Merge pull request #37 from vailkyte/backport2

Backport Mesen2 fixes

Core/BaseMapper.cpp

@@ -154,18 +154,13 @@ void BaseMapper::SetPpuMemoryMapping(uint16_t startAddr, uint16_t endAddr, uint1
 	uint32_t pageCount = 0;
 	uint32_t pageSize = 0;
 	uint8_t defaultAccessType = MemoryAccessType::Read;
+
+	if(type == ChrMemoryType::Default) {
+		type = _chrRomSize > 0 ? ChrMemoryType::ChrRom : ChrMemoryType::ChrRam;
+	}
+
 	switch(type) {
 		case ChrMemoryType::Default:
-			pageSize = InternalGetChrPageSize();
-			if(pageSize == 0) {
-				return;
-			}
-			pageCount = GetCHRPageCount();
-			if(_onlyChrRam) {
-				defaultAccessType |= MemoryAccessType::Write;
-			}
-			break;
-
 		case ChrMemoryType::ChrRom:
 			pageSize = InternalGetChrPageSize();
 			if(pageSize == 0)
@@ -194,7 +189,6 @@ void BaseMapper::SetPpuMemoryMapping(uint16_t startAddr, uint16_t endAddr, uint1
 	pageNumber = pageNumber % pageCount;
 
 	if((uint16_t)(endAddr - startAddr) >= pageSize) {
-
 		uint32_t addr = startAddr;
 		while(addr <= endAddr - pageSize + 1) {
 			SetPpuMemoryMapping(addr, addr + pageSize - 1, type, pageNumber * pageSize, accessType);
@@ -209,17 +203,22 @@ void BaseMapper::SetPpuMemoryMapping(uint16_t startAddr, uint16_t endAddr, uint1
 void BaseMapper::SetPpuMemoryMapping(uint16_t startAddr, uint16_t endAddr, ChrMemoryType type, uint32_t sourceOffset, int8_t accessType)
 {
 	uint8_t* sourceMemory = nullptr;
+
+	if(type == ChrMemoryType::Default) {
+		type = _chrRomSize > 0 ? ChrMemoryType::ChrRom : ChrMemoryType::ChrRam;
+	}
+
 	switch(type) {
 		default:
 		case ChrMemoryType::Default: 
-			sourceMemory = _onlyChrRam ? _chrRam : _chrRom; 
+		case ChrMemoryType::ChrRom:
-			type = _onlyChrRam ? ChrMemoryType::ChrRam : ChrMemoryType::ChrRom;
+			sourceMemory = _chrRom;
 			break;
 
-		case ChrMemoryType::ChrRom: sourceMemory = _chrRom; break;
 		case ChrMemoryType::ChrRam: sourceMemory = _chrRam; break;
 		case ChrMemoryType::NametableRam: sourceMemory = _nametableRam; break;
 	}
+
 	int firstSlot = startAddr >> 8;
 	int slotCount = (endAddr - startAddr + 1) >> 8;
 	for(int i = 0; i < slotCount; i++) {
@@ -323,6 +322,9 @@ void BaseMapper::SelectCHRPage(uint16_t slot, uint16_t page, ChrMemoryType memor
 	if(memoryType == ChrMemoryType::NametableRam) {
 		pageSize = BaseMapper::NametableSize;
 	} else {
+		if(memoryType == ChrMemoryType::Default) {
+			memoryType = _chrRomSize > 0 ? ChrMemoryType::ChrRom : ChrMemoryType::ChrRam;
+		}
 		pageSize = memoryType == ChrMemoryType::ChrRam ? InternalGetChrRamPageSize() : InternalGetChrPageSize();
 	}
 
