调优性能

CMakeLists.txt

@@ -12,7 +12,9 @@ elseif ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
 endif()
 
 if (UNIX)
-	set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -pthread")
+	find_package(mimalloc 1.4 REQUIRED)
+
+	set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fPIC -pthread")
 endif()
 
 if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "MSVC")
@@ -112,4 +114,9 @@ ${PROJECT_SOURCE_DIR}/test_librf.cpp
 
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${PROJECT_SOURCE_DIR}/bin)
 add_executable(rf_tutorial ${TSRC})
-target_link_libraries(rf_tutorial ${LIB_NAME})
+
+if (UNIX)
+	target_link_libraries(rf_tutorial ${LIB_NAME} mimalloc)
+else()
+	target_link_libraries(rf_tutorial ${LIB_NAME})
+endif()

librf/src/config.h

@@ -24,4 +24,8 @@
 /* #undef _WITH_LOCK_FREE_Q_KEEP_REAL_SIZE */
 #endif	//_WITH_LOCK_FREE_Q_KEEP_REAL_SIZE
 
+#ifndef RESUMEF_DISABLE_MULT_THREAD
+/* #undef RESUMEF_DISABLE_MULT_THREAD */
+#endif	//RESUMEF_DISABLE_MULT_THREAD
+
 /* #undef RESUMEF_USE_CUSTOM_SPINLOCK */

librf/src/macro_def.inl

@@ -15,3 +15,19 @@
 #define current_scheduler() (co_await ::resumef::get_current_scheduler())
 #define root_state() (co_await ::resumef::get_root_state())
 #define current_task() (co_await ::resumef::get_current_task())
+
+#ifdef _MSC_VER
+#ifndef likely
+#define likely(x) x
+#endif // likely
+#ifndef unlikely
+#define unlikely(x) x
+#endif // unlikely
+#else  // _MSC_VER
+#ifndef likely
+#define likely(x) __builtin_expect(!!(x), 1)
+#endif // likely
+#ifndef unlikely
+#define unlikely(x) __builtin_expect(!!(x), 0)
+#endif // unlikely
+#endif // _MSC_VER

librf/src/scheduler.cpp

@@ -113,24 +113,6 @@ namespace resumef
 		return task;
 	}
 
-	void scheduler_t::add_generator(state_base_t* sptr)
-	{
-		assert(sptr != nullptr);
-
-#if !RESUMEF_DISABLE_MULT_THREAD
-		scoped_lock<spinlock> __guard(_lock_running);
-#endif
-		_runing_states.emplace_back(sptr);
-	}
-
-	void scheduler_t::del_final(state_base_t* sptr)
-	{
-#if !RESUMEF_DISABLE_MULT_THREAD
-		scoped_lock<spinlock> __guard(_lock_ready);
-#endif
-		this->_ready_task.erase(sptr);
-	}
-
 	std::unique_ptr<task_t> scheduler_t::del_switch(state_base_t* sptr)
 	{
 #if !RESUMEF_DISABLE_MULT_THREAD
@@ -149,28 +131,6 @@ namespace resumef
 		return task_ptr;
 	}
 
-	void scheduler_t::add_switch(std::unique_ptr<task_t> task)
-	{
-		state_base_t* sptr = task->_state.get();
-
-#if !RESUMEF_DISABLE_MULT_THREAD
-		scoped_lock<spinlock> __guard(_lock_ready);
-#endif
-		this->_ready_task.emplace(sptr, std::move(task));
-	}
-
-	task_t* scheduler_t::find_task(state_base_t* sptr) const noexcept
-	{
-#if !RESUMEF_DISABLE_MULT_THREAD
-		scoped_lock<spinlock> __guard(_lock_ready);
-#endif
-
-		auto iter = this->_ready_task.find(sptr);
-		if (iter != this->_ready_task.end())
-			return iter->second.get();
-		return nullptr;
-	}
-
 /*
 	void scheduler_t::cancel_all_task_()
 	{
@@ -195,7 +155,7 @@ namespace resumef
 #if !RESUMEF_DISABLE_MULT_THREAD
 			scoped_lock<spinlock> __guard(_lock_running);
 #endif
-			if (_runing_states.empty())
+			if (likely(_runing_states.empty()))
 				return false;
 
 			std::swap(_cached_states, _runing_states);
@@ -215,15 +175,15 @@ namespace resumef
 			//介于网上有人做评测，导致单协程切换数据很难看，那就注释掉吧。
 			//std::this_thread::yield();
 
-			if (this->run_one_batch()) continue;	//当前运行了一个state，则认为还可能有任务未完成
+			if (likely(this->run_one_batch())) continue;	//当前运行了一个state，则认为还可能有任务未完成
 
