merge with viewer-release

1 files changed, 127 insertions, 190 deletions

diff --git a/indra/llcommon/llfasttimer.h b/indra/llcommon/llfasttimer.h
index e42e549df5..32a0629a87 100644
--- a/indra/llcommon/llfasttimer.h
+++ b/indra/llcommon/llfasttimer.h
@@ -28,214 +28,98 @@
 #define LL_FASTTIMER_H
 
 #include "llinstancetracker.h"
+#include "lltrace.h"
 
 #define FAST_TIMER_ON 1
-#define DEBUG_FAST_TIMER_THREADS 1
+#define LL_FASTTIMER_USE_RDTSC 1
 
 class LLMutex;
 
-#include <queue>
-#include "llsd.h"
-
-#define LL_FASTTIMER_USE_RDTSC 1
-
+namespace LLTrace
+{
 
-LL_COMMON_API void assert_main_thread();
+struct BlockTimerStackRecord
+{
+	class BlockTimer*	mActiveTimer;
+	class TimeBlock*	mTimeBlock;
+	U64					mChildTime;
+};
 
-class LL_COMMON_API LLFastTimer
+class ThreadTimerStack 
+:	public BlockTimerStackRecord, 
+	public LLThreadLocalSingleton<ThreadTimerStack>
 {
-public:
-	class NamedTimer;
+	friend class LLThreadLocalSingleton<ThreadTimerStack>;
+	ThreadTimerStack() 
+	{}
 
-	struct LL_COMMON_API FrameState
-	{
-		FrameState();
-		void setNamedTimer(NamedTimer* timerp) { mTimer = timerp; }
-
-		U32 				mSelfTimeCounter;
-		U32 				mCalls;
-		FrameState*			mParent;		// info for caller timer
-		FrameState*			mLastCaller;	// used to bootstrap tree construction
-		NamedTimer*			mTimer;
-		U16					mActiveCount;	// number of timers with this ID active on stack
-		bool				mMoveUpTree;	// needs to be moved up the tree of timers at the end of frame
-	};
-
-	// stores a "named" timer instance to be reused via multiple LLFastTimer stack instances
-	class LL_COMMON_API NamedTimer
-	:	public LLInstanceTracker<NamedTimer>
+public:
+	ThreadTimerStack& operator=(const BlockTimerStackRecord& other)
 	{
-		friend class DeclareTimer;
-	public:
-		~NamedTimer();
-
-		enum { HISTORY_NUM = 300 };
-
-		const std::string& getName() const { return mName; }
-		NamedTimer* getParent() const { return mParent; }
-		void setParent(NamedTimer* parent);
-		S32 getDepth();
-		std::string getToolTip(S32 history_index = -1);
-
-		typedef std::vector<NamedTimer*>::const_iterator child_const_iter;
-		child_const_iter beginChildren();
-		child_const_iter endChildren();
-		std::vector<NamedTimer*>& getChildren();
-
-		void setCollapsed(bool collapsed) { mCollapsed = collapsed; }
-		bool getCollapsed() const { return mCollapsed; }
-
-		U32 getCountAverage() const { return mCountAverage; }
-		U32 getCallAverage() const { return mCallAverage; }
-
-		U32 getHistoricalCount(S32 history_index = 0) const;
-		U32 getHistoricalCalls(S32 history_index = 0) const;
-
-		void setFrameState(FrameState* state) { mFrameState = state; state->setNamedTimer(this); }
-		FrameState& getFrameState() const;
-
-	private:
-		friend class LLFastTimer;
-		friend class NamedTimerFactory;
-
-		//
-		// methods
-		//
-		NamedTimer(const std::string& name);
-		// recursive call to gather total time from children
-		static void accumulateTimings();
-
-		// updates cumulative times and hierarchy,
-		// can be called multiple times in a frame, at any point
-		static void processTimes();
-
-		static void buildHierarchy();
-		static void resetFrame();
-		static void reset();
-
-		//
-		// members
-		//
-		FrameState*		mFrameState;
-
-		std::string	mName;
-
-		U32 		mTotalTimeCounter;
-
-		U32 		mCountAverage;
-		U32			mCallAverage;
+		BlockTimerStackRecord::operator=(other);
+		return *this;
+	}
+};
 
-		U32*		mCountHistory;
-		U32*		mCallHistory;
+class BlockTimer
+{
+public:
+	friend class TimeBlock;
+	typedef BlockTimer self_t;
+	typedef class TimeBlock DeclareTimer;
 
-		// tree structure
-		NamedTimer*					mParent;				// NamedTimer of caller(parent)
-		std::vector<NamedTimer*>	mChildren;
-		bool						mCollapsed;				// don't show children
-		bool						mNeedsSorting;			// sort children whenever child added
-	};
+	BlockTimer(TimeBlock& timer);
+	~BlockTimer();
 