@@ -415,7 +417,7 @@ bool BaseMapper::HasChrRam()
 
 bool BaseMapper::HasChrRom()
 {
-	return !_onlyChrRam;
+	return _chrRomSize > 0;
 }
 
 void BaseMapper::AddRegisterRange(uint16_t startAddr, uint16_t endAddr, MemoryOperation operation)
@@ -516,6 +518,7 @@ void BaseMapper::Initialize(RomData &romData)
 
 	_prgRom = new uint8_t[_prgSize];
 	_chrRom = new uint8_t[_chrRomSize];
+
 	memcpy(_prgRom, romData.PrgRom.data(), _prgSize);
 	if(_chrRomSize > 0) {
 		memcpy(_chrRom, romData.ChrRom.data(), _chrRomSize);
@@ -554,12 +557,10 @@ void BaseMapper::Initialize(RomData &romData)
 
 	if(_chrRomSize == 0) {
 		//Assume there is CHR RAM if no CHR ROM exists
-		_onlyChrRam = true;
 		InitializeChrRam(romData.ChrRamSize);
 
 		//Map CHR RAM to 0x0000-0x1FFF by default when no CHR ROM exists
 		SetPpuMemoryMapping(0x0000, 0x1FFF, 0, ChrMemoryType::ChrRam);
-		_chrRomSize = _chrRamSize;
 	} else if(romData.ChrRamSize >= 0) {
 		InitializeChrRam(romData.ChrRamSize);
 	} else if(GetChrRamSize()) {
@@ -852,7 +853,7 @@ uint32_t BaseMapper::GetMemorySize(DebugMemoryType type)
 {
 	switch(type) {
 		default: return 0;
-		case DebugMemoryType::ChrRom: return _onlyChrRam ? 0 : _chrRomSize;
+		case DebugMemoryType::ChrRom: return _chrRomSize;
 		case DebugMemoryType::ChrRam: return _chrRamSize;
 		case DebugMemoryType::NametableRam: return _nametableCount * BaseMapper::NametableSize;
 		case DebugMemoryType::SaveRam: return _saveRamSize;
@@ -1016,7 +1017,7 @@ int32_t BaseMapper::ToAbsoluteChrRomAddress(uint16_t addr)
 
 int32_t BaseMapper::FromAbsoluteChrAddress(uint32_t addr)
 {
-	uint8_t* ptrAddress = (_onlyChrRam ? _chrRam : _chrRom) + (addr & 0x3FFF);
+	uint8_t* ptrAddress = _chrRom + (addr & 0x3FFF);
 
 	for(int i = 0; i < 64; i++) {
 		uint8_t* pageAddress = _chrPages[i];
@@ -1095,14 +1096,14 @@ CartridgeState BaseMapper::GetState()
 	state.HasBattery = _romInfo.HasBattery;
 
 	state.PrgRomSize = _prgSize;
-	state.ChrRomSize = _onlyChrRam ? 0 : _chrRomSize;
+	state.ChrRomSize = _chrRomSize;
 	state.ChrRamSize = _chrRamSize;
 
 	state.PrgPageCount = GetPRGPageCount();
 	state.PrgPageSize = InternalGetPrgPageSize();
 	state.ChrPageCount = GetCHRPageCount();
 	state.ChrPageSize = InternalGetChrPageSize();
-	state.ChrRamPageSize = _onlyChrRam ? InternalGetChrPageSize() : InternalGetChrRamPageSize();
+	state.ChrRamPageSize = InternalGetChrRamPageSize();
 	for(int i = 0; i < 0x100; i++) {
 		state.PrgMemoryOffset[i] = _prgMemoryOffset[i];
 		state.PrgType[i] = _prgMemoryType[i];
@@ -1157,20 +1158,16 @@ void BaseMapper::GetRomFileData(vector<uint8_t> &out, bool asIpsFile, uint8_t* h
 vector<uint8_t> BaseMapper::GetPrgChrCopy()
 {
 	vector<uint8_t> data;
-	data.resize(_prgSize + (_onlyChrRam ? 0 : _chrRomSize));
+	data.resize(_prgSize + _chrRomSize);
 	memcpy(data.data(), _prgRom, _prgSize);
-	if(!_onlyChrRam) {
+	memcpy(data.data() + _prgSize, _chrRom, _chrRomSize);
-		memcpy(data.data() + _prgSize, _chrRom, _chrRomSize);
-	}
 	return data;
 }
 
