diff --git a/rpcs3/Emu/ARMv7/ARMv7Thread.cpp b/rpcs3/Emu/ARMv7/ARMv7Thread.cpp
index 8c57c78485..7f98249d9d 100644
--- a/rpcs3/Emu/ARMv7/ARMv7Thread.cpp
+++ b/rpcs3/Emu/ARMv7/ARMv7Thread.cpp
@@ -102,6 +102,33 @@ void ARMv7Thread::DoCode()
 {
 }
 
+void ARMv7Thread::FastCall(u32 addr)
+{
+	auto old_status = m_status;
+	auto old_PC = PC;
+	auto old_stack = SP;
+	auto old_LR = LR;
+	auto old_thread = GetCurrentNamedThread();
+
+	m_status = Running;
+	PC = addr;
+	LR = Emu.m_cpu_thr_stop;
+	SetCurrentNamedThread(this);
+
+	CPUThread::Task();
+
+	m_status = old_status;
+	PC = old_PC;
+	SP = old_stack;
+	LR = old_LR;
+	SetCurrentNamedThread(old_thread);
+}
+
+void ARMv7Thread::FastStop()
+{
+	m_status = Stopped;
+}
+
 arm7_thread::arm7_thread(u32 entry, const std::string& name, u32 stack_size, u32 prio)
 {
 	thread = &Emu.GetCPU().AddThread(CPU_THREAD_ARMv7);
@@ -112,4 +139,4 @@ arm7_thread::arm7_thread(u32 entry, const std::string& name, u32 stack_size, u32
 	thread->SetPrio(prio ? prio : Emu.GetInfo().GetProcParam().primary_prio);
 
 	argc = 0;
-}
\ No newline at end of file
+}
diff --git a/rpcs3/Emu/ARMv7/ARMv7Thread.h b/rpcs3/Emu/ARMv7/ARMv7Thread.h
index 1eeefff09c..8737454857 100644
--- a/rpcs3/Emu/ARMv7/ARMv7Thread.h
+++ b/rpcs3/Emu/ARMv7/ARMv7Thread.h
@@ -123,6 +123,8 @@ public:
 	virtual void InitRegs(); 
 	virtual void InitStack();
 	u32 GetStackArg(u32 pos);
+	void FastCall(u32 addr);
+	void FastStop();
 	virtual void DoRun();
 
 public:
diff --git a/rpcs3/Emu/ARMv7/Modules/sceLibKernel.cpp b/rpcs3/Emu/ARMv7/Modules/sceLibKernel.cpp
index 18a29f85eb..151d12ae21 100644
--- a/rpcs3/Emu/ARMv7/Modules/sceLibKernel.cpp
+++ b/rpcs3/Emu/ARMv7/Modules/sceLibKernel.cpp
@@ -140,9 +140,12 @@ s32 sceKernelStartThread(s32 threadId, u32 argSize, vm::psv::ptr<const void> pAr
 	return SCE_OK;
 }
 
-s32 sceKernelExitThread(s32 exitStatus)
+s32 sceKernelExitThread(ARMv7Thread& CPU, s32 exitStatus)
 {
-	sceLibKernel.Todo("sceKernelExitThread(exitStatus=0x%x)", exitStatus);
+	sceLibKernel.Error("sceKernelExitThread(exitStatus=0x%x)", exitStatus);
+
+	// exit status is stored in r0
+	CPU.Stop();
 
 	return SCE_OK;
 }
diff --git a/rpcs3/Emu/ARMv7/PSVFuncList.cpp b/rpcs3/Emu/ARMv7/PSVFuncList.cpp
index bde15b0dc9..5f644a6d10 100644
--- a/rpcs3/Emu/ARMv7/PSVFuncList.cpp
+++ b/rpcs3/Emu/ARMv7/PSVFuncList.cpp
@@ -6,20 +6,32 @@ std::vector<psv_func> g_psv_func_list = []() -> std::vector<psv_func>
 {
 	std::vector<psv_func> v;
 
-	psv_func f =
+	psv_func unimplemented =
 	{
-		0xdeadbeef,
+		0x00000000, // must not be a valid id
 		"INVALID FUNCTION",
 		new psv_func_detail::func_binder<u32>([]() -> u32
 		{
-			LOG_ERROR(HLE, "Unimplemented function found");
+			LOG_ERROR(HLE, "Unimplemented function executed");
 			Emu.Pause();
 
 			return 0xffffffffu;
 		}),
 		nullptr,
 	};
-	v.push_back(f);
+	v.push_back(unimplemented);
+
+	psv_func hle_return =
+	{
+		0x00000001, // must not be a valid id
+		"INVALID FUNCTION",
+		new psv_func_detail::func_binder<void, ARMv7Thread&>([](ARMv7Thread& CPU)
+		{
+			CPU.FastStop();
+		}),
+		nullptr,
+	};
+	v.push_back(hle_return);
 
 	return v;
 }();
diff --git a/rpcs3/Emu/Cell/PPUInterpreter.h b/rpcs3/Emu/Cell/PPUInterpreter.h
index c24509ea0f..85f86ef76c 100644
--- a/rpcs3/Emu/Cell/PPUInterpreter.h
+++ b/rpcs3/Emu/Cell/PPUInterpreter.h
@@ -1840,15 +1840,15 @@ private:
 	{
 		for (uint b = 0; b < 16; b++)
 		{
-				s16 result = (s16)CPU.VPR[va]._u8[b] - (s16)CPU.VPR[vb]._u8[b];
+			s16 result = (s16)CPU.VPR[va]._u8[b] - (s16)CPU.VPR[vb]._u8[b];
 
