00001 /* 00002 * puff.c 00003 * Copyright (C) 2002-2010 Mark Adler 00004 * For conditions of distribution and use, see copyright notice in puff.h 00005 * version 2.1, 4 Apr 2010 00006 * 00007 * puff.c is a simple inflate written to be an unambiguous way to specify the 00008 * deflate format. It is not written for speed but rather simplicity. As a 00009 * side benefit, this code might actually be useful when small code is more 00010 * important than speed, such as bootstrap applications. For typical deflate 00011 * data, zlib's inflate() is about four times as fast as puff(). zlib's 00012 * inflate compiles to around 20K on my machine, whereas puff.c compiles to 00013 * around 4K on my machine (a PowerPC using GNU cc). If the faster decode() 00014 * function here is used, then puff() is only twice as slow as zlib's 00015 * inflate(). 00016 * 00017 * All dynamically allocated memory comes from the stack. The stack required 00018 * is less than 2K bytes. This code is compatible with 16-bit int's and 00019 * assumes that long's are at least 32 bits. puff.c uses the short data type, 00020 * assumed to be 16 bits, for arrays in order to to conserve memory. The code 00021 * works whether integers are stored big endian or little endian. 00022 * 00023 * In the comments below are "Format notes" that describe the inflate process 00024 * and document some of the less obvious aspects of the format. This source 00025 * code is meant to supplement RFC 1951, which formally describes the deflate 00026 * format: 00027 * 00028 * http://www.zlib.org/rfc-deflate.html 00029 */ 00030 00031 /* 00032 * Change history: 00033 * 00034 * 1.0 10 Feb 2002 - First version 00035 * 1.1 17 Feb 2002 - Clarifications of some comments and notes 00036 * - Update puff() dest and source pointers on negative 00037 * errors to facilitate debugging deflators 00038 * - Remove longest from struct huffman -- not needed 00039 * - Simplify offs[] index in construct() 00040 * - Add input size and checking, using longjmp() to 00041 * maintain easy readability 00042 * - Use short data type for large arrays 00043 * - Use pointers instead of long to specify source and 00044 * destination sizes to avoid arbitrary 4 GB limits 00045 * 1.2 17 Mar 2002 - Add faster version of decode(), doubles speed (!), 00046 * but leave simple version for readabilty 00047 * - Make sure invalid distances detected if pointers 00048 * are 16 bits 00049 * - Fix fixed codes table error 00050 * - Provide a scanning mode for determining size of 00051 * uncompressed data 00052 * 1.3 20 Mar 2002 - Go back to lengths for puff() parameters [Jean-loup] 00053 * - Add a puff.h file for the interface 00054 * - Add braces in puff() for else do [Jean-loup] 00055 * - Use indexes instead of pointers for readability 00056 * 1.4 31 Mar 2002 - Simplify construct() code set check 00057 * - Fix some comments 00058 * - Add FIXLCODES #define 00059 * 1.5 6 Apr 2002 - Minor comment fixes 00060 * 1.6 7 Aug 2002 - Minor format changes 00061 * 1.7 3 Mar 2003 - Added test code for distribution 00062 * - Added zlib-like license 00063 * 1.8 9 Jan 2004 - Added some comments on no distance codes case 00064 * 1.9 21 Feb 2008 - Fix bug on 16-bit integer architectures [Pohland] 00065 * - Catch missing end-of-block symbol error 00066 * 2.0 25 Jul 2008 - Add #define to permit distance too far back 00067 * - Add option in TEST code for puff to write the data 00068 * - Add option in TEST code to skip input bytes 00069 * - Allow TEST code to read from piped stdin 00070 * 2.1 4 Apr 2010 - Avoid variable initialization for happier compilers 00071 * - Avoid unsigned comparisons for even happier compilers 00072 */ 00073 00074 #include <setjmp.h> /* for setjmp(), longjmp(), and jmp_buf */ 00075 #include "puff.h" /* prototype for puff() */ 00076 00077 #define local static /* for local function definitions */ 00078 #define NIL ((unsigned char *)0) /* for no output option */ 00079 00080 /* 00081 * Maximums for allocations and loops. It is not useful to change these -- 00082 * they are fixed by the deflate format. 00083 */ 00084 #define MAXBITS 15 /* maximum bits in a code */ 00085 #define MAXLCODES 286 /* maximum number of literal/length codes */ 00086 #define MAXDCODES 30 /* maximum number of distance codes */ 00087 #define MAXCODES (MAXLCODES+MAXDCODES) /* maximum codes lengths to read */ 00088 #define FIXLCODES 288 /* number of fixed literal/length codes */ 00089 00090 /* input and output state */ 00091 struct state { 00092 /* output state */ 00093 unsigned char *out; /* output buffer */ 00094 unsigned long outlen; /* available space at out */ 00095 unsigned long outcnt; /* bytes written to out so far */ 00096 00097 /* input state */ 00098 unsigned char *in; /* input buffer */ 00099 unsigned long inlen; /* available input at in */ 00100 unsigned long incnt; /* bytes read so far */ 00101 int bitbuf; /* bit buffer */ 00102 int bitcnt; /* number of bits in bit buffer */ 00103 00104 /* input limit error return state for bits() and decode() */ 00105 jmp_buf env; 00106 }; 00107 00108 /* 00109 * Return need bits from the input stream. This always leaves less than 00110 * eight bits in the buffer. bits() works properly for need == 0. 