This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Optimal Optimizations: A Very Preliminary Report
- From: <tm_gccmail at mail dot kloo dot net>
- To: Nathanael Nerode <neroden at twcny dot rr dot com>
- Cc: gcc at gcc dot gnu dot org
- Date: Tue, 12 Aug 2003 14:06:19 -0700 (PDT)
- Subject: Re: Optimal Optimizations: A Very Preliminary Report
On Tue, 12 Aug 2003, Nathanael Nerode wrote:
> Toshi <tm_gccmail@mail.kloo.net> said:
> >Please don't suggest abandoning -fschedule-insns unless you understand
> >the full intricacies of the situation and have a viable alternative.
>
> I'm not suggesting ripping it out of the code base.
Okay, understood.
> However, it should *not* be activated by default at any optimization
> level (until it is fixed), since it consistently makes things worse on
> at least x86 and PowerPC. Currently it is activated by default at -O2,
> which is stupid given that it's known to make things worse most of the
> time.
We already noticed this on the SH, and so it's been disabled since 1996 or
so. I suspected it was the case for other targets as well.
> >The current problem with sched1 isn't the concept of scheduling before
> >register allocation; rather, it's allowing the scheduler unrestrained
> >freedom to move instructions before register allocation which can
> >raise register pressure unnecessarily. People are working on a solution
> >for this already.
>
> Yeah, but it doesn't look like the fix will be in for 3.4 or even 3.5.
> It should be turned off.
I won't disagree with that.
Toshi