Fix profile updating after complette loop unrolling
Zdenek Dvorak
rakdver@atrey.karlin.mff.cuni.cz
Wed Aug 3 09:25:00 GMT 2005
Hello,
> other common source of weird profiles is complette loop unrolling. At the
> moemnt updating is done via generic loop peeling mechanizm, that expect that
> last copy of loop to still loop at the average same number of times as did the
> original loop causing the other more frequent copies relatively cold
> causing some relatively minnor regressions in gzip.
>
> This patch don't get the updating 100% right either whent the loop has nonzero
> probability of exiting by other edges too, but I quite convinced myself that
> there is no way to do much saner job (ie one might try to guess how histogram
> of number of iterations of the loop but without that it is not quite possibly
> to get counts of loop exit edges right)
>
> Zdenek, does this code seems to make sense to you?
seems OK to me.
Zdenek
More information about the Gcc-patches
mailing list