Merge pull request #19193 from hrydgard/ir-interpreter-opts

IRInterpreter: Enable some optimizations that accidentally were only enabled on non-ARM64.

Core/MIPS/ARM64/Arm64IRJit.cpp

@@ -44,6 +44,7 @@ Arm64JitBackend::Arm64JitBackend(JitOptions &jitopt, IRBlockCache &blocks)
 	if (((intptr_t)Memory::base & 0x00000000FFFFFFFFUL) != 0) {
 		jo.enablePointerify = false;
 	}
+	jo.optimizeForInterpreter = false;
 #ifdef MASKED_PSP_MEMORY
 	jo.enablePointerify = false;
 #endif

Core/MIPS/IR/IRFrontend.cpp

@@ -277,7 +277,7 @@ void IRFrontend::DoJit(u32 em_address, std::vector<IRInst> &instructions, u32 &m
 	IRWriter simplified;
 	IRWriter *code = &ir;
 	if (!js.hadBreakpoints) {
-		static const IRPassFunc passes[] = {
+		std::vector<IRPassFunc> passes{
 			&ApplyMemoryValidation,
 			&RemoveLoadStoreLeftRight,
 			&OptimizeFPMoves,
@@ -288,7 +288,12 @@ void IRFrontend::DoJit(u32 em_address, std::vector<IRInst> &instructions, u32 &m
 			// &MergeLoadStore,
 			// &ThreeOpToTwoOp,
 		};
-		if (IRApplyPasses(passes, ARRAY_SIZE(passes), ir, simplified, opts))
+
+		if (opts.optimizeForInterpreter) {
+			// Add special passes here.
+			// passes.push_back(&ReorderLoadStore);
+		}
+		if (IRApplyPasses(passes.data(), passes.size(), ir, simplified, opts))
 			logBlocks = 1;
 		code = &simplified;
 		//if (ir.GetInstructions().size() >= 24)

Core/MIPS/IR/IRInst.h

@@ -405,6 +405,7 @@ struct IROptions {
 	bool unalignedLoadStoreVec4;
 	bool preferVec4;
 	bool preferVec4Dot;
+	bool optimizeForInterpreter;
 };
 
 const IRMeta *GetIRMeta(IROp op);

Core/MIPS/IR/IRJit.cpp

@@ -48,6 +48,8 @@ IRJit::IRJit(MIPSState *mipsState) : frontend_(mipsState->HasDefaultPrefix()), m
 	// blTrampolines_ = kernelMemory.Alloc(size, true, "trampoline");
 	InitIR();
 
+	jo.optimizeForInterpreter = true;
+
 	IROptions opts{};
 	opts.disableFlags = g_Config.uJitDisableFlags;
 #if PPSSPP_ARCH(RISCV64)
@@ -55,7 +57,7 @@ IRJit::IRJit(MIPSState *mipsState) : frontend_(mipsState->HasDefaultPrefix()), m
 	opts.unalignedLoadStore = false;
 	opts.unalignedLoadStoreVec4 = true;
 	opts.preferVec4 = cpu_info.RiscV_V;
-#elif PPSSPP_ARCH(ARM)
+#elif PPSSPP_ARCH(ARM) || PPSSPP_ARCH(ARM64)
 	opts.unalignedLoadStore = (opts.disableFlags & (uint32_t)JitDisable::LSU_UNALIGNED) == 0;
 	opts.unalignedLoadStoreVec4 = true;
 	opts.preferVec4 = cpu_info.bASIMD || cpu_info.bNEON;
@@ -65,6 +67,7 @@ IRJit::IRJit(MIPSState *mipsState) : frontend_(mipsState->HasDefaultPrefix()), m
 	opts.unalignedLoadStoreVec4 = false;
 	opts.preferVec4 = true;
 #endif
+	opts.optimizeForInterpreter = jo.optimizeForInterpreter;
 	frontend_.SetOptions(opts);
 }
 
@@ -143,7 +146,7 @@ bool IRJit::CompileBlock(u32 em_address, std::vector<IRInst> &instructions, u32
 
 	IRBlock *b = blocks_.GetBlock(block_num);
 	b->SetInstructions(instructions);
-	b->SetOriginalSize(mipsBytes);
+	b->SetOriginalAddrSize(em_address, mipsBytes);
 	if (preload) {
 		// Hash, then only update page stats, don't link yet.
 		// TODO: Should we always hash?  Then we can reuse blocks.

