1 /*
2 * puff.c
3 * Copyright (C) 2002-2010 Mark Adler
4 * For conditions of distribution and use, see copyright notice in puff.h
5 * version 2.2, 25 Apr 2010
6 *
7 * puff.c is a simple inflate written to be an unambiguous way to specify the
8 * deflate format. It is not written for speed but rather simplicity. As a
9 * side benefit, this code might actually be useful when small code is more
10 * important than speed, such as bootstrap applications. For typical deflate
11 * data, zlib's inflate() is about four times as fast as puff(). zlib's
12 * inflate compiles to around 20K on my machine, whereas puff.c compiles to
13 * around 4K on my machine (a PowerPC using GNU cc). If the faster decode()
14 * function here is used, then puff() is only twice as slow as zlib's
15 * inflate().
16 *
17 * All dynamically allocated memory comes from the stack. The stack required
18 * is less than 2K bytes. This code is compatible with 16-bit int's and
19 * assumes that long's are at least 32 bits. puff.c uses the short data type,
20 * assumed to be 16 bits, for arrays in order to to conserve memory. The code
21 * works whether integers are stored big endian or little endian.
22 *
23 * In the comments below are "Format notes" that describe the inflate process
24 * and document some of the less obvious aspects of the format. This source
25 * code is meant to supplement RFC 1951, which formally describes the deflate
26 * format:
27 *
28 * http://www.zlib.org/rfc-deflate.html
29 */
30
31 /*
32 * Change history:
33 *
34 * 1.0 10 Feb 2002 - First version
35 * 1.1 17 Feb 2002 - Clarifications of some comments and notes
36 * - Update puff() dest and source pointers on negative
37 * errors to facilitate debugging deflators
38 * - Remove longest from struct huffman -- not needed
39 * - Simplify offs[] index in construct()
40 * - Add input size and checking, using longjmp() to
41 * maintain easy readability
42 * - Use short data type for large arrays
43 * - Use pointers instead of long to specify source and
44 * destination sizes to avoid arbitrary 4 GB limits
45 * 1.2 17 Mar 2002 - Add faster version of decode(), doubles speed (!),
46 * but leave simple version for readabilty
47 * - Make sure invalid distances detected if pointers
48 * are 16 bits
49 * - Fix fixed codes table error
50 * - Provide a scanning mode for determining size of
51 * uncompressed data
52 * 1.3 20 Mar 2002 - Go back to lengths for puff() parameters [Gailly]
53 * - Add a puff.h file for the interface
54 * - Add braces in puff() for else do [Gailly]
55 * - Use indexes instead of pointers for readability
56 * 1.4 31 Mar 2002 - Simplify construct() code set check
57 * - Fix some comments
58 * - Add FIXLCODES #define
59 * 1.5 6 Apr 2002 - Minor comment fixes
60 * 1.6 7 Aug 2002 - Minor format changes
61 * 1.7 3 Mar 2003 - Added test code for distribution
62 * - Added zlib-like license
63 * 1.8 9 Jan 2004 - Added some comments on no distance codes case
64 * 1.9 21 Feb 2008 - Fix bug on 16-bit integer architectures [Pohland]
65 * - Catch missing end-of-block symbol error
66 * 2.0 25 Jul 2008 - Add #define to permit distance too far back
67 * - Add option in TEST code for puff to write the data
68 * - Add option in TEST code to skip input bytes
69 * - Allow TEST code to read from piped stdin
70 * 2.1 4 Apr 2010 - Avoid variable initialization for happier compilers
71 * - Avoid unsigned comparisons for even happier compilers
72 * 2.2 25 Apr 2010 - Fix bug in variable initializations [Oberhumer]
73 * - Add const where appropriate [Oberhumer]
74 * - Split if's and ?'s for coverage testing
75 * - Break out test code to separate file
76 * - Move NIL to puff.h
77 * - Allow incomplete code only if single code length is 1
78 * - Add full code coverage test to Makefile
79 */
80
81 #include <setjmp.h> /* for setjmp(), longjmp(), and jmp_buf */
82 #include "puff.h" /* prototype for puff() */
83
84 #define local static /* for local function definitions */
85
86 /*
87 * Maximums for allocations and loops. It is not useful to change these --
88 * they are fixed by the deflate format.
