diff --git a/Source/Core/Common/x64ABI.cpp b/Source/Core/Common/x64ABI.cpp
index 2815ab8495..44818ceb67 100644
--- a/Source/Core/Common/x64ABI.cpp
+++ b/Source/Core/Common/x64ABI.cpp
@@ -207,18 +207,18 @@ void XEmitter::MOVTwo(int bits, Gen::X64Reg dst1, Gen::X64Reg src1, Gen::X64Reg
 	}
 }
 
-void XEmitter::ABI_CallFunctionAC(const void *func, const Gen::OpArg &arg1, u32 param2)
+void XEmitter::ABI_CallFunctionAC(int bits, const void *func, const Gen::OpArg &arg1, u32 param2)
 {
 	if (!arg1.IsSimpleReg(ABI_PARAM1))
-		MOV(32, R(ABI_PARAM1), arg1);
+		MOV(bits, R(ABI_PARAM1), arg1);
 	MOV(32, R(ABI_PARAM2), Imm32(param2));
 	ABI_CallFunction(func);
 }
 
-void XEmitter::ABI_CallFunctionA(const void *func, const Gen::OpArg &arg1)
+void XEmitter::ABI_CallFunctionA(int bits, const void *func, const Gen::OpArg &arg1)
 {
 	if (!arg1.IsSimpleReg(ABI_PARAM1))
-		MOV(32, R(ABI_PARAM1), arg1);
+		MOV(bits, R(ABI_PARAM1), arg1);
 	ABI_CallFunction(func);
 }
 
diff --git a/Source/Core/Common/x64Emitter.cpp b/Source/Core/Common/x64Emitter.cpp
index 3716d922aa..cbdae83c91 100644
--- a/Source/Core/Common/x64Emitter.cpp
+++ b/Source/Core/Common/x64Emitter.cpp
@@ -1184,7 +1184,7 @@ void OpArg::WriteNormalOp(XEmitter *emit, bool toRM, NormalOp op, const OpArg &o
 		}
 		else
 		{
-			_assert_msg_(DYNA_REC, 0, "WriteNormalOp - Unhandled case");
+			_assert_msg_(DYNA_REC, 0, "WriteNormalOp - Unhandled case %d %d", operand.scale, bits);
 		}
 		_operandReg = (X64Reg)normalops[op].ext; //pass extension in REG of ModRM
 	}
diff --git a/Source/Core/Common/x64Emitter.h b/Source/Core/Common/x64Emitter.h
index 7f98497456..9df47e10eb 100644
--- a/Source/Core/Common/x64Emitter.h
+++ b/Source/Core/Common/x64Emitter.h
@@ -871,8 +871,8 @@ public:
 	void ABI_CallFunctionCCCP(const void *func, u32 param1, u32 param2,u32 param3, void *param4);
 	void ABI_CallFunctionPC(const void *func, void *param1, u32 param2);
 	void ABI_CallFunctionPPC(const void *func, void *param1, void *param2, u32 param3);
-	void ABI_CallFunctionAC(const void *func, const OpArg &arg1, u32 param2);
-	void ABI_CallFunctionA(const void *func, const OpArg &arg1);
+	void ABI_CallFunctionAC(int bits, const void *func, const OpArg &arg1, u32 param2);
+	void ABI_CallFunctionA(int bits, const void *func, const OpArg &arg1);
 
 	// Pass a register as a parameter.
 	void ABI_CallFunctionR(const void *func, X64Reg reg1);
diff --git a/Source/Core/Core/PowerPC/Jit64/JitAsm.cpp b/Source/Core/Core/PowerPC/Jit64/JitAsm.cpp
index ae1d149e56..38c1d52ab8 100644
--- a/Source/Core/Core/PowerPC/Jit64/JitAsm.cpp
+++ b/Source/Core/Core/PowerPC/Jit64/JitAsm.cpp
@@ -130,7 +130,7 @@ void Jit64AsmRoutineManager::Generate()
 
 			//Ok, no block, let's jit
 			ABI_PushRegistersAndAdjustStack({}, 0);
-			ABI_CallFunctionA((void *)&Jit, PPCSTATE(pc));
+			ABI_CallFunctionA(32, (void *)&Jit, PPCSTATE(pc));
 			ABI_PopRegistersAndAdjustStack({}, 0);
 
 			// Jit might have cleared the code cache
@@ -186,6 +186,8 @@ void Jit64AsmRoutineManager::GenerateCommon()
 	GenFifoWrite(16);
 	fifoDirectWrite32 = AlignCode4();
 	GenFifoWrite(32);
+	fifoDirectWrite64 = AlignCode4();
+	GenFifoWrite(64);
 	frsqrte = AlignCode4();
 	GenFrsqrte();
 	fres = AlignCode4();
diff --git a/Source/Core/Core/PowerPC/JitCommon/JitAsmCommon.h b/Source/Core/Core/PowerPC/JitCommon/JitAsmCommon.h
index f38df7f266..847d318230 100644
--- a/Source/Core/Core/PowerPC/JitCommon/JitAsmCommon.h
+++ b/Source/Core/Core/PowerPC/JitCommon/JitAsmCommon.h
@@ -13,6 +13,7 @@ public:
 	const u8 *fifoDirectWrite8;
 	const u8 *fifoDirectWrite16;
 	const u8 *fifoDirectWrite32;
+	const u8 *fifoDirectWrite64;
 
 	const u8 *enterCode;
 
