4 * Functions used for decompression.
8 * Copyright (C) 2012, 2013 Eric Biggers
10 * This file is part of wimlib, a library for working with WIM files.
12 * wimlib is free software; you can redistribute it and/or modify it under the
13 * terms of the GNU General Public License as published by the Free
14 * Software Foundation; either version 3 of the License, or (at your option)
17 * wimlib is distributed in the hope that it will be useful, but WITHOUT ANY
18 * WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR
19 * A PARTICULAR PURPOSE. See the GNU General Public License for more
22 * You should have received a copy of the GNU General Public License
23 * along with wimlib; if not, see http://www.gnu.org/licenses/.
26 #include "decompress.h"
30 * make_huffman_decode_table: - Builds a fast huffman decoding table from an
31 * array that gives the length of the codeword for each symbol in the alphabet.
32 * Originally based on code written by David Tritscher (taken the original LZX
33 * decompression code); also heavily modified to add some optimizations used in
34 * the zlib code, as well as more comments.
36 * @decode_table: The array in which to create the fast huffman decoding
37 * table. It must have a length of at least
38 * (2**table_bits) + 2 * num_syms to guarantee
39 * that there is enough space.
41 * @num_syms: Number of symbols in the alphabet, including symbols
42 * that do not appear in this particular input chunk.
44 * @table_bits: Any symbols with a code length of table_bits or less can
45 * be decoded in one lookup of the table. 2**table_bits
46 * must be greater than or equal to @num_syms if there are
47 * any Huffman codes longer than @table_bits.
49 * @lens: An array of length @num_syms, indexable by symbol, that
50 * gives the length of the Huffman codeword for that
51 * symbol. Because the Huffman tree is in canonical form,
52 * it can be reconstructed by only knowing the length of
53 * the codeword for each symbol. It is assumed, but not
54 * checked, that every length is less than
57 * @max_codeword_len: The longest codeword length allowed in the compression
60 * Returns 0 on success; returns -1 if the length values do not correspond to a
63 * The format of the Huffamn decoding table is as follows. The first (1 <<
64 * table_bits) entries of the table are indexed by chunks of the input of size
65 * @table_bits. If the next Huffman codeword in the input happens to have a
66 * length of exactly @table_bits, the symbol is simply read directly from the
67 * decoding table. Alternatively, if the next Huffman codeword has length _less
68 * than_ @table_bits, the symbol is also read directly from the decode table;
69 * this is possible because every entry in the table that is indexed by an
70 * integer that has the shorter codeword as a binary prefix is filled in with
71 * the appropriate symbol. If a codeword has length n <= table_bits, it will
72 * have 2**(table_bits - n) possible suffixes, and thus that many entries in the
75 * It's a bit more complicated if the next Huffman codeword has length of more
76 * than @table_bits. The table entry indexed by the first @table_bits of that
77 * codeword cannot give the appropriate symbol directly, because that entry is
78 * guaranteed to be referenced by the Huffman codewords of multiple symbols.
79 * And while the LZX compression format does not allow codes longer than 16
80 * bits, a table of size (2 ** 16) = 65536 entries would be too slow to create.
82 * There are several different ways to make it possible to look up the symbols
83 * for codewords longer than @table_bits. One way is to make the entries for
84 * the prefixes of length @table_bits of those entries be pointers to additional
85 * decoding tables that are indexed by some number of additional bits of the
86 * codeword. The technique used here is a bit simpler, however: just store the
87 * needed subtrees of the Huffman tree in the decoding table after the lookup
88 * entries, beginning at index (2**table_bits). Real pointers are replaced by
89 * indices into the decoding table, and symbol entries are distinguished from
90 * pointers by the fact that values less than @num_syms must be symbol values.
