From df34d43936411582ac1a1553cd1cc6a32bf03dd7 Mon Sep 17 00:00:00 2001
From: Tillmann Karras <tilkax@gmail.com>
Date: Thu, 21 May 2015 12:33:37 +0200
Subject: [PATCH] Jit64: merge ps_sign into fsign

---
 Source/Core/Core/PowerPC/Jit64/Jit.h          |  1 -
 .../Core/Core/PowerPC/Jit64/Jit64_Tables.cpp  |  6 ++--
 .../Core/PowerPC/Jit64/Jit_FloatingPoint.cpp  | 24 +++++++-------
 Source/Core/Core/PowerPC/Jit64/Jit_Paired.cpp | 31 -------------------
 4 files changed, 15 insertions(+), 47 deletions(-)

diff --git a/Source/Core/Core/PowerPC/Jit64/Jit.h b/Source/Core/Core/PowerPC/Jit64/Jit.h
index f068030b61..3c7e80bcec 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit.h
+++ b/Source/Core/Core/PowerPC/Jit64/Jit.h
@@ -192,7 +192,6 @@ public:
 	void reg_imm(UGeckoInstruction inst);
 
 	void ps_mr(UGeckoInstruction inst);
-	void ps_sign(UGeckoInstruction inst); //aggregate
 	void ps_mergeXX(UGeckoInstruction inst);
 	void ps_res(UGeckoInstruction inst);
 	void ps_rsqrte(UGeckoInstruction inst);
diff --git a/Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp b/Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp
index 8dd66132ca..5da1deb704 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp
+++ b/Source/Core/Core/PowerPC/Jit64/Jit64_Tables.cpp
@@ -102,9 +102,9 @@ static GekkoOPTemplate table4[] =
 {    //SUBOP10
 	{0,    &Jit64::ps_cmpXX},              // ps_cmpu0
 	{32,   &Jit64::ps_cmpXX},              // ps_cmpo0
-	{40,   &Jit64::ps_sign},               // ps_neg
-	{136,  &Jit64::ps_sign},               // ps_nabs
-	{264,  &Jit64::ps_sign},               // ps_abs
+	{40,   &Jit64::fsign},               // ps_neg
+	{136,  &Jit64::fsign},               // ps_nabs
+	{264,  &Jit64::fsign},               // ps_abs
 	{64,   &Jit64::ps_cmpXX},              // ps_cmpu1
 	{72,   &Jit64::ps_mr},                 // ps_mr
 	{96,   &Jit64::ps_cmpXX},              // ps_cmpo1
diff --git a/Source/Core/Core/PowerPC/Jit64/Jit_FloatingPoint.cpp b/Source/Core/Core/PowerPC/Jit64/Jit_FloatingPoint.cpp
index a848c72e6e..954b2410c1 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit_FloatingPoint.cpp
+++ b/Source/Core/Core/PowerPC/Jit64/Jit_FloatingPoint.cpp
@@ -13,6 +13,7 @@ using namespace Gen;
 static const u64 GC_ALIGNED16(psSignBits[2]) = {0x8000000000000000ULL, 0x0000000000000000ULL};
 static const u64 GC_ALIGNED16(psSignBits2[2]) = {0x8000000000000000ULL, 0x8000000000000000ULL};
 static const u64 GC_ALIGNED16(psAbsMask[2])  = {0x7FFFFFFFFFFFFFFFULL, 0xFFFFFFFFFFFFFFFFULL};
+static const u64 GC_ALIGNED16(psAbsMask2[2]) = {0x7FFFFFFFFFFFFFFFULL, 0x7FFFFFFFFFFFFFFFULL};
 static const double GC_ALIGNED16(half_qnan_and_s32_max[2]) = {0x7FFFFFFF, -0x80000};
 
 void Jit64::fp_tri_op(int d, int a, int b, bool reversible, bool single, void (XEmitter::*avxOp)(X64Reg, X64Reg, OpArg),
@@ -238,23 +239,22 @@ void Jit64::fsign(UGeckoInstruction inst)
 
 	int d = inst.FD;
 	int b = inst.FB;
-	fpr.Lock(b, d);
-	fpr.BindToRegister(d);
+	bool packed = inst.OPCD == 4;
+
+	fpr.Lock(b, d);
+	OpArg src = fpr.R(b);
+	fpr.BindToRegister(d, false);
 
