1 /* vi: set sw=4 ts=4: */
2 /* Small bzip2 deflate implementation, by Rob Landley (rob@landley.net).
4 Based on bzip2 decompression code by Julian R Seward (jseward@acm.org),
5 which also acknowledges contributions by Mike Burrows, David Wheeler,
6 Peter Fenwick, Alistair Moffat, Radford Neal, Ian H. Witten,
7 Robert Sedgewick, and Jon L. Bentley.
9 This code is licensed under the LGPLv2:
10 LGPL (http://www.gnu.org/copyleft/lgpl.html
14 Size and speed optimizations by Manuel Novoa III (mjn3@codepoet.org).
16 More efficient reading of Huffman codes, a streamlined read_bunzip()
17 function, and various other tweaks. In (limited) tests, approximately
18 20% faster than bzcat on x86 and about 10% faster on arm.
20 Note that about 2/3 of the time is spent in read_unzip() reversing
21 the Burrows-Wheeler transformation. Much of that time is delay
22 resulting from cache misses.
24 I would ask that anyone benefiting from this work, especially those
25 using it in commercial products, consider making a donation to my local
26 non-profit hospice organization in the name of the woman I loved, who
27 passed away Feb. 12, 2003.
29 In memory of Toni W. Hagan
31 Hospice of Acadiana, Inc.
32 2600 Johnston St., Suite 200
33 Lafayette, LA 70503-3240
35 Phone (337) 232-1234 or 1-800-738-2226
38 http://www.hospiceacadiana.com/
52 /* Constants for Huffman coding */
54 #define GROUP_SIZE 50 /* 64 would have been more efficient */
55 #define MAX_HUFCODE_BITS 20 /* Longest Huffman code allowed */
56 #define MAX_SYMBOLS 258 /* 256 literals + RUNA + RUNB */
60 /* Status return values */
62 #define RETVAL_LAST_BLOCK (-1)
63 #define RETVAL_NOT_BZIP_DATA (-2)
64 #define RETVAL_UNEXPECTED_INPUT_EOF (-3)
65 #define RETVAL_UNEXPECTED_OUTPUT_EOF (-4)
66 #define RETVAL_DATA_ERROR (-5)
67 #define RETVAL_OUT_OF_MEMORY (-6)
68 #define RETVAL_OBSOLETE_INPUT (-7)
70 /* Other housekeeping constants */
71 #define IOBUF_SIZE 4096
73 /* This is what we know about each Huffman coding group */
75 /* We have an extra slot at the end of limit[] for a sentinal value. */
76 int limit[MAX_HUFCODE_BITS+1],base[MAX_HUFCODE_BITS],permute[MAX_SYMBOLS];
80 /* Structure holding all the housekeeping data, including IO buffers and
81 memory that persists between calls to bunzip */
83 /* State for interrupting output loop */
84 int writeCopies,writePos,writeRunCountdown,writeCount,writeCurrent;
85 /* I/O tracking data (file handles, buffers, positions, etc.) */
86 int in_fd,out_fd,inbufCount,inbufPos /*,outbufPos*/;
87 unsigned char *inbuf /*,*outbuf*/;
88 unsigned int inbufBitCount, inbufBits;
89 /* The CRC values stored in the block header and calculated from the data */
90 unsigned int crc32Table[256],headerCRC, totalCRC, writeCRC;
91 /* Intermediate buffer and its size (in bytes) */
92 unsigned int *dbuf, dbufSize;
93 /* These things are a bit too big to go on the stack */
94 unsigned char selectors[32768]; /* nSelectors=15 bits */
95 struct group_data groups[MAX_GROUPS]; /* Huffman coding tables */
96 /* For I/O error handling */
100 /* Return the next nnn bits of input. All reads from the compressed input
101 are done through this function. All reads are big endian */
102 static unsigned int get_bits(bunzip_data *bd, char bits_wanted)
106 /* If we need to get more data from the byte buffer, do so. (Loop getting
107 one byte at a time to enforce endianness and avoid unaligned access.) */
108 while (bd->inbufBitCount<bits_wanted) {
109 /* If we need to read more data from file into byte buffer, do so */
110 if(bd->inbufPos==bd->inbufCount) {
