diff --git a/rpcs3/Emu/CPU/CPUThread.cpp b/rpcs3/Emu/CPU/CPUThread.cpp
index fe9bda7676..48bd8bc6c7 100644
--- a/rpcs3/Emu/CPU/CPUThread.cpp
+++ b/rpcs3/Emu/CPU/CPUThread.cpp
@@ -466,6 +466,7 @@ bool cpu_thread::check_state() noexcept
 
 	bool cpu_sleep_called = false;
 	bool cpu_flag_memory = false;
+	bool escape, retval;
 
 	if (!(state & cpu_flag::wait))
 	{
@@ -474,54 +475,73 @@ bool cpu_thread::check_state() noexcept
 
 	while (true)
 	{
-		if (state & cpu_flag::memory)
+		// Process all flags in a single atomic op
+		const auto state0 = state.fetch_op([&](bs_t<cpu_flag>& flags)
 		{
-			if (auto& ptr = vm::g_tls_locked)
+			bool store = false;
+
+			if (flags & cpu_flag::signal)
 			{
-				ptr->compare_and_swap(this, nullptr);
-				ptr = nullptr;
+				flags -= cpu_flag::signal;
+				cpu_sleep_called = false;
+				store = true;
 			}
 
-			cpu_flag_memory = true;
-			state -= cpu_flag::memory;
-		}
-
-		if (state & (cpu_flag::exit + cpu_flag::dbg_global_stop))
-		{
-			state += cpu_flag::wait;
-			return true;
-		}
-
-		const auto [state0, escape] = state.fetch_op([&](bs_t<cpu_flag>& flags)
-		{
 			// Atomically clean wait flag and escape
 			if (!(flags & (cpu_flag::exit + cpu_flag::dbg_global_stop + cpu_flag::ret + cpu_flag::stop)))
 			{
 				// Check pause flags which hold thread inside check_state
 				if (flags & (cpu_flag::pause + cpu_flag::suspend + cpu_flag::dbg_global_pause + cpu_flag::dbg_pause))
 				{
-					return false;
+					escape = false;
+					return store;
 				}
 
-				flags -= cpu_flag::wait;
+				if (flags & cpu_flag::memory)
+				{
+					// wait flag will be cleared later (optimization)
+					cpu_flag_memory = true;
+				}
+				else
+				{
+					AUDIT(!cpu_flag_memory);
+					flags -= cpu_flag::wait;
+					store = true;
+				}
+
+				retval = false;
+			}
+			else
+			{
+				cpu_flag_memory = false;
+				retval = true;
 			}
 
-			return true;
-		});
+			if (flags & cpu_flag::dbg_step)
+			{
+				flags += cpu_flag::dbg_pause;
+				flags -= cpu_flag::dbg_step;
+				store = true;
+			}
 
-		if (state & cpu_flag::signal && state.test_and_reset(cpu_flag::signal))
-		{
-			cpu_sleep_called = false;
-		}
+			escape = true;
+			return store;
+		}).first;
 
 		if (escape)
 		{
 			if (cpu_flag_memory)
 			{
 				cpu_mem();
+
+				if (state & (cpu_flag::pause + cpu_flag::memory)) [[unlikely]]
+				{
+					state += cpu_flag::wait;
+					continue;
+				}
 			}
 
-			break;
+			return retval;
 		}
 		else if (!cpu_sleep_called && state0 & cpu_flag::suspend)
 		{
@@ -540,21 +560,6 @@ bool cpu_thread::check_state() noexcept
 			g_fxo->get<cpu_counter>()->cpu_suspend_lock.lock_unlock();
 		}
 	}
-
-	const auto state_ = state.load();
-
-	if (state_ & (cpu_flag::ret + cpu_flag::stop))
-	{
-		return true;
-	}
-
-	if (state_ & cpu_flag::dbg_step)
-	{
-		state += cpu_flag::dbg_pause;
-		state -= cpu_flag::dbg_step;
-	}
-
-	return false;
 }
 
 void cpu_thread::notify()
diff --git a/rpcs3/Emu/Memory/vm.cpp b/rpcs3/Emu/Memory/vm.cpp
index 118988fc71..24f4c0651f 100644
--- a/rpcs3/Emu/Memory/vm.cpp
+++ b/rpcs3/Emu/Memory/vm.cpp
@@ -129,9 +129,30 @@ namespace vm
 	{
 		if (g_tls_locked && *g_tls_locked == &cpu) [[unlikely]]
 		{
+			if (cpu.state & cpu_flag::wait)
+			{
+				while (true)
+				{
+					g_mutex.lock_unlock();
+					cpu.state -= cpu_flag::wait + cpu_flag::memory;
+
+					if (g_mutex.is_lockable()) [[likely]]
+					{
+						return;
+					}
+
+					cpu.state += cpu_flag::wait;
+				}
+			}
+
 			return;
 		}
 
+		if (cpu.state & cpu_flag::memory)
+		{
+			cpu.state -= cpu_flag::memory;
+		}
+
 		if (g_mutex.is_lockable()) [[likely]]
 		{
 			// Optimistic path (hope that mutex is not exclusively locked)
@@ -355,7 +376,8 @@ namespace vm
 
 			for (auto lock = g_locks.cbegin(), end = lock + g_cfg.core.ppu_threads; lock != end; lock++)
 			{
-				while (*lock)
+				cpu_thread* ptr;
+				while ((ptr = *lock) && !(ptr->state & cpu_flag::wait))
 				{
 					_mm_pause();
 				}