Re: [ast-optimizer-branch] Simplification is pretty expensive

[Jan Hubicka <jh at suse dot cz>]

> On Tue, 2002-06-04 at 13:35, Jan Hubicka wrote:
> > > On Tue, 2002-06-04 at 13:22, Jan Hubicka wrote:
> > > 
> > > > I tought it just enabled the simplification pass?
> > > > 
> > > Yes, but simplification was not really enabled until we fixed the
> > > bootstrap problems.
> > Even more interesting...
> > Why was the builds failing with -ftree-ssa then?
> > 
> Because simplification had bugs, obviously.  But I was a bit misleading
> in my previous post.
> 
> Prior to last week, much of the simplifier was ifdef'd out in the
> branch.  So, -ftree-ssa would have some effect, but it would only
> re-write some types of trees.
> 
> Since it wasn't re-writing everything, the performance degradation
> wasn't very noticeable.  Once we fixed all the bugs and fully enabled
> simplification the degradation became very noticeable.

Thanks, now I understand better.
> 
> In any case, I wouldn't be surprised if tree simplification has some
> negative effect on SPEC scores (remember that none of the tree
> optimizers are enabled now).  The hope is of course to start climbing up
> the performance ladder with the new tree optimizers.  But we must pay
> attention to the compile time degradation.  It's pretty significant.

Agreed.  Looking at the results again, the overall scores are wrong,
since the perlbmk build has been still failing last time, but still
there seems to be more than 5% degradations in some tests that looks
serious enought to take a look at.

I have to take a closer look at ast code to have some fun :)
Hope today new results will appear on the tester page.

Honza