 			{
 #if !RESUMEF_DISABLE_MULT_THREAD
 				scoped_lock<spinlock> __guard(_lock_ready);
 #endif
-				if (!_ready_task.empty()) continue;	//当前还存在task，则必然还有任务未完成
+				if (likely(!_ready_task.empty())) continue;	//当前还存在task，则必然还有任务未完成
 			}
-			if (!_timer->empty()) continue;			//定时器不为空，也需要等待定时器触发
+			if (unlikely(!_timer->empty())) continue;			//定时器不为空，也需要等待定时器触发
 
 			break;
 		};

librf/src/scheduler.h

@@ -146,4 +146,44 @@ namespace resumef
 	private:
 		scheduler_t* _scheduler_ptr;
 	};
+
+	inline void scheduler_t::add_generator(state_base_t* sptr)
+	{
+		assert(sptr != nullptr);
+
+#if !RESUMEF_DISABLE_MULT_THREAD
+		scoped_lock<spinlock> __guard(_lock_running);
+#endif
+		_runing_states.emplace_back(sptr);
+	}
+
+	inline void scheduler_t::del_final(state_base_t* sptr)
+	{
+#if !RESUMEF_DISABLE_MULT_THREAD
+		scoped_lock<spinlock> __guard(_lock_ready);
+#endif
+		this->_ready_task.erase(sptr);
+	}
+
+	inline void scheduler_t::add_switch(std::unique_ptr<task_t> task)
+	{
+		state_base_t* sptr = task->_state.get();
+
+#if !RESUMEF_DISABLE_MULT_THREAD
+		scoped_lock<spinlock> __guard(_lock_ready);
+#endif
+		this->_ready_task.emplace(sptr, std::move(task));
+	}
+
+	inline task_t* scheduler_t::find_task(state_base_t* sptr) const noexcept
+	{
+#if !RESUMEF_DISABLE_MULT_THREAD
+		scoped_lock<spinlock> __guard(_lock_ready);
+#endif
+
+		auto iter = this->_ready_task.find(sptr);
+		if (iter != this->_ready_task.end())
+			return iter->second.get();
+		return nullptr;
+	}
 }

librf/src/state.cpp

@@ -57,10 +57,14 @@ namespace resumef
 
 	void state_generator_t::resume()
 	{
-		if (_coro)
+		if (likely(_coro))
 		{
 			_coro.resume();
-			if (_coro.done())
+			if (likely(!_coro.done()))
+			{
+				_scheduler->add_generator(this);
+			}
+			else
 			{
 				coroutine_handle<> handler = _coro;
 				_coro = nullptr;
@@ -68,10 +72,6 @@ namespace resumef
 
 				handler.destroy();
 			}
-			else
-			{
-				_scheduler->add_generator(this);
-			}
 		}
 	}
 

librf/src/timer.cpp

@@ -25,9 +25,13 @@ namespace resumef
 
 	void timer_manager::clear()
 	{
+#if !RESUMEF_DISABLE_MULT_THREAD
 		std::unique_lock<spinlock> __lock(_added_mtx);
+#endif
 		auto _atimer = std::move(_added_timers);
+#if !RESUMEF_DISABLE_MULT_THREAD
 		__lock.unlock();
+#endif
 
 		for (auto& sptr : _atimer)
 			call_target_(sptr, true);
@@ -42,7 +46,9 @@ namespace resumef
 		assert(sptr);
 		assert(sptr->st == timer_target::State::Invalid);
 
+#if !RESUMEF_DISABLE_MULT_THREAD
 		scoped_lock<spinlock> __lock(_added_mtx);
+#endif
 #if _DEBUG
 		assert(sptr->_manager == nullptr);
 		sptr->_manager = this;
@@ -69,13 +75,17 @@ namespace resumef
 	void timer_manager::update()
 	{
 		{
+#if !RESUMEF_DISABLE_MULT_THREAD
 			std::unique_lock<spinlock> __lock(_added_mtx);
+#endif
 
-			if (_added_timers.size() > 0)
+			if (unlikely(_added_timers.size() > 0))
 			{
 				auto _atimer = std::move(_added_timers);
 				_added_timers.reserve(128);
+#if !RESUMEF_DISABLE_MULT_THREAD
 				__lock.unlock();
+#endif
 
 				for (auto& sptr : _atimer)
 				{
@@ -93,7 +103,7 @@ namespace resumef
 			}
 		}
 
-		if (_runing_timers.size() > 0)
+		if (unlikely(_runing_timers.size() > 0))
 		{
 			auto now_ = clock_type::now();
 

librf/src/timer.h

@@ -139,7 +139,9 @@ namespace resumef
 			return add_(std::make_shared<timer_target>(tp_, std::forward<_Cb>(cb_)));
 		}
 	private:
+#if !RESUMEF_DISABLE_MULT_THREAD
 		spinlock _added_mtx;
+#endif
 		timer_vector_type	_added_timers;
 		timer_map_type		_runing_timers;
 

vs_proj/librf.vcxproj

@@ -166,13 +166,13 @@
       </SDLCheck>
       <EnablePREfast>true</EnablePREfast>
       <InlineFunctionExpansion>AnySuitable</InlineFunctionExpansion>
-      <OmitFramePointers>true</OmitFramePointers>
       <BufferSecurityCheck>false</BufferSecurityCheck>
       <EnableParallelCodeGeneration>true</EnableParallelCodeGeneration>
       <DisableSpecificWarnings>4834;4505</DisableSpecificWarnings>
       <FunctionLevelLinking>true</FunctionLevelLinking>
       <RuntimeTypeInfo>true</RuntimeTypeInfo>
       <ExceptionHandling>Async</ExceptionHandling>
+      <EnableFiberSafeOptimizations>true</EnableFiberSafeOptimizations>
     </ClCompile>
     <Link>
       <SubSystem>Console</SubSystem>
@@ -180,8 +180,7 @@
       <OptimizeReferences>true</OptimizeReferences>
       <GenerateDebugInformation>true</GenerateDebugInformation>
       <LinkTimeCodeGeneration>UseLinkTimeCodeGeneration</LinkTimeCodeGeneration>
-      <Profile>
-      </Profile>
+      <Profile>true</Profile>
     </Link>
   </ItemDefinitionGroup>
   <ItemGroup>