Chris@4: /* Chris@4: * puff.c Chris@4: * Copyright (C) 2002-2010 Mark Adler Chris@4: * For conditions of distribution and use, see copyright notice in puff.h Chris@4: * version 2.2, 25 Apr 2010 Chris@4: * Chris@4: * puff.c is a simple inflate written to be an unambiguous way to specify the Chris@4: * deflate format. It is not written for speed but rather simplicity. As a Chris@4: * side benefit, this code might actually be useful when small code is more Chris@4: * important than speed, such as bootstrap applications. For typical deflate Chris@4: * data, zlib's inflate() is about four times as fast as puff(). zlib's Chris@4: * inflate compiles to around 20K on my machine, whereas puff.c compiles to Chris@4: * around 4K on my machine (a PowerPC using GNU cc). If the faster decode() Chris@4: * function here is used, then puff() is only twice as slow as zlib's Chris@4: * inflate(). Chris@4: * Chris@4: * All dynamically allocated memory comes from the stack. The stack required Chris@4: * is less than 2K bytes. This code is compatible with 16-bit int's and Chris@4: * assumes that long's are at least 32 bits. puff.c uses the short data type, Chris@4: * assumed to be 16 bits, for arrays in order to to conserve memory. The code Chris@4: * works whether integers are stored big endian or little endian. Chris@4: * Chris@4: * In the comments below are "Format notes" that describe the inflate process Chris@4: * and document some of the less obvious aspects of the format. This source Chris@4: * code is meant to supplement RFC 1951, which formally describes the deflate Chris@4: * format: Chris@4: * Chris@4: * http://www.zlib.org/rfc-deflate.html Chris@4: */ Chris@4: Chris@4: /* Chris@4: * Change history: Chris@4: * Chris@4: * 1.0 10 Feb 2002 - First version Chris@4: * 1.1 17 Feb 2002 - Clarifications of some comments and notes Chris@4: * - Update puff() dest and source pointers on negative Chris@4: * errors to facilitate debugging deflators Chris@4: * - Remove longest from struct huffman -- not needed Chris@4: * - Simplify offs[] index in construct() Chris@4: * - Add input size and checking, using longjmp() to Chris@4: * maintain easy readability Chris@4: * - Use short data type for large arrays Chris@4: * - Use pointers instead of long to specify source and Chris@4: * destination sizes to avoid arbitrary 4 GB limits Chris@4: * 1.2 17 Mar 2002 - Add faster version of decode(), doubles speed (!), Chris@4: * but leave simple version for readabilty Chris@4: * - Make sure invalid distances detected if pointers Chris@4: * are 16 bits Chris@4: * - Fix fixed codes table error Chris@4: * - Provide a scanning mode for determining size of Chris@4: * uncompressed data Chris@4: * 1.3 20 Mar 2002 - Go back to lengths for puff() parameters [Gailly] Chris@4: * - Add a puff.h file for the interface Chris@4: * - Add braces in puff() for else do [Gailly] Chris@4: * - Use indexes instead of pointers for readability Chris@4: * 1.4 31 Mar 2002 - Simplify construct() code set check Chris@4: * - Fix some comments Chris@4: * - Add FIXLCODES #define Chris@4: * 1.5 6 Apr 2002 - Minor comment fixes Chris@4: * 1.6 7 Aug 2002 - Minor format changes Chris@4: * 1.7 3 Mar 2003 - Added test code for distribution Chris@4: * - Added zlib-like license Chris@4: * 1.8 9 Jan 2004 - Added some comments on no distance codes case Chris@4: * 1.9 21 Feb 2008 - Fix bug on 16-bit integer architectures [Pohland] Chris@4: * - Catch missing end-of-block symbol error Chris@4: * 2.0 25 Jul 2008 - Add #define to permit distance too far back Chris@4: * - Add option in TEST code for puff to write the data Chris@4: * - Add option in TEST code to skip input bytes Chris@4: * - Allow TEST code to read from piped stdin Chris@4: * 2.1 4 Apr 2010 - Avoid variable initialization for happier compilers Chris@4: * - Avoid unsigned comparisons for even happier compilers Chris@4: * 2.2 25 Apr 2010 - Fix bug in variable initializations [Oberhumer] Chris@4: * - Add const where appropriate [Oberhumer] Chris@4: * - Split if's and ?'s for coverage testing Chris@4: * - Break out test code to separate file Chris@4: * - Move NIL to puff.h Chris@4: * - Allow incomplete code only if single code length is 1 Chris@4: * - Add full code coverage test to Makefile Chris@4: */ Chris@4: Chris@4: #include /* for setjmp(), longjmp(), and jmp_buf */ Chris@4: #include "puff.h" /* prototype for puff() */ Chris@4: Chris@4: #define local static /* for local function definitions */ Chris@4: Chris@4: /* Chris@4: * Maximums for allocations and loops. It is not useful to change these -- Chris@4: * they are fixed by the deflate format. Chris@4: */ Chris@4: #define MAXBITS 15 /* maximum bits in a code */ Chris@4: #define MAXLCODES 286 /* maximum number of literal/length codes */ Chris@4: #define MAXDCODES 30 /* maximum number of distance codes */ Chris@4: #define MAXCODES (MAXLCODES+MAXDCODES) /* maximum codes lengths to read */ Chris@4: #define FIXLCODES 288 /* number of fixed