111 if((bd->inbufCount = read(bd->in_fd, bd->inbuf, IOBUF_SIZE)) <= 0)
112 longjmp(bd->jmpbuf,RETVAL_UNEXPECTED_INPUT_EOF);
115 /* Avoid 32-bit overflow (dump bit buffer to top of output) */
116 if(bd->inbufBitCount>=24) {
117 bits=bd->inbufBits&((1<<bd->inbufBitCount)-1);
118 bits_wanted-=bd->inbufBitCount;
122 /* Grab next 8 bits of input from buffer. */
123 bd->inbufBits=(bd->inbufBits<<8)|bd->inbuf[bd->inbufPos++];
124 bd->inbufBitCount+=8;
126 /* Calculate result */
127 bd->inbufBitCount-=bits_wanted;
128 bits|=(bd->inbufBits>>bd->inbufBitCount)&((1<<bits_wanted)-1);
133 /* Unpacks the next block and sets up for the inverse burrows-wheeler step. */
135 static int get_next_block(bunzip_data *bd)
137 /* Note: Ignore the warning about hufGroup, base and limit being used uninitialized.
138 * They will be initialized on the fist pass of the loop. */
139 struct group_data *hufGroup;
140 int dbufCount,nextSym,dbufSize,groupCount,*base,*limit,selector,
141 i,j,k,t,runPos,symCount,symTotal,nSelectors,byteCount[256];
142 unsigned char uc, symToByte[256], mtfSymbol[256], *selectors;
143 unsigned int *dbuf,origPtr;
146 dbufSize=bd->dbufSize;
147 selectors=bd->selectors;
148 /* Reset longjmp I/O error handling */
149 i=setjmp(bd->jmpbuf);
151 /* Read in header signature and CRC, then validate signature.
152 (last block signature means CRC is for whole file, return now) */
155 bd->headerCRC=get_bits(bd,32);
156 if ((i == 0x177245) && (j == 0x385090)) return RETVAL_LAST_BLOCK;
157 if ((i != 0x314159) || (j != 0x265359)) return RETVAL_NOT_BZIP_DATA;
158 /* We can add support for blockRandomised if anybody complains. There was
159 some code for this in busybox 1.0.0-pre3, but nobody ever noticed that
160 it didn't actually work. */
161 if(get_bits(bd,1)) return RETVAL_OBSOLETE_INPUT;
162 if((origPtr=get_bits(bd,24)) > dbufSize) return RETVAL_DATA_ERROR;
163 /* mapping table: if some byte values are never used (encoding things
164 like ascii text), the compression code removes the gaps to have fewer
165 symbols to deal with, and writes a sparse bitfield indicating which
166 values were present. We make a translation table to convert the symbols
167 back to the corresponding bytes. */
174 if(k&(1<<(15-j))) symToByte[symTotal++]=(16*i)+j;
177 /* How many different Huffman coding groups does this block use? */
178 groupCount=get_bits(bd,3);
179 if (groupCount<2 || groupCount>MAX_GROUPS) return RETVAL_DATA_ERROR;
180 /* nSelectors: Every GROUP_SIZE many symbols we select a new Huffman coding
181 group. Read in the group selector list, which is stored as MTF encoded
182 bit runs. (MTF=Move To Front, as each value is used it's moved to the
183 start of the list.) */
184 if(!(nSelectors=get_bits(bd, 15))) return RETVAL_DATA_ERROR;
185 for(i=0; i<groupCount; i++) mtfSymbol[i] = i;
186 for(i=0; i<nSelectors; i++) {
188 for(j=0;get_bits(bd,1);j++) if (j>=groupCount) return RETVAL_DATA_ERROR;
189 /* Decode MTF to get the next selector */
191 for(;j;j--) mtfSymbol[j] = mtfSymbol[j-1];
192 mtfSymbol[0]=selectors[i]=uc;
194 /* Read the Huffman coding tables for each group, which code for symTotal
195 literal symbols, plus two run symbols (RUNA, RUNB) */
197 for (j=0; j<groupCount; j++) {
198 unsigned char length[MAX_SYMBOLS],temp[MAX_HUFCODE_BITS+1];
199 int minLen, maxLen, pp;
200 /* Read Huffman code lengths for each symbol. They're stored in
201 a way similar to mtf; record a starting value for the first symbol,
202 and an offset from the previous value for everys symbol after that.
