summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
Diffstat
-rw-r--r--indra/llcommon/llfasttimer.cpp40
-rw-r--r--indra/llcommon/llthreadlocalstorage.h8
-rw-r--r--indra/llcommon/lltrace.h48
-rw-r--r--indra/llcommon/lltracerecording.cpp22
-rw-r--r--indra/llcommon/lltracerecording.h6
-rw-r--r--indra/llcommon/lltracethreadrecorder.h5
-rw-r--r--indra/newview/llappviewer.cpp1
7 files changed, 71 insertions, 59 deletions
diff --git a/indra/llcommon/llfasttimer.cpp b/indra/llcommon/llfasttimer.cpp
index 5dc5fdd5be..3fdd33959d 100644
--- a/indra/llcommon/llfasttimer.cpp
+++ b/indra/llcommon/llfasttimer.cpp
@@ -180,6 +180,7 @@ TimeBlockTreeNode& TimeBlock::getTreeNode() const
static LLFastTimer::DeclareTimer FTM_PROCESS_TIMES("Process FastTimer Times");
+// not thread safe, so only call on main thread
//static
void TimeBlock::processTimes()
{
@@ -195,8 +196,8 @@ void TimeBlock::processTimes()
TimeBlock& timer = *it;
if (&timer == &TimeBlock::getRootTimeBlock()) continue;
- // bootstrap tree construction by attaching to last timer to be on stack
- // when this timer was called
+ // bootstrap tree construction by attaching to last timer to be on stack
+ // when this timer was called
if (timer.getParent() == &TimeBlock::getRootTimeBlock())
{
TimeBlockAccumulator* accumulator = timer.getPrimaryAccumulator();
@@ -233,30 +234,30 @@ void TimeBlock::processTimes()
TimeBlockAccumulator* accumulator = timerp->getPrimaryAccumulator();
if (accumulator->mMoveUpTree)
- {
+ {
// since ancestors have already been visited, re-parenting won't affect tree traversal
- //step up tree, bringing our descendants with us
- LL_DEBUGS("FastTimers") << "Moving " << timerp->getName() << " from child of " << timerp->getParent()->getName() <<
- " to child of " << timerp->getParent()->getParent()->getName() << LL_ENDL;
- timerp->setParent(timerp->getParent()->getParent());
- accumulator->mParent = timerp->getParent();
- accumulator->mMoveUpTree = false;
-
- // don't bubble up any ancestors until descendants are done bubbling up
- // as ancestors may call this timer only on certain paths, so we want to resolve
- // child-most block locations before their parents
- it.skipAncestors();
+ //step up tree, bringing our descendants with us
+ LL_DEBUGS("FastTimers") << "Moving " << timerp->getName() << " from child of " << timerp->getParent()->getName() <<
+ " to child of " << timerp->getParent()->getParent()->getName() << LL_ENDL;
+ timerp->setParent(timerp->getParent()->getParent());
+ accumulator->mParent = timerp->getParent();
+ accumulator->mMoveUpTree = false;
+
+ // don't bubble up any ancestors until descendants are done bubbling up
+ // as ancestors may call this timer only on certain paths, so we want to resolve
+ // child-most block locations before their parents
+ it.skipAncestors();
+ }
}
}
-}
// walk up stack of active timers and accumulate current time while leaving timing structures active
BlockTimerStackRecord* stack_record = ThreadTimerStack::getInstance();
BlockTimer* cur_timer = stack_record->mActiveTimer;
TimeBlockAccumulator* accumulator = stack_record->mTimeBlock->getPrimaryAccumulator();
- // root defined by parent pointing to self
- while(cur_timer && cur_timer->mParentTimerData.mActiveTimer != cur_timer)
+ while(cur_timer
+ && cur_timer->mParentTimerData.mActiveTimer != cur_timer) // root defined by parent pointing to self
{
U64 cumulative_time_delta = cur_time - cur_timer->mStartTime;