literal/length codes */ Chris@4: Chris@4: /* input and output state */ Chris@4: struct state { Chris@4: /* output state */ Chris@4: unsigned char *out; /* output buffer */ Chris@4: unsigned long outlen; /* available space at out */ Chris@4: unsigned long outcnt; /* bytes written to out so far */ Chris@4: Chris@4: /* input state */ Chris@4: const unsigned char *in; /* input buffer */ Chris@4: unsigned long inlen; /* available input at in */ Chris@4: unsigned long incnt; /* bytes read so far */ Chris@4: int bitbuf; /* bit buffer */ Chris@4: int bitcnt; /* number of bits in bit buffer */ Chris@4: Chris@4: /* input limit error return state for bits() and decode() */ Chris@4: jmp_buf env; Chris@4: }; Chris@4: Chris@4: /* Chris@4: * Return need bits from the input stream. This always leaves less than Chris@4: * eight bits in the buffer. bits() works properly for need == 0. Chris@4: * Chris@4: * Format notes: Chris@4: * Chris@4: * - Bits are stored in bytes from the least significant bit to the most Chris@4: * significant bit. Therefore bits are dropped from the bottom of the bit Chris@4: * buffer, using shift right, and new bytes are appended to the top of the Chris@4: * bit buffer, using shift left. Chris@4: */ Chris@4: local int bits(struct state *s, int need) Chris@4: { Chris@4: long val; /* bit accumulator (can use up to 20 bits) */ Chris@4: Chris@4: /* load at least need bits into val */ Chris@4: val = s->bitbuf; Chris@4: while (s->bitcnt < need) { Chris@4: if (s->incnt == s->inlen) Chris@4: longjmp(s->env, 1); /* out of input */ Chris@4: val |= (long)(s->in[s->incnt++]) << s->bitcnt; /* load eight bits */ Chris@4: s->bitcnt += 8; Chris@4: } Chris@4: Chris@4: /* drop need bits and update buffer, always zero to seven bits left */ Chris@4: s->bitbuf = (int)(val >> need); Chris@4: s->bitcnt -= need; Chris@4: Chris@4: /* return need bits, zeroing the bits above that */ Chris@4: return (int)(val & ((1L << need) - 1)); Chris@4: } Chris@4: Chris@4: /* Chris@4: * Process a stored block. Chris@4: * Chris@4: * Format notes: Chris@4: * Chris@4: * - After the two-bit stored block type (00), the stored block length and Chris@4: * stored bytes are byte-aligned for fast copying. Therefore any leftover Chris@4: * bits in the byte that has the last bit of the type, as many as seven, are Chris@4: * discarded. The value of the discarded bits are not defined and should not Chris@4: * be checked against any expectation. Chris@4: * Chris@4: * - The second inverted copy of the stored block length does not have to be Chris@4: * checked, but it's probably a good idea to do so anyway. Chris@4: * Chris@4: * - A stored block can have zero length. This is sometimes used to byte-align Chris@4: * subsets of the compressed data for random access or partial recovery. Chris@4: */ Chris@4: local int stored(struct state *s) Chris@4: { Chris@4: unsigned len; /* length of stored block */ Chris@4: Chris@4: /* discard leftover bits from current byte (assumes s->bitcnt < 8) */ Chris@4: s->bitbuf = 0; Chris@4: s->bitcnt = 0; Chris@4: Chris@4: /* get length and check against its one's complement */ Chris@4: if (s->incnt + 4 > s->inlen) Chris@4: return 2; /* not enough input */ Chris@4: len = s->in[s->incnt++]; Chris@4: len |= s->in[s->incnt++] << 8; Chris@4: if (s->in[s->incnt++] != (~len & 0xff) || Chris@4: s->in[s->incnt++] != ((~len >> 8) & 0xff)) Chris@4: return -2; /* didn't match complement! */ Chris@4: Chris@4: /* copy len bytes from in to out */ Chris@4: if (s->incnt + len > s->inlen) Chris@4: return 2; /* not enough input */ Chris@4: if (s->out != NIL) { Chris@4: if (s->outcnt + len > s->outlen) Chris@4: return 1; /* not enough output space */ Chris@4: while (len--) Chris@4: s->out[s->outcnt++] = s->in[s->incnt++]; Chris@4: } Chris@4: else { /* just scanning */ Chris@4: s->outcnt += len; Chris@4: s->incnt += len; Chris@4: } Chris@4: Chris@4: /* done with a valid stored block */ Chris@4: return 0; Chris@4: } Chris@4: Chris@4: /* Chris@4: * Huffman code decoding tables. count[1..MAXBITS] is the number of symbols of Chris@4: * each length, which for a canonical code are stepped through in order. Chris@4: * symbol[] are the symbol values in canonical order, where the number of Chris@4: * entries is the sum of the counts in count[]. The decoding process can be Chris@4: * seen in the function decode() below. Chris@4: */ Chris@4: struct huffman { Chris@4: short *count; /* number of symbols of each length */ Chris@4: short *symbol; /* canonically ordered symbols */ Chris@4: }; Chris@4: Chris@4: /* Chris@4: * Decode a code from the stream s using huffman table h. Return the symbol or Chris@4: * a negative value if there is an error. If all of the lengths are zero, i.e. Chris@4: * an empty code, or if the code is incomplete and an invalid code is received, Chris@4: * then -10 is returned after reading MAXBITS bits. Chris@4: * Chris@4: * Format notes: Chris@4: * Chris@4: * - The codes as stored in the compressed data are