Re: Patch fixing 3.3 bug 9745 /Â?10021

[Geoff Keating <geoffk at geoffk dot org>]

Mark Mitchell <mark@codesourcery.com> writes:

> On Wed, 2003-07-02 at 07:53, David Edelsohn wrote:
> > >>>>> Marcus Meissner writes:
> > 
> > Marcus> Bugzilla entry 10021 contains a patch fixing the 'critical' 3.3 regression
> > Marcus> in bug 9745 from Jim Wilson.
> > 
> > Marcus> Can someone please review the patch for inclusion into the 3.3 branch
> > Marcus> (and 3.4)?
> > 
> > Marcus> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=10021
> > 
> > Marcus> Patch is in attached file:
> > 
> > Marcus> http://gcc.gnu.org/bugzilla/attachment.cgi?id=4306&action=view
> > 
> > 	If the patch goes in, I will immediately open a critical
> > performance regression bug against 3.3 and 3.4.  As with my earlier patch
> > for PR 9475, the proposed patch is too conservative and disables too much
> > alias-based optimizations.
> 
> It's up to you and Geoff (as RS6000 maintainers) to decide whether the
> RS6000 port is better off generating bad code or slow code.
> 
> We are going to be stuck in one state or the other until someone is
> willing to invest some serious effort in this problem.

In the current situation, the bug can be avoided by using
-fno-strict-aliasing.  If this patch was applied, the performance
regression could not be avoided.  I therefore think that the current
situation is least worst of these two choices.

However, we need to fix the bug properly.  Experience has shown that
the bug is more likely to appear in complex functions, for instance
those created by intermodule inlining.  In those situations, neither
the bug nor a performance regression is acceptable.

-- 
- Geoffrey Keating <geoffk@geoffk.org>