diff --git a/rpcs3/Emu/CPU/CPUThread.h b/rpcs3/Emu/CPU/CPUThread.h
index bde7140c8b..9568e80dbd 100644
--- a/rpcs3/Emu/CPU/CPUThread.h
+++ b/rpcs3/Emu/CPU/CPUThread.h
@@ -1,6 +1,6 @@
 #pragma once
 #include "Emu/Memory/MemoryBlock.h"
-#include "Emu/Cell/PPCDecoder.h"
+#include "Emu/CPU/CPUDecoder.h"
 
 enum CPUThreadType
 {
diff --git a/rpcs3/Emu/Cell/MFC.h b/rpcs3/Emu/Cell/MFC.h
index 7031505600..c928078c46 100644
--- a/rpcs3/Emu/Cell/MFC.h
+++ b/rpcs3/Emu/Cell/MFC.h
@@ -163,16 +163,16 @@ struct DMAC
 		if(proxy_pos)
 		{
 			DMAC_Proxy p = proxy[0];
-			memcpy(proxy, proxy + 1, proxy_pos--);
+			memcpy(proxy, proxy + 1, --proxy_pos * sizeof(DMAC_Proxy));
 
 			switch(p.cmd)
 			{
 			case MFC_PUT_CMD:
-				memcpy(Memory + ls_offset + p.lsa, Memory + p.ea, p.size);
+				memcpy(Memory + p.ea, Memory + ls_offset + p.lsa, p.size);
 			break;
 
 			case MFC_GET_CMD:
-				memcpy(Memory + p.ea, Memory + ls_offset + p.lsa, p.size);
+				memcpy(Memory + ls_offset + p.lsa, Memory + p.ea, p.size);
 			break;
 
 			default:
diff --git a/rpcs3/Emu/Cell/RawSPUThread.cpp b/rpcs3/Emu/Cell/RawSPUThread.cpp
index 6cf6b212f8..86d8addf0c 100644
--- a/rpcs3/Emu/Cell/RawSPUThread.cpp
+++ b/rpcs3/Emu/Cell/RawSPUThread.cpp
@@ -11,7 +11,14 @@ RawSPUThread::RawSPUThread(u32 index, CPUThreadType type)
 
 RawSPUThread::~RawSPUThread()
 {
-	MemoryBlock::Delete();
+	for(int i=0; i<Memory.MemoryBlocks.GetCount(); ++i)
+	{
+		if(&Memory.MemoryBlocks[i] == this)
+		{
+			Memory.MemoryBlocks.RemoveFAt(i);
+			return;
+		}
+	}
 }
 