00111 * 00112 * Format notes: 00113 * 00114 * - Bits are stored in bytes from the least significant bit to the most 00115 * significant bit. Therefore bits are dropped from the bottom of the bit 00116 * buffer, using shift right, and new bytes are appended to the top of the 00117 * bit buffer, using shift left. 00118 */ 00119 local int bits(struct state *s, int need) 00120 { 00121 long val; /* bit accumulator (can use up to 20 bits) */ 00122 00123 /* load at least need bits into val */ 00124 val = s->bitbuf; 00125 while (s->bitcnt < need) { 00126 if (s->incnt == s->inlen) longjmp(s->env, 1); /* out of input */ 00127 val |= (long)(s->in[s->incnt++]) << s->bitcnt; /* load eight bits */ 00128 s->bitcnt += 8; 00129 } 00130 00131 /* drop need bits and update buffer, always zero to seven bits left */ 00132 s->bitbuf = (int)(val >> need); 00133 s->bitcnt -= need; 00134 00135 /* return need bits, zeroing the bits above that */ 00136 return (int)(val & ((1L << need) - 1)); 00137 } 00138 00139 /* 00140 * Process a stored block. 00141 * 00142 * Format notes: 00143 * 00144 * - After the two-bit stored block type (00), the stored block length and 00145 * stored bytes are byte-aligned for fast copying. Therefore any leftover 00146 * bits in the byte that has the last bit of the type, as many as seven, are 00147 * discarded. The value of the discarded bits are not defined and should not 00148 * be checked against any expectation. 00149 * 00150 * - The second inverted copy of the stored block length does not have to be 00151 * checked, but it's probably a good idea to do so anyway. 00152 * 00153 * - A stored block can have zero length. This is sometimes used to byte-align 00154 * subsets of the compressed data for random access or partial recovery. 00155 */ 00156 local int stored(struct state *s) 00157 { 00158 unsigned len; /* length of stored block */ 00159 00160 /* discard leftover bits from current byte (assumes s->bitcnt < 8) */ 00161 s->bitbuf = 0; 00162 s->bitcnt = 0; 00163 00164 /* get length and check against its one's complement */ 00165 if (s->incnt + 4 > s->inlen) return 2; /* not enough input */ 00166 len = s->in[s->incnt++]; 00167 len |= s->in[s->incnt++] << 8; 00168 if (s->in[s->incnt++] != (~len & 0xff) || 00169 s->in[s->incnt++] != ((~len >> 8) & 0xff)) 00170 return -2; /* didn't match complement! */ 00171 00172 /* copy len bytes from in to out */ 00173 if (s->incnt + len > s->inlen) return 2; /* not enough input */ 00174 if (s->out != NIL) { 00175 if (s->outcnt + len > s->outlen) 00176 return 1; /* not enough output space */ 00177 while (len--) 00178 s->out[s->outcnt++] = s->in[s->incnt++]; 00179 } 00180 else { /* just scanning */ 00181 s->outcnt += len; 00182 s->incnt += len; 00183 } 00184 00185 /* done with a valid stored block */ 00186 return 0; 00187 } 00188 00189 /* 00190 * Huffman code decoding tables. count[1..MAXBITS] is the number of symbols of 00191 * each length, which for a canonical code are stepped through in order. 00192 * symbol[] are the symbol values in canonical order, where the number of 00193 * entries is the sum of the counts in count[]. The decoding process can be 00194 * seen in the function decode() below. 00195 */ 00196 struct huffman { 00197 short *count; /* number of symbols of each length */ 00198 short *symbol; /* canonically ordered symbols */ 00199 }; 00200 00201 /* 00202 * Decode a code from the stream s using huffman table h. Return the symbol or 00203 * a negative value if there is an error. If all of the lengths are zero, i.e. 00204 * an empty code, or if the code is incomplete and an invalid code is received, 00205 * then -10 is returned after reading MAXBITS bits. 00206 * 00207 * Format notes: 00208 * 00209 * - The codes as stored in the compressed data are bit-reversed relative to 00210 * a simple integer ordering of codes of the same lengths. Hence below the 00211 * bits are pulled from the compressed data one at a time and used to 00212 * build the code value reversed from what is in the stream in order to 00213 * permit simple integer comparisons for decoding. A table-based decoding 00214 * scheme (as used in zlib) does not need to do this reversal. 00215 * 00216 * - The first code for the shortest length is all zeros. Subsequent codes of 00217 * the same length are simply integer increments of the previous code. When 00218 * moving up a length, a zero bit is appended to the code. For a complete 00219 * code, the last code of the longest length will be all ones. 00220 * 00221 * - Incomplete codes are handled by this decoder, since they are permitted 00222 * in the deflate format. See the format notes for fixed() and dynamic(). 