This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: Quadratic bottleneck in SSA propagation engine
- From: Jan Hubicka <jh at suse dot cz>
- To: Diego Novillo <dnovillo at redhat dot com>
- Cc: Jan Hubicka <jh at suse dot cz>, gcc-patches at gcc dot gnu dot org, Ken Zadeck <zadeck at naturalbridge dot com>
- Date: Fri, 22 Sep 2006 00:15:51 +0200
- Subject: Re: Quadratic bottleneck in SSA propagation engine
- References: <20060915084653.GB27957@kam.mff.cuni.cz> <4512B132.8000208@redhat.com>
> Yes, please. Let's wait for 4.3, unless you can prove good speedups
> across the board.
OK, I will get some numbers tomorrow.
> simulate_block will already ignore blocks that have already been
> executed. I'm not sure why you need this.
I see, I missed that bit - the problem is that simulate_block uses it's
own queue meaning that the PHI bypass my expensive/nonexpensive stmt
list trick. (other alternative would be to have expensive/nonexpensive
BB queue as well).
>
> Spelling nit: s/incomming/incoming/
>
> I'm interested in the VRP patch. This propagator patch will paper over
> the real problem. It's useful, though.
I will clean it up and send too (I got part of anti-intervals wrong
causing some regresisons). It is not solving the whole problem - it
saves about 20% overhead (if my memory is correct) but the quadratic
behaviour of ssa_propagate keeps VRP very top in profile, so both
patches are needed.
Honza