4 * Functions too long to declare as inline in decomp.h.
6 * Copyright (C) 2012 Eric Biggers
8 * wimlib - Library for working with WIM files
10 * This library is free software; you can redistribute it and/or modify it under
11 * the terms of the GNU Lesser General Public License as published by the Free
12 * Software Foundation; either version 2.1 of the License, or (at your option) any
15 * This library is distributed in the hope that it will be useful, but WITHOUT ANY
16 * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A
17 * PARTICULAR PURPOSE. See the GNU Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public License along
20 * with this library; if not, write to the Free Software Foundation, Inc., 59
21 * Temple Place, Suite 330, Boston, MA 02111-1307 USA
27 /* Reads @n bytes from the bitstream @stream into the location pointed to by @dest.
28 * The bitstream must be 16-bit aligned. */
29 int bitstream_read_bytes(struct input_bitstream *stream, size_t n, void *dest)
31 /* Precondition: The bitstream is 16-byte aligned. */
32 wimlib_assert(stream->bitsleft % 16 == 0);
36 /* Get the bytes currently in the buffer variable. */
37 while (stream->bitsleft != 0) {
40 *p++ = bitstream_peek_bits(stream, 8);
41 bitstream_remove_bits(stream, 8);
44 /* Get the rest directly from the pointer to the data. Of course, it's
45 * necessary to check there are really n bytes available. */
46 if (n > stream->data_bytes_left) {
47 ERROR("Unexpected end of input when "
48 "reading %zu bytes from bitstream "
49 "(only have %u bytes left)\n", n,
50 stream->data_bytes_left);
53 memcpy(p, stream->data, n);
55 stream->data_bytes_left -= n;
57 /* It's possible to copy an odd number of bytes and leave the stream in
58 * an inconsistent state. Fix it by reading the next byte, if it is
60 if ((n & 1) && stream->data_bytes_left != 0) {
62 stream->data_bytes_left--;
63 stream->bitbuf |= (input_bitbuf_t)(*stream->data) <<
64 (sizeof(input_bitbuf_t) * 8 - 8);
70 /* Aligns the bitstream on a 16-bit boundary.
72 * Note: M$'s idea of "alignment" means that for some reason, a 16-bit word
73 * should be skipped over if the buffer happens to already be aligned on such a
74 * boundary. This only applies for realigning the stream after the blocktype
75 * and length fields of an uncompressed block, however; it does not apply when
76 * realigning the stream after the end of the uncompressed block.
78 int align_input_bitstream(struct input_bitstream *stream,
79 bool skip_word_if_aligned)
82 if (stream->bitsleft % 16 != 0) {
83 bitstream_remove_bits(stream, stream->bitsleft % 16);
84 } else if (skip_word_if_aligned) {
85 if (stream->bitsleft == 0) {
86 ret = bitstream_ensure_bits(stream, 16);
88 ERROR("Unexpected end of input when "
89 "aligning bitstream!\n");
93 bitstream_remove_bits(stream, 16);
99 * Builds a fast huffman decoding table from a canonical huffman code lengths
100 * table. Based on code written by David Tritscher.
102 * @decode_table: The array in which to create the fast huffman decoding
103 * table. It must have a length of at least
104 * (2**num_bits) + 2 * num_syms to guarantee
105 * that there is enough space.
107 * @num_syms: Total number of symbols in the Huffman tree.
109 * @num_bits: Any symbols with a code length of num_bits or less can be
110 * decoded in one lookup of the table. 2**num_bits
111 * must be greater than or equal to @num_syms if there are
112 * any Huffman codes longer than @num_bits.
114 * @lens: An array of length @num_syms, indexable by symbol, that
115 * gives the length of that symbol. Because the Huffman
116 * tree is in canonical form, it can be reconstructed by
117 * only knowing the length of the code for each symbol.
119 * @make_codeword_len: An integer that gives the longest possible codeword
122 * Returns 0 on success; returns 1 if the length values do not correspond to a
123 * valid Huffman tree, or if there are codes of length greater than @num_bits
124 * but 2**num_bits < num_syms.
126 * What exactly is the format of the fast Huffman decoding table? The first
127 * (1 << num_bits) entries of the table are indexed by chunks of the input of
128 * size @num_bits. If the next Huffman code in the input happens to have a
129 * length of exactly @num_bits, the symbol is simply read directly from the
130 * decoding table. Alternatively, if the next Huffman code has length _less
131 * than_ @num_bits, the symbol is also read directly from the decode table; this
132 * is possible because every entry in the table that is indexed by an integer
133 * that has the shorter code as a binary prefix is filled in with the
134 * appropriate symbol. If a code has length n <= num_bits, it will have
135 * 2**(num_bits - n) possible suffixes, and thus that many entries in the
138 * It's a bit more complicated if the next Huffman code has length of more than
139 * @num_bits. The table entry indexed by the first @num_bits of that code
140 * cannot give the appropriate symbol directly, because that entry is guaranteed
141 * to be referenced by the Huffman codes for multiple symbols. And while the
142 * LZX compression format does not allow codes longer than 16 bits, a table of
143 * size (2 ** 16) = 65536 entries would be too slow to create.
