This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [patch] for PR 18316 (updated)
Hello,
> > > > 171.swim 3100 653 475 * 3100 710 437 *
> > > > 172.mgrid 1800 508 355 * 1800 709 254 *
> > >
> > > Will TARGET_MEM_REF fix this regression? It's not that we haven't seen
> > > mgrid jump up and down in scores... it seems very sensitive to IV choices.
> >
> > I have no idea (I haven't checked what's going on there yet), but I
> > doubt that.
>
> It'd be good to know whether this 30% degradation in mgrid is consistent
> across platforms.
> If you don't have access to another platform, i'm happy to SPEC test
> this patch on ppc for you.
I am quite sure now that the regression would show there also. What
happens is that we strength reduce and create a new iv for every single
memory reference in one of the loops (ending up with some 40 induction
variables). It seems hard to do anything sane with ivs in the loops
that occur in mgrid; I am now trying to persuade the patch to just do nothing on
this type of weird testcases, which seems to be the best choice to me.
Zdenek