From dc220fa13dff7caaf02061e391f47d9ee0b6a5ed Mon Sep 17 00:00:00 2001
From: Tillmann Karras <tilkax@gmail.com>
Date: Thu, 21 May 2015 12:33:36 +0200
Subject: [PATCH] Jit64: merge ps_sel into fselx

---
 Source/Core/Core/PowerPC/Jit64/Jit.h          |  1 -
 .../Core/Core/PowerPC/Jit64/Jit64_Tables.cpp  |  2 +-
 .../Core/PowerPC/Jit64/Jit_FloatingPoint.cpp  | 17 +++++++---
 Source/Core/Core/PowerPC/Jit64/Jit_Paired.cpp | 34 -------------------
 4 files changed, 14 insertions(+), 40 deletions(-)

diff --git a/Source/Core/Core/PowerPC/Jit64/Jit.h b/Source/Core/Core/PowerPC/Jit64/Jit.h
index 3fb54dce73..bbaf971191 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit.h
+++ b/Source/Core/Core/PowerPC/Jit64/Jit.h
@@ -193,7 +193,6 @@ public:
 
 	void reg_imm(UGeckoInstruction inst);
 
-	void ps_sel(UGeckoInstruction inst);
 	void ps_mr(UGeckoInstruction inst);
 	void ps_sign(UGeckoInstruction inst); //aggregate
 	void ps_arith(UGeckoInstruction inst); //aggregate
diff --git a/Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp b/Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp
index 5b9ac6ae23..0d8cd10cf3 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp
+++ b/Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp
@@ -127,7 +127,7 @@ static GekkoOPTemplate table4_2[] =
 	{18, &Jit64::ps_arith},  // ps_div
 	{20, &Jit64::ps_arith},  // ps_sub
 	{21, &Jit64::ps_arith},  // ps_add
-	{23, &Jit64::ps_sel},    // ps_sel
+	{23, &Jit64::fselx},     // ps_sel
 	{24, &Jit64::ps_res},    // ps_res
 	{25, &Jit64::ps_arith},  // ps_mul
 	{26, &Jit64::ps_rsqrte}, // ps_rsqrte
diff --git a/Source/Core/Core/PowerPC/Jit64/Jit_FloatingPoint.cpp b/Source/Core/Core/PowerPC/Jit64/Jit_FloatingPoint.cpp
index 8f45a3f983..3a97f5b439 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit_FloatingPoint.cpp
+++ b/Source/Core/Core/PowerPC/Jit64/Jit_FloatingPoint.cpp
@@ -256,13 +256,18 @@ void Jit64::fselx(UGeckoInstruction inst)
 	int b = inst.FB;
 	int c = inst.FC;
 
+	bool packed = inst.OPCD == 4; // ps_sel
+
 	fpr.Lock(a, b, c, d);
-	MOVAPD(XMM1, fpr.R(a));
 	PXOR(XMM0, R(XMM0));
 	// This condition is very tricky; there's only one right way to handle both the case of
 	// negative/positive zero and NaN properly.
 	// (a >= -0.0 ? c : b) transforms into (0 > a ? b : c), hence the NLE.
-	CMPSD(XMM0, R(XMM1), NLE);
+	if (packed)
+		CMPPD(XMM0, fpr.R(a), NLE);
+	else
+		CMPSD(XMM0, fpr.R(a), NLE);
+
 	if (cpu_info.bSSE4_1)
 	{
 		MOVAPD(XMM1, fpr.R(c));
@@ -275,8 +280,12 @@ void Jit64::fselx(UGeckoInstruction inst)
 		PANDN(XMM1, fpr.R(c));
 		POR(XMM1, R(XMM0));
 	}
-	fpr.BindToRegister(d);
-	MOVSD(fpr.RX(d), R(XMM1));
+
+	fpr.BindToRegister(d, !packed);
+	if (packed)
+		MOVAPD(fpr.RX(d), R(XMM1));
+	else
+		MOVSD(fpr.RX(d), R(XMM1));
 	fpr.UnlockAll();
 }
 
diff --git a/Source/Core/Core/PowerPC/Jit64/Jit_Paired.cpp b/Source/Core/Core/PowerPC/Jit64/Jit_Paired.cpp
index d440d16e90..bbf6e0e8e9 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit_Paired.cpp
+++ b/Source/Core/Core/PowerPC/Jit64/Jit_Paired.cpp
@@ -28,40 +28,6 @@ void Jit64::ps_mr(UGeckoInstruction inst)
 	MOVAPD(fpr.RX(d), fpr.R(b));
 }
 
-void Jit64::ps_sel(UGeckoInstruction inst)
-{
-	INSTRUCTION_START
-	JITDISABLE(bJITPairedOff);
-	FALLBACK_IF(inst.Rc);
-
-	int d = inst.FD;
-	int a = inst.FA;
-	int b = inst.FB;
-	int c = inst.FC;
-
-	fpr.Lock(a, b, c, d);
-
-	if (cpu_info.bSSE4_1)
-	{
-		PXOR(XMM0, R(XMM0));
-		CMPPD(XMM0, fpr.R(a), NLE);
-		MOVAPD(XMM1, fpr.R(c));
-		BLENDVPD(XMM1, fpr.R(b));
-	}
-	else
-	{
-		PXOR(XMM1, R(XMM1));
-		CMPPD(XMM1, fpr.R(a), NLE);
-		MOVAPD(XMM0, R(XMM1));
-		PAND(XMM1, fpr.R(b));
-		PANDN(XMM0, fpr.R(c));
-		POR(XMM1, R(XMM0));
-	}
-	fpr.BindToRegister(d, false);
-	MOVAPD(fpr.RX(d), R(XMM1));
-	fpr.UnlockAll();
-}
-
 void Jit64::ps_sign(UGeckoInstruction inst)
 {
 	INSTRUCTION_START