From ef2ea31eacefcaa02aeda653f3ff3da4fa7e46e4 Mon Sep 17 00:00:00 2001
From: Ryan Houdek <Sonicadvance1@Gmail.com>
Date: Sat, 14 Sep 2013 04:12:57 +0000
Subject: [PATCH] [ARM] ps_madds0 and ps_madds1 implementations.

---
 Source/Core/Core/Src/PowerPC/JitArm32/Jit.h   |  2 +
 .../Src/PowerPC/JitArm32/JitArm_Paired.cpp    | 68 ++++++++++++++++++-
 .../Src/PowerPC/JitArm32/JitArm_Tables.cpp    |  4 +-
 3 files changed, 71 insertions(+), 3 deletions(-)

diff --git a/Source/Core/Core/Src/PowerPC/JitArm32/Jit.h b/Source/Core/Core/Src/PowerPC/JitArm32/Jit.h
index 75136ec155..143f21b5ee 100644
--- a/Source/Core/Core/Src/PowerPC/JitArm32/Jit.h
+++ b/Source/Core/Core/Src/PowerPC/JitArm32/Jit.h
@@ -210,6 +210,8 @@ public:
 	void ps_sum0(UGeckoInstruction _inst);
 	void ps_sum1(UGeckoInstruction _inst);
 	void ps_madd(UGeckoInstruction _inst);
+	void ps_madds0(UGeckoInstruction _inst);
+	void ps_madds1(UGeckoInstruction _inst);
 	void ps_sub(UGeckoInstruction _inst);
 	void ps_mul(UGeckoInstruction _inst);
 	void ps_muls0(UGeckoInstruction _inst);
diff --git a/Source/Core/Core/Src/PowerPC/JitArm32/JitArm_Paired.cpp b/Source/Core/Core/Src/PowerPC/JitArm32/JitArm_Paired.cpp
index 019e0563e6..9f8c90b7df 100644
--- a/Source/Core/Core/Src/PowerPC/JitArm32/JitArm_Paired.cpp
+++ b/Source/Core/Core/Src/PowerPC/JitArm32/JitArm_Paired.cpp
@@ -47,7 +47,6 @@ void JitArm::ps_add(UGeckoInstruction inst)
 	VADD(vD1, vA1, vB1);
 }
 
-// Wrong, THP videos like SMS and Ikaruga show artifacts
 void JitArm::ps_madd(UGeckoInstruction inst)
 {
 	INSTRUCTION_START
@@ -83,6 +82,73 @@ void JitArm::ps_madd(UGeckoInstruction inst)
 	fpr.Unlock(V1);
 }
 
+void JitArm::ps_madds0(UGeckoInstruction inst)
+{
+	INSTRUCTION_START
+	JITDISABLE(bJITPairedOff)
+	
+	u32 a = inst.FA, b = inst.FB, c = inst.FC, d = inst.FD;
+
+	if (inst.Rc) {
+		Default(inst); return;
+	}
+	ARMReg vA0 = fpr.R0(a);
+	ARMReg vA1 = fpr.R1(a);
+	ARMReg vB0 = fpr.R0(b);
+	ARMReg vB1 = fpr.R1(b);
+	ARMReg vC0 = fpr.R0(c);
+	ARMReg vD0 = fpr.R0(d, false);
+	ARMReg vD1 = fpr.R1(d, false);
+
+	ARMReg V0 = fpr.GetReg();
+	ARMReg V1 = fpr.GetReg();
+
+	VMOV(V0, vB0);
+	VMOV(V1, vB1);
+	
+	VMLA(V0, vA0, vC0);
+	VMLA(V1, vA1, vC0);
+
+	VMOV(vD0, V0);
+	VMOV(vD1, V1);
+
+	fpr.Unlock(V0);
+	fpr.Unlock(V1);
+}
+
+void JitArm::ps_madds1(UGeckoInstruction inst)
+{
+	INSTRUCTION_START
+	JITDISABLE(bJITPairedOff)
+	
+	u32 a = inst.FA, b = inst.FB, c = inst.FC, d = inst.FD;
+
+	if (inst.Rc) {
+		Default(inst); return;
+	}
+	ARMReg vA0 = fpr.R0(a);
+	ARMReg vA1 = fpr.R1(a);
+	ARMReg vB0 = fpr.R0(b);
+	ARMReg vB1 = fpr.R1(b);
+	ARMReg vC1 = fpr.R1(c);
+	ARMReg vD0 = fpr.R0(d, false);
+	ARMReg vD1 = fpr.R1(d, false);
+
+	ARMReg V0 = fpr.GetReg();
+	ARMReg V1 = fpr.GetReg();
+
+	VMOV(V0, vB0);
+	VMOV(V1, vB1);
+	
+	VMLA(V0, vA0, vC1);
+	VMLA(V1, vA1, vC1);
+
+	VMOV(vD0, V0);
+	VMOV(vD1, V1);
+
+	fpr.Unlock(V0);
+	fpr.Unlock(V1);
+}
 void JitArm::ps_sum0(UGeckoInstruction inst)
 {
 	INSTRUCTION_START
diff --git a/Source/Core/Core/Src/PowerPC/JitArm32/JitArm_Tables.cpp b/Source/Core/Core/Src/PowerPC/JitArm32/JitArm_Tables.cpp
index da6d6b9dc8..4569eb54c4 100644
--- a/Source/Core/Core/Src/PowerPC/JitArm32/JitArm_Tables.cpp
+++ b/Source/Core/Core/Src/PowerPC/JitArm32/JitArm_Tables.cpp
@@ -147,8 +147,8 @@ static GekkoOPTemplate table4_2[] =
 	{11, &JitArm::ps_sum1}, //"ps_sum1",   OPTYPE_PS, 0}},
 	{12, &JitArm::ps_muls0}, //"ps_muls0",  OPTYPE_PS, 0}},
 	{13, &JitArm::ps_muls1}, //"ps_muls1",  OPTYPE_PS, 0}},
-	{14, &JitArm::Default}, //"ps_madds0", OPTYPE_PS, 0}},
-	{15, &JitArm::Default}, //"ps_madds1", OPTYPE_PS, 0}},
+	{14, &JitArm::ps_madds0}, //"ps_madds0", OPTYPE_PS, 0}},
+	{15, &JitArm::ps_madds1}, //"ps_madds1", OPTYPE_PS, 0}},
 	{18, &JitArm::Default}, //"ps_div",    OPTYPE_PS, 0, 16}},
 	{20, &JitArm::ps_sub}, //"ps_sub",    OPTYPE_PS, 0}},
 	{21, &JitArm::ps_add}, //"ps_add",    OPTYPE_PS, 0}},