-	// used to statically declare a new named timer
-	class LL_COMMON_API DeclareTimer
-	:	public LLInstanceTracker<DeclareTimer>
-	{
-		friend class LLFastTimer;
-	public:
-		DeclareTimer(const std::string& name, bool open);
-		DeclareTimer(const std::string& name);
+	LLUnit<LLUnits::Seconds, F64> getElapsedTime();
 
-		NamedTimer& getNamedTimer() { return mTimer; }
+private:
 
-	private:
-		FrameState		mFrameState;
-		NamedTimer&		mTimer;
-	};
+	U64						mStartTime;
+	U64						mBlockStartTotalTimeCounter;
+	BlockTimerStackRecord	mParentTimerData;
+};
 
+// stores a "named" timer instance to be reused via multiple BlockTimer stack instances
+class TimeBlock 
+:	public TraceType<TimeBlockAccumulator>,
+	public LLInstanceTracker<TimeBlock>
+{
 public:
-	LLFastTimer(LLFastTimer::FrameState* state);
-
-	LL_FORCE_INLINE LLFastTimer(LLFastTimer::DeclareTimer& timer)
-	:	mFrameState(&timer.mFrameState)
-	{
-#if FAST_TIMER_ON
-		LLFastTimer::FrameState* frame_state = mFrameState;
-		mStartTime = getCPUClockCount32();
-
-		frame_state->mActiveCount++;
-		frame_state->mCalls++;
-		// keep current parent as long as it is active when we are
-		frame_state->mMoveUpTree |= (frame_state->mParent->mActiveCount == 0);
-
-		LLFastTimer::CurTimerData* cur_timer_data = &LLFastTimer::sCurTimerData;
-		mLastTimerData = *cur_timer_data;
-		cur_timer_data->mCurTimer = this;
-		cur_timer_data->mFrameState = frame_state;
-		cur_timer_data->mChildTime = 0;
-#endif
-#if DEBUG_FAST_TIMER_THREADS
-#if !LL_RELEASE
-		assert_main_thread();
-#endif
-#endif
-	}
+	TimeBlock(const char* name, bool open = false, TimeBlock* parent = &getRootTimeBlock());
 
-	LL_FORCE_INLINE ~LLFastTimer()
-	{
-#if FAST_TIMER_ON
-		LLFastTimer::FrameState* frame_state = mFrameState;
-		U32 total_time = getCPUClockCount32() - mStartTime;
-
-		frame_state->mSelfTimeCounter += total_time - LLFastTimer::sCurTimerData.mChildTime;
-		frame_state->mActiveCount--;
+	TimeBlockTreeNode& getTreeNode() const;
+	TimeBlock* getParent() const { return getTreeNode().getParent(); }
+	void setParent(TimeBlock* parent) { getTreeNode().setParent(parent); }
 
-		// store last caller to bootstrap tree creation
-		// do this in the destructor in case of recursion to get topmost caller
-		frame_state->mLastCaller = mLastTimerData.mFrameState;
+	typedef std::vector<TimeBlock*>::iterator child_iter;
+	typedef std::vector<TimeBlock*>::const_iterator child_const_iter;
+	child_iter beginChildren();
+	child_iter endChildren();
+	std::vector<TimeBlock*>& getChildren();
 
-		// we are only tracking self time, so subtract our total time delta from parents
-		mLastTimerData.mChildTime += total_time;
+	void setCollapsed(bool collapsed)	{ mCollapsed = collapsed; }
+	bool getCollapsed() const			{ return mCollapsed; }
 
-		LLFastTimer::sCurTimerData = mLastTimerData;
-#endif
+	TraceType<TimeBlockAccumulator::CallCountAspect>& callCount() 
+	{ 
+		return static_cast<TraceType<TimeBlockAccumulator::CallCountAspect>&>(*(TraceType<TimeBlockAccumulator>*)this);
 	}
 
-public:
-	static LLMutex*			sLogLock;
-	static std::queue<LLSD> sLogQueue;
-	static BOOL				sLog;
-	static BOOL				sMetricLog;
-	static std::string		sLogName;
-	static bool 			sPauseHistory;
-	static bool 			sResetHistory;
+	TraceType<TimeBlockAccumulator::SelfTimeAspect>& selfTime() 
+	{ 
+		return static_cast<TraceType<TimeBlockAccumulator::SelfTimeAspect>&>(*(TraceType<TimeBlockAccumulator>*)this);
+	}
 
-	// call this once a frame to reset timers
-	static void nextFrame();
+	static TimeBlock& getRootTimeBlock();
+	static void pushLog(LLSD sd);
+	static void setLogLock(LLMutex* mutex);
+	static void writeLog(std::ostream& os);
 
 	// dumps current cumulative frame stats to log
 	// call nextFrame() to reset timers
 	static void dumpCurTimes();
 