bit-reversed relative to Chris@4: * a simple integer ordering of codes of the same lengths. Hence below the Chris@4: * bits are pulled from the compressed data one at a time and used to Chris@4: * build the code value reversed from what is in the stream in order to Chris@4: * permit simple integer comparisons for decoding. A table-based decoding Chris@4: * scheme (as used in zlib) does not need to do this reversal. Chris@4: * Chris@4: * - The first code for the shortest length is all zeros. Subsequent codes of Chris@4: * the same length are simply integer increments of the previous code. When Chris@4: * moving up a length, a zero bit is appended to the code. For a complete Chris@4: * code, the last code of the longest length will be all ones. Chris@4: * Chris@4: * - Incomplete codes are handled by this decoder, since they are permitted Chris@4: * in the deflate format. See the format notes for fixed() and dynamic(). Chris@4: */ Chris@4: #ifdef SLOW Chris@4: local int decode(struct state *s, const struct huffman *h) Chris@4: { Chris@4: int len; /* current number of bits in code */ Chris@4: int code; /* len bits being decoded */ Chris@4: int first; /* first code of length len */ Chris@4: int count; /* number of codes of length len */ Chris@4: int index; /* index of first code of length len in symbol table */ Chris@4: Chris@4: code = first = index = 0; Chris@4: for (len = 1; len <= MAXBITS; len++) { Chris@4: code |= bits(s, 1); /* get next bit */ Chris@4: count = h->count[len]; Chris@4: if (code - count < first) /* if length len, return symbol */ Chris@4: return h->symbol[index + (code - first)]; Chris@4: index += count; /* else update for next length */ Chris@4: first += count; Chris@4: first <<= 1; Chris@4: code <<= 1; Chris@4: } Chris@4: return -10; /* ran out of codes */ Chris@4: } Chris@4: Chris@4: /* Chris@4: * A faster version of decode() for real applications of this code. It's not Chris@4: * as readable, but it makes puff() twice as fast. And it only makes the code Chris@4: * a few percent larger. Chris@4: */ Chris@4: #else /* !SLOW */ Chris@4: local int decode(struct state *s, const struct huffman *h) Chris@4: { Chris@4: int len; /* current number of bits in code */ Chris@4: int code; /* len bits being decoded */ Chris@4: int first; /* first code of length len */ Chris@4: int count; /* number of codes of length len */ Chris@4: int index; /* index of first code of length len in symbol table */ Chris@4: int bitbuf; /* bits from stream */ Chris@4: int left; /* bits left in next or left to process */ Chris@4: short *next; /* next number of codes */ Chris@4: Chris@4: bitbuf = s->bitbuf; Chris@4: left = s->bitcnt; Chris@4: code = first = index = 0; Chris@4: len = 1; Chris@4: next = h->count + 1; Chris@4: while (1) { Chris@4: while (left--) { Chris@4: code |= bitbuf & 1; Chris@4: bitbuf >>= 1; Chris@4: count = *next++; Chris@4: if (code - count < first) { /* if length len, return symbol */ Chris@4: s->bitbuf = bitbuf; Chris@4: s->bitcnt = (s->bitcnt - len) & 7; Chris@4: return h->symbol[index + (code - first)]; Chris@4: } Chris@4: index += count; /* else update for next length */ Chris@4: first += count; Chris@4: first <<= 1; Chris@4: code <<= 1; Chris@4: len++; Chris@4: } Chris@4: left = (MAXBITS+1) - len; Chris@4: if (left == 0) Chris@4: break; Chris@4: if (s->incnt == s->inlen) Chris@4: longjmp(s->env, 1); /* out of input */ Chris@4: bitbuf = s->in[s->incnt++]; Chris@4: if (left > 8) Chris@4: left = 8; Chris@4: } Chris@4: return -10; /* ran out of codes */ Chris@4: } Chris@4: #endif /* SLOW */ Chris@4: Chris@4: /* Chris@4: * Given the list of code lengths length[0..n-1] representing a canonical Chris@4: * Huffman code for n symbols, construct the tables required to decode those Chris@4: * codes. Those tables are the number of codes of each length, and the symbols Chris@4: * sorted by length, retaining their original order within each length. The Chris@4: * return value is zero for a complete code set, negative for an over- Chris@4: * subscribed code set, and positive for an incomplete code set. The tables Chris@4: * can be used if the return value is zero or positive, but they cannot be used Chris@4: * if the return value is negative. If the return value is zero, it is not Chris@4: * possible for decode() using that table to return an error--any stream of Chris@4: * enough bits will resolve to a symbol. If the return value is positive, then Chris@4: * it is possible for decode() using that table to return an error for received Chris@4: * codes past the end of the incomplete lengths. Chris@4: * Chris@4: * Not used by decode(), but used for error checking, h->count[0] is the number Chris@4: * of the n symbols not in the code. So n - h->count[0] is the number of Chris@4: * codes. This is useful for checking for incomplete codes that have more than Chris@4: * one symbol, which is an error in a dynamic block. Chris@4: * Chris@4: * Assumption: for all i in 0..n-1, 0 <= length[i] <= MAXBITS Chris@4: * This is assured by the construction of the length arrays in dynamic() and Chris@4: * fixed() and is not verified by construct(). Chris@4: * Chris@4: * Format notes: Chris@4: * Chris@4: * - Permitted and expected examples of incomplete codes are one of the fixed Chris@4: * codes and any code with a single symbol which in deflate is coded as one Chris@4: * bit instead of zero bits. See the format notes for fixed() and dynamic(). Chris@4: * Chris@4: * - Within a given code length, the symbols are kept in ascending order for Chris@4: * the code bits definition. Chris@4: */ Chris@4: local int construct(struct huffman *h, const short *length, int n) Chris@4: { Chris@4: int symbol; /* current symbol when stepping through length[] */ Chris@4: int len; /* current length when stepping through h->count[] */ Chris@4: int left; /* number of possible codes left of current length */ Chris@4: short offs[MAXBITS+1]; /* offsets in symbol table for each length */ Chris@4: Chris@4: /* count number of codes of each length */ Chris@4: for (len = 0; len <= MAXBITS; len++) Chris@4: h->count[len] = 0; Chris@4: for (symbol = 0; symbol < n; symbol++) Chris@4: (h->count[length[symbol]])++; /* assumes lengths are within bounds */ Chris@4: if (h->count[0] == n) /* no codes! */ Chris@4: return 0; /* complete, but decode() will fail */ Chris@4: Chris@4: /* check for an over-subscribed or incomplete set of lengths */ Chris@4: left = 1; /* one possible code of zero length */ Chris@4: for (len = 1; len <= MAXBITS; len++) { Chris@4: left <<= 1; /* one more bit, double codes left */ Chris@4: left -= h->count[len]; /* deduct count from possible codes */ Chris@4: if (left < 0) Chris@4: return left; /* over-subscribed--return negative */ Chris@4: } /* left > 0 means incomplete */ Chris@4: Chris@4: /* generate offsets into symbol table for each length for sorting */ Chris@4: offs[1] = 0; Chris@4: for (len = 1; len < MAXBITS; len++) Chris@4: offs[len + 1] = offs[len] + h->count[len]; Chris@4: Chris@4: /* Chris@4: * put symbols in table sorted by length, by symbol order within each Chris@4: * length Chris@4: */ Chris@4: for (symbol = 0; symbol < n; symbol++) Chris@4: if (length[symbol] != 0) Chris@4: h->symbol[offs[length[symbol]]++] = symbol; Chris@4: Chris@4: /* return zero for complete set, positive for incomplete set */ Chris@4: return left; Chris@4: } Chris@4: Chris@4: /* Chris@4: * Decode literal/length and distance codes until an end-of-block code. Chris@4: * Chris@4: * Format notes: Chris@4: * Chris@4: * - Compressed data that is after the block type if fixed or after the code Chris@4: * description if dynamic is a combination of literals and length/distance Chris@4: * pairs terminated by and end-of-block code. Literals are simply Huffman Chris@4: * coded bytes. A length/distance pair is a coded length followed by a Chris@4: * coded distance to represent a string that occurs earlier in the Chris@4: * uncompressed data that occurs again at the current location. Chris@4: * Chris@4: * - Literals, lengths, and the end-of-block code are combined into a single Chris@4: * code of up to 286 symbols. They are 256 literals (0..255), 29 length Chris@4: * symbols (257..285), and the end-of-block symbol (256). Chris@4: * Chris@4: * - There are 256 possible lengths (3..258), and so 29 symbols are not enough Chris@4: * to represent all of those. Lengths 3..10 and 258 are in fact represented Chris@4: * by just a length symbol. Lengths 11..257 are represented as a symbol and Chris@4: * some number of extra bits that are added as an integer to the base length Chris@4: * of the length symbol. The number of extra bits is determined by the base Chris@4: * length symbol. These are in the static arrays below, lens[] for the base Chris@4: * lengths and lext[] for the corresponding number of extra bits. Chris@4: * Chris@4: * - The reason that 258 gets its own symbol is that the longest length is used Chris@4: * often in highly redundant files. Note that 258 can also be coded as the Chris@4: * base value 227 plus the maximum extra value of 31. While a good deflate Chris@4: * should never do this, it is not an error, and should be decoded properly. Chris@4: * Chris@4: * - If a length is decoded, including its extra bits if any, then it is Chris@4: * followed a distance code. There are up to 30 distance symbols. Again Chris@4: * there are many more possible distances (1..32768), so extra bits are added Chris@4: * to a base value represented by the symbol. The distances 1..4 get their Chris@4: * own symbol, but the rest require extra bits. The base distances and Chris@4: * corresponding number of extra bits are below in the static arrays dist[] Chris@4: * and dext[]. Chris@4: * Chris@4: * - Literal bytes are simply written to the output. A length/distance pair is Chris@4: * an instruction to copy previously uncompressed bytes to the output. The Chris@4: * copy is from distance bytes back in the output