00223 */ 00224 #ifdef SLOW 00225 local int decode(struct state *s, struct huffman *h) 00226 { 00227 int len; /* current number of bits in code */ 00228 int code; /* len bits being decoded */ 00229 int first; /* first code of length len */ 00230 int count; /* number of codes of length len */ 00231 int index; /* index of first code of length len in symbol table */ 00232 00233 code = first = index = 0; 00234 for (len = 1; len <= MAXBITS; len++) { 00235 code |= bits(s, 1); /* get next bit */ 00236 count = h->count[len]; 00237 if (code - count < first) /* if length len, return symbol */ 00238 return h->symbol[index + (code - first)]; 00239 index += count; /* else update for next length */ 00240 first += count; 00241 first <<= 1; 00242 code <<= 1; 00243 } 00244 return -10; /* ran out of codes */ 00245 } 00246 00247 /* 00248 * A faster version of decode() for real applications of this code. It's not 00249 * as readable, but it makes puff() twice as fast. And it only makes the code 00250 * a few percent larger. 00251 */ 00252 #else /* !SLOW */ 00253 local int decode(struct state *s, struct huffman *h) 00254 { 00255 int len; /* current number of bits in code */ 00256 int code; /* len bits being decoded */ 00257 int first; /* first code of length len */ 00258 int count; /* number of codes of length len */ 00259 int index; /* index of first code of length len in symbol table */ 00260 int bitbuf; /* bits from stream */ 00261 int left; /* bits left in next or left to process */ 00262 short *next; /* next number of codes */ 00263 00264 bitbuf = s->bitbuf; 00265 left = s->bitcnt; 00266 code = first = index = 0; 00267 len = 1; 00268 next = h->count + 1; 00269 while (1) { 00270 while (left--) { 00271 code |= bitbuf & 1; 00272 bitbuf >>= 1; 00273 count = *next++; 00274 if (code - count < first) { /* if length len, return symbol */ 00275 s->bitbuf = bitbuf; 00276 s->bitcnt = (s->bitcnt - len) & 7; 00277 return h->symbol[index + (code - first)]; 00278 } 00279 index += count; /* else update for next length */ 00280 first += count; 00281 first <<= 1; 00282 code <<= 1; 00283 len++; 00284 } 00285 left = (MAXBITS+1) - len; 00286 if (left == 0) break; 00287 if (s->incnt == s->inlen) longjmp(s->env, 1); /* out of input */ 00288 bitbuf = s->in[s->incnt++]; 00289 if (left > 8) left = 8; 00290 } 00291 return -10; /* ran out of codes */ 00292 } 00293 #endif /* SLOW */ 00294 00295 /* 00296 * Given the list of code lengths length[0..n-1] representing a canonical 00297 * Huffman code for n symbols, construct the tables required to decode those 00298 * codes. Those tables are the number of codes of each length, and the symbols 00299 * sorted by length, retaining their original order within each length. The 00300 * return value is zero for a complete code set, negative for an over- 00301 * subscribed code set, and positive for an incomplete code set. The tables 00302 * can be used if the return value is zero or positive, but they cannot be used 00303 * if the return value is negative. If the return value is zero, it is not 00304 * possible for decode() using that table to return an error--any stream of 00305 * enough bits will resolve to a symbol. If the return value is positive, then 00306 * it is possible for decode() using that table to return an error for received 00307 * codes past the end of the incomplete lengths. 00308 * 00309 * Not used by decode(), but used for error checking, h->count[0] is the number 00310 * of the n symbols not in the code. So n - h->count[0] is the number of 00311 * codes. This is useful for checking for incomplete codes that have more than 00312 * one symbol, which is an error in a dynamic block. 00313 * 00314 * Assumption: for all i in 0..n-1, 0 <= length[i] <= MAXBITS 00315 * This is assured by the construction of the length arrays in dynamic() and 00316 * fixed() and is not verified by construct(). 00317 * 00318 * Format notes: 00319 * 00320 * - Permitted and expected examples of incomplete codes are one of the fixed 00321 * codes and any code with a single symbol which in deflate is coded as one 00322 * bit instead of zero bits. See the format notes for fixed() and dynamic(). 00323 * 00324 * - Within a given code length, the symbols are kept in ascending order for 00325 * the code bits definition. 00326 */ 00327 local int construct(struct huffman *h, short *length, int n) 00328 { 00329 int symbol; /* current symbol when stepping through length[] */ 00330 int len; /* current length when stepping through h->count[] */ 00331 int left; /* number of possible codes left of current length */ 00332 short offs[MAXBITS+1]; /* offsets in symbol table for each length */ 00333 00334 /* count number of codes of each length */ 00335 for (len = 0; len <= MAXBITS; len++) 00336 h->count[len] = 0; 00337 for (symbol = 0; symbol < n; symbol++) 00338 (h->count[length[symbol]])++; /* assumes lengths are within bounds */ 00339 if (h->count[0] == n) /* no codes! */ 00340 return 0; /* complete, but decode() will fail */ 00341 00342 /* check for an over-subscribed or incomplete set of lengths */ 00343 left = 1; /* one possible code of zero length */ 00344 for (len = 1; len <= MAXBITS; len++) { 00345 left <<= 1; /* one more bit, double codes left */ 00346 left -= h->count[len]; /* deduct count from possible codes */ 00347 if (left < 0) return left; /* over-subscribed--return negative */ 00348 } /* left > 0 means incomplete */ 00349 00350 /* generate offsets into symbol table for each length for sorting */ 00351 offs[1] = 