93 make_huffman_decode_table(u16 decode_table[], unsigned num_syms,
94 unsigned table_bits, const u8 lens[],
95 unsigned max_codeword_len)
97 unsigned len_counts[max_codeword_len + 1];
98 u16 sorted_syms[num_syms];
99 unsigned offsets[max_codeword_len + 1];
100 const unsigned table_num_entries = 1 << table_bits;
102 /* accumulate lengths for codes */
103 for (unsigned i = 0; i <= max_codeword_len; i++)
106 for (unsigned sym = 0; sym < num_syms; sym++) {
107 wimlib_assert2(lens[sym] <= max_codeword_len);
108 len_counts[lens[sym]]++;
111 /* check for an over-subscribed or incomplete set of lengths */
113 for (unsigned len = 1; len <= max_codeword_len; len++) {
115 left -= len_counts[len];
116 if (left < 0) { /* over-subscribed */
117 ERROR("Invalid Huffman code (over-subscribed)");
121 if (left != 0) /* incomplete set */{
122 if (left == 1 << max_codeword_len) {
123 /* Empty code--- okay in XPRESS and LZX */
124 memset(decode_table, 0,
125 table_num_entries * sizeof(decode_table[0]));
128 ERROR("Invalid Huffman code (incomplete set)");
133 /* Generate offsets into symbol table for each length for sorting */
135 for (unsigned len = 1; len < max_codeword_len; len++)
136 offsets[len + 1] = offsets[len] + len_counts[len];
138 /* Sort symbols primarily by length and secondarily by symbol order.
139 * This is basically a count-sort over the codeword lengths.
140 * In the process, calculate the number of symbols that have nonzero
141 * length and are therefore used in the symbol stream. */
142 unsigned num_used_syms = 0;
143 for (unsigned sym = 0; sym < num_syms; sym++) {
144 if (lens[sym] != 0) {
145 sorted_syms[offsets[lens[sym]]++] = sym;
150 /* Fill entries for codewords short enough for a direct mapping. We can
151 * take advantage of the ordering of the codewords, since the Huffman
152 * code is canonical. It must be the case that all the codewords of
153 * some length L numerically precede all the codewords of length L + 1.
154 * Furthermore, if we have 2 symbols A and B with the same codeword
155 * length but symbol A is sorted before symbol B, then then we know that
156 * the codeword for A numerically precedes the codeword for B. */
157 unsigned decode_table_pos = 0;
160 wimlib_assert2(num_used_syms != 0);
162 unsigned sym = sorted_syms[i];
163 unsigned codeword_len = lens[sym];
164 if (codeword_len > table_bits)
167 unsigned num_entries = 1 << (table_bits - codeword_len);
168 const unsigned entries_per_long = sizeof(unsigned long) /
169 sizeof(decode_table[0]);
170 if (num_entries >= entries_per_long) {
171 /* Fill in the Huffman decode table entries one unsigned
172 * long at a time. On 32-bit systems this is 2 entries
173 * per store, while on 64-bit systems this is 4 entries
175 wimlib_assert2(decode_table_pos % entries_per_long == 0);
176 BUILD_BUG_ON(sizeof(unsigned long) != 4 &&
177 sizeof(unsigned long) != 8);
179 unsigned long *p = (unsigned long *)&decode_table[decode_table_pos];
180 unsigned n = num_entries / entries_per_long;
181 unsigned long v = sym;
182 if (sizeof(unsigned long) >= 4)
184 if (sizeof(unsigned long) >= 8) {
185 /* This may produce a compiler warning if an
186 * unsigned long is 32 bits, but this won't be
187 * executed unless an unsigned long is at least
195 decode_table_pos += num_entries;
197 /* Fill in the Huffman decode table entries one 16-bit
198 * integer at a time. */
200 decode_table[decode_table_pos++] = sym;
201 } while (--num_entries);
203 wimlib_assert2(decode_table_pos <= table_num_entries);
204 if (++i == num_used_syms) {
205 wimlib_assert2(decode_table_pos == table_num_entries);
206 /* No codewords were longer than @table_bits, so the
207 * table is now entirely filled with the codewords. */