 void BaseMapper::RestorePrgChrBackup(vector<uint8_t> &backupData)
 {
 	memcpy(_prgRom, backupData.data(), _prgSize);
-	if(!_onlyChrRam) {
+	memcpy(_chrRom, backupData.data() + _prgSize, _chrRomSize);
-		memcpy(_chrRom, backupData.data() + _prgSize, _chrRomSize);
-	}
 }
 
 void BaseMapper::RevertPrgChrChanges()
@@ -1198,8 +1195,6 @@ void BaseMapper::CopyPrgChrRom(shared_ptr<BaseMapper> mapper)
 {
 	if(_prgSize == mapper->_prgSize && _chrRomSize == mapper->_chrRomSize) {
 		memcpy(_prgRom, mapper->_prgRom, _prgSize);
-		if(!_onlyChrRam) {
+		memcpy(_chrRom, mapper->_chrRom, _chrRomSize);
-			memcpy(_chrRom, mapper->_chrRom, _chrRomSize);
-		}
 	}
 }

Core/BaseMapper.h

@@ -26,7 +26,6 @@ private:
 	uint8_t *_nametableRam = nullptr;
 	uint8_t _nametableCount = 2;
 
-	bool _onlyChrRam = false;
 	bool _hasBusConflicts = false;
 	
 	bool _allowRegisterRead = false;
@@ -78,7 +77,7 @@ protected:
 
 	bool IsNes20();
 
-	virtual uint16_t GetChrRamPageSize() { return 0x2000; }
+	virtual uint16_t GetChrRamPageSize() { return GetCHRPageSize(); }
 
 	//Save ram is battery backed and saved to disk
 	virtual uint32_t GetSaveRamSize() { return HasBattery() ? 0x2000 : 0; }

Core/CPU.cpp

@@ -6,6 +6,7 @@
 #include "APU.h"
 #include "DeltaModulationChannel.h"
 #include "MemoryManager.h"
+#include "ControlManager.h"
 #include "Console.h"
 
 CPU::CPU(shared_ptr<Console> console)
@@ -306,14 +307,26 @@ void CPU::ProcessPendingDma(uint16_t readAddress)
 		return;
 	}
 
+	uint16_t prevReadAddress = readAddress;
+	bool enableInternalRegReads = (readAddress & 0xFFE0) == 0x4000;
+	bool skipFirstInputClock = false;
+	if(enableInternalRegReads && _dmcDmaRunning && (readAddress == 0x4016 || readAddress == 0x4017)) {
+		uint16_t dmcAddress = _console->GetApu()->GetDmcReadAddress();
+		if((dmcAddress & 0x1F) == (readAddress & 0x1F)) {
+			//DMC will cause a read on the same address as the CPU was reading from
+			//This will hide the reads from the controllers because /OE will be active the whole time
+			skipFirstInputClock = true;
+		}
+	}
+	
 	//On PAL, the dummy/idle reads done by the DMA don't appear to be done on the
 	//address that the CPU was about to read. This prevents the 2+x reads on registers issues.
 	//The exact specifics of where the CPU reads instead aren't known yet - so just disable read side-effects entirely on PAL
-	bool allowReadSideEffects = _console->GetModel() != NesModel::PAL;
+	bool isNtscInputBehavior = _console->GetModel() != NesModel::PAL;
 