145 * There are several different ways to make it possible to look up the symbols
146 * for codes longer than @num_bits. A common way is to make the entries for the
147 * prefixes of length @num_bits of those entries be pointers to additional
148 * decoding tables that are indexed by some number of additional bits of the
149 * code symbol. The technique used here is a bit simpler, however. We just
150 * store the needed subtrees of the Huffman tree in the decoding table after the
151 * lookup entries, beginning at index (2**num_bits). Real pointers are
152 * replaced by indices into the decoding table, and we distinguish symbol
153 * entries from pointers by the fact that values less than @num_syms must be
156 int make_huffman_decode_table(u16 decode_table[], uint num_syms,
157 uint num_bits, const u8 lens[],
160 /* Number of entries in the decode table. */
161 u32 table_num_entries = 1 << num_bits;
163 /* Current position in the decode table. */
164 u32 decode_table_pos = 0;
166 /* Fill entries for codes short enough for a direct mapping. Here we
167 * are taking advantage of the ordering of the codes, since they are for
168 * a canonical Huffman tree. It must be the case that all the codes of
169 * some length @code_length, zero-extended or one-extended, numerically
170 * precede all the codes of length @code_length + 1. Furthermore, if we
171 * have 2 symbols A and B, such that A is listed before B in the lens
172 * array, and both symbols have the same code length, then we know that
173 * the code for A numerically precedes the code for B.
175 for (uint code_len = 1; code_len <= num_bits; code_len++) {
177 /* Number of entries that a code of length @code_length would
179 u32 code_num_entries = 1 << (num_bits - code_len);
182 /* For each symbol of length @code_len, fill in its entries in
183 * the decode table. */
184 for (uint sym = 0; sym < num_syms; sym++) {
186 if (lens[sym] != code_len)
190 /* Check for table overrun. This can only happen if the
191 * given lengths do not correspond to a valid Huffman
193 if (decode_table_pos >= table_num_entries) {
194 ERROR("Huffman decoding table overrun: "
195 "pos = %u, num_entries = %u\n",
201 /* Fill all possible lookups of this symbol with
202 * the symbol itself. */
203 for (uint i = 0; i < code_num_entries; i++)
204 decode_table[decode_table_pos + i] = sym;
206 /* Increment the position in the decode table by
207 * the number of entries that were just filled
209 decode_table_pos += code_num_entries;
213 /* If all entries of the decode table have been filled in, there are no
214 * codes longer than num_bits, so we are done filling in the decode
216 if (decode_table_pos == table_num_entries)
219 /* Otherwise, fill in the remaining entries, which correspond to codes longer
223 /* First, zero out the rest of the entries; this is necessary so
224 * that the entries appear as "unallocated" in the next part. */
225 for (uint i = decode_table_pos; i < table_num_entries; i++)
228 /* Assert that 2**num_bits is at least num_syms. If this wasn't the
229 * case, we wouldn't be able to distinguish pointer entries from symbol
231 wimlib_assert((1 << num_bits) >= num_syms);
234 /* The current Huffman code. */
235 uint current_code = decode_table_pos;
237 /* The tree nodes are allocated starting at
238 * decode_table[table_num_entries]. Remember that the full size of the
239 * table, including the extra space for the tree nodes, is actually
240 * 2**num_bits + 2 * num_syms slots, while table_num_entries is only
242 uint next_free_tree_slot = table_num_entries;
244 /* Go through every codeword of length greater than @num_bits. Note:
245 * the LZX format guarantees that the codeword length can be at most 16
247 for (uint code_len = num_bits + 1; code_len <= max_code_len;
251 for (uint sym = 0; sym < num_syms; sym++) {
252 if (lens[sym] != code_len)
256 /* i is the index of the current node; find it from the
257 * prefix of the current Huffman code. */
258 uint i = current_code >> (code_len - num_bits);
260 if (i >= (1 << num_bits)) {
261 ERROR("Invalid canonical Huffman code!\n");
265 /* Go through each bit of the current Huffman code
266 * beyond the prefix of length num_bits and walk the
267 * tree, "allocating" slots that have not yet been
269 for (int bit_num = num_bits + 1; bit_num <= code_len; bit_num++) {
271 /* If the current tree node points to nowhere
