LLE JIT: Added simple block linking at immediate CALL's and JUMP's. The code checks if the block being jump to has already been compiled and jumps there if it has, bypassing the dispatcher. This results in a speed-up of around 7 - 10%.

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@6585 8ced0084-cf51-0410-be5f-012b33b47a6e

Source/Core/DSPCore/Src/DSPEmitter.cpp

@@ -39,6 +39,7 @@ DSPEmitter::DSPEmitter() : storeIndex(-1), storeIndex2(-1)
 	AllocCodeSpace(COMPILED_CODE_SIZE);
 
 	blocks = new CompiledCode[MAX_BLOCKS];
+	blockLinks = new CompiledCode[MAX_BLOCKS];
 	blockSize = new u16[0x10000];
 	
 	compileSR = 0;
@@ -52,6 +53,7 @@ DSPEmitter::DSPEmitter() : storeIndex(-1), storeIndex2(-1)
 	for(int i = 0x0000; i < MAX_BLOCKS; i++)
 	{
 		blocks[i] = (CompiledCode)stubEntryPoint;
+		blockLinks[i] = 0;
 		blockSize[i] = 0;
 	}
 }
@@ -59,6 +61,7 @@ DSPEmitter::DSPEmitter() : storeIndex(-1), storeIndex2(-1)
 DSPEmitter::~DSPEmitter() 
 {
 	delete[] blocks;
+	delete[] blockLinks;
 	delete[] blockSize;
 	FreeCodeSpace();
 }
@@ -68,6 +71,7 @@ void DSPEmitter::ClearIRAM() {
 	for(int i = 0x0000; i < 0x1000; i++)
 	{
 		blocks[i] = (CompiledCode)stubEntryPoint;
+		blockLinks[i] = 0;
 		blockSize[i] = 0;
 	}
 }
@@ -202,6 +206,9 @@ void DSPEmitter::Compile(int start_addr)
 	if (g_dsp.exceptions == 0)
 		return;	
 	*/
+
+	blockLinkEntry = GetCodePtr();
+
 	ABI_CallFunction((void *)&DSPCore_CheckExternalInterrupt);
 
 	compilePC = start_addr;
@@ -327,6 +334,7 @@ void DSPEmitter::Compile(int start_addr)
 	}
 
 	blocks[start_addr] = (CompiledCode)entryPoint;
+	blockLinks[start_addr] = (CompiledCode)blockLinkEntry;
 	if (blockSize[start_addr] == 0) 
 	{
 		// just a safeguard, should never happen anymore.

Source/Core/DSPCore/Src/DSPEmitter.h

@@ -156,9 +156,11 @@ public:
 	// CALL this to start the dispatcher
 	const u8 *enterDispatcher;
 	u16 compilePC;
+	CompiledCode *blockLinks;
 
 private:
 	CompiledCode *blocks;
+	const u8 *blockLinkEntry;
 	u16 *blockSize;
 	u16 compileSR;
 

Source/Core/DSPCore/Src/Jit/DSPJitBranch.cpp

@@ -139,9 +139,24 @@ void r_jcc(const UDSPInstruction opc, DSPEmitter& emitter)
 	u16 dest = dsp_imem_read(emitter.compilePC + 1);
 #ifdef _M_IX86 // All32
 	emitter.MOV(16, M(&(g_dsp.pc)), Imm16(dest));
+
+	// Jump directly to the called block if it has already been compiled.
+	// TODO: Subtract cycles from cyclesLeft
+	if (emitter.blockLinks[dest])
+	{
+		emitter.JMPptr(M(&emitter.blockLinks[dest]));
+	}
 #else
 	emitter.MOV(64, R(RAX), ImmPtr(&(g_dsp.pc)));
 	emitter.MOV(16, MDisp(RAX,0), Imm16(dest));
+
+	// Jump directly to the next block if it has already been compiled.
+	// TODO: Subtract cycles from cyclesLeft
+	if (emitter.blockLinks[dest])
+	{
+		emitter.MOV(64, R(RAX), ImmPtr(emitter.blockLinks[dest]));
+		emitter.JMPptr(R(RAX));
+	}
 #endif
 }
 // Generic jmp implementation
@@ -190,9 +205,24 @@ void r_call(const UDSPInstruction opc, DSPEmitter& emitter)
 	emitter.ABI_CallFunctionCC16((void *)dsp_reg_store_stack, DSP_STACK_C, emitter.compilePC + 2);
 #ifdef _M_IX86 // All32
 	emitter.MOV(16, M(&(g_dsp.pc)), Imm16(dest));
+
+	// Jump directly to the called block if it has already been compiled.
+	// TODO: Subtract cycles from cyclesLeft
+	if (emitter.blockLinks[dest])
+	{
+		emitter.JMPptr(M(&emitter.blockLinks[dest]));
+	}
 #else
 	emitter.MOV(64, R(RAX), ImmPtr(&(g_dsp.pc)));
 	emitter.MOV(16, MDisp(RAX,0), Imm16(dest));
+
+	// Jump directly to the called block if it has already been compiled.
+	// TODO: Subtract cycles from cyclesLeft
+	if (emitter.blockLinks[dest])
+	{
+		emitter.MOV(64, R(RAX), ImmPtr(emitter.blockLinks[dest]));
+		emitter.JMPptr(R(RAX));
+	}
 #endif
 }
 // Generic call implementation