Rewrite cond_variable to use waitable atomics

Increase max_timeout and fix max_timeout usage

Utilities/Thread.cpp

@@ -1778,11 +1778,13 @@ void thread_base::finalize() noexcept
 	--g_thread_count;
 }
 
-bool thread_ctrl::_wait_for(u64 usec)
+void thread_ctrl::_wait_for(u64 usec)
 {
 	auto _this = g_tls_this_thread;
 
-	do
+	std::unique_lock lock(_this->m_mutex, std::defer_lock);
+
+	while (true)
 	{
 		// Mutex is unlocked at the start and after the waiting
 		if (u32 sig = _this->m_signal.load())
@@ -1790,17 +1792,20 @@ bool thread_ctrl::_wait_for(u64 usec)
 			if (sig & 1)
 			{
 				_this->m_signal &= ~1;
-				return true;
+				return;
 			}
 		}
 
 		if (usec == 0)
 		{
 			// No timeout: return immediately
-			return false;
+			return;
 		}
 
-		_this->m_mutex.lock();
+		if (!lock)
+		{
+			lock.lock();
+		}
 
 		// Double-check the value
 		if (u32 sig = _this->m_signal.load())
@@ -1808,15 +1813,17 @@ bool thread_ctrl::_wait_for(u64 usec)
 			if (sig & 1)
 			{
 				_this->m_signal &= ~1;
-				_this->m_mutex.unlock();
-				return true;
+				return;
 			}
 		}
-	}
-	while (_this->m_cond.wait_unlock(std::exchange(usec, usec > cond_variable::max_timeout ? -1 : 0), _this->m_mutex));
 
-	// Timeout
-	return false;
+		_this->m_cond.wait_unlock(usec, lock);
+
+		if (usec < cond_variable::max_timeout)
+		{
+			usec = 0;
+		}
+	}
 }
 
 thread_base::thread_base(std::string_view name)

Utilities/Thread.h

@@ -185,7 +185,7 @@ class thread_ctrl final
 	static atomic_t<native_core_arrangement> g_native_core_layout;
 
 	// Internal waiting function, may throw. Infinite value is -1.
-	static bool _wait_for(u64 usec);
+	static void _wait_for(u64 usec);
 
 	friend class thread_base;
 
@@ -235,9 +235,9 @@ public:
 	}
 
 	// Wait once with timeout. May spuriously return false.
-	static inline bool wait_for(u64 usec)
+	static inline void wait_for(u64 usec)
 	{
-		return _wait_for(usec);
+		_wait_for(usec);
 	}
 
 	// Wait.

Utilities/cond.cpp

@@ -8,50 +8,57 @@
 #include <thread>
 #endif
 
-bool cond_variable::imp_wait(u32 _old, u64 _timeout) noexcept
+// use constants, increase signal space
+
+void cond_variable::imp_wait(u32 _old, u64 _timeout) noexcept
 {
-	verify("cond_variable overflow" HERE), (_old & 0xffff) != 0xffff; // Very unlikely: it requires 65535 distinct threads to wait simultaneously
+	// Not supposed to fail
+	verify(HERE), _old;
 
-	return balanced_wait_until(m_value, _timeout, [&](u32& value, auto... ret) -> int
+	// Wait with timeout
+	m_value.wait(_old, atomic_wait_timeout{_timeout > max_timeout ? UINT64_MAX : _timeout * 1000});
+
+	// Cleanup
+	m_value.atomic_op([](u32& value)
 	{
-		if (value >> 16)
-		{
-			// Success
-			value -= 0x10001;
-			return +1;
-		}
+		value -= c_waiter_mask & -c_waiter_mask;
 
-		if constexpr (sizeof...(ret))
+		if ((value & c_waiter_mask) == 0)
 		{
-			// Retire
-			value -= 1;
-			return -1;
+			// Last waiter removed, clean signals
+			value = 0;
 		}
-
-		return 0;
 	});
-
-#ifdef _WIN32
-	if (_old >= 0x10000 && !OptWaitOnAddress && m_value)
-	{
-		// Workaround possibly stolen signal
-		imp_wake(1);
-	}
-#endif
 }
 
 void cond_variable::imp_wake(u32 _count) noexcept
 {
-	// TODO (notify_one)
-	balanced_awaken<true>(m_value, m_value.atomic_op([&](u32& value) -> u32
+	const auto [_old, ok] = m_value.fetch_op([](u32& value)
 	{
-		// Subtract already signaled number from total amount of waiters
-		const u32 can_sig = (value & 0xffff) - (value >> 16);
-		const u32 num_sig = std::min<u32>(can_sig, _count);
+		if (!value || (value & c_signal_mask) == c_signal_mask)
+		{
+			return false;
+		}
 
