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 struct group_data *hufGroup;
138 int dbufCount,nextSym,dbufSize,groupCount,*base,*limit,selector,
139 i,j,k,t,runPos,symCount,symTotal,nSelectors,byteCount[256];
140 unsigned char uc, symToByte[256], mtfSymbol[256], *selectors;
141 unsigned int *dbuf,origPtr;
144 dbufSize=bd->dbufSize;
145 selectors=bd->selectors;
146 /* Reset longjmp I/O error handling */
147 i=setjmp(bd->jmpbuf);
149 /* Read in header signature and CRC, then validate signature.
150 (last block signature means CRC is for whole file, return now) */
153 bd->headerCRC=get_bits(bd,32);
154 if ((i == 0x177245) && (j == 0x385090)) return RETVAL_LAST_BLOCK;
155 if ((i != 0x314159) || (j != 0x265359)) return RETVAL_NOT_BZIP_DATA;
156 /* We can add support for blockRandomised if anybody complains. There was
157 some code for this in busybox 1.0.0-pre3, but nobody ever noticed that
158 it didn't actually work. */
159 if(get_bits(bd,1)) return RETVAL_OBSOLETE_INPUT;
160 if((origPtr=get_bits(bd,24)) > dbufSize) return RETVAL_DATA_ERROR;
161 /* mapping table: if some byte values are never used (encoding things
162 like ascii text), the compression code removes the gaps to have fewer
163 symbols to deal with, and writes a sparse bitfield indicating which
164 values were present. We make a translation table to convert the symbols
165 back to the corresponding bytes. */
172 if(k&(1<<(15-j))) symToByte[symTotal++]=(16*i)+j;
175 /* How many different Huffman coding groups does this block use? */
176 groupCount=get_bits(bd,3);
177 if (groupCount<2 || groupCount>MAX_GROUPS) return RETVAL_DATA_ERROR;
178 /* nSelectors: Every GROUP_SIZE many symbols we select a new Huffman coding
179 group. Read in the group selector list, which is stored as MTF encoded
180 bit runs. (MTF=Move To Front, as each value is used it's moved to the
181 start of the list.) */
182 if(!(nSelectors=get_bits(bd, 15))) return RETVAL_DATA_ERROR;
183 for(i=0; i<groupCount; i++) mtfSymbol[i] = i;
184 for(i=0; i<nSelectors; i++) {
186 for(j=0;get_bits(bd,1);j++) if (j>=groupCount) return RETVAL_DATA_ERROR;
187 /* Decode MTF to get the next selector */
189 for(;j;j--) mtfSymbol[j] = mtfSymbol[j-1];
190 mtfSymbol[0]=selectors[i]=uc;
192 /* Read the Huffman coding tables for each group, which code for symTotal
193 literal symbols, plus two run symbols (RUNA, RUNB) */
195 for (j=0; j<groupCount; j++) {
196 unsigned char length[MAX_SYMBOLS],temp[MAX_HUFCODE_BITS+1];
197 int minLen, maxLen, pp;
198 /* Read Huffman code lengths for each symbol. They're stored in
199 a way similar to mtf; record a starting value for the first symbol,
200 and an offset from the previous value for everys symbol after that.
201 (Subtracting 1 before the loop and then adding it back at the end is
202 an optimization that makes the test inside the loop simpler: symbol
203 length 0 becomes negative, so an unsigned inequality catches it.) */
205 for (i = 0; i < symCount; i++) {
207 if (((unsigned)t) > (MAX_HUFCODE_BITS-1))
208 return RETVAL_DATA_ERROR;
209 /* If first bit is 0, stop. Else second bit indicates whether
210 to increment or decrement the value. Optimization: grab 2
211 bits and unget the second if the first was 0. */
217 /* Add one if second bit 1, else subtract 1. Avoids if/else */
220 /* Correct for the initial -1, to get the final symbol length */
223 /* Find largest and smallest lengths in this group */
224 minLen=maxLen=length[0];
225 for(i = 1; i < symCount; i++) {
226 if(length[i] > maxLen) maxLen = length[i];
227 else if(length[i] < minLen) minLen = length[i];
229 /* Calculate permute[], base[], and limit[] tables from length[].
231 * permute[] is the lookup table for converting Huffman coded symbols
232 * into decoded symbols. base[] is the amount to subtract from the
233 * value of a Huffman symbol of a given length when using permute[].
235 * limit[] indicates the largest numerical value a symbol with a given
236 * number of bits can have. This is how the Huffman codes can vary in
237 * length: each code with a value>limit[length] needs another bit.
