Optimal Optimizations: A Very Preliminary Report

[tm_gccmail@mail.kloo.net]

On Sun, 17 Aug 2003, Nathanael Nerode wrote:

> Michael Matz wrote:
> > Hi,
> > 
> > On Tue, 12 Aug 2003, Nathanael Nerode wrote:
> > 
> > 
> >>(Incidentally, the evidence for that particular case indicates that
> >>-fschedule-insns should be abandoned entirely, while -fschedule-insns2
> >>should be standard.)
> > 
> > 
> > For this particular case!  Scott is testing ooo architectures, which
> > generally do not benefit that greatly from scheduling anyway.  The picture
> > is wholly different with in-order archs where pre-regalloc scheduling can
> > make a horribly high difference (if it's given that there is a reasonable
> > number of registers, so that scheduling does not interfere with regalloc
> > that much).
> 
> Hmm.  Which of our arches are out-of-order and which of them are 
> in-order?  I believe that -fschedule-insns has been reported to make 
> things worse on sh, powerpc, x86, and ia64.  :-/

As I've mentioned before, -fschedule-insns with throttling results
in up to a 20% gain on the SH architecture.

> Is there a way to set the default optimizations differently on a 
> per-architecture basis?  If there isn't, should there be? ;-)

We already do this. For the H8/300 we disable bbro.

Toshi