203 (Subtracting 1 before the loop and then adding it back at the end is
204 an optimization that makes the test inside the loop simpler: symbol
205 length 0 becomes negative, so an unsigned inequality catches it.) */
207 for (i = 0; i < symCount; i++) {
209 if (((unsigned)t) > (MAX_HUFCODE_BITS-1))
210 return RETVAL_DATA_ERROR;
211 /* If first bit is 0, stop. Else second bit indicates whether
212 to increment or decrement the value. Optimization: grab 2
213 bits and unget the second if the first was 0. */
219 /* Add one if second bit 1, else subtract 1. Avoids if/else */
222 /* Correct for the initial -1, to get the final symbol length */
225 /* Find largest and smallest lengths in this group */
226 minLen=maxLen=length[0];
227 for(i = 1; i < symCount; i++) {
228 if(length[i] > maxLen) maxLen = length[i];
229 else if(length[i] < minLen) minLen = length[i];
231 /* Calculate permute[], base[], and limit[] tables from length[].
233 * permute[] is the lookup table for converting Huffman coded symbols
234 * into decoded symbols. base[] is the amount to subtract from the
235 * value of a Huffman symbol of a given length when using permute[].
237 * limit[] indicates the largest numerical value a symbol with a given
238 * number of bits can have. This is how the Huffman codes can vary in
239 * length: each code with a value>limit[length] needs another bit.
241 hufGroup=bd->groups+j;
242 hufGroup->minLen = minLen;
243 hufGroup->maxLen = maxLen;
244 /* Note that minLen can't be smaller than 1, so we adjust the base
245 and limit array pointers so we're not always wasting the first
246 entry. We do this again when using them (during symbol decoding).*/
247 base=hufGroup->base-1;
248 limit=hufGroup->limit-1;
249 /* Calculate permute[]. Concurently, initialize temp[] and limit[]. */
251 for(i=minLen;i<=maxLen;i++) {
253 for(t=0;t<symCount;t++)
254 if(length[t]==i) hufGroup->permute[pp++] = t;
256 /* Count symbols coded for at each bit length */
257 for (i=0;i<symCount;i++) temp[length[i]]++;
258 /* Calculate limit[] (the largest symbol-coding value at each bit
259 * length, which is (previous limit<<1)+symbols at this level), and
260 * base[] (number of symbols to ignore at each bit length, which is
261 * limit minus the cumulative count of symbols coded for already). */
263 for (i=minLen; i<maxLen; i++) {
265 /* We read the largest possible symbol size and then unget bits
266 after determining how many we need, and those extra bits could
267 be set to anything. (They're noise from future symbols.) At
268 each level we're really only interested in the first few bits,
269 so here we set all the trailing to-be-ignored bits to 1 so they
270 don't affect the value>limit[length] comparison. */
271 limit[i]= (pp << (maxLen - i)) - 1;
273 base[i+1]=pp-(t+=temp[i]);
275 limit[maxLen+1] = INT_MAX; /* Sentinal value for reading next sym. */
276 limit[maxLen]=pp+temp[maxLen]-1;
279 /* We've finished reading and digesting the block header. Now read this
280 block's Huffman coded symbols from the file and undo the Huffman coding
281 and run length encoding, saving the result into dbuf[dbufCount++]=uc */
283 /* Initialize symbol occurrence counters and symbol Move To Front table */
286 mtfSymbol[i]=(unsigned char)i;
288 /* Loop through compressed symbols. */
289 runPos=dbufCount=symCount=selector=0;
291 /* Determine which Huffman coding group to use. */
293 symCount=GROUP_SIZE-1;
294 if(selector>=nSelectors) return RETVAL_DATA_ERROR;
295 hufGroup=bd->groups+selectors[selector++];
296 base=hufGroup->base-1;
297 limit=hufGroup->limit-1;
299 /* Read next Huffman-coded symbol. */
300 /* Note: It is far cheaper to read maxLen bits and back up than it is
