4 * Functions used for decompression.
8 * Copyright (C) 2012 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 * Builds a fast huffman decoding table from an array that gives the length of
31 * the codeword for each symbol in the alphabet. Originally based on code
32 * written by David Tritscher (taken the original LZX decompression code); also
33 * heavily modified to add some optimizations used in the zlib code, as well as
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: Total number of symbols in the Huffman tree.
43 * @table_bits: Any symbols with a code length of table_bits or less can
44 * be decoded in one lookup of the table. 2**table_bits
45 * must be greater than or equal to @num_syms if there are
46 * any Huffman codes longer than @table_bits.
48 * @lens: An array of length @num_syms, indexable by symbol, that
49 * gives the length of the Huffman codeward for that
50 * symbol. Because the Huffman tree is in canonical form,
51 * it can be reconstructed by only knowing the length of
52 * the codeword for each symbol. It is assumed, but not
53 * checked, that every length is less than
56 * @max_codeword_len: The longest codeword length allowed in the compression
59 * Returns 0 on success; returns -1 if the length values do not correspond to a
62 * The format of the Huffamn decoding table is as follows. The first (1 <<
63 * table_bits) entries of the table are indexed by chunks of the input of size
64 * @table_bits. If the next Huffman codeword in the input happens to have a
65 * length of exactly @table_bits, the symbol is simply read directly from the
66 * decoding table. Alternatively, if the next Huffman codeword has length _less
67 * than_ @table_bits, the symbol is also read directly from the decode table;
68 * this is possible because every entry in the table that is indexed by an
69 * integer that has the shorter codeword as a binary prefix is filled in with
70 * the appropriate symbol. If a codeword has length n <= table_bits, it will
71 * have 2**(table_bits - n) possible suffixes, and thus that many entries in the
74 * It's a bit more complicated if the next Huffman codeword has length of more
75 * than @table_bits. The table entry indexed by the first @table_bits of that
76 * codeword cannot give the appropriate symbol directly, because that entry is
77 * guaranteed to be referenced by the Huffman codewords of multiple symbols.
78 * And while the LZX compression format does not allow codes longer than 16
79 * bits, a table of size (2 ** 16) = 65536 entries would be too slow to create.
81 * There are several different ways to make it possible to look up the symbols
82 * for codewords longer than @table_bits. One way is to make the entries for
83 * the prefixes of length @table_bits of those entries be pointers to additional
84 * decoding tables that are indexed by some number of additional bits of the
85 * codeword. The technique used here is a bit simpler, however: just store the
86 * needed subtrees of the Huffman tree in the decoding table after the lookup
87 * entries, beginning at index (2**table_bits). Real pointers are replaced by
88 * indices into the decoding table, and symbol entries are distinguished from
89 * pointers by the fact that values less than @num_syms must be symbol values.
91 int make_huffman_decode_table(u16 decode_table[], unsigned num_syms,
92 unsigned table_bits, const u8 lens[],
93 unsigned max_codeword_len)
95 unsigned len_counts[max_codeword_len + 1];
96 u16 sorted_syms[num_syms];
97 unsigned offsets[max_codeword_len + 1];
98 const unsigned table_num_entries = 1 << table_bits;
100 /* accumulate lengths for codes */
101 for (unsigned i = 0; i <= max_codeword_len; i++)
104 for (unsigned sym = 0; sym < num_syms; sym++) {
105 wimlib_assert2(lens[sym] <= max_codeword_len);
106 len_counts[lens[sym]]++;
109 /* check for an over-subscribed or incomplete set of lengths */
111 for (unsigned len = 1; len <= max_codeword_len; len++) {
113 left -= len_counts[len];
114 if (left < 0) { /* over-subscribed */
115 ERROR("Invalid Huffman code (over-subscribed)");
119 if (left != 0) /* incomplete set */{
120 if (left == 1 << max_codeword_len) {
121 /* Empty code--- okay in XPRESS and LZX */
122 memset(decode_table, 0,
123 table_num_entries * sizeof(decode_table[0]));
126 ERROR("Invalid Huffman code (incomplete set)");
131 /* Generate offsets into symbol table for each length for sorting */
133 for (unsigned len = 1; len < max_codeword_len; len++)
134 offsets[len + 1] = offsets[len] + len_counts[len];
136 /* Sort symbols primarily by length and secondarily by symbol order.
137 * This is basically a count-sort over the codeword lengths.
138 * In the process, calculate the number of symbols that have nonzero
139 * length and are therefore used in the symbol stream. */
140 unsigned num_used_syms = 0;
141 for (unsigned sym = 0; sym < num_syms; sym++) {
142 if (lens[sym] != 0) {
143 sorted_syms[offsets[lens[sym]]++] = sym;
148 /* Fill entries for codewords short enough for a direct mapping. We can
149 * take advantage of the ordering of the codewords, since the Huffman
150 * code is canonical. It must be the case that all the codewords of
151 * some length L numerically precede all the codewords of length L + 1.
