3 more activated instructions in the jit = maybe 1% more speed ;P

git-svn-id: https://dolphin-emu.googlecode.com/svn/trunk@163 8ced0084-cf51-0410-be5f-012b33b47a6e

Source/Core/Core/Src/PowerPC/Interpreter/Interpreter_Integer.cpp

@@ -391,7 +391,7 @@ void CInterpreter::srawix(UGeckoInstruction _inst)
 void CInterpreter::srwx(UGeckoInstruction _inst)
 {
 	u32 amount = m_GPR[_inst.RB];
-	m_GPR[_inst.RA] = (amount&0x20) ? 0 : m_GPR[_inst.RS] >> amount;
+	m_GPR[_inst.RA] = (amount & 0x20) ? 0 : (m_GPR[_inst.RS] >> amount);
 
 	if (_inst.Rc) Helper_UpdateCR0(m_GPR[_inst.RA]);
 }

Source/Core/Core/Src/PowerPC/Jit64/Jit.h

@@ -147,6 +147,8 @@ namespace Jit64
 	void rlwinmx(UGeckoInstruction inst);
 	void rlwimix(UGeckoInstruction inst);
 	void rlwnmx(UGeckoInstruction inst);
+	void negx(UGeckoInstruction inst);
+	void slwx(UGeckoInstruction inst);
 	void srwx(UGeckoInstruction inst);
 	void dcbz(UGeckoInstruction inst);
 	void lfsx(UGeckoInstruction inst);

Source/Core/Core/Src/PowerPC/Jit64/Jit_Integer.cpp

@@ -67,7 +67,7 @@ namespace Jit64
 		}
 		else
 		{
-			_assert_msg_(DYNA_REC,0,"WTF");
+			_assert_msg_(DYNA_REC, 0, "WTF");
 		}
 		if (Rc)
 		{
@@ -93,10 +93,10 @@ namespace Jit64
 		case 25: regimmop(a, s, true, inst.UIMM << 16, Or,  OR, false); break;//oris
 		case 28: regimmop(a, s, true, inst.UIMM,       And, AND, true); break;
 		case 29: regimmop(a, s, true, inst.UIMM << 16, And, AND, true); break;
+		case 26: regimmop(a, s, true, inst.UIMM,       Xor, XOR, false); break; //xori
+		case 27: regimmop(a, s, true, inst.UIMM << 16, Xor, XOR, false); break; //xoris
 		case 12: //addic
 		case 13: //addic_rc
-		case 26: //xori
-		case 27: //xoris
 		default:
 			Default(inst);
 			break;
@@ -633,7 +633,6 @@ namespace Jit64
 		}
 
 		u32 mask = Helper_Mask(inst.MB, inst.ME);
-		
 		gpr.FlushR(ECX);
 		gpr.LockX(ECX);
 		gpr.Lock(a, b, s);
@@ -652,21 +651,34 @@ namespace Jit64
 		}
 	}
 
+	void negx(UGeckoInstruction inst)
+	{
+		INSTRUCTION_START;
+		int a = inst.RA;
+		int d = inst.RD;
+		gpr.Lock(a, d);
+		gpr.LoadToX64(d, a == d, true);
+		if (a != d)
+			MOV(32, gpr.R(d), gpr.R(a));
+		NEG(32, gpr.R(d));
+		gpr.UnlockAll();
+		if (inst.Rc)
+		{
+			MOV(32, R(EAX), gpr.R(a));
+			CALL((u8*)Asm::computeRc);
+		}
+	}
+
 	void srwx(UGeckoInstruction inst)
 	{
 		INSTRUCTION_START;
-		// BUGGY?
-		Default(inst); return;
 		int a = inst.RA;
 		int b = inst.RB;
 		int s = inst.RS;
 		gpr.FlushR(ECX);
 		gpr.LockX(ECX);
 		gpr.Lock(a, b, s);
-		if (a == s)
+		gpr.LoadToX64(a, a == s || a == b || s == b, true);
-			gpr.LoadToX64(a, true, true);
-		else
-			gpr.LoadToX64(a, false, true);
 		MOV(32, R(ECX), gpr.R(b));
 		AND(32, R(ECX), Imm8(63));
 		XOR(32, R(EAX), R(EAX));
@@ -685,6 +697,34 @@ namespace Jit64
 		}
 	}
 
+	void slwx(UGeckoInstruction inst)
+	{
+		INSTRUCTION_START;
+		int a = inst.RA;
+		int b = inst.RB;
+		int s = inst.RS;
+		gpr.FlushR(ECX);
+		gpr.LockX(ECX);
+		gpr.Lock(a, b, s);
+		gpr.LoadToX64(a, a == s || a == b || s == b, true);
+		MOV(32, R(ECX), gpr.R(b));
+		AND(32, R(ECX), Imm8(63));
+		XOR(32, R(EAX), R(EAX));
+		TEST(32, R(ECX), Imm32(32));
+		FixupBranch branch = J_CC(CC_NZ);
+		MOV(32, R(EAX), gpr.R(s));
+		SHL(32, R(EAX), R(ECX));
+		SetJumpTarget(branch);
+		MOV(32, gpr.R(a), R(EAX));
+		gpr.UnlockAll();
+		gpr.UnlockAllX();
+		if (inst.Rc) 
+		{
+			MOV(32, R(EAX), gpr.R(a));
+			CALL((u8*)Asm::computeRc);
+		}
+	}
+
 	// another crazy instruction :(
 	void srawix(UGeckoInstruction inst)
 	{