239 hufGroup=bd->groups+j;
240 hufGroup->minLen = minLen;
241 hufGroup->maxLen = maxLen;
242 /* Note that minLen can't be smaller than 1, so we adjust the base
243 and limit array pointers so we're not always wasting the first
244 entry. We do this again when using them (during symbol decoding).*/
245 base=hufGroup->base-1;
246 limit=hufGroup->limit-1;
247 /* Calculate permute[]. Concurently, initialize temp[] and limit[]. */
249 for(i=minLen;i<=maxLen;i++) {
251 for(t=0;t<symCount;t++)
252 if(length[t]==i) hufGroup->permute[pp++] = t;
254 /* Count symbols coded for at each bit length */
255 for (i=0;i<symCount;i++) temp[length[i]]++;
256 /* Calculate limit[] (the largest symbol-coding value at each bit
257 * length, which is (previous limit<<1)+symbols at this level), and
258 * base[] (number of symbols to ignore at each bit length, which is
259 * limit minus the cumulative count of symbols coded for already). */
261 for (i=minLen; i<maxLen; i++) {
263 /* We read the largest possible symbol size and then unget bits
264 after determining how many we need, and those extra bits could
265 be set to anything. (They're noise from future symbols.) At
266 each level we're really only interested in the first few bits,
267 so here we set all the trailing to-be-ignored bits to 1 so they
268 don't affect the value>limit[length] comparison. */
269 limit[i]= (pp << (maxLen - i)) - 1;
271 base[i+1]=pp-(t+=temp[i]);
273 limit[maxLen+1] = INT_MAX; /* Sentinal value for reading next sym. */
274 limit[maxLen]=pp+temp[maxLen]-1;
277 /* We've finished reading and digesting the block header. Now read this
278 block's Huffman coded symbols from the file and undo the Huffman coding
279 and run length encoding, saving the result into dbuf[dbufCount++]=uc */
281 /* Initialize symbol occurrence counters and symbol Move To Front table */
284 mtfSymbol[i]=(unsigned char)i;
286 /* Loop through compressed symbols. */
287 runPos=dbufCount=selector=0;
289 /* fetch next Huffman coding group from list. */
290 symCount=GROUP_SIZE-1;
291 if(selector>=nSelectors) return RETVAL_DATA_ERROR;
292 hufGroup=bd->groups+selectors[selector++];
293 base=hufGroup->base-1;
294 limit=hufGroup->limit-1;
296 /* Read next Huffman-coded symbol. */
297 /* Note: It is far cheaper to read maxLen bits and back up than it is
298 to read minLen bits and then an additional bit at a time, testing
299 as we go. Because there is a trailing last block (with file CRC),
300 there is no danger of the overread causing an unexpected EOF for a
301 valid compressed file. As a further optimization, we do the read
302 inline (falling back to a call to get_bits if the buffer runs
303 dry). The following (up to got_huff_bits:) is equivalent to
304 j=get_bits(bd,hufGroup->maxLen);
306 while (bd->inbufBitCount<hufGroup->maxLen) {
307 if(bd->inbufPos==bd->inbufCount) {
308 j = get_bits(bd,hufGroup->maxLen);
311 bd->inbufBits=(bd->inbufBits<<8)|bd->inbuf[bd->inbufPos++];
312 bd->inbufBitCount+=8;
314 bd->inbufBitCount-=hufGroup->maxLen;
315 j = (bd->inbufBits>>bd->inbufBitCount)&((1<<hufGroup->maxLen)-1);
317 /* Figure how how many bits are in next symbol and unget extras */
319 while(j>limit[i]) ++i;
320 bd->inbufBitCount += (hufGroup->maxLen - i);
321 /* Huffman decode value to get nextSym (with bounds checking) */
322 if ((i > hufGroup->maxLen)
323 || (((unsigned)(j=(j>>(hufGroup->maxLen-i))-base[i]))
325 return RETVAL_DATA_ERROR;
326 nextSym = hufGroup->permute[j];
327 /* We have now decoded the symbol, which indicates either a new literal
328 byte, or a repeated run of the most recent literal byte. First,
329 check if nextSym indicates a repeated run, and if so loop collecting
330 how many times to repeat the last literal. */
331 if (((unsigned)nextSym) <= SYMBOL_RUNB) { /* RUNA or RUNB */
332 /* If this is the start of a new run, zero out counter */
337 /* Neat trick that saves 1 symbol: instead of or-ing 0 or 1 at
338 each bit position, add 1 or 2 instead. For example,
339 1011 is 1<<0 + 1<<1 + 2<<2. 1010 is 2<<0 + 2<<1 + 1<<2.
