Improve emulation stopping speed

Split phases of signalling threads and joining them.

Utilities/Thread.h

@@ -612,20 +612,24 @@ public:
 		return static_cast<thread_state>(thread::m_sync.load() & 3);
 	}
 
-	// Try to abort by assigning thread_state::aborting (UB if assigning different state)
+	// Try to abort by assigning thread_state::aborting/finished
+	// Join thread by thread_state::finished
 	named_thread& operator=(thread_state s)
 	{
-		if (s == thread_state::aborting && thread::m_sync.fetch_op([](u64& v){ return !(v & 3) && (v |= 1); }).second)
+		if (s >= thread_state::aborting && thread::m_sync.fetch_op([](u64& v){ return !(v & 3) && (v |= 1); }).second)
-		{
-			if (s == thread_state::aborting)
 		{
 			thread::m_sync.notify_one(1);
-			}
 
 			if constexpr (std::is_base_of_v<need_wakeup, Context>)
 			{
 				this->wake_up();
 			}
+
+			if (s == thread_state::finished)
+			{
+				// This participates in emulation stopping, use destruction-alike semantics
+				thread::join(true);
+			}
 		}
 
 		return *this;
@@ -634,9 +638,8 @@ public:
 	// Context type doesn't need virtual destructor
 	~named_thread()
 	{
-		// Assign aborting state forcefully
+		// Assign aborting state forcefully and join thread
-		operator=(thread_state::aborting);
+		operator=(thread_state::finished);
-		thread::join(true);
 
 		if constexpr (!result::empty)
 		{

rpcs3/Emu/CPU/CPUThread.cpp

@@ -827,26 +827,6 @@ void cpu_thread::notify()
 	}
 }
 
-void cpu_thread::abort()
-{
-	state += cpu_flag::exit;
-	state.notify_one(cpu_flag::exit);
-
-	// Downcast to correct type
-	if (id_type() == 1)
-	{
-		*static_cast<named_thread<ppu_thread>*>(this) = thread_state::aborting;
-	}
-	else if (id_type() == 2)
-	{
-		*static_cast<named_thread<spu_thread>*>(this) = thread_state::aborting;
-	}
-	else
-	{
-		fmt::throw_exception("Invalid cpu_thread type");
-	}
-}
-
 std::string cpu_thread::get_name() const
 {
 	// Downcast to correct type
@@ -1131,7 +1111,8 @@ void cpu_thread::stop_all() noexcept
 	{
 		auto on_stop = [](u32, cpu_thread& cpu)
 		{
-			cpu.abort();
+			cpu.state += cpu_flag::exit;
+			cpu.state.notify_one(cpu_flag::exit);
 		};
 
 		idm::select<named_thread<ppu_thread>>(on_stop);
@@ -1139,12 +1120,12 @@ void cpu_thread::stop_all() noexcept
 	}
 
 	sys_log.notice("All CPU threads have been signaled.");
-
-	while (s_cpu_counter)
-	{
-		std::this_thread::sleep_for(1ms);
 }
 
+void cpu_thread::cleanup() noexcept
+{
+	ensure(!s_cpu_counter);
+
 	sys_log.notice("All CPU threads have been stopped. [+: %u]", +g_threads_created);
 
 	g_threads_deleted -= g_threads_created.load();

rpcs3/Emu/CPU/CPUThread.h

@@ -140,9 +140,6 @@ public:
 
 	void notify();
 
-private:
-	void abort();
-
 public:
 	// Thread stats for external observation
 	static atomic_t<u64> g_threads_created, g_threads_deleted, g_suspend_counter;
@@ -269,6 +266,9 @@ public:
 	// Stop all threads with cpu_flag::exit
 	static void stop_all() noexcept;
 
+	// Cleanup thread counting information
+	static void cleanup() noexcept;
+
 	// Send signal to the profiler(s) to flush results
 	static void flush_profilers() noexcept;
 

rpcs3/Emu/IdManager.h

@@ -12,6 +12,8 @@ extern stx::manual_typemap<void, 0x20'00000, 128> g_fixed_typemap;
 
 constexpr auto* g_fxo = &g_fixed_typemap;
 