 	//"If this cycle is a read, hijack the read, discard the value, and prevent all other actions that occur on this cycle (PC not incremented, etc)"
 	StartCpuCycle(true);
-	if(allowReadSideEffects) {
+	if(isNtscInputBehavior && !skipFirstInputClock) {
 		_memoryManager->Read(readAddress, MemoryOperationType::DummyRead);
 	}
 	EndCpuCycle(true);
@@ -327,7 +340,7 @@ void CPU::ProcessPendingDma(uint16_t readAddress)
 	//On NES (or AV Famicom), only the first dummy/idle read causes side effects (e.g only a single bit is lost)
 	ConsoleType type = _console->GetSettings()->GetConsoleType();
 	bool isNesBehavior = type != ConsoleType::Famicom;
-	bool skipDummyReads = !allowReadSideEffects || (isNesBehavior && (readAddress == 0x4016 || readAddress == 0x4017));
+	bool skipDummyReads = !isNtscInputBehavior || (isNesBehavior && (readAddress == 0x4016 || readAddress == 0x4017));
 
 	auto processCycle = [this] {
 		//Sprite DMA cycles count as halt/dummy cycles for the DMC DMA when both run at the same time
@@ -345,14 +358,14 @@ void CPU::ProcessPendingDma(uint16_t readAddress)
 			if(_dmcDmaRunning && !_needHalt && !_needDummyRead) {
 				//DMC DMA is ready to read a byte (both halt and dummy read cycles were performed before this)
 				processCycle();
-				readValue = _memoryManager->Read(_console->GetApu()->GetDmcReadAddress(), MemoryOperationType::DmcRead);
+				readValue = ProcessDmaRead(_console->GetApu()->GetDmcReadAddress(), prevReadAddress, enableInternalRegReads, isNesBehavior);
 				EndCpuCycle(true); 
 				_console->GetApu()->SetDmcReadBuffer(readValue);
 				_dmcDmaRunning = false;
 			} else if(_spriteDmaTransfer) {
 				//DMC DMA is not running, or not ready, run sprite DMA
 				processCycle();
-				readValue = _memoryManager->Read(_spriteDmaOffset * 0x100 + spriteReadAddr);
+				readValue = ProcessDmaRead(_spriteDmaOffset * 0x100 + spriteReadAddr, prevReadAddress, enableInternalRegReads, isNesBehavior);
 				EndCpuCycle(true);
 				spriteReadAddr++;
 				spriteDmaCounter++;
@@ -361,7 +374,7 @@ void CPU::ProcessPendingDma(uint16_t readAddress)
 				assert(_needHalt || _needDummyRead);
 				processCycle();
 				if(!skipDummyReads) {
-					_memoryManager->Read(readAddress, MemoryOperationType::DummyRead);
+					ProcessDmaRead(readAddress, prevReadAddress, enableInternalRegReads, isNesBehavior);
 				}
 				EndCpuCycle(true);
 			}
@@ -379,7 +392,7 @@ void CPU::ProcessPendingDma(uint16_t readAddress)
 				//Align to read cycle before starting sprite DMA (or align to perform DMC read)
 				processCycle();
 				if(!skipDummyReads) {
-					_memoryManager->Read(readAddress, MemoryOperationType::DummyRead);
+					ProcessDmaRead(readAddress, prevReadAddress, enableInternalRegReads, isNesBehavior);
 				}
 				EndCpuCycle(true);
 			}
@@ -387,6 +400,77 @@ void CPU::ProcessPendingDma(uint16_t readAddress)
 	}
 }
 
