On Mon, Jun 15, 2009 at 06:04:15PM -0400, Bill Nottingham wrote:
Gregory Maxwell (gmaxwell(a)gmail.com) said:
> Is dropping hardware worth this instruction? I suggested some
> benchmarking and demonstrated that it's a very small improvement for
> Theora. I'm not sure if anyone got around to testing freetype,
> firefox, or any of the other couple CPU heavy desktop apps.
>
> 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.
Wasn't code paths in glibc and possibly other packages a concern? I seem to
remember the glibc maintainer not being happy with using i586 at all, but
conceded to it as a path to i686.
josh