accumulator->mTotalTimeCounter += cumulative_time_delta - (accumulator->mTotalTimeCounter - cur_timer->mBlockStartTotalTimeCounter);
@@ -413,8 +414,11 @@ TimeBlockAccumulator::TimeBlockAccumulator()
mParent(NULL)
{}
-void TimeBlockAccumulator::addSamples( const TimeBlockAccumulator& other )
+void TimeBlockAccumulator::addSamples( const TimeBlockAccumulator& other, bool append )
{
+ // we can't merge two unrelated time block samples, as that will screw with the nested timings
+ // due to the call hierarchy of each thread
+ llassert(append);
mTotalTimeCounter += other.mTotalTimeCounter - other.mStartTotalTimeCounter;
mSelfTimeCounter += other.mSelfTimeCounter;
mCalls += other.mCalls;
diff --git a/indra/llcommon/llthreadlocalstorage.h b/indra/llcommon/llthreadlocalstorage.h
index 471784749b..cc67248124 100644
--- a/indra/llcommon/llthreadlocalstorage.h
+++ b/indra/llcommon/llthreadlocalstorage.h
@@ -313,11 +313,6 @@ template<typename DERIVED_TYPE>
class LLThreadLocalSingletonPointer
{
public:
- void operator =(DERIVED_TYPE* value)
- {
- setInstance(value);
- }
-
LL_FORCE_INLINE static DERIVED_TYPE* getInstance()
{
#if LL_DARWIN
@@ -328,7 +323,7 @@ public:
#endif
}
- LL_FORCE_INLINE static void setInstance(DERIVED_TYPE* instance)
+ static void setInstance(DERIVED_TYPE* instance)
{
#if LL_DARWIN
createTLSKey();
@@ -339,6 +334,7 @@ public:
}
private:
+
#if LL_WINDOWS
static __declspec(thread) DERIVED_TYPE* sInstance;
#elif LL_LINUX
diff --git a/indra/llcommon/lltrace.h b/indra/llcommon/lltrace.h
index e950a119d3..6dfe9e4b4e 100644
--- a/indra/llcommon/lltrace.h
+++ b/indra/llcommon/lltrace.h
@@ -35,6 +35,7 @@
#include "llunit.h"
#include "llapr.h"
#include "llthreadlocalstorage.h"
+#include "lltimer.h"
#include <list>
@@ -75,7 +76,6 @@ void set_thread_recorder(class ThreadRecorder*);
class MasterThreadRecorder& getUIThreadRecorder();
-// one per thread per type
template<typename ACCUMULATOR>
class AccumulatorBuffer : public LLRefCount
{
@@ -104,9 +104,9 @@ public:
~AccumulatorBuffer()
{
- if (LLThreadLocalSingletonPointer<ACCUMULATOR>::getInstance() == mStorage)
+ if (isPrimary())
{
- LLThreadLocalSingletonPointer<ACCUMULATOR>::setInstance(getDefaultBuffer()->mStorage);
+ LLThreadLocalSingletonPointer<ACCUMULATOR>::setInstance(NULL);
}
delete[] mStorage;
}
@@ -121,12 +121,12 @@ public:
return mStorage[index];
}
- void addSamples(const AccumulatorBuffer<ACCUMULATOR>& other)
+ void addSamples(const AccumulatorBuffer<ACCUMULATOR>& other, bool append = true)
{
llassert(mStorageSize >= sNextStorageSlot && other.mStorageSize > sNextStorageSlot);
for (size_t i = 0; i < sNextStorageSlot; i++)
{
- mStorage[i].addSamples(other.mStorage[i]);
+ mStorage[i].addSamples(other.mStorage[i], append);
}
}
@@ -169,7 +169,8 @@ public:
LL_FORCE_INLINE static ACCUMULATOR* getPrimaryStorage()
{
- return LLThreadLocalSingletonPointer<ACCUMULATOR>::getInstance();
+ ACCUMULATOR* accumulator = LLThreadLocalSingletonPointer<ACCUMULATOR>::getInstance();
+ return accumulator ? accumulator : sDefaultBuffer->mStorage;
}
// NOTE: this is not thread-safe. We assume that slots are reserved in the main thread before any child threads are spawned
@@ -222,25 +223,27 @@ public:
static self_t* getDefaultBuffer()
{
- // this buffer is allowed to leak so that trace calls from global destructors have somewhere to put their data
- // so as not to trigger an access violation
- static self_t* sBuffer = new AccumulatorBuffer(StaticAllocationMarker());
static bool sInitialized = false;
if (!sInitialized)
