mtune=nocona

[Dan Williams]

On Thu, 2005-06-09 at 10:03 -0400, Horst von Brand wrote:
> Arjan van de Ven <arjanv at redhat.com> wrote:
> 
> > > Of course, on through -m64 was as expected.  I'm just a bit surprised
> > > that it didn't say mtune=k8 since it was definitely building on one ...
> > > which begs the question ... is GCC (and hence the entire x86_64
> > > distribution) being built this way?  What are the K8 performance
> > > implications?  I've seen a few things that indicate that Nocona cores
> > > running K8 code look pretty bad in some areas compared to optimized
> > > code...however, I find nothing to indicate how nocona optimized code
> > > runs on the K8.
> 
> > AMD cpus are very good at running just about anything you throw at it
> > (think about it; with their marketshare... most code out there is
> > optimized for intel cpus so their cpus are effectively required to deal
> > well with such code). The reverse unfortunatly isn't the case.
> 
> AFAIU, they are talking about /64 bit/ code here. There the situation is
> almost exactly the other way around, isn't it?

What Arjan is saying is this:  AMD Opteron/Athlon64 CPUs do much, much
better with -mtune=nocona than Intel EM64T CPUs do with -mtune=k8.  It
appears that the Opteron/Althon64 are more flexible in this regard.
Therefore, since the hit for -mtune=nocona on Opteron is much smaller
than -mtune=k8 on EM64T, the distro uses -mtune=nocona.

The question, distilled:

Would you like: (a) Small speed hit for Opteron users, OR (b) much
larger speed hit for EM64T users?

(a), ie -mtune=nocona everywhere, seems to be the fairest option here
for everybody since we obviously cannot make 1 build of Fedora Core for
Opteron and 1 comletely separate build of Fedora Core for EM64T.

Dan