SPU: Optimize concurrent PUTLLUC

rpcs3/Emu/Cell/PPUThread.cpp

@@ -1285,7 +1285,7 @@ static bool ppu_store_reservation(ppu_thread& ppu, u32 addr, T reg_value)
 		}
 		}
 
-		if (res == ppu.rtime && vm::reservation_trylock(res, ppu.rtime))
+		if (res == ppu.rtime && vm::reservation_trylock(res, ppu.rtime).first)
 		{
 			if (data.compare_and_swap_test(old_data, reg_value))
 			{
@@ -1300,7 +1300,7 @@ static bool ppu_store_reservation(ppu_thread& ppu, u32 addr, T reg_value)
 		return false;
 	}
 
-	if (!vm::reservation_trylock(res, ppu.rtime))
+	if (!vm::reservation_trylock(res, ppu.rtime).first)
 	{
 		return false;
 	}

rpcs3/Emu/Cell/SPUThread.cpp

@@ -1883,7 +1883,7 @@ bool spu_thread::do_putllc(const spu_mfc_cmd& args)
 			}
 		}
 
-		if (!vm::reservation_trylock(res, rtime))
+		if (!vm::reservation_trylock(res, rtime).first)
 		{
 			return false;
 		}
@@ -2003,7 +2003,34 @@ void spu_thread::do_putlluc(const spu_mfc_cmd& args)
 	else
 	{
 		auto& data = vm::_ref<decltype(rdata)>(addr);
-		auto [res, time0] = vm::reservation_lock(addr, 128);
+		auto& res = vm::reservation_acquire(addr, 128);
+
+		for(u64 time0 = UINT64_MAX & -128;;)
+		{
+			auto [ok, time1] = vm::reservation_trylock(res, res & -128, vm::putlluc_lockb);
+
+			if (time0 & vm::putlluc_lockb && time1 > time0)
+			{
+				// Other PUTLLUC completed, don't execute
+
+				if (ok)
+				{
+					res.release(time1 & -128);
+				}
+
+				return;
+			}
+
+			if (ok)
+			{
+				break;
+			}
+
+			if (time1 & vm::putlluc_lockb)
+			{
+				time0 = time1;
+			}
+		}
 
 		*reinterpret_cast<atomic_t<u32>*>(&data) += 0;
 

rpcs3/Emu/Memory/vm.cpp

@@ -444,7 +444,7 @@ namespace vm
 	{
 		for (u64 i = 0;; i++)
 		{
-			if (u64 rtime = res; !(rtime & 127) && reservation_trylock(res, rtime, lock_bits)) [[likely]]
+			if (u64 rtime = res; !(rtime & 127) && reservation_trylock(res, rtime, lock_bits).first) [[likely]]
 			{
 				return rtime;
 			}

rpcs3/Emu/Memory/vm_reservation.h

@@ -40,22 +40,22 @@ namespace vm
 
 	u64 reservation_lock_internal(u32, atomic_t<u64>&, u64);
 
-	inline bool reservation_trylock(atomic_t<u64>& res, u64 rtime, u64 lock_bits = stcx_lockb)
+	inline std::pair<bool, u64> reservation_trylock(atomic_t<u64>& res, u64 rtime, u64 lock_bits = stcx_lockb)
 	{
-		if (res.compare_and_swap_test(rtime, rtime + lock_bits)) [[likely]]
+		if (res.compare_exchange(rtime, rtime + lock_bits)) [[likely]]
 		{
-			return true;
+			return {true, rtime};
 		}
 
-		return false;
+		return {false, rtime};
 	}
 
 	inline std::pair<atomic_t<u64>&, u64> reservation_lock(u32 addr, u32 size, u64 lock_bits = stcx_lockb)
 	{
 		auto res = &vm::reservation_acquire(addr, size);
 		auto rtime = res->load();
 
-		if (rtime & 127 || !reservation_trylock(*res, rtime, lock_bits)) [[unlikely]]
+		if (rtime & 127 || !reservation_trylock(*res, rtime, lock_bits).first) [[unlikely]]
 		{
 			static atomic_t<u64> no_lock{};