h"
18 /* Structure that encapsulates a block of in-memory data being interpreted as a
19 * stream of bits, optionally with interwoven literal bytes. Bits are assumed
20 * to be stored in little endian 16-bit coding units, with the bits ordered high
22 struct input_bitstream {
24 /* Bits that have been read from the input buffer. The bits are
25 * left-justified; the next bit is always bit 31. */
28 /* Number of bits currently held in @bitbuf. */
31 /* Pointer to the next byte to be retrieved from the input buffer. */
34 /* Pointer past the end of the input buffer. */
38 /* Initialize a bitstream to read from the specified input buffer. */
40 init_input_bitstream(struct input_bitstream *is, const void *buffer, u32 size)
45 is->end = is->next + size;
48 /* Note: for performance reasons, the following methods don't return error codes
49 * to the caller if the input buffer is overrun. Instead, they just assume that
50 * all overrun data is zeroes. This has no effect on well-formed compressed
51 * data. The only disadvantage is that bad compressed data may go undetected,
52 * but even this is irrelevant if higher level code checksums the uncompressed
55 /* Ensure the bit buffer variable for the bitstream contains at least @num_bits
56 * bits. Following this, bitstream_peek_bits() and/or bitstream_remove_bits()
57 * may be called on the bitstream to peek or remove up to @num_bits bits. */
59 bitstream_ensure_bits(struct input_bitstream *is, const unsigned num_bits)
61 /* This currently works for at most 17 bits. */
63 if (is->bitsleft >= num_bits)
66 if (unlikely(is->end - is->next < 2))
69 is->bitbuf |= (u32)get_unaligned_u16_le(is->next) << (16 - is->bitsleft);
73 if (unlikely(num_bits == 17 && is->bitsleft == 16)) {
74 if (unlikely(is->end - is->next < 2))
77 is->bitbuf |= (u32)get_unaligned_u16_le(is->next);
88 /* Return the next @num_bits bits from the bitstream, without removing them.
89 * There must be at least @num_bits remaining in the buffer variable, from a
90 * previous call to bitstream_ensure_bits(). */
92 bitstream_peek_bits(const struct input_bitstream *is, const unsigned num_bits)
94 if (unlikely(num_bits == 0))
96 return is->bitbuf >> (32 - num_bits);
99 /* Remove @num_bits from the bitstream. There must be at least @num_bits
100 * remaining in the buffer variable, from a previous call to
101 * bitstream_ensure_bits(). */
103 bitstream_remove_bits(struct input_bitstream *is, unsigned num_bits)
105 is->bitbuf <<= num_bits;
106 is->bitsleft -= num_bits;
109 /* Remove and return @num_bits bits from the bitstream. There must be at least
110 * @num_bits remaining in the buffer variable, from a previous call to
111 * bitstream_ensure_bits(). */
113 bitstream_pop_bits(struct input_bitstream *is, unsigned num_bits)
115 u32 bits = bitstream_peek_bits(is, num_bits);
116 bitstream_remove_bits(is, num_bits);
120 /* Read and return the next @num_bits bits from the bitstream. */
122 bitstream_read_bits(struct input_bitstream *is, unsigned num_bits)
124 bitstream_ensure_bits(is, num_bits);
125 return bitstream_pop_bits(is, num_bits);
128 /* Read and return the next literal byte embedded in the bitstream. */
130 bitstream_read_byte(struct input_bitstream *is)
132 if (unlikely(is->end == is->next))
137 /* Read and return the next 16-bit integer embedded in the bitstream. */
139 bitstream_read_u16(struct input_bitstream *is)
143 if (unlikely(is->end - is->next < 2))
145 v = get_unaligned_u16_le(is->next);
150 /* Read and return the next 32-bit integer embedded in the bitstream. */
152 bitstream_read_u32(struct input_bitstream *is)
156 if (unlikely(is->end - is->next < 4))
158 v = get_unaligned_u32_le(is->next);
163 /* Read an array of literal bytes embedded in the bitstream. Return a pointer
164 * to the resulting array, or NULL if the read overflows the input buffer. */
165 static inline const u8 *
166 bitstream_read_bytes(struct input_bitstream *is, size_t count)
170 if (unlikely(is->end - is->next < count))
177 /* Align the input bitstream on a coding-unit boundary. */
179 bitstream_align(struct input_bitstream *is)
185 /* Needed alignment of decode_table parameter to make_huffman_decode_table().
187 * Reason: We may fill the entries with SSE instructions without worrying
188 * about dealing with the unaligned case. */
189 #define DECODE_TABLE_ALIGNMENT 16
191 /* Maximum supported symbol count for make_huffman_decode_table().
193 * Reason: In direct mapping entries, we store the symbol in 11 bits. */
194 #define DECODE_TABLE_MAX_SYMBOLS 2048
196 /* Maximum supported table bits for make_huffman_decode_table().
198 * Reason: In internal binary tree nodes, offsets are encoded in 14 bits.
