Re: gcc compile-time performance take 2

[Jan Hubicka <jh at suse dot cz>]

> On Mon, 20 May 2002, Jan Hubicka wrote:
> 
> > Looks definitly very interesting.  Do you think you can also
> > modify your scripts to oprofile cc1 binarry on the SPEC build?
> > Then we will have logged when particular function became
> > expensive.  You are using PentiumPro class machine that has
> > better counters than Athlons...
> >
> It depends on oprofile's overhead.  The SPEC95 runs are already
> taking about 9-10 hours.  I don't have too many cycles to spare.
> We can always save the compressed cc1 binaries and re-run with
> oprofile when we notice a slowdown.

oprofile has de-facto no overhead (ie 3-5%) to my experience, unless
the values are badly tunned...
> 
> > The graph looks interesting, definitly.  I am surprised it
> > grows so consistently over the development cycle (ie even in
> > the freezes). Perhaps it can be explained by filesystem
> > overhead or something...
> > 
> But the compiler is built on the same filesystem, it should be
> affecting bottstrap times the same way.  In any case, I'll be
> keeping better track of these numbers from now on.  The SPEC95
> runs should resume in a few more days.

Thanks!

Honza
> 
> Diego.