0; 00352 for (len = 1; len < MAXBITS; len++) 00353 offs[len + 1] = offs[len] + h->count[len]; 00354 00355 /* 00356 * put symbols in table sorted by length, by symbol order within each 00357 * length 00358 */ 00359 for (symbol = 0; symbol < n; symbol++) 00360 if (length[symbol] != 0) 00361 h->symbol[offs[length[symbol]]++] = symbol; 00362 00363 /* return zero for complete set, positive for incomplete set */ 00364 return left; 00365 } 00366 00367 /* 00368 * Decode literal/length and distance codes until an end-of-block code. 00369 * 00370 * Format notes: 00371 * 00372 * - Compressed data that is after the block type if fixed or after the code 00373 * description if dynamic is a combination of literals and length/distance 00374 * pairs terminated by and end-of-block code. Literals are simply Huffman 00375 * coded bytes. A length/distance pair is a coded length followed by a 00376 * coded distance to represent a string that occurs earlier in the 00377 * uncompressed data that occurs again at the current location. 00378 * 00379 * - Literals, lengths, and the end-of-block code are combined into a single 00380 * code of up to 286 symbols. They are 256 literals (0..255), 29 length 00381 * symbols (257..285), and the end-of-block symbol (256). 00382 * 00383 * - There are 256 possible lengths (3..258), and so 29 symbols are not enough 00384 * to represent all of those. Lengths 3..10 and 258 are in fact represented 00385 * by just a length symbol. Lengths 11..257 are represented as a symbol and 00386 * some number of extra bits that are added as an integer to the base length 00387 * of the length symbol. The number of extra bits is determined by the base 00388 * length symbol. These are in the static arrays below, lens[] for the base 00389 * lengths and lext[] for the corresponding number of extra bits. 00390 * 00391 * - The reason that 258 gets its own symbol is that the longest length is used 00392 * often in highly redundant files. Note that 258 can also be coded as the 00393 * base value 227 plus the maximum extra value of 31. While a good deflate 00394 * should never do this, it is not an error, and should be decoded properly. 00395 * 00396 * - If a length is decoded, including its extra bits if any, then it is 00397 * followed a distance code. There are up to 30 distance symbols. Again 00398 * there are many more possible distances (1..32768), so extra bits are added 00399 * to a base value represented by the symbol. The distances 1..4 get their 00400 * own symbol, but the rest require extra bits. The base distances and 00401 * corresponding number of extra bits are below in the static arrays dist[] 00402 * and dext[]. 00403 * 00404 * - Literal bytes are simply written to the output. A length/distance pair is 00405 * an instruction to copy previously uncompressed bytes to the output. The 00406 * copy is from distance bytes back in the output stream, copying for length 00407 * bytes. 00408 * 00409 * - Distances pointing before the beginning of the output data are not 00410 * permitted. 00411 * 00412 * - Overlapped copies, where the length is greater than the distance, are 00413 * allowed and common. For example, a distance of one and a length of 258 00414 * simply copies the last byte 258 times. A distance of four and a length of 00415 * twelve copies the last four bytes three times. A simple forward copy 00416 * ignoring whether the length is greater than the distance or not implements 00417 * this correctly. You should not use memcpy() since its behavior is not 00418 * defined for overlapped arrays. You should not use memmove() or bcopy() 00419 * since though their behavior -is- defined for overlapping arrays, it is 00420 * defined to do the wrong thing in this case. 00421 */ 00422 local int codes(struct state *s, 00423 struct huffman *lencode, 00424 struct huffman *distcode) 00425 { 00426 int symbol; /* decoded symbol */ 00427 int len; /* length for copy */ 00428 unsigned dist; /* distance for copy */ 00429 static const short lens[29] = { /* Size base for length codes 257..285 */ 00430 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 15, 17, 19, 23, 27, 31, 00431 35, 43, 51, 59, 67, 83, 99, 115, 131, 163, 195, 227, 258}; 00432 static const short lext[29] = { /* Extra bits for length codes 257..285 */ 00433 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, 00434 3, 3, 3, 3, 4, 4, 4, 4, 5, 5, 5, 5, 0}; 00435 static const short dists[30] = { /* Offset base for distance codes 0..29 */ 00436 1, 2, 3, 4, 5, 7, 9, 13, 17, 25, 33, 49, 65, 97, 129, 193, 00437 257, 385, 513, 769, 1025, 1537, 2049, 3073, 4097, 6145, 00438 8193, 12289, 16385, 24577}; 00439 static const short dext[30] = { /* Extra bits for distance codes 0..29 */ 00440 0, 0, 0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 00441 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, 00442 12, 12, 13, 13}; 00443 00444 /* decode literals and length/distance pairs */ 00445 do { 00446 symbol = decode(s, lencode); 00447 if (symbol < 0) return symbol; /* invalid symbol */ 00448 if (symbol < 256) { /* literal: symbol is the byte */ 00449 /* write out the