Re: PR 10776

[Jan Hubicka <hubicka at ucw dot cz>]

> On Sat, Jan 10, 2004 at 07:12:29PM +0100, Jan Hubicka wrote:
> > Since the test is inherently quadratic it appears to make sense to
> > simply limit number of memory locations considered.  This patch makes
> > the testcase at -O1 to compile in 22 seconds.  Changing the value to 10
> > (instead of 500 locations) reduce it to 17 seconds.
> 
> Huh?  It still takes > 600 seconds here.  It's all spent in reload
> proper, and not in cselib.

Hmm?  I did used -O1 i686 compiler.  This seems to depend on particular
code generated.  When optimizing for Athlon there are no temporaries and
thus reload doesn't degenerate.
It all depends on whether we do direct memory stores or go via register.

There is yet another quadratic loop in reload itself when looking for
equal registers, but I didn't find any testcase for it, perhaps it fires
for you?
What settings do you use?

Honza
> 
> 
> 
> r~