-		value += num_sig << 16;
-		return num_sig;
-	}));
+		// Add signal
+		value += c_signal_mask & -c_signal_mask;
+		return true;
+	});
+
+	if (!ok || !_count)
+	{
+		return;
+	}
+
+	if (_count > 1 || ((_old + (c_signal_mask & -c_signal_mask)) & c_signal_mask) == c_signal_mask)
+	{
+		// Resort to notify_all if signal count reached max
+		m_value.notify_all();
+	}
+	else
+	{
+		m_value.notify_one();
+	}
 }
 
 bool shared_cond::imp_wait(u32 slot, u64 _timeout) noexcept

Utilities/cond.h

@@ -11,9 +11,31 @@ class cond_variable
 	// Internal waiter counter
 	atomic_t<u32> m_value{0};
 
+	enum : u32
+	{
+		c_waiter_mask = 0x1fff,
+		c_signal_mask = 0xffffffff & ~c_waiter_mask,
+	};
+
 protected:
+	// Increment waiter count
+	u32 add_waiter() noexcept
+	{
+		return m_value.atomic_op([](u32& value) -> u32
+		{
+			if ((value & c_signal_mask) == c_signal_mask || (value & c_waiter_mask) == c_waiter_mask)
+			{
+				// Signal or waiter overflow, return immediately
+				return 0;
+			}
+
+			value += c_waiter_mask & -c_waiter_mask;
+			return value;
+		});
+	}
+
 	// Internal waiting function
-	bool imp_wait(u32 _old, u64 _timeout) noexcept;
+	void imp_wait(u32 _old, u64 _timeout) noexcept;
 
 	// Try to notify up to _count threads
 	void imp_wake(u32 _count) noexcept;
@@ -23,22 +45,33 @@ public:
 
 	// Intrusive wait algorithm for lockable objects
 	template <typename T>
-	bool wait(T& object, u64 usec_timeout = -1)
+	void wait(T& object, u64 usec_timeout = -1) noexcept
 	{
-		const u32 _old = m_value.fetch_add(1); // Increment waiter counter
+		const u32 _old = add_waiter();
+
+		if (!_old)
+		{
+			return;
+		}
+
 		object.unlock();
-		const bool res = imp_wait(_old, usec_timeout);
+		imp_wait(_old, usec_timeout);
 		object.lock();
-		return res;
 	}
 
 	// Unlock all specified objects but don't lock them again
 	template <typename... Locks>
-	bool wait_unlock(u64 usec_timeout, Locks&&... locks)
+	void wait_unlock(u64 usec_timeout, Locks&&... locks)
 	{
-		const u32 _old = m_value.fetch_add(1); // Increment waiter counter
+		const u32 _old = add_waiter();
 		(..., std::forward<Locks>(locks).unlock());
-		return imp_wait(_old, usec_timeout);
+
+		if (!_old)
+		{
+			return;
+		}
+
+		imp_wait(_old, usec_timeout);
 	}
 
 	// Wake one thread
@@ -55,11 +88,11 @@ public:
 	{
 		if (m_value)
 		{
-			imp_wake(65535);
+			imp_wake(-1);
 		}
 	}
 
-	static constexpr u64 max_timeout = u64{UINT32_MAX} / 1000 * 1000000;
+	static constexpr u64 max_timeout = UINT64_MAX / 1000;
 };
 
 // Condition variable fused with a pseudo-mutex supporting only reader locks (up to 32 readers).

rpcs3/Emu/Cell/lv2/sys_sync.h

@@ -236,10 +236,10 @@ public:
 		static_assert(UINT64_MAX / cond_variable::max_timeout >= g_cfg.core.clocks_scale.max, "timeout may overflow during scaling");
 
 		// Clamp to max timeout accepted
-		if (usec > cond_variable::max_timeout) usec = cond_variable::max_timeout;
+		const u64 max_usec = cond_variable::max_timeout * 100 / g_cfg.core.clocks_scale.max;
 
 		// Now scale the result
-		usec = (usec * g_cfg.core.clocks_scale) / 100;
+		usec = (std::min<u64>(usec, max_usec) * g_cfg.core.clocks_scale) / 100;
 
 #ifdef __linux__
 		// TODO: Confirm whether Apple or any BSD can benefit from this as well
@@ -271,7 +271,7 @@ public:
 					// Do not wait for the last quantum to avoid loss of accuracy
 					thread_ctrl::wait_for(remaining - ((remaining % host_min_quantum) + host_min_quantum));
 #else
-					// Wait on multiple of min quantum for large durations to avoid overloading low thread cpus 
+					// Wait on multiple of min quantum for large durations to avoid overloading low thread cpus
 					thread_ctrl::wait_for(remaining - (remaining % host_min_quantum));
 #endif
 				}