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"
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. */
62 wimlib_assert2(num_bits <= 17);
64 if (is->bitsleft >= num_bits)
67 if (unlikely(is->end - is->next < 2))
70 is->bitbuf |= (u32)le16_to_cpu(*(const le16 *)is->next)
71 << (16 - is->bitsleft);
75 if (unlikely(num_bits == 17 && is->bitsleft == 16)) {
76 if (unlikely(is->end - is->next < 2))
79 is->bitbuf |= (u32)le16_to_cpu(*(const le16 *)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 < 1))
139 /* Read and return the next 32-bit integer embedded in the bitstream. */
141 bitstream_read_u32(struct input_bitstream *is)
145 if (unlikely(is->end - is->next < 4))
147 v = le32_to_cpu(*(const le32 *)is->next);
152 /* Read an array of literal bytes embedded in the bitstream. Return a pointer
153 * to the resulting array, or NULL if the read overflows the input buffer. */
154 static inline const u8 *
155 bitstream_read_bytes(struct input_bitstream *is, size_t count)
159 if (unlikely(is->end - is->next < count))
166 /* Align the input bitstream on a coding-unit boundary. */
168 bitstream_align(struct input_bitstream *is)
174 /* Needed alignment of decode_table parameter to make_huffman_decode_table().
176 * Reason: We may fill the entries with SSE instructions without worrying
177 * about dealing with the unaligned case. */
178 #define DECODE_TABLE_ALIGNMENT 16
180 /* Maximum supported symbol count for make_huffman_decode_table().
182 * Reason: In direct mapping entries, we store the symbol in 11 bits. */
183 #define DECODE_TABLE_MAX_SYMBOLS 2048
185 /* Maximum supported table bits for make_huffman_decode_table().
187 * Reason: In internal binary tree nodes, offsets are encoded in 14 bits.
188 * But the real limit is 13, because we allocate entries past the end of
189 * the direct lookup part of the table for binary tree nodes. (Note: if
190 * needed this limit could be removed by encoding the offsets relative to
191 * &decode_table[1 << table_bits].) */
192 #define DECODE_TABLE_MAX_TABLE_BITS 13
194 /* Maximum supported codeword length for make_huffman_decode_table().
196 * Reason: In direct mapping entries, we encode the codeword length in 5
197 * bits, and the top 2 bits can't both be set because that has special
199 #define DECODE_TABLE_MAX_CODEWORD_LEN 23
201 /* Reads and returns the next Huffman-encoded symbol from a bitstream. If the
202 * input data is exhausted, the Huffman symbol is decoded as if the missing bits
205 * XXX: This is mostly duplicated in lzms_huffman_decode_symbol() in
206 * lzms-decompress.c. */
208 read_huffsym(struct input_bitstream *istream, const u16 decode_table[],
209 unsigned table_bits, unsigned max_codeword_len)
214 bitstream_ensure_bits(istream, max_codeword_len);
216 /* Index the decode table by the next table_bits bits of the input. */
217 key_bits = bitstream_peek_bits(istream, table_bits);
218 entry = decode_table[key_bits];
219 if (likely(entry < 0xC000)) {
220 /* Fast case: The decode table directly provided the
221 * symbol and codeword length. The low 11 bits are the
222 * symbol, and the high 5 bits are the codeword length. */
223 bitstream_remove_bits(istream, entry >> 11);
224 return entry & 0x7FF;
226 /* Slow case: The codeword for the symbol is longer than
227 * table_bits, so the symbol does not have an entry
228 * directly in the first (1 << table_bits) entries of the
229 * decode table. Traverse the appropriate binary tree
230 * bit-by-bit to decode the symbol. */
231 bitstream_remove_bits(istream, table_bits);
233 key_bits = (entry & 0x3FFF) + bitstream_pop_bits(istream, 1);
234 } while ((entry = decode_table[key_bits]) >= 0xC000);
240 make_huffman_decode_table(u16 decode_table[], unsigned num_syms,
241 unsigned num_bits, const u8 lens[],
242 unsigned max_codeword_len);
246 * Copy a LZ77 match at (dst - offset) to dst.
248 * The length and offset must be already validated --- that is, (dst - offset)
249 * can't underrun the output buffer, and (dst + length) can't overrun the output
250 * buffer. Also, the length cannot be 0.
252 * @winend points to the byte past the end of the output buffer.
253 * This function won't write any data beyond this position.
256 lz_copy(u8 *dst, u32 length, u32 offset, const u8 *winend)
258 const u8 *src = dst - offset;
259 #if defined(__x86_64__) || defined(__i386__)
260 /* Copy one 'unsigned long' at a time. On i386 and x86_64 this is
261 * faster than copying one byte at a time, unless the data is
262 * near-random and all the matches have very short lengths. Note that
263 * since this requires unaligned memory accesses, it won't necessarily
264 * be faster on every architecture.
266 * Also note that we might copy more than the length of the match. For
267 * example, if an 'unsigned long' is 8 bytes and the match is of length
268 * 5, then we'll simply copy 8 bytes. This is okay as long as we don't
269 * write beyond the end of the output buffer, hence the check for
270 * (winend - (dst + length) >= sizeof(unsigned long) - 1). */
271 if (offset >= sizeof(unsigned long) &&
272 winend - (dst + length) >= sizeof(unsigned long) - 1)
274 /* Access memory through a packed struct. This tricks the
275 * compiler into allowing unaligned memory accesses. */
276 struct ulong_wrapper {
280 const u8 * const end = dst + length;
283 v = ((struct ulong_wrapper *)src)->v;
284 ((struct ulong_wrapper *)dst)->v = v;
285 dst += sizeof(unsigned long);
286 src += sizeof(unsigned long);
290 v = ((struct ulong_wrapper *)src)->v;
291 ((struct ulong_wrapper *)dst)->v = v;
292 dst += sizeof(unsigned long);
293 src += sizeof(unsigned long);
305 #endif /* _WIMLIB_DECOMPRESS_COMMON_H */