89 */
90 #define MAXBITS 15 /* maximum bits in a code */
91 #define MAXLCODES 286 /* maximum number of literal/length codes */
92 #define MAXDCODES 30 /* maximum number of distance codes */
93 #define MAXCODES (MAXLCODES+MAXDCODES) /* maximum codes lengths to read */
94 #define FIXLCODES 288 /* number of fixed literal/length codes */
95
96 /* input and output state */
97 struct state {
98 /* output state */
99 unsigned char *out; /* output buffer */
100 unsigned long outlen; /* available space at out */
101 unsigned long outcnt; /* bytes written to out so far */
102
103 /* input state */
104 const unsigned char *in; /* input buffer */
105 unsigned long inlen; /* available input at in */
106 unsigned long incnt; /* bytes read so far */
107 int bitbuf; /* bit buffer */
108 int bitcnt; /* number of bits in bit buffer */
109
110 /* input limit error return state for bits() and decode() */
111 jmp_buf env;
112 };
113
114 /*
115 * Return need bits from the input stream. This always leaves less than
116 * eight bits in the buffer. bits() works properly for need == 0.
117 *
118 * Format notes:
119 *
120 * - Bits are stored in bytes from the least significant bit to the most
121 * significant bit. Therefore bits are dropped from the bottom of the bit
122 * buffer, using shift right, and new bytes are appended to the top of the
123 * bit buffer, using shift left.
124 */
bits(struct state * s,int need)125 local int bits(struct state *s, int need)
126 {
127 long val; /* bit accumulator (can use up to 20 bits) */
128
129 /* load at least need bits into val */
130 val = s->bitbuf;
131 while (s->bitcnt < need) {
132 if (s->incnt == s->inlen)
133 longjmp(s->env, 1); /* out of input */
134 val |= (long)(s->in[s->incnt++]) << s->bitcnt; /* load eight bits */
135 s->bitcnt += 8;
136 }
137
138 /* drop need bits and update buffer, always zero to seven bits left */
139 s->bitbuf = (int)(val >> need);
140 s->bitcnt -= need;
141
142 /* return need bits, zeroing the bits above that */
143 return (int)(val & ((1L << need) - 1));
144 }
145
146 /*
147 * Process a stored block.
148 *
149 * Format notes:
150 *
151 * - After the two-bit stored block type (00), the stored block length and
152 * stored bytes are byte-aligned for fast copying. Therefore any leftover
153 * bits in the byte that has the last bit of the type, as many as seven, are
154 * discarded. The value of the discarded bits are not defined and should not
155 * be checked against any expectation.
156 *
157 * - The second inverted copy of the stored block length does not have to be
158 * checked, but it's probably a good idea to do so anyway.
159 *
160 * - A stored block can have zero length. This is sometimes used to byte-align
161 * subsets of the compressed data for random access or partial recovery.
162 */
stored(struct state * s)163 local int stored(struct state *s)
164 {
165 unsigned len; /* length of stored block */
166
167 /* discard leftover bits from current byte (assumes s->bitcnt < 8) */
168 s->bitbuf = 0;
169 s->bitcnt = 0;
170
171 /* get length and check against its one's complement */
172 if (s->incnt + 4 > s->inlen)
173 return 2; /* not enough input */
174 len = s->in[s->incnt++];
175 len |= s->in[s->incnt++] << 8;
176 if (s->in[s->incnt++] != (~len & 0xff) ||
177 s->in[s->incnt++] != ((~len >> 8) & 0xff))
178 return -2; /* didn't match complement! */
179
180 /* copy len bytes from in to out */
181 if (s->incnt + len > s->inlen)
182 return 2; /* not enough input */
183 if (s->out != NIL) {
184 if (s->outcnt + len > s->outlen)
185 return 1; /* not enough output space */
186 while (len--)
187 s->out[s->outcnt++] = s->in[s->incnt++];
188 }
189 else { /* just scanning */
190 s->outcnt += len;
191 s->incnt += len;
192 }
193
194 /* done with a valid stored block */
195 return 0;
196 }
197
198 /*
199 * Huffman code decoding tables. count[1..MAXBITS] is the number of symbols of
200 * each length, which for a canonical code are stepped through in order.
201 * symbol[] are the symbol values in canonical order, where the number of
202 * entries is the sum of the counts in count[]. The decoding process can be
203 * seen in the function decode() below.
204 */
205 struct huffman {
206 short *count; /* number of symbols of each length */
207 short *symbol; /* canonically ordered symbols */
208 };
209
210 /*
211 * Decode a code from the stream s using huffman table h. Return the symbol or
212 * a negative value if there is an error. If all of the lengths are zero, i.e.
