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
15 #include "wimlib/compiler.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. */
64 if (is->bitsleft >= num_bits)
67 if (unlikely(is->end - is->next < 2))
70 is->bitbuf |= (u32)get_unaligned_u16_le(is->next) << (16 - is->bitsleft);
74 if (unlikely(num_bits == 17 && is->bitsleft == 16)) {
75 if (unlikely(is->end - is->next < 2))
78 is->bitbuf |= (u32)get_unaligned_u16_le(is->next);
89 /* Return the next @num_bits bits from the bitstream, without removing them.
90 * There must be at least @num_bits remaining in the buffer variable, from a
91 * previous call to bitstream_ensure_bits(). */
93 bitstream_peek_bits(const struct input_bitstream *is, const unsigned num_bits)
95 return (is->bitbuf >> 1) >> (sizeof(is->bitbuf) * 8 - num_bits - 1);
98 /* Remove @num_bits from the bitstream. There must be at least @num_bits
99 * remaining in the buffer variable, from a previous call to
100 * bitstream_ensure_bits(). */
102 bitstream_remove_bits(struct input_bitstream *is, unsigned num_bits)
104 is->bitbuf <<= num_bits;
105 is->bitsleft -= num_bits;
108 /* Remove and return @num_bits bits from the bitstream. There must be at least
109 * @num_bits remaining in the buffer variable, from a previous call to
110 * bitstream_ensure_bits(). */
112 bitstream_pop_bits(struct input_bitstream *is, unsigned num_bits)
114 u32 bits = bitstream_peek_bits(is, num_bits);
115 bitstream_remove_bits(is, num_bits);
119 /* Read and return the next @num_bits bits from the bitstream. */
121 bitstream_read_bits(struct input_bitstream *is, unsigned num_bits)
123 bitstream_ensure_bits(is, num_bits);
124 return bitstream_pop_bits(is, num_bits);
127 /* Read and return the next literal byte embedded in the bitstream. */
129 bitstream_read_byte(struct input_bitstream *is)
131 if (unlikely(is->end == is->next))
136 /* Read and return the next 16-bit integer embedded in the bitstream. */
138 bitstream_read_u16(struct input_bitstream *is)
142 if (unlikely(is->end - is->next < 2))
144 v = get_unaligned_u16_le(is->next);
149 /* Read and return the next 32-bit integer embedded in the bitstream. */
151 bitstream_read_u32(struct input_bitstream *is)
155 if (unlikely(is->end - is->next < 4))
157 v = get_unaligned_u32_le(is->next);
162 /* Read into @dst_buffer an array of literal bytes embedded in the bitstream.
163 * Return either a pointer to the byte past the last written, or NULL if the
164 * read overflows the input buffer. */
166 bitstream_read_bytes(struct input_bitstream *is, void *dst_buffer, size_t count)
168 if (unlikely(is->end - is->next < count))
170 memcpy(dst_buffer, is->next, count);
172 return (u8 *)dst_buffer + count;
175 /* Align the input bitstream on a coding-unit boundary. */
177 bitstream_align(struct input_bitstream *is)
183 /* Needed alignment of decode_table parameter to make_huffman_decode_table().
185 * Reason: We may fill the entries with SSE instructions without worrying
186 * about dealing with the unaligned case. */
187 #define DECODE_TABLE_ALIGNMENT 16
189 /* Maximum supported symbol count for make_huffman_decode_table().
191 * Reason: In direct mapping entries, we store the symbol in 11 bits. */
192 #define DECODE_TABLE_MAX_SYMBOLS 2048
194 /* Maximum supported table bits for make_huffman_decode_table().
196 * Reason: In internal binary tree nodes, offsets are encoded in 14 bits.
197 * But the real limit is 13, because we allocate entries past the end of
198 * the direct lookup part of the table for binary tree nodes. (Note: if
199 * needed this limit could be removed by encoding the offsets relative to
200 * &decode_table[1 << table_bits].) */
201 #define DECODE_TABLE_MAX_TABLE_BITS 13
203 /* Maximum supported codeword length for make_huffman_decode_table().
