Merge pull request #212 from unknownbrackets/perf

Improve memory access performance a little

rpcs3/Emu/Cell/PPCDecoder.cpp

@@ -3,7 +3,9 @@
 
 u8 PPCDecoder::DecodeMemory(const u64 address)
 {
-	Decode(Memory.Read32(address));
+	u32 instr;
+	Memory.Read32ByAddr(address, &instr);
+	Decode(instr);
 
 	return 4;
 }

rpcs3/Emu/Memory/Memory.cpp

@@ -90,32 +90,17 @@ bool MemoryBlock::IsMyAddress(const u64 addr)
 	return mem && addr >= GetStartAddr() && addr < GetEndAddr();
 }
 
-__forceinline const u8 MemoryBlock::FastRead8(const u64 addr) const
+template <typename T>
+__forceinline const T MemoryBlock::FastRead(const u64 addr) const
 {
-	return *GetMem(addr);
+	volatile const T data = *(const T *)GetMem(addr);
-}
-
-__forceinline const u16 MemoryBlock::FastRead16(const u64 addr) const
-{
-	volatile const u16 data = *(u16*)GetMem(addr);
 	return re(data);
 }
 
-__forceinline const u32 MemoryBlock::FastRead32(const u64 addr) const
+template <>
+__forceinline const u128 MemoryBlock::FastRead<u128>(const u64 addr) const
 {
-	volatile const u32 data = *(u32*)GetMem(addr);
+	volatile const u128 data = *(const u128 *)GetMem(addr);
-	return re(data);
-}
-
-__forceinline const u64 MemoryBlock::FastRead64(const u64 addr) const
-{
-	volatile const u64 data = *(u64*)GetMem(addr);
-	return re(data);
-}
-
-__forceinline const u128 MemoryBlock::FastRead128(const u64 addr)
-{
-	volatile const u128 data = *(u128*)GetMem(addr);
 	u128 ret;
 	ret.lo = re(data.hi);
 	ret.hi = re(data.lo);
@@ -131,7 +116,7 @@ bool MemoryBlock::Read8(const u64 addr, u8* value)
 	}
 
 	//*value = std::atomic_load((volatile std::atomic<u8>*)GetMem(FixAddr(addr)));
-	*value = FastRead8(FixAddr(addr));
+	*value = FastRead<u8>(FixAddr(addr));
 	return true;
 }
 
@@ -144,7 +129,7 @@ bool MemoryBlock::Read16(const u64 addr, u16* value)
 	}
 
 	//se_t<u16>::func(*value, std::atomic_load((volatile std::atomic<u16>*)GetMem(FixAddr(addr))));
-	*value = FastRead16(FixAddr(addr));
+	*value = FastRead<u16>(FixAddr(addr));
 	return true;
 }
 
@@ -157,7 +142,7 @@ bool MemoryBlock::Read32(const u64 addr, u32* value)
 	}
 
 	//se_t<u32>::func(*value, std::atomic_load((volatile std::atomic<u32>*)GetMem(FixAddr(addr))));
-	*value = FastRead32(FixAddr(addr));
+	*value = FastRead<u32>(FixAddr(addr));
 	return true;
 }
 
@@ -170,7 +155,7 @@ bool MemoryBlock::Read64(const u64 addr, u64* value)
 	}
 
 	//se_t<u64>::func(*value, std::atomic_load((volatile std::atomic<u64>*)GetMem(FixAddr(addr))));
-	*value = FastRead64(FixAddr(addr));
+	*value = FastRead<u64>(FixAddr(addr));
 	return true;
 }
 
@@ -185,31 +170,18 @@ bool MemoryBlock::Read128(const u64 addr, u128* value)
 	//u64 f_addr = FixAddr(addr);
 	//se_t<u64>::func(value->lo, std::atomic_load((volatile std::atomic<u64>*)GetMem(f_addr)));
 	//se_t<u64>::func(value->hi, std::atomic_load((volatile std::atomic<u64>*)GetMem(f_addr + 8)));
-	*value = FastRead128(FixAddr(addr));
+	*value = FastRead<u128>(FixAddr(addr));
 	return true;
 }
 
-__forceinline void MemoryBlock::FastWrite8(const u64 addr, const u8 value)
+template <typename T>
+__forceinline void MemoryBlock::FastWrite(const u64 addr, const T value)
 {
-	*GetMem(addr) = value;
+	*(T *)GetMem(addr) = re(value);
 }
 
-__forceinline void MemoryBlock::FastWrite16(const u64 addr, const u16 value)
+template <>
-{
+__forceinline void MemoryBlock::FastWrite<u128>(const u64 addr, const u128 value)
-	*(u16*)GetMem(addr) = re(value);
-}
-
-__forceinline void MemoryBlock::FastWrite32(const u64 addr, const u32 value)
-{
-	*(u32*)GetMem(addr) = re(value);
-}
-
-__forceinline void MemoryBlock::FastWrite64(const u64 addr, const u64 value)
-{
-	*(u64*)GetMem(addr) = re(value);
-}
-
-__forceinline void MemoryBlock::FastWrite128(const u64 addr, const u128 value)
 {
 	u128 res;
 	res.lo = re(value.hi);
@@ -222,7 +194,7 @@ bool MemoryBlock::Write8(const u64 addr, const u8 value)
 	if(!IsMyAddress(addr) || IsLocked(addr)) return false;
 
