Re: [PATCH] Fix RTL DSE compile time hog (PR rtl-optimization/48141)

[Paolo Bonzini <bonzini at gnu dot org>]

On 03/16/2011 12:59 PM, Richard Guenther wrote:
> On Wed, Mar 16, 2011 at 12:52 PM, Jakub Jelinek<jakub@redhat.com>  wrote:
> > On Wed, Mar 16, 2011 at 09:46:20AM +0100, Paolo Bonzini wrote:
> > > On 03/16/2011 12:12 AM, Kenneth Zadeck wrote:
> > > > so how much time does this save?
> > > >
> > > > I agree that this is a useful simplification, but it seems unlikely to
> > > > be that important in real code.
> > > > it seems like the 5000 store test would in general provide a better
> > > > safety valve.
> > >
> > > I think having both is a good idea.
> >
> > Here is the second patch, ok for both to trunk if this one passes
> > bootstrap/regtest?
> >
> > With just this patch alone on the testcase with the default
> > --param=max-dse-active-local-stores=5000 cc1 spends 18.9 seconds
> > in DSE1+DSE2, with =1000 just 4 seconds and with =10 0.4 seconds.
> > With the earlier patch in addition to this one the time in DSE1+DSE2
> > is 0.3 seconds no matter what the parameter is.
>
> That suggests we can avoid the new param?  (btw, we should start
> thinking of emitting a diagnostic if we run into such artificial limits)

Jakub's behavior is expected, since the number of active local stores 
should be 1 with the earlier patch in that testcase.  However, similar 
artificial testcases can probably be constructed that still trigger the 
ill behavior.

But I agree that the first patch is enough for now, and in general for 
release branches, especially without a testcase for which DSE is really 
the top hog.

Thanks,

Paolo