213 * an empty code, or if the code is incomplete and an invalid code is received,
214 * then -10 is returned after reading MAXBITS bits.
215 *
216 * Format notes:
217 *
218 * - The codes as stored in the compressed data are bit-reversed relative to
219 * a simple integer ordering of codes of the same lengths. Hence below the
220 * bits are pulled from the compressed data one at a time and used to
221 * build the code value reversed from what is in the stream in order to
222 * permit simple integer comparisons for decoding. A table-based decoding
223 * scheme (as used in zlib) does not need to do this reversal.
224 *
225 * - The first code for the shortest length is all zeros. Subsequent codes of
226 * the same length are simply integer increments of the previous code. When
227 * moving up a length, a zero bit is appended to the code. For a complete
228 * code, the last code of the longest length will be all ones.
229 *
230 * - Incomplete codes are handled by this decoder, since they are permitted
231 * in the deflate format. See the format notes for fixed() and dynamic().
232 */
233 #ifdef SLOW
decode(struct state * s,const struct huffman * h)234 local int decode(struct state *s, const struct huffman *h)
235 {
236 int len; /* current number of bits in code */
237 int code; /* len bits being decoded */
238 int first; /* first code of length len */
239 int count; /* number of codes of length len */
240 int index; /* index of first code of length len in symbol table */
241
242 code = first = index = 0;
243 for (len = 1; len <= MAXBITS; len++) {
244 code |= bits(s, 1); /* get next bit */
245 count = h->count[len];
246 if (code - count < first) /* if length len, return symbol */
247 return h->symbol[index + (code - first)];
248 index += count; /* else update for next length */
249 first += count;
250 first <<= 1;
251 code <<= 1;
252 }
253 return -10; /* ran out of codes */
254 }
255
256 /*
257 * A faster version of decode() for real applications of this code. It's not
258 * as readable, but it makes puff() twice as fast. And it only makes the code
259 * a few percent larger.
260 */
261 #else /* !SLOW */
decode(struct state * s,const struct huffman * h)262 local int decode(struct state *s, const struct huffman *h)
263 {
264 int len; /* current number of bits in code */
265 int code; /* len bits being decoded */
266 int first; /* first code of length len */
267 int count; /* number of codes of length len */
268 int index; /* index of first code of length len in symbol table */
269 int bitbuf; /* bits from stream */
270 int left; /* bits left in next or left to process */
271 short *next; /* next number of codes */
272
273 bitbuf = s->bitbuf;
274 left = s->bitcnt;
275 code = first = index = 0;
276 len = 1;
277 next = h->count + 1;
278 while (1) {
279 while (left--) {
280 code |= bitbuf & 1;
281 bitbuf >>= 1;
282 count = *next++;
283 if (code - count < first) { /* if length len, return symbol */
284 s->bitbuf = bitbuf;
285 s->bitcnt = (s->bitcnt - len) & 7;
286 return h->symbol[index + (code - first)];
287 }
288 index += count; /* else update for next length */
289 first += count;
290 first <<= 1;
291 code <<= 1;
292 len++;
293 }
294 left = (MAXBITS+1) - len;
295 if (left == 0)
296 break;
297 if (s->incnt == s->inlen)
298 longjmp(s->env, 1); /* out of input */
299 bitbuf = s->in[s->incnt++];
300 if (left > 8)
301 left = 8;
302 }
303 return -10; /* ran out of codes */
304 }
305 #endif /* SLOW */
306
307 /*
308 * Given the list of code lengths length[0..n-1] representing a canonical
309 * Huffman code for n symbols, construct the tables required to decode those
310 * codes. Those tables are the number of codes of each length, and the symbols
311 * sorted by length, retaining their original order within each length. The
312 * return value is zero for a complete code set, negative for an over-
313 * subscribed code set, and positive for an incomplete code set. The tables
314 * can be used if the return value is zero or positive, but they cannot be used
315 * if the return value is negative. If the return value is zero, it is not
316 * possible for decode() using that table to return an error--any stream of
317 * enough bits will resolve to a symbol. If the return value is positive, then
318 * it is possible for decode() using that table to return an error for received
319 * codes past the end of the incomplete lengths.
320 *
321 * Not used by decode(), but used for error checking, h->count[0] is the number
322 * of the n symbols not in the code. So n - h->count[0] is the number of
323 * codes. This is useful for checking for incomplete codes that have more than
324 * one symbol, which is an error in a dynamic block.
