samplerjit: Prevent thread local stale cache read.

If the generation count happens to match, would still get a stale pointer
and crash.  Let's just make the generation count static so it always
increases.

GPU/Software/DrawPixel.cpp

@@ -802,10 +802,12 @@ SingleFunc PixelJitCache::GenericSingle(const PixelFuncID &id) {
 }
 
 thread_local PixelJitCache::LastCache PixelJitCache::lastSingle_;
+int PixelJitCache::clearGen_ = 0;
 
 // 256k should be plenty of space for plenty of variations.
 PixelJitCache::PixelJitCache() : CodeBlock(1024 * 64 * 4), cache_(64) {
 	lastSingle_.gen = -1;
+	clearGen_++;
 }
 
 void PixelJitCache::Clear() {

GPU/Software/DrawPixel.h

@@ -128,7 +128,7 @@ private:
 	DenseHashMap<size_t, SingleFunc, nullptr> cache_;
 	std::unordered_map<PixelFuncID, const u8 *> addresses_;
 	std::unordered_set<PixelFuncID> compileQueue_;
-	int clearGen_ = 0;
+	static int clearGen_;
 	static thread_local LastCache lastSingle_;
 
 	const u8 *constBlendHalf_11_4s_ = nullptr;

GPU/Software/Sampler.cpp

@@ -109,12 +109,14 @@ FetchFunc GetFetchFunc(SamplerID id, BinManager *binner) {
 thread_local SamplerJitCache::LastCache SamplerJitCache::lastFetch_;
 thread_local SamplerJitCache::LastCache SamplerJitCache::lastNearest_;
 thread_local SamplerJitCache::LastCache SamplerJitCache::lastLinear_;
+int SamplerJitCache::clearGen_ = 0;
 
 // 256k should be enough.
 SamplerJitCache::SamplerJitCache() : Rasterizer::CodeBlock(1024 * 64 * 4), cache_(64) {
 	lastFetch_.gen = -1;
 	lastNearest_.gen = -1;
 	lastLinear_.gen = -1;
+	clearGen_++;
 }
 
 void SamplerJitCache::Clear() {

GPU/Software/Sampler.h

@@ -147,7 +147,7 @@ private:
 	DenseHashMap<size_t, NearestFunc, nullptr> cache_;
 	std::unordered_map<SamplerID, const u8 *> addresses_;
 	std::unordered_set<SamplerID> compileQueue_;
-	int clearGen_ = 0;
+	static int clearGen_;
 	static thread_local LastCache lastFetch_;
 	static thread_local LastCache lastNearest_;
 	static thread_local LastCache lastLinear_;