Re: tree-ssa memory and compile-time regressions

[Daniel Berlin <dberlin at dberlin dot org>]

On Jan 8, 2004, at 11:13 AM, Richard Guenther wrote:

> On Thu, 8 Jan 2004, Daniel Berlin wrote:
>
> > On Jan 8, 2004, at 10:19 AM, Richard Guenther wrote:
> >
> > > On Thu, 8 Jan 2004, Diego Novillo wrote:
> > >
> > > > On Thu, 2004-01-08 at 09:53, Richard Guenther wrote:
> > > >
> > > > > -O0 220MB, 0m17s
> > > > > -O1 260MB, 0m24s
> > > > > -O2 800MB, 7m35s  (-ftime-report on a faster machine attached 
> > > > > below)
> > > > > -O2 -fno-unit-at-a-time 260MB, 0m28s
> > > > Hmm, based on what the time report says, could you try with
> > > > -fno-tree-pre?
> > >
> > > loop times go up the roof then, but the tree stuff looks ok.
> >
> > Also, how much memory with -fno-tree-pre at -O2?
>
> This time without "fix" on the fast machine (has 1gig ram, too):
>
> 560MB
Thanks a bunch!
that helps me do some calculations.

It looks like the memory usage for PRE given the number of expressions 
(you don't even want to know what the number was) it had to process was 
normal.
There are certain unavoidable costs because we have to build new trees 
*during* PRE in order to determine whether we can optimize the 
expression.
Without changing the algorithm somewhat (which i've been considering), 
this can't be done.

I've been looking at C-SSAPRE: 
http://www-2.cs.cmu.edu/~tom7/ssapre/final.pdf
and seeing if i could reuse the coloring algorithm to avoid having to 
build trees in rename2.
--Dan