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/assert.h"
14 #include "wimlib/compiler.h"
15 #include "wimlib/endianness.h"
16 #include "wimlib/types.h"
17 #include "wimlib/unaligned.h"
19 /* Structure that encapsulates a block of in-memory data being interpreted as a
20 * stream of bits, optionally with interwoven literal bytes. Bits are assumed
21 * to be stored in little endian 16-bit coding units, with the bits ordered high
23 struct input_bitstream {
25 /* Bits that have been read from the input buffer. The bits are
26 * left-justified; the next bit is always bit 31. */
29 /* Number of bits currently held in @bitbuf. */
32 /* Pointer to the next byte to be retrieved from the input buffer. */
35 /* Pointer past the end of the input buffer. */
39 /* Initialize a bitstream to read from the specified input buffer. */
41 init_input_bitstream(struct input_bitstream *is, const void *buffer, u32 size)
46 is->end = is->next + size;
49 /* Note: for performance reasons, the following methods don't return error codes
50 * to the caller if the input buffer is overrun. Instead, they just assume that
51 * all overrun data is zeroes. This has no effect on well-formed compressed
52 * data. The only disadvantage is that bad compressed data may go undetected,
53 * but even this is irrelevant if higher level code checksums the uncompressed
56 /* Ensure the bit buffer variable for the bitstream contains at least @num_bits
57 * bits. Following this, bitstream_peek_bits() and/or bitstream_remove_bits()
58 * may be called on the bitstream to peek or remove up to @num_bits bits. */
60 bitstream_ensure_bits(struct input_bitstream *is, const unsigned num_bits)
62 /* This currently works for at most 17 bits. */
63 wimlib_assert2(num_bits <= 17);
65 if (is->bitsleft >= num_bits)
68 if (unlikely(is->end - is->next < 2))
71 is->bitbuf |= (u32)get_unaligned_u16_le(is->next) << (16 - is->bitsleft);
75 if (unlikely(num_bits == 17 && is->bitsleft == 16)) {
76 if (unlikely(is->end - is->next < 2))
79 is->bitbuf |= (u32)get_unaligned_u16_le(is->next);
90 /* Return the next @num_bits bits from the bitstream, without removing them.
91 * There must be at least @num_bits remaining in the buffer variable, from a
92 * previous call to bitstream_ensure_bits(). */
94 bitstream_peek_bits(const struct input_bitstream *is, const unsigned num_bits)
96 if (unlikely(num_bits == 0))
98 return is->bitbuf >> (32 - num_bits);
101 /* Remove @num_bits from the bitstream. There must be at least @num_bits
102 * remaining in the buffer variable, from a previous call to
103 * bitstream_ensure_bits(). */
105 bitstream_remove_bits(struct input_bitstream *is, unsigned num_bits)
107 is->bitbuf <<= num_bits;
108 is->bitsleft -= num_bits;
111 /* Remove and return @num_bits bits from the bitstream. There must be at least
112 * @num_bits remaining in the buffer variable, from a previous call to
113 * bitstream_ensure_bits(). */
115 bitstream_pop_bits(struct input_bitstream *is, unsigned num_bits)
117 u32 bits = bitstream_peek_bits(is, num_bits);
118 bitstream_remove_bits(is, num_bits);
122 /* Read and return the next @num_bits bits from the bitstream. */
124 bitstream_read_bits(struct input_bitstream *is, unsigned num_bits)
126 bitstream_ensure_bits(is, num_bits);
127 return bitstream_pop_bits(is, num_bits);
130 /* Read and return the next literal byte embedded in the bitstream. */
132 bitstream_read_byte(struct input_bitstream *is)
134 if (unlikely(is->end == is->next))
139 /* Read and return the next 16-bit integer embedded in the bitstream. */
141 bitstream_read_u16(struct input_bitstream *is)
145 if (unlikely(is->end - is->next < 2))
147 v = get_unaligned_u16_le(is->next);
152 /* Read and return the next 32-bit integer embedded in the bitstream. */
154 bitstream_read_u32(struct input_bitstream *is)
158 if (unlikely(is->end - is->next < 4))
160 v = get_unaligned_u32_le(is->next);
165 /* Read an array of literal bytes embedded in the bitstream. Return a pointer
166 * to the resulting array, or NULL if the read overflows the input buffer. */
167 static inline const u8 *
168 bitstream_read_bytes(struct input_bitstream *is, size_t count)
172 if (unlikely(is->end - is->next < count))
179 /* Align the input bitstream on a coding-unit boundary. */
181 bitstream_align(struct input_bitstream *is)
187 /* Needed alignment of decode_table parameter to make_huffman_decode_table().
189 * Reason: We may fill the entries with SSE instructions without worrying
190 * about dealing with the unaligned case. */
191 #define DECODE_TABLE_ALIGNMENT 16
193 /* Maximum supported symbol count for make_huffman_decode_table().