325 *
326 * Assumption: for all i in 0..n-1, 0 <= length[i] <= MAXBITS
327 * This is assured by the construction of the length arrays in dynamic() and
328 * fixed() and is not verified by construct().
329 *
330 * Format notes:
331 *
332 * - Permitted and expected examples of incomplete codes are one of the fixed
333 * codes and any code with a single symbol which in deflate is coded as one
334 * bit instead of zero bits. See the format notes for fixed() and dynamic().
335 *
336 * - Within a given code length, the symbols are kept in ascending order for
337 * the code bits definition.
338 */
construct(struct huffman * h,const short * length,int n)339 local int construct(struct huffman *h, const short *length, int n)
340 {
341 int symbol; /* current symbol when stepping through length[] */
342 int len; /* current length when stepping through h->count[] */
343 int left; /* number of possible codes left of current length */
344 short offs[MAXBITS+1]; /* offsets in symbol table for each length */
345
346 /* count number of codes of each length */
347 for (len = 0; len <= MAXBITS; len++)
348 h->count[len] = 0;
349 for (symbol = 0; symbol < n; symbol++)
350 (h->count[length[symbol]])++; /* assumes lengths are within bounds */
351 if (h->count[0] == n) /* no codes! */
352 return 0; /* complete, but decode() will fail */
353
354 /* check for an over-subscribed or incomplete set of lengths */
355 left = 1; /* one possible code of zero length */
356 for (len = 1; len <= MAXBITS; len++) {
357 left <<= 1; /* one more bit, double codes left */
358 left -= h->count[len]; /* deduct count from possible codes */
359 if (left < 0)
360 return left; /* over-subscribed--return negative */
361 } /* left > 0 means incomplete */
362
363 /* generate offsets into symbol table for each length for sorting */
364 offs[1] = 0;
365 for (len = 1; len < MAXBITS; len++)
366 offs[len + 1] = offs[len] + h->count[len];
367
368 /*
369 * put symbols in table sorted by length, by symbol order within each
370 * length
371 */
372 for (symbol = 0; symbol < n; symbol++)
373 if (length[symbol] != 0)
374 h->symbol[offs[length[symbol]]++] = symbol;
375
376 /* return zero for complete set, positive for incomplete set */
377 return left;
378 }
379
380 /*
381 * Decode literal/length and distance codes until an end-of-block code.
382 *
383 * Format notes:
384 *
385 * - Compressed data that is after the block type if fixed or after the code
386 * description if dynamic is a combination of literals and length/distance
387 * pairs terminated by and end-of-block code. Literals are simply Huffman
388 * coded bytes. A length/distance pair is a coded length followed by a
389 * coded distance to represent a string that occurs earlier in the
390 * uncompressed data that occurs again at the current location.
391 *
392 * - Literals, lengths, and the end-of-block code are combined into a single
393 * code of up to 286 symbols. They are 256 literals (0..255), 29 length
394 * symbols (257..285), and the end-of-block symbol (256).
395 *
396 * - There are 256 possible lengths (3..258), and so 29 symbols are not enough
397 * to represent all of those. Lengths 3..10 and 258 are in fact represented
398 * by just a length symbol. Lengths 11..257 are represented as a symbol and
399 * some number of extra bits that are added as an integer to the base length
400 * of the length symbol. The number of extra bits is determined by the base
401 * length symbol. These are in the static arrays below, lens[] for the base
402 * lengths and lext[] for the corresponding number of extra bits.
403 *
404 * - The reason that 258 gets its own symbol is that the longest length is used
405 * often in highly redundant files. Note that 258 can also be coded as the
406 * base value 227 plus the maximum extra value of 31. While a good deflate
407 * should never do this, it is not an error, and should be decoded properly.
408 *
409 * - If a length is decoded, including its extra bits if any, then it is
410 * followed a distance code. There are up to 30 distance symbols. Again
411 * there are many more possible distances (1..32768), so extra bits are added
412 * to a base value represented by the symbol. The distances 1..4 get their
413 * own symbol, but the rest require extra bits. The base distances and
414 * corresponding number of extra bits are below in the static arrays dist[]
415 * and dext[].
416 *
417 * - Literal bytes are simply written to the output. A length/distance pair is
418 * an instruction to copy previously uncompressed bytes to the output. The
419 * copy is from distance bytes back in the output stream, copying for length
420 * bytes.
421 *
422 * - Distances pointing before the beginning of the output data are not
423 * permitted.