@@ -55,8 +56,6 @@ protected:
 
 public:
 	void GenFifoWrite(int size);
-	void GenFifoXmm64Write();
-	void GenFifoFloatWrite();
 	void GenFrsqrte();
 	void GenFres();
 };
diff --git a/Source/Core/Core/PowerPC/JitCommon/Jit_Util.cpp b/Source/Core/Core/PowerPC/JitCommon/Jit_Util.cpp
index 6066e334d2..fe6d6b2f2c 100644
--- a/Source/Core/Core/PowerPC/JitCommon/Jit_Util.cpp
+++ b/Source/Core/Core/PowerPC/JitCommon/Jit_Util.cpp
@@ -464,6 +464,17 @@ u8 *EmuCodeBlock::UnsafeWriteRegToReg(OpArg reg_value, X64Reg reg_addr, int acce
 	return result;
 }
 
+static OpArg FixImmediate(int accessSize, OpArg arg)
+{
+	if (arg.IsImm())
+	{
+		arg = accessSize == 8  ? Imm8((u8)arg.offset) :
+		      accessSize == 16 ? Imm16((u16)arg.offset) :
+		                         Imm32((u32)arg.offset);
+	}
+	return arg;
+}
+
 void EmuCodeBlock::UnsafeWriteGatherPipe(int accessSize)
 {
 	// No need to protect these, they don't touch any state
@@ -479,18 +490,23 @@ void EmuCodeBlock::UnsafeWriteGatherPipe(int accessSize)
 	case 32:
 		CALL((void *)jit->GetAsmRoutines()->fifoDirectWrite32);
 		break;
+	case 64:
+		CALL((void *)jit->GetAsmRoutines()->fifoDirectWrite64);
+		break;
 	}
 	jit->js.fifoBytesThisBlock += accessSize >> 3;
 }
 
 bool EmuCodeBlock::WriteToConstAddress(int accessSize, OpArg arg, u32 address, BitSet32 registersInUse)
 {
+	arg = FixImmediate(accessSize, arg);
+
 	// If we already know the address through constant folding, we can do some
 	// fun tricks...
-	if ((address & 0xFFFFF000) == 0xCC008000 && jit->jo.optimizeGatherPipe && accessSize <= 32)
+	if ((address & 0xFFFFF000) == 0xCC008000 && jit->jo.optimizeGatherPipe)
 	{
 		if (!arg.IsSimpleReg() || arg.GetSimpleReg() != RSCRATCH)
-			MOV(32, R(RSCRATCH), arg);
+			MOV(accessSize, R(RSCRATCH), arg);
 
 		UnsafeWriteGatherPipe(accessSize);
 		return false;
@@ -509,16 +525,16 @@ bool EmuCodeBlock::WriteToConstAddress(int accessSize, OpArg arg, u32 address, B
 		switch (accessSize)
 		{
 		case 64:
-			ABI_CallFunctionAC((void *)&Memory::Write_U64, arg, address);
+			ABI_CallFunctionAC(64, (void *)&Memory::Write_U64, arg, address);
 			break;
 		case 32:
-			ABI_CallFunctionAC((void *)&Memory::Write_U32, arg, address);
+			ABI_CallFunctionAC(32, (void *)&Memory::Write_U32, arg, address);
 			break;
 		case 16:
-			ABI_CallFunctionAC((void *)&Memory::Write_U16, arg, address);
+			ABI_CallFunctionAC(16, (void *)&Memory::Write_U16, arg, address);
 			break;
 		case 8:
-			ABI_CallFunctionAC((void *)&Memory::Write_U8, arg, address);
+			ABI_CallFunctionAC(8, (void *)&Memory::Write_U8, arg, address);
 			break;
 		}
 		ABI_PopRegistersAndAdjustStack(registersInUse, 0);
@@ -529,12 +545,7 @@ bool EmuCodeBlock::WriteToConstAddress(int accessSize, OpArg arg, u32 address, B
 void EmuCodeBlock::SafeWriteRegToReg(OpArg reg_value, X64Reg reg_addr, int accessSize, s32 offset, BitSet32 registersInUse, int flags)
 {
 	// set the correct immediate format
-	if (reg_value.IsImm())
-	{
-		reg_value = accessSize == 32 ? Imm32((u32)reg_value.offset) :
-		            accessSize == 16 ? Imm16((u16)reg_value.offset) :
-		                               Imm8((u8)reg_value.offset);
-	}
+	reg_value = FixImmediate(accessSize, reg_value);
 
 	// TODO: support byte-swapped non-immediate fastmem stores
 	if (!SConfig::GetInstance().m_LocalCoreStartupParameter.bMMU &&