Core/MIPS/IR/IRJit.h

@@ -69,7 +69,8 @@ public:
 	bool HasOriginalFirstOp() const;
 	bool RestoreOriginalFirstOp(int number);
 	bool IsValid() const { return origAddr_ != 0 && origFirstOpcode_.encoding != 0x68FFFFFF; }
-	void SetOriginalSize(u32 size) {
+	void SetOriginalAddrSize(u32 address, u32 size) {
+		origAddr_ = address;
 		origSize_ = size;
 	}
 	void SetTargetOffset(int offset) {
@@ -114,25 +115,28 @@ public:
 	IRBlockCache() {}
 	void Clear();
 	std::vector<int> FindInvalidatedBlockNumbers(u32 address, u32 length);
-	void FinalizeBlock(int i, bool preload = false);
+	void FinalizeBlock(int blockNum, bool preload = false);
 	int GetNumBlocks() const override { return (int)blocks_.size(); }
 	int AllocateBlock(int emAddr) {
 		blocks_.push_back(IRBlock(emAddr));
 		return (int)blocks_.size() - 1;
 	}
-	IRBlock *GetBlock(int i) {
+	IRBlock *GetBlock(int blockNum) {
-		if (i >= 0 && i < (int)blocks_.size()) {
+		if (blockNum >= 0 && blockNum < (int)blocks_.size()) {
-			return &blocks_[i];
+			return &blocks_[blockNum];
 		} else {
 			return nullptr;
 		}
 	}
-	IRBlock *GetBlockUnchecked(int i) {
+	bool IsValidBlock(int blockNum) const override {
-		return &blocks_[i];
+		return blockNum < (int)blocks_.size() && blocks_[blockNum].IsValid();
 	}
-	const IRBlock *GetBlock(int i) const {
+	IRBlock *GetBlockUnchecked(int blockNum) {
-		if (i >= 0 && i < (int)blocks_.size()) {
+		return &blocks_[blockNum];
-			return &blocks_[i];
+	}
+	const IRBlock *GetBlock(int blockNum) const {
+		if (blockNum >= 0 && blockNum < (int)blocks_.size()) {
+			return &blocks_[blockNum];
 		} else {
 			return nullptr;
 		}

Core/MIPS/IR/IRNativeCommon.cpp

@@ -714,6 +714,10 @@ void IRNativeBlockCacheDebugInterface::Init(const IRNativeBackend *backend) {
 	backend_ = backend;
 }
 
+bool IRNativeBlockCacheDebugInterface::IsValidBlock(int blockNum) const {
+	return irBlocks_.IsValidBlock(blockNum);
+}
+
 int IRNativeBlockCacheDebugInterface::GetNumBlocks() const {
 	return irBlocks_.GetNumBlocks();
 }

Core/MIPS/IR/IRNativeCommon.h

@@ -162,10 +162,11 @@ class IRNativeBlockCacheDebugInterface : public JitBlockCacheDebugInterface {
 public:
 	IRNativeBlockCacheDebugInterface(const MIPSComp::IRBlockCache &irBlocks);
 	void Init(const IRNativeBackend *backend);
-	int GetNumBlocks() const;
+	int GetNumBlocks() const override;
-	int GetBlockNumberFromStartAddress(u32 em_address, bool realBlocksOnly = true) const;
+	int GetBlockNumberFromStartAddress(u32 em_address, bool realBlocksOnly = true) const override;
-	JitBlockDebugInfo GetBlockDebugInfo(int blockNum) const;
+	JitBlockDebugInfo GetBlockDebugInfo(int blockNum) const override;
-	void ComputeStats(BlockCacheStats &bcStats) const;
+	void ComputeStats(BlockCacheStats &bcStats) const override;
+	bool IsValidBlock(int blockNum) const override;
 
 private:
 	void GetBlockCodeRange(int blockNum, int *startOffset, int *size) const;