stream, copying for length Chris@4: * bytes. Chris@4: * Chris@4: * - Distances pointing before the beginning of the output data are not Chris@4: * permitted. Chris@4: * Chris@4: * - Overlapped copies, where the length is greater than the distance, are Chris@4: * allowed and common. For example, a distance of one and a length of 258 Chris@4: * simply copies the last byte 258 times. A distance of four and a length of Chris@4: * twelve copies the last four bytes three times. A simple forward copy Chris@4: * ignoring whether the length is greater than the distance or not implements Chris@4: * this correctly. You should not use memcpy() since its behavior is not Chris@4: * defined for overlapped arrays. You should not use memmove() or bcopy() Chris@4: * since though their behavior -is- defined for overlapping arrays, it is Chris@4: * defined to do the wrong thing in this case. Chris@4: */ Chris@4: local int codes(struct state *s, Chris@4: const struct huffman *lencode, Chris@4: const struct huffman *distcode) Chris@4: { Chris@4: int symbol; /* decoded symbol */ Chris@4: int len; /* length for copy */ Chris@4: unsigned dist; /* distance for copy */ Chris@4: static const short lens[29] = { /* Size base for length codes 257..285 */ Chris@4: 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 15, 17, 19, 23, 27, 31, Chris@4: 35, 43, 51, 59, 67, 83, 99, 115, 131, 163, 195, 227, 258}; Chris@4: static const short lext[29] = { /* Extra bits for length codes 257..285 */ Chris@4: 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2, Chris@4: 3, 3, 3, 3, 4, 4, 4, 4, 5, 5, 5, 5, 0}; Chris@4: static const short dists[30] = { /* Offset base for distance codes 0..29 */ Chris@4: 1, 2, 3, 4, 5, 7, 9, 13, 17, 25, 33, 49, 65, 97, 129, 193, Chris@4: 257, 385, 513, 769, 1025, 1537, 2049, 3073, 4097, 6145, Chris@4: 8193, 12289, 16385, 24577}; Chris@4: static const short dext[30] = { /* Extra bits for distance codes 0..29 */ Chris@4: 0, 0, 0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, Chris@4: 7, 7, 8, 8, 9, 9, 10, 10, 11, 11, Chris@4: 12, 12, 13, 13}; Chris@4: Chris@4: /* decode literals and length/distance pairs */ Chris@4: do { Chris@4: symbol = decode(s, lencode); Chris@4: if (symbol < 0) Chris@4: return symbol; /* invalid symbol */ Chris@4: if (symbol < 256) { /* literal: symbol is the byte */ Chris@4: /* write out the literal */ Chris@4: if (s->out != NIL) { Chris@4: if (s->outcnt == s->outlen) Chris@4: return 1; Chris@4: s->out[s->outcnt] = symbol; Chris@4: } Chris@4: s->outcnt++; Chris@4: } Chris@4: else if (symbol > 256) { /* length */ Chris@4: /* get and compute length */ Chris@4: symbol -= 257; Chris@4: if (symbol >= 29) Chris@4: return -10; /* invalid fixed code */ Chris@4: len = lens[symbol] + bits(s, lext[symbol]); Chris@4: Chris@4: /* get and check distance */ Chris@4: symbol = decode(s, distcode); Chris@4: if (symbol < 0) Chris@4: return symbol; /* invalid symbol */ Chris@4: dist = dists[symbol] + bits(s, dext[symbol]); Chris@4: #ifndef INFLATE_ALLOW_INVALID_DISTANCE_TOOFAR_ARRR Chris@4: if (dist > s->outcnt) Chris@4: return -11; /* distance too far back */ Chris@4: #endif Chris@4: Chris@4: /* copy length bytes from distance bytes back */ Chris@4: if (s->out != NIL) { Chris@4: if (s->outcnt + len > s->outlen) Chris@4: return 1; Chris@4: while (len--) { Chris@4: s->out[s->outcnt] = Chris@4: #ifdef INFLATE_ALLOW_INVALID_DISTANCE_TOOFAR_ARRR Chris@4: dist > s->outcnt ? Chris@4: 0 : Chris@4: #endif Chris@4: s->out[s->outcnt - dist]; Chris@4: s->outcnt++; Chris@4: } Chris@4: } Chris@4: else Chris@4: s->outcnt += len; Chris@4: } Chris@4: } while (symbol != 256); /* end of block symbol */ Chris@4: Chris@4: /* done with a valid fixed or dynamic block */ Chris@4: return 0; Chris@4: } Chris@4: Chris@4: /* Chris@4: * Process a fixed codes block. Chris@4: * Chris@4: * Format notes: Chris@4: * Chris@4: * - This block type can be useful for compressing small amounts of data for Chris@4: * which the size of the code descriptions in a dynamic block exceeds the Chris@4: * benefit of custom codes for that block. For fixed codes, no bits are Chris@4: * spent on code descriptions. Instead the code lengths for literal/length Chris@4: * codes and distance codes are fixed. The specific lengths for each symbol Chris@4: * can be seen in the "for" loops below. Chris@4: * Chris@4: * - The literal/length code is complete, but has two symbols that are invalid Chris@4: * and should result in an error if received. This cannot be implemented Chris@4: * simply as an incomplete code since those two symbols are in the "middle" Chris@4: * of the code. They are eight bits long and the longest literal/length\ Chris@4: * code is nine bits. Therefore the code must be constructed with those Chris@4: * symbols, and the invalid symbols must be detected after decoding. Chris@4: * Chris@4: * - The fixed distance codes also have two invalid symbols that should result Chris@4: * in an error if received. Since all of the