SNES: Performance improvements

This makes performance equal or slightly better to what it was before all of the recent accuracy improvements

Core/SNES/InternalRegisters.cpp

@@ -408,5 +408,7 @@ void InternalRegisters::Serialize(Serializer& s)
 
 	SV(_hCounter);
 	SV(_vCounter);
-	SV(_irqEnabled);
+	if(!s.IsSaving()) {
+		_irqEnabled = _state.EnableHorizontalIrq || _state.EnableVerticalIrq;
+	}
 }

Core/SNES/SnesCpu.cpp

@@ -121,7 +121,11 @@ void SnesCpu::IdleTakeBranch()
 void SnesCpu::ProcessCpuCycle()
 {
 	_state.CycleCount++;
-	_state.IrqLock = _dmaController->ProcessPendingTransfers();
+	if(_dmaController->HasPendingTransfer()){
+		_state.IrqLock = _dmaController->ProcessPendingTransfers();
+	} else {
+		_state.IrqLock = false;
+	}
 	DetectNmiSignalEdge();
 }
 

Core/SNES/SnesCpu.h

@@ -323,7 +323,7 @@ class SnesCpu : public ISerializable
 	void AddrMode_StkRel();
 	void AddrMode_StkRelIndIdxY();
 
-	void RunOp();
+	__forceinline void RunOp();
 	__noinline void ProcessHaltedState();
 	__forceinline void CheckForInterrupts();
 

Core/SNES/SnesDmaController.h

@@ -49,6 +49,8 @@ class SnesDmaController final : public ISerializable
 	void BeginHdmaTransfer();
 	void BeginHdmaInit();
 
+	__forceinline bool HasPendingTransfer() { return _needToProcess; }
+
 	bool ProcessPendingTransfers();
 
 	void Write(uint16_t addr, uint8_t value);

Core/SNES/SnesMemoryManager.cpp

@@ -21,7 +21,10 @@ void SnesMemoryManager::Initialize(SnesConsole *console)
 {
 	_masterClock = 0;
 	_openBus = 0;
+
 	_cpuSpeed = 8;
+	UpdateExecCallbacks();
+
 	_console = console;
 	_emu = console->GetEmulator();
 	_regs = console->GetInternalRegisters();
@@ -125,6 +128,33 @@ void SnesMemoryManager::GenerateMasterClockTable()
 	}
 }
 
+template<uint8_t clocks>
+void SnesMemoryManager::IncMasterClock()
+{
+	if constexpr(clocks == 2) {
+		Exec();
+	} else if constexpr(clocks == 4) {
+		Exec();
+		Exec();
+	} else if constexpr(clocks == 6) {
+		Exec();
+		Exec();
+		Exec();
+	} else if constexpr(clocks == 8) {
+		Exec();
+		Exec();
+		Exec();
+		Exec();
+	} else if constexpr(clocks == 12) {
+		Exec();
+		Exec();
+		Exec();
+		Exec();
+		Exec();
+		Exec();
+	}
+}
+
 void SnesMemoryManager::IncMasterClock4()
 {
 	Exec();
@@ -236,7 +266,7 @@ void SnesMemoryManager::ProcessEvent()
 
 uint8_t SnesMemoryManager::Read(uint32_t addr, MemoryOperationType type)
 {
-	IncrementMasterClockValue(_cpuSpeed - 4);
+	(this->*_execRead)();
 
 	uint8_t value;
 	IMemoryHandler *handler = _mappings.GetHandler(addr);
@@ -311,7 +341,8 @@ void SnesMemoryManager::PeekBlock(uint32_t addr, uint8_t *dest)
 
 void SnesMemoryManager::Write(uint32_t addr, uint8_t value, MemoryOperationType type)
 {
-	IncrementMasterClockValue(_cpuSpeed);
+	(this->*_execWrite)();
+
 	if(_emu->ProcessMemoryWrite<CpuType::Snes>(addr, value, type)) {
 		IMemoryHandler* handler = _mappings.GetHandler(addr);
 		if(handler) {
@@ -389,7 +420,30 @@ uint8_t SnesMemoryManager::GetCpuSpeed()
 
 void SnesMemoryManager::SetCpuSpeed(uint8_t speed)
 {
-	_cpuSpeed = speed;
+	if(_cpuSpeed != speed) {
+		_cpuSpeed = speed;
+		UpdateExecCallbacks();
+	}
+}
+
+void SnesMemoryManager::UpdateExecCallbacks()
+{
+	switch(_cpuSpeed) {
+		case 6:
+			_execRead = &SnesMemoryManager::IncMasterClock<2>;
+			_execWrite = &SnesMemoryManager::IncMasterClock<6>;
+			break;
+
+		case 8:
+			_execRead = &SnesMemoryManager::IncMasterClock<4>;
+			_execWrite = &SnesMemoryManager::IncMasterClock<8>;
+			break;
+
+		case 12:
+			_execRead = &SnesMemoryManager::IncMasterClock<8>;
+			_execWrite = &SnesMemoryManager::IncMasterClock<12>;
+			break;
+	}
 }
 