Core/MIPS/JitCommon/JitBlockCache.h

@@ -109,6 +109,7 @@ public:
 	virtual int GetBlockNumberFromStartAddress(u32 em_address, bool realBlocksOnly = true) const = 0;
 	virtual JitBlockDebugInfo GetBlockDebugInfo(int blockNum) const = 0;
 	virtual void ComputeStats(BlockCacheStats &bcStats) const = 0;
+	virtual bool IsValidBlock(int blockNum) const = 0;
 
 	virtual ~JitBlockCacheDebugInterface() {}
 };
@@ -164,6 +165,7 @@ public:
 	void RestoreSavedEmuHackOps(const std::vector<u32> &saved);
 
 	int GetNumBlocks() const override { return num_blocks_; }
+	bool IsValidBlock(int blockNum) const override { return blockNum < num_blocks_ && !blocks_[blockNum].invalid; }
 
 	static int GetBlockExitSize();
 

Core/MIPS/JitCommon/JitState.h

@@ -237,6 +237,8 @@ namespace MIPSComp {
 		// ARM64 and RV64
 		bool useStaticAlloc;
 		bool enablePointerify;
+		// IR Interpreter
+		bool optimizeForInterpreter;
 
 		// Common
 		bool enableBlocklink;
@@ -245,6 +247,4 @@ namespace MIPSComp {
 		bool continueJumps;
 		int continueMaxInstructions;
 	};
-
 }
-

Core/MIPS/RiscV/RiscVJit.cpp

@@ -39,6 +39,7 @@ RiscVJitBackend::RiscVJitBackend(JitOptions &jitopt, IRBlockCache &blocks)
 	if (((intptr_t)Memory::base & 0x00000000FFFFFFFFUL) != 0) {
 		jo.enablePointerify = false;
 	}
+	jo.optimizeForInterpreter = false;
 
 	// Since we store the offset, this is as big as it can be.
 	// We could shift off one bit to double it, would need to change RiscVAsm.

Core/MIPS/x86/X64IRJit.cpp

@@ -41,6 +41,7 @@ X64JitBackend::X64JitBackend(JitOptions &jitopt, IRBlockCache &blocks)
 	if (((intptr_t)Memory::base & 0x00000000FFFFFFFFUL) != 0) {
 		jo.enablePointerify = false;
 	}
+	jo.optimizeForInterpreter = false;
 
 	// Since we store the offset, this is as big as it can be.
 	AllocCodeSpace(1024 * 1024 * 16);

UI/DevScreens.cpp

@@ -1060,6 +1060,9 @@ void JitCompareScreen::UpdateDisasm() {
 	}
 
 	JitBlockCacheDebugInterface *blockCacheDebug = MIPSComp::jit->GetBlockCacheDebugInterface();
+	if (!blockCacheDebug->IsValidBlock(currentBlock_)) {
+		return;
+	}
 
 	char temp[256];
 	snprintf(temp, sizeof(temp), "%i/%i", currentBlock_, blockCacheDebug->GetNumBlocks());
@@ -1205,7 +1208,13 @@ UI::EventReturn JitCompareScreen::OnRandomBlock(UI::EventParams &e) {
 
 	int numBlocks = blockCache->GetNumBlocks();
 	if (numBlocks > 0) {
-		currentBlock_ = rand() % numBlocks;
+		int tries = 100;
+		while (tries-- > 0) {
+			currentBlock_ = rand() % numBlocks;
+			if (blockCache->IsValidBlock(currentBlock_)) {
+				break;
+			}
+		}
 	}
 	UpdateDisasm();
 	return UI::EVENT_DONE;