301 to read minLen bits and then an additional bit at a time, testing
302 as we go. Because there is a trailing last block (with file CRC),
303 there is no danger of the overread causing an unexpected EOF for a
304 valid compressed file. As a further optimization, we do the read
305 inline (falling back to a call to get_bits if the buffer runs
306 dry). The following (up to got_huff_bits:) is equivalent to
307 j=get_bits(bd,hufGroup->maxLen);
309 while (bd->inbufBitCount<hufGroup->maxLen) {
310 if(bd->inbufPos==bd->inbufCount) {
311 j = get_bits(bd,hufGroup->maxLen);
314 bd->inbufBits=(bd->inbufBits<<8)|bd->inbuf[bd->inbufPos++];
315 bd->inbufBitCount+=8;
317 bd->inbufBitCount-=hufGroup->maxLen;
318 j = (bd->inbufBits>>bd->inbufBitCount)&((1<<hufGroup->maxLen)-1);
320 /* Figure how how many bits are in next symbol and unget extras */
322 while(j>limit[i]) ++i;
323 bd->inbufBitCount += (hufGroup->maxLen - i);
324 /* Huffman decode value to get nextSym (with bounds checking) */
325 if ((i > hufGroup->maxLen)
326 || (((unsigned)(j=(j>>(hufGroup->maxLen-i))-base[i]))
328 return RETVAL_DATA_ERROR;
329 nextSym = hufGroup->permute[j];
330 /* We have now decoded the symbol, which indicates either a new literal
331 byte, or a repeated run of the most recent literal byte. First,
332 check if nextSym indicates a repeated run, and if so loop collecting
333 how many times to repeat the last literal. */
334 if (((unsigned)nextSym) <= SYMBOL_RUNB) { /* RUNA or RUNB */
335 /* If this is the start of a new run, zero out counter */
340 /* Neat trick that saves 1 symbol: instead of or-ing 0 or 1 at
341 each bit position, add 1 or 2 instead. For example,
342 1011 is 1<<0 + 1<<1 + 2<<2. 1010 is 2<<0 + 2<<1 + 1<<2.
343 You can make any bit pattern that way using 1 less symbol than
344 the basic or 0/1 method (except all bits 0, which would use no
345 symbols, but a run of length 0 doesn't mean anything in this
346 context). Thus space is saved. */
347 t += (runPos << nextSym); /* +runPos if RUNA; +2*runPos if RUNB */
351 /* When we hit the first non-run symbol after a run, we now know
352 how many times to repeat the last literal, so append that many
353 copies to our buffer of decoded symbols (dbuf) now. (The last
354 literal used is the one at the head of the mtfSymbol array.) */
357 if(dbufCount+t>=dbufSize) return RETVAL_DATA_ERROR;
359 uc = symToByte[mtfSymbol[0]];
361 while(t--) dbuf[dbufCount++]=uc;
363 /* Is this the terminating symbol? */
364 if(nextSym>symTotal) break;
365 /* At this point, nextSym indicates a new literal character. Subtract
366 one to get the position in the MTF array at which this literal is
367 currently to be found. (Note that the result can't be -1 or 0,
368 because 0 and 1 are RUNA and RUNB. But another instance of the
369 first symbol in the mtf array, position 0, would have been handled
370 as part of a run above. Therefore 1 unused mtf position minus
371 2 non-literal nextSym values equals -1.) */
372 if(dbufCount>=dbufSize) return RETVAL_DATA_ERROR;
375 /* Adjust the MTF array. Since we typically expect to move only a
376 * small number of symbols, and are bound by 256 in any case, using
377 * memmove here would typically be bigger and slower due to function
378 * call overhead and other assorted setup costs. */
380 mtfSymbol[i] = mtfSymbol[i-1];
384 /* We have our literal byte. Save it into dbuf. */
386 dbuf[dbufCount++] = (unsigned int)uc;
388 /* At this point, we've read all the Huffman-coded symbols (and repeated
389 runs) for this block from the input stream, and decoded them into the
390 intermediate buffer. There are dbufCount many decoded bytes in dbuf[].
