[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.

diff --git a/lib/compress/huf_compress.c b/lib/compress/huf_compress.c
index 7af0789a9c58ef548e85a522ecd01499cf65cad6..beb4fdb6012eb0eec0bc503d08659c87aee095e4 100644 (file)
--- a/lib/compress/huf_compress.c
+++ b/lib/compress/huf_compress.c
@@ -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 */

diff --git a/tests/fuzzer.c b/tests/fuzzer.c
index 904ce6fd8d2f6f11afa36f307901c47b4cfdf611..3099f346e0b3de09099525387c633d8ee628eb28 100644 (file)
--- a/tests/fuzzer.c
+++ b/tests/fuzzer.c
@@ -1003,6 +1003,42 @@ static int basicUnitTests(U32 seed, double compressibility)
       if (r != _3BYTESTESTLENGTH) goto _output_error; }
     DISPLAYLEVEL(4, "OK \n");
 
+    DISPLAYLEVEL(4, "test%3i : incompressible data and ill suited dictionary : ", testNb++);
+    RDG_genBuffer(CNBuffer, CNBuffSize, 0.0, 0.1, seed);
+    {   /* Train a dictionary on low characters */
+        size_t dictSize = 16 KB;
+        void* const dictBuffer = malloc(dictSize);
+        size_t const totalSampleSize = 1 MB;
+        size_t const sampleUnitSize = 8 KB;
+        U32 const nbSamples = (U32)(totalSampleSize / sampleUnitSize);
+        size_t* const samplesSizes = (size_t*) malloc(nbSamples * sizeof(size_t));
+        if (!dictBuffer || !samplesSizes) goto _output_error;
+        { U32 u; for (u=0; u<nbSamples; u++) samplesSizes[u] = sampleUnitSize; }
+        dictSize = ZDICT_trainFromBuffer(dictBuffer, dictSize, CNBuffer, samplesSizes, nbSamples);
+        if (ZDICT_isError(dictSize)) goto _output_error;
+        /* Reverse the characters to make the dictionary ill suited */
+        {   U32 u;
+            for (u = 0; u < CNBuffSize; ++u) {
+              ((BYTE*)CNBuffer)[u] = 255 - ((BYTE*)CNBuffer)[u];
+            }
+        }
+        {   /* Compress the data */
+            size_t const inputSize = 500;
+            size_t const outputSize = ZSTD_compressBound(inputSize);
+            void* const outputBuffer = malloc(outputSize);
+            ZSTD_CCtx* const cctx = ZSTD_createCCtx();
+            if (!outputBuffer || !cctx) goto _output_error;
+            CHECK(ZSTD_compress_usingDict(cctx, outputBuffer, outputSize, CNBuffer, inputSize, dictBuffer, dictSize, 1));
+            free(outputBuffer);
+            ZSTD_freeCCtx(cctx);
+        }
+
+        free(dictBuffer);
+        free(samplesSizes);
+    }
+    DISPLAYLEVEL(4, "OK \n");
+
+
     /* findFrameCompressedSize on skippable frames */
     DISPLAYLEVEL(4, "test%3i : frame compressed size of skippable frame : ", testNb++);
     {   const char* frame = "\x50\x2a\x4d\x18\x05\x0\x0\0abcde";