212 wimlib_assert2(i < num_used_syms);
213 wimlib_assert2(decode_table_pos < table_num_entries);
215 /* Fill in the remaining entries, which correspond to codes longer than
218 * First, zero out the rest of the entries. This is necessary so that
219 * the entries appear as "unallocated" in the next part. */
221 unsigned j = decode_table_pos;
224 } while (++j != table_num_entries);
227 /* Assert that 2**table_bits is at least num_syms. If this wasn't the
228 * case, we wouldn't be able to distinguish pointer entries from symbol
230 wimlib_assert2(table_num_entries >= num_syms);
232 /* The current Huffman codeword */
233 unsigned cur_codeword = decode_table_pos;
235 /* The tree nodes are allocated starting at decode_table[1 <<
236 * table_bits]. Remember that the full size of the table, including the
237 * extra space for the tree nodes, is actually 2**table_bits + 2 *
238 * num_syms slots, while table_num_entries is only 2**table_Bits. */
239 unsigned next_free_tree_slot = table_num_entries;
241 /* Go through every codeword of length greater than @table_bits,
242 * primarily in order of codeword length and secondarily in order of
244 unsigned prev_codeword_len = table_bits;
246 unsigned sym = sorted_syms[i];
247 unsigned codeword_len = lens[sym];
248 unsigned extra_bits = codeword_len - table_bits;
250 cur_codeword <<= (codeword_len - prev_codeword_len);
251 prev_codeword_len = codeword_len;
253 /* index of the current node; find it from the prefix of the
254 * current Huffman codeword. */
255 unsigned node_idx = cur_codeword >> extra_bits;
256 wimlib_assert2(node_idx < table_num_entries);
258 /* Go through each bit of the current Huffman codeword beyond
259 * the prefix of length @table_bits and walk the tree,
260 * allocating any slots that have not yet been allocated. */
263 /* If the current tree node points to nowhere
264 * but we need to follow it, allocate a new node
265 * for it to point to. */
266 if (decode_table[node_idx] == 0) {
267 decode_table[node_idx] = next_free_tree_slot;
268 decode_table[next_free_tree_slot++] = 0;
269 decode_table[next_free_tree_slot++] = 0;
270 wimlib_assert2(next_free_tree_slot <=
271 table_num_entries + 2 * num_syms);
274 /* Set node_idx to left child */
275 node_idx = decode_table[node_idx];
277 /* Is the next bit 0 or 1? If 0, go left (already done).
278 * If 1, go right by incrementing node_idx. */
280 node_idx += (cur_codeword >> extra_bits) & 1;
281 } while (extra_bits != 0);
283 /* node_idx is now the index of the leaf entry into which the
284 * actual symbol will go. */
285 decode_table[node_idx] = sym;
287 /* cur_codeword is always incremented because this is
288 * how canonical Huffman codes are generated (add 1 for
289 * each code, then left shift whenever the code length
292 } while (++i != num_used_syms);
296 /* Reads a Huffman-encoded symbol from the bistream when the number of remaining
297 * bits is less than the maximum codeword length. */
299 read_huffsym_near_end_of_input(struct input_bitstream *istream,
300 const u16 decode_table[],
306 unsigned bitsleft = istream->bitsleft;
311 if (table_bits > bitsleft) {
314 key_bits = bitstream_peek_bits(istream, key_size) <<
315 (table_bits - key_size);
317 key_size = table_bits;
318 bitsleft -= table_bits;
319 key_bits = bitstream_peek_bits(istream, table_bits);
322 sym = decode_table[key_bits];
323 if (sym >= num_syms) {
324 bitstream_remove_bits(istream, key_size);
327 ERROR("Input stream exhausted");
330 key_bits = sym + bitstream_peek_bits(istream, 1);
331 bitstream_remove_bits(istream, 1);
333 } while ((sym = decode_table[key_bits]) >= num_syms);
335 bitstream_remove_bits(istream, lens[sym]);