391 Now undo the Burrows-Wheeler transform on dbuf.
392 See http://dogma.net/markn/articles/bwt/bwt.htm
394 /* Turn byteCount into cumulative occurrence counts of 0 to n-1. */
401 /* Figure out what order dbuf would be in if we sorted it. */
402 for (i=0;i<dbufCount;i++) {
403 uc=(unsigned char)(dbuf[i] & 0xff);
404 dbuf[byteCount[uc]] |= (i << 8);
407 /* Decode first byte by hand to initialize "previous" byte. Note that it
408 doesn't get output, and if the first three characters are identical
409 it doesn't qualify as a run (hence writeRunCountdown=5). */
411 if(origPtr>=dbufCount) return RETVAL_DATA_ERROR;
412 bd->writePos=dbuf[origPtr];
413 bd->writeCurrent=(unsigned char)(bd->writePos&0xff);
415 bd->writeRunCountdown=5;
417 bd->writeCount=dbufCount;
422 /* Undo burrows-wheeler transform on intermediate buffer to produce output.
423 If start_bunzip was initialized with out_fd=-1, then up to len bytes of
424 data are written to outbuf. Return value is number of bytes written or
425 error (all errors are negative numbers). If out_fd!=-1, outbuf and len
426 are ignored, data is written to out_fd and return is RETVAL_OK or error.
429 static int read_bunzip(bunzip_data *bd, char *outbuf, int len)
431 const unsigned int *dbuf;
432 int pos,current,previous,gotcount;
434 /* If last read was short due to end of file, return last block now */
435 if(bd->writeCount<0) return bd->writeCount;
440 current=bd->writeCurrent;
442 /* We will always have pending decoded data to write into the output
443 buffer unless this is the very first call (in which case we haven't
444 Huffman-decoded a block into the intermediate buffer yet). */
446 if (bd->writeCopies) {
447 /* Inside the loop, writeCopies means extra copies (beyond 1) */
449 /* Loop outputting bytes */
451 /* If the output buffer is full, snapshot state and return */
452 if(gotcount >= len) {
454 bd->writeCurrent=current;
458 /* Write next byte into output buffer, updating CRC */
459 outbuf[gotcount++] = current;
460 bd->writeCRC=(((bd->writeCRC)<<8)
461 ^bd->crc32Table[((bd->writeCRC)>>24)^current]);
462 /* Loop now if we're outputting multiple copies of this byte */
463 if (bd->writeCopies) {
468 if (!bd->writeCount--) break;
469 /* Follow sequence vector to undo Burrows-Wheeler transform */
474 /* After 3 consecutive copies of the same byte, the 4th is a repeat
475 count. We count down from 4 instead
476 * of counting up because testing for non-zero is faster */
477 if(--bd->writeRunCountdown) {
478 if(current!=previous) bd->writeRunCountdown=4;
480 /* We have a repeated run, this byte indicates the count */
481 bd->writeCopies=current;
483 bd->writeRunCountdown=5;
484 /* Sometimes there are just 3 bytes (run length 0) */
485 if(!bd->writeCopies) goto decode_next_byte;
486 /* Subtract the 1 copy we'd output anyway to get extras */
490 /* Decompression of this block completed successfully */
491 bd->writeCRC=~bd->writeCRC;
492 bd->totalCRC=((bd->totalCRC<<1) | (bd->totalCRC>>31)) ^ bd->writeCRC;