195 * Reason: In direct mapping entries, we store the symbol in 11 bits. */
196 #define DECODE_TABLE_MAX_SYMBOLS 2048
198 /* Maximum supported table bits for make_huffman_decode_table().
200 * Reason: In internal binary tree nodes, offsets are encoded in 14 bits.
201 * But the real limit is 13, because we allocate entries past the end of
202 * the direct lookup part of the table for binary tree nodes. (Note: if
203 * needed this limit could be removed by encoding the offsets relative to
204 * &decode_table[1 << table_bits].) */
205 #define DECODE_TABLE_MAX_TABLE_BITS 13
207 /* Maximum supported codeword length for make_huffman_decode_table().
209 * Reason: In direct mapping entries, we encode the codeword length in 5
210 * bits, and the top 2 bits can't both be set because that has special
212 #define DECODE_TABLE_MAX_CODEWORD_LEN 23
214 /* Reads and returns the next Huffman-encoded symbol from a bitstream. If the
215 * input data is exhausted, the Huffman symbol is decoded as if the missing bits
218 * XXX: This is mostly duplicated in lzms_huffman_decode_symbol() in
219 * lzms-decompress.c. */
221 read_huffsym(struct input_bitstream *istream, const u16 decode_table[],
222 unsigned table_bits, unsigned max_codeword_len)
227 bitstream_ensure_bits(istream, max_codeword_len);
229 /* Index the decode table by the next table_bits bits of the input. */
230 key_bits = bitstream_peek_bits(istream, table_bits);
231 entry = decode_table[key_bits];
232 if (likely(entry < 0xC000)) {
233 /* Fast case: The decode table directly provided the
234 * symbol and codeword length. The low 11 bits are the
235 * symbol, and the high 5 bits are the codeword length. */
236 bitstream_remove_bits(istream, entry >> 11);
237 return entry & 0x7FF;
239 /* Slow case: The codeword for the symbol is longer than
240 * table_bits, so the symbol does not have an entry
241 * directly in the first (1 << table_bits) entries of the
242 * decode table. Traverse the appropriate binary tree
243 * bit-by-bit to decode the symbol. */
244 bitstream_remove_bits(istream, table_bits);
246 key_bits = (entry & 0x3FFF) + bitstream_pop_bits(istream, 1);
247 } while ((entry = decode_table[key_bits]) >= 0xC000);
253 make_huffman_decode_table(u16 decode_table[], unsigned num_syms,
254 unsigned num_bits, const u8 lens[],
255 unsigned max_codeword_len);
259 * Copy an LZ77 match at (dst - offset) to dst.
261 * The length and offset must be already validated --- that is, (dst - offset)
262 * can't underrun the output buffer, and (dst + length) can't overrun the output
263 * buffer. Also, the length cannot be 0.
265 * @winend points to the byte past the end of the output buffer.
266 * This function won't write any data beyond this position.
269 lz_copy(u8 *dst, u32 length, u32 offset, const u8 *winend, u32 min_length)
271 const u8 *src = dst - offset;
272 const u8 * const end = dst + length;
275 * Try to copy one machine word at a time. On i386 and x86_64 this is
276 * faster than copying one byte at a time, unless the data is
277 * near-random and all the matches have very short lengths. Note that
278 * since this requires unaligned memory accesses, it won't necessarily
279 * be faster on every architecture.
281 * Also note that we might copy more than the length of the match. For
282 * example, if a word is 8 bytes and the match is of length 5, then
283 * we'll simply copy 8 bytes. This is okay as long as we don't write
284 * beyond the end of the output buffer, hence the check for (winend -
285 * end >= WORDSIZE - 1).
287 if (UNALIGNED_ACCESS_IS_VERY_FAST &&
288 likely(winend - end >= WORDSIZE - 1))
291 if (offset >= WORDSIZE) {
292 /* The source and destination words don't overlap. */
294 /* To improve branch prediction, one iteration of this
295 * loop is unrolled. Most matches are short and will
296 * fail the first check. But if that check passes, then
297 * it becomes increasing likely that the match is long
298 * and we'll need to continue copying. */
300 copy_word_unaligned(src, dst);
306 copy_word_unaligned(src, dst);
312 } else if (offset == 1) {
314 /* Offset 1 matches are equivalent to run-length
315 * encoding of the previous byte. This case is common
316 * if the data contains many repeated bytes. */
318 machine_word_t v = repeat_byte(*(dst - 1));
320 store_word_unaligned(v, dst);
327 * We don't bother with special cases for other 'offset <
328 * WORDSIZE', which are usually rarer than 'offset == 1'. Extra
329 * checks will just slow things down. Actually, it's possible
330 * to handle all the 'offset < WORDSIZE' cases using the same
331 * code, but it still becomes more complicated doesn't seem any
332 * faster overall; it definitely slows down the more common
333 * 'offset == 1' case.
337 /* Fall back to a bytewise copy. */
339 if (min_length >= 2) {
343 if (min_length >= 3) {
347 if (min_length >= 4) {
356 #endif /* _WIMLIB_DECOMPRESS_COMMON_H */