diff --git a/rpcs3/Emu/ARMv7/ARMv7Context.h b/rpcs3/Emu/ARMv7/ARMv7Context.h
index 876e55fb36..2ab58ec3d4 100644
--- a/rpcs3/Emu/ARMv7/ARMv7Context.h
+++ b/rpcs3/Emu/ARMv7/ARMv7Context.h
@@ -1,7 +1,5 @@
 #pragma once
 
-class ARMv7Thread;
-
 enum ARMv7InstructionSet
 {
 	ARM,
@@ -122,13 +120,9 @@ struct ARMv7Context
 
 	std::array<perf_counter, 6> counters;
 
-	ARMv7Thread& thread;
-
-	u32 debug; // debug flags
+	u32 debug;
 	std::string debug_str;
 
-	ARMv7Context(ARMv7Thread& thread) : thread(thread), debug(/*DF_DISASM | DF_PRINT*/ 0) {}
-
 	void write_pc(u32 value);
 	u32 read_pc();
 	u32 get_stack_arg(u32 pos);
diff --git a/rpcs3/Emu/ARMv7/ARMv7Interpreter.cpp b/rpcs3/Emu/ARMv7/ARMv7Interpreter.cpp
index a064fc768f..1c51b2de71 100644
--- a/rpcs3/Emu/ARMv7/ARMv7Interpreter.cpp
+++ b/rpcs3/Emu/ARMv7/ARMv7Interpreter.cpp
@@ -287,17 +287,19 @@ namespace ARMv7_instrs
 	{
 		if (context.debug & DF_PRINT)
 		{
+			ARMv7Thread& CPU = static_cast<ARMv7Thread&>(context);
+
 			auto pos = context.debug_str.find(' ');
 			if (pos != std::string::npos && pos < 8)
 			{
 				context.debug_str.insert(pos, 8 - pos, ' ');
 			}
 
-			context.fmt_debug_str("0x%08x: %s", context.thread.PC, context.debug_str);
+			context.fmt_debug_str("0x%08x: %s", CPU.PC, context.debug_str);
 
 			LV2_LOCK;
 
-			auto found = g_armv7_dump.find(context.thread.PC);
+			auto found = g_armv7_dump.find(CPU.PC);
 			if (found != g_armv7_dump.end())
 			{
 				if (found->second != context.debug_str)
@@ -307,7 +309,7 @@ namespace ARMv7_instrs
 			}
 			else
 			{
-				g_armv7_dump[context.thread.PC] = context.debug_str;
+				g_armv7_dump[CPU.PC] = context.debug_str;
 			}
 		}
 
@@ -1258,7 +1260,7 @@ void ARMv7_instrs::B(ARMv7Context& context, const ARMv7Code code, const ARMv7_en
 
 	if (ConditionPassed(context, cond))
 	{
-		context.thread.SetBranch(context.read_pc() + imm32);
+		static_cast<ARMv7Thread&>(context).SetBranch(context.read_pc() + imm32);
 	}
 }
 
@@ -1438,12 +1440,14 @@ void ARMv7_instrs::BL(ARMv7Context& context, const ARMv7Code code, const ARMv7_e
 	if (ConditionPassed(context, cond))
 	{
 		context.LR = lr;
-		context.thread.SetBranch(pc);
+		static_cast<ARMv7Thread&>(context).SetBranch(pc);
 	}
 }
 
 void ARMv7_instrs::BLX(ARMv7Context& context, const ARMv7Code code, const ARMv7_encoding type)
 {
+	ARMv7Thread& thread = static_cast<ARMv7Thread&>(context);
+
 	u32 cond, target, newLR;
 