493 /* If this block had a CRC error, force file level CRC error. */
494 if(bd->writeCRC!=bd->headerCRC) {
495 bd->totalCRC=bd->headerCRC+1;
496 return RETVAL_LAST_BLOCK;
500 /* Refill the intermediate buffer by Huffman-decoding next block of input */
501 /* (previous is just a convenient unused temp variable here) */
502 previous=get_next_block(bd);
504 bd->writeCount=previous;
505 return (previous!=RETVAL_LAST_BLOCK) ? previous : gotcount;
507 bd->writeCRC=0xffffffffUL;
509 current=bd->writeCurrent;
510 goto decode_next_byte;
513 /* Allocate the structure, read file header. If in_fd==-1, inbuf must contain
514 a complete bunzip file (len bytes long). If in_fd!=-1, inbuf and len are
515 ignored, and data is read from file handle into temporary buffer. */
516 static int start_bunzip(bunzip_data **bdp, int in_fd, char *inbuf, int len)
520 const unsigned int BZh0=(((unsigned int)'B')<<24)+(((unsigned int)'Z')<<16)
521 +(((unsigned int)'h')<<8)+(unsigned int)'0';
523 /* Figure out how much data to allocate */
524 i=sizeof(bunzip_data);
525 if(in_fd!=-1) i+=IOBUF_SIZE;
526 /* Allocate bunzip_data. Most fields initialize to zero. */
528 memset(bd,0,sizeof(bunzip_data));
529 /* Setup input buffer */
530 if(-1==(bd->in_fd=in_fd)) {
533 } else bd->inbuf=(unsigned char *)(bd+1);
534 /* Init the CRC32 table (big endian) */
538 c=c&0x80000000 ? (c<<1)^0x04c11db7 : (c<<1);
541 /* Setup for I/O error handling via longjmp */
542 i=setjmp(bd->jmpbuf);
545 /* Ensure that file starts with "BZh['1'-'9']." */
547 if (((unsigned int)(i-BZh0-1)) >= 9) return RETVAL_NOT_BZIP_DATA;
549 /* Fourth byte (ascii '1'-'9'), indicates block size in units of 100k of
550 uncompressed data. Allocate intermediate buffer for block. */
551 bd->dbufSize=100000*(i-BZh0);
553 bd->dbuf=xmalloc(bd->dbufSize * sizeof(int));
557 /* Example usage: decompress src_fd to dst_fd. (Stops at end of bzip data,
559 extern int uncompressStream(int src_fd, int dst_fd)
565 outbuf=xmalloc(IOBUF_SIZE);
566 if(!(i=start_bunzip(&bd,src_fd,0,0))) {
568 if((i=read_bunzip(bd,outbuf,IOBUF_SIZE)) <= 0) break;
569 if(i!=write(dst_fd,outbuf,i)) {
570 i=RETVAL_UNEXPECTED_OUTPUT_EOF;
575 /* Check CRC and release memory */
576 if(i==RETVAL_LAST_BLOCK) {
577 if (bd->headerCRC!=bd->totalCRC) {
578 bb_error_msg("Data integrity error when decompressing.");
583 else if (i==RETVAL_UNEXPECTED_OUTPUT_EOF) {
584 bb_error_msg("Compressed file ends unexpectedly");
586 bb_error_msg("Decompression failed");
588 if(bd->dbuf) free(bd->dbuf);
597 static char * const bunzip_errors[]={NULL,"Bad file checksum","Not bzip data",
598 "Unexpected input EOF","Unexpected output EOF","Data error",
599 "Out of memory","Obsolete (pre 0.9.5) bzip format not supported."};
601 /* Dumb little test thing, decompress stdin to stdout */
602 int main(int argc, char *argv[])
604 int i=uncompressStream(0,1);
607 if(i) fprintf(stderr,"%s\n", bunzip_errors[-i]);
608 else if(read(0,&c,1)) fprintf(stderr,"Trailing garbage ignored\n");