Merge branch 'master' into GLSL-master

Conflicts:
	Source/Core/VideoCommon/Src/PixelShaderGen.cpp
	Source/Plugins/Plugin_VideoDX11/Src/VertexManager.cpp
	Source/Plugins/Plugin_VideoDX9/Src/VertexManager.cpp

CMakeLists.txt

@@ -3,8 +3,9 @@
 #
 cmake_minimum_required(VERSION 2.6)
 
-option(USE_GLES "Enables GLES And EGL, disables OGL" OFF)
+option(ANDROID "Enables a build for Android" OFF)
 option(USE_EGL "Enables EGL OpenGL Interface" OFF)
+option(USE_GLES "Enables GLES And EGL, disables OGL" OFF)
 option(DISABLE_WX "Disable wxWidgets (use CLI interface)" OFF)
 
 option(FASTLOG "Enable all logs" OFF)
@@ -271,7 +272,6 @@ if(USE_EGL)
 endif()
 add_definitions(-D_FILE_OFFSET_BITS=64 -D_LARGEFILE_SOURCE)
 
-option(ANDROID "Enables a build for Android" OFF)
 if(ANDROID)
 	message("Building for Android")
 	add_definitions(-DANDROID)
@@ -394,7 +394,6 @@ if(NOT ANDROID)
 		set(PORTAUDIO_FOUND FALSE)
 	endif(PORTAUDIO)
 
-	option(OPROFILING "Enable profiling" OFF)
 	if(OPROFILING)
 		check_lib(OPROFILE opagent opagent.h)
 		check_lib(BFD bfd bfd.h)

Data/User/GameConfig/GAFE01.ini

@@ -319,4 +319,4 @@ PH_ZFar =
 [Video_Settings]
 SafeTextureCacheColorSamples = 512
 [Core]
-FastDiscSpeed = 1
+

Data/User/GameConfig/GAFJ01.ini

@@ -16,4 +16,4 @@ PH_ZFar =
 [Video_Settings]
 SafeTextureCacheColorSamples = 512
 [Core]
-FastDiscSpeed = 1
+

Data/User/GameConfig/GAFP01.ini

@@ -16,4 +16,4 @@ PH_ZFar =
 [Video_Settings]
 SafeTextureCacheColorSamples = 512
 [Core]
-FastDiscSpeed = 1
+

Data/User/GameConfig/GAFU01.ini

@@ -16,4 +16,4 @@ PH_ZFar =
 [Video_Settings]
 SafeTextureCacheColorSamples = 512
 [Core]
-FastDiscSpeed = 1
+

Data/User/GameConfig/GFZE01.ini

@@ -2,10 +2,10 @@
 [Core] Values set here will override the main dolphin settings.
 EnableFPRF = True
 TLBHack = 1
-CPUThread = 0
+SyncGPU = 1
 [EmuState] The Emulation State. 1 is worst, 5 is best, 0 is not set.
 EmulationStateId = 4
-EmulationIssues = Dual core is very unstable/doesn't even boot with most video plugins. Efb to ram is needed for proper shadows.
+EmulationIssues = Needs Synchronize GPU thread for stability.
 [OnFrame] Add memory patches to be applied every frame here.
 [Video]
 ProjectionHack = 1

Data/User/GameConfig/GFZJ01.ini

@@ -0,0 +1,26 @@
+# GFZJ01 - F-ZERO GX 
+[Core] Values set here will override the main dolphin settings.
+EnableFPRF = True
+TLBHack = 1
+SyncGPU = 1
+[EmuState] The Emulation State. 1 is worst, 5 is best, 0 is not set.
+EmulationStateId = 4
+EmulationIssues = Needs Synchronize GPU thread for stability.
+[OnFrame] Add memory patches to be applied every frame here.
+[Video]
+ProjectionHack = 1
+PH_SZNear = 1
+PH_SZFar = 1
+PH_ExtraParam = 0
+PH_ZNear = 20
+PH_ZFar = 1.7555555
+[ActionReplay]
+$Make Save Copyable
+0C031514 909C0028
+04031518 48BFFBE8
+04C31100 38000004
+04C31104 981C0034
+04C31108 38000000
+04C3110C 4B400410
+[Gecko]
+

Data/User/GameConfig/GFZP01.ini

@@ -2,10 +2,10 @@
 [Core] Values set here will override the main dolphin settings.
 EnableFPRF = True
 TLBHack = 1
-CPUThread = 0
+SyncGPU = 1
 [EmuState] The Emulation State. 1 is worst, 5 is best, 0 is not set.
 EmulationStateId = 4
-EmulationIssues = Dual core is very unstable/doesn't even boot with most video plugins. Efb to ram is needed for proper shadows.
+EmulationIssues = Needs Synchronize GPU thread for stability.
 [OnFrame] Add memory patches to be applied every frame here.
 [Video]
 ProjectionHack = 1

Data/User/GameConfig/GJXE51.ini

@@ -15,4 +15,4 @@ EmulationIssues = Slow,needs mmu and lle audio plugin for proper audio(r7411).
 [Video_Settings]
 [Core]
 MMU = 1
-FastDiscSpeed = 1
+

Data/User/GameConfig/GNOE78.ini

@@ -1,7 +1,6 @@
 # GNOE78 - Nicktoons Unite!
 [Core] Values set here will override the main dolphin settings.
 TLBHack = 1
-FastDiscSpeed = 1
 [EmuState] The Emulation State. 1 is worst, 5 is best, 0 is not set.
 EmulationStateId = 4
 EmulationIssues = 

Data/User/GameConfig/GOSE41.ini

@@ -1,7 +1,6 @@
 # GOSE41 - Open Season
 [Core] Values set here will override the main dolphin settings.
 MMU = 1
-FastDiscSpeed = 1
 BlockMerging = 1
 [EmuState] The Emulation State. 1 is worst, 5 is best, 0 is not set.
 EmulationStateId = 4

Data/User/GameConfig/GOSP41.ini

@@ -1,7 +1,6 @@
 # GOSP41 - Open Season
 [Core] Values set here will override the main dolphin settings.
 MMU = 1
-FastDiscSpeed = 1
 BlockMerging = 1
 [EmuState] The Emulation State. 1 is worst, 5 is best, 0 is not set.
 EmulationStateId = 4

Data/User/GameConfig/GOSX41.ini

@@ -1,7 +1,6 @@
 # GOSX41 - Open Season
 [Core] Values set here will override the main dolphin settings.
 MMU = 1
-FastDiscSpeed = 1
 BlockMerging = 1
 [EmuState] The Emulation State. 1 is worst, 5 is best, 0 is not set.
 EmulationStateId = 4

Data/User/GameConfig/GWAE8P.ini

@@ -1,6 +1,5 @@
 # GWAE8P - Spartan: Total Warrior
 [Core] Values set here will override the main dolphin settings.
-FastDiscSpeed = 1
 [EmuState] The Emulation State. 1 is worst, 5 is best, 0 is not set.
 EmulationStateId = 0
 EmulationIssues = 

Data/User/GameConfig/GWAF8P.ini

@@ -1,6 +1,5 @@
 # GWAF8P - Spartan: Total Warrior
 [Core] Values set here will override the main dolphin settings.
-FastDiscSpeed = 1
 [EmuState] The Emulation State. 1 is worst, 5 is best, 0 is not set.
 EmulationStateId = 0
 EmulationIssues = 