-	if (d != b)
-		MOVSD(fpr.RX(d), fpr.R(b));
 	switch (inst.SUBOP10)
 	{
-	case 40:  // fnegx
-		// We can cheat and not worry about clobbering the top half by using masks
-		// that don't modify the top half.
-		PXOR(fpr.RX(d), M(psSignBits));
+	case 40: // neg
+		avx_op(&XEmitter::VPXOR, &XEmitter::PXOR, fpr.RX(d), src, M(packed ? psSignBits2 : psSignBits), packed);
 		break;
-	case 264: // fabsx
-		PAND(fpr.RX(d), M(psAbsMask));
+	case 136: // nabs
+		avx_op(&XEmitter::VPOR, &XEmitter::POR, fpr.RX(d), src, M(packed ? psSignBits2 : psSignBits), packed);
 		break;
-	case 136: // fnabs
-		POR(fpr.RX(d), M(psSignBits));
+	case 264: // abs
+		avx_op(&XEmitter::VPAND, &XEmitter::PAND, fpr.RX(d), src, M(packed ? psAbsMask2 : psAbsMask), packed);
 		break;
 	default:
 		PanicAlert("fsign bleh");
diff --git a/Source/Core/Core/PowerPC/Jit64/Jit_Paired.cpp b/Source/Core/Core/PowerPC/Jit64/Jit_Paired.cpp
index 9a4a6186bf..c60e5c8d58 100644
--- a/Source/Core/Core/PowerPC/Jit64/Jit_Paired.cpp
+++ b/Source/Core/Core/PowerPC/Jit64/Jit_Paired.cpp
@@ -10,9 +10,6 @@
 
 using namespace Gen;
 
-static const u64 GC_ALIGNED16(psSignBits[2]) = {0x8000000000000000ULL, 0x8000000000000000ULL};
-static const u64 GC_ALIGNED16(psAbsMask[2])  = {0x7FFFFFFFFFFFFFFFULL, 0x7FFFFFFFFFFFFFFFULL};
-
 void Jit64::ps_mr(UGeckoInstruction inst)
 {
 	INSTRUCTION_START
@@ -28,34 +25,6 @@ void Jit64::ps_mr(UGeckoInstruction inst)
 	MOVAPD(fpr.RX(d), fpr.R(b));
 }
 
-void Jit64::ps_sign(UGeckoInstruction inst)
-{
-	INSTRUCTION_START
-	JITDISABLE(bJITPairedOff);
-	FALLBACK_IF(inst.Rc);
-
-	int d = inst.FD;
-	int b = inst.FB;
-
-	fpr.Lock(d, b);
-	fpr.BindToRegister(d, d == b);
-
-	switch (inst.SUBOP10)
-	{
-	case 40: //neg
-		avx_op(&XEmitter::VPXOR, &XEmitter::PXOR, fpr.RX(d), fpr.R(b), M(psSignBits));
-		break;
-	case 136: //nabs
-		avx_op(&XEmitter::VPOR, &XEmitter::POR, fpr.RX(d), fpr.R(b), M(psSignBits));
-		break;
-	case 264: //abs
-		avx_op(&XEmitter::VPAND, &XEmitter::PAND, fpr.RX(d), fpr.R(b), M(psAbsMask));
-		break;
-	}
-
-	fpr.UnlockAll();
-}
-
 void Jit64::ps_sum(UGeckoInstruction inst)
 {
 	INSTRUCTION_START