 bool RawSPUThread::Read8(const u64 addr, u8* value)
@@ -59,9 +66,13 @@ bool RawSPUThread::Read32(const u64 addr, u32* value)
 	case Prxy_QueryType_offs:		ConLog.Warning("RawSPUThread[%d]: Read32(Prxy_QueryType)", m_index);	*value = Prxy.QueryType.GetValue(); break;
 	case Prxy_QueryMask_offs:		ConLog.Warning("RawSPUThread[%d]: Read32(Prxy_QueryMask)", m_index);	*value = Prxy.QueryMask.GetValue(); break;
 	case Prxy_TagStatus_offs:		ConLog.Warning("RawSPUThread[%d]: Read32(Prxy_TagStatus)", m_index);	*value = Prxy.TagStatus.GetValue(); break;
-	case SPU_Out_MBox_offs:			ConLog.Warning("RawSPUThread[%d]: Read32(SPU_Out_MBox)", m_index);		*value = SPU.Out_MBox.GetValue(); break;
+	case SPU_Out_MBox_offs:			ConLog.Warning("RawSPUThread[%d]: Read32(SPU_Out_MBox)", m_index);		while(!SPU.Out_MBox.Pop(*value) && !Emu.IsStopped()) Sleep(1); break;
 	case SPU_In_MBox_offs:			ConLog.Warning("RawSPUThread[%d]: Read32(SPU_In_MBox)", m_index);		*value = SPU.In_MBox.GetValue(); break;
-	case SPU_MBox_Status_offs:		ConLog.Warning("RawSPUThread[%d]: Read32(SPU_MBox_Status)", m_index);	*value = SPU.MBox_Status.GetValue(); break;
+	case SPU_MBox_Status_offs:		//ConLog.Warning("RawSPUThread[%d]: Read32(SPU_MBox_Status)", m_index);
+		SPU.MBox_Status.SetValue(SPU.Out_MBox.GetCount() ? SPU.MBox_Status.GetValue() | 1 : SPU.MBox_Status.GetValue() & ~1);
+		SPU.MBox_Status.SetValue((SPU.MBox_Status.GetValue() & ~0xff00) | (SPU.In_MBox.GetCount() << 8));
+		*value = SPU.MBox_Status.GetValue();
+		break;
 	case SPU_RunCntl_offs:			ConLog.Warning("RawSPUThread[%d]: Read32(SPU_RunCntl)", m_index);		*value = SPU.RunCntl.GetValue(); break;
 	case SPU_Status_offs:			ConLog.Warning("RawSPUThread[%d]: Read32(SPU_Status)", m_index);		*value = SPU.Status.GetValue(); break;
 	case SPU_NPC_offs:				ConLog.Warning("RawSPUThread[%d]: Read32(SPU_NPC)", m_index);			*value = SPU.NPC.GetValue(); break;
@@ -201,7 +212,7 @@ bool RawSPUThread::Write32(const u64 addr, const u32 value)
 	break;
 	case Prxy_QueryMask_offs:		ConLog.Warning("RawSPUThread[%d]: Write32(Prxy_QueryMask, 0x%x)", m_index, value);		Prxy.QueryMask.SetValue(value); break;
 	case Prxy_TagStatus_offs:		ConLog.Warning("RawSPUThread[%d]: Write32(Prxy_TagStatus, 0x%x)", m_index, value);		Prxy.TagStatus.SetValue(value); break;
-	case SPU_Out_MBox_offs:			ConLog.Warning("RawSPUThread[%d]: Write32(SPU_Out_MBox, 0x%x)", m_index, value);		SPU.Out_MBox.SetValue(value); break;
+	case SPU_Out_MBox_offs:			ConLog.Warning("RawSPUThread[%d]: Write32(SPU_Out_MBox, 0x%x)", m_index, value);		while(!SPU.Out_MBox.Push(value) && !Emu.IsStopped()) Sleep(1); break;
 	case SPU_In_MBox_offs:			ConLog.Warning("RawSPUThread[%d]: Write32(SPU_In_MBox, 0x%x)", m_index, value);			SPU.In_MBox.SetValue(value); break;
 	case SPU_MBox_Status_offs:		ConLog.Warning("RawSPUThread[%d]: Write32(SPU_MBox_Status, 0x%x)", m_index, value);		SPU.MBox_Status.SetValue(value); break;
 	case SPU_RunCntl_offs:			ConLog.Warning("RawSPUThread[%d]: Write32(SPU_RunCntl, 0x%x)", m_index, value);			SPU.RunCntl.SetValue(value); break;
@@ -247,7 +258,7 @@ bool RawSPUThread::Write128(const u64 addr, const u128 value)
 
 void RawSPUThread::InitRegs()
 {
-	m_offset = GetStartAddr() + RAW_SPU_LS_OFFSET;
+	dmac.ls_offset = m_offset = GetStartAddr() + RAW_SPU_LS_OFFSET;
 	SPUThread::InitRegs();
 }
 
@@ -289,6 +300,8 @@ void RawSPUThread::Task()
 				continue;
 			}
 
+			dmac.DoCmd();
+
 			if(SPU.RunCntl.GetValue() != SPU_RUNCNTL_RUNNABLE)
 			{
 				if(!is_last_paused)
@@ -311,6 +324,7 @@ void RawSPUThread::Task()
 				SPU.Status.SetValue(SPU_STATUS_RUNNING);
 			}
 
+			Step();
 			NextPc(m_dec->DecodeMemory(PC + m_offset));
 
