Rewrite Utilitis/sema.cpp

Utilities/sema.cpp

@@ -8,72 +8,84 @@ void semaphore_base::imp_wait()
 	{
 		busy_wait();
 
-		const s32 value = m_value.load();
+		const u32 value = m_value.load();
 
-		if (value > 0 && m_value.compare_and_swap_test(value, value - 1))
+		if (value & c_value_mask && m_value.compare_and_swap_test(value, value - c_value))
 		{
 			return;
 		}
 	}
 
+	bool waits = false;
+
 	while (true)
 	{
 		// Try hard way
-		const s32 value = m_value.atomic_op([](s32& value)
+		const u32 value = m_value.fetch_op([&](u32& value)
 		{
-			// Use sign bit to acknowledge waiter presence
+			ensure(value != c_waiter_mask); // "semaphore_base: overflow"
-			if (value && value > smin)
-			{
-				value--;
 
-				if (value < 0)
+			if (value & c_value_mask)
+			{
+				// Obtain signal
+				value -= c_value;
+
+				if (waits)
 				{
-					// Remove sign bit
+					// Remove waiter
-					value -= s32{smin};
+					value -= c_waiter;
 				}
 			}
-			else
+			else if (!waits)
 			{
-				// Set sign bit
+				// Add waiter
-				value = smin;
+				value += c_waiter;
 			}
-
-			return value;
 		});
 
-		if (value >= 0)
+		if (value & c_value_mask)
 		{
-			// Signal other waiter to wake up or to restore sign bit
+			break;
-			m_value.notify_one();
-			return;
 		}
 
-		m_value.wait(value);
+		m_value.wait(value + (waits ? 0 : c_waiter));
+		waits = true;
 	}
 }
 
-void semaphore_base::imp_post(s32 _old)
+void semaphore_base::imp_post(u32 _old)
 {
-	ensure(_old < 0); // "semaphore_base: overflow"
+	ensure(~_old & c_value_mask); // "semaphore_base: overflow"
 
-	m_value.notify_one();
+	if ((_old & c_waiter_mask) / c_waiter > (_old & c_value_mask) / c_value)
+	{
+		m_value.notify_one();
+	}
 }
 
-bool semaphore_base::try_post(s32 _max)
+bool semaphore_base::try_post(u32 _max)
 {
 	// Conditional increment
-	const s32 value = m_value.fetch_op([&](s32& value)
+	const auto [value, ok] = m_value.fetch_op([&](u32& value)
 	{
-		if (value < _max)
+		if ((value & c_value_mask) <= _max)
 		{
-			value += 1;
+			value += c_value;
+			return true;
 		}
+
+		return false;
 	});
 
-	if (value < 0)
+	if (!ok)
+	{
+		return false;
+	}
+
+	if (value & c_waiter_mask)
 	{
 		imp_post(value);
 	}
 
-	return value < _max;
+	return true;
 }

Utilities/sema.h

@@ -7,14 +7,22 @@
 class semaphore_base
 {
 	// Semaphore value
-	atomic_t<s32> m_value;
+	atomic_t<u32> m_value;
+
+	enum : u32
+	{
+		c_value = 1u << 0,
+		c_value_mask = +c_value * 0xffff,
+		c_waiter = 1u << 16,
+		c_waiter_mask = +c_waiter * 0xffff,
+	};
 
 	void imp_wait();
 
-	void imp_post(s32 _old);
+	void imp_post(u32 _old);
 
 protected:
-	explicit constexpr semaphore_base(s32 value)
+	explicit constexpr semaphore_base(u32 value) noexcept
 		: m_value{value}
 	{
 	}
@@ -22,10 +30,10 @@ protected:
 	void wait()
 	{
 		// Load value
-		const s32 value = m_value.load();
+		const u32 value = m_value.load();
 
 		// Conditional decrement
-		if (value <= 0 || !m_value.compare_and_swap_test(value, value - 1)) [[unlikely]]
+		if ((value & c_value_mask) == 0 || !m_value.compare_and_swap_test(value, value - c_value)) [[unlikely]]
 		{
 			imp_wait();
 		}
@@ -33,36 +41,47 @@ protected:
 
 	bool try_wait()
 	{
-		return m_value.try_dec(0);
+		return m_value.fetch_op([](u32& value)
+		{
+			if (value & c_value_mask)
+			{
+				value -= c_value;
+				return true;
+			}
+
+			return false;
+		}).second;
 	}
 
-	void post(s32 _max)
+	void post(u32 _max)
 	{
 		// Unconditional increment
-		const s32 value = m_value.fetch_add(1);
+		const u32 value = m_value.fetch_add(c_value);
 
-		if (value < 0 || value >= _max) [[unlikely]]
+		if (value & c_waiter_mask || (value & c_value_mask) >= std::min<u32>(c_value_mask, _max)) [[unlikely]]
 		{
 			imp_post(value);
 		}
 	}
 
-	bool try_post(s32 _max);
+	bool try_post(u32 _max);
 
 public:
 	// Get current semaphore value
 	s32 get() const
 	{
 		// Load value
-		const s32 value = m_value;
+		const u32 raw_value = m_value;
+		const u32 waiters = (raw_value & c_waiter_mask) / c_waiter;
+		const u32 value = (raw_value & c_value_mask) / c_value;
 
 		// Return only positive value
-		return value < 0 ? 0 : value;
+		return static_cast<s32>(waiters >= value ? 0 : value - waiters);
 	}
 };
 
 // Lightweight semaphore template (default arguments define binary semaphore and Def == Max)
-template <s32 Max = 1, s32 Def = Max>
+template <s16 Max = 1, s16 Def = Max>
 class semaphore final : public semaphore_base
 {
 	static_assert(Max >= 0, "semaphore<>: Max is out of bounds");
@@ -73,13 +92,13 @@ class semaphore final : public semaphore_base
 
 public:
 	// Default constructor (recommended)
-	constexpr semaphore()
+	constexpr semaphore() noexcept
 		: base(Def)
 	{
 	}
 
 	// Explicit value constructor (not recommended)
-	explicit constexpr semaphore(s32 value)
+	explicit constexpr semaphore(s16 value) noexcept
 		: base(value)
 	{
 	}

rpcs3/Emu/Cell/PPUThread.cpp

@@ -3420,14 +3420,14 @@ extern bool ppu_stdcx(ppu_thread& ppu, u32 addr, u64 reg_value)
 
 struct jit_core_allocator
 {
-	const s32 thread_count = g_cfg.core.llvm_threads ? std::min<s32>(g_cfg.core.llvm_threads, limit()) : limit();
+	const s16 thread_count = g_cfg.core.llvm_threads ? std::min<s32>(g_cfg.core.llvm_threads, limit()) : limit();
 
 	// Initialize global semaphore with the max number of threads
-	::semaphore<0x7fffffff> sem{std::max<s32>(thread_count, 1)};
+	::semaphore<0x7fff> sem{std::max<s16>(thread_count, 1)};
 
-	static s32 limit()
+	static s16 limit()
 	{
-		return static_cast<s32>(utils::get_thread_count());
+		return static_cast<s16>(std::min<s32>(0x7fff, utils::get_thread_count()));
 	}
 };