+enum class thread_state : u32;
+
 // Helper namespace
 namespace id_manager
 {
@@ -141,7 +143,7 @@ namespace id_manager
 	template <typename T>
 	struct id_map
 	{
-		std::vector<std::pair<id_key, std::shared_ptr<void>>> vec{};
+		std::vector<std::pair<id_key, std::shared_ptr<void>>> vec{}, private_copy{};
 		shared_mutex mutex{}; // TODO: Use this instead of global mutex
 
 		id_map()
@@ -149,6 +151,29 @@ namespace id_manager
 			// Preallocate memory
 			vec.reserve(T::id_count);
 		}
+
+		template <bool dummy = false> requires (std::is_assignable_v<T&, thread_state>)
+		id_map& operator=(thread_state state)
+		{
+			if (private_copy.empty())
+			{
+				reader_lock lock(g_mutex);
+
+				// Save all entries
+				private_copy = vec;
+			}
+
+			// Signal or join threads
+			for (const auto& [key, ptr] : private_copy)
+			{
+				if (ptr)
+				{
+					*static_cast<T*>(ptr.get()) = state;
+				}
+			}
+
+			return *this;
+		}
 	};
 }
 

rpcs3/Emu/System.cpp

@@ -1543,6 +1543,35 @@ void Emulator::Stop(bool restart)
 	}
 
 	cpu_thread::stop_all();
+
+	using fxo_t = std::remove_pointer_t<decltype(g_fxo)>;
+
+	// Signal threads
+	for (const auto& type : fxo_t::view_typelist())
+	{
+		if (type.stop)
+		{
+			if (auto data = g_fxo->try_get(type))
+			{
+				type.stop(data, thread_state::aborting);
+			}
+		}
+	}
+
+	// Join threads
+	for (const auto& type : fxo_t::view_typelist())
+	{
+		if (type.stop)
+		{
+			if (auto data = g_fxo->try_get(type))
+			{
+				type.stop(data, thread_state::finished);
+			}
+		}
+	}
+
+	cpu_thread::cleanup();
+
 	g_fxo->reset();
 
 	sys_log.notice("All threads have been stopped.");

rpcs3/util/fixed_typemap.hpp

@@ -8,6 +8,8 @@
 #include <utility>
 #include <type_traits>
 
+enum class thread_state : u32;
+
 namespace stx
 {
 	// Simplified typemap with exactly one object of each used type, non-moveable. Initialized on init(). Destroyed on clear().
@@ -53,10 +55,11 @@ namespace stx
 #endif
 		}
 
-		// Save default constructor and destructor
+		// Save default constructor and destructor and optional joining operation
 		struct typeinfo
 		{
 			bool(*create)(uchar* ptr, manual_typemap&) noexcept = nullptr;
+			void(*stop)(void* ptr, thread_state) noexcept = nullptr;
 			void(*destroy)(void* ptr) noexcept = nullptr;
 			std::string_view name{};
 
@@ -86,6 +89,13 @@ namespace stx
 				std::launder(static_cast<T*>(ptr))->~T();
 			}
 
+			template <typename T>
+			static void call_stop(void* ptr, thread_state state) noexcept
+			{
+				// Abort and/or join (expected thread_state::aborting or thread_state::finished)
+				*std::launder(static_cast<T*>(ptr)) = state;
+			}
+
 			template <typename T>
 			static typeinfo make_typeinfo()
 			{
@@ -94,6 +104,12 @@ namespace stx
 				typeinfo r;
 				r.create = &call_ctor<T>;
 				r.destroy = &call_dtor<T>;
+
+				if constexpr (std::is_assignable_v<T&, thread_state>)
+				{
+					r.stop = &call_stop<T>;
+				}
+
 #ifdef _MSC_VER
 				constexpr std::string_view name = parse_type(__FUNCSIG__);
 #else
@@ -327,5 +343,21 @@ namespace stx
 
 			[[unlikely]] return nullptr;
 		}
+
+		static const auto& view_typelist() noexcept
+		{
+			return stx::typelist<typeinfo>();
+		}
+
+		// Get type-erased raw pointer to storage of type
+		uchar* try_get(const type_info<typeinfo>& type) const noexcept
+		{
+			if (m_init[type.index()])
+			{
+				[[likely]] return (Size ? +m_data : m_list) + type.pos();
+			}
+
+			[[unlikely]] return nullptr;
+		}
 	};
 }