Data/User/GameConfig/GWAP8P.ini

@@ -1,6 +1,5 @@
 # GWAP8P - Spartan : Total  Warrior (TM)
 [Core] Values set here will override the main dolphin settings.
-FastDiscSpeed = 1
 [EmuState] The Emulation State. 1 is worst, 5 is best, 0 is not set.
 EmulationStateId = 0
 EmulationIssues = 

Data/User/GameConfig/RRBE41.ini

@@ -1,9 +1,9 @@
 # RRBE41 - Rayman Raving Rabbids
 [Core] Values set here will override the main dolphin settings.
-CPUThread = 0
+SyncGPU = 1
 [EmuState] The Emulation State. 1 is worst, 5 is best, 0 is not set.
 EmulationStateId = 4
-EmulationIssues = Dual core is unstable.Use direct3d11 for less glitches. Slow.(r7687)
+EmulationIssues = Needs Synchronise GPU thread for stability. Use direct3d11 for less glitches.
 [OnFrame] Add memory patches to be applied every frame here.
 [ActionReplay] Add action replay cheats here.
 [Video]

Data/User/GameConfig/RRBJ41.ini

@@ -1,9 +1,9 @@
 # RRBJ41 - Rayman Raving Rabbids
 [Core] Values set here will override the main dolphin settings.
-CPUThread = 0
+SyncGPU = 1
 [EmuState] The Emulation State. 1 is worst, 5 is best, 0 is not set.
 EmulationStateId = 4
-EmulationIssues = Dual core is unstable.Use direct3d11 for less glitches. Slow.(r7687)
+EmulationIssues = Needs Synchronise GPU thread for stability. Use direct3d11 for less glitches.
 [OnFrame] Add memory patches to be applied every frame here.
 [ActionReplay] Add action replay cheats here.
 [Video]

Data/User/GameConfig/RRBP41.ini

@@ -1,9 +1,9 @@
 # RRBP41 - Rayman Raving Rabbids
 [Core] Values set here will override the main dolphin settings.
-CPUThread = 0
+SyncGPU = 1
 [EmuState] The Emulation State. 1 is worst, 5 is best, 0 is not set.
 EmulationStateId = 4
-EmulationIssues = Dual core is unstable.Use direct3d11 for less glitches. Slow.(r7687)
+EmulationIssues = Needs Synchronise GPU thread for stability. Use direct3d11 for less glitches.
 [OnFrame] Add memory patches to be applied every frame here.
 [ActionReplay] Add action replay cheats here.
 [Video]

Externals/soundtouch/SoundTouch.vcxproj

@@ -174,7 +174,7 @@
       <CompileAs>Default</CompileAs>
     </ClCompile>
     <Lib>
-      <OutputFile>Win32\SoundTouch.lib</OutputFile>
+      <OutputFile>Win32\SoundTouchD.lib</OutputFile>
       <SuppressStartupBanner>true</SuppressStartupBanner>
     </Lib>
     <PostBuildEvent>

Source/Core/Common/Src/ArmCPUDetect.cpp

@@ -122,6 +122,12 @@ void CPUInfo::Detect()
 	bVFPv4 = CheckCPUFeature("vfpv4");
 	bIDIVa = CheckCPUFeature("idiva");
 	bIDIVt = CheckCPUFeature("idivt");
+	
+	// On some buggy kernels(Qualcomm) they show that they support VFPv4 but not IDIVa
+	// All VFPv4 CPUs will support IDIVa
+	if (bVFPv4)
+		bIDIVa = bIDIVt = true;
+
 	// These two are ARMv8 specific.
 	bFP = CheckCPUFeature("fp");
 	bASIMD = CheckCPUFeature("asimd");

Source/Core/Common/Src/ArmEmitter.cpp

@@ -176,7 +176,7 @@ void ARMXEmitter::MOVI2R(ARMReg reg, u32 val, bool optimize)
 		} else {
 			// Use literal pool for ARMv6.
 			AddNewLit(val);
-			LDRLIT(reg, 0, 0); // To be backpatched later
+			LDR(reg, _PC); // To be backpatched later
 		}
 	}
 }
@@ -608,39 +608,6 @@ void ARMXEmitter::MRS (ARMReg dest)
 {
 	Write32(condition | (16 << 20) | (15 << 16) | (dest << 12));
 }
-
-void ARMXEmitter::WriteStoreOp(u32 op, ARMReg src, ARMReg dest, s16 op2)
-{
-       // Qualcomm chipsets get /really/ angry if you don't use index, even if the offset is zero.
-       // bool Index = op2 != 0 ? true : false;
-       bool Index = true;
-       bool Add = op2 >= 0 ? true : false;
-       u32 imm = abs(op2);
-       Write32(condition | (op << 20) | (Index << 24) | (Add << 23) | (src << 16) | (dest << 12) | imm);
-}
-void ARMXEmitter::STR  (ARMReg result, ARMReg base, s16 op) { WriteStoreOp(0x40, base, result, op);}
-void ARMXEmitter::STRH (ARMReg result, ARMReg base, Operand2 op)
-{
-	u8 Imm = op.Imm8();
-	Write32(condition | (0x04 << 20) | (base << 16) | (result << 12) | ((Imm >> 4) << 8) | (0xB << 4) | (Imm & 0x0F));
-}
-void ARMXEmitter::STRB (ARMReg result, ARMReg base, s16 op) { WriteStoreOp(0x44, base, result, op);}
-void ARMXEmitter::STR  (ARMReg result, ARMReg base, Operand2 op2, bool Index, bool Add)
-{
-	Write32(condition | (0x60 << 20) | (Index << 24) | (Add << 23) | (base << 16) | (result << 12) | op2.IMMSR());
-}
-void ARMXEmitter::STR  (ARMReg result, ARMReg base, ARMReg offset, bool Index, bool Add)
-{
-	Write32(condition | (0x60 << 20) | (Index << 24) | (Add << 23) | (base << 16) | (result << 12) | offset);
-}
-void ARMXEmitter::STRH (ARMReg result, ARMReg base, ARMReg offset, bool Index, bool Add)
-{
-	Write32(condition | (0x00 << 20) | (Index << 24) | (Add << 23) | (base << 16) | (result << 12) | (0xB << 4) | offset);
-}
-void ARMXEmitter::STRB (ARMReg result, ARMReg base, ARMReg offset, bool Index, bool Add)
-{
-	Write32(condition | (0x64 << 20) | (Index << 24) | (Add << 23) | (base << 16) | (result << 12) | offset);
-}
 void ARMXEmitter::LDREX(ARMReg dest, ARMReg base)
 {
 	Write32(condition | (25 << 20) | (base << 16) | (dest << 12) | 0xF9F);
@@ -659,49 +626,113 @@ void ARMXEmitter::SVC(Operand2 op)
 	Write32(condition | (0x0F << 24) | op.Imm24());
 }
 
