From c234dc97c00cf5d47313478ef3896a52b461fef1 Mon Sep 17 00:00:00 2001
From: Tillmann Karras <tilkax@gmail.com>
Date: Tue, 15 Oct 2013 23:47:35 +0200
Subject: [PATCH] Jit64: optimize floating-point/paired-single arith

The "else if (b != d)" branch was dead code and only works if b == d. Now
the last else block with two temporary XMM registers is not needed anymore.
---
 .../Src/PowerPC/Jit64/Jit_FloatingPoint.cpp   | 35 ++++++++-----------
 .../Core/Src/PowerPC/Jit64/Jit_Paired.cpp     | 33 ++++++++---------
 2 files changed, 29 insertions(+), 39 deletions(-)

diff --git a/Source/Core/Core/Src/PowerPC/Jit64/Jit_FloatingPoint.cpp b/Source/Core/Core/Src/PowerPC/Jit64/Jit_FloatingPoint.cpp
index a04114d110..c157903726 100644
--- a/Source/Core/Core/Src/PowerPC/Jit64/Jit_FloatingPoint.cpp
+++ b/Source/Core/Core/Src/PowerPC/Jit64/Jit_FloatingPoint.cpp
@@ -25,33 +25,28 @@ void Jit64::fp_tri_op(int d, int a, int b, bool reversible, bool dupe, void (XEm
 		fpr.BindToRegister(d, true);
 		(this->*op)(fpr.RX(d), fpr.R(b));
 	}
-	else if (d == b && reversible)
+	else if (d == b)
 	{
-		fpr.BindToRegister(d, true);
-		(this->*op)(fpr.RX(d), fpr.R(a));
+		if (reversible)
+		{
+			fpr.BindToRegister(d, true);
+			(this->*op)(fpr.RX(d), fpr.R(a));
+		}
+		else
+		{
+			MOVSD(XMM0, fpr.R(b));
+			fpr.BindToRegister(d, !dupe);
+			MOVSD(fpr.RX(d), fpr.R(a));
+			(this->*op)(fpr.RX(d), Gen::R(XMM0));
+		}
 	}
-	else if (a != d && b != d) 
+	else
 	{
 		// Sources different from d, can use rather quick solution
 		fpr.BindToRegister(d, !dupe);
 		MOVSD(fpr.RX(d), fpr.R(a));
 		(this->*op)(fpr.RX(d), fpr.R(b));
 	}
-	else if (b != d)
-	{
-		fpr.BindToRegister(d, !dupe);
-		MOVSD(XMM0, fpr.R(b));
-		MOVSD(fpr.RX(d), fpr.R(a));
-		(this->*op)(fpr.RX(d), Gen::R(XMM0));
-	}
-	else // Other combo, must use two temps :(
-	{
-		MOVSD(XMM0, fpr.R(a));
-		MOVSD(XMM1, fpr.R(b));
-		fpr.BindToRegister(d, !dupe);
-		(this->*op)(XMM0, Gen::R(XMM1));
-		MOVSD(fpr.RX(d), Gen::R(XMM0));
-	}
 	if (dupe)
 	{
 		ForceSinglePrecisionS(fpr.RX(d));
@@ -83,7 +78,7 @@ void Jit64::fp_arith_s(UGeckoInstruction inst)
 	// Causing problems for GC - Starfox Assault (invisible boss at the end of level 1)
 	if (inst.SUBOP5 == 21) {
 		Default(inst); return;
-	}	
+	}
 
 	if (inst.SUBOP5 == 26) {
 		// frsqrtex
diff --git a/Source/Core/Core/Src/PowerPC/Jit64/Jit_Paired.cpp b/Source/Core/Core/Src/PowerPC/Jit64/Jit_Paired.cpp
index b58a140497..074a79740d 100644
--- a/Source/Core/Core/Src/PowerPC/Jit64/Jit_Paired.cpp
+++ b/Source/Core/Core/Src/PowerPC/Jit64/Jit_Paired.cpp
@@ -152,33 +152,28 @@ void Jit64::tri_op(int d, int a, int b, bool reversible, void (XEmitter::*op)(X6
 		fpr.BindToRegister(d, true);
 		(this->*op)(fpr.RX(d), fpr.R(b));
 	}
-	else if (d == b && reversible)
+	else if (d == b)
 	{
-		fpr.BindToRegister(d, true);
-		(this->*op)(fpr.RX(d), fpr.R(a));
+		if (reversible)
+		{
+			fpr.BindToRegister(d, true);
+			(this->*op)(fpr.RX(d), fpr.R(a));
+		}
+		else
+		{
+			MOVAPD(XMM0, fpr.R(b));
+			fpr.BindToRegister(d, false);
+			MOVAPD(fpr.RX(d), fpr.R(a));
+			(this->*op)(fpr.RX(d), Gen::R(XMM0));
+		}
 	}
-	else if (a != d && b != d) 
+	else
 	{
 		//sources different from d, can use rather quick solution
 		fpr.BindToRegister(d, false);
 		MOVAPD(fpr.RX(d), fpr.R(a));
 		(this->*op)(fpr.RX(d), fpr.R(b));
 	}
-	else if (b != d)
-	{
-		fpr.BindToRegister(d, false);
-		MOVAPD(XMM0, fpr.R(b));
-		MOVAPD(fpr.RX(d), fpr.R(a));
-		(this->*op)(fpr.RX(d), Gen::R(XMM0));
-	}
-	else //Other combo, must use two temps :(
-	{
-		MOVAPD(XMM0, fpr.R(a));
-		MOVAPD(XMM1, fpr.R(b));
-		fpr.BindToRegister(d, false);
-		(this->*op)(XMM0, Gen::R(XMM1));
-		MOVAPD(fpr.RX(d), Gen::R(XMM0));
-	}
 	ForceSinglePrecisionP(fpr.RX(d));
 	fpr.UnlockAll();
 }