{
- sBuffer->resize(DEFAULT_ACCUMULATOR_BUFFER_SIZE);
+ // this buffer is allowed to leak so that trace calls from global destructors have somewhere to put their data
+ // so as not to trigger an access violation
+ sDefaultBuffer = new AccumulatorBuffer(StaticAllocationMarker());
sInitialized = true;
+ sDefaultBuffer->resize(DEFAULT_ACCUMULATOR_BUFFER_SIZE);
}
- return sBuffer;
+ return sDefaultBuffer;
}
private:
ACCUMULATOR* mStorage;
size_t mStorageSize;
static size_t sNextStorageSlot;
+ static self_t* sDefaultBuffer;
};
template<typename ACCUMULATOR> size_t AccumulatorBuffer<ACCUMULATOR>::sNextStorageSlot = 0;
+template<typename ACCUMULATOR> AccumulatorBuffer<ACCUMULATOR>* AccumulatorBuffer<ACCUMULATOR>::sDefaultBuffer = NULL;
template<typename ACCUMULATOR>
class TraceType
@@ -307,7 +310,7 @@ public:
mLastValue = value;
}
- void addSamples(const self_t& other)
+ void addSamples(const self_t& other, bool append)
{
if (other.mNumSamples)
{
@@ -347,7 +350,7 @@ public:
F64 weight = (F64)mNumSamples / (F64)(mNumSamples + other.mNumSamples);
mNumSamples += other.mNumSamples;
mMean = mMean * weight + other.mMean * (1.f - weight);
- mLastValue = other.mLastValue;
+ if (append) mLastValue = other.mLastValue;
}
}
@@ -431,7 +434,7 @@ public:
mHasValue = true;
}
- void addSamples(const self_t& other)
+ void addSamples(const self_t& other, bool append)
{
if (other.mTotalSamplingTime)
{
@@ -473,9 +476,12 @@ public:
mNumSamples += other.mNumSamples;
mTotalSamplingTime += other.mTotalSamplingTime;
mMean = (mMean * weight) + (other.mMean * (1.0 - weight));
- mLastValue = other.mLastValue;
- mLastSampleTimeStamp = other.mLastSampleTimeStamp;
- mHasValue |= other.mHasValue;
+ if (append)
+ {
+ mLastValue = other.mLastValue;
+ mLastSampleTimeStamp = other.mLastSampleTimeStamp;
+ mHasValue |= other.mHasValue;
+ }
}
}
@@ -548,7 +554,7 @@ public:
mSum += value;
}
- void addSamples(const CountAccumulator<T>& other)
+ void addSamples(const CountAccumulator<T>& other, bool /*append*/)
{
mSum += other.mSum;
mNumSamples += other.mNumSamples;
@@ -593,7 +599,7 @@ public:
};
TimeBlockAccumulator();
- void addSamples(const self_t& other);
+ void addSamples(const self_t& other, bool /*append*/);
void reset(const self_t* other);
void flush() {}
@@ -713,6 +719,8 @@ void add(CountStatHandle<T>& count, VALUE_T value)
struct MemStatAccumulator
{
+ typedef MemStatAccumulator self_t;
+
MemStatAccumulator()
: mSize(0),
mChildSize(0),
@@ -720,7 +728,7 @@ struct MemStatAccumulator
mDeallocatedCount(0)
{}
- void addSamples(const MemStatAccumulator& other)
+ void addSamples(const MemStatAccumulator& other, bool /*append*/)
{
mSize += other.mSize;
mChildSize += other.mChildSize;
diff --git a/indra/llcommon/lltracerecording.cpp b/indra/llcommon/lltracerecording.cpp
index 5b0b74524f..3994e4f521 100644
--- a/indra/llcommon/lltracerecording.cpp
+++ b/indra/llcommon/lltracerecording.cpp
@@ -97,13 +97,15 @@ void RecordingBuffers::append( const RecordingBuffers& other )
void RecordingBuffers::merge( const RecordingBuffers& other)
{
- mCountsFloat.addSamples(other.mCountsFloat);
- mCounts.addSamples(other.mCounts);
- mSamplesFloat.addSamples(other.mSamplesFloat);
- mSamples.addSamples(other.mSamples);
- mEventsFloat.addSamples(other.mEventsFloat);
- mEvents.addSamples(other.mEvents);
- mMemStats.addSamples(other.mMemStats);
+ mCountsFloat.addSamples(other.mCountsFloat, false);
+ mCounts.addSamples(other.mCounts, false);
+ mSamplesFloat.addSamples(other.mSamplesFloat, false);
