Re: [PATCH][libgcc-math] Vectorized intrinsics for x86_64

[Richard Guenther <rguenther at suse dot de>]

On Sun, 9 Apr 2006, H. J. Lu wrote:

> On Sun, Apr 09, 2006 at 03:37:03PM +0100, Andrew Haley wrote:
> > Steven Bosscher writes:
> >  > On Friday 07 April 2006 15:47, H. J. Lu wrote:
> >  > > The code is specific to Opteron. Shouldn't it be neutral to Opteron
> >  > > and Nocona?
> >  > 
> >  > The assembly code was "contributed by AMD" (see the original posting
> >  > in http://gcc.gnu.org/ml/gcc-patches/2006-03/msg01611.html).  Shouldn't
> >  > Intel first make sure that ICC works as well for Opteron as it does for
> >  > Nocona before Intel asks from AMD to write Nocona-optimized code? :-P
> > 
> > This is a bizarre argument.
> > 
> > Does the code work?  Yes.
> > Is it an improvement over current gcc?  Yes.
> > So check it in.
> > 
> > If Intel wants to contribute patches to make this code better on
> > Nocona, fine.  But it isn't AMD's problem.
> 
> FWIW, the current change is tuned for Opteron. Any future changes for
> Nocona may have negative impact on Opteron. Will that be a problem
> when Intel submits such a patch?

Now, technically that will be a regression for Opteron :P  But on a side
note we are currently negotiating what a reasonable change might look 
like.  Of course you are encouraged to share your superior knowledge about 
the Nocona architecture by providing a patch on top of the submission 
showing what you would think would be a good compromise for the two 
different architectures.  Instead of just talking.

Thanks,
Richard.