272 * but we need to follow it, allocate a new node
273 * for it to point to. */
274 if (decode_table[i] == 0) {
275 decode_table[i] = next_free_tree_slot;
276 decode_table[next_free_tree_slot++] = 0;
277 decode_table[next_free_tree_slot++] = 0;
282 /* Is the next bit 0 or 1? If 0, go left;
283 * otherwise, go right (by incrementing i by 1) */
284 int bit_pos = code_len - bit_num;
286 int bit = (current_code & (1 << bit_pos)) >>
291 /* i is now the index of the leaf entry into which the
292 * actual symbol will go. */
293 decode_table[i] = sym;
295 /* Increment decode_table_pos only if the prefix of the
296 * Huffman code changes. */
297 if (current_code >> (code_len - num_bits) !=
298 (current_code + 1) >> (code_len - num_bits))
301 /* current_code is always incremented because this is
302 * how canonical Huffman codes are generated (add 1 for
303 * each code, then left shift whenever the code length
310 /* If the lengths really represented a valid Huffman tree, all
311 * @table_num_entries in the table will have been filled. However, it
312 * is also possible that the tree is completely empty (as noted
313 * earlier) with all 0 lengths, and this is expected to succeed. */
315 if (decode_table_pos != table_num_entries) {
317 for (uint i = 0; i < num_syms; i++) {
319 ERROR("Lengths do not form a valid "
320 "canonical Huffman tree "
321 "(only filled %u of %u decode "
322 "table slots)!\n", decode_table_pos,
331 /* Reads a Huffman-encoded symbol when it is known there are less than
332 * MAX_CODE_LEN bits remaining in the bitstream. */
333 static int read_huffsym_near_end_of_input(struct input_bitstream *istream,
334 const u16 decode_table[],
340 uint bitsleft = istream->bitsleft;
345 if (table_bits > bitsleft) {
348 key_bits = bitstream_peek_bits(istream, key_size) <<
349 (table_bits - key_size);
351 key_size = table_bits;
352 bitsleft -= table_bits;
353 key_bits = bitstream_peek_bits(istream, table_bits);
356 sym = decode_table[key_bits];
357 if (sym >= num_syms) {
358 bitstream_remove_bits(istream, key_size);
361 ERROR("Input stream exhausted!\n");
364 key_bits = sym + bitstream_peek_bits(istream, 1);
365 bitstream_remove_bits(istream, 1);
367 } while ((sym = decode_table[key_bits]) >= num_syms);
369 bitstream_remove_bits(istream, lens[sym]);
376 * Reads a Huffman-encoded symbol from a bitstream.
378 * This function may be called hundreds of millions of times when extracting a
379 * large WIM file. I'm not sure it could be made much faster that it is,
380 * especially since there isn't enough time to make a big table that allows
381 * decoding multiple symbols per lookup. But if extracting files to a hard
382 * disk, the IO will be the bottleneck anyway.
384 * @buf: The input buffer from which the symbol will be read.
385 * @decode_table: The fast Huffman decoding table for the Huffman tree.
386 * @lengths: The table that gives the length of the code for each
388 * @num_symbols: The number of symbols in the Huffman code.
389 * @table_bits: Huffman codes this length or less can be looked up
390 * directory in the decode_table, as the
391 * decode_table contains 2**table_bits entries.
393 int read_huffsym(struct input_bitstream *stream,
394 const u16 decode_table[],
396 unsigned num_symbols,
399 unsigned max_codeword_len)
401 /* In the most common case, there are at least max_codeword_len bits
402 * remaining in the stream. */
403 if (bitstream_ensure_bits(stream, max_codeword_len) == 0) {
405 /* Use the next table_bits of the input as an index into the
407 u16 key_bits = bitstream_peek_bits(stream, table_bits);
409 u16 sym = decode_table[key_bits];
411 /* If the entry in the decode table is not a valid symbol, it is
412 * the offset of the root of its Huffman subtree. */
413 if (sym >= num_symbols) {
414 bitstream_remove_bits(stream, table_bits);
416 key_bits = sym + bitstream_peek_bits(stream, 1);
417 bitstream_remove_bits(stream, 1);
419 wimlib_assert(key_bits < num_symbols * 2 +
421 } while ((sym = decode_table[key_bits]) >= num_symbols);
423 wimlib_assert(lengths[sym] <= table_bits);
424 bitstream_remove_bits(stream, lengths[sym]);
429 /* Otherwise, we must be careful to use only the bits that are
430 * actually remaining. Don't inline this part since it is very
432 return read_huffsym_near_end_of_input(stream, decode_table, lengths,
433 num_symbols, table_bits, n);