On Mon, Jun 15, 2009 at 6:01 PM, Bill Nottingham<notting(a)redhat.com> wrote:
Gregory Maxwell (gmaxwell(a)gmail.com) said:
> I doubt having consistently lower FP precision is anything many users
> are asking for. The few that do can usually take care of themselves.
And yet you say we should push them all to x86_64, which has
the same lower precision?
You misunderstand me.
I don't think people care. Code that depends on the order of FPU
register spills on it is broken.
I'm saying it's not a feature. It's neutral. Or nearly so.
> > - More clearly delineates our support set of targets,
sticking true
> > to forwards innovation, not necessarily legacy support
>
> If thats the case why maintain x86 at all?
Because it's 58% of our userbase (source: F11 torrent stats.)
The relevant break down is what percentage is x86 no sse2 vs x86 no
x86_64 vs x86 w/ x86_64.
> Take your answer and now apply it to why fedora should maintain
> support for x86 CPUs without SSE2.
Because that's significantly less of our userbase. I'd love to have
harder numbers, but we're still talking about a set of CPUs that
(outside of corner cases like the Geode and C3) ceased production
anywhere from 4 (Athlon) to 6 (P3) to 10 (P2) years ago.
'outside'. Please don't just dismiss these recent systems, they are a
real issue.
I have low power x86 which I purchased *recently* (geode box for audio
conferencing, this year; and a via box from ~two years ago) without
SSE2. Most of my systems are x86_64, but what is x86 is not sse
capable. Anything of mine that needed performance became x86_64 a long
time ago.
On Mon, Jun 15, 2009 at 6:04 PM, Bill Nottingham<notting(a)redhat.com> wrote:
> It would probably be most interesting to perform that test on
the
> x86-only ATOM, since I can see CMOV being a bigger win on an in-order
> CPU.
> (I can't personally protest: I think all the x86 stuff I have has CMOV).
Last time I profiled it (using the bits of openbench that actually worked),
i586 -> i686 was an improvement of 1% (Core2Duo) to 2% (Atom). Athlon64 was
essentially equal.
SPECCPU results showed similar.
That sounds about right to me.