LV2: Re-add dropped optimization by previous commit

Currently only for lock-free syscalls.

Utilities/bit_set.h

@@ -371,12 +371,12 @@ public:
 	bool bit_test_reset(uint bit) = delete;
 	bool bit_test_invert(uint bit) = delete;
 
-	bool all_of(bs_t arg)
+	bool all_of(bs_t arg) const
 	{
 		return base::load().all_of(arg);
 	}
 
-	bool none_of(bs_t arg)
+	bool none_of(bs_t arg) const
 	{
 		return base::load().none_of(arg);
 	}

rpcs3/Emu/Cell/lv2/sys_lwmutex.cpp

@@ -173,6 +173,7 @@ error_code _sys_lwmutex_lock(ppu_thread& ppu, u32 lwmutex_id, u64 timeout)
 		}
 
 		mutex.sleep(ppu, timeout);
+		notify.cleanup();
 		return false;
 	});
 
@@ -308,7 +309,7 @@ error_code _sys_lwmutex_unlock(ppu_thread& ppu, u32 lwmutex_id)
 			}
 
 			mutex.awake(cpu);
-			return;
+			notify.cleanup(); // lv2_lwmutex::mutex is not really active 99% of the time, can be ignored
 		}
 	});
 
@@ -345,7 +346,7 @@ error_code _sys_lwmutex_unlock2(ppu_thread& ppu, u32 lwmutex_id)
 
 			static_cast<ppu_thread*>(cpu)->gpr[3] = CELL_EBUSY;
 			mutex.awake(cpu);
-			return;
+			notify.cleanup(); // lv2_lwmutex::mutex is not really active 99% of the time, can be ignored
 		}
 	});
 

rpcs3/Emu/Cell/lv2/sys_mutex.cpp

@@ -169,6 +169,7 @@ error_code sys_mutex_lock(ppu_thread& ppu, u32 mutex_id, u64 timeout)
 			else
 			{
 				mutex.sleep(ppu, timeout);
+				notify.cleanup();
 			}
 		}
 
@@ -314,6 +315,7 @@ error_code sys_mutex_unlock(ppu_thread& ppu, u32 mutex_id)
 			result = {};
 		}
 
+		notify.cleanup();
 		return result;
 	});
 

rpcs3/Emu/Cell/lv2/sys_sync.h

@@ -507,11 +507,14 @@ public:
 				return;
 			}
 
+			if (cpu != &g_to_notify)
+			{
 				// Note: by the time of notification the thread could have been deallocated which is why the direct function is used
 				// TODO: Pass a narrower mask
 				atomic_wait_engine::notify_one(cpu, 4, atomic_wait::default_mask<atomic_bs_t<cpu_flag>>);
 			}
 		}
+	}
 
 	// Can be called before the actual sleep call in order to move it out of mutex scope
 	static inline void prepare_for_sleep(cpu_thread& cpu)
@@ -527,6 +530,26 @@ public:
 			g_postpone_notify_barrier = true;
 		}
 
+		notify_all_t(const notify_all_t&) = delete;
+	
+		static void cleanup()
+		{
+			for (auto& cpu : g_to_notify)
+			{
+				if (!cpu)
+				{
+					return;
+				}
+
+				// While IDM mutex is still locked (this function assumes so) check if the notification is still needed
+				if (cpu != &g_to_notify && !static_cast<const decltype(cpu_thread::state)*>(cpu)->all_of(cpu_flag::signal + cpu_flag::wait))
+				{
+					// Omit it (this is a void pointer, it can hold anything)
+					cpu = &g_to_notify;
+				}
+			}
+		}
+
 		~notify_all_t() noexcept
 		{
 			lv2_obj::notify_all();