+uint8_t CPU::ProcessDmaRead(uint16_t addr, uint16_t& prevReadAddress, bool enableInternalRegReads, bool isNesBehavior)
+{
+	//This is to reproduce a CPU bug that can occur during DMA which can cause the 2A03 to read from
+	//its internal registers (4015, 4016, 4017) at the same time as the DMA unit reads a byte from 
+	//the bus. This bug occurs if the CPU is halted while it's reading a value in the $4000-$401F range.
+	//
+	//This has a number of side effects:
+	// -It can cause a read of $4015 to occur without the program's knowledge, which would clear the frame counter's IRQ flag
+	// -It can cause additional bit deletions while reading the input (e.g more than the DMC glitch usually causes)
+	// -It can also *prevent* bit deletions from occurring at all in another scenario
+	// -It can replace/corrupt the byte that the DMA is reading, causing DMC to play the wrong sample
+
+	uint8_t val;
+	if(!enableInternalRegReads) {
+		if(addr >= 0x4000 && addr <= 0x401F) {
+			//Nothing will respond on $4000-$401F on the external bus - return open bus value
+			val = _memoryManager->GetOpenBus();
+		} else {
+			val = _memoryManager->Read(addr, MemoryOperationType::DmcRead);
+			
+		}
+		prevReadAddress = addr;
+		return val;
+	} else {
+		//This glitch causes the CPU to read from the internal APU/Input registers
+		//regardless of the address the DMA unit is trying to read
+		uint16_t internalAddr = 0x4000 | (addr & 0x1F);
+		bool isSameAddress = internalAddr == addr;
+
+		switch(internalAddr) {
+			case 0x4015:
+				val = _memoryManager->Read(internalAddr, MemoryOperationType::DmcRead);
+				if(!isSameAddress) {
+					//Also trigger a read from the actual address the CPU was supposed to read from (external bus)
+					_memoryManager->Read(addr, MemoryOperationType::DmcRead);
+				}
+				break;
+
+			case 0x4016:
+			case 0x4017:
+				if(_console->GetModel() == NesModel::PAL || (isNesBehavior && prevReadAddress == internalAddr)) {
+					//Reading from the same input register twice in a row, skip the read entirely to avoid
+					//triggering a bit loss from the read, since the controller won't react to this read
+					//Return the same value as the last read, instead
+					//On PAL, the behavior is unknown - for now, don't cause any bit deletions
+					val = _memoryManager->GetOpenBus();
+				} else {
+					val = _memoryManager->Read(internalAddr, MemoryOperationType::DmcRead);
+				}
+
+				if(!isSameAddress) {
+					//The DMA unit is reading from a different address, read from it too (external bus)
+					uint8_t obMask = _console->GetControlManager()->GetOpenBusMask(internalAddr - 0x4016);
+					uint8_t externalValue = _memoryManager->Read(addr, MemoryOperationType::DmcRead);
+
+					//Merge values, keep the external value for all open bus pins on the 4016/4017 port
+					//AND all other bits together (bus conflict)
+					val = (externalValue & obMask) | ((val & ~obMask) & (externalValue & ~obMask));
+				}
+				break;
+
+			default:
+				val = _memoryManager->Read(addr, MemoryOperationType::DmcRead);
+				break;
+		}
+
+		prevReadAddress = internalAddr;
+		return val;
+	}
+}
+
 void CPU::RunDMATransfer(uint8_t offsetValue)
 {
 	_spriteDmaTransfer = true;

Core/CPU.h

@@ -80,6 +80,7 @@ private:
 
 	__forceinline void StartCpuCycle(bool forRead);
 	__forceinline void ProcessPendingDma(uint16_t readAddress);
+	uint8_t ProcessDmaRead(uint16_t addr, uint16_t& prevReadAddress, bool enableInternalRegReads, bool isNesBehavior);
 	__forceinline uint16_t FetchOperand();
 	__forceinline void EndCpuCycle(bool forRead);
 	void IRQ();

Core/ControlManager.h

@@ -39,12 +39,13 @@ protected:
 	virtual void StreamState(bool saving) override;
 	virtual ControllerType GetControllerType(uint8_t port);
 	virtual void RemapControllerButtons();
-	virtual uint8_t GetOpenBusMask(uint8_t port);
 
 public:
 	ControlManager(std::shared_ptr<Console> console, std::shared_ptr<BaseControlDevice> systemActionManager, std::shared_ptr<BaseControlDevice> mapperControlDevice);
 	virtual ~ControlManager();
 
