inftrees.c revision 1.11
1/* $OpenBSD: inftrees.c,v 1.11 2003/12/16 23:57:48 millert Exp $ */ 2/* inftrees.c -- generate Huffman trees for efficient decoding 3 * Copyright (C) 1995-2003 Mark Adler 4 * For conditions of distribution and use, see copyright notice in zlib.h 5 */ 6 7#include "zutil.h" 8#include "inftrees.h" 9 10#define MAXBITS 15 11 12const char inflate_copyright[] = 13 " inflate 1.2.1 Copyright 1995-2003 Mark Adler "; 14/* 15 If you use the zlib library in a product, an acknowledgment is welcome 16 in the documentation of your product. If for some reason you cannot 17 include such an acknowledgment, I would appreciate that you keep this 18 copyright string in the executable of your product. 19 */ 20 21/* 22 Build a set of tables to decode the provided canonical Huffman code. 23 The code lengths are lens[0..codes-1]. The result starts at *table, 24 whose indices are 0..2^bits-1. work is a writable array of at least 25 lens shorts, which is used as a work area. type is the type of code 26 to be generated, CODES, LENS, or DISTS. On return, zero is success, 27 -1 is an invalid code, and +1 means that ENOUGH isn't enough. table 28 on return points to the next available entry's address. bits is the 29 requested root table index bits, and on return it is the actual root 30 table index bits. It will differ if the request is greater than the 31 longest code or if it is less than the shortest code. 32 */ 33int inflate_table(type, lens, codes, table, bits, work) 34codetype type; 35unsigned short FAR *lens; 36unsigned codes; 37code FAR * FAR *table; 38unsigned FAR *bits; 39unsigned short FAR *work; 40{ 41 unsigned len; /* a code's length in bits */ 42 unsigned sym; /* index of code symbols */ 43 unsigned min, max; /* minimum and maximum code lengths */ 44 unsigned root; /* number of index bits for root table */ 45 unsigned curr; /* number of index bits for current table */ 46 unsigned drop; /* code bits to drop for sub-table */ 47 int left; /* number of prefix codes available */ 48 unsigned used; /* code entries in table used */ 49 unsigned huff; /* Huffman code */ 50 unsigned incr; /* for incrementing code, index */ 51 unsigned fill; /* index for replicating entries */ 52 unsigned low; /* low bits for current root entry */ 53 unsigned mask; /* mask for low root bits */ 54 code this; /* table entry for duplication */ 55 code FAR *next; /* next available space in table */ 56 const unsigned short FAR *base; /* base value table to use */ 57 const unsigned short FAR *extra; /* extra bits table to use */ 58 int end; /* use base and extra for symbol > end */ 59 unsigned short count[MAXBITS+1]; /* number of codes of each length */ 60 unsigned short offs[MAXBITS+1]; /* offsets in table for each length */ 61 static const unsigned short lbase[31] = { /* Length codes 257..285 base */ 62 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 15, 17, 19, 23, 27, 31, 63 35, 43, 51, 59, 67, 83, 99, 115, 131, 163, 195, 227, 258, 0, 0}; 64 static const unsigned short lext[31] = { /* Length codes 257..285 extra */ 65 16, 16, 16, 16, 16, 16, 16, 16, 17, 17, 17, 17, 18, 18, 18, 18, 66 19, 19, 19, 19, 20, 20, 20, 20, 21, 21, 21, 21, 16, 76, 66}; 67 static const unsigned short dbase[32] = { /* Distance codes 0..29 base */ 68 1, 2, 3, 4, 5, 7, 9, 13, 17, 25, 33, 49, 65, 97, 129, 193, 69 257, 385, 513, 769, 1025, 1537, 2049, 3073, 4097, 6145, 70 8193, 12289, 16385, 24577, 0, 0}; 71 static const unsigned short dext[32] = { /* Distance codes 0..29 extra */ 72 16, 16, 16, 16, 17, 17, 18, 18, 19, 19, 20, 20, 21, 21, 22, 22, 73 23, 23, 24, 24, 25, 25, 26, 26, 27, 27, 74 28, 28, 29, 29, 64, 64}; 75 76 /* 77 Process a set of code lengths to create a canonical Huffman code. The 78 code lengths are lens[0..codes-1]. Each length corresponds to the 79 symbols 0..codes-1. The Huffman code is generated by first sorting the 80 symbols by length from short to long, and retaining the symbol order 81 for codes with equal lengths. Then the code starts with all zero bits 82 for the first code of the shortest length, and the codes are integer 83 increments for the same length, and zeros are appended as the length 84 increases. For the deflate format, these bits are stored backwards 85 from their more natural integer increment ordering, and so when the 86 decoding tables are built in the large loop below, the integer codes 87 are incremented backwards. 