-void ARMXEmitter::LDR (ARMReg dest, ARMReg src, s16 op) { WriteStoreOp(0x41, src, dest, op);}
+// IMM, REG, IMMSREG, RSR
-void ARMXEmitter::LDRH(ARMReg dest, ARMReg src, Operand2 op)
+// -1 for invalid if the instruction doesn't support that
+const s32 LoadStoreOps[][4] = { {0x40, 0x60, 0x60, -1}, // STR
+				{0x41, 0x61, 0x61, -1}, // LDR
+				{0x44, 0x64, 0x64, -1}, // STRB
+				{0x45, 0x65, 0x65, -1}, // LDRB
+				// Special encodings
+				{ 0x4,  0x0,  -1, -1}, // STRH
+				{ 0x5,  0x1,  -1, -1}, // LDRH
+				{ 0x5,  0x1,  -1, -1}, // LDRSB
+				{ 0x5,  0x1,  -1, -1}, // LDRSH
+				};
+const char *LoadStoreNames[] = { "STR",
+				 "LDR",
+				 "STRB",
+				 "LDRB",
+				 "STRH",
+				 "LDRH",
+				 "LDRSB",
+				 "LDRSH",
+				};
+
+void ARMXEmitter::WriteStoreOp(u32 Op, ARMReg Rt, ARMReg Rn, Operand2 Rm, bool RegAdd)
 {
-	u8 Imm = op.Imm8();
+	s32 op = LoadStoreOps[Op][Rm.GetType()]; // Type always decided by last operand
-	Write32(condition | (0x05 << 20) | (src << 16) | (dest << 12) | ((Imm >> 4) << 8) | (0xB << 4) | (Imm & 0x0F));
+	u32 Data;
+	
+	// Qualcomm chipsets get /really/ angry if you don't use index, even if the offset is zero.
+	// Some of these encodings require Index at all times anyway. Doesn't really matter.
+	// bool Index = op2 != 0 ? true : false;
+	bool Index = true;
+	bool Add = false; 	
+
+	// Special Encoding	
+	bool SpecialOp = false;
+	bool Half = false;
+	bool SignedLoad = false;
+
+	if (op == -1)
+		_assert_msg_(DYNA_REC, false, "%s does not support %d", LoadStoreNames[Op], Rm.GetType()); 
+
+	switch (Op)
+	{
+		case 4: // STRH
+			SpecialOp = true;
+			Half = true;
+			SignedLoad = false;
+		break;
+		case 5: // LDRH
+			SpecialOp = true;
+			Half = true;
+			SignedLoad = false;
+		break;
+		case 6: // LDRSB
+			SpecialOp = true;
+			Half = false;
+			SignedLoad = true;
+		break;
+		case 7: // LDRSH
+			SpecialOp = true;
+			Half = true;
+			SignedLoad = true;
+		break;
 	}
-void ARMXEmitter::LDRSH(ARMReg dest, ARMReg src, Operand2 op)
+	switch (Rm.GetType())
 	{
-	u8 Imm = op.Imm8();
+		case TYPE_IMM:
-	Write32(condition | (0x05 << 20) | (src << 16) | (dest << 12) | ((Imm >> 4) << 8) | (0xF << 4) | (Imm & 0x0F));
+		{
+			s32 Temp = (s32)Rm.Value;
+			Data = abs(Temp);
+			// The offset is encoded differently on this one.
+			if (SpecialOp)
+				Data = (Data & 0xF0 << 4) | (Data & 0xF);
+			if (Temp >= 0) Add = true;	
 		}
-void ARMXEmitter::LDRB(ARMReg dest, ARMReg src, s16 op) { WriteStoreOp(0x45, src, dest, op);}
+		break;
-void ARMXEmitter::LDRSB(ARMReg dest, ARMReg src, Operand2 op)
+		case TYPE_REG:
+		case TYPE_IMMSREG:
+			if (!SpecialOp)
 			{
-	u8 Imm = op.Imm8();
+				Data = Rm.GetData();
-	Write32(condition | (0x05 << 20) | (src << 16) | (dest << 12) | ((Imm >> 4) << 8) | (0xD << 4) | (Imm & 0x0F));
+				Add = RegAdd;
+				break;
+			}
+		default:
+			// RSR not supported for any of these
+			// We already have the warning above
+			BKPT(0x2);
+			return;
+		break;  
+	}
+	if (SpecialOp)
+	{
+		// Add SpecialOp things
+		Data = (0x5 << 4) | (SignedLoad << 6) | (Half << 5) | Data;
+	}
+	Write32(condition | (op << 20) | (Index << 24) | (Add << 23) | (Rn << 16) | (Rt << 12) | Data);
 }
 
-void ARMXEmitter::LDR  (ARMReg dest, ARMReg base, Operand2 op2, bool Index, bool Add)
+void ARMXEmitter::LDR (ARMReg dest, ARMReg base, Operand2 op2, bool RegAdd) { WriteStoreOp(1, dest, base, op2, RegAdd);}
-{
+void ARMXEmitter::LDRB(ARMReg dest, ARMReg base, Operand2 op2, bool RegAdd) { WriteStoreOp(3, dest, base, op2, RegAdd);}
-	Write32(condition | (0x61 << 20) | (Index << 24) | (Add << 23) | (base << 16) | (dest << 12) | op2.IMMSR());
+void ARMXEmitter::LDRH(ARMReg dest, ARMReg base, Operand2 op2, bool RegAdd) { WriteStoreOp(5, dest, base, op2, RegAdd);}
-}
+void ARMXEmitter::LDRSB(ARMReg dest, ARMReg base, Operand2 op2, bool RegAdd) { WriteStoreOp(6, dest, base, op2, RegAdd);}
-void ARMXEmitter::LDR  (ARMReg dest, ARMReg base, ARMReg offset, bool Index, bool Add)
+void ARMXEmitter::LDRSH(ARMReg dest, ARMReg base, Operand2 op2, bool RegAdd) { WriteStoreOp(7, dest, base, op2, RegAdd);}
-{
+void ARMXEmitter::STR  (ARMReg result, ARMReg base, Operand2 op2, bool RegAdd) { WriteStoreOp(0, result, base, op2, RegAdd);}
-	Write32(condition | (0x61 << 20) | (Index << 24) | (Add << 23) | (base << 16) | (dest << 12) | offset);
+void ARMXEmitter::STRH (ARMReg result, ARMReg base, Operand2 op2, bool RegAdd) { WriteStoreOp(4, result, base, op2, RegAdd);}
-}
+void ARMXEmitter::STRB (ARMReg result, ARMReg base, Operand2 op2, bool RegAdd) { WriteStoreOp(2, result, base, op2, RegAdd);}
-void ARMXEmitter::LDRH (ARMReg dest, ARMReg base, ARMReg offset, bool Index, bool Add)
-{
-	Write32(condition | (0x01 << 20) | (Index << 24) | (Add << 23) | (base << 16) | (dest << 12) | (0xB << 4) | offset);
-}
-void ARMXEmitter::LDRSH(ARMReg dest, ARMReg base, ARMReg offset, bool Index, bool Add)
-{
-	Write32(condition | (0x01 << 20) | (Index << 24) | (Add << 23) | (base << 16) | (dest << 12) | (0xF << 4) | offset);
-}
-void ARMXEmitter::LDRB (ARMReg dest, ARMReg base, ARMReg offset, bool Index, bool Add)
-{
-	Write32(condition | (0x65 << 20) | (Index << 24) | (Add << 23) | (base << 16) | (dest << 12) | offset);
-}
-void ARMXEmitter::LDRSB(ARMReg dest, ARMReg base, ARMReg offset, bool Index, bool Add)
-{
-	Write32(condition | (0x01 << 20) | (Index << 24) | (Add << 23) | (base << 16) | (dest << 12) | (0xD << 4) | offset);
-}
-void ARMXEmitter::LDRLIT (ARMReg dest, u32 offset, bool Add) { Write32(condition | 0x05 << 24 | Add << 23 | 0x1F << 16 | dest << 12 | offset);}
 
 void ARMXEmitter::WriteRegStoreOp(u32 op, ARMReg dest, bool WriteBack, u16 RegList)
 {
@@ -787,6 +818,100 @@ void ARMXEmitter::VSUB(IntegerSize Size, ARMReg Vd, ARMReg Vn, ARMReg Vm)
 }
 
