Re: [patch] Fix PR39976, sixtrack degradation

[Luis Machado <luisgpm at linux dot vnet dot ibm dot com>]

Hi Michael,

Thanks for the patch. It does fix the problem on 64-bits bringing the
performance numbers back.

On 32-bit we're still ~9% down, though we only have a single BB in the
loop and the loop structure is the same as the good case.

I'll do some more digging and will update the ticket.

Regards,
Luis

On Wed, 2009-05-13 at 21:28 +0200, Michael Matz wrote:
> Hi,
> 
> On Wed, 13 May 2009, Richard Guenther wrote:
> 
> > > +      for (gsi = gsi_start_bb (bb); gsi_stmt (gsi) != use_stmt; gsi_next (&gsi))
> > > +       if (gsi_stmt (gsi) == defa)
> > > +         return true;
> > 
> > Use gimple_uid here to check ordering.
> 
> Ah, yes, much better, thanks.
> 
> > They can be globally initialized by renumber_gimple_stmt_uids().
> 
> Instead of this I'll lazily initialize those only in blocks where it's 
> necessary, there are enough walks over all insns already.  I'm 
> regstrapping this currently and commit then.
> 
> 
> Ciao,
> Michael.