88 89 This routine assumes, but does not check, that all of the entries in 90 lens[] are in the range 0..MAXBITS. The caller must assure this. 91 1..MAXBITS is interpreted as that code length. zero means that that 92 symbol does not occur in this code. 93 94 The codes are sorted by computing a count of codes for each length, 95 creating from that a table of starting indices for each length in the 96 sorted table, and then entering the symbols in order in the sorted 97 table. The sorted table is work[], with that space being provided by 98 the caller. 99 100 The length counts are used for other purposes as well, i.e. finding 101 the minimum and maximum length codes, determining if there are any 102 codes at all, checking for a valid set of lengths, and looking ahead 103 at length counts to determine sub-table sizes when building the 104 decoding tables. 105 */ 106 107 /* accumulate lengths for codes (assumes lens[] all in 0..MAXBITS) */ 108 for (len = 0; len <= MAXBITS; len++) 109 count[len] = 0; 110 for (sym = 0; sym < codes; sym++) 111 count[lens[sym]]++; 112 113 /* bound code lengths, force root to be within code lengths */ 114 root = *bits; 115 for (max = MAXBITS; max >= 1; max--) 116 if (count[max] != 0) break; 117 if (root > max) root = max; 118 if (max == 0) return -1; /* no codes! */ 119 for (min = 1; min <= MAXBITS; min++) 120 if (count[min] != 0) break; 121 if (root < min) root = min; 122 123 /* check for an over-subscribed or incomplete set of lengths */ 124 left = 1; 125 for (len = 1; len <= MAXBITS; len++) { 126 left <<= 1; 127 left -= count[len]; 128 if (left < 0) return -1; /* over-subscribed */ 129 } 130 if (left > 0 && (type == CODES || (codes - count[0] != 1))) 131 return -1; /* incomplete set */ 132 133 /* generate offsets into symbol table for each length for sorting */ 134 offs[1] = 0; 135 for (len = 1; len < MAXBITS; len++) 136 offs[len + 1] = offs[len] + count[len]; 137 138 /* sort symbols by length, by symbol order within each length */ 139 for (sym = 0; sym < codes; sym++) 140 if (lens[sym] != 0) work[offs[lens[sym]]++] = (unsigned short)sym; 141 142 /* 143 Create and fill in decoding tables. In this loop, the table being 144 filled is at next and has curr index bits. The code being used is huff 145 with length len. That code is converted to an index by dropping drop 146 bits off of the bottom. For codes where len is less than drop + curr, 147 those top drop + curr - len bits are incremented through all values to 148 fill the table with replicated entries. 149 150 root is the number of index bits for the root table. When len exceeds 151 root, sub-tables are created pointed to by the root entry with an index 152 of the low root bits of huff. This is saved in low to check for when a 153 new sub-table should be started. drop is zero when the root table is 154 being filled, and drop is root when sub-tables are being filled. 155 156 When a new sub-table is needed, it is necessary to look ahead in the 157 code lengths to determine what size sub-table is needed. The length 158 counts are used for this, and so count[] is decremented as codes are 159 entered in the tables. 160 161 used keeps track of how many table entries have been allocated from the 162 provided *table space. It is checked when a LENS table is being made 163 against the space in *table, ENOUGH, minus the maximum space needed by 164 the worst case distance code, MAXD. This should never happen, but the 165 sufficiency of ENOUGH has not been proven exhaustively, hence the check. 166 This assumes that when type == LENS, bits == 9. 167 168 sym increments through all symbols, and the loop terminates when 169 all codes of length max, i.e. all codes, have been processed. This 170 routine permits incomplete codes, so another loop after this one fills 171 in the rest of the decoding tables with invalid code markers. 