TRACY_ON_DEMAND causes "Zone is ended twice" instrumentation failure

[davidvodnik]

Description

I keep getting Zone is ended twice failures when compiled with TRACY_ON_DEMAND. Without this flag it works fine.

I debugged tracy profiler and in the failing cases I am getting QueueType::ZoneEnd before any QueueType::ZoneBegin.

Workaround

I can patch void Worker::ProcessZoneEnd( const QueueZoneEnd& ev ) function to ignore the mismatched QueueType::ZoneEnd, and the issue goes away.

    if( td->zoneIdStack.empty() )
    {
        //ZoneDoubleEndFailure( td->id, td->timeline.empty() ? nullptr : td->timeline.back() );
        return;
    }

I also needed to ignore timeEnd smaller than zone->Start().

    //assert( timeEnd >= zone->Start() );
    if (timeEnd < zone->Start()) {
        return;
    }

Questions

• Where are mismatched QueueType::ZoneBegin and QueueType::ZoneEnd supposed to be handled?
• Is the client supposed to send a coherent stream?
• Or is the server supposed to handle that case?

Additional Information

• Tracy version: 0.11.0, also 0.10.0
• Platform: Windows, OSX

[Ggjorven]

I have had this exact issue for a while now. Would be nice if it got fixed.

[ccp-serpent]

I encountered this issue during very early profiler startup in my process. A workaround I found was to delay instrumentation until the profiler is connected. E.g. guarding against tracy::GetProfiler().IsConnected(). The telemetry context active state should be providing this exact guard, but that doesn't seem to cover all cases during early profiler startup.