 // VFP Specific
+struct VFPEnc
+{
+	u16 opc1;
+	u16 opc2;
+};
+// Double/single, Neon
+const VFPEnc VFPOps[][2] = {
+	{{0xE0, 0xA0}, {0x20, 0xD1}}, // VMLA
+	{{0xE0, 0xA4}, {0x22, 0xD1}}, // VMLS
+	{{0xE3, 0xA0}, {0x20, 0xD0}}, // VADD
+	{{0xE3, 0xA4}, {0x22, 0xD0}}, // VSUB
+	{{0xE2, 0xA0}, {0x30, 0xD1}}, // VMUL
+	{{0xEB, 0xAC}, {0x3B, 0x30}}, // VABS	
+	{{0xE8, 0xA0}, {  -1,   -1}}, // VDIV
+	{{0xEB, 0xA4}, {  -1,   -1}}, // VNEG
+	{{0xEB, 0xAC}, {  -1,   -1}}, // VSQRT
+	{{0xED, 0xA2}, {  -1,   -1}}, // VCMP
+	{{0xED, 0xAA}, {  -1,   -1}}, // VCMPE
+	{{  -1,   -1}, {0x3B, 0x70}}, // VABSi
+	};
+const char *VFPOps[] = {
+	"VMLA",
+	"VMLS",
+	"VADD",
+	"VSUB",
+	"VMUL",
+	"VABS",
+	"VDIV",
+	"VNEG",
+	"VSQRT",
+	"VCMP",
+	"VCMPE",
+	"VABSi",
+};
+
+u32 ARMXEmitter::EncodeVd(ARMReg Vd)
+{
+	bool quad_reg = Vd >= Q0;
+	bool double_reg = Vd >= D0;
+
+	ARMReg Reg = SubBase(Vd);
+
+	if (quad_reg)
+		((Reg & 0x10) << 18) | ((Reg & 0xF) << 12);
+	else
+		if (double_reg)
+			return ((Reg & 0x10) << 18) | ((Reg & 0xF) << 16);
+		else
+			return ((Reg & 0x1) << 22) | ((Reg & 0x1E) << 11);
+}
+u32 ARMXEmitter::EncodeVn(ARMReg Vn)
+{
+	bool quad_reg = Vn >= Q0;
+	bool double_reg = Vn >= D0;
+	
+	ARMReg Reg = SubBase(Vn);
+	if (quad_reg)
+		((Reg & 0xF) << 16) | ((Reg & 0x10) << 3);
+	else
+		if (double_reg)
+			return ((Reg & 0xF) << 16) | ((Reg & 0x10) << 3);
+		else
+			return ((Reg & 0x1E) << 15) | ((Reg & 0x1) << 7);
+}
+u32 ARMXEmitter::EncodeVm(ARMReg Vm)
+{
+	bool quad_reg = Vm >= Q0;
+	bool double_reg = Vm >= D0;
+
+	ARMReg Reg = SubBase(Vm);
+
+	if (quad_reg)
+		((Reg & 0x10) << 2) | (Reg & 0xF);
+	else
+		if (double_reg)
+			return ((Reg & 0x10) << 2) | (Reg & 0xF);
+		else
+			return ((Reg & 0x1) << 5) | (Reg >> 1);
+}
+void ARMXEmitter::WriteVFPDataOp(u32 Op, ARMReg Vd, ARMReg Vn, ARMReg Vm)
+{
+	bool quad_reg = Vd >= Q0;
+	bool double_reg = Vd >= D0;
+
+	VFPEnc enc = VFPOps[Op][quad_reg];
+	if (enc.opc1 == (u16)-1 && enc.opc1 == (u16)-1)
+		_assert_msg_(DYNA_REC, false, "%s does not support %s", VFPOps[Op], quad_reg ? "NEON" : "VFP"); 
+	u32 VdEnc = EncodeVd(Vd);
+	u32 VnEnc = EncodeVn(Vn);
+	u32 VmEnc = EncodeVm(Vm);
+	u32 cond = quad_reg ? (0xF << 28) : condition;
+
+	Write32(cond | enc.opc1 | VnEnc | VdEnc | enc.opc2 | VmEnc);
+}
 
 void ARMXEmitter::VLDR(ARMReg Dest, ARMReg Base, s16 offset)
 {

Source/Core/Common/Src/ArmEmitter.h

@@ -352,7 +352,7 @@ private:
 	u32 condition;
 	std::vector<LiteralPool> currentLitPool;
 
-	void WriteStoreOp(u32 op, ARMReg src, ARMReg dest, s16 op2);
+	void WriteStoreOp(u32 Op, ARMReg Rt, ARMReg Rn, Operand2 op2, bool RegAdd);
 	void WriteRegStoreOp(u32 op, ARMReg dest, bool WriteBack, u16 RegList);
 	void WriteShiftedDataOp(u32 op, bool SetFlags, ARMReg dest, ARMReg src, ARMReg op2);
 	void WriteShiftedDataOp(u32 op, bool SetFlags, ARMReg dest, ARMReg src, Operand2 op2);
@@ -495,28 +495,14 @@ public:
 	void MRS  (ARMReg dest);
 
