From b0d271db3c7e79e3cf27131393d7a490f5e122d1 Mon Sep 17 00:00:00 2001
From: skidau <skidau@gmail.com>
Date: Sat, 2 Jun 2012 10:06:47 +1000
Subject: [PATCH] Changed MOVDDUP to use MOVSD on non-SSE3 CPU's. Added DMA
 wait time under DSP HLE mode.

Fixes Knockout Kings 2003.
---
 Source/Core/Common/Src/x64Emitter.cpp             |  2 +-
 Source/Core/Core/Src/HW/DSP.cpp                   | 10 ++++++----
 .../Core/Src/PowerPC/Jit64/Jit_FloatingPoint.cpp  | 15 +++++++++++++--
 3 files changed, 20 insertions(+), 7 deletions(-)

diff --git a/Source/Core/Common/Src/x64Emitter.cpp b/Source/Core/Common/Src/x64Emitter.cpp
index 00c58a0fbc..d4bd1adabf 100644
--- a/Source/Core/Common/Src/x64Emitter.cpp
+++ b/Source/Core/Common/Src/x64Emitter.cpp
@@ -1321,7 +1321,7 @@ void XEmitter::MOVDDUP(X64Reg regOp, OpArg arg)
 	{
 		// Simulate this instruction with SSE2 instructions
 		if (!arg.IsSimpleReg(regOp))
-			MOVQ_xmm(regOp, arg);   // MOVSD better?
+			MOVSD(regOp, arg);
 		UNPCKLPD(regOp, R(regOp));
 	}
 }
diff --git a/Source/Core/Core/Src/HW/DSP.cpp b/Source/Core/Core/Src/HW/DSP.cpp
index 0eb19b006f..f2509555ed 100644
--- a/Source/Core/Core/Src/HW/DSP.cpp
+++ b/Source/Core/Core/Src/HW/DSP.cpp
@@ -436,8 +436,9 @@ void Write16(const u16 _Value, const u32 _Address)
 			if (tmpControl.ARAM) g_dspState.DSPControl.ARAM = 0;
 			if (tmpControl.DSP)  g_dspState.DSPControl.DSP  = 0;
 
-			// g_ARAM (line below should be commented out to emulate the DMA wait time)
-			g_dspState.DSPControl.DMAState = 0;	// keep g_ARAM DMA State zero
+			// Tracking DMAState fixes Knockout Kings 2003 in DSP HLE mode
+			if (GetDSPEmulator()->IsLLE())
+				g_dspState.DSPControl.DMAState = 0;	// keep g_ARAM DMA State zero
 
 			// unknown
 			g_dspState.DSPControl.unk3	= tmpControl.unk3;
@@ -700,8 +701,9 @@ void Do_ARAM_DMA()
 	// seems like a good estimate
 	CoreTiming::ScheduleEvent_Threadsafe(g_arDMA.Cnt.count >> 1, et_GenerateDSPInterrupt, INT_ARAM | (1<<16));
 
-	// Uncomment the line below to emulate the DMA wait time.
-	//g_dspState.DSPControl.DMAState = 1;
+	// Emulating the DMA wait time fixes Knockout Kings 2003 in DSP HLE mode
+	if (!GetDSPEmulator()->IsLLE())
+		g_dspState.DSPControl.DMAState = 1;
 
 	// Real hardware DMAs in 32byte chunks, but we can get by with 8byte chunks
 	if (g_arDMA.Cnt.dir)
diff --git a/Source/Core/Core/Src/PowerPC/Jit64/Jit_FloatingPoint.cpp b/Source/Core/Core/Src/PowerPC/Jit64/Jit_FloatingPoint.cpp
index 5fcd2f9e29..fd197d0e84 100644
--- a/Source/Core/Core/Src/PowerPC/Jit64/Jit_FloatingPoint.cpp
+++ b/Source/Core/Core/Src/PowerPC/Jit64/Jit_FloatingPoint.cpp
@@ -24,6 +24,7 @@
 
 #include "Jit.h"
 #include "JitRegCache.h"
+#include "CPUDetect.h"
 
 const u64 GC_ALIGNED16(psSignBits2[2]) = {0x8000000000000000ULL, 0x8000000000000000ULL};
 const u64 GC_ALIGNED16(psAbsMask2[2])  = {0x7FFFFFFFFFFFFFFFULL, 0x7FFFFFFFFFFFFFFFULL};
@@ -64,9 +65,19 @@ void Jit64::fp_tri_op(int d, int a, int b, bool reversible, bool dupe, void (XEm
 		(this->*op)(XMM0, Gen::R(XMM1));
 		MOVSD(fpr.RX(d), Gen::R(XMM0));
 	}
-	if (dupe) {
+	if (dupe)
+	{
 		ForceSinglePrecisionS(fpr.RX(d));
-		MOVDDUP(fpr.RX(d), fpr.R(d));
+		if (cpu_info.bSSE3)
+		{
+			MOVDDUP(fpr.RX(d), fpr.R(d));
+		}
+		else
+		{
+			if (!fpr.R(d).IsSimpleReg(fpr.RX(d)))
+				MOVQ_xmm(fpr.RX(d), fpr.R(d));
+			UNPCKLPD(fpr.RX(d), R(fpr.RX(d)));
+		}
 	}
 	fpr.UnlockAll();
 }