199 * But the real limit is 13, because we allocate entries past the end of
200 * the direct lookup part of the table for binary tree nodes. (Note: if
201 * needed this limit could be removed by encoding the offsets relative to
202 * &decode_table[1 << table_bits].) */
203 #define DECODE_TABLE_MAX_TABLE_BITS 13
205 /* Maximum supported codeword length for make_huffman_decode_table().
207 * Reason: In direct mapping entries, we encode the codeword length in 5
208 * bits, and the top 2 bits can't both be set because that has special
210 #define DECODE_TABLE_MAX_CODEWORD_LEN 23
212 /* Reads and returns the next Huffman-encoded symbol from a bitstream. If the
213 * input data is exhausted, the Huffman symbol is decoded as if the missing bits
216 * XXX: This is mostly duplicated in lzms_decode_huffman_symbol() in
217 * lzms_decompress.c. */
219 read_huffsym(struct input_bitstream *istream, const u16 decode_table[],
220 unsigned table_bits, unsigned max_codeword_len)
225 bitstream_ensure_bits(istream, max_codeword_len);
227 /* Index the decode table by the next table_bits bits of the input. */
228 key_bits = bitstream_peek_bits(istream, table_bits);
229 entry = decode_table[key_bits];
230 if (likely(entry < 0xC000)) {
231 /* Fast case: The decode table directly provided the
232 * symbol and codeword length. The low 11 bits are the
233 * symbol, and the high 5 bits are the codeword length. */
234 bitstream_remove_bits(istream, entry >> 11);
235 return entry & 0x7FF;
237 /* Slow case: The codeword for the symbol is longer than
238 * table_bits, so the symbol does not have an entry
239 * directly in the first (1 << table_bits) entries of the
240 * decode table. Traverse the appropriate binary tree
241 * bit-by-bit to decode the symbol. */
242 bitstream_remove_bits(istream, table_bits);
244 key_bits = (entry & 0x3FFF) + bitstream_pop_bits(istream, 1);
245 } while ((entry = decode_table[key_bits]) >= 0xC000);
251 make_huffman_decode_table(u16 decode_table[], unsigned num_syms,
252 unsigned num_bits, const u8 lens[],
253 unsigned max_codeword_len);
256 copy_word_unaligned(const void *src, void *dst)
258 store_word_unaligned(load_word_unaligned(src), dst);
261 static inline machine_word_t
266 BUILD_BUG_ON(WORDSIZE != 4 && WORDSIZE != 8);
271 v |= v << ((WORDSIZE == 8) ? 32 : 0);
276 * Copy an LZ77 match at (dst - offset) to dst.
278 * The length and offset must be already validated --- that is, (dst - offset)
279 * can't underrun the output buffer, and (dst + length) can't overrun the output
280 * buffer. Also, the length cannot be 0.
282 * @winend points to the byte past the end of the output buffer.
283 * This function won't write any data beyond this position.
286 lz_copy(u8 *dst, u32 length, u32 offset, const u8 *winend, u32 min_length)
288 const u8 *src = dst - offset;
289 const u8 * const end = dst + length;
292 * Try to copy one machine word at a time. On i386 and x86_64 this is
293 * faster than copying one byte at a time, unless the data is
294 * near-random and all the matches have very short lengths. Note that
295 * since this requires unaligned memory accesses, it won't necessarily
296 * be faster on every architecture.
298 * Also note that we might copy more than the length of the match. For
299 * example, if a word is 8 bytes and the match is of length 5, then
300 * we'll simply copy 8 bytes. This is okay as long as we don't write
301 * beyond the end of the output buffer, hence the check for (winend -
302 * end >= WORDSIZE - 1).
304 if (UNALIGNED_ACCESS_IS_VERY_FAST &&
305 likely(winend - end >= WORDSIZE - 1))
308 if (offset >= WORDSIZE) {
309 /* The source and destination words don't overlap. */
311 /* To improve branch prediction, one iteration of this
312 * loop is unrolled. Most matches are short and will
313 * fail the first check. But if that check passes, then
314 * it becomes increasing likely that the match is long
315 * and we'll need to continue copying. */
317 copy_word_unaligned(src, dst);
323 copy_word_unaligned(src, dst);
329 } else if (offset == 1) {
331 /* Offset 1 matches are equivalent to run-length
332 * encoding of the previous byte. This case is common
333 * if the data contains many repeated bytes. */
335 machine_word_t v = repeat_byte(*(dst - 1));
337 store_word_unaligned(v, dst);
344 * We don't bother with special cases for other 'offset <
345 * WORDSIZE', which are usually rarer than 'offset == 1'. Extra
346 * checks will just slow things down. Actually, it's possible
347 * to handle all the 'offset < WORDSIZE' cases using the same
348 * code, but it still becomes more complicated doesn't seem any
349 * faster overall; it definitely slows down the more common
350 * 'offset == 1' case.
354 /* Fall back to a bytewise copy. */
356 if (min_length >= 2) {
360 if (min_length >= 3) {
364 if (min_length >= 4) {
373 #endif /* _WIMLIB_DECOMPRESS_COMMON_H */