Patch fixing 3.3 bug PR 9745 and PR 10021

[Mark Mitchell]

On Wed, 2003-07-02 at 11:17, Michael Matz wrote:
> Hi,
> 
> On Wed, 2 Jul 2003, Franz Sirl wrote:
> 
> > compiled correctly, he has to compile with -O1 (this is what Suse does for
> > their _whole_ build system now) or -O2 -fno-strict-aliasing anyway.
> 
> No, in the packages which bite us we used -fno-strict-aliasing or similar.
> And now we have this patch applied to our gcc.  If it creates performance
> problems, well, so be it for now.  It's really sad, but broken code is
> unacceptable.

I agree.

> > And certainly this is slower than half-working aliasing, or? BTW, do you
> > have any hard numbers on how badly performance is affected (compared to
> > 3.2.3 and 3.3)?
> 
> This would be interesting indeed, yes.

I agree with the fact that data would be helpful; benchmarks with and
without Jim's patch would be nice.

Jim's patch is logical, and seems to be driven from a correct reading of
the interfaces in our aliasing code.  I am leaning towards installing
that patch for 3.3.1 at this point.  Dale's alternative is another
possibility -- as it pushes the failing cases further into the dark
corners, while preserving the current performance.

-- 
Mark Mitchell
CodeSourcery, LLC
mark@codesourcery.com