Merge pull request #1218 from hthh/trampolinecaching

JIT: reuse trampolines when possible

4 files changed, 100 insertions, 11 deletions

diff --git a/Source/Core/Common/x64Analyzer.cpp b/Source/Core/Common/x64Analyzer.cpp
index 5be91ea585..7e8b40e6cd 100644
--- a/Source/Core/Common/x64Analyzer.cpp
+++ b/Source/Core/Common/x64Analyzer.cpp
@@ -223,3 +223,19 @@ bool DisassembleMov(const unsigned char *codePtr, InstructionInfo *info)
 	info->instructionSize = (int)(codePtr - startCodePtr);
 	return true;
 }
+
+bool InstructionInfo::operator==(const InstructionInfo &other) const
+{
+	return operandSize     == other.operandSize     &&
+	       instructionSize == other.instructionSize &&
+	       regOperandReg   == other.regOperandReg   &&
+	       otherReg        == other.otherReg        &&
+	       scaledReg       == other.scaledReg       &&
+	       zeroExtend      == other.zeroExtend      &&
+	       signExtend      == other.signExtend      &&
+	       hasImmediate    == other.hasImmediate    &&
+	       isMemoryWrite   == other.isMemoryWrite   &&
+	       byteSwap        == other.byteSwap        &&
+	       immediate       == other.immediate       &&
+	       displacement    == other.displacement;
+}
diff --git a/Source/Core/Common/x64Analyzer.h b/Source/Core/Common/x64Analyzer.h
index 4858f099da..c87a89a51f 100644
--- a/Source/Core/Common/x64Analyzer.h
+++ b/Source/Core/Common/x64Analyzer.h
@@ -20,6 +20,8 @@ struct InstructionInfo
 	bool byteSwap;
 	u64 immediate;
 	s32 displacement;
+
+	bool operator==(const InstructionInfo &other) const;
 };
 
 struct ModRM
diff --git a/Source/Core/Core/PowerPC/JitCommon/TrampolineCache.cpp b/Source/Core/Core/PowerPC/JitCommon/TrampolineCache.cpp
index ae83eec2d5..7b23ac1427 100644
--- a/Source/Core/Core/PowerPC/JitCommon/TrampolineCache.cpp
+++ b/Source/Core/Core/PowerPC/JitCommon/TrampolineCache.cpp
@@ -19,25 +19,42 @@
 
 using namespace Gen;
 
-extern u8 *trampolineCodePtr;
-
 void TrampolineCache::Init()
 {
-	AllocCodeSpace(4 * 1024 * 1024);
+	AllocCodeSpace(8 * 1024 * 1024);
+}
+
+void TrampolineCache::ClearCodeSpace()
+{
+	X64CodeBlock::ClearCodeSpace();
+	cachedTrampolines.clear();
 }
 
 void TrampolineCache::Shutdown()
 {
 	FreeCodeSpace();
+	cachedTrampolines.clear();
+}
+
+const u8* TrampolineCache::GetReadTrampoline(const InstructionInfo &info, u32 registersInUse)
+{
+	TrampolineCacheKey key = { registersInUse, 0, info };
+
+	auto it = cachedTrampolines.find(key);
+	if (it != cachedTrampolines.end())
+		return it->second;
+
+	const u8* trampoline = GenerateReadTrampoline(info, registersInUse);
+	cachedTrampolines[key] = trampoline;
+	return trampoline;
 }
 
-// Extremely simplistic - just generate the requested trampoline. May reuse them in the future.
-const u8 *TrampolineCache::GetReadTrampoline(const InstructionInfo &info, u32 registersInUse)
+const u8* TrampolineCache::GenerateReadTrampoline(const InstructionInfo &info, u32 registersInUse)
 {
 	if (GetSpaceLeft() < 1024)
 		PanicAlert("Trampoline cache full");
 
-	const u8 *trampoline = GetCodePtr();
+	const u8* trampoline = GetCodePtr();
 	X64Reg addrReg = (X64Reg)info.scaledReg;
 	X64Reg dataReg = (X64Reg)info.regOperandReg;
 
@@ -80,13 +97,25 @@ const u8 *TrampolineCache::GetReadTrampoline(const InstructionInfo &info, u32 re
 	return trampoline;
 }
 