205 * Reason: In direct mapping entries, we encode the codeword length in 5
206 * bits, and the top 2 bits can't both be set because that has special
208 #define DECODE_TABLE_MAX_CODEWORD_LEN 23
210 /* Reads and returns the next Huffman-encoded symbol from a bitstream. If the
211 * input data is exhausted, the Huffman symbol is decoded as if the missing bits
214 * XXX: This is mostly duplicated in lzms_decode_huffman_symbol() in
215 * lzms_decompress.c. */
216 static inline unsigned
217 read_huffsym(struct input_bitstream *istream, const u16 decode_table[],
218 unsigned table_bits, unsigned max_codeword_len)
223 bitstream_ensure_bits(istream, max_codeword_len);
225 /* Index the decode table by the next table_bits bits of the input. */
226 key_bits = bitstream_peek_bits(istream, table_bits);
227 entry = decode_table[key_bits];
228 if (likely(entry < 0xC000)) {
229 /* Fast case: The decode table directly provided the
230 * symbol and codeword length. The low 11 bits are the
231 * symbol, and the high 5 bits are the codeword length. */
232 bitstream_remove_bits(istream, entry >> 11);
233 return entry & 0x7FF;
235 /* Slow case: The codeword for the symbol is longer than
236 * table_bits, so the symbol does not have an entry
237 * directly in the first (1 << table_bits) entries of the
238 * decode table. Traverse the appropriate binary tree
239 * bit-by-bit to decode the symbol. */
240 bitstream_remove_bits(istream, table_bits);
242 key_bits = (entry & 0x3FFF) + bitstream_pop_bits(istream, 1);
243 } while ((entry = decode_table[key_bits]) >= 0xC000);
249 make_huffman_decode_table(u16 decode_table[], unsigned num_syms,
250 unsigned num_bits, const u8 lens[],
251 unsigned max_codeword_len);
254 copy_word_unaligned(const void *src, void *dst)
256 store_word_unaligned(load_word_unaligned(src), dst);
259 static inline machine_word_t
264 BUILD_BUG_ON(WORDSIZE != 4 && WORDSIZE != 8);
269 v |= v << ((WORDSIZE == 8) ? 32 : 0);
274 * Copy an LZ77 match at (dst - offset) to dst.
276 * The length and offset must be already validated --- that is, (dst - offset)
277 * can't underrun the output buffer, and (dst + length) can't overrun the output
278 * buffer. Also, the length cannot be 0.
280 * @winend points to the byte past the end of the output buffer.
281 * This function won't write any data beyond this position.
284 lz_copy(u8 *dst, u32 length, u32 offset, const u8 *winend, u32 min_length)
286 const u8 *src = dst - offset;
287 const u8 * const end = dst + length;
290 * Try to copy one machine word at a time. On i386 and x86_64 this is
291 * faster than copying one byte at a time, unless the data is
292 * near-random and all the matches have very short lengths. Note that
293 * since this requires unaligned memory accesses, it won't necessarily
294 * be faster on every architecture.
296 * Also note that we might copy more than the length of the match. For
297 * example, if a word is 8 bytes and the match is of length 5, then
298 * we'll simply copy 8 bytes. This is okay as long as we don't write
299 * beyond the end of the output buffer, hence the check for (winend -
300 * end >= WORDSIZE - 1).
302 if (UNALIGNED_ACCESS_IS_VERY_FAST &&
303 likely(winend - end >= WORDSIZE - 1))
306 if (offset >= WORDSIZE) {
307 /* The source and destination words don't overlap. */
309 /* To improve branch prediction, one iteration of this
310 * loop is unrolled. Most matches are short and will
311 * fail the first check. But if that check passes, then
312 * it becomes increasing likely that the match is long
313 * and we'll need to continue copying. */
315 copy_word_unaligned(src, dst);
321 copy_word_unaligned(src, dst);
327 } else if (offset == 1) {
329 /* Offset 1 matches are equivalent to run-length
330 * encoding of the previous byte. This case is common
331 * if the data contains many repeated bytes. */
333 machine_word_t v = repeat_byte(*(dst - 1));
335 store_word_unaligned(v, dst);
342 * We don't bother with special cases for other 'offset <
343 * WORDSIZE', which are usually rarer than 'offset == 1'. Extra
344 * checks will just slow things down. Actually, it's possible
345 * to handle all the 'offset < WORDSIZE' cases using the same
346 * code, but it still becomes more complicated doesn't seem any
347 * faster overall; it definitely slows down the more common
348 * 'offset == 1' case.
352 /* Fall back to a bytewise copy. */
354 if (min_length >= 2) {
358 if (min_length >= 3) {
362 if (min_length >= 4) {
371 #endif /* _WIMLIB_DECOMPRESS_COMMON_H */