340 You can make any bit pattern that way using 1 less symbol than
341 the basic or 0/1 method (except all bits 0, which would use no
342 symbols, but a run of length 0 doesn't mean anything in this
343 context). Thus space is saved. */
344 t += (runPos << nextSym); /* +runPos if RUNA; +2*runPos if RUNB */
346 goto end_of_huffman_loop;
348 /* When we hit the first non-run symbol after a run, we now know
349 how many times to repeat the last literal, so append that many
350 copies to our buffer of decoded symbols (dbuf) now. (The last
351 literal used is the one at the head of the mtfSymbol array.) */
354 if(dbufCount+t>=dbufSize) return RETVAL_DATA_ERROR;
356 uc = symToByte[mtfSymbol[0]];
358 while(t--) dbuf[dbufCount++]=uc;
360 /* Is this the terminating symbol? */
361 if(nextSym>symTotal) break;
362 /* At this point, nextSym indicates a new literal character. Subtract
363 one to get the position in the MTF array at which this literal is
364 currently to be found. (Note that the result can't be -1 or 0,
365 because 0 and 1 are RUNA and RUNB. But another instance of the
366 first symbol in the mtf array, position 0, would have been handled
367 as part of a run above. Therefore 1 unused mtf position minus
368 2 non-literal nextSym values equals -1.) */
369 if(dbufCount>=dbufSize) return RETVAL_DATA_ERROR;
372 /* Adjust the MTF array. Since we typically expect to move only a
373 * small number of symbols, and are bound by 256 in any case, using
374 * memmove here would typically be bigger and slower due to function
375 * call overhead and other assorted setup costs. */
377 mtfSymbol[i] = mtfSymbol[i-1];
381 /* We have our literal byte. Save it into dbuf. */
383 dbuf[dbufCount++] = (unsigned int)uc;
384 /* Skip group initialization if we're not done with this group. Done this
385 * way to avoid compiler warning. */
387 if(symCount--) goto continue_this_group;
389 /* At this point, we've read all the Huffman-coded symbols (and repeated
390 runs) for this block from the input stream, and decoded them into the
391 intermediate buffer. There are dbufCount many decoded bytes in dbuf[].
392 Now undo the Burrows-Wheeler transform on dbuf.
393 See http://dogma.net/markn/articles/bwt/bwt.htm
395 /* Turn byteCount into cumulative occurrence counts of 0 to n-1. */
402 /* Figure out what order dbuf would be in if we sorted it. */
403 for (i=0;i<dbufCount;i++) {
404 uc=(unsigned char)(dbuf[i] & 0xff);
405 dbuf[byteCount[uc]] |= (i << 8);
408 /* Decode first byte by hand to initialize "previous" byte. Note that it
409 doesn't get output, and if the first three characters are identical
410 it doesn't qualify as a run (hence writeRunCountdown=5). */
412 if(origPtr>=dbufCount) return RETVAL_DATA_ERROR;
413 bd->writePos=dbuf[origPtr];
414 bd->writeCurrent=(unsigned char)(bd->writePos&0xff);
416 bd->writeRunCountdown=5;
418 bd->writeCount=dbufCount;
423 /* Undo burrows-wheeler transform on intermediate buffer to produce output.
424 If start_bunzip was initialized with out_fd=-1, then up to len bytes of
425 data are written to outbuf. Return value is number of bytes written or
426 error (all errors are negative numbers). If out_fd!=-1, outbuf and len
427 are ignored, data is written to out_fd and return is RETVAL_OK or error.
