On Sun, 2006-07-09 at 04:50 -0400, Dave Jones wrote:
On Sun, Jul 09, 2006 at 10:43:32AM +0200, Arjan van de Ven wrote:
> On Sat, 2006-07-08 at 16:38 -0400, Dave Jones wrote:
> > On Sat, Jul 08, 2006 at 10:27:55PM +0200, Andreas Thienemann wrote:
> >
> > > Several legacy i586 systems are used as routers and I'd hate to
leave them
> > > on FC5.
> >
> > The buildsys was tweaked a few days ago, and 586 kernels have been popping
out
> > since then. test2 should be fine.
>
> now to go a step further; cmov isn't actually a performance gain on any
> cpu sold in the last year or two. Why ship the i686 kernel at all over
> an i586 kernel? Why not ship ONLY a i586 kernel?
This has crossed my mind, and I did some investigation on this a month or two ago.
You'd still want gcc to schedule instructions optimally for newer cpus
(-mtune-generic seems to handle this),
yes absolutely; no doubt about that
but iirc adding -march=i586 somehow
affects the behaviour of -mtune in a negative way.
hmmm that's a good question for the gcc guys... I'm actually surprised
by this. Maybe -mtune=i686 ?