 			for(uint i=0; i<bp.GetCount(); ++i)
diff --git a/rpcs3/Emu/Cell/SPUInterpreter.h b/rpcs3/Emu/Cell/SPUInterpreter.h
index 69eb99009b..f225baa5a5 100644
--- a/rpcs3/Emu/Cell/SPUInterpreter.h
+++ b/rpcs3/Emu/Cell/SPUInterpreter.h
@@ -30,6 +30,11 @@ private:
 			CPU.SetExitStatus(code & 0xfff);
 			CPU.Stop();
 		}
+		else
+		{
+			ConLog.Warning("STOP: 0x%x", code);
+			Emu.Pause();
+		}
 	}
 	void LNOP()
 	{
@@ -272,23 +277,23 @@ private:
 	}
 	void BIZ(u32 rt, u32 ra)
 	{
-		if(CPU.GPR[rt]._u32[0] == 0)
-			CPU.SetBranch(CPU.GPR[ra]._u32[0] & 0xfffffffc);
+		if(CPU.GPR[rt]._u32[3] == 0)
+			CPU.SetBranch(CPU.GPR[ra]._u32[3] & 0xfffffffc);
 	}
 	void BINZ(u32 rt, u32 ra)
 	{
-		if(CPU.GPR[rt]._u32[0] != 0)
-			CPU.SetBranch(CPU.GPR[ra]._u32[0] & 0xfffffffc);
+		if(CPU.GPR[rt]._u32[3] != 0)
+			CPU.SetBranch(CPU.GPR[ra]._u32[3] & 0xfffffffc);
 	}
 	void BIHZ(u32 rt, u32 ra)
 	{
-		if(CPU.GPR[rt]._u16[0] == 0)
-			CPU.SetBranch(CPU.GPR[ra]._u32[0] & 0xfffffffc);
+		if(CPU.GPR[rt]._u16[7] == 0)
+			CPU.SetBranch(CPU.GPR[ra]._u32[3] & 0xfffffffc);
 	}
 	void BIHNZ(u32 rt, u32 ra)
 	{
-		if(CPU.GPR[rt]._u16[0] != 0)
-			CPU.SetBranch(CPU.GPR[ra]._u32[0] & 0xfffffffc);
+		if(CPU.GPR[rt]._u16[7] != 0)
+			CPU.SetBranch(CPU.GPR[ra]._u32[3] & 0xfffffffc);
 	}
 	void STOPD(u32 rc, u32 ra, u32 rb)
 	{
@@ -296,7 +301,15 @@ private:
 	}
 	void STQX(u32 rt, u32 ra, u32 rb)
 	{
-		CPU.WriteLS128(CPU.GPR[ra]._u32[0] + CPU.GPR[rb]._u32[0], CPU.GPR[rt]._u128);
+		u32 lsa = CPU.GPR[ra]._u32[3] + CPU.GPR[rb]._u32[3];
+		if(!CPU.IsGoodLSA(lsa))
+		{
+			ConLog.Error("STQX: bad lsa (0x%x)", lsa);
+			Emu.Pause();
+			return;
+		}
+
+		CPU.WriteLS128(lsa, CPU.GPR[rt]._u128);
 	}
 	void BI(u32 ra)
 	{
@@ -368,7 +381,15 @@ private:
 	}
 	void LQX(u32 rt, u32 ra, u32 rb)
 	{
-		CPU.GPR[rt]._u128 = CPU.ReadLS128(CPU.GPR[ra]._u32[0] + CPU.GPR[rb]._u32[0]);
+		u32 lsa = CPU.GPR[ra]._u32[3] + CPU.GPR[rb]._u32[3];
+		if(!CPU.IsGoodLSA(lsa))
+		{
+			ConLog.Error("LQX: bad lsa (0x%x)", lsa);
+			Emu.Pause();
+			return;
+		}
+
+		CPU.GPR[rt]._u128 = CPU.ReadLS128(lsa);
 	}
 	void ROTQBYBI(u32 rt, u32 ra, u32 rb)
 	{
@@ -957,7 +978,15 @@ private:
 	}
 	void STQA(u32 rt, s32 i16)
 	{
-		CPU.WriteLS128(i16 << 2, CPU.GPR[rt]._u128);
+		u32 lsa = i16 << 2;
+		if(!CPU.IsGoodLSA(lsa))
+		{
+			ConLog.Error("STQA: bad lsa (0x%x)", lsa);
+			Emu.Pause();
+			return;
+		}
+