-// Extremely simplistic - just generate the requested trampoline. May reuse them in the future.
-const u8 *TrampolineCache::GetWriteTrampoline(const InstructionInfo &info, u32 registersInUse, u32 pc)
+const u8* TrampolineCache::GetWriteTrampoline(const InstructionInfo &info, u32 registersInUse, u32 pc)
+{
+	TrampolineCacheKey key = { registersInUse, pc, info };
+
+	auto it = cachedTrampolines.find(key);
+	if (it != cachedTrampolines.end())
+		return it->second;
+
+	const u8* trampoline = GenerateWriteTrampoline(info, registersInUse, pc);
+	cachedTrampolines[key] = trampoline;
+	return trampoline;
+}
+
+const u8* TrampolineCache::GenerateWriteTrampoline(const InstructionInfo &info, u32 registersInUse, u32 pc)
 {
 	if (GetSpaceLeft() < 1024)
 		PanicAlert("Trampoline cache full");
 
-	const u8 *trampoline = GetCodePtr();
+	const u8* trampoline = GetCodePtr();
 
 	X64Reg dataReg = (X64Reg)info.regOperandReg;
 	X64Reg addrReg = (X64Reg)info.scaledReg;
@@ -153,4 +182,25 @@ const u8 *TrampolineCache::GetWriteTrampoline(const InstructionInfo &info, u32 r
 	return trampoline;
 }
 
+size_t TrampolineCacheKeyHasher::operator()(const TrampolineCacheKey& k) const
+{
+	size_t res = std::hash<int>()(k.registersInUse);
+	res ^= std::hash<int>()(k.info.operandSize)    >> 1;
+	res ^= std::hash<int>()(k.info.regOperandReg)  >> 2;
+	res ^= std::hash<int>()(k.info.scaledReg)      >> 3;
+	res ^= std::hash<u64>()(k.info.immediate)      >> 4;
+	res ^= std::hash<int>()(k.pc)                  >> 5;
+	res ^= std::hash<int>()(k.info.displacement)   << 1;
+	res ^= std::hash<bool>()(k.info.signExtend)    << 2;
+	res ^= std::hash<bool>()(k.info.hasImmediate)  << 3;
+	res ^= std::hash<bool>()(k.info.isMemoryWrite) << 4;
+
+	return res;
+}
 
+bool TrampolineCacheKey::operator==(const TrampolineCacheKey &other) const
+{
+	return pc == other.pc &&
+	       registersInUse == other.registersInUse &&
+	       info == other.info;
+}
diff --git a/Source/Core/Core/PowerPC/JitCommon/TrampolineCache.h b/Source/Core/Core/PowerPC/JitCommon/TrampolineCache.h
index 516a071ac2..bf7109d6d9 100644
--- a/Source/Core/Core/PowerPC/JitCommon/TrampolineCache.h
+++ b/Source/Core/Core/PowerPC/JitCommon/TrampolineCache.h
@@ -11,12 +11,33 @@
 // We need at least this many bytes for backpatching.
 const int BACKPATCH_SIZE = 5;
 
+struct TrampolineCacheKey
+{
+	u32 registersInUse;
+	u32 pc;
+	InstructionInfo info;
+
+	bool operator==(const TrampolineCacheKey &other) const;
+};
+
+struct TrampolineCacheKeyHasher
+{
+	size_t operator()(const TrampolineCacheKey& k) const;
+};
+
 class TrampolineCache : public Gen::X64CodeBlock
 {
 public:
 	void Init();
 	void Shutdown();
 
-	const u8 *GetReadTrampoline(const InstructionInfo &info, u32 registersInUse);
-	const u8 *GetWriteTrampoline(const InstructionInfo &info, u32 registersInUse, u32 pc);
+	const u8* GetReadTrampoline(const InstructionInfo &info, u32 registersInUse);
+	const u8* GetWriteTrampoline(const InstructionInfo &info, u32 registersInUse, u32 pc);
+	void ClearCodeSpace();
+
+private:
+	const u8* GenerateReadTrampoline(const InstructionInfo &info, u32 registersInUse);
+	const u8* GenerateWriteTrampoline(const InstructionInfo &info, u32 registersInUse, u32 pc);
+
+	std::unordered_map<TrampolineCacheKey, const u8*, TrampolineCacheKeyHasher> cachedTrampolines;
 };