 	// Memory load/store operations
-	void LDR  (ARMReg dest, ARMReg src, s16 op2 = 0);
+	void LDR  (ARMReg dest, ARMReg base, Operand2 op2 = 0, bool RegAdd = true);
-	void LDRH (ARMReg dest, ARMReg src, Operand2 op2 = 0);
+	void LDRB (ARMReg dest, ARMReg base, Operand2 op2 = 0, bool RegAdd = true);
-	void LDRSH(ARMReg dest, ARMReg src, Operand2 op2 = 0);
+	void LDRH (ARMReg dest, ARMReg base, Operand2 op2 = 0, bool RegAdd = true);
-	void LDRB (ARMReg dest, ARMReg src, s16 op2 = 0);
+	void LDRSB(ARMReg dest, ARMReg base, Operand2 op2 = 0, bool RegAdd = true);
-	void LDRSB(ARMReg dest, ARMReg src, Operand2 op2 = 0);
+	void LDRSH(ARMReg dest, ARMReg base, Operand2 op2 = 0, bool RegAdd = true);
-	// Offset adds to the base register in LDR
+	void STR  (ARMReg result, ARMReg base, Operand2 op2 = 0, bool RegAdd = true);
-	void LDR  (ARMReg dest, ARMReg base, Operand2 op2, bool Index, bool Add);
+	void STRB (ARMReg result, ARMReg base, Operand2 op2 = 0, bool RegAdd = true);
-	void LDR  (ARMReg dest, ARMReg base, ARMReg offset, bool Index, bool Add);
+	void STRH (ARMReg result, ARMReg base, Operand2 op2 = 0, bool RegAdd = true);
-	void LDRH (ARMReg dest, ARMReg base, ARMReg offset, bool Index, bool Add);
-	void LDRSH(ARMReg dest, ARMReg base, ARMReg offset, bool Index, bool Add);
-	void LDRB (ARMReg dest, ARMReg base, ARMReg offset, bool Index, bool Add);
-	void LDRSB(ARMReg dest, ARMReg base, ARMReg offset, bool Index, bool Add);
-	void LDRLIT(ARMReg dest, u32 offset, bool Add);
-
-	void STR  (ARMReg result, ARMReg base, s16 op2 = 0);
-	void STRH (ARMReg result, ARMReg base, Operand2 op2 = 0);
-	void STRB (ARMReg result, ARMReg base, s16 op2 = 0);
-	// Offset adds on to the destination register in STR
-	void STR  (ARMReg result, ARMReg base, Operand2 op2, bool Index, bool Add);
-	void STR  (ARMReg result, ARMReg base, ARMReg offset, bool Index, bool Add);
-	void STRH (ARMReg result, ARMReg base, ARMReg offset, bool Index, bool Add);
-	void STRB (ARMReg result, ARMReg base, ARMReg offset, bool Index, bool Add);
 
 	void STMFD(ARMReg dest, bool WriteBack, const int Regnum, ...);
 	void LDMFD(ARMReg dest, bool WriteBack, const int Regnum, ...);

Source/Core/Core/Src/HW/WII_IPC.cpp

@@ -28,6 +28,7 @@
 
 #include "../IPC_HLE/WII_IPC_HLE.h"
 #include "WII_IPC.h"
+#include "CoreTiming.h"
 
 
 // This is the intercommunication between ARM and PPC. Currently only PPC actually uses it, because of the IOS HLE
@@ -104,6 +105,8 @@ static u32 arm_irq_masks;
 
 static u32 sensorbar_power; // do we need to care about this?
 
+int updateInterrupts;
+
 void DoState(PointerWrap &p)
 {
 	p.Do(ppc_msg);
@@ -130,6 +133,8 @@ void Init()
 	sensorbar_power = 0;
 
 	ppc_irq_masks |= INT_CAUSE_IPC_BROADWAY;
+
+	updateInterrupts = CoreTiming::RegisterEvent("IPCInterrupt", UpdateInterrupts);
 }
 
 void Reset()
@@ -222,10 +227,10 @@ void Write32(const u32 _Value, const u32 _Address)
 	}	
 
 	WII_IPC_HLE_Interface::Update();
-	UpdateInterrupts();
+	CoreTiming::ScheduleEvent_Threadsafe(0, updateInterrupts, 0);
 }
 
-void UpdateInterrupts()
+void UpdateInterrupts(u64 userdata, int cyclesLate)
 {
 	if ((ctrl.Y1 & ctrl.IY1) || (ctrl.Y2 & ctrl.IY2))
 	{
@@ -247,7 +252,7 @@ void GenerateAck(u32 _Address)
 	ctrl.Y2 = 1;
 	INFO_LOG(WII_IPC, "GenerateAck: %08x | %08x [R:%i A:%i E:%i]",
 		ppc_msg,_Address, ctrl.Y1, ctrl.Y2, ctrl.X1);
-	UpdateInterrupts();
+	CoreTiming::ScheduleEvent_Threadsafe(0, updateInterrupts, 0);
 }
 
 void GenerateReply(u32 _Address)

Source/Core/Core/Src/HW/WII_IPC.h

@@ -52,7 +52,7 @@ void DoState(PointerWrap &p);
 void Read32(u32& _rReturnValue, const u32 _Address);
 void Write32(const u32 _Value, const u32 _Address);
 
-void UpdateInterrupts();
+void UpdateInterrupts(u64 userdata = 0, int cyclesLate = 0);
 void GenerateAck(u32 _Address);
 void GenerateReply(u32 _Address);
 

Source/Core/Core/Src/PowerPC/JitArm32/JitAsm.cpp

@@ -72,7 +72,7 @@ void JitArmAsmRoutineManager::Generate()
 			MOVI2R(R14, (u32)jit->GetBlockCache()->GetICache());
 			// Confirmed That this loads the base iCache Location correctly 08-04-12
 
-			LDR(R12, R14, R12, true, true); // R12 contains iCache[PC & JIT_ICACHE_MASK] here
+			LDR(R12, R14, R12); // R12 contains iCache[PC & JIT_ICACHE_MASK] here
 			// R12 Confirmed this is the correct iCache Location loaded.
 			TST(R12, 0xFC); // Test  to see if it is a JIT block.
 
@@ -82,7 +82,7 @@ void JitArmAsmRoutineManager::Generate()
 			// LDR R14 right here to get CodePointers()[0] pointer.
 			REV(R12, R12); // Reversing this gives us our JITblock.
 			LSL(R12, R12, 2); // Multiply by four because address locations are u32 in size 
-			LDR(R14, R14, R12, true, true); // Load the block address in to R14 
+			LDR(R14, R14, R12); // Load the block address in to R14 
 
 			B(R14);
 			

Source/Core/DolphinWX/Src/FrameTools.cpp

@@ -74,6 +74,20 @@ Core::GetWindowHandle().
 
 #include <wx/datetime.h> // wxWidgets
 
+#ifdef _WIN32
+#ifndef SM_YVIRTUALSCREEN
+#define SM_YVIRTUALSCREEN 77
+#endif
+#ifndef SM_XVIRTUALSCREEN
+#define SM_XVIRTUALSCREEN 76
+#endif
+#ifndef SM_CXVIRTUALSCREEN
+#define SM_CXVIRTUALSCREEN 78
+#endif
+#ifndef SM_CYVIRTUALSCREEN
+#define SM_CYVIRTUALSCREEN 79
+#endif
+#endif
 
 // Resources
 extern "C" {
@@ -861,14 +875,23 @@ void CFrame::StartGame(const std::string& filename)
 			position = wxDefaultPosition;
 #endif
 
+		wxSize size(SConfig::GetInstance().m_LocalCoreStartupParameter.iRenderWindowWidth,
+				SConfig::GetInstance().m_LocalCoreStartupParameter.iRenderWindowHeight);
+#ifdef _WIN32
+		// Out of desktop check
+		int height = GetSystemMetrics(SM_CYVIRTUALSCREEN);
+		int width =  GetSystemMetrics(SM_CXVIRTUALSCREEN);
+		int leftPos = GetSystemMetrics(SM_XVIRTUALSCREEN);
+		int topPos = GetSystemMetrics(SM_YVIRTUALSCREEN);
+		if ((leftPos + width) < (position.x + size.GetWidth()) || leftPos > position.x || (topPos + height) < (position.y + size.GetHeight()) || topPos > position.y)
+			position.x = position.y = wxDefaultCoord;
+#endif
 		m_RenderFrame = new CRenderFrame((wxFrame*)this, wxID_ANY, _("Dolphin"), position);
 		if (SConfig::GetInstance().m_LocalCoreStartupParameter.bKeepWindowOnTop)
 			m_RenderFrame->SetWindowStyle(m_RenderFrame->GetWindowStyle() | wxSTAY_ON_TOP);
 		else
 			m_RenderFrame->SetWindowStyle(m_RenderFrame->GetWindowStyle() & ~wxSTAY_ON_TOP);
 