+		CPU.WriteLS128(lsa, CPU.GPR[rt]._u128);
 	}
 	void BRNZ(u32 rt, s32 i16)
 	{
@@ -974,7 +1003,15 @@ private:
 	}
 	void STQR(u32 rt, s32 i16)
 	{
-		CPU.WriteLS128(branchTarget(CPU.PC, i16), CPU.GPR[rt]._u128);
+		u32 lsa = branchTarget(CPU.PC, i16);
+		if(!CPU.IsGoodLSA(lsa))
+		{
+			ConLog.Error("STQR: bad lsa (0x%x)", lsa);
+			Emu.Pause();
+			return;
+		}
+
+		CPU.WriteLS128(lsa, CPU.GPR[rt]._u128);
 	}
 	void BRA(s32 i16)
 	{
@@ -982,7 +1019,15 @@ private:
 	}
 	void LQA(u32 rt, s32 i16)
 	{
-		CPU.GPR[rt]._u128 = CPU.ReadLS128(i16 << 2);
+		u32 lsa = i16 << 2;
+		if(!CPU.IsGoodLSA(lsa))
+		{
+			ConLog.Error("LQA: bad lsa (0x%x)", lsa);
+			Emu.Pause();
+			return;
+		}
+
+		CPU.GPR[rt]._u128 = CPU.ReadLS128(lsa);
 	}
 	void BRASL(u32 rt, s32 i16)
 	{
@@ -1018,7 +1063,15 @@ private:
 	}
 	void LQR(u32 rt, s32 i16)
 	{
-		CPU.GPR[rt]._u128 = CPU.ReadLS128(branchTarget(CPU.PC, i16));
+		u32 lsa = branchTarget(CPU.PC, i16);
+		if(!CPU.IsGoodLSA(lsa))
+		{
+			ConLog.Error("LQR: bad lsa (0x%x)", lsa);
+			Emu.Pause();
+			return;
+		}
+
+		CPU.GPR[rt]._u128 = CPU.ReadLS128(lsa);
 	}
 	void IL(u32 rt, s32 i16)
 	{
@@ -1103,11 +1156,26 @@ private:
 	}
 	void STQD(u32 rt, s32 i10, u32 ra)
 	{
-		CPU.WriteLS128(CPU.GPR[ra]._u32[3] + i10, CPU.GPR[rt]._u128);
+		u32 lsa = CPU.GPR[ra]._u32[3] + i10;
+		if(!CPU.IsGoodLSA(lsa))
+		{
+			ConLog.Error("STQD: bad lsa (0x%x)", lsa);
+			Emu.Pause();
+			return;
+		}
+		CPU.WriteLS128(lsa, CPU.GPR[rt]._u128);
 	}
 	void LQD(u32 rt, s32 i10, u32 ra)
 	{
-		CPU.GPR[rt]._u128 = CPU.ReadLS128(CPU.GPR[ra]._u32[3] + i10);
+		u32 lsa = CPU.GPR[ra]._u32[3] + i10;
+		if(!CPU.IsGoodLSA(lsa))
+		{
+			ConLog.Error("LQD: bad lsa (0x%x)", lsa);
+			Emu.Pause();
+			return;
+		}
+
+		CPU.GPR[rt]._u128 = CPU.ReadLS128(lsa);
 	}
 	void XORI(u32 rt, u32 ra, s32 i10)
 	{
@@ -1203,7 +1271,6 @@ private:
 	void HBRA(s32 ro, s32 i16)
 	{
 		UNIMPLEMENTED();
-		UNIMPLEMENTED();
 	}
 	void HBRR(s32 ro, s32 i16)
 	{
@@ -1228,7 +1295,7 @@ private:
 	}
 	void SHUFB(u32 rt, u32 ra, u32 rb, u32 rc)
 	{
-		ConLog.Warning("SHUFB");
+		//ConLog.Warning("SHUFB");
 	}
 	void MPYA(u32 rc, u32 ra, u32 rb, u32 rt)
 	{
diff --git a/rpcs3/Emu/Cell/SPUThread.cpp b/rpcs3/Emu/Cell/SPUThread.cpp
index 332aa605ae..8bc5a30dbc 100644
--- a/rpcs3/Emu/Cell/SPUThread.cpp
+++ b/rpcs3/Emu/Cell/SPUThread.cpp
@@ -37,7 +37,7 @@ void SPUThread::DoReset()
 