literal */ 00450 if (s->out != NIL) { 00451 if (s->outcnt == s->outlen) return 1; 00452 s->out[s->outcnt] = symbol; 00453 } 00454 s->outcnt++; 00455 } 00456 else if (symbol > 256) { /* length */ 00457 /* get and compute length */ 00458 symbol -= 257; 00459 if (symbol >= 29) return -10; /* invalid fixed code */ 00460 len = lens[symbol] + bits(s, lext[symbol]); 00461 00462 /* get and check distance */ 00463 symbol = decode(s, distcode); 00464 if (symbol < 0) return symbol; /* invalid symbol */ 00465 dist = dists[symbol] + bits(s, dext[symbol]); 00466 #ifndef INFLATE_ALLOW_INVALID_DISTANCE_TOOFAR_ARRR 00467 if (dist > s->outcnt) 00468 return -11; /* distance too far back */ 00469 #endif 00470 00471 /* copy length bytes from distance bytes back */ 00472 if (s->out != NIL) { 00473 if (s->outcnt + len > s->outlen) return 1; 00474 while (len--) { 00475 s->out[s->outcnt] = 00476 #ifdef INFLATE_ALLOW_INVALID_DISTANCE_TOOFAR_ARRR 00477 dist > s->outcnt ? 0 : 00478 #endif 00479 s->out[s->outcnt - dist]; 00480 s->outcnt++; 00481 } 00482 } 00483 else 00484 s->outcnt += len; 00485 } 00486 } while (symbol != 256); /* end of block symbol */ 00487 00488 /* done with a valid fixed or dynamic block */ 00489 return 0; 00490 } 00491 00492 /* 00493 * Process a fixed codes block. 00494 * 00495 * Format notes: 00496 * 00497 * - This block type can be useful for compressing small amounts of data for 00498 * which the size of the code descriptions in a dynamic block exceeds the 00499 * benefit of custom codes for that block. For fixed codes, no bits are 00500 * spent on code descriptions. Instead the code lengths for literal/length 00501 * codes and distance codes are fixed. The specific lengths for each symbol 00502 * can be seen in the "for" loops below. 00503 * 00504 * - The literal/length code is complete, but has two symbols that are invalid 00505 * and should result in an error if received. This cannot be implemented 00506 * simply as an incomplete code since those two symbols are in the "middle" 00507 * of the code. They are eight bits long and the longest literal/length\ 00508 * code is nine bits. Therefore the code must be constructed with those 00509 * symbols, and the invalid symbols must be detected after decoding. 00510 * 00511 * - The fixed distance codes also have two invalid symbols that should result 00512 * in an error if received. Since all of the distance codes are the same 00513 * length, this can be implemented as an incomplete code. Then the invalid 00514 * codes are detected while decoding. 00515 */ 00516 local int fixed(struct state *s) 00517 { 00518 static int virgin = 1; 00519 static short lencnt[MAXBITS+1], lensym[FIXLCODES]; 00520 static short distcnt[MAXBITS+1], distsym[MAXDCODES]; 00521 static struct huffman lencode, distcode; 00522 00523 /* build fixed huffman tables if first call (may not be thread safe) */ 00524 if (virgin) { 00525 int symbol; 00526 short lengths[FIXLCODES]; 00527 00528 /* literal/length table */ 00529 for (symbol = 0; symbol < 144; symbol++) 00530 lengths[symbol] = 8; 00531 for (; symbol < 256; symbol++) 00532 lengths[symbol] = 9; 00533 for (; symbol < 280; symbol++) 00534 lengths[symbol] = 7; 00535 for (; symbol < FIXLCODES; symbol++) 00536 lengths[symbol] = 8; 00537 construct(&lencode, lengths, FIXLCODES); 00538 00539 /* distance table */ 00540 for (symbol = 0; symbol < MAXDCODES; symbol++) 00541 lengths[symbol] = 5; 00542 construct(&distcode, lengths, MAXDCODES); 00543 00544 /* construct lencode and distcode */ 00545 lencode.count = lencnt; 00546 lencode.symbol = lensym; 00547 distcode.count = distcnt; 00548 distcode.symbol = distsym; 00549 00550 /* do this just once */ 00551 virgin = 0; 00552 } 00553 00554 /* decode data until end-of-block code */ 00555 return codes(s, &lencode, &distcode); 00556 } 00557 00558 /* 00559 * Process a dynamic codes block. 00560 * 00561 * Format notes: 00562 * 00563 * - A dynamic block starts with a description of the literal/length and 00564 * distance codes for that block. New dynamic blocks allow the compressor to 00565 * rapidly adapt to changing data with new codes optimized for that data. 00566 * 00567 * - The codes used by the deflate format are "canonical", which means that 00568 * the actual bits of the codes are generated in an unambiguous way simply 00569 * from the number of bits in each code. Therefore the code descriptions 00570 * are simply a list of code lengths for each symbol. 00571 * 00572 * - The code lengths are stored in order for the symbols, so lengths are 00573 * provided for each of the literal/length symbols, and for each of the 00574 * distance symbols. 00575 * 00576 * - If a symbol is not used in the block, this is represented by a zero as 00577 * as the code length. This does not mean a zero-length code, but rather 00578 * that no code should be created for this symbol. There is no way in the 00579 * deflate format to represent a zero-length code. 00580 * 00581 * - The maximum number of bits in a code is 15, so the possible lengths for 00582 * any code are 1..15. 