Source/Core/Core/Src/PowerPC/Jit64/Jit_LoadStoreFloating.cpp

@@ -93,6 +93,7 @@ void lfs(UGeckoInstruction inst)
 	gpr.UnlockAll();
 	fpr.UnlockAll();
 }
+
 void lfd(UGeckoInstruction inst)
 {
 	INSTRUCTION_START;
@@ -112,8 +113,7 @@ void lfd(UGeckoInstruction inst)
 	MOV(64, M(&temp64), R(EAX));
 	fpr.Lock(d);
 	fpr.LoadToX64(d, false);
-	MOVSD(fpr.RX(d), M(&temp64));
+	MOVDDUP(fpr.RX(d), M(&temp64));
-	MOVDDUP(fpr.RX(d), fpr.R(d));
 	gpr.UnlockAll();
 	fpr.UnlockAll();
 }

Source/Core/Core/Src/PowerPC/PPCTables.cpp

@@ -155,8 +155,8 @@ GekkoOPTemplate primarytable[] =
 
 	{24, CInterpreter::ori,       Jit64::reg_imm,      {"ori",      OPTYPE_INTEGER, 0}},
 	{25, CInterpreter::oris,      Jit64::reg_imm,      {"oris",     OPTYPE_INTEGER, 0}},
-	{26, CInterpreter::xori,      Jit64::Default,      {"xori",     OPTYPE_INTEGER, 0}},
+	{26, CInterpreter::xori,      Jit64::reg_imm,      {"xori",     OPTYPE_INTEGER, 0}},
-	{27, CInterpreter::xoris,     Jit64::Default,      {"xoris",    OPTYPE_INTEGER, 0}},
+	{27, CInterpreter::xoris,     Jit64::reg_imm,      {"xoris",    OPTYPE_INTEGER, 0}},
 	{28, CInterpreter::andi_rc,   Jit64::reg_imm,      {"andi_rc",  OPTYPE_INTEGER, FL_SET_CR0}},
 	{29, CInterpreter::andis_rc,  Jit64::reg_imm,      {"andis_rc", OPTYPE_INTEGER, FL_SET_CR0}},
 
@@ -291,7 +291,7 @@ GekkoOPTemplate table31[] =
 	{536, CInterpreter::srwx,    Jit64::srwx,     {"srwx",   OPTYPE_INTEGER, FL_RC_BIT}},
 	{792, CInterpreter::srawx,   Jit64::Default,  {"srawx",  OPTYPE_INTEGER, FL_RC_BIT}},
 	{824, CInterpreter::srawix,  Jit64::srawix,   {"srawix", OPTYPE_INTEGER, FL_RC_BIT}},
-	{24,  CInterpreter::slwx,    Jit64::Default,  {"slwx",   OPTYPE_INTEGER, FL_RC_BIT}},
+	{24,  CInterpreter::slwx,    Jit64::slwx,     {"slwx",   OPTYPE_INTEGER, FL_RC_BIT}},
 
 	{54,   CInterpreter::dcbst,  Jit64::Default,  {"dcbst",  OPTYPE_DCACHE, 0, 4}},
 	{86,   CInterpreter::dcbf,   Jit64::Default,  {"dcbf",   OPTYPE_DCACHE, 0, 4}},
@@ -396,7 +396,7 @@ GekkoOPTemplate table31_2[] =
 	{75,  CInterpreter::mulhwx,  Jit64::Default,      {"mulhwx",  OPTYPE_INTEGER, FL_OUT_D | FL_IN_A | FL_IN_B | FL_RC_BIT, 4}},
 	{11,  CInterpreter::mulhwux, Jit64::mulhwux,      {"mulhwux", OPTYPE_INTEGER, FL_OUT_D | FL_IN_A | FL_IN_B | FL_RC_BIT, 4}},
 	{235, CInterpreter::mullwx,  Jit64::mullwx,       {"mullwx",  OPTYPE_INTEGER, FL_OUT_D | FL_IN_A | FL_IN_B | FL_RC_BIT, 4}},
-	{104, CInterpreter::negx,    Jit64::Default,      {"negx",    OPTYPE_INTEGER, FL_OUT_D | FL_IN_A | FL_IN_B | FL_RC_BIT}},
+	{104, CInterpreter::negx,    Jit64::negx,         {"negx",    OPTYPE_INTEGER, FL_OUT_D | FL_IN_A | FL_IN_B | FL_RC_BIT}},
 	{40,  CInterpreter::subfx,   Jit64::subfx,        {"subfx",   OPTYPE_INTEGER, FL_OUT_D | FL_IN_A | FL_IN_B | FL_RC_BIT}},
 	{8,   CInterpreter::subfcx,  Jit64::Default,      {"subfcx",  OPTYPE_INTEGER, FL_OUT_D | FL_IN_A | FL_IN_B | FL_SET_CA | FL_RC_BIT}},
 	{136, CInterpreter::subfex,  Jit64::Default,      {"subfex",  OPTYPE_INTEGER, FL_OUT_D | FL_IN_A | FL_IN_B | FL_READ_CA | FL_SET_CA | FL_RC_BIT}},