tools: Add fuzz driver for all codecs

All codecs must be resistant to malicious inputs; while the decoders
aren't guaranteed to provide a valid result, they must be able to decode
the input without reading past the input bounds or writing past the
output bounds.

This change adds a fuzzing driver that attempts to prove this for both
SIMD and scalar implementations of the decoders.

Makefile

@@ -121,6 +121,9 @@ codecbench.js codecbench.wasm: tools/codecbench.cpp ${LIBRARY_SOURCES}
 codecbench-simd.js codecbench-simd.wasm: tools/codecbench.cpp ${LIBRARY_SOURCES}
 	emcc $^ -O3 -g -DNDEBUG -s TOTAL_MEMORY=268435456 -msimd128 -o $@
 
+codecfuzz: tools/codecfuzz.cpp src/vertexcodec.cpp src/indexcodec.cpp
+	$(CXX) $^ -fsanitize=fuzzer,address,undefined -O1 -g -o $@
+
 $(LIBRARY): $(LIBRARY_OBJECTS)
 	ar rcs $@ $^
 

tools/codecfuzz.cpp

@@ -0,0 +1,55 @@
+#include "../src/meshoptimizer.h"
+
+#include <stdint.h>
+#include <stdlib.h>
+
+void fuzzDecoder(const uint8_t* data, size_t size, size_t stride, int (*decode)(void*, size_t, size_t, const unsigned char*, size_t))
+{
+	size_t count = 66; // must be divisible by 3 for decodeIndexBuffer; should be >=64 to cover large vertex blocks
+
+	void* destination = malloc(count * stride);
+	assert(destination);
+
+	int rc = decode(destination, count, stride, reinterpret_cast<const unsigned char*>(data), size);
+	(void)rc;
+
+	free(destination);
+}
+
+namespace meshopt
+{
+extern unsigned int cpuid;
+}
+
+extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size)
+{
+	// decodeIndexBuffer supports 2 and 4-byte indices
+	fuzzDecoder(data, size, 2, meshopt_decodeIndexBuffer);
+	fuzzDecoder(data, size, 4, meshopt_decodeIndexBuffer);
+
+	// decodeIndexSequence supports 2 and 4-byte indices
+	fuzzDecoder(data, size, 2, meshopt_decodeIndexSequence);
+	fuzzDecoder(data, size, 4, meshopt_decodeIndexSequence);
+
+	// decodeVertexBuffer supports any strides divisible by 4 in 4-256 interval
+	// It's a waste of time to check all of them, so we'll just check a few with different alignment mod 16
+	fuzzDecoder(data, size, 4, meshopt_decodeVertexBuffer);
+	fuzzDecoder(data, size, 16, meshopt_decodeVertexBuffer);
+	fuzzDecoder(data, size, 24, meshopt_decodeVertexBuffer);
+	fuzzDecoder(data, size, 32, meshopt_decodeVertexBuffer);
+
+#if !(defined(__AVX__) || defined(__SSSE3__)) && (defined(_M_IX86) || defined(_M_X64) || defined(__i386__) || defined(__x86_64__))
+	// When SSSE3/AVX support isn't enabled unconditionally, we use a cpuid-based fallback
+	// It's useful to be able to test scalar code in this case, so we temporarily fake the feature bits
+	// and restore them later
+	unsigned int cpuid = meshopt::cpuid;
+	meshopt::cpuid = 0;
+
+	// Note that scalar code doesn't have odd edge conditions around 16b vertices so we could just test it for one stride
+	fuzzDecoder(data, size, 4, meshopt_decodeVertexBuffer);
+
+	meshopt::cpuid = cpuid;
+#endif
+
+	return 0;
+}