 	//std::atomic_store((std::atomic<u8>*)GetMem(FixAddr(addr)), value);
-	FastWrite8(FixAddr(addr), value);
+	FastWrite<u8>(FixAddr(addr), value);
 	return true;
 }
 
@@ -233,7 +205,7 @@ bool MemoryBlock::Write16(const u64 addr, const u16 value)
 	//u16 re_value;
 	//se_t<u16>::func(re_value, value);
 	//std::atomic_store((std::atomic<u16>*)GetMem(FixAddr(addr)), re_value);
-	FastWrite16(FixAddr(addr), value);
+	FastWrite<u16>(FixAddr(addr), value);
 	return true;
 }
 
@@ -244,7 +216,7 @@ bool MemoryBlock::Write32(const u64 addr, const u32 value)
 	//u32 re_value;
 	//se_t<u32>::func(re_value, value);
 	//std::atomic_store((std::atomic<u32>*)GetMem(FixAddr(addr)), re_value);
-	FastWrite32(FixAddr(addr), value);
+	FastWrite<u32>(FixAddr(addr), value);
 	return true;
 }
 
@@ -255,7 +227,7 @@ bool MemoryBlock::Write64(const u64 addr, const u64 value)
 	//u64 re_value;
 	//se_t<u64>::func(re_value, value);
 	//std::atomic_store((std::atomic<u64>*)GetMem(FixAddr(addr)), re_value);
-	FastWrite64(FixAddr(addr), value);
+	FastWrite<u64>(FixAddr(addr), value);
 	return true;
 }
 
@@ -269,7 +241,7 @@ bool MemoryBlock::Write128(const u64 addr, const u128 value)
 	//std::atomic_store((std::atomic<u64>*)GetMem(f_addr), re_value);
 	//se_t<u64>::func(re_value, value.hi);
 	//std::atomic_store((std::atomic<u64>*)GetMem(f_addr + 8), re_value);
-	FastWrite128(FixAddr(addr), value);
+	FastWrite<u128>(FixAddr(addr), value);
 	return true;
 }
 
@@ -498,35 +470,35 @@ bool MemoryBase::Write128NN(u64 addr, const u128 data)
 u8 MemoryBase::Read8(u64 addr)
 {
 	u8 res;
-	GetMemByAddr(addr).Read8(addr, &res);
+	Read8ByAddr(addr, &res);
 	return res;
 }
 
 u16 MemoryBase::Read16(u64 addr)
 {
 	u16 res;
-	GetMemByAddr(addr).Read16(addr, &res);
+	Read16ByAddr(addr, &res);
 	return res;
 }
 
 u32 MemoryBase::Read32(u64 addr)
 {
 	u32 res;
-	GetMemByAddr(addr).Read32(addr, &res);
+	Read32ByAddr(addr, &res);
 	return res;
 }
 
 u64 MemoryBase::Read64(u64 addr)
 {
 	u64 res;
-	GetMemByAddr(addr).Read64(addr, &res);
+	Read64ByAddr(addr, &res);
 	return res;
 }
 
 u128 MemoryBase::Read128(u64 addr)
 {
 	u128 res;
-	GetMemByAddr(addr).Read128(addr, &res);
+	Read128ByAddr(addr, &res);
 	return res;
 }
 
@@ -655,9 +627,11 @@ bool VirtualMemoryBlock::Read16(const u64 addr, u16* value)
 
 bool VirtualMemoryBlock::Read32(const u64 addr, u32* value)
 {
-	u64 realAddr;
+	u64 realAddr = getRealAddr(addr);
-	*value = Memory.Read32(realAddr = getRealAddr(addr));
+	if (realAddr == 0)
-	return realAddr != 0;
+		return false;
+	*value = Memory.Read32(realAddr);
+	return true;
 }
 
 bool VirtualMemoryBlock::Read64(const u64 addr, u64* value)

rpcs3/Emu/Memory/Memory.h

@@ -105,14 +105,70 @@ public:
 