-		wxSize size(SConfig::GetInstance().m_LocalCoreStartupParameter.iRenderWindowWidth,
-				SConfig::GetInstance().m_LocalCoreStartupParameter.iRenderWindowHeight);
 		m_RenderFrame->SetClientSize(size.GetWidth(), size.GetHeight());
 		m_RenderFrame->Bind(wxEVT_CLOSE_WINDOW, &CFrame::OnRenderParentClose, this);
 		m_RenderFrame->Bind(wxEVT_ACTIVATE, &CFrame::OnActive, this);

Source/Core/DolphinWX/Src/ISOProperties.cpp

@@ -316,7 +316,7 @@ void CISOProperties::CreateGUIControls(bool IsWad)
 	DCBZOFF->SetToolTip(_("Bypass the clearing of the data cache by the DCBZ instruction. Usually leave this option disabled."));
 	VBeam = new wxCheckBox(m_GameConfig, ID_VBEAM, _("Accurate VBeam emulation"), wxDefaultPosition, wxDefaultSize, wxCHK_3STATE|wxCHK_ALLOW_3RD_STATE_FOR_USER, wxDefaultValidator);
 	VBeam->SetToolTip(_("If the FPS is erratic, this option may help. (ON = Compatible, OFF = Fast)"));
-	SyncGPU = new wxCheckBox(m_GameConfig, ID_SYNCGPU, _("Sychronise GPU thread"), wxDefaultPosition, wxDefaultSize, wxCHK_3STATE|wxCHK_ALLOW_3RD_STATE_FOR_USER, wxDefaultValidator);
+	SyncGPU = new wxCheckBox(m_GameConfig, ID_SYNCGPU, _("Synchronise GPU thread"), wxDefaultPosition, wxDefaultSize, wxCHK_3STATE|wxCHK_ALLOW_3RD_STATE_FOR_USER, wxDefaultValidator);
 	SyncGPU->SetToolTip(_("Synchonises the GPU and CPU threads to help prevent random freezes in Dual Core mode. (ON = Compatible, OFF = Fast)"));
 	FastDiscSpeed = new wxCheckBox(m_GameConfig, ID_DISCSPEED, _("Speed up Disc Transfer Rate"), wxDefaultPosition, wxDefaultSize, wxCHK_3STATE|wxCHK_ALLOW_3RD_STATE_FOR_USER, wxDefaultValidator);
 	FastDiscSpeed->SetToolTip(_("Enable fast disc access.  Needed for a few games. (ON = Fast, OFF = Compatible)"));
@@ -659,8 +659,8 @@ void CISOProperties::OnExtractFile(wxCommandEvent& WXUNUSED (event))
 
 	if (DiscIO::IsVolumeWiiDisc(OpenISO))
 	{
-		int partitionNum = wxAtoi(File.SubString(10, 11));
+		int partitionNum = wxAtoi(File.Mid(File.find_first_of("/"), 1));
-		File.Remove(0, 12); // Remove "Partition x/"
+		File.Remove(0, File.find_first_of("/") +1); // Remove "Partition x/"
 		WiiDisc.at(partitionNum).FileSystem->ExportFile(WxStrToStr(File).c_str(), WxStrToStr(Path).c_str());
 	}
 	else
@@ -794,8 +794,8 @@ void CISOProperties::OnExtractDir(wxCommandEvent& event)
 
 	if (DiscIO::IsVolumeWiiDisc(OpenISO))
 	{
-		int partitionNum = wxAtoi(Directory.SubString(10, 11));
+		int partitionNum = wxAtoi(Directory.Mid(Directory.find_first_of("/"), 1));
-		Directory.Remove(0, 12); // Remove "Partition x/"
+		Directory.Remove(0, Directory.find_first_of("/") +1); // Remove "Partition x/"
 		ExportDir(WxStrToStr(Directory).c_str(), WxStrToStr(Path).c_str(), partitionNum);
 	}
 	else
@@ -860,7 +860,7 @@ void CISOProperties::CheckPartitionIntegrity(wxCommandEvent& event)
 		return;
 
 	// Get the partition number from the item text ("Partition N")
-	int PartitionNum = wxAtoi(PartitionName.SubString(10, 11));
+	int PartitionNum = wxAtoi(PartitionName.Mid(PartitionName.find_first_of("0123456789"), 1));
 	const WiiPartition& Partition = WiiDisc[PartitionNum];
 
 	wxProgressDialog* dialog = new wxProgressDialog(

Source/Core/DolphinWX/Src/Main.cpp

@@ -48,6 +48,20 @@
 
 #ifdef _WIN32
 #include <shellapi.h>
+
+#ifndef SM_YVIRTUALSCREEN
+#define SM_YVIRTUALSCREEN 77
+#endif
+#ifndef SM_XVIRTUALSCREEN
+#define SM_XVIRTUALSCREEN 76
+#endif
+#ifndef SM_CXVIRTUALSCREEN
+#define SM_CXVIRTUALSCREEN 78
+#endif
+#ifndef SM_CYVIRTUALSCREEN
+#define SM_CYVIRTUALSCREEN 79
+#endif
+
 #endif
 
 #ifdef __APPLE__
@@ -166,6 +180,11 @@ bool DolphinApp::OnInit()
 			"Low level (LLE) or high level (HLE) audio",
 			wxCMD_LINE_VAL_STRING, wxCMD_LINE_PARAM_OPTIONAL
 		},
+		{
+			wxCMD_LINE_OPTION, "m", "movie",
+			"Play a movie file",
+			wxCMD_LINE_VAL_STRING, wxCMD_LINE_PARAM_OPTIONAL
+		},
 		{
 			wxCMD_LINE_NONE, NULL, NULL, NULL, wxCMD_LINE_VAL_NONE, 0
 		}
@@ -186,6 +205,7 @@ bool DolphinApp::OnInit()
 		&videoBackendName);
 	selectAudioEmulation = parser.Found(wxT("audio_emulation"),
 		&audioEmulationName);
+	playMovie = parser.Found(wxT("movie"), &movieFile);
 #endif // wxUSE_CMDLINE_PARSER
 
 #if defined _DEBUG && defined _WIN32
@@ -303,10 +323,11 @@ bool DolphinApp::OnInit()
 	// do not allow windows to be created off the desktop.
 #ifdef _WIN32
 	// Out of desktop check
-	HWND hDesktop = GetDesktopWindow();
+	int height = GetSystemMetrics(SM_CYVIRTUALSCREEN);
-	RECT rc;
+	int width =  GetSystemMetrics(SM_CXVIRTUALSCREEN);
-	GetWindowRect(hDesktop, &rc);
+	int leftPos = GetSystemMetrics(SM_XVIRTUALSCREEN);
-	if (rc.right < x + w || rc.bottom < y + h)
+	int topPos = GetSystemMetrics(SM_YVIRTUALSCREEN);
+	if ((leftPos + width) < (x + w) || leftPos > x || (topPos + height) < (y + h) || topPos > y)
 		x = y = wxDefaultCoord;
 #elif defined __APPLE__
 	if (y < 1)
@@ -346,28 +367,31 @@ void DolphinApp::AfterInit(wxTimerEvent& WXUNUSED(event))
 	if (!BatchMode)
 		main_frame->UpdateGameList();
 
