Fixed profiled ReJIT. It's very experimental though. Makes a great performance boost, but is way too unsafe. However it seems to work fine on games I've tested. Can be enabled from ini-file (ProfiledReJIT = True in [Core]).

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@3321 8ced0084-cf51-0410-be5f-012b33b47a6e

Source/Core/Core/Src/ConfigManager.cpp

@@ -223,6 +223,7 @@ void SConfig::LoadSettings()
 		ini.Get("Core", "SlotA",			(int*)&m_EXIDevice[0], EXIDEVICE_MEMORYCARD_A);
 		ini.Get("Core", "SlotB",			(int*)&m_EXIDevice[1], EXIDEVICE_MEMORYCARD_B);
 		ini.Get("Core", "SerialPort1",		(int*)&m_EXIDevice[2], EXIDEVICE_DUMMY);
+		ini.Get("Core", "ProfiledReJIT", &m_LocalCoreStartupParameter.bJITProfiledReJIT, false);
 		char sidevicenum[16];
 		for (int i = 0; i < 4; ++i)
 		{

Source/Core/Core/Src/CoreParameter.h

@@ -55,6 +55,7 @@ struct SCoreStartupParameter
 	bool bJITPairedOff;
 	bool bJITSystemRegistersOff;
 	bool bJITBranchOff;
+	bool bJITProfiledReJIT;
 
 	bool bUseDualCore;
 	bool bDSPThread;

Source/Core/Core/Src/PowerPC/Jit64IL/IR.cpp

@@ -138,6 +138,7 @@ Fix profiled loads/stores to work safely.  On 32-bit, one solution is to
 #include "JitAsm.h"
 #include "Jit.h"
 #include "../../HW/GPFifo.h"
+#include "../../Core.h"
 using namespace Gen;
 
 namespace IREmitter {
@@ -1287,17 +1288,18 @@ static void regWriteExit(RegInfo& RI, InstLoc dest) {
 	if (isImm(*dest)) {
 		RI.Jit->WriteExit(RI.Build->GetImmValue(dest), RI.exitNumber++);
 	} else {
-		RI.Jit->MOV(32, R(EAX), regLocForInst(RI, dest));
+		if (!regLocForInst(RI, dest).IsSimpleReg(EAX))
+			RI.Jit->MOV(32, R(EAX), regLocForInst(RI, dest));
 		RI.Jit->WriteExitDestInEAX(RI.exitNumber++);
 	}
 }
 
-static void DoWriteCode(IRBuilder* ibuild, Jit64* Jit, bool UseProfile) {
+static void DoWriteCode(IRBuilder* ibuild, Jit64* Jit, bool UseProfile, bool MakeProfile) {
 	//printf("Writing block: %x\n", js.blockStart);
 	RegInfo RI(Jit, ibuild->getFirstInst(), ibuild->getNumInsts());
 	RI.Build = ibuild;
 	RI.UseProfile = UseProfile;
-	RI.MakeProfile = false;//!RI.UseProfile;
+	RI.MakeProfile = MakeProfile;
 	// Pass to compute liveness
 	ibuild->StartBackPass();
 	for (unsigned int index = (unsigned int)RI.IInfo.size() - 1; index != -1U; --index) {
@@ -2268,13 +2270,12 @@ static void DoWriteCode(IRBuilder* ibuild, Jit64* Jit, bool UseProfile) {
 }
 
 void Jit64::WriteCode() {
-	DoWriteCode(&ibuild, this, false);
+	DoWriteCode(&ibuild, this, false, Core::GetStartupParameter().bJITProfiledReJIT);
 }
 
 void ProfiledReJit() {
-	u8* x = (u8*)jit.GetCodePtr();
 	jit.SetCodePtr(jit.js.rewriteStart);
-	DoWriteCode(&jit.ibuild, &jit, true);
+	DoWriteCode(&jit.ibuild, &jit, true, false);
 	jit.js.curBlock->codeSize = (int)(jit.GetCodePtr() - jit.js.rewriteStart);
-	jit.SetCodePtr(x);
+	jit.GetBlockCache()->FinalizeBlock(jit.js.curBlock->blockNum, jit.jo.enableBlocklink, jit.js.curBlock->normalEntry);
 }

Source/Core/Core/Src/PowerPC/Jit64IL/Jit.cpp

@@ -428,6 +428,7 @@ namespace CPUCompare
 		SetJumpTarget(skip);
 
 		const u8 *normalEntry = GetCodePtr();
+		b->normalEntry = normalEntry;
 		
 		if (ImHereDebug)
 			ABI_CallFunction((void *)&ImHere); //Used to get a trace of the last few blocks before a crash, sometimes VERY useful

Source/Core/Core/Src/PowerPC/JitCommon/JitCache.cpp

@@ -168,6 +168,7 @@ bool JitBlock::ContainsAddress(u32 em_address)
 		b.exitPtrs[1] = 0;
 		b.linkStatus[0] = false;
 		b.linkStatus[1] = false;
+		b.blockNum = num_blocks;
 		num_blocks++; //commit the current block
 		return num_blocks - 1;
 	}

Source/Core/Core/Src/PowerPC/JitCommon/JitCache.h

@@ -46,6 +46,7 @@ struct JitBlock
 	u32 codeSize; 
 	u32 originalSize;
 	int runCount;  // for profiling.
+	int blockNum;
 
 #ifdef _WIN32
 	// we don't really need to save start and stop
@@ -55,6 +56,7 @@ struct JitBlock
 	LARGE_INTEGER ticCounter;	// for profiling - time.
 #endif
 	const u8 *checkedEntry;
+	const u8 *normalEntry;
 	bool invalid;
 	int flags;