This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: [RFC] Enabling loop unrolls at -O3?
On 11/5/05, Giovanni Bajo <rasky@develer.com> wrote:
> Richard Guenther <richard.guenther@gmail.com> wrote:
>
> >>>> I guess the issue is what does "huge" mean, it is hard to discuss based
> >>>> on loaded adjectives taking the place of data :-)
> >>>
> >>> Huge here means 15-20% on x86* hosts.
> >>
> >> I don't consider this huge for -O3. I think -O3 can be slower if it
> >> achieves better code, and -funroll-loops makes it do just that.
> >
> > First of all, this is too late for 4.1, second, we enable loop-unrolling
> > with FDO. And without FDO, -funroll-loops is not necessarily an
> > improvement.
>
> What is FDO?
Feedback Directed Optimization, i.e. -fprofile-generate/-fprofile-use.
> My feeling is that -funroll-loops is almost always an improvement. It might
> be false sometimes, but this can be said of many optimization passes.
Can you back up this feeling with numbers? SPEC -O2 vs. -O2 -funroll-loops
and/or -O3 vs. -O3 -funroll-loops would do it.
Richard.