00583 * 00584 * - The fact that a length of zero is not permitted for a code has an 00585 * interesting consequence. Normally if only one symbol is used for a given 00586 * code, then in fact that code could be represented with zero bits. However 00587 * in deflate, that code has to be at least one bit. So for example, if 00588 * only a single distance base symbol appears in a block, then it will be 00589 * represented by a single code of length one, in particular one 0 bit. This 00590 * is an incomplete code, since if a 1 bit is received, it has no meaning, 00591 * and should result in an error. So incomplete distance codes of one symbol 00592 * should be permitted, and the receipt of invalid codes should be handled. 00593 * 00594 * - It is also possible to have a single literal/length code, but that code 00595 * must be the end-of-block code, since every dynamic block has one. This 00596 * is not the most efficient way to create an empty block (an empty fixed 00597 * block is fewer bits), but it is allowed by the format. So incomplete 00598 * literal/length codes of one symbol should also be permitted. 00599 * 00600 * - If there are only literal codes and no lengths, then there are no distance 00601 * codes. This is represented by one distance code with zero bits. 00602 * 00603 * - The list of up to 286 length/literal lengths and up to 30 distance lengths 00604 * are themselves compressed using Huffman codes and run-length encoding. In 00605 * the list of code lengths, a 0 symbol means no code, a 1..15 symbol means 00606 * that length, and the symbols 16, 17, and 18 are run-length instructions. 00607 * Each of 16, 17, and 18 are follwed by extra bits to define the length of 00608 * the run. 16 copies the last length 3 to 6 times. 17 represents 3 to 10 00609 * zero lengths, and 18 represents 11 to 138 zero lengths. Unused symbols 00610 * are common, hence the special coding for zero lengths. 00611 * 00612 * - The symbols for 0..18 are Huffman coded, and so that code must be 00613 * described first. This is simply a sequence of up to 19 three-bit values 00614 * representing no code (0) or the code length for that symbol (1..7). 00615 * 00616 * - A dynamic block starts with three fixed-size counts from which is computed 00617 * the number of literal/length code lengths, the number of distance code 00618 * lengths, and the number of code length code lengths (ok, you come up with 00619 * a better name!) in the code descriptions. For the literal/length and 00620 * distance codes, lengths after those provided are considered zero, i.e. no 00621 * code. The code length code lengths are received in a permuted order (see 00622 * the order[] array below) to make a short code length code length list more 00623 * likely. As it turns out, very short and very long codes are less likely 00624 * to be seen in a dynamic code description, hence what may appear initially 00625 * to be a peculiar ordering. 00626 * 00627 * - Given the number of literal/length code lengths (nlen) and distance code 00628 * lengths (ndist), then they are treated as one long list of nlen + ndist 00629 * code lengths. Therefore run-length coding can and often does cross the 00630 * boundary between the two sets of lengths. 00631 * 00632 * - So to summarize, the code description at the start of a dynamic block is 00633 * three counts for the number of code lengths for the literal/length codes, 00634 * the distance codes, and the code length codes. This is followed by the 00635 * code length code lengths, three bits each. This is used to construct the 00636 * code length code which is used to read the remainder of the lengths. Then 00637 * the literal/length code lengths and distance lengths are read as a single 00638 * set of lengths using the code length codes. Codes are constructed from 00639 * the resulting two sets of lengths, and then finally you can start 00640 * decoding actual compressed data in the block. 00641 * 00642 * - For reference, a "typical" size for the code description in a dynamic 00643 * block is around 80 bytes. 00644 */ 00645 local int dynamic(struct state *s) 00646 { 00647 int nlen, ndist, ncode; /* number of lengths in descriptor */ 00648 int index; /* index of lengths[] */ 00649 int err; /* construct() return value */ 00650 short lengths[MAXCODES]; /* descriptor code lengths */ 00651 short lencnt[MAXBITS+1], lensym[MAXLCODES]; /* lencode memory */ 00652 short distcnt[MAXBITS+1], distsym[MAXDCODES]; /* distcode memory */ 00653 struct huffman lencode, distcode; /* length and distance codes */ 00654 static const short order[19] = /* permutation of code length codes */ 00655 {16, 17, 18, 0, 8, 7, 9, 6, 10, 5, 11, 4, 12, 3, 13, 2, 14, 1, 15}; 00656 00657 /* construct lencode and distcode */ 00658 lencode.count = lencnt; 00659 lencode.symbol = lensym; 00660 distcode.count = distcnt; 00661 distcode.symbol = distsym; 00662 00663 /* get number of lengths in each table, check lengths */ 00664 nlen = bits(s, 5) + 257; 00665 ndist = bits(s, 5) + 1; 00666 ncode = bits(s, 4) + 4; 00667 if (nlen > MAXLCODES || ndist > MAXDCODES) 00668 return -3; /* bad counts */ 00669 00670 /* read code length code lengths (really), missing lengths