distance codes are the same Chris@4: * length, this can be implemented as an incomplete code. Then the invalid Chris@4: * codes are detected while decoding. Chris@4: */ Chris@4: local int fixed(struct state *s) Chris@4: { Chris@4: static int virgin = 1; Chris@4: static short lencnt[MAXBITS+1], lensym[FIXLCODES]; Chris@4: static short distcnt[MAXBITS+1], distsym[MAXDCODES]; Chris@4: static struct huffman lencode, distcode; Chris@4: Chris@4: /* build fixed huffman tables if first call (may not be thread safe) */ Chris@4: if (virgin) { Chris@4: int symbol; Chris@4: short lengths[FIXLCODES]; Chris@4: Chris@4: /* construct lencode and distcode */ Chris@4: lencode.count = lencnt; Chris@4: lencode.symbol = lensym; Chris@4: distcode.count = distcnt; Chris@4: distcode.symbol = distsym; Chris@4: Chris@4: /* literal/length table */ Chris@4: for (symbol = 0; symbol < 144; symbol++) Chris@4: lengths[symbol] = 8; Chris@4: for (; symbol < 256; symbol++) Chris@4: lengths[symbol] = 9; Chris@4: for (; symbol < 280; symbol++) Chris@4: lengths[symbol] = 7; Chris@4: for (; symbol < FIXLCODES; symbol++) Chris@4: lengths[symbol] = 8; Chris@4: construct(&lencode, lengths, FIXLCODES); Chris@4: Chris@4: /* distance table */ Chris@4: for (symbol = 0; symbol < MAXDCODES; symbol++) Chris@4: lengths[symbol] = 5; Chris@4: construct(&distcode, lengths, MAXDCODES); Chris@4: Chris@4: /* do this just once */ Chris@4: virgin = 0; Chris@4: } Chris@4: Chris@4: /* decode data until end-of-block code */ Chris@4: return codes(s, &lencode, &distcode); Chris@4: } Chris@4: Chris@4: /* Chris@4: * Process a dynamic codes block. Chris@4: * Chris@4: * Format notes: Chris@4: * Chris@4: * - A dynamic block starts with a description of the literal/length and Chris@4: * distance codes for that block. New dynamic blocks allow the compressor to Chris@4: * rapidly adapt to changing data with new codes optimized for that data. Chris@4: * Chris@4: * - The codes used by the deflate format are "canonical", which means that Chris@4: * the actual bits of the codes are generated in an unambiguous way simply Chris@4: * from the number of bits in each code. Therefore the code descriptions Chris@4: * are simply a list of code lengths for each symbol. Chris@4: * Chris@4: * - The code lengths are stored in order for the symbols, so lengths are Chris@4: * provided for each of the literal/length symbols, and for each of the Chris@4: * distance symbols. Chris@4: * Chris@4: * - If a symbol is not used in the block, this is represented by a zero as Chris@4: * as the code length. This does not mean a zero-length code, but rather Chris@4: * that no code should be created for this symbol. There is no way in the Chris@4: * deflate format to represent a zero-length code. Chris@4: * Chris@4: * - The maximum number of bits in a code is 15, so the possible lengths for Chris@4: * any code are 1..15. Chris@4: * Chris@4: * - The fact that a length of zero is not permitted for a code has an Chris@4: * interesting consequence. Normally if only one symbol is used for a given Chris@4: * code, then in fact that code could be represented with zero bits. However Chris@4: * in deflate, that code has to be at least one bit. So for example, if Chris@4: * only a single distance base symbol appears in a block, then it will be Chris@4: * represented by a single code of length one, in particular one 0 bit. This Chris@4: * is an incomplete code, since if a 1 bit is received, it has no meaning, Chris@4: * and should result in an error. So incomplete distance codes of one symbol Chris@4: * should be permitted, and the receipt of invalid codes should be handled. Chris@4: * Chris@4: * - It is also possible to have a single literal/length code, but that code Chris@4: * must be the end-of-block code, since every dynamic block has one. This Chris@4: * is not the most efficient way to create an empty block (an empty fixed Chris@4: * block is fewer bits), but it is allowed by the format. So incomplete Chris@4: * literal/length codes of one symbol should also be permitted. Chris@4: * Chris@4: * - If there are only literal codes and no lengths, then there are no distance Chris@4: * codes. This is represented by one distance code with zero bits. Chris@4: * Chris@4: * - The list of up to 286 length/literal lengths and up to 30 distance lengths Chris@4: * are themselves compressed using Huffman codes and run-length encoding. In Chris@4: * the list of code lengths, a 0 symbol means no code, a 1..15 symbol means Chris@4: * that length, and the symbols 16, 17, and 18 are run-length instructions. Chris@4: * Each of 16, 17, and 18 are follwed by extra bits to define the length of Chris@4: * the run. 16 copies the last length 3 to 6 times. 