JIT64: optimize sign-extend

Remove some code duplication.

Also remove some comments that no longer apply since x86_32 was dropped.

3 files changed, 20 insertions, 40 deletions

diff --git a/Source/Core/Core/PowerPC/Jit64/Jit.h b/Source/Core/Core/PowerPC/Jit64/Jit.h
index 0b726521a6..a2b261d7b3 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit.h
+++ b/Source/Core/Core/PowerPC/Jit64/Jit.h
@@ -153,8 +153,7 @@ public:
 	void addmex(UGeckoInstruction inst);
 	void addzex(UGeckoInstruction inst);
 
-	void extsbx(UGeckoInstruction inst);
-	void extshx(UGeckoInstruction inst);
+	void extsXx(UGeckoInstruction inst);
 
 	void sc(UGeckoInstruction _inst);
 	void rfi(UGeckoInstruction _inst);
diff --git a/Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp b/Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp
index fa7c19aec8..927e83353f 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp
+++ b/Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp
@@ -190,8 +190,8 @@ static GekkoOPTemplate table31[] =
 	{0,   &Jit64::cmpXX},                  //"cmp",    OPTYPE_INTEGER, FL_IN_AB | FL_SET_CRn}},
 	{32,  &Jit64::cmpXX},                  //"cmpl",   OPTYPE_INTEGER, FL_IN_AB | FL_SET_CRn}},
 	{26,  &Jit64::cntlzwx},                //"cntlzwx",OPTYPE_INTEGER, FL_OUT_A | FL_IN_S | FL_RC_BIT}},
-	{922, &Jit64::extshx},                 //"extshx", OPTYPE_INTEGER, FL_OUT_A | FL_IN_S | FL_RC_BIT}},
-	{954, &Jit64::extsbx},                 //"extsbx", OPTYPE_INTEGER, FL_OUT_A | FL_IN_S | FL_RC_BIT}},
+	{922, &Jit64::extsXx},                 //"extshx", OPTYPE_INTEGER, FL_OUT_A | FL_IN_S | FL_RC_BIT}},
+	{954, &Jit64::extsXx},                 //"extsbx", OPTYPE_INTEGER, FL_OUT_A | FL_IN_S | FL_RC_BIT}},
 	{536, &Jit64::srwx},                   //"srwx",   OPTYPE_INTEGER, FL_OUT_A | FL_IN_B | FL_IN_S | FL_RC_BIT}},
 	{792, &Jit64::srawx},                  //"srawx",  OPTYPE_INTEGER, FL_OUT_A | FL_IN_B | FL_IN_S | FL_RC_BIT}},
 	{824, &Jit64::srawix},                 //"srawix", OPTYPE_INTEGER, FL_OUT_A | FL_IN_B | FL_IN_S | FL_RC_BIT}},
diff --git a/Source/Core/Core/PowerPC/Jit64/Jit_Integer.cpp b/Source/Core/Core/PowerPC/Jit64/Jit_Integer.cpp
index ec5f68cd18..01212e3065 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit_Integer.cpp
+++ b/Source/Core/Core/PowerPC/Jit64/Jit_Integer.cpp
@@ -796,56 +796,37 @@ void Jit64::boolX(UGeckoInstruction inst)
 	}
 }
 
-void Jit64::extsbx(UGeckoInstruction inst)
+void Jit64::extsXx(UGeckoInstruction inst)
 {
 	INSTRUCTION_START
 	JITDISABLE(bJITIntegerOff);
 	int a = inst.RA, s = inst.RS;
+	int size = inst.SUBOP10 == 922 ? 16 : 8;
 
 	if (gpr.R(s).IsImm())
 	{
-		gpr.SetImmediate32(a, (u32)(s32)(s8)gpr.R(s).offset);
-	}
-	else
-	{
-		gpr.Lock(a, s);
-		gpr.BindToRegister(a, a == s, true);
-		MOVSX(32, 8, gpr.RX(a), gpr.R(s));
-		gpr.UnlockAll();
-	}
-
-	if (inst.Rc)
-	{
-		ComputeRC(gpr.R(a));
-	}
-}
-
-void Jit64::extshx(UGeckoInstruction inst)
-{
-	INSTRUCTION_START
-	JITDISABLE(bJITIntegerOff);
-	int a = inst.RA, s = inst.RS;
-
-	if (gpr.R(s).IsImm())
-	{
-		gpr.SetImmediate32(a, (u32)(s32)(s16)gpr.R(s).offset);
+		gpr.SetImmediate32(a, (u32)(s32)(size == 16 ? (s16)gpr.R(s).offset : (s8)gpr.R(s).offset));
+		if (inst.Rc)
+			ComputeRC(gpr.R(a));
 	}
 	else
 	{
 		gpr.Lock(a, s);
-		gpr.KillImmediate(s, true, false);
 		gpr.BindToRegister(a, a == s, true);
-		// This looks a little dangerous, but it's safe because
-		// every 32-bit register has a 16-bit half at the same index
-		// as the 32-bit register.
-		MOVSX(32, 16, gpr.RX(a), gpr.R(s));
+		// exts is moderately commonly used with inst.Rc, so try to optimize it.
+		if (inst.Rc)
+		{
+			// Only do one movsx; the movzx is free on most modern CPUs.
+			MOVSX(64, size, gpr.RX(a), gpr.R(s));
+			MOV(64, PPCSTATE(cr_val[0]), gpr.R(a));
+			MOVZX(64, 32, gpr.RX(a), gpr.R(a));
+		}
+		else
+		{
+			MOVSX(32, size, gpr.RX(a), gpr.R(s));
+		}
 		gpr.UnlockAll();
 	}
-
-	if (inst.Rc)
-	{
-		ComputeRC(gpr.R(a));
-	}
 }
 
 void Jit64::subfic(UGeckoInstruction inst)