are zero */ 00671 for (index = 0; index < ncode; index++) 00672 lengths[order[index]] = bits(s, 3); 00673 for (; index < 19; index++) 00674 lengths[order[index]] = 0; 00675 00676 /* build huffman table for code lengths codes (use lencode temporarily) */ 00677 err = construct(&lencode, lengths, 19); 00678 if (err != 0) return -4; /* require complete code set here */ 00679 00680 /* read length/literal and distance code length tables */ 00681 index = 0; 00682 while (index < nlen + ndist) { 00683 int symbol; /* decoded value */ 00684 int len; /* last length to repeat */ 00685 00686 symbol = decode(s, &lencode); 00687 if (symbol < 16) /* length in 0..15 */ 00688 lengths[index++] = symbol; 00689 else { /* repeat instruction */ 00690 len = 0; /* assume repeating zeros */ 00691 if (symbol == 16) { /* repeat last length 3..6 times */ 00692 if (index == 0) return -5; /* no last length! */ 00693 len = lengths[index - 1]; /* last length */ 00694 symbol = 3 + bits(s, 2); 00695 } 00696 else if (symbol == 17) /* repeat zero 3..10 times */ 00697 symbol = 3 + bits(s, 3); 00698 else /* == 18, repeat zero 11..138 times */ 00699 symbol = 11 + bits(s, 7); 00700 if (index + symbol > nlen + ndist) 00701 return -6; /* too many lengths! */ 00702 while (symbol--) /* repeat last or zero symbol times */ 00703 lengths[index++] = len; 00704 } 00705 } 00706 00707 /* check for end-of-block code -- there better be one! */ 00708 if (lengths[256] == 0) 00709 return -9; 00710 00711 /* build huffman table for literal/length codes */ 00712 err = construct(&lencode, lengths, nlen); 00713 if (err < 0 || (err > 0 && nlen - lencode.count[0] != 1)) 00714 return -7; /* only allow incomplete codes if just one code */ 00715 00716 /* build huffman table for distance codes */ 00717 err = construct(&distcode, lengths + nlen, ndist); 00718 if (err < 0 || (err > 0 && ndist - distcode.count[0] != 1)) 00719 return -8; /* only allow incomplete codes if just one code */ 00720 00721 /* decode data until end-of-block code */ 00722 return codes(s, &lencode, &distcode); 00723 } 00724 00725 /* 00726 * Inflate source to dest. On return, destlen and sourcelen are updated to the 00727 * size of the uncompressed data and the size of the deflate data respectively. 00728 * On success, the return value of puff() is zero. If there is an error in the 00729 * source data, i.e. it is not in the deflate format, then a negative value is 00730 * returned. If there is not enough input available or there is not enough 00731 * output space, then a positive error is returned. In that case, destlen and 00732 * sourcelen are not updated to facilitate retrying from the beginning with the 00733 * provision of more input data or more output space. In the case of invalid 00734 * inflate data (a negative error), the dest and source pointers are updated to 00735 * facilitate the debugging of deflators. 00736 * 00737 * puff() also has a mode to determine the size of the uncompressed output with 00738 * no output written. For this dest must be (unsigned char *)0. In this case, 00739 * the input value of *destlen is ignored, and on return *destlen is set to the 00740 * size of the uncompressed output. 00741 * 00742 * The return codes are: 00743 * 00744 * 2: available inflate data did not terminate 00745 * 1: output space exhausted before completing inflate 00746 * 0: successful inflate 00747 * -1: invalid block type (type == 3) 00748 * -2: stored block length did not match one's complement 00749 * -3: dynamic block code description: too many length or distance codes 00750 * -4: dynamic block code description: code lengths codes incomplete 00751 * -5: dynamic block code description: repeat lengths with no first length 00752 * -6: dynamic block code description: repeat more than specified lengths 00753 * -7: dynamic block code description: invalid literal/length code lengths 00754 * -8: dynamic block code description: invalid distance code lengths 00755 * -9: dynamic block code description: missing end-of-block code 00756 * -10: invalid literal/length or distance code in fixed or dynamic block 00757 * -11: distance is too far back in fixed or dynamic block 00758 * 00759 * Format notes: 00760 * 00761 * - Three bits are read for each block to determine the kind of block and 00762 * whether or not it is the last block. Then the block is decoded and the 00763 * process repeated if it was not the last block. 00764 * 00765 * - The leftover bits in the last byte of the deflate data after the last 00766 * block (if it was a fixed or dynamic block) are undefined and have no 00767 * expected values to check. 