This is the mail archive of the gcc@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Some (small) c++ compilation profiling data (oprofile)


John Levon wrote:-

> > > /* time count, time %age, mem ref count, mem reg %age */
> > > 
> > >     141 /* ht_lookup 149193 3.769% 27240 2.39% */
> > 
> > I looked on the oprofile site but couldn't see this format of
> > output, nor an explanation of what it meant.
> 
> Well I added the comment above, I suppose I wasn't clear enough. The
> first two numbers are the count and percentage of time spent , the
> second two are for the number/percentage of memory references. They are
> both statistical not exact.

Heh, can't believe I missed it.  Thanks.

> Like I briefly said, it's subject to some inherent unreliable ness. In
> fact this number probably represents that this time was spent at the
> /end/ of this for loop: by the time the profiler could enquire into the
> EIP, it had reached the start of the loop again (due to interrupt
> delivery latency). See the example here :
> http://oprofile.sourceforge.net/doc/interpreting.html

That worries me then.  If 1% of compile time is simply looping back
to the start of the for loop that's an awful lot of collisions.

Could you try -fmem-report?

Neil.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]