Re: [tree-ssa] Speeding up the dominator optimizer

[Daniel Berlin <dberlin at dberlin dot org>]

On Sun, 21 Sep 2003, Andrew MacLeod wrote:

> On Sun, 2003-09-21 at 16:25, Daniel Berlin wrote:
> >
> >
> > On Sun, 21 Sep 2003 law@redhat.com wrote:
> >
> > >
> > >
> > > I should not that for the one test in question (java/interpret.cc) the
> > > dominator optimizer went from consuming from 50 seconds to ~5 with that patch.
> > > It ought to get the dominator optimizer off the radar for a while (with PRE
> > > and CCP being the biggest hogs for java/interpret.cc).
> >
> > Is this with or without checking?
> >
> The following is with checking disabled. A preprocessed
> libjava/interpret.cc give me the following results on x86.
>
> PRE is #1 now, and CCP is #2. I have more CCP improvements in mind. I
> haven't really lookeed at PRE yet.

I'm about to patch PRE to deal with at least one of the issues.
These are the mainly O(n^2) traversals that I pointed out, which are
done to put the expression references in pre-order, dominator tree order.
Technically, we could also qsort the varray to do this, with the right
comparison function, and it might be even faster than what i'm doing now.

This shaves a bit of time off PRE.

I'm also working on the memory usage issues, too. Open64 caches the
expressions it generates for a given phi pred when doing expression ssa renaming. We
don't, so we generate a ton more garbage (at least, this is the cause
of the memory usage spike, AFAICT).

 --Dan