From 84395b65f6f1c4e8f9579aab63b5c8c0584dab1f Mon Sep 17 00:00:00 2001
From: degasus <wickmarkus@web.de>
Date: Fri, 12 Feb 2016 11:22:44 +0100
Subject: [PATCH] JitArm64: Track singles in psq_st.

---
 .../PowerPC/JitArm64/JitArm64_BackPatch.cpp   | 12 +++++++
 .../JitArm64/JitArm64_LoadStorePaired.cpp     | 24 ++++++++++---
 .../Core/PowerPC/JitArmCommon/BackPatch.h     | 34 ++++++++++---------
 3 files changed, 49 insertions(+), 21 deletions(-)

diff --git a/Source/Core/Core/PowerPC/JitArm64/JitArm64_BackPatch.cpp b/Source/Core/Core/PowerPC/JitArm64/JitArm64_BackPatch.cpp
index 0a69bf2c9c..5f4421ee93 100644
--- a/Source/Core/Core/PowerPC/JitArm64/JitArm64_BackPatch.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/JitArm64_BackPatch.cpp
@@ -73,6 +73,11 @@ void JitArm64::EmitBackpatchRoutine(u32 flags, bool fastmem, bool do_farcode,
 				m_float_emit.REV32(8, D0, D0);
 				m_float_emit.STR(64, Q0, X28, addr);
 			}
+			else if (flags & BackPatchInfo::FLAG_SIZE_F32X2I)
+			{
+				m_float_emit.REV32(8, D0, RS);
+				m_float_emit.STR(64, Q0, X28, addr);
+			}
 			else
 			{
 				m_float_emit.REV64(8, Q0, RS);
@@ -197,6 +202,13 @@ void JitArm64::EmitBackpatchRoutine(u32 flags, bool fastmem, bool do_farcode,
 				MOVI2R(X30, (u64)PowerPC::Write_U64);
 				BLR(X30);
 			}
+			else if (flags & BackPatchInfo::FLAG_SIZE_F32X2I)
+			{
+				m_float_emit.UMOV(64, X0, RS, 0);
+				ORR(X0, SP, X0, ArithOption(X0, ST_ROR, 32));
+				MOVI2R(X30, (u64)PowerPC::Write_U64);
+				BLR(X30);
+			}
 			else
 			{
 				MOVI2R(X30, (u64)&PowerPC::Write_U64);
diff --git a/Source/Core/Core/PowerPC/JitArm64/JitArm64_LoadStorePaired.cpp b/Source/Core/Core/PowerPC/JitArm64/JitArm64_LoadStorePaired.cpp
index 9878d98511..dde52cfad5 100644
--- a/Source/Core/Core/PowerPC/JitArm64/JitArm64_LoadStorePaired.cpp
+++ b/Source/Core/Core/PowerPC/JitArm64/JitArm64_LoadStorePaired.cpp
@@ -115,8 +115,10 @@ void JitArm64::psq_st(UGeckoInstruction inst)
 	gpr.Lock(W0, W1, W2, W30);
 	fpr.Lock(Q0, Q1);
 
+	bool single = fpr.IsSingle(inst.RS);
+
 	ARM64Reg arm_addr = gpr.R(inst.RA);
-	ARM64Reg VS = fpr.R(inst.RS, REG_REG);
+	ARM64Reg VS = fpr.R(inst.RS, single ? REG_REG_SINGLE : REG_REG);
 
 	ARM64Reg scale_reg = W0;
 	ARM64Reg addr_reg = W1;
@@ -150,7 +152,12 @@ void JitArm64::psq_st(UGeckoInstruction inst)
 	if (js.assumeNoPairedQuantize)
 	{
 		u32 flags = BackPatchInfo::FLAG_STORE;
-		flags |= (inst.W ? BackPatchInfo::FLAG_SIZE_F32 : BackPatchInfo::FLAG_SIZE_F32X2);
+
+		if (single)
+			flags |= (inst.W ? BackPatchInfo::FLAG_SIZE_F32I : BackPatchInfo::FLAG_SIZE_F32X2I);
+		else
+			flags |= (inst.W ? BackPatchInfo::FLAG_SIZE_F32 : BackPatchInfo::FLAG_SIZE_F32X2);
+
 		EmitBackpatchRoutine(flags,
 			jo.fastmem,
 			jo.fastmem,
@@ -160,10 +167,17 @@ void JitArm64::psq_st(UGeckoInstruction inst)
 	}
 	else
 	{
-		if (inst.W)
-			m_float_emit.FCVT(32, 64, D0, VS);
+		if (single)
+		{
+			m_float_emit.ORR(D0, VS, VS);
+		}
 		else
-			m_float_emit.FCVTN(32, D0, VS);
+		{
+			if (inst.W)
+				m_float_emit.FCVT(32, 64, D0, VS);
+			else
+				m_float_emit.FCVTN(32, D0, VS);
+		}
 
 		LDR(INDEX_UNSIGNED, scale_reg, X29, PPCSTATE_OFF(spr[SPR_GQR0 + inst.I]));
 		UBFM(type_reg, scale_reg, 0, 2); // Type
diff --git a/Source/Core/Core/PowerPC/JitArmCommon/BackPatch.h b/Source/Core/Core/PowerPC/JitArmCommon/BackPatch.h
index c712918883..4e9f7ca199 100644
--- a/Source/Core/Core/PowerPC/JitArmCommon/BackPatch.h
+++ b/Source/Core/Core/PowerPC/JitArmCommon/BackPatch.h
@@ -9,22 +9,24 @@ struct BackPatchInfo
 {
 	enum
 	{
-		FLAG_STORE      = (1 << 0),
-		FLAG_LOAD       = (1 << 1),
-		FLAG_SIZE_8     = (1 << 2),
-		FLAG_SIZE_16    = (1 << 3),
-		FLAG_SIZE_32    = (1 << 4),
-		FLAG_SIZE_F32   = (1 << 5),
-		FLAG_SIZE_F32X2 = (1 << 6),
-		FLAG_SIZE_F64   = (1 << 7),
-		FLAG_REVERSE    = (1 << 8),
-		FLAG_EXTEND     = (1 << 9),
-		FLAG_SIZE_F32I  = (1 << 10),
-		FLAG_ZERO_256   = (1 << 11),
-		FLAG_MASK_FLOAT = FLAG_SIZE_F32 |
-		                  FLAG_SIZE_F32X2 |
-		                  FLAG_SIZE_F64 |
-		                  FLAG_SIZE_F32I,
+		FLAG_STORE       = (1 << 0),
+		FLAG_LOAD        = (1 << 1),
+		FLAG_SIZE_8      = (1 << 2),
+		FLAG_SIZE_16     = (1 << 3),
+		FLAG_SIZE_32     = (1 << 4),
+		FLAG_SIZE_F32    = (1 << 5),
+		FLAG_SIZE_F32X2  = (1 << 6),
+		FLAG_SIZE_F32X2I = (1 << 7),
+		FLAG_SIZE_F64    = (1 << 8),
+		FLAG_REVERSE     = (1 << 9),
+		FLAG_EXTEND      = (1 << 10),
+		FLAG_SIZE_F32I   = (1 << 11),
+		FLAG_ZERO_256    = (1 << 12),
+		FLAG_MASK_FLOAT  = FLAG_SIZE_F32 |
+		                   FLAG_SIZE_F32X2 |
+		                   FLAG_SIZE_F32X2I |
+		                   FLAG_SIZE_F64 |
+		                   FLAG_SIZE_F32I,
 	};
 
 	static u32 GetFlagSize(u32 flags)