424 *
425 * - Overlapped copies, where the length is greater than the distance, are
426 * allowed and common. For example, a distance of one and a length of 258
427 * simply copies the last byte 258 times. A distance of four and a length of
428 * twelve copies the last four bytes three times. A simple forward copy
429 * ignoring whether the length is greater than the distance or not implements
430 * this correctly. You should not use memcpy() since its behavior is not
431 * defined for overlapped arrays. You should not use memmove() or bcopy()
432 * since though their behavior -is- defined for overlapping arrays, it is
433 * defined to do the wrong thing in this case.
434 */
codes(struct state * s,const struct huffman * lencode,const struct huffman * distcode)435 local int codes(struct state *s,
436 const struct huffman *lencode,
437 const struct huffman *distcode)
438 {
439 int symbol; /* decoded symbol */
440 int len; /* length for copy */
441 unsigned dist; /* distance for copy */
442 static const short lens[29] = { /* Size base for length codes 257..285 */
443 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 15, 17, 19, 23, 27, 31,
444 35, 43, 51, 59, 67, 83, 99, 115, 131, 163, 195, 227, 258};
445 static const short lext[29] = { /* Extra bits for length codes 257..285 */
446 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2,
447 3, 3, 3, 3, 4, 4, 4, 4, 5, 5, 5, 5, 0};
448 static const short dists[30] = { /* Offset base for distance codes 0..29 */
449 1, 2, 3, 4, 5, 7, 9, 13, 17, 25, 33, 49, 65, 97, 129, 193,
450 257, 385, 513, 769, 1025, 1537, 2049, 3073, 4097, 6145,
451 8193, 12289, 16385, 24577};
452 static const short dext[30] = { /* Extra bits for distance codes 0..29 */
453 0, 0, 0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6,
454 7, 7, 8, 8, 9, 9, 10, 10, 11, 11,
455 12, 12, 13, 13};
456
457 /* decode literals and length/distance pairs */
458 do {
459 symbol = decode(s, lencode);
460 if (symbol < 0)
461 return symbol; /* invalid symbol */
462 if (symbol < 256) { /* literal: symbol is the byte */
463 /* write out the literal */
464 if (s->out != NIL) {
465 if (s->outcnt == s->outlen)
466 return 1;
467 s->out[s->outcnt] = symbol;
468 }
469 s->outcnt++;
470 }
471 else if (symbol > 256) { /* length */
472 /* get and compute length */
473 symbol -= 257;
474 if (symbol >= 29)
475 return -10; /* invalid fixed code */
476 len = lens[symbol] + bits(s, lext[symbol]);
477
478 /* get and check distance */
479 symbol = decode(s, distcode);
480 if (symbol < 0)
481 return symbol; /* invalid symbol */
482 dist = dists[symbol] + bits(s, dext[symbol]);
483 #ifndef INFLATE_ALLOW_INVALID_DISTANCE_TOOFAR_ARRR
484 if (dist > s->outcnt)
485 return -11; /* distance too far back */
486 #endif
487
488 /* copy length bytes from distance bytes back */
489 if (s->out != NIL) {
490 if (s->outcnt + len > s->outlen)
491 return 1;
492 while (len--) {
493 s->out[s->outcnt] =
494 #ifdef INFLATE_ALLOW_INVALID_DISTANCE_TOOFAR_ARRR
495 dist > s->outcnt ?
496 0 :
497 #endif
498 s->out[s->outcnt - dist];
499 s->outcnt++;
500 }
501 }
502 else
503 s->outcnt += len;
504 }
505 } while (symbol != 256); /* end of block symbol */
506
507 /* done with a valid fixed or dynamic block */
508 return 0;
509 }
510
511 /*
512 * Process a fixed codes block.
513 *
514 * Format notes:
515 *
516 * - This block type can be useful for compressing small amounts of data for
517 * which the size of the code descriptions in a dynamic block exceeds the
518 * benefit of custom codes for that block. For fixed codes, no bits are
519 * spent on code descriptions. Instead the code lengths for literal/length
520 * codes and distance codes are fixed. The specific lengths for each symbol
521 * can be seen in the "for" loops below.
522 *
523 * - The literal/length code is complete, but has two symbols that are invalid
524 * and should result in an error if received. This cannot be implemented
525 * simply as an incomplete code since those two symbols are in the "middle"
526 * of the code. They are eight bits long and the longest literal/length\
527 * code is nine bits. Therefore the code must be constructed with those
528 * symbols, and the invalid symbols must be detected after decoding.