 	switch (type)
@@ -1451,7 +1455,7 @@ void ARMv7_instrs::BLX(ARMv7Context& context, const ARMv7Code code, const ARMv7_
 	case T1:
 	{
 		cond = context.ITSTATE.advance();
-		newLR = (context.thread.PC + 2) | 1;
+		newLR = (thread.PC + 2) | 1;
 		{
 			const u32 m = (code.data >> 3) & 0xf;
 			reject(m == 15, "UNPREDICTABLE");
@@ -1464,12 +1468,12 @@ void ARMv7_instrs::BLX(ARMv7Context& context, const ARMv7Code code, const ARMv7_
 	case T2:
 	{
 		cond = context.ITSTATE.advance();
-		newLR = (context.thread.PC + 4) | 1;
+		newLR = (thread.PC + 4) | 1;
 		{
 			const u32 s = (code.data >> 26) & 0x1;
 			const u32 i1 = (code.data >> 13) & 0x1 ^ s ^ 1;
 			const u32 i2 = (code.data >> 11) & 0x1 ^ s ^ 1;
-			target = ~3 & context.thread.PC + 4 + sign<25, u32>(s << 24 | i2 << 23 | i1 << 22 | (code.data & 0x3ff0000) >> 4 | (code.data & 0x7ff) << 1);
+			target = ~3 & thread.PC + 4 + sign<25, u32>(s << 24 | i2 << 23 | i1 << 22 | (code.data & 0x3ff0000) >> 4 | (code.data & 0x7ff) << 1);
 		}
 
 		reject(context.ITSTATE, "UNPREDICTABLE");
@@ -1478,15 +1482,15 @@ void ARMv7_instrs::BLX(ARMv7Context& context, const ARMv7Code code, const ARMv7_
 	case A1:
 	{
 		cond = code.data >> 28;
-		newLR = context.thread.PC + 4;
+		newLR = thread.PC + 4;
 		target = context.read_gpr(code.data & 0xf);
 		break;
 	}
 	case A2:
 	{
 		cond = 0xe; // always true
-		newLR = context.thread.PC + 4;
-		target = 1 | context.thread.PC + 8 + sign<25, u32>((code.data & 0xffffff) << 2 | (code.data & 0x1000000) >> 23);
+		newLR = thread.PC + 4;
+		target = 1 | thread.PC + 8 + sign<25, u32>((code.data & 0xffffff) << 2 | (code.data & 0x1000000) >> 23);
 		break;
 	}
 	default: throw __FUNCTION__;
@@ -1577,7 +1581,7 @@ void ARMv7_instrs::CB_Z(ARMv7Context& context, const ARMv7Code code, const ARMv7
 
 	if ((context.read_gpr(n) == 0) ^ nonzero)
 	{
-		context.thread.SetBranch(context.read_pc() + imm32);
+		static_cast<ARMv7Thread&>(context).SetBranch(context.read_pc() + imm32);
 	}
 }
 
diff --git a/rpcs3/Emu/ARMv7/ARMv7Thread.cpp b/rpcs3/Emu/ARMv7/ARMv7Thread.cpp
index 90a56e2187..eed779a987 100644
--- a/rpcs3/Emu/ARMv7/ARMv7Thread.cpp
+++ b/rpcs3/Emu/ARMv7/ARMv7Thread.cpp
@@ -12,12 +12,16 @@
 
 void ARMv7Context::write_pc(u32 value)
 {
+	ARMv7Thread& thread = *static_cast<ARMv7Thread*>(this);
+
 	ISET = value & 1 ? Thumb : ARM;
 	thread.SetBranch(value & ~1);
 }
 
 u32 ARMv7Context::read_pc()
 {
+	ARMv7Thread& thread = *static_cast<ARMv7Thread*>(this);
+
 	return ISET == ARM ? thread.PC + 8 : thread.PC + 4;
 }
 
@@ -28,7 +32,7 @@ u32 ARMv7Context::get_stack_arg(u32 pos)
 
 void ARMv7Context::fast_call(u32 addr)
 {
-	return thread.FastCall(addr);
+	return static_cast<ARMv7Thread*>(this)->FastCall(addr);
 }
 
 #define TLS_MAX 128
@@ -96,7 +100,6 @@ void armv7_free_tls(u32 thread)
 