-	// First check if we have an exec command line.
+	if (playMovie && movieFile != wxEmptyString)
+	{
+		if (Movie::PlayInput(movieFile.char_str()))
+		{
 			if (LoadFile && FileToLoad != wxEmptyString)
 			{
 				main_frame->BootGame(WxStrToStr(FileToLoad));
 			}
+			else
+				main_frame->BootGame(std::string(""));
+		}
+	}
+
+	// First check if we have an exec command line.
+	else if (LoadFile && FileToLoad != wxEmptyString)
+	{
+		main_frame->BootGame(WxStrToStr(FileToLoad));
+	}
 	// If we have selected Automatic Start, start the default ISO,
 	// or if no default ISO exists, start the last loaded ISO
 	else if (main_frame->g_pCodeWindow)
 	{
 		if (main_frame->g_pCodeWindow->AutomaticStart())
 		{
-			if(!SConfig::GetInstance().m_LocalCoreStartupParameter.m_strDefaultGCM.empty()
+			main_frame->BootGame("");
-					&& File::Exists(SConfig::GetInstance().m_LocalCoreStartupParameter.m_strDefaultGCM))
-			{
-				main_frame->BootGame(SConfig::GetInstance().m_LocalCoreStartupParameter.
-						m_strDefaultGCM);
-			}
-			else if(!SConfig::GetInstance().m_LastFilename.empty()
-					&& File::Exists(SConfig::GetInstance().m_LastFilename))
-			{
-				main_frame->BootGame(SConfig::GetInstance().m_LastFilename);
-			}	
 		}
 	}
 }

Source/Core/DolphinWX/Src/Main.h

@@ -40,7 +40,9 @@ private:
 	wxTimer *m_afterinit;
 	bool BatchMode;
 	bool LoadFile;
+	bool playMovie;
 	wxString FileToLoad;
+	wxString movieFile;
 	wxLocale *m_locale;
 
 	void AfterInit(wxTimerEvent& WXUNUSED(event));

Source/Core/VideoCommon/Src/BPMemory.h

@@ -93,7 +93,7 @@
 #define BPMEM_TEV_ALPHA_ENV    0xC1 // 0xC1 + (2 * 16)
 #define BPMEM_TEV_REGISTER_L   0xE0 // 0xE0 + (2 * 4)
 #define BPMEM_TEV_REGISTER_H   0xE1 // 0xE1 + (2 * 4)
-#define BPMEM_FOGRANGE         0xE8
+#define BPMEM_FOGRANGE         0xE8 // 0xE8 + 6
 #define BPMEM_FOGPARAM0        0xEE
 #define BPMEM_FOGBMAGNITUDE    0xEF
 #define BPMEM_FOGBEXPONENT     0xF0
@@ -988,7 +988,7 @@ struct BPMemory
     FourTexUnits tex[2]; //80-bf
     TevStageCombiner combiners[16]; //0xC0-0xDF
     TevReg tevregs[4];  //0xE0
-    FogRangeParams fogRange;
+    FogRangeParams fogRange;  // 0xE8
     FogParams fog; //0xEE,0xEF,0xF0,0xF1,0xF2
     AlphaTest alpha_test; //0xF3
     ZTex1 ztex1; //0xf4,0xf5

Source/Core/VideoCommon/Src/PixelShaderGen.cpp

@@ -276,7 +276,7 @@ void ValidatePixelShaderIDs(API_TYPE api, PIXELSHADERUIDSAFE old_id, const std::
 static void WriteStage(char *&p, int n, API_TYPE ApiType);
 static void SampleTexture(char *&p, const char *destination, const char *texcoords, const char *texswap, int texmap, API_TYPE ApiType);
 // static void WriteAlphaCompare(char *&p, int num, int comp);
-static void WriteAlphaTest(char *&p, API_TYPE ApiType,DSTALPHA_MODE dstAlphaMode, bool depthTextureEnable);
+static void WriteAlphaTest(char *&p, API_TYPE ApiType,DSTALPHA_MODE dstAlphaMode, bool per_pixel_depth);
 static void WriteFog(char *&p);
 
 static const char *tevKSelTableC[] = // KCSEL
@@ -512,7 +512,7 @@ const char *GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType
 	int numStages = bpmem.genMode.numtevstages + 1;
 	int numTexgen = bpmem.genMode.numtexgens;
 
-	bool depthTextureEnable = bpmem.ztex2.op != ZTEXTURE_DISABLE && !bpmem.zcontrol.early_ztest && bpmem.zmode.testenable;
+	bool per_pixel_depth = bpmem.ztex2.op != ZTEXTURE_DISABLE && !bpmem.zcontrol.early_ztest && bpmem.zmode.testenable;
 
 	char *p = text;
 	WRITE(p, "//Pixel Shader for TEV stages\n");
@@ -602,7 +602,7 @@ const char *GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType
 		if (dstAlphaMode == DSTALPHA_DUAL_SOURCE_BLEND)
 			WRITE(p, "out float4 ocol1;\n");
 		
-		if (depthTextureEnable)
+		if (per_pixel_depth)
 			WRITE(p, "#define depth gl_FragDepth\n");
 		WRITE(p, "float4 rawpos = gl_FragCoord;\n");
 
@@ -658,14 +658,14 @@ const char *GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType
 		{
 			WRITE(p, "  out float4 ocol0 : COLOR0,%s%s\n  in float4 rawpos : %s,\n",
 				dstAlphaMode == DSTALPHA_DUAL_SOURCE_BLEND ? "\n  out float4 ocol1 : COLOR1," : "",
-				depthTextureEnable ? "\n  out float depth : DEPTH," : "",
+				per_pixel_depth ? "\n  out float depth : DEPTH," : "",
 				ApiType & API_D3D9_SM20 ? "POSITION" : "VPOS");
 		}
 		else
 		{
 			WRITE(p, "  out float4 ocol0 : SV_Target0,%s%s\n  in float4 rawpos : SV_Position,\n",
 				dstAlphaMode == DSTALPHA_DUAL_SOURCE_BLEND ? "\n  out float4 ocol1 : SV_Target1," : "",
-				depthTextureEnable ? "\n  out float depth : SV_Depth," : "");
+				per_pixel_depth ? "\n  out float depth : SV_Depth," : "");
 		}
 
 		WRITE(p, "  in float4 colors_0 : COLOR0,\n");
@@ -808,7 +808,7 @@ const char *GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType
 
 	AlphaTest::TEST_RESULT Pretest = bpmem.alpha_test.TestResult();
 	if (Pretest == AlphaTest::UNDETERMINED)
-		WriteAlphaTest(p, ApiType, dstAlphaMode, depthTextureEnable);
+		WriteAlphaTest(p, ApiType, dstAlphaMode, per_pixel_depth);
 
