Savestates/SPU: Kill emulation when its safe to save SPU state

rpcs3/Emu/CPU/CPUThread.cpp

@@ -1196,3 +1196,72 @@ u32 CPUDisAsm::DisAsmBranchTarget(s32 /*imm*/)
 	// Unused
 	return 0;
 }
+
+extern bool try_lock_spu_threads_in_a_state_compatible_with_savestates()
+{
+	const u64 start = get_system_time();
+
+	// Attempt to lock for half a second, if somehow takes longer abort it
+	do
+	{
+		if (cpu_thread::suspend_all(nullptr, {}, []()
+		{
+			return idm::select<named_thread<spu_thread>>([](u32, spu_thread& spu)
+			{
+				if (!spu.unsavable)
+				{
+					if (Emu.IsPaused())
+					{
+						// If emulation is paused, we can only hope it's already in a state compatible with savestates
+						return !!(spu.state & (cpu_flag::dbg_global_pause + cpu_flag::dbg_pause));
+					}
+					else
+					{
+						ensure(!spu.state.test_and_set(cpu_flag::dbg_global_pause));
+					}
+
+					return false;
+				}
+
+				return true;
+			}).ret;
+		}))
+		{
+			if (Emu.IsPaused())
+			{
+				return false;
+			}
+
+			// It's faster to lock once
+			reader_lock lock(id_manager::g_mutex);
+
+			idm::select<named_thread<spu_thread>>([](u32, spu_thread& spu)
+			{
+				spu.state -= cpu_flag::dbg_global_pause;
+			}, idm::unlocked);
+
+			// For faster signalling, first remove state flags then batch notifications
+			idm::select<named_thread<spu_thread>>([](u32, spu_thread& spu)
+			{
+				if (spu.state & cpu_flag::wait)
+				{
+					spu.state.notify_one();
+				}
+			}, idm::unlocked);
+
+			continue;
+		}
+
+		return true;
+	} while (std::this_thread::yield(), get_system_time() - start <= 500'000);
+
+	idm::select<named_thread<spu_thread>>([&](u32, named_thread<spu_thread>& spu)
+	{
+		if (spu.state.test_and_reset(cpu_flag::dbg_global_pause))
+		{
+			spu.state.notify_one();
+		}
+	});
+
+	return false;
+}

rpcs3/Emu/Cell/PPUThread.cpp

@@ -24,6 +24,7 @@
 #include "lv2/sys_prx.h"
 #include "lv2/sys_overlay.h"
 #include "lv2/sys_process.h"
+#include "lv2/sys_spu.h"
 
 #ifdef LLVM_AVAILABLE
 #ifdef _MSC_VER
@@ -1400,6 +1401,21 @@ void ppu_thread::cpu_task()
 			// (the farther it's postponed, the less accuracy of guest time has been lost)
 			Emu.FixGuestTime();
 
+			// Run SPUs waiting on a syscall (savestates related)
+			idm::select<named_thread<spu_thread>>([&](u32, named_thread<spu_thread>& spu)
+			{
+				if (spu.group && spu.index == spu.group->waiter_spu_index)
+				{
+					if (std::exchange(spu.stop_flag_removal_protection, false))
+					{
+						return;
+					}
+
+					ensure(spu.state.test_and_reset(cpu_flag::stop));
+					spu.state.notify_one(cpu_flag::stop);
+				}
+			});
+
 			// Check if this is the only PPU left to initialize (savestates related)
 			if (lv2_obj::is_scheduler_ready())
 			{

rpcs3/Emu/Cell/SPURecompiler.cpp

@@ -4310,7 +4310,7 @@ class spu_llvm_recompiler : public spu_recompiler_base, public cpu_translator
 	}
 
