Re: [pooma-dev] Using gcc3.0 to compile POOMA

[Richard Guenther]

On 3 Jan 2003, Tarjei Knapstad wrote:

> On Thu, 2002-12-19 at 16:59, Richard Guenther wrote:
> > On Thu, 19 Dec 2002, Richard Guenther wrote:
> >
> > > Hi!
> > >
> > > I was just using gcc3.0 (CVS head) for POOMA and now noticed that
> > > even at -O the NearestNeighbors test from the Field testsuite is
> > > failing (works at -O0).
> >
> > Same problem with current gcc3.2 CVS (-O2 -fomit-frame-pointer
> > -march=athlon) :/
> >
> Those nasties in 3.0 are apparently fixed in 3.2. Have you tried running

No - I see lots of extra failures on powerpc comapred to a ia32 run.

> that test through Valgrind (http://devel-home.kde.org/~sewardj/) and see
> if there's anything going bad in code execution? In my experience such

Yes - there is nothing unusual going on (tested ia32 only, of course).

> behaviour as described, is (at least in my case :) almost allways due to
> bad code, not bad optimization on the compilers part.

Maybe its time to try to construct smaller testcases for this... (from
just looking at the faulty values, it seems gcc misoptimizes simple loops
of the form

  for (int i=0; i<3; i++)
      a[i] = b[i];

but that would cause failures all over the place, I think, so there must
be something else triggering the failure.

Richard.