 MemoryType SnesMemoryManager::GetMemoryTypeBusA()
@@ -420,4 +474,8 @@ void SnesMemoryManager::Serialize(Serializer &s)
 	SV(_memTypeBusA); SV(_nextEvent); SV(_nextEventClock);
 	SVArray(_workRam, SnesMemoryManager::WorkRamSize);
 	SV(_registerHandlerB);
+
+	if(!s.IsSaving()) {
+		UpdateExecCallbacks();
+	}
 }

Core/SNES/SnesMemoryManager.h

@@ -59,7 +59,14 @@ class SnesMemoryManager : public ISerializable
 	vector<unique_ptr<IMemoryHandler>> _workRamHandlers;
 	uint8_t _masterClockTable[0x800] = {};
 
-	void Exec();
+	typedef void(SnesMemoryManager::*Func)();
+	Func _execRead = nullptr;
+	Func _execWrite = nullptr;
+
+	template<uint8_t clocks> void IncMasterClock();
+	void UpdateExecCallbacks();
+
+	__forceinline void Exec();
 
 	void ProcessEvent();
 

Core/SNES/Spc.Instructions.cpp

@@ -1,7 +1,51 @@
 #include "pch.h"
 #include "SNES/Spc.h"
+#include "SNES/SnesMemoryManager.h"
+#include "Shared/Emulator.h"
 #include "Utilities/HexUtilities.h"
 
+void Spc::Run()
+{
+	if(!_enabled) {
+		//Used to temporarily disable the SPC when overclocking is enabled
+		return;
+	} else if(_state.StopState != SnesCpuStopState::Running) {
+		//STOP or SLEEP were executed - execution is stopped forever.
+		_emu->ProcessHaltedCpu<CpuType::Spc>();
+		return;
+	}
+
+	uint64_t targetCycle = (uint64_t)(_memoryManager->GetMasterClock() * _clockRatio);
+	while(_state.Cycle < targetCycle) {
+		ProcessCycle();
+	}
+}
+
+void Spc::ProcessCycle()
+{
+	if(_opStep == SpcOpStep::ReadOpCode) {
+#ifndef DUMMYSPC
+		_emu->ProcessInstruction<CpuType::Spc>();
+#endif 
+		_opCode = GetOpCode();
+		_opStep = SpcOpStep::Addressing;
+		_opSubStep = 0;
+	} else {
+		Exec();
+	}
+
+	if(_pendingCpuRegUpdate) {
+		//There appears to be a delay between the moment the CPU writes
+		//to a register and when the SPC can see the new value
+		//This makes the SPC see the newly written value after a 1 SPC cycle delay
+		//This allows Kishin Douji Zenki to boot without freezing
+		for(int i = 0; i < 4; i++) {
+			_state.CpuRegs[i] = _state.NewCpuRegs[i];
+		}
+		_pendingCpuRegUpdate = false;
+	}
+}
+
 void Spc::EndOp()
 {
 	_opStep = SpcOpStep::ReadOpCode;

Core/SNES/Spc.cpp

@@ -378,48 +378,6 @@ void Spc::DspWriteRam(uint16_t addr, uint8_t value)
 	_ram[addr] = value;
 }
 
-void Spc::Run()
-{
-	if(!_enabled) {
-		//Used to temporarily disable the SPC when overclocking is enabled
-		return;
-	} else if(_state.StopState != SnesCpuStopState::Running) {
-		//STOP or SLEEP were executed - execution is stopped forever.
-		_emu->ProcessHaltedCpu<CpuType::Spc>();
-		return;
-	}
-
-	uint64_t targetCycle = (uint64_t)(_memoryManager->GetMasterClock() * _clockRatio);
-	while(_state.Cycle < targetCycle) {
-		ProcessCycle();
-	}
-}
-
-void Spc::ProcessCycle()
-{
-	if(_opStep == SpcOpStep::ReadOpCode) {
-#ifndef DUMMYSPC
-		_emu->ProcessInstruction<CpuType::Spc>();
-#endif 
-		_opCode = GetOpCode();
-		_opStep = SpcOpStep::Addressing;
-		_opSubStep = 0;
-	} else {
-		Exec();
-	}
-
-	if(_pendingCpuRegUpdate) {
-		//There appears to be a delay between the moment the CPU writes
-		//to a register and when the SPC can see the new value
-		//This makes the SPC see the newly written value after a 1 SPC cycle delay
-		//This allows Kishin Douji Zenki to boot without freezing
-		for(int i = 0; i < 4; i++) {
-			_state.CpuRegs[i] = _state.NewCpuRegs[i];
-		}
-		_pendingCpuRegUpdate = false;
-	}
-}
-
 void Spc::ProcessEndFrame()
 {
 	Run();

Core/SNES/Spc.h

@@ -279,11 +279,11 @@ class Spc : public ISerializable
 	uint8_t GetOpCode();
 	uint8_t ReadOperandByte();
 
-	void IncCycleCount(int32_t addr);
+	__forceinline void IncCycleCount(int32_t addr);
 	void EndOp();
 	void EndAddr();
-	void ProcessCycle();
-	void Exec();
+	__forceinline void ProcessCycle();
+	__forceinline void Exec();
 
 	void UpdateClockRatio();
 	void ExitExecLoop();