 void SPUThread::InitRegs()
 {
-	//GPR[1]._u64[0] = stack_point;
+	GPR[1]._u32[3] = 0x40000 - 120;
 	GPR[3]._u64[1] = m_args[0];
 	GPR[4]._u64[1] = m_args[1];
 	GPR[5]._u64[1] = m_args[2];
@@ -56,7 +56,7 @@ void SPUThread::InitRegs()
 
 u64 SPUThread::GetFreeStackSize() const
 {
-	return (GetStackAddr() + GetStackSize()) - GPR[1]._u64[0];
+	return (GetStackAddr() + GetStackSize()) - GPR[1]._u64[3];
 }
 
 void SPUThread::DoRun()
@@ -72,6 +72,9 @@ void SPUThread::DoRun()
 		m_dec = new SPUDecoder(*new SPUInterpreter(*this));
 	break;
 	}
+
+	Pause();
+	//Emu.Pause();
 }
 
 void SPUThread::DoResume()
@@ -85,5 +88,5 @@ void SPUThread::DoPause()
 void SPUThread::DoStop()
 {
 	delete m_dec;
-	m_dec = 0;
+	m_dec = nullptr;
 }
diff --git a/rpcs3/Emu/Cell/SPUThread.h b/rpcs3/Emu/Cell/SPUThread.h
index 52d3c2bb8e..2a016e27c4 100644
--- a/rpcs3/Emu/Cell/SPUThread.h
+++ b/rpcs3/Emu/Cell/SPUThread.h
@@ -359,7 +359,7 @@ public:
 			return SPU.Out_MBox.GetFreeCount();
 
 		case SPU_RdInMbox:
-			return SPU.In_MBox.GetFreeCount();
+			return SPU.In_MBox.GetCount();
 
 		case SPU_WrOutIntrMbox:
 			return 0;//return SPU.OutIntr_Mbox.GetFreeCount();
@@ -380,19 +380,20 @@ public:
 		{
 		case SPU_WrOutIntrMbox:
 			ConLog.Warning("SPU_WrOutIntrMbox = 0x%x", v);
-			if(!SPU.OutIntr_Mbox.Push(v))
+
+			while(!SPU.OutIntr_Mbox.Push(v) && !Emu.IsStopped())
 			{
-				ConLog.Warning("Not enought free rooms.");
+				Sleep(1);
 			}
 		break;
 
 		case SPU_WrOutMbox:
 			ConLog.Warning("SPU_WrOutMbox = 0x%x", v);
-			if(!SPU.Out_MBox.Push(v))
+
+			while(!SPU.Out_MBox.Push(v) && !Emu.IsStopped())
 			{
-				ConLog.Warning("Not enought free rooms.");
+				Sleep(1);
 			}
-			SPU.Status.SetValue((SPU.Status.GetValue() & ~0xff) | 1);
 		break;
 
 		default:
@@ -410,7 +411,7 @@ public:
 		{
 		case SPU_RdInMbox:
 			if(!SPU.In_MBox.Pop(v)) v = 0;
-			SPU.Status.SetValue((SPU.Status.GetValue() & ~0xff00) | (SPU.In_MBox.GetCount() << 8));
+			ConLog.Warning("%s: SPU_RdInMbox(0x%x).", __FUNCTION__, v);
 		break;
 
 		default:
@@ -419,7 +420,7 @@ public:
 		}
 	}
 
-	bool IsGoodLSA(const u32 lsa) const { return Memory.IsGoodAddr(lsa + m_offset); }
+	bool IsGoodLSA(const u32 lsa) const { return Memory.IsGoodAddr(lsa + m_offset) && lsa < 0x40000; }
 	virtual u8   ReadLS8  (const u32 lsa) const { return Memory.Read8  (lsa + (m_offset & 0x3fffc)); }
 	virtual u16  ReadLS16 (const u32 lsa) const { return Memory.Read16 (lsa + m_offset); }
 	virtual u32  ReadLS32 (const u32 lsa) const { return Memory.Read32 (lsa + m_offset); }
@@ -478,7 +479,7 @@ public:
 	}
 
 public:
-	virtual void InitRegs(); 
+	virtual void InitRegs();
 	virtual u64 GetFreeStackSize() const;
 
 protected:
diff --git a/rpcs3/Emu/SysCalls/SC_FUNC.h b/rpcs3/Emu/SysCalls/SC_FUNC.h
index 88171718d5..3c8f6aa362 100644
--- a/rpcs3/Emu/SysCalls/SC_FUNC.h
+++ b/rpcs3/Emu/SysCalls/SC_FUNC.h
@@ -255,7 +255,7 @@ class binder_func_10<void, T1, T2, T3, T4, T5, T6, T7, T8, T9, T10> : public fun
 
 public:
 	binder_func_10(func_t call) : func_caller(), m_call(call) {}
-	virtual void operator()() { declCPU(); m_call(ARG(1), ARG(2), ARG(3), ARG(4), ARG(5), ARG(6), ARG(7), ARG(8), ARG(9), ARG(10), ARG(11)); }
+	virtual void operator()() { declCPU(); m_call(ARG(1), ARG(2), ARG(3), ARG(4), ARG(5), ARG(6), ARG(7), ARG(8), ARG(9), ARG(10)); }
 };
 
 template<typename TR, typename T1, typename T2, typename T3, typename T4, typename T5, typename T6, typename T7, typename T8, typename T9, typename T10, typename T11>