Re: [PATCH] Fix PR23744, VRP not merging discontinuous ranges of PHIs

["Richard Guenther" <richard dot guenther at gmail dot com>]

On 5/16/06, Duncan Sands <baldrick@free.fr> wrote:
> On Monday 15 May 2006 18:16, Diego Novillo wrote:
> > Duncan Sands wrote on 05/05/06 05:10:
> >
> > >     PR tree-optimization/23744
> > >     * tree-vrp.c (vrp_meet): do not require ranges to intersect.
> > >
> > This will be more appropriate for the next stage 1.  In the meantime, I
> > would like to see compile time impact analysis on real code (I can help
> > out with the usual suspects I have around: cc1-i-files, DLV, MICO, SPEC,
> > etc).
> >
> > You are lengthening the simulation by slowing down the decay into
> > VR_VARYING.
>
> You are correct, this may be a problem.  I will try to do some benchmarking,
> but it may take a few days.  In the meantime, Richard Guenther has kindly
> agreed run the patch through his C++ tester.

The results show compile time changes in the noise and at worst a slight <0.1%
memory usage increase for tramp3d.  Runtimes for libstdc++ performance testsuite
improves, other tests are in the noise for runtime effects.  Polyhedron showed
no effect on compile time or runtime.

Richard.