On Fri, 2009-01-30 at 19:47 -0500, Gregory Maxwell wrote: > On Fri, Jan 30, 2009 at 6:18 PM, Jan Kratochvil > <jan.kratochvil@xxxxxxxxxx> wrote: > [snip] > > Guessing some software (codecs?) may fallback on 64-bit to unoptimized C code > > instead of some assembly one but Fedora would at least show the usefulness of > > the updates of such software for the x86_64 arch. > > Try GLIBC: For example, x86_64 exp() (and expf()) is *much* slower > than x86. (At least as of my testing in Fedora 9, I haven't measured > F10… the point being that these differences can lurk in unexpected > places. Last I checked, gcc was unable to translate libm calls into nice inline instructions on x86_64 like it can on i386, as it's apparently only implemented for x87 and not SSE. x86-64 ends up with function calls inside inner loops instead of an instruction or two. I've done a lot of eyeballing of gcc's asm output... > They need to be fixed in anycase, but useful to keep in mind. Yes, if someone could get gcc's SSE FP optimization on par with x87, it would probably speed up quite a few inner loops... :)
Attachment:
signature.asc
Description: This is a digitally signed message part
-- fedora-devel-list mailing list fedora-devel-list@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/fedora-devel-list