17 represents 3 to 10 Chris@4: * zero lengths, and 18 represents 11 to 138 zero lengths. Unused symbols Chris@4: * are common, hence the special coding for zero lengths. Chris@4: * Chris@4: * - The symbols for 0..18 are Huffman coded, and so that code must be Chris@4: * described first. This is simply a sequence of up to 19 three-bit values Chris@4: * representing no code (0) or the code length for that symbol (1..7). Chris@4: * Chris@4: * - A dynamic block starts with three fixed-size counts from which is computed Chris@4: * the number of literal/length code lengths, the number of distance code Chris@4: * lengths, and the number of code length code lengths (ok, you come up with Chris@4: * a better name!) in the code descriptions. For the literal/length and Chris@4: * distance codes, lengths after those provided are considered zero, i.e. no Chris@4: * code. The code length code lengths are received in a permuted order (see Chris@4: * the order[] array below) to make a short code length code length list more Chris@4: * likely. As it turns out, very short and very long codes are less likely Chris@4: * to be seen in a dynamic code description, hence what may appear initially Chris@4: * to be a peculiar ordering. Chris@4: * Chris@4: * - Given the number of literal/length code lengths (nlen) and distance code Chris@4: * lengths (ndist), then they are treated as one long list of nlen + ndist Chris@4: * code lengths. Therefore run-length coding can and often does cross the Chris@4: * boundary between the two sets of lengths. Chris@4: * Chris@4: * - So to summarize, the code description at the start of a dynamic block is Chris@4: * three counts for the number of code lengths for the literal/length codes, Chris@4: * the distance codes, and the code length codes. This is followed by the Chris@4: * code length code lengths, three bits each. This is used to construct the Chris@4: * code length code which is used to read the remainder of the lengths. Then Chris@4: * the literal/length code lengths and distance lengths are read as a single Chris@4: * set of lengths using the code length codes. Codes are constructed from Chris@4: * the resulting two sets of lengths, and then finally you can start Chris@4: * decoding actual compressed data in the block. Chris@4: * Chris@4: * - For reference, a "typical" size for the code description in a dynamic Chris@4: * block is around 80 bytes. Chris@4: */ Chris@4: local int dynamic(struct state *s) Chris@4: { Chris@4: int nlen, ndist, ncode; /* number of lengths in descriptor */ Chris@4: int index; /* index of lengths[] */ Chris@4: int err; /* construct() return value */ Chris@4: short lengths[MAXCODES]; /* descriptor code lengths */ Chris@4: short lencnt[MAXBITS+1], lensym[MAXLCODES]; /* lencode memory */ Chris@4: short distcnt[MAXBITS+1], distsym[MAXDCODES]; /* distcode memory */ Chris@4: struct huffman lencode, distcode; /* length and distance codes */ Chris@4: static const short order[19] = /* permutation of code length codes */ Chris@4: {16, 17, 18, 0, 8, 7, 9, 6, 10, 5, 11, 4, 12, 3, 13, 2, 14, 1, 15}; Chris@4: Chris@4: /* construct lencode and distcode */ Chris@4: lencode.count = lencnt; Chris@4: lencode.symbol = lensym; Chris@4: distcode.count = distcnt; Chris@4: distcode.symbol = distsym; Chris@4: Chris@4: /* get number of lengths in each table, check lengths */ Chris@4: nlen = bits(s, 5) + 257; Chris@4: ndist = bits(s, 5) + 1; Chris@4: ncode = bits(s, 4) + 4; Chris@4: if (nlen > MAXLCODES || ndist > MAXDCODES) Chris@4: return -3; /* bad counts */ Chris@4: Chris@4: /* read code length code lengths (really), missing lengths are zero */ Chris@4: for (index = 0; index < ncode; index++) Chris@4: lengths[order[index]] = bits(s, 3); Chris@4: for (; index < 19; index++) Chris@4: lengths[order[index]] = 0; Chris@4: Chris@4: /* build huffman table for code lengths codes (use lencode temporarily) */ Chris@4: err = construct(&lencode, lengths, 19); Chris@4: if (err != 0) /* require complete code set here */ Chris@4: return -4; Chris@4: Chris@4: /* read length/literal and distance code length tables */ Chris@4: index = 0; Chris@4: while (index < nlen + ndist) { Chris@4: int symbol; /* decoded value */ Chris@4: int len; /* last length to repeat */ Chris@4: Chris@4: symbol = decode(s, &lencode); Chris@4: if (symbol < 16) /* length in 0..15 */ Chris@4: lengths[index++] = symbol; Chris@4: else { /* repeat instruction */ Chris@4: len = 0; /* assume repeating zeros */ Chris@4: if (symbol == 16) { /* repeat last length 3..6 times */ Chris@4: if (index == 0) Chris@4: return -5; /* no last length! */ Chris@4: len = lengths[index - 1]; /* last length */ Chris@4: symbol = 3 + bits(s, 2); Chris@4: } Chris@4: else if (symbol == 17) /* repeat zero 3..10 times */ Chris@4: symbol = 3 + bits(s, 3); Chris@4: else /* == 18, repeat zero 11..138 times */ Chris@4: symbol = 11 + bits(s, 7); Chris@4: if (index + symbol > nlen + ndist) Chris@4: return -6; /* too many lengths! */ Chris@4: while (symbol--) /* repeat last or zero symbol times */ Chris@4: lengths[index++] = len; Chris@4: } Chris@4: } Chris@4: Chris@4: /* check for end-of-block code -- there better be one! */ Chris@4: if (lengths[256] == 0) Chris@4: return -9; Chris@4: