Re: Fix IMA ICE in get_tmt_for really (PR 16954)

[Andrew Pinski <pinskia at physics dot uc dot edu>]

On Aug 22, 2004, at 10:49 AM, Dale Johannesen wrote:

> On Aug 12, 2004, at 8:54 AM, Diego Novillo wrote:
> > > > Sounds good.  Remember to change Andrew's patch to only reject
> > > > coalescing pointers.
> > >
> > > OK, I put in code to report mismatched alias sets (but not disable 
> > > the
> > > renaming)
> > > with these results (IMA -O3...)
> > >
> > > vortex 25
> > > mesa 21  (but exited at the first error)
> > > eon 5913
> > OK, but what I really wanted to know is whether it affects performance
> > at all, and why.
>
> I finally got enough free cycles to try this in a controlled 
> environment.
>
> There is a 1.5% drop on eon by applying Andrew's (second) patch.
> No difference on anything else.  (So the data above is actually
> a pretty good predictor.)


I think this is caused by the fact that reference types and
pointers types were given two different aliasing sets which is
wrong. See patch at
<http://gcc.gnu.org/ml/gcc-patches/2004-09/msg01141.html> which
fixes this issue and might also fix the 1.5% drop in eon with the
patch.

Thanks,
Andrew Pinski