Re: [PATCH] profile feedback: -fprofile-use= and -fprofile-correction, correctness fixes and option semantic changes.

[Jan Hubicka <jh at suse dot cz>]

Hi,
note that if we make -fprofile-use= shortcut, we probably could have
also -fprofile-gen=
> (or maybe even submitted, don't remember now).
> But using atomic increment on highly multithreaded programs on high
> core count systems
> incur significant performance overhead.
> So I'd rather like to see TLS counters than atomic increments.
> Regardless, this option (-fprofile-correction) is completely
> orthogonal to either atomic increment or TLS.

Zdenek also had equivalent of TLS counters (before TLS architecture was
added to GCC).  The problem here that the counters can get pretty big,
so for instance for Mozilla, you can end up producing quite few
megabytes of extra data.

Lets see where the patches are and try to get some to mainline.
Profiling threaded applications is definitly important thing.
> 
> > > Making counters TLS adds more memory overhead
> >  > and more complicated interaction with threading system.
> >  > This ad-hoc correction algorithm is an interim solution
> >  > until we have more proper algorithm implemented
> >  > (based on variants of minimum cost network flow algorithms),
> >
> >  Do you have some reference here?
> 
> http://www.springerlink.com/content/j444437052584j16/

Interesting, thanks!
Honza
> 
> >  Thanks and sorry for the delay,
> >  Honza
> 
> Thanks for the review.
> 
> Seongbae