Chris@4: /* build huffman table for literal/length codes */ Chris@4: err = construct(&lencode, lengths, nlen); Chris@4: if (err && (err < 0 || nlen != lencode.count[0] + lencode.count[1])) Chris@4: return -7; /* incomplete code ok only for single length 1 code */ Chris@4: Chris@4: /* build huffman table for distance codes */ Chris@4: err = construct(&distcode, lengths + nlen, ndist); Chris@4: if (err && (err < 0 || ndist != distcode.count[0] + distcode.count[1])) Chris@4: return -8; /* incomplete code ok only for single length 1 code */ Chris@4: Chris@4: /* decode data until end-of-block code */ Chris@4: return codes(s, &lencode, &distcode); Chris@4: } Chris@4: Chris@4: /* Chris@4: * Inflate source to dest. On return, destlen and sourcelen are updated to the Chris@4: * size of the uncompressed data and the size of the deflate data respectively. Chris@4: * On success, the return value of puff() is zero. If there is an error in the Chris@4: * source data, i.e. it is not in the deflate format, then a negative value is Chris@4: * returned. If there is not enough input available or there is not enough Chris@4: * output space, then a positive error is returned. In that case, destlen and Chris@4: * sourcelen are not updated to facilitate retrying from the beginning with the Chris@4: * provision of more input data or more output space. In the case of invalid Chris@4: * inflate data (a negative error), the dest and source pointers are updated to Chris@4: * facilitate the debugging of deflators. Chris@4: * Chris@4: * puff() also has a mode to determine the size of the uncompressed output with Chris@4: * no output written. For this dest must be (unsigned char *)0. In this case, Chris@4: * the input value of *destlen is ignored, and on return *destlen is set to the Chris@4: * size of the uncompressed output. Chris@4: * Chris@4: * The return codes are: Chris@4: * Chris@4: * 2: available inflate data did not terminate Chris@4: * 1: output space exhausted before completing inflate Chris@4: * 0: successful inflate Chris@4: * -1: invalid block type (type == 3) Chris@4: * -2: stored block length did not match one's complement Chris@4: * -3: dynamic block code description: too many length or distance codes Chris@4: * -4: dynamic block code description: code lengths codes incomplete Chris@4: * -5: dynamic block code description: repeat lengths with no first length Chris@4: * -6: dynamic block code description: repeat more than specified lengths Chris@4: * -7: dynamic block code description: invalid literal/length code lengths Chris@4: * -8: dynamic block code description: invalid distance code lengths Chris@4: * -9: dynamic block code description: missing end-of-block code Chris@4: * -10: invalid literal/length or distance code in fixed or dynamic block Chris@4: * -11: distance is too far back in fixed or dynamic block Chris@4: * Chris@4: * Format notes: Chris@4: * Chris@4: * - Three bits are read for each block to determine the kind of block and Chris@4: * whether or not it is the last block. Then the block is decoded and the Chris@4: * process repeated if it was not the last block. Chris@4: * Chris@4: * - The leftover bits in the last byte of the deflate data after the last Chris@4: * block (if it was a fixed or dynamic block) are undefined and have no Chris@4: * expected values to check. Chris@4: */ Chris@4: int puff(unsigned char *dest, /* pointer to destination pointer */ Chris@4: unsigned long *destlen, /* amount of output space */ Chris@4: const unsigned char *source, /* pointer to source data pointer */ Chris@4: unsigned long *sourcelen) /* amount of input available */ Chris@4: { Chris@4: struct state s; /* input/output state */ Chris@4: int last, type; /* block information */ Chris@4: int err; /* return value */ Chris@4: Chris@4: /* initialize output state */ Chris@4: s.out = dest; Chris@4: s.outlen = *destlen; /* ignored if dest is NIL */ Chris@4: s.outcnt = 0; Chris@4: Chris@4: /* initialize input state */ Chris@4: s.in = source; Chris@4: s.inlen = *sourcelen; Chris@4: s.incnt = 0; Chris@4: s.bitbuf = 0; Chris@4: s.bitcnt = 0; Chris@4: Chris@4: /* return if bits() or decode() tries to read past available input */ Chris@4: if (setjmp(s.env) != 0) /* if came back here via longjmp() */ Chris@4: err = 2; /* then skip do-loop, return error */ Chris@4: else { Chris@4: /* process blocks until last block or error */ Chris@4: do { Chris@4: last = bits(&s, 1); /* one if last block */ Chris@4: type = bits(&s, 2); /* block type 0..3 */ Chris@4: err = type == 0 ? Chris@4: stored(&s) : Chris@4: (type == 1 ? Chris@4: fixed(&s) : Chris@4: (type == 2 ? Chris@4: dynamic(&s) : Chris@4: -1)); /* type == 3, invalid */ Chris@4: if (err != 0) Chris@4: break; /* return with error */ Chris@4: } while (!last); Chris@4: } Chris@4: Chris@4: /* update the lengths and return */ Chris@4: if (err <= 0) { Chris@4: *destlen = s.outcnt; Chris@4: *sourcelen = s.incnt; Chris@4: } Chris@4: return err; Chris@4: }