00768 */ 00769 int puff(unsigned char *dest, /* pointer to destination pointer */ 00770 unsigned long *destlen, /* amount of output space */ 00771 unsigned char *source, /* pointer to source data pointer */ 00772 unsigned long *sourcelen) /* amount of input available */ 00773 { 00774 struct state s; /* input/output state */ 00775 int last, type; /* block information */ 00776 int err; /* return value */ 00777 00778 /* initialize output state */ 00779 s.out = dest; 00780 s.outlen = *destlen; /* ignored if dest is NIL */ 00781 s.outcnt = 0; 00782 00783 /* initialize input state */ 00784 s.in = source; 00785 s.inlen = *sourcelen; 00786 s.incnt = 0; 00787 s.bitbuf = 0; 00788 s.bitcnt = 0; 00789 00790 /* return if bits() or decode() tries to read past available input */ 00791 if (setjmp(s.env) != 0) /* if came back here via longjmp() */ 00792 err = 2; /* then skip do-loop, return error */ 00793 else { 00794 /* process blocks until last block or error */ 00795 do { 00796 last = bits(&s, 1); /* one if last block */ 00797 type = bits(&s, 2); /* block type 0..3 */ 00798 err = type == 0 ? stored(&s) : 00799 (type == 1 ? fixed(&s) : 00800 (type == 2 ? dynamic(&s) : 00801 -1)); /* type == 3, invalid */ 00802 if (err != 0) break; /* return with error */ 00803 } while (!last); 00804 } 00805 00806 /* update the lengths and return */ 00807 if (err <= 0) { 00808 *destlen = s.outcnt; 00809 *sourcelen = s.incnt; 00810 } 00811 return err; 00812 } 00813 00814 #ifdef TEST 00815 /* Examples of how to use puff(). 00816 00817 Usage: puff [-w] [-nnn] file 00818 ... | puff [-w] [-nnn] 00819 00820 where file is the input file with deflate data, nnn is the number of bytes 00821 of input to skip before inflating (e.g. to skip a zlib or gzip header), and 00822 -w is used to write the decompressed data to stdout */ 00823 00824 #include <stdio.h> 00825 #include <stdlib.h> 00826 00827 /* Return size times approximately the cube root of 2, keeping the result as 1, 00828 3, or 5 times a power of 2 -- the result is always > size, until the result 00829 is the maximum value of an unsigned long, where it remains. This is useful 00830 to keep reallocations less than ~33% over the actual data. */ 00831 local size_t bythirds(size_t size) 00832 { 00833 int n; 00834 size_t m; 00835 00836 m = size; 00837 for (n = 0; m; n++) 00838 m >>= 1; 00839 if (n < 3) 00840 return size + 1; 00841 n -= 3; 00842 m = size >> n; 00843 m += m == 6 ? 2 : 1; 00844 m <<= n; 00845 return m > size ? m : (size_t)(-1); 00846 } 00847 00848 /* Read the input file *name, or stdin if name is NULL, into allocated memory. 00849 Reallocate to larger buffers until the entire file is read in. Return a 00850 pointer to the allocated data, or NULL if there was a memory allocation 00851 failure. *len is the number of bytes of data read from the input file (even 00852 if load() returns NULL). If the input file was empty or could not be opened 00853 or read, *len is zero. */ 00854 local void *load(char *name, size_t *len) 00855 { 00856 size_t size; 00857 void *buf, *swap; 00858 FILE *in; 00859 00860 *len = 0; 00861 buf = malloc(size = 4096); 00862 if (buf == NULL) 00863 return NULL; 00864 in = name == NULL ? stdin : fopen(name, "rb"); 00865 if (in != NULL) { 00866 for (;;) { 00867 *len += fread((char *)buf + *len, 1, size - *len, in); 00868 if (*len < size) break; 00869 size = bythirds(size); 00870 if (size == *len || (swap = realloc(buf, size)) == NULL) { 00871 free(buf); 00872 buf = NULL; 00873 break; 00874 } 00875 buf = swap; 00876 } 00877 fclose(in); 00878 } 00879 return buf; 00880 } 00881 00882 int main(int argc, char **argv) 00883 { 00884 int ret, put = 0; 00885 unsigned skip = 0; 00886 char *arg, *name = NULL; 00887 unsigned char *source = NULL, *dest; 00888 size_t len = 0; 00889 unsigned long sourcelen, destlen; 00890 00891 /* process arguments */ 00892 while (arg = *++argv, --argc) 00893 if (arg[0] == '-') { 00894 if (arg[1] == 'w' && arg[2] == 0) 00895 put = 1; 00896 else if (arg[1] >= '0' && arg[1] <= '9') 00897 skip = (unsigned)atoi(arg + 1); 00898 else { 00899 fprintf(stderr, "invalid option %s\n", arg); 00900 return 3; 00901 } 00902 } 00903 else if (name != NULL) { 00904 fprintf(stderr, "only one file name allowed\n"); 00905 return 3; 00906 } 00907 else 00908 name = arg; 00909 source = load(name, &len); 00910 if (source == NULL) { 00911 fprintf(stderr, "memory allocation failure\n"); 00912 return 4; 00913 } 00914 if (len == 0) { 00915 fprintf(stderr, "could not read %s, or it was empty\n", 00916 name == NULL ? "<stdin>" : name); 00917 free(source); 00918 return 3; 00919 } 00920 if (skip >= len) { 00921 fprintf(stderr, "skip request of %d leaves no input\n", skip); 00922 free(source); 00923 return 3; 00924 } 00925 00926 /* test inflate data with offset skip */ 00927 len -= skip; 00928 sourcelen = (unsigned long)len; 00929 ret = puff(NIL, &destlen, source + skip, &sourcelen); 00930 if (ret) 00931 fprintf(stderr, "puff() failed with return code %d\n", ret); 00932 else { 00933 fprintf(stderr, "puff() succeeded uncompressing %lu bytes\n", destlen); 00934 if (sourcelen < len) fprintf(stderr, "%lu compressed bytes unused\n", 00935 len - sourcelen); 00936 } 00937 00938 /* if requested, inflate again and write decompressd data to stdout */ 00939 if (put) { 00940 dest = malloc(destlen); 00941 if (dest == NULL) { 00942 fprintf(stderr, "memory allocation failure\n"); 00943 free(source); 00944 return 4; 00945 } 00946 puff(dest, &destlen, source + skip, &sourcelen); 00947 fwrite(dest, 1, destlen, stdout); 00948 free(dest); 00949 } 00950 00951 /* clean up */ 00952 free(source); 00953 return ret; 00954 } 00955 #endif