-	// call this to reset timer hierarchy, averages, etc.
-	static void reset();
-
-	static U64 countsPerSecond();
-	static S32 getLastFrameIndex() { return sLastFrameIndex; }
-	static S32 getCurFrameIndex() { return sCurFrameIndex; }
-
-	static void writeLog(std::ostream& os);
-	static const NamedTimer* getTimerByName(const std::string& name);
-
-	struct CurTimerData
-	{
-		LLFastTimer*	mCurTimer;
-		FrameState*		mFrameState;
-		U32				mChildTime;
-	};
-	static CurTimerData		sCurTimerData;
-
-private:
-
-
 	//////////////////////////////////////////////////////////////////////////////
 	//
 	// Important note: These implementations must be FAST!
@@ -258,14 +142,14 @@ private:
 	//#undef _interlockedbittestandset
 	//#undef _interlockedbittestandreset
 
-	//inline U32 LLFastTimer::getCPUClockCount32()
+	//inline U32 TimeBlock::getCPUClockCount32()
 	//{
 	//	U64 time_stamp = __rdtsc();
 	//	return (U32)(time_stamp >> 8);
 	//}
 	//
 	//// return full timer value, *not* shifted by 8 bits
-	//inline U64 LLFastTimer::getCPUClockCount64()
+	//inline U64 TimeBlock::getCPUClockCount64()
 	//{
 	//	return __rdtsc();
 	//}
@@ -305,7 +189,7 @@ private:
 	}
 
 #else
-	//LL_COMMON_API U64 get_clock_count(); // in lltimer.cpp
+	//U64 get_clock_count(); // in lltimer.cpp
 	// These use QueryPerformanceCounter, which is arguably fine and also works on AMD architectures.
 	static U32 getCPUClockCount32()
 	{
@@ -372,18 +256,71 @@ private:
 
 #endif
 
-	static U64 sClockResolution;
+	static U64 countsPerSecond();
+
+	// updates cumulative times and hierarchy,
+	// can be called multiple times in a frame, at any point
+	static void processTimes();
 
-	static S32				sCurFrameIndex;
-	static S32				sLastFrameIndex;
-	static U64				sLastFrameTime;
+	// call this once a frame to periodically log timers
+	static void logStats();
 
-	U32							mStartTime;
-	LLFastTimer::FrameState*	mFrameState;
-	LLFastTimer::CurTimerData	mLastTimerData;
+	bool						mCollapsed;				// don't show children
 
+	// statics
+	static std::string							sLogName;
+	static bool									sMetricLog,
+												sLog;	
+	static U64									sClockResolution;
 };
 
-typedef class LLFastTimer LLFastTimer;
+LL_FORCE_INLINE BlockTimer::BlockTimer(TimeBlock& timer)
+{
+#if FAST_TIMER_ON
+	mStartTime = TimeBlock::getCPUClockCount64();
+
+	BlockTimerStackRecord* cur_timer_data = ThreadTimerStack::getIfExists();
+	TimeBlockAccumulator* accumulator = timer.getPrimaryAccumulator();
+	accumulator->mActiveCount++;
+	mBlockStartTotalTimeCounter = accumulator->mTotalTimeCounter;
+	// keep current parent as long as it is active when we are
+	accumulator->mMoveUpTree |= (accumulator->mParent->getPrimaryAccumulator()->mActiveCount == 0);
+
+	// store top of stack
+	mParentTimerData = *cur_timer_data;
+	// push new information
+	cur_timer_data->mActiveTimer = this;
+	cur_timer_data->mTimeBlock = &timer;
+	cur_timer_data->mChildTime = 0;
+#endif
+}
+
+LL_FORCE_INLINE BlockTimer::~BlockTimer()
+{
+#if FAST_TIMER_ON
+	U64 total_time = TimeBlock::getCPUClockCount64() - mStartTime;
+	BlockTimerStackRecord* cur_timer_data = ThreadTimerStack::getIfExists();
+	TimeBlockAccumulator* accumulator = cur_timer_data->mTimeBlock->getPrimaryAccumulator();
+
+	accumulator->mCalls++;
+	accumulator->mTotalTimeCounter += total_time - (accumulator->mTotalTimeCounter - mBlockStartTotalTimeCounter);
+	accumulator->mSelfTimeCounter += total_time - cur_timer_data->mChildTime;
+	accumulator->mActiveCount--;
+
+	// store last caller to bootstrap tree creation
+	// do this in the destructor in case of recursion to get topmost caller
+	accumulator->mLastCaller = mParentTimerData.mTimeBlock;
+
+	// we are only tracking self time, so subtract our total time delta from parents
+	mParentTimerData.mChildTime += total_time;
+
+	//pop stack
+	*cur_timer_data = mParentTimerData;
+#endif
+}
+
+}
+
+typedef LLTrace::BlockTimer LLFastTimer; 
 
 #endif // LL_LLFASTTIMER_H