atomic.cpp: use new thread alerting API (Win8+)

Win7 will remain using old API (keyed events).

rpcs3/util/atomic.cpp

@@ -104,8 +104,17 @@ namespace
 				// Find lowest clear bit
 				const auto sema = &sema_data[std::countr_one(bits)];
 
-#if defined(USE_FUTEX) || defined(_WIN32)
+#if defined(USE_FUTEX)
 				sema->release(1);
+#elif defined(_WIN32)
+				if (NtWaitForAlertByThreadId)
+				{
+					sema->release(GetCurrentThreadId());
+				}
+				else
+				{
+					sema->release(1);
+				}
 #endif
 
 				return sema;
@@ -396,10 +405,16 @@ void atomic_storage_futex::wait(const void* data, std::size_t size, u64 old_valu
 	}
 
 	// Can skip unqueue process if true
-#ifdef USE_FUTEX
-	bool fallback = true;
-#else
+#if defined(USE_FUTEX)
+	const bool fallback = true;
+#elif defined(_WIN32)
 	bool fallback = false;
+
+	if (NtWaitForAlertByThreadId)
+	{
+		// Similarly to futex, never required
+		fallback = true;
+	}
 #endif
 
 	while (ptr_cmp(data, size, old_value, mask))
@@ -428,18 +443,33 @@ void atomic_storage_futex::wait(const void* data, std::size_t size, u64 old_valu
 			qw.QuadPart -= 1;
 		}
 
-		if (fallback)
+		if (NtWaitForAlertByThreadId)
 		{
-			// Restart waiting
-			verify(HERE), sema->load() == 2;
-			sema->release(1);
-			fallback = false;
+			if (sema->load() == umax) [[unlikely]]
+			{
+				// Signaled prematurely
+				sema->release(GetCurrentThreadId());
+			}
+			else
+			{
+				NtWaitForAlertByThreadId(const_cast<void*>(data), timeout + 1 ? &qw : nullptr);
+			}
 		}
-
-		if (!NtWaitForKeyedEvent(nullptr, sema, false, timeout + 1 ? &qw : nullptr))
+		else
 		{
-			// Error code assumed to be timeout
-			fallback = true;
+			if (fallback)
+			{
+				// Restart waiting
+				verify(HERE), sema->load() == 2;
+				sema->release(1);
+				fallback = false;
+			}
+
+			if (!NtWaitForKeyedEvent(nullptr, sema, false, timeout + 1 ? &qw : nullptr))
+			{
+				// Error code assumed to be timeout
+				fallback = true;
+			}
 		}
 #endif
 
@@ -487,6 +517,19 @@ static inline bool alert_sema(atomic_t<u32>* sema)
 		return true;
 	}
 #elif defined(_WIN32)
+	if (NtWaitForAlertByThreadId)
+	{
+		u32 tid = sema->load();
+
+		if (tid + 1 <= 1 && sema->compare_and_swap_test(tid, -1))
+		{
+			verify(HERE), NtAlertThreadByThreadId(tid) == STATUS_SUCCESS;
+			return true;
+		}
+
+		return false;
+	}
+
 	if (sema->load() == 1 && sema->compare_and_swap_test(1, 2))
 	{
 		// Can wait in rare cases, which is its annoying weakness
@@ -548,7 +591,7 @@ void atomic_storage_futex::notify_all(const void* data)
 	}
 
 #if defined(_WIN32) && !defined(USE_FUTEX)
-	if (true)
+	if (!NtAlertThreadByThreadId)
 	{
 		// Make a copy to filter out waiters that fail some checks
 		u64 copy = slot->sema_bits.load();