 	// Call cpu_thread::check_state if necessary and return or continue (full check)
-	void check_state(u32 addr)
+	void check_state(u32 addr, bool may_be_unsafe_for_savestate = true)
 	{
 		const auto pstate = spu_ptr<u32>(&spu_thread::state);
 		const auto _body = llvm::BasicBlock::Create(m_context, "", m_function);
@@ -4318,7 +4318,24 @@ class spu_llvm_recompiler : public spu_recompiler_base, public cpu_translator
 		m_ir->CreateCondBr(m_ir->CreateICmpEQ(m_ir->CreateLoad(pstate, true), m_ir->getInt32(0)), _body, check, m_md_likely);
 		m_ir->SetInsertPoint(check);
 		update_pc(addr);
+
+		if (may_be_unsafe_for_savestate && std::none_of(std::begin(m_block->phi), std::end(m_block->phi), FN(!!x)))
+		{
+			may_be_unsafe_for_savestate = false;
+		}
+
+		if (may_be_unsafe_for_savestate)
+		{
+			m_ir->CreateStore(m_ir->getInt8(1), spu_ptr<u8>(&spu_thread::unsavable), true);
+		}
+
 		m_ir->CreateStore(m_ir->getFalse(), m_fake_global1, true);
+
+		if (may_be_unsafe_for_savestate)
+		{
+			m_ir->CreateStore(m_ir->getInt8(0), spu_ptr<u8>(&spu_thread::unsavable), true);
+		}
+
 		m_ir->CreateBr(_body);
 		m_ir->SetInsertPoint(_body);
 	}

rpcs3/Emu/Cell/SPUThread.cpp

@@ -1468,6 +1468,8 @@ void spu_thread::cpu_task()
 			spu_runtime::g_gateway(*this, _ptr<u8>(0), nullptr);
 		}
 
+		unsavable = false;
+
 		// Print some stats
 		spu_log.notice("Stats: Block Weight: %u (Retreats: %u);", block_counter, block_failure);
 	}

rpcs3/Emu/Cell/SPUThread.h

@@ -847,6 +847,7 @@ public:
 
 	const char* current_func{}; // Current STOP or RDCH blocking function
 	u64 start_time{}; // Starting time of STOP or RDCH bloking function
+	bool unsavable = false; // Flag indicating whether saving the spu thread state is currently unsafe
 
 	atomic_t<u8> debugger_float_mode = 0;
 

rpcs3/Emu/RSX/RSXThread.cpp

@@ -47,7 +47,8 @@ bool serialize<rsx::rsx_state>(utils::serial& ar, rsx::rsx_state& o)
 {
 	ar(o.transform_program);
 
-	if (GET_SERIALIZATION_VERSION(global_version))
+	// Hack for compatiblity with previous RSX captures
+	if (rsx::get_current_renderer()->state & cpu_flag::exit || GET_SERIALIZATION_VERSION(global_version))
 	{
 		ar(o.transform_constants);
 	}

rpcs3/Emu/System.cpp

@@ -1878,22 +1878,6 @@ void Emulator::RunPPU()
 		signalled_thread = true;
 	});
 
-	// Run SPUs waiting on a syscall (savestates related)
-	idm::select<named_thread<spu_thread>>([&](u32, named_thread<spu_thread>& spu)
-	{
-		if (spu.group && spu.index == spu.group->waiter_spu_index)
-		{
-			if (std::exchange(spu.stop_flag_removal_protection, false))
-			{
-				return;
-			}
-
-			ensure(spu.state.test_and_reset(cpu_flag::stop));
-			spu.state.notify_one(cpu_flag::stop);
-			signalled_thread = true;
-		}
-	});
-
 	if (!signalled_thread)
 	{
 		FixGuestTime();
@@ -2169,9 +2153,15 @@ void Emulator::GracefulShutdown(bool allow_autoexit, bool async_op, bool savesta
 }
 
 extern bool try_lock_vdec_context_creation();
+extern bool try_lock_spu_threads_in_a_state_compatible_with_savestates();
 
 void Emulator::Kill(bool allow_autoexit, bool savestate)
 {
+	if (savestate && !try_lock_spu_threads_in_a_state_compatible_with_savestates())
+	{
+		sys_log.error("Failed to savestate: failed to lock SPU threads execution.");
+	}
+
 	if (savestate && !try_lock_vdec_context_creation())
 	{
 		sys_log.error("Failed to savestate: HLE VDEC (video decoder) context(s) exist."