This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: [tree-profiling] Fix return predictors
> On Wednesday 11 August 2004 15:29, Jan Hubicka wrote:
> > On tree profiloing branch we now get better SPEC scores, faster
> > compilation even with the loop disabled, so I will start merging
> > remaining neccesary infrastructure bits at the end of week and in the
> > case loop won't be replaced for 3.5, we can simply reconstruct the
> > profile estimate from REG_BR_PROB notes (I tried that on tree profiling
> > branch and it works just as it is written now).
>
> I think we're slow enough as it is, not another pass thank you ;-)
Hmm, we do all neccesary passes on mainline anyway (ie build profile
info early and then build it again after loop ;) The idea is to
gradually kill the second pass once the loop is dead.
>
> We already have to reconstruct loop notes for the old loop optimizer,
> and to hack around loop with profile information too, that's just
> stupid. IMHO it is an absolute _must_ that the old loop optimizer can
> be removed before GCC 3.5 is released.
Yep, that would be nice tought ;)
Honza
>
> The missing bits are, iiuc, strength reduction and prefetching. For
> the latter, the tree pass on the LNO branch is already superior.
> Strength reduction is a bit of an issue because it needs some hacks
> (ie. the REF_ORIGINAL bits) to get alias information right.
>
> Gr.
> Steven
>
>
>