 ARMv7Thread::ARMv7Thread()
 	: CPUThread(CPU_THREAD_ARMv7)
-	, context(*this)
 	//, m_arg(0)
 	//, m_last_instr_size(0)
 	//, m_last_instr_name("UNK")
@@ -110,15 +113,15 @@ ARMv7Thread::~ARMv7Thread()
 
 void ARMv7Thread::InitRegs()
 {
-	memset(context.GPR, 0, sizeof(context.GPR));
-	context.APSR.APSR = 0;
-	context.IPSR.IPSR = 0;
-	context.ISET = PC & 1 ? Thumb : ARM; // select instruction set
-	context.thread.SetPc(PC & ~1); // and fix PC
-	context.ITSTATE.IT = 0;
-	context.SP = m_stack_addr + m_stack_size;
-	context.TLS = armv7_get_tls(GetId());
-	context.debug |= DF_DISASM | DF_PRINT;
+	memset(GPR, 0, sizeof(GPR));
+	APSR.APSR = 0;
+	IPSR.IPSR = 0;
+	ISET = PC & 1 ? Thumb : ARM; // select instruction set
+	SetPc(PC & ~1); // and fix PC
+	ITSTATE.IT = 0;
+	SP = m_stack_addr + m_stack_size;
+	TLS = armv7_get_tls(GetId());
+	debug = DF_DISASM | DF_PRINT;
 }
 
 void ARMv7Thread::InitStack()
@@ -144,16 +147,16 @@ std::string ARMv7Thread::RegsToString()
 	std::string result = "Registers:\n=========\n";
 	for(int i=0; i<15; ++i)
 	{
-		result += fmt::Format("%s\t= 0x%08x\n", g_arm_reg_name[i], context.GPR[i]);
+		result += fmt::Format("%s\t= 0x%08x\n", g_arm_reg_name[i], GPR[i]);
 	}
 
 	result += fmt::Format("APSR\t= 0x%08x [N: %d, Z: %d, C: %d, V: %d, Q: %d]\n", 
-		context.APSR.APSR,
-		fmt::by_value(context.APSR.N),
-		fmt::by_value(context.APSR.Z),
-		fmt::by_value(context.APSR.C),
-		fmt::by_value(context.APSR.V),
-		fmt::by_value(context.APSR.Q));
+		APSR.APSR,
+		fmt::by_value(APSR.N),
+		fmt::by_value(APSR.Z),
+		fmt::by_value(APSR.C),
+		fmt::by_value(APSR.V),
+		fmt::by_value(APSR.Q));
 	
 	return result;
 }
@@ -180,7 +183,7 @@ void ARMv7Thread::DoRun()
 	{
 	case 0:
 	case 1:
-		m_dec = new ARMv7Decoder(context);
+		m_dec = new ARMv7Decoder(*this);
 		break;
 	default:
 		LOG_ERROR(PPU, "Invalid CPU decoder mode: %d", Ini.CPUDecoderMode.GetValue());
@@ -208,21 +211,21 @@ void ARMv7Thread::FastCall(u32 addr)
 {
 	auto old_status = m_status;
 	auto old_PC = PC;
-	auto old_stack = context.SP;
-	auto old_LR = context.LR;
+	auto old_stack = SP;
+	auto old_LR = LR;
 	auto old_thread = GetCurrentNamedThread();
 
 	m_status = Running;
 	PC = addr;
-	context.LR = Emu.GetCPUThreadStop();
+	LR = Emu.GetCPUThreadStop();
 	SetCurrentNamedThread(this);
 
 	CPUThread::Task();
 
 	m_status = old_status;
 	PC = old_PC;
-	context.SP = old_stack;
-	context.LR = old_LR;
+	SP = old_stack;
+	LR = old_LR;
 	SetCurrentNamedThread(old_thread);
 }
 
@@ -284,8 +287,8 @@ cpu_thread& armv7_thread::run()
 	armv7.Run();
 