+	virtual uint8_t GetOpenBusMask(uint8_t port);
+
 	virtual void UpdateControlDevices();
 	void UpdateInputState();
 

Core/PPU.cpp

@@ -983,19 +983,16 @@ void PPU::ProcessScanline()
 			}
 		}
 	} else if(_cycle >= 321 && _cycle <= 336) {
+		LoadTileInfo();
 		if(_cycle == 321) {
 			if(IsRenderingEnabled()) {
 				LoadExtraSprites();
 				_oamCopybuffer = _secondarySpriteRAM[0];
 			}
-			LoadTileInfo();
 		} else if(_prevRenderingEnabled && (_cycle == 328 || _cycle == 336)) {
-			LoadTileInfo();
 			_state.LowBitShift <<= 8;
 			_state.HighBitShift <<= 8;
 			IncHorizontalScrolling();
-		} else {
-			LoadTileInfo();
 		}
 	} else if(_cycle == 337 || _cycle == 339) {
 		if(IsRenderingEnabled()) {
@@ -1247,52 +1244,8 @@ void PPU::ProcessOamCorruption()
 
 void PPU::Exec()
 {
-	if(_cycle > 339) {
+	if(_cycle < 340) {
-		_cycle = 0;
+		//Process cycles 1 to 340
-		if(++_scanline > _vblankEnd) {
-			_lastUpdatedPixel = -1;
-			_scanline = -1;
-
-			//Force prerender scanline sprite fetches to load the dummy $FF tiles (fixes shaking in Ninja Gaiden 3 stage 1 after beating boss)
-			_spriteCount = 0;
-
-			if(_renderingEnabled) {
-				ProcessOamCorruption();
-			}
-
-			UpdateMinimumDrawCycles();
-		}
-
-		UpdateApuStatus();
-		
-		if(_scanline == _settings->GetInputPollScanline()) {
-			_console->GetControlManager()->UpdateInputState();
-		}
-
-		//Cycle = 0
-		if(_scanline < 240) {
-			if(_scanline == -1) {
-				_statusFlags.SpriteOverflow = false;
-				_statusFlags.Sprite0Hit = false;
-
-				//Switch to alternate output buffer (VideoDecoder may still be decoding the last frame buffer)
-				_currentOutputBuffer = (_currentOutputBuffer == _outputBuffers[0]) ? _outputBuffers[1] : _outputBuffers[0];
-			} else if(_prevRenderingEnabled) {
-				if(_scanline > 0 || (!(_frameCount & 0x01) || _nesModel != NesModel::NTSC || _settings->GetPpuModel() != PpuModel::Ppu2C02)) {
-					//Set bus address to the tile address calculated from the unused NT fetches at the end of the previous scanline
-					//This doesn't happen on scanline 0 if the last dot of the previous frame was skipped
-					SetBusAddress((_nextTile.TileAddr << 4) | (_state.VideoRamAddr >> 12) | _flags.BackgroundPatternAddr);
-				}
-			}
-		} else if(_scanline == 240) {
-			//At the start of vblank, the bus address is set back to VideoRamAddr.
-			//According to Visual NES, this occurs on scanline 240, cycle 1, but is done here on cycle for performance reasons
-			SetBusAddress(_state.VideoRamAddr);
-			SendFrame();
-			_frameCount++;
-		}
-	} else {
-		//Cycle > 0
 		_cycle++;
 
 		if(_scanline < 240) {
@@ -1314,6 +1267,9 @@ void PPU::Exec()
 				_state.SpriteRamAddr = 0;
 			}
 		}
+	} else {
+		//Process cycle 0
+		ProcessScanlineFirstCycle();
 	}
 
