This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Compilation performance comparison of 3.5.0 and TreeSSA treeson MICO sources as requested in: [tree-ssa] Merge status 2004-05-03
- From: Andrew MacLeod <amacleod at redhat dot com>
- To: Jeff Law <law at redhat dot com>
- Cc: Diego Novillo <dnovillo at redhat dot com>, Jan Hubicka <hubicka at ucw dot cz>, KarelGardas <kgardas at objectsecurity dot com>, gcc mailing list <gcc at gcc dot gnu dot org>
- Date: 17 May 2004 15:33:19 -0400
- Subject: Re: Compilation performance comparison of 3.5.0 and TreeSSA treeson MICO sources as requested in: [tree-ssa] Merge status 2004-05-03
- References: <200405171910.i4HJAGXq008163@speedy.slc.redhat.com>
On Mon, 2004-05-17 at 15:10, law@redhat.com wrote:
> In message <1083800290.3350.6.camel@localhost.localdomain>, Diego Novillo write
> s:
> >On Wed, 2004-05-05 at 19:33, Jan Hubicka wrote:
> >
> >> You need ssa form form DCE. Going in/out just for DCE might be bit
> >> expensive...
> >>
> >Well, is it? You probably don't even need the full out-of-ssa pass. If
> >at -O0 we only schedule passes the will never create overlapping LRs,
> >you can just drop the SSA_NAMEs on the way out.
> Right. And DCE certainly doesn't create overlapping lifetimes.
>
>
> You know, we could actually make the "may create overlapping lifetime" a
> property of the optimization passes. That way we just check the property
> at out-of-ssa time, if the property is clear, then we just drop the SSA_NAMES.
>
That sounds like a very good idea. I would think that having
remove_ssa_form() check the property and decide whether to do what it
does now, or a small subset. You'd still want to perform TER on the way
out. You would also lose any live range splitting that we currently get
(but at -O0 that wouldnt matter).
Andrew