 	// set arguments
-	armv7.context.GPR[0] = argc;
-	armv7.context.GPR[1] = argv;
+	armv7.GPR[0] = argc;
+	armv7.GPR[1] = argv;
 
 	return *this;
 }
diff --git a/rpcs3/Emu/ARMv7/ARMv7Thread.h b/rpcs3/Emu/ARMv7/ARMv7Thread.h
index 5bbc4a564c..05e8e8bb96 100644
--- a/rpcs3/Emu/ARMv7/ARMv7Thread.h
+++ b/rpcs3/Emu/ARMv7/ARMv7Thread.h
@@ -2,11 +2,9 @@
 #include "Emu/CPU/CPUThread.h"
 #include "ARMv7Context.h"
 
-class ARMv7Thread : public CPUThread
+class ARMv7Thread : public CPUThread, public ARMv7Context
 {
 public:
-	ARMv7Context context;
-
 	ARMv7Thread();
 	~ARMv7Thread();
 
diff --git a/rpcs3/Emu/ARMv7/Modules/sceLibKernel.cpp b/rpcs3/Emu/ARMv7/Modules/sceLibKernel.cpp
index 3fccfec929..602b60ca0b 100644
--- a/rpcs3/Emu/ARMv7/Modules/sceLibKernel.cpp
+++ b/rpcs3/Emu/ARMv7/Modules/sceLibKernel.cpp
@@ -79,12 +79,12 @@ s32 sceKernelStartThread(s32 threadId, u32 argSize, vm::ptr<const void> pArgBloc
 	ARMv7Thread& thread = static_cast<ARMv7Thread&>(*t);
 
 	// push arg block onto the stack
-	const u32 pos = (thread.context.SP -= argSize);
+	const u32 pos = (thread.SP -= argSize);
 	memcpy(vm::get_ptr<void>(pos), pArgBlock.get_ptr(), argSize);
 
 	// set SceKernelThreadEntry function arguments
-	thread.context.GPR[0] = argSize;
-	thread.context.GPR[1] = pos;
+	thread.GPR[0] = argSize;
+	thread.GPR[1] = pos;
 
 	thread.Exec();
 	return SCE_OK;
@@ -95,7 +95,7 @@ s32 sceKernelExitThread(ARMv7Context& context, s32 exitStatus)
 	sceLibKernel.Warning("sceKernelExitThread(exitStatus=0x%x)", exitStatus);
 
 	// exit status is stored in r0
-	context.thread.Stop();
+	static_cast<ARMv7Thread&>(context).Stop();
 
 	return SCE_OK;
 }
@@ -127,10 +127,10 @@ s32 sceKernelExitDeleteThread(ARMv7Context& context, s32 exitStatus)
 	sceLibKernel.Warning("sceKernelExitDeleteThread(exitStatus=0x%x)", exitStatus);
 
 	// exit status is stored in r0
-	context.thread.Stop();
+	static_cast<ARMv7Thread&>(context).Stop();
 
 	// current thread should be deleted
-	const u32 id = context.thread.GetId();
+	const u32 id = static_cast<ARMv7Thread&>(context).GetId();
 	CallAfter([id]()
 	{
 		Emu.GetCPU().RemoveThread(id);
@@ -171,7 +171,7 @@ u32 sceKernelGetThreadId(ARMv7Context& context)
 {
 	sceLibKernel.Log("sceKernelGetThreadId()");
 
-	return context.thread.GetId();
+	return static_cast<ARMv7Thread&>(context).GetId();
 }
 
 s32 sceKernelChangeCurrentThreadAttr(u32 clearAttr, u32 setAttr)
@@ -287,7 +287,7 @@ s32 sceKernelWaitThreadEnd(s32 threadId, vm::ptr<s32> pExitStatus, vm::ptr<u32>
 
 	if (pExitStatus)
 	{
-		*pExitStatus = thread.context.GPR[0];
+		*pExitStatus = thread.GPR[0];
 	}
 
 	return SCE_OK;
diff --git a/rpcs3/Emu/ARMv7/Modules/sceLibc.cpp b/rpcs3/Emu/ARMv7/Modules/sceLibc.cpp
index a29f44dccf..897ce4eec8 100644
--- a/rpcs3/Emu/ARMv7/Modules/sceLibc.cpp
+++ b/rpcs3/Emu/ARMv7/Modules/sceLibc.cpp
@@ -243,12 +243,11 @@ namespace sce_libc_func
 		::memset(dst.get_ptr(), value, size);
 	}
 