172 */ 173 174 /* set up for code type */ 175 switch (type) { 176 case CODES: 177 base = extra = work; /* dummy value--not used */ 178 end = 19; 179 break; 180 case LENS: 181 base = lbase; 182 base -= 257; 183 extra = lext; 184 extra -= 257; 185 end = 256; 186 break; 187 default: /* DISTS */ 188 base = dbase; 189 extra = dext; 190 end = -1; 191 } 192 193 /* initialize state for loop */ 194 huff = 0; /* starting code */ 195 sym = 0; /* starting code symbol */ 196 len = min; /* starting code length */ 197 next = *table; /* current table to fill in */ 198 curr = root; /* current table index bits */ 199 drop = 0; /* current bits to drop from code for index */ 200 low = (unsigned)(-1); /* trigger new sub-table when len > root */ 201 used = 1U << root; /* use root table entries */ 202 mask = used - 1; /* mask for comparing low */ 203 204 /* check available table space */ 205 if (type == LENS && used >= ENOUGH - MAXD) 206 return 1; 207 208 /* process all codes and make table entries */ 209 for (;;) { 210 /* create table entry */ 211 this.bits = (unsigned char)(len - drop); 212 if ((int)(work[sym]) < end) { 213 this.op = (unsigned char)0; 214 this.val = work[sym]; 215 } 216 else if ((int)(work[sym]) > end) { 217 this.op = (unsigned char)(extra[work[sym]]); 218 this.val = base[work[sym]]; 219 } 220 else { 221 this.op = (unsigned char)(32 + 64); /* end of block */ 222 this.val = 0; 223 } 224 225 /* replicate for those indices with low len bits equal to huff */ 226 incr = 1U << (len - drop); 227 fill = 1U << curr; 228 do { 229 fill -= incr; 230 next[(huff >> drop) + fill] = this; 231 } while (fill != 0); 232 233 /* backwards increment the len-bit code huff */ 234 incr = 1U << (len - 1); 235 while (huff & incr) 236 incr >>= 1; 237 if (incr != 0) { 238 huff &= incr - 1; 239 huff += incr; 240 } 241 else 242 huff = 0; 243 244 /* go to next symbol, update count, len */ 245 sym++; 246 if (--(count[len]) == 0) { 247 if (len == max) break; 248 len = lens[work[sym]]; 249 } 250 251 /* create new sub-table if needed */ 252 if (len > root && (huff & mask) != low) { 253 /* if first time, transition to sub-tables */ 254 if (drop == 0) 255 drop = root; 256 257 /* increment past last table */ 258 next += 1U << curr; 259 260 /* determine length of next table */ 261 curr = len - drop; 262 left = (int)(1 << curr); 263 while (curr + drop < max) { 264 left -= count[curr + drop]; 265 if (left <= 0) break; 266 curr++; 267 left <<= 1; 268 } 269 270 /* check for enough space */ 271 used += 1U << curr; 272 if (type == LENS && used >= ENOUGH - MAXD) 273 return 1; 274 275 /* point entry in root table to sub-table */ 276 low = huff & mask; 277 (*table)[low].op = (unsigned char)curr; 278 (*table)[low].bits = (unsigned char)root; 279 (*table)[low].val = (unsigned short)(next - *table); 280 } 281 } 282 283 /* 284 Fill in rest of table for incomplete codes. This loop is similar to the 285 loop above in incrementing huff for table indices. It is assumed that 286 len is equal to curr + drop, so there is no loop needed to increment 287 through high index bits. When the current sub-table is filled, the loop 288 drops back to the root table to fill in any remaining entries there. 289 */ 290 this.op = (unsigned char)64; /* invalid code marker */ 291 this.bits = (unsigned char)(len - drop); 292 this.val = (unsigned short)0; 293 while (huff != 0) { 294 /* when done with sub-table, drop back to root table */ 295 if (drop != 0 && (huff & mask) != low) { 296 drop = 0; 297 len = root; 298 next = *table; 299 curr = root; 300 this.bits = (unsigned char)len; 301 } 302 303 /* put invalid code marker in table */ 304 next[huff >> drop] = this; 305 306 /* backwards increment the len-bit code huff */ 307 incr = 1U << (len - 1); 308 while (huff & incr) 309 incr >>= 1; 310 if (incr != 0) { 311 huff &= incr - 1; 312 huff += incr; 313 } 314 else 315 huff = 0; 316 } 317 318 /* set return parameters */ 319 *table += used; 320 *bits = root; 321 return 0; 322} 323