 	
 	// the screen space depth value = far z + (clip z / clip w) * z range
@@ -818,8 +818,9 @@ const char *GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType
 		// dx9 doesn't support 4 component position, so we have to calculate it again
 		WRITE(p, "float zCoord = " I_ZBIAS"[1].x + (clipPos.z / clipPos.w) * " I_ZBIAS"[1].y;\n");
 
-	// Note: depth textures are disabled if early depth test is enabled
+	// depth texture can safely be ignored if the result won't be written to the depth buffer (early_ztest) and isn't used for fog either
-	if (depthTextureEnable)
+	bool skip_ztexture = !per_pixel_depth && !bpmem.fog.c_proj_fsel.fsel;
+	if (bpmem.ztex2.op != ZTEXTURE_DISABLE && !skip_ztexture)
 	{
 		// use the texture input of the last texture stage (textemp), hopefully this has been read and is in correct format...
 		WRITE(p, "zCoord = dot(" I_ZBIAS"[0].xyzw, textemp.xyzw) + " I_ZBIAS"[1].w %s;\n",
@@ -830,6 +831,8 @@ const char *GeneratePixelShaderCode(DSTALPHA_MODE dstAlphaMode, API_TYPE ApiType
 		WRITE(p, "zCoord = frac(zCoord);\n");
 		WRITE(p, "zCoord = zCoord * (16777216.0f/16777215.0f);\n");
 
+		// Note: depth texture out put is only written to depth buffer if late depth test is used
+		if (per_pixel_depth)
 			WRITE(p, "depth = zCoord;\n");
 	}
 
@@ -1242,7 +1245,7 @@ static const char *tevAlphaFunclogicTable[] =
 	" == "  // xnor
 };
 
-static void WriteAlphaTest(char *&p, API_TYPE ApiType,DSTALPHA_MODE dstAlphaMode, bool depthTextureEnable)
+static void WriteAlphaTest(char *&p, API_TYPE ApiType,DSTALPHA_MODE dstAlphaMode, bool per_pixel_depth)
 {
 	static const char *alphaRef[2] =
 	{
@@ -1266,7 +1269,7 @@ static void WriteAlphaTest(char *&p, API_TYPE ApiType,DSTALPHA_MODE dstAlphaMode
 	WRITE(p, "\t\tocol0 = float4(0.0f, 0.0f, 0.0f, 0.0f);\n");
 	if (dstAlphaMode == DSTALPHA_DUAL_SOURCE_BLEND)
 		WRITE(p, "\t\tocol1 = float4(0.0f, 0.0f, 0.0f, 0.0f);\n");
-	if(depthTextureEnable)
+	if(per_pixel_depth)
 		WRITE(p, "depth = 1.f;\n");
 
 	// HAXX: zcomploc (aka early_ztest) is a way to control whether depth test is done before

Source/Core/VideoCommon/Src/PixelShaderManager.cpp

@@ -415,6 +415,7 @@ void PixelShaderManager::SetZTextureBias(u32 bias)
 void PixelShaderManager::SetViewportChanged()
 {
 	s_bDepthRangeChanged = true;
+	s_bFogRangeAdjustChanged = true; // TODO: Shouldn't be necessary with an accurate fog range adjust implementation
 }
 
 void PixelShaderManager::SetIndTexScaleChanged(u8 stagemask)

Source/Core/VideoCommon/Src/VertexManagerBase.h

@@ -2,6 +2,7 @@
 #ifndef _VERTEXMANAGERBASE_H
 #define _VERTEXMANAGERBASE_H
 
+#include "Common.h"
 #include <vector>
 
 class NativeVertexFormat;
@@ -10,17 +11,16 @@ class PointerWrap;
 class VertexManager
 {
 private:
-	// What are the actual values?
+	static const u32 SMALLEST_POSSIBLE_VERTEX = sizeof(float)*3;                 // 3 pos
-	static const u32 SMALLEST_POSSIBLE_VERTEX = 1;
+	static const u32 LARGEST_POSSIBLE_VERTEX = sizeof(float)*45 + sizeof(u32)*2; // 3 pos, 3*3 normal, 2*u32 color, 8*4 tex, 1 posMat 
-	static const u32 LARGEST_POSSIBLE_VERTEX = 188;
 	
 	static const u32 MAX_PRIMITIVES_PER_COMMAND = (u16)-1;
 	
 public:
-	static const u32 MAXVBUFFERSIZE = MAX_PRIMITIVES_PER_COMMAND * LARGEST_POSSIBLE_VERTEX;
+	static const u32 MAXVBUFFERSIZE = ROUND_UP_POW2 (MAX_PRIMITIVES_PER_COMMAND * LARGEST_POSSIBLE_VERTEX);
 	
 	// We may convert triangle-fans to triangle-lists, almost 3x as many indices.
-	static const u32 MAXIBUFFERSIZE = MAX_PRIMITIVES_PER_COMMAND * 3;
+	static const u32 MAXIBUFFERSIZE = ROUND_UP_POW2 (MAX_PRIMITIVES_PER_COMMAND * 3);
 
 	VertexManager();
 	// needs to be virtual for DX11's dtor

Source/Plugins/Plugin_VideoDX11/Src/VertexManager.cpp

@@ -39,9 +39,9 @@ namespace DX11
 {
 
 // TODO: Find sensible values for these two
-const UINT IBUFFER_SIZE = VertexManager::MAXIBUFFERSIZE * 16 * sizeof(u16);
+const UINT IBUFFER_SIZE = VertexManager::MAXIBUFFERSIZE * sizeof(u16) * 8;
-const UINT VBUFFER_SIZE = VertexManager::MAXVBUFFERSIZE * 16;
+const UINT VBUFFER_SIZE = VertexManager::MAXVBUFFERSIZE;
-const UINT MAX_VBUFFER_COUNT = 2;
+const UINT MAXVBUFFER_COUNT = 2;
 
 void VertexManager::CreateDeviceObjects()
 {

Source/Plugins/Plugin_VideoDX9/Src/VertexManager.cpp

@@ -42,9 +42,9 @@ extern NativeVertexFormat *g_nativeVertexFmt;
 namespace DX9
 {
 //This are the initially requeted size for the buffers expresed in elements
-const u32 IBUFFER_SIZE = VertexManager::MAXIBUFFERSIZE * 16;
+const u32 IBUFFER_SIZE = VertexManager::MAXIBUFFERSIZE * sizeof(u16) * 8;
-const u32 VBUFFER_SIZE = VertexManager::MAXVBUFFERSIZE * 16;
+const u32 VBUFFER_SIZE = VertexManager::MAXVBUFFERSIZE;
-const u32 MAX_VBUFFER_COUNT = 2;
+const u32 MAXVBUFFER_COUNT = 2;
 
 inline void DumpBadShaders()
 {

Source/Plugins/Plugin_VideoSoftware/Src/EfbInterface.cpp

@@ -463,6 +463,7 @@ namespace EfbInterface
 
 	void SetDepth(u16 x, u16 y, u32 depth)
 	{
+		if (bpmem.zmode.updateenable)
 			SetPixelDepth(GetDepthOffset(x, y), depth);
 	}