+ mSamples.addSamples(other.mSamples, false);
+ mEventsFloat.addSamples(other.mEventsFloat, false);
+ mEvents.addSamples(other.mEvents, false);
+ mMemStats.addSamples(other.mMemStats, false);
+ // for now, hold out timers from merge, need to be displayed per thread
+ //mStackTimers.addSamples(other.mStackTimers, false);
}
void RecordingBuffers::reset(RecordingBuffers* other)
@@ -190,7 +192,6 @@ void Recording::handleStop()
{
mElapsedSeconds += mSamplingTimer.getElapsedTimeF64();
mBuffers.write()->flush();
- LLTrace::TimeBlock::processTimes();
LLTrace::get_thread_recorder()->deactivate(this);
}
@@ -490,8 +491,9 @@ void PeriodicRecording::appendPeriodicRecording( PeriodicRecording& other )
{
if (mAutoResize)
{
- for (S32 other_index = (other.mCurPeriod + 2) % other_recording_count;
- other_index != other.mCurPeriod;
+ for (S32 other_index = (other.mCurPeriod + 2) % other_recording_count,
+ end_index = (other.mCurPeriod + 1) % other_recording_count;
+ other_index != end_index;
other_index = (other_index + 1) % other_recording_count)
{
llassert(other.mRecordingPeriods[other_index].getDuration() != 0.f
diff --git a/indra/llcommon/lltracerecording.h b/indra/llcommon/lltracerecording.h
index 19a4fae737..aaeb32e891 100644
--- a/indra/llcommon/lltracerecording.h
+++ b/indra/llcommon/lltracerecording.h
@@ -336,12 +336,12 @@ namespace LLTrace
}
template <typename T>
- typename T getPeriodMin(const TraceType<SampleAccumulator<T> >& stat, size_t num_periods = U32_MAX) const
+ T getPeriodMin(const TraceType<SampleAccumulator<T> >& stat, size_t num_periods = U32_MAX) const
{
size_t total_periods = mRecordingPeriods.size();
num_periods = llmin(num_periods, total_periods);
- typename T min_val = std::numeric_limits<T>::max();
+ T min_val = std::numeric_limits<T>::max();
for (S32 i = 1; i <= num_periods; i++)
{
S32 index = (mCurPeriod + total_periods - i) % total_periods;
@@ -351,7 +351,7 @@ namespace LLTrace
}
template <typename T>
- typename T getPeriodMin(const TraceType<EventAccumulator<T> >& stat, size_t num_periods = U32_MAX) const
+ T getPeriodMin(const TraceType<EventAccumulator<T> >& stat, size_t num_periods = U32_MAX) const
{
size_t total_periods = mRecordingPeriods.size();
num_periods = llmin(num_periods, total_periods);
diff --git a/indra/llcommon/lltracethreadrecorder.h b/indra/llcommon/lltracethreadrecorder.h
index 17a2d4a9a9..bf3701304f 100644
--- a/indra/llcommon/lltracethreadrecorder.h
+++ b/indra/llcommon/lltracethreadrecorder.h
@@ -87,12 +87,13 @@ namespace LLTrace
LLMutex* getSlaveListMutex() { return &mSlaveListMutex; }
+
private:
typedef std::list<class SlaveThreadRecorder*> slave_thread_recorder_list_t;
- slave_thread_recorder_list_t mSlaveThreadRecorders;
- LLMutex mSlaveListMutex;
+ slave_thread_recorder_list_t mSlaveThreadRecorders; // list of slave thread recorders associated with this master
+ LLMutex mSlaveListMutex; // protects access to slave list
};
class LL_COMMON_API SlaveThreadRecorder : public ThreadRecorder
diff --git a/indra/newview/llappviewer.cpp b/indra/newview/llappviewer.cpp
index a049aabe74..ff481d6278 100644
--- a/indra/newview/llappviewer.cpp
+++ b/indra/newview/llappviewer.cpp
@@ -1241,6 +1241,7 @@ bool LLAppViewer::mainLoop()
while (!LLApp::isExiting())
{
LLFastTimer _(FTM_FRAME);
+ LLTrace::TimeBlock::processTimes();
LLTrace::get_frame_recording().nextPeriod();
LLTrace::TimeBlock::logStats();
generated by cgit v1.2.3 (git 2.25.1) at 2025-05-30 16:31:49 +0000