 	if(_needStateUpdate) {
@@ -1321,6 +1277,53 @@ void PPU::Exec()
 	}
 }
 
+void PPU::ProcessScanlineFirstCycle()
+{
+	_cycle = 0;
+	if(++_scanline > _vblankEnd) {
+		_lastUpdatedPixel = -1;
+		_scanline = -1;
+
+		//Force prerender scanline sprite fetches to load the dummy $FF tiles (fixes shaking in Ninja Gaiden 3 stage 1 after beating boss)
+		_spriteCount = 0;
+
+		if(_renderingEnabled) {
+			ProcessOamCorruption();
+		}
+
+		UpdateMinimumDrawCycles();
+	}
+
+	UpdateApuStatus();
+		
+	if(_scanline == _settings->GetInputPollScanline()) {
+		_console->GetControlManager()->UpdateInputState();
+	}
+
+	//Cycle = 0
+	if(_scanline < 240) {
+		if(_scanline == -1) {
+			_statusFlags.SpriteOverflow = false;
+			_statusFlags.Sprite0Hit = false;
+
+			//Switch to alternate output buffer (VideoDecoder may still be decoding the last frame buffer)
+			_currentOutputBuffer = (_currentOutputBuffer == _outputBuffers[0]) ? _outputBuffers[1] : _outputBuffers[0];
+		} else if(_prevRenderingEnabled) {
+			if(_scanline > 0 || (!(_frameCount & 0x01) || _nesModel != NesModel::NTSC || _settings->GetPpuModel() != PpuModel::Ppu2C02)) {
+				//Set bus address to the tile address calculated from the unused NT fetches at the end of the previous scanline
+				//This doesn't happen on scanline 0 if the last dot of the previous frame was skipped
+				SetBusAddress((_nextTile.TileAddr << 4) | (_state.VideoRamAddr >> 12) | _flags.BackgroundPatternAddr);
+			}
+		}
+	} else if(_scanline == 240) {
+		//At the start of vblank, the bus address is set back to VideoRamAddr.
+		//According to Visual NES, this occurs on scanline 240, cycle 1, but is done here on cycle for performance reasons
+		SetBusAddress(_state.VideoRamAddr);
+		SendFrame();
+		_frameCount++;
+	}
+}
+
 void PPU::UpdateState()
 {
 	_needStateUpdate = false;

Core/PPU.h

@@ -129,16 +129,17 @@ class PPU : public IMemoryHandler, public Snapshotable
 		uint16_t GetNameTableAddr();
 		uint16_t GetAttributeAddr();
 
+		void ProcessScanlineFirstCycle(); 
 		__forceinline void ProcessScanline();
 		__forceinline void ProcessSpriteEvaluation();
 
 		void BeginVBlank();
 		void TriggerNmi();
 
-		void LoadTileInfo();
+		__forceinline void LoadTileInfo();
-		void LoadSprite(uint8_t spriteY, uint8_t tileIndex, uint8_t attributes, uint8_t spriteX, bool extraSprite);
+		__forceinline void LoadSprite(uint8_t spriteY, uint8_t tileIndex, uint8_t attributes, uint8_t spriteX, bool extraSprite);
-		void LoadSpriteTileInfo();
+		__forceinline void LoadSpriteTileInfo();
-		void LoadExtraSprites();
+		__forceinline void LoadExtraSprites();
 		__forceinline void ShiftTileRegisters();
 
 		__forceinline uint8_t ReadSpriteRam(uint8_t addr);

Core/SoundMixer.h

@@ -73,8 +73,8 @@ private:
 
 	double _previousTargetRate;
 
-	double GetChannelOutput(AudioChannel channel, bool forRightChannel);
+	__forceinline double GetChannelOutput(AudioChannel channel, bool forRightChannel);
-	int16_t GetOutputVolume(bool forRightChannel);
+	__forceinline int16_t GetOutputVolume(bool forRightChannel);
 	void EndFrame(uint32_t time);
 
 	void UpdateRates(bool forceUpdate);