152 * Furthermore, if we have 2 symbols A and B with the same codeword
153 * length but symbol A is sorted before symbol B, then then we know that
154 * the codeword for A numerically precedes the codeword for B. */
155 unsigned decode_table_pos = 0;
158 wimlib_assert2(num_used_syms != 0);
160 unsigned sym = sorted_syms[i];
161 unsigned codeword_len = lens[sym];
162 if (codeword_len > table_bits)
165 unsigned num_entries = 1 << (table_bits - codeword_len);
167 (sizeof(unsigned long) / sizeof(decode_table[0])))
169 wimlib_assert2(decode_table_pos % 4 == 0);
170 BUILD_BUG_ON(sizeof(unsigned long) != 4 &&
171 sizeof(unsigned long) != 8);
173 unsigned long *p = (unsigned long *)&decode_table[decode_table_pos];
174 unsigned long n = num_entries /
175 (sizeof(unsigned long) /
176 sizeof(decode_table[0]));
177 unsigned long v = sym;
178 if (sizeof(unsigned long) >= 4)
180 if (sizeof(unsigned long) >= 8)
186 decode_table_pos += num_entries;
189 decode_table[decode_table_pos++] = sym;
190 } while (--num_entries);
192 wimlib_assert2(decode_table_pos <= table_num_entries);
193 if (++i == num_used_syms) {
194 wimlib_assert2(decode_table_pos == table_num_entries);
195 /* No codewords were longer than @table_bits, so the
196 * table is now entirely filled with the codewords. */
201 wimlib_assert2(i < num_used_syms);
202 wimlib_assert2(decode_table_pos < table_num_entries);
204 /* Fill in the remaining entries, which correspond to codes longer than
207 * First, zero out the rest of the entries. This is necessary so that
208 * the entries appear as "unallocated" in the next part. */
210 unsigned j = decode_table_pos;
213 } while (++j != table_num_entries);
216 /* Assert that 2**table_bits is at least num_syms. If this wasn't the
217 * case, we wouldn't be able to distinguish pointer entries from symbol
219 wimlib_assert2(table_num_entries >= num_syms);
221 /* The current Huffman codeword */
222 unsigned cur_codeword = decode_table_pos;
224 /* The tree nodes are allocated starting at decode_table[1 <<
225 * table_bits]. Remember that the full size of the table, including the
226 * extra space for the tree nodes, is actually 2**table_bits + 2 *
227 * num_syms slots, while table_num_entries is only 2**table_Bits. */
228 unsigned next_free_tree_slot = table_num_entries;
230 /* Go through every codeword of length greater than @table_bits,
231 * primarily in order of codeword length and secondarily in order of
233 unsigned prev_codeword_len = table_bits;
235 unsigned sym = sorted_syms[i];
236 unsigned codeword_len = lens[sym];
237 unsigned extra_bits = codeword_len - table_bits;
240 cur_codeword <<= (codeword_len - prev_codeword_len);
241 prev_codeword_len = codeword_len;
243 /* index of the current node; find it from the prefix of the
244 * current Huffman codeword. */
245 unsigned node_idx = cur_codeword >> extra_bits;
246 wimlib_assert2(node_idx < table_num_entries);
248 /* Go through each bit of the current Huffman codeword beyond
249 * the prefix of length @table_bits and walk the tree,
250 * allocating any slots that have not yet been allocated. */
253 /* If the current tree node points to nowhere
254 * but we need to follow it, allocate a new node
255 * for it to point to. */
256 if (decode_table[node_idx] == 0) {
257 decode_table[node_idx] = next_free_tree_slot;
258 decode_table[next_free_tree_slot++] = 0;
259 decode_table[next_free_tree_slot++] = 0;
260 wimlib_assert2(next_free_tree_slot <=
261 table_num_entries + 2 * num_syms);
264 /* Set node_idx to left child */
265 node_idx = decode_table[node_idx];
267 /* Is the next bit 0 or 1? If 0, go left (already done).
268 * If 1, go right by incrementing node_idx. */
270 node_idx += (cur_codeword >> extra_bits) & 1;
271 } while (extra_bits != 0);
273 /* node_idx is now the index of the leaf entry into which the
274 * actual symbol will go. */
275 decode_table[node_idx] = sym;
277 /* cur_codeword is always incremented because this is
278 * how canonical Huffman codes are generated (add 1 for
279 * each code, then left shift whenever the code length
282 } while (++i != num_used_syms);
286 /* Reads a Huffman-encoded symbol when it is known there are less than
287 * MAX_CODE_LEN bits remaining in the bitstream. */
288 int read_huffsym_near_end_of_input(struct input_bitstream *istream,
289 const u16 decode_table[],
295 unsigned bitsleft = istream->bitsleft;
300 if (table_bits > bitsleft) {
303 key_bits = bitstream_peek_bits(istream, key_size) <<
304 (table_bits - key_size);
306 key_size = table_bits;
307 bitsleft -= table_bits;
308 key_bits = bitstream_peek_bits(istream, table_bits);
311 sym = decode_table[key_bits];
312 if (sym >= num_syms) {
313 bitstream_remove_bits(istream, key_size);
316 ERROR("Input stream exhausted");
319 key_bits = sym + bitstream_peek_bits(istream, 1);
320 bitstream_remove_bits(istream, 1);
322 } while ((sym = decode_table[key_bits]) >= num_syms);
324 bitstream_remove_bits(istream, lens[sym]);