-				if (result < 0)
-				{
+			if (result < 0)
+			{
 				CPU.VPR[vd]._u8[b] = 0;
 				CPU.VSCR.SAT = 1;
-				}
-				else
-					CPU.VPR[vd]._u8[b] = (u8)result;
+			}
+			else
+				CPU.VPR[vd]._u8[b] = (u8)result;
 		}
 	}
 	void VSUBUHM(u32 vd, u32 va, u32 vb)
diff --git a/rpcs3/Emu/Cell/PPUThread.cpp b/rpcs3/Emu/Cell/PPUThread.cpp
index 06b2c8b926..c9703c10ce 100644
--- a/rpcs3/Emu/Cell/PPUThread.cpp
+++ b/rpcs3/Emu/Cell/PPUThread.cpp
@@ -202,7 +202,7 @@ u64 PPUThread::FastCall2(u32 addr, u32 rtoc)
 	m_status = Running;
 	PC = addr;
 	GPR[2] = rtoc;
-	LR = Emu.m_ppu_thr_stop;
+	LR = Emu.m_cpu_thr_stop;
 	SetCurrentNamedThread(this);
 
 	CPUThread::Task();
@@ -288,4 +288,4 @@ ppu_thread& ppu_thread::gpr(uint index, u64 value)
 	static_cast<PPUThread*>(thread)->GPR[index] = value;
 
 	return *this;
-}
\ No newline at end of file
+}
diff --git a/rpcs3/Emu/Cell/SPURecompiler.h b/rpcs3/Emu/Cell/SPURecompiler.h
index 437ebbcd97..1b977a6e52 100644
--- a/rpcs3/Emu/Cell/SPURecompiler.h
+++ b/rpcs3/Emu/Cell/SPURecompiler.h
@@ -1715,13 +1715,10 @@ private:
 		}
 		else
 		{
+			// rotate left
 			const XmmLink& va = XmmGet(ra, rt);
-			const XmmLink& v1 = XmmCopy(va);
-			c.pslldq(va.get(), s);
-			c.psrldq(v1.get(), 16 - s);
-			c.por(va.get(), v1.get());
+			c.palignr(va.get(), va.get(), 16 - s);
 			XmmFinalize(va, rt);
-			XmmFinalize(v1);
 		}
 		LOG_OPCODE();
 	}
diff --git a/rpcs3/Emu/SysCalls/Modules/cellGame.cpp b/rpcs3/Emu/SysCalls/Modules/cellGame.cpp
index e480edc50f..3a11eedc17 100644
--- a/rpcs3/Emu/SysCalls/Modules/cellGame.cpp
+++ b/rpcs3/Emu/SysCalls/Modules/cellGame.cpp
@@ -181,9 +181,8 @@ int cellGameContentPermit(vm::ptr<char[CELL_GAME_PATH_MAX]> contentInfoPath, vm:
 	cellGame->Warning("cellGameContentPermit(contentInfoPath_addr=0x%x, usrdirPath_addr=0x%x)",
 		contentInfoPath.addr(), usrdirPath.addr());
 
-	if (!contentInfoPath && !usrdirPath)
+	if (!contentInfoPath || !usrdirPath)
 	{
-		cellGame->Warning("cellGameContentPermit(): CELL_GAME_ERROR_PARAM");
 		return CELL_GAME_ERROR_PARAM;
 	}
 	
diff --git a/rpcs3/Loader/ELF32.cpp b/rpcs3/Loader/ELF32.cpp
index 3e68915494..fb83221f96 100644
--- a/rpcs3/Loader/ELF32.cpp
+++ b/rpcs3/Loader/ELF32.cpp
@@ -86,16 +86,7 @@ namespace loader
 			{
 			case MACHINE_MIPS: break;
 			case MACHINE_ARM: arm7_thread(m_ehdr.is_le() ? m_ehdr.data_le.e_entry : m_ehdr.data_be.e_entry, "main_thread").args({ Emu.GetPath()/*, "-emu"*/ }).run(); break;
-			case MACHINE_SPU:
-			{
-				if (!Emu.IsStopped())
-				{
-					break;
-				}
-
-				spu_thread(m_ehdr.is_le() ? m_ehdr.data_le.e_entry : m_ehdr.data_be.e_entry, "main_thread").args({ Emu.GetPath()/*, "-emu"*/ }).run();
-				break;
-			}
+			case MACHINE_SPU: spu_thread(m_ehdr.is_le() ? m_ehdr.data_le.e_entry : m_ehdr.data_be.e_entry, "main_thread").args({ Emu.GetPath()/*, "-emu"*/ }).run(); break;
 			}
 
 			return ok;
@@ -117,7 +108,7 @@ namespace loader
 				case 0x00000001: //LOAD
 					if (phdr.data_le.p_memsz)
 					{
-						if (machine != MACHINE_SPU && !vm::alloc(vaddr, memsz, vm::main))
+						if ((machine != MACHINE_SPU || Emu.IsStopped()) && !vm::alloc(vaddr, memsz, vm::main))
 						{
 							LOG_ERROR(LOADER, "%s(): AllocFixed(0x%llx, 0x%x) failed", __FUNCTION__, vaddr, memsz);