430 static int read_bunzip(bunzip_data *bd, char *outbuf, int len)
432 const unsigned int *dbuf;
433 int pos,current,previous,gotcount;
435 /* If last read was short due to end of file, return last block now */
436 if(bd->writeCount<0) return bd->writeCount;
441 current=bd->writeCurrent;
443 /* We will always have pending decoded data to write into the output
444 buffer unless this is the very first call (in which case we haven't
445 Huffman-decoded a block into the intermediate buffer yet). */
447 if (bd->writeCopies) {
448 /* Inside the loop, writeCopies means extra copies (beyond 1) */
450 /* Loop outputting bytes */
452 /* If the output buffer is full, snapshot state and return */
453 if(gotcount >= len) {
455 bd->writeCurrent=current;
459 /* Write next byte into output buffer, updating CRC */
460 outbuf[gotcount++] = current;
461 bd->writeCRC=(((bd->writeCRC)<<8)
462 ^bd->crc32Table[((bd->writeCRC)>>24)^current]);
463 /* Loop now if we're outputting multiple copies of this byte */
464 if (bd->writeCopies) {
469 if (!bd->writeCount--) break;
470 /* Follow sequence vector to undo Burrows-Wheeler transform */
475 /* After 3 consecutive copies of the same byte, the 4th is a repeat
476 count. We count down from 4 instead
477 * of counting up because testing for non-zero is faster */
478 if(--bd->writeRunCountdown) {
479 if(current!=previous) bd->writeRunCountdown=4;
481 /* We have a repeated run, this byte indicates the count */
482 bd->writeCopies=current;
484 bd->writeRunCountdown=5;
485 /* Sometimes there are just 3 bytes (run length 0) */
486 if(!bd->writeCopies) goto decode_next_byte;
487 /* Subtract the 1 copy we'd output anyway to get extras */
491 /* Decompression of this block completed successfully */
492 bd->writeCRC=~bd->writeCRC;
493 bd->totalCRC=((bd->totalCRC<<1) | (bd->totalCRC>>31)) ^ bd->writeCRC;
494 /* If this block had a CRC error, force file level CRC error. */
495 if(bd->writeCRC!=bd->headerCRC) {
496 bd->totalCRC=bd->headerCRC+1;
497 return RETVAL_LAST_BLOCK;
501 /* Refill the intermediate buffer by Huffman-decoding next block of input */
502 /* (previous is just a convenient unused temp variable here) */
503 previous=get_next_block(bd);
505 bd->writeCount=previous;
506 return (previous!=RETVAL_LAST_BLOCK) ? previous : gotcount;
508 bd->writeCRC=0xffffffffUL;
510 current=bd->writeCurrent;
511 goto decode_next_byte;
514 /* Allocate the structure, read file header. If in_fd==-1, inbuf must contain
515 a complete bunzip file (len bytes long). If in_fd!=-1, inbuf and len are
516 ignored, and data is read from file handle into temporary buffer. */
517 static int start_bunzip(bunzip_data **bdp, int in_fd, unsigned char *inbuf,
522 const unsigned int BZh0=(((unsigned int)'B')<<24)+(((unsigned int)'Z')<<16)
523 +(((unsigned int)'h')<<8)+(unsigned int)'0';
525 /* Figure out how much data to allocate */
526 i=sizeof(bunzip_data);
527 if(in_fd!=-1) i+=IOBUF_SIZE;
528 /* Allocate bunzip_data. Most fields initialize to zero. */
530 memset(bd,0,sizeof(bunzip_data));
531 /* Setup input buffer */
532 if(-1==(bd->in_fd=in_fd)) {
535 } else bd->inbuf=(unsigned char *)(bd+1);
536 /* Init the CRC32 table (big endian) */
540 c=c&0x80000000 ? (c<<1)^0x04c11db7 : (c<<1);
543 /* Setup for I/O error handling via longjmp */
544 i=setjmp(bd->jmpbuf);
547 /* Ensure that file starts with "BZh['1'-'9']." */
549 if (((unsigned int)(i-BZh0-1)) >= 9) return RETVAL_NOT_BZIP_DATA;
551 /* Fourth byte (ascii '1'-'9'), indicates block size in units of 100k of
552 uncompressed data. Allocate intermediate buffer for block. */
553 bd->dbufSize=100000*(i-BZh0);
555 bd->dbuf=xmalloc(bd->dbufSize * sizeof(int));
559 /* Example usage: decompress src_fd to dst_fd. (Stops at end of bzip data,
561 extern int uncompressStream(int src_fd, int dst_fd)
567 outbuf=xmalloc(IOBUF_SIZE);
568 if(!(i=start_bunzip(&bd,src_fd,0,0))) {
570 if((i=read_bunzip(bd,outbuf,IOBUF_SIZE)) <= 0) break;
571 if(i!=write(dst_fd,outbuf,i)) {
572 i=RETVAL_UNEXPECTED_OUTPUT_EOF;
577 /* Check CRC and release memory */
578 if(i==RETVAL_LAST_BLOCK) {
579 if (bd->headerCRC!=bd->totalCRC) {
580 bb_error_msg("Data integrity error when decompressing.");
585 else if (i==RETVAL_UNEXPECTED_OUTPUT_EOF) {
586 bb_error_msg("Compressed file ends unexpectedly");
588 bb_error_msg("Decompression failed");
590 if(bd->dbuf) free(bd->dbuf);
599 static char * const bunzip_errors[]={NULL,"Bad file checksum","Not bzip data",
600 "Unexpected input EOF","Unexpected output EOF","Data error",
601 "Out of memory","Obsolete (pre 0.9.5) bzip format not supported."};
603 /* Dumb little test thing, decompress stdin to stdout */
604 int main(int argc, char *argv[])
606 int i=uncompressStream(0,1);
609 if(i) fprintf(stderr,"%s\n", bunzip_errors[-i]);
610 else if(read(0,&c,1)) fprintf(stderr,"Trailing garbage ignored\n");