 	MemoryBlock& GetMemByAddr(const u64 addr)
 	{
-		for(uint i=0; i<MemoryBlocks.size(); ++i)
+		for (auto block : MemoryBlocks)
 		{
-			if(MemoryBlocks[i]->IsMyAddress(addr)) return *MemoryBlocks[i];
+			if (block->IsMyAddress(addr))
+				return *block;
 		}
 
 		return NullMem;
 	}
 
+	bool Read8ByAddr(const u64 addr, u8 *value)
+	{
+		for (auto block : MemoryBlocks)
+		{
+			if (block->Read8(addr, value))
+				return true;
+		}
+
+		return NullMem.Read8(addr, value);
+	}
+
+	bool Read16ByAddr(const u64 addr, u16 *value)
+	{
+		for (auto block : MemoryBlocks)
+		{
+			if (block->Read16(addr, value))
+				return true;
+		}
+
+		return NullMem.Read16(addr, value);
+	}
+
+	bool Read32ByAddr(const u64 addr, u32 *value)
+	{
+		for (auto block : MemoryBlocks)
+		{
+			if (block->Read32(addr, value))
+				return true;
+		}
+
+		return NullMem.Read32(addr, value);
+	}
+
+	bool Read64ByAddr(const u64 addr, u64 *value)
+	{
+		for (auto block : MemoryBlocks)
+		{
+			if (block->Read64(addr, value))
+				return true;
+		}
+
+		return NullMem.Read64(addr, value);
+	}
+
+	bool Read128ByAddr(const u64 addr, u128 *value)
+	{
+		for (auto block : MemoryBlocks)
+		{
+			if (block->Read128(addr, value))
+				return true;
+		}
+
+		return NullMem.Read128(addr, value);
+	}
+
 	u8* GetMemFromAddr(const u64 addr)
 	{
 		return GetMemByAddr(addr).GetMemFromAddr(addr);
@@ -126,9 +182,9 @@ public:
 	u64 RealToVirtualAddr(const void* addr)
 	{
 		const u64 raddr = (u64)addr;
-		for(u32 i=0; i<MemoryBlocks.size(); ++i)
+		for (auto block : MemoryBlocks)
 		{
-			MemoryBlock& b = *MemoryBlocks[i];
+			MemoryBlock& b = *block;
 			const u64 baddr = (u64)b.GetMem();
 
 			if(raddr >= baddr && raddr < baddr + b.GetSize())
@@ -188,9 +244,10 @@ public:
 
 	bool IsGoodAddr(const u64 addr)
 	{
-		for(uint i=0; i<MemoryBlocks.size(); ++i)
+		for (auto block : MemoryBlocks)
 		{
-			if(MemoryBlocks[i]->IsMyAddress(addr)) return true;
+			if (block->IsMyAddress(addr))
+				return true;
 		}
 
 		return false;
@@ -198,10 +255,11 @@ public:
 
 	bool IsGoodAddr(const u64 addr, const u32 size)
 	{
-		for(uint i=0; i<MemoryBlocks.size(); ++i)
+		const u64 end = addr + size - 1;
+		for (auto block : MemoryBlocks)
 		{
-			if( MemoryBlocks[i]->IsMyAddress(addr) &&
+			if (block->IsMyAddress(addr) && block->IsMyAddress(end))
-				MemoryBlocks[i]->IsMyAddress(addr + size - 1) ) return true;
+				return true;
 		}
 
 		return false;
@@ -214,9 +272,9 @@ public:
 
 		ConLog.Write("Closing memory...");
 
-		for(uint i=0; i<MemoryBlocks.size(); ++i)
+		for (auto block : MemoryBlocks)
 		{
-			MemoryBlocks[i]->Delete();
+			block->Delete();
 		}
 
 		MemoryBlocks.clear();

rpcs3/Emu/Memory/MemoryBlock.h

@@ -105,11 +105,8 @@ public:
 	virtual bool IsMyAddress(const u64 addr);
 	virtual bool IsLocked(const u64 addr) { return false; }
 
-	__forceinline const u8 FastRead8(const u64 addr) const;
+	template <typename T>
-	__forceinline const u16 FastRead16(const u64 addr) const;
+	__forceinline const T FastRead(const u64 addr) const;
-	__forceinline const u32 FastRead32(const u64 addr) const;
-	__forceinline const u64 FastRead64(const u64 addr) const;
-	__forceinline const u128 FastRead128(const u64 addr);
 
 	virtual bool Read8(const u64 addr, u8* value);
 	virtual bool Read16(const u64 addr, u16* value);
@@ -117,11 +114,8 @@ public:
 	virtual bool Read64(const u64 addr, u64* value);
 	virtual bool Read128(const u64 addr, u128* value);
 
-	__forceinline void FastWrite8(const u64 addr, const u8 value);
+	template <typename T>
-	__forceinline void FastWrite16(const u64 addr, const u16 value);
+	__forceinline void FastWrite(const u64 addr, const T value);
-	__forceinline void FastWrite32(const u64 addr, const u32 value);
-	__forceinline void FastWrite64(const u64 addr, const u64 value);
-	__forceinline void FastWrite128(const u64 addr, const u128 value);
 
 	virtual bool Write8(const u64 addr, const u8 value);
 	virtual bool Write16(const u64 addr, const u16 value);
@@ -187,6 +181,7 @@ public:
 
 class NullMemoryBlock : public MemoryBlock
 {
+public:
 	virtual bool IsNULL() { return true; }
 	virtual bool IsMyAddress(const u64 addr) { return true; }