-	void _Assert(ARMv7Context& context, vm::cptr<char> text, vm::cptr<char> func)
+	void _Assert(vm::cptr<char> text, vm::cptr<char> func)
 	{
 		sceLibc.Error("_Assert(text=*0x%x, func=*0x%x)", text, func);
 
 		LOG_ERROR(TTY, "%s : %s\n", func.get_ptr(), text.get_ptr());
-		LOG_NOTICE(ARMv7, context.thread.RegsToString());
 		Emu.Pause();
 	}
 }
diff --git a/rpcs3/Emu/ARMv7/PSVFuncList.cpp b/rpcs3/Emu/ARMv7/PSVFuncList.cpp
index de2b1d9c62..cf2ff21730 100644
--- a/rpcs3/Emu/ARMv7/PSVFuncList.cpp
+++ b/rpcs3/Emu/ARMv7/PSVFuncList.cpp
@@ -70,8 +70,10 @@ void execute_psv_func_by_index(ARMv7Context& context, u32 index)
 {
 	if (auto func = get_psv_func_by_index(index))
 	{
-		auto old_last_syscall = context.thread.m_last_syscall;
-		context.thread.m_last_syscall = func->nid;
+		ARMv7Thread& CPU = static_cast<ARMv7Thread&>(context);
+
+		auto old_last_syscall = CPU.m_last_syscall;
+		CPU.m_last_syscall = func->nid;
 
 		if (func->func)
 		{
@@ -88,7 +90,7 @@ void execute_psv_func_by_index(ARMv7Context& context, u32 index)
 			func->module->on_error(context.GPR[0], func);
 		}
 
-		context.thread.m_last_syscall = old_last_syscall;
+		CPU.m_last_syscall = old_last_syscall;
 	}
 	else
 	{
@@ -231,7 +233,7 @@ void initialize_psv_modules()
 	hle_return.name = "HLE_RETURN";
 	hle_return.func = [](ARMv7Context& context)
 	{
-		context.thread.FastStop();
+		static_cast<ARMv7Thread&>(context).FastStop();
 	};
 
 	// load functions
diff --git a/rpcs3/Emu/Cell/SPUContext.h b/rpcs3/Emu/Cell/SPUContext.h
index 93a2464038..2088497285 100644
--- a/rpcs3/Emu/Cell/SPUContext.h
+++ b/rpcs3/Emu/Cell/SPUContext.h
@@ -1,10 +1,6 @@
 #pragma once
 
-class SPUThread;
-
 struct SPUContext
 {
 	u128 gpr[128];
-
-	SPUThread& thread;
 };
diff --git a/rpcs3/Emu/Memory/vm.cpp b/rpcs3/Emu/Memory/vm.cpp
index ef10b43fe6..769e63dacd 100644
--- a/rpcs3/Emu/Memory/vm.cpp
+++ b/rpcs3/Emu/Memory/vm.cpp
@@ -659,7 +659,7 @@ namespace vm
 
 		case CPU_THREAD_ARMv7:
 		{
-			ARMv7Context& context = static_cast<ARMv7Thread&>(CPU).context;
+			ARMv7Context& context = static_cast<ARMv7Thread&>(CPU);
 
 			old_pos = context.SP;
 			context.SP -= align(size, 4); // room minimal possible size
@@ -711,7 +711,7 @@ namespace vm
 
 		case CPU_THREAD_ARMv7:
 		{
-			ARMv7Context& context = static_cast<ARMv7Thread&>(CPU).context;
+			ARMv7Context& context = static_cast<ARMv7Thread&>(CPU);
 
 			if (context.SP != addr && !Emu.IsStopped())
 			{