[libzstd] Fix bug in Huffman encoding

Summary:
Huffman encoding with a bad dictionary can encode worse than the
HUF_BLOCKBOUND(srcSize), since we don't filter out incompressible
input, and even if we did, the dictionaries Huffman table could be
ill suited to compressing actual data.

The fast optimization doesn't seem to improve compression speed,
even when I hard coded fast = 1, the speed didn't improve over hard coding
it to 0.

Benchmarks:
$ ./zstd.dev -b1e5
Benchmarking levels from 1 to 5
 1#Synthetic 50%     :  10000000 ->   3139163 (3.186), 524.8 MB/s ,1890.0 MB/s
 2#Synthetic 50%     :  10000000 ->   3115138 (3.210), 372.6 MB/s ,1830.2 MB/s
 3#Synthetic 50%     :  10000000 ->   3222672 (3.103), 223.3 MB/s ,1400.2 MB/s
 4#Synthetic 50%     :  10000000 ->   3276678 (3.052), 198.0 MB/s ,1280.1 MB/s
 5#Synthetic 50%     :  10000000 ->   3271570 (3.057), 107.8 MB/s ,1200.0 MB/s
$ ./zstd -b1e5
Benchmarking levels from 1 to 5
 1#Synthetic 50%     :  10000000 ->   3139163 (3.186), 524.8 MB/s ,1870.2 MB/s
 2#Synthetic 50%     :  10000000 ->   3115138 (3.210), 370.0 MB/s ,1810.3 MB/s
 3#Synthetic 50%     :  10000000 ->   3222672 (3.103), 223.3 MB/s ,1380.1 MB/s
 4#Synthetic 50%     :  10000000 ->   3276678 (3.052), 196.1 MB/s ,1270.0 MB/s
 5#Synthetic 50%     :  10000000 ->   3271570 (3.057), 106.8 MB/s ,1180.1 MB/s
$ ./zstd.dev -b1e5 ../silesia.tar
Benchmarking levels from 1 to 5
 1#silesia.tar       : 211988480 ->  73651685 (2.878), 429.7 MB/s ,1096.5 MB/s
 2#silesia.tar       : 211988480 ->  70158785 (3.022), 321.2 MB/s ,1029.1 MB/s
 3#silesia.tar       : 211988480 ->  66993813 (3.164), 243.7 MB/s , 981.4 MB/s
 4#silesia.tar       : 211988480 ->  66306481 (3.197), 226.7 MB/s , 972.4 MB/s
 5#silesia.tar       : 211988480 ->  64757852 (3.274), 150.3 MB/s , 963.6 MB/s
$ ./zstd -b1e5 ../silesia.tar
Benchmarking levels from 1 to 5
 1#silesia.tar       : 211988480 ->  73651685 (2.878), 429.7 MB/s ,1087.1 MB/s
 2#silesia.tar       : 211988480 ->  70158785 (3.022), 318.8 MB/s ,1029.1 MB/s
 3#silesia.tar       : 211988480 ->  66993813 (3.164), 246.5 MB/s , 981.4 MB/s
 4#silesia.tar       : 211988480 ->  66306481 (3.197), 229.2 MB/s , 972.4 MB/s
 5#silesia.tar       : 211988480 ->  64757852 (3.274), 149.3 MB/s , 963.6 MB/s

Test Plan:
I added a test case to the fuzzer which crashed with ASAN before the patch
and succeeded after.
This commit is contained in:
Nick Terrell
2017-07-18 13:20:40 -07:00
parent 77d67fb167
commit cc1522351f
2 changed files with 37 additions and 2 deletions
+1 -2
View File
@@ -436,7 +436,7 @@ static void HUF_encodeSymbol(BIT_CStream_t* bitCPtr, U32 symbol, const HUF_CElt*
size_t HUF_compressBound(size_t size) { return HUF_COMPRESSBOUND(size); }
#define HUF_FLUSHBITS(s) (fast ? BIT_flushBitsFast(s) : BIT_flushBits(s))
#define HUF_FLUSHBITS(s) BIT_flushBits(s)
#define HUF_FLUSHBITS_1(stream) \
if (sizeof((stream)->bitContainer)*8 < HUF_TABLELOG_MAX*2+7) HUF_FLUSHBITS(stream)
@@ -451,7 +451,6 @@ size_t HUF_compress1X_usingCTable(void* dst, size_t dstSize, const void* src, si
BYTE* const oend = ostart + dstSize;
BYTE* op = ostart;
size_t n;
const unsigned fast = (dstSize >= HUF_BLOCKBOUND(srcSize));
BIT_CStream_t bitC;
/* init */