This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: Final intermodule patch
> Jan Hubicka <jh@suse.cz> writes:
>
> > > Andreas Jaeger <aj@suse.de> writes:
> > >
> > > > Mike Stump <mrs@apple.com> writes:
> > > >
> > > > > On Friday, July 11, 2003, at 03:36 AM, Andi Kleen wrote:
> > > > >> Can you expand a bit on how usable this is yet?
> > > > >
> > > > > Should be fairly usable, though, it hasn't seen tons of use. It does
> > > > > all of SPEC (well, except eon) for example, just to give you a flavor
> > > > > of what has already been though it.
> > > >
> > > > Do you have any SPEC numbers that you can share? How much improvement
> > > > does it bring?
> > >
> > > If you just use this patch, you get a decrease in SPEC performance,
> > > because GCC's current inlining heuristics don't work very well in that
> > > situation. With improved heuristics (Dale's and Stuart's work), we
> > > got significant improvement instead of a decrease.
> >
> > What kind of improvements have you implemented? How does it relate to
> > the new inlining heuristics code I commited last week?
>
> The most significant part is feedback-directed inlining, which is a
I see, feedback based inlining ineed must be easy winner. Did you
implemneted some other improvements?
> big win, as you might expect since it suppresses the existing
> heuristics and uses actual feedback data. Unfortunately, as I
> understand it, your new code makes our changes need a rewrite, so it
> might be some time before we can get them into FSF GCC. There are
Yes, it is shame that we didn't syncrhonized better. It should be easy
as the API on cgraph side is pretty straigh forward.
> also some unsolved issues, like how to do feedback-directed inlining
> at the same time as effective feedback-directed branch prediction.
This is ineed important issue. The only way I see is to make inlining
CFG aware on AST branch, manage CFG to survive during RTL lowering pass
and load branch prediction data before inlining. THis is not impossible
to implement, but bit dificult and I am trying to move forward in this
direction.
Honza
>
> --
> - Geoffrey Keating <geoffk@geoffk.org>