529 *
530 * - The fixed distance codes also have two invalid symbols that should result
531 * in an error if received. Since all of the distance codes are the same
532 * length, this can be implemented as an incomplete code. Then the invalid
533 * codes are detected while decoding.
534 */
fixed(struct state * s)535 local int fixed(struct state *s)
536 {
537 static int virgin = 1;
538 static short lencnt[MAXBITS+1], lensym[FIXLCODES];
539 static short distcnt[MAXBITS+1], distsym[MAXDCODES];
540 static struct huffman lencode, distcode;
541
542 /* build fixed huffman tables if first call (may not be thread safe) */
543 if (virgin) {
544 int symbol;
545 short lengths[FIXLCODES];
546
547 /* construct lencode and distcode */
548 lencode.count = lencnt;
549 lencode.symbol = lensym;
550 distcode.count = distcnt;
551 distcode.symbol = distsym;
552
553 /* literal/length table */
554 for (symbol = 0; symbol < 144; symbol++)
555 lengths[symbol] = 8;
556 for (; symbol < 256; symbol++)
557 lengths[symbol] = 9;
558 for (; symbol < 280; symbol++)
559 lengths[symbol] = 7;
560 for (; symbol < FIXLCODES; symbol++)
561 lengths[symbol] = 8;
562 construct(&lencode, lengths, FIXLCODES);
563
564 /* distance table */
565 for (symbol = 0; symbol < MAXDCODES; symbol++)
566 lengths[symbol] = 5;
567 construct(&distcode, lengths, MAXDCODES);
568
569 /* do this just once */
570 virgin = 0;
571 }
572
573 /* decode data until end-of-block code */
574 return codes(s, &lencode, &distcode);
575 }
576
577 /*
578 * Process a dynamic codes block.
579 *
580 * Format notes:
581 *
582 * - A dynamic block starts with a description of the literal/length and
583 * distance codes for that block. New dynamic blocks allow the compressor to
584 * rapidly adapt to changing data with new codes optimized for that data.
585 *
586 * - The codes used by the deflate format are "canonical", which means that
587 * the actual bits of the codes are generated in an unambiguous way simply
588 * from the number of bits in each code. Therefore the code descriptions
589 * are simply a list of code lengths for each symbol.
590 *
591 * - The code lengths are stored in order for the symbols, so lengths are
592 * provided for each of the literal/length symbols, and for each of the
593 * distance symbols.
594 *
595 * - If a symbol is not used in the block, this is represented by a zero as
596 * as the code length. This does not mean a zero-length code, but rather
597 * that no code should be created for this symbol. There is no way in the
598 * deflate format to represent a zero-length code.
599 *
600 * - The maximum number of bits in a code is 15, so the possible lengths for
601 * any code are 1..15.
602 *
603 * - The fact that a length of zero is not permitted for a code has an
604 * interesting consequence. Normally if only one symbol is used for a given
605 * code, then in fact that code could be represented with zero bits. However
606 * in deflate, that code has to be at least one bit. So for example, if
607 * only a single distance base symbol appears in a block, then it will be
608 * represented by a single code of length one, in particular one 0 bit. This
609 * is an incomplete code, since if a 1 bit is received, it has no meaning,
610 * and should result in an error. So incomplete distance codes of one symbol
611 * should be permitted, and the receipt of invalid codes should be handled.
612 *
613 * - It is also possible to have a single literal/length code, but that code
614 * must be the end-of-block code, since every dynamic block has one. This
615 * is not the most efficient way to create an empty block (an empty fixed
616 * block is fewer bits), but it is allowed by the format. So incomplete
617 * literal/length codes of one symbol should also be permitted.
618 *
619 * - If there are only literal codes and no lengths, then there are no distance
620 * codes. This is represented by one distance code with zero bits.
621 *
622 * - The list of up to 286 length/literal lengths and up to 30 distance lengths
623 * are themselves compressed using Huffman codes and run-length encoding. In
624 * the list of code lengths, a 0 symbol means no code, a 1..15 symbol means
625 * that length, and the symbols 16, 17, and 18 are run-length instructions.
626 * Each of 16, 17, and 18 are follwed by extra bits to define the length of
627 * the run. 16 copies the last length 3 to 6 times. 17 represents 3 to 10
628 * zero lengths, and 18 represents 11 to 138 zero lengths. Unused symbols
629 * are common, hence the special coding for zero lengths.
630 *
631 * - The symbols for 0..18 are Huffman coded, and so that code must be
632 * described first. This is simply a sequence of up to 19 three-bit values
633 * representing no code (0) or the code length for that symbol (1..7).
634 *
635 * - A dynamic block starts with three fixed-size counts from which is computed
636 * the number of literal/length code lengths, the number of distance code
637 * lengths, and the number of code length code lengths (ok, you come up with
638 * a better name!) in the code descriptions. For the literal/length and
639 * distance codes, lengths after those provided are considered zero, i.e. no
640 * code. The code length code lengths are received in a permuted order (see
641 * the order[] array below) to make a short code length code length list more
642 * likely. As it turns out, very short and very long codes are less likely
643 * to be seen in a dynamic code description, hence what may appear initially
644 * to be a peculiar ordering.
645 *
646 * - Given the number of literal/length code lengths (nlen) and distance code
647 * lengths (ndist), then they are treated as one long list of nlen + ndist
648 * code lengths. Therefore run-length coding can and often does cross the
649 * boundary between the two sets of lengths.
650 *
651 * - So to summarize, the code description at the start of a dynamic block is
652 * three counts for the number of code lengths for the literal/length codes,
653 * the distance codes, and the code length codes. This is followed by the
654 * code length code lengths, three bits each. This is used to construct the
655 * code length code which is used to read the remainder of the lengths. Then
656 * the literal/length code lengths and distance lengths are read as a single
657 * set of lengths using the code length codes. Codes are constructed from
658 * the resulting two sets of lengths, and then finally you can start
659 * decoding actual compressed data in the block.
660 *
661 * - For reference, a "typical" size for the code description in a dynamic
662 * block is around 80 bytes.
663 */
dynamic(struct state * s)664 local int dynamic(struct state *s)
665 {
666 int nlen, ndist, ncode; /* number of lengths in descriptor */
667 int index; /* index of lengths[] */
668 int err; /* construct() return value */
669 short lengths[MAXCODES]; /* descriptor code lengths */
670 short lencnt[MAXBITS+1], lensym[MAXLCODES]; /* lencode memory */
671 short distcnt[MAXBITS+1], distsym[MAXDCODES]; /* distcode memory */
672 struct huffman lencode, distcode; /* length and distance codes */
673 static const short order[19] = /* permutation of code length codes */
674 {16, 17, 18, 0, 8, 7, 9, 6, 10, 5, 11, 4, 12, 3, 13, 2, 14, 1, 15};
675
676 /* construct lencode and distcode */
677 lencode.count = lencnt;
678 lencode.symbol = lensym;
679 distcode.count = distcnt;
680 distcode.symbol = distsym;
681
682 /* get number of lengths in each table, check lengths */
683 nlen = bits(s, 5) + 257;
684 ndist = bits(s, 5) + 1;
685 ncode = bits(s, 4) + 4;
686 if (nlen > MAXLCODES || ndist > MAXDCODES)
687 return -3; /* bad counts */
688
689 /* read code length code lengths (really), missing lengths are zero */
690 for (index = 0; index < ncode; index++)
691 lengths[order[index]] = bits(s, 3);
692 for (; index < 19; index++)
693 lengths[order[index]] = 0;
694
695 /* build huffman table for code lengths codes (use lencode temporarily) */
696 err = construct(&lencode, lengths, 19);
697 if (err != 0) /* require complete code set here */
698 return -4;
699
700 /* read length/literal and distance code length tables */
701 index = 0;
702 while (index < nlen + ndist) {
703 int symbol; /* decoded value */
704 int len; /* last length to repeat */
705
706 symbol = decode(s, &lencode);
707 if (symbol < 16) /* length in 0..15 */
708 lengths[index++] = symbol;
709 else { /* repeat instruction */
710 len = 0; /* assume repeating zeros */
711 if (symbol == 16) { /* repeat last length 3..6 times */
712 if (index == 0)
713 return -5; /* no last length! */
714 len = lengths[index - 1]; /* last length */
715 symbol = 3 + bits(s, 2);
716 }
717 else if (symbol == 17) /* repeat zero 3..10 times */
718 symbol = 3 + bits(s, 3);
719 else /* == 18, repeat zero 11..138 times */
720 symbol = 11 + bits(s, 7);
721 if (index + symbol > nlen + ndist)
722 return -6; /* too many lengths! */
723 while (symbol--) /* repeat last or zero symbol times */
724 lengths[index++] = len;
725 }
726 }
727
728 /* check for end-of-block code -- there better be one! */
729 if (lengths[256] == 0)
730 return -9;
731
732 /* build huffman table for literal/length codes */
733 err = construct(&lencode, lengths, nlen);
734 if (err && (err < 0 || nlen != lencode.count[0] + lencode.count[1]))
735 return -7; /* incomplete code ok only for single length 1 code */
736
737 /* build huffman table for distance codes */
738 err = construct(&distcode, lengths + nlen, ndist);
739 if (err && (err < 0 || ndist != distcode.count[0] + distcode.count[1]))
740 return -8; /* incomplete code ok only for single length 1 code */
741
742 /* decode data until end-of-block code */
743 return codes(s, &lencode, &distcode);
744 }
745
746 /*
747 * Inflate source to dest. On return, destlen and sourcelen are updated to the
748 * size of the uncompressed data and the size of the deflate data respectively.
749 * On success, the return value of puff() is zero. If there is an error in the
750 * source data, i.e. it is not in the deflate format, then a negative value is
751 * returned. If there is not enough input available or there is not enough
752 * output space, then a positive error is returned. In that case, destlen and
753 * sourcelen are not updated to facilitate retrying from the beginning with the
754 * provision of more input data or more output space. In the case of invalid
755 * inflate data (a negative error), the dest and source pointers are updated to
756 * facilitate the debugging of deflators.
757 *
758 * puff() also has a mode to determine the size of the uncompressed output with
759 * no output written. For this dest must be (unsigned char *)0. In this case,
760 * the input value of *destlen is ignored, and on return *destlen is set to the
761 * size of the uncompressed output.
762 *
763 * The return codes are:
764 *
765 * 2: available inflate data did not terminate
766 * 1: output space exhausted before completing inflate
767 * 0: successful inflate
768 * -1: invalid block type (type == 3)
769 * -2: stored block length did not match one's complement
770 * -3: dynamic block code description: too many length or distance codes
771 * -4: dynamic block code description: code lengths codes incomplete
772 * -5: dynamic block code description: repeat lengths with no first length
773 * -6: dynamic block code description: repeat more than specified lengths
774 * -7: dynamic block code description: invalid literal/length code lengths
775 * -8: dynamic block code description: invalid distance code lengths
776 * -9: dynamic block code description: missing end-of-block code
777 * -10: invalid literal/length or distance code in fixed or dynamic block
778 * -11: distance is too far back in fixed or dynamic block
779 *
780 * Format notes:
781 *
782 * - Three bits are read for each block to determine the kind of block and
783 * whether or not it is the last block. Then the block is decoded and the
784 * process repeated if it was not the last block.
785 *
786 * - The leftover bits in the last byte of the deflate data after the last
787 * block (if it was a fixed or dynamic block) are undefined and have no
788 * expected values to check.
789 */
puff(unsigned char * dest,unsigned long * destlen,const unsigned char * source,unsigned long * sourcelen)790 int puff(unsigned char *dest, /* pointer to destination pointer */
791 unsigned long *destlen, /* amount of output space */
792 const unsigned char *source, /* pointer to source data pointer */
793 unsigned long *sourcelen) /* amount of input available */
794 {
795 struct state s; /* input/output state */
796 int last, type; /* block information */
797 int err; /* return value */
798
799 /* initialize output state */
800 s.out = dest;
801 s.outlen = *destlen; /* ignored if dest is NIL */
802 s.outcnt = 0;
803
804 /* initialize input state */
805 s.in = source;
806 s.inlen = *sourcelen;
807 s.incnt = 0;
808 s.bitbuf = 0;
809 s.bitcnt = 0;
810
811 /* return if bits() or decode() tries to read past available input */
812 if (setjmp(s.env) != 0) /* if came back here via longjmp() */
813 err = 2; /* then skip do-loop, return error */
814 else {
815 /* process blocks until last block or error */
816 do {
817 last = bits(&s, 1); /* one if last block */
818 type = bits(&s, 2); /* block type 0..3 */
819 err = type == 0 ?
820 stored(&s) :
821 (type == 1 ?
822 fixed(&s) :
823 (type == 2 ?
824 dynamic(&s) :
825 -1)); /* type == 3, invalid */
826 if (err != 0)
827 break; /* return with error */
828 } while (!last);
829 }
830
831 /* update the lengths and return */
832 if (err <= 0) {
833 *destlen = s.outcnt;
834 *sourcelen = s.incnt;
835 }
836 return err;
837 }
838