[Bug optimization/13653] [3.3 regression] -O2 -funroll-loop miscompiles POOMA testcase

[rguenth at tat dot physik dot uni-tuebingen dot de]

------- Additional Comments From rguenth at tat dot physik dot uni-tuebingen dot de  2004-01-29 22:54 -------
Subject: Re:  [3.3 regression] -O2 -funroll-loop
 miscompiles POOMA testcase

On Thu, 29 Jan 2004, Richard Guenther wrote:

> On Thu, 29 Jan 2004, Richard Guenther wrote:
>
> > On Thu, 29 Jan 2004, dberlin at gcc dot gnu dot org wrote:
> >
> > > I'll attach my even *more* reduced testcase, but that only passes at -O2 and
> > > fails at -O2 -funroll-loops (IE it doesn't pass at -O0 or -O1)
> >
> > That does hint to maybe a stack corruption or somehow.  That would be
> > consistent with my observation that the original testcase fails only, if I
> > enable assertions in the code, if I disable them and they get #ifdef'ed
> > away, the failure does not occour.
> >
> > So it may be interesting to look at the generated assembly of the main
> > function and looking for corrupt stack after calls?
>
> Additionally to the noreturn attribute, I can reduce the main function to
>
> int main(int argc, char *argv[]) {
>         Interval<1> D(6);
>         Interval<3> I3(D,D,D);
>         Loc<3> blocks(2,2,2);
>         UniformGridPartition<3> partition(blocks);
>         UniformGridLayout<3> layout(I3, partition,ReplicatedTag());
>         Array<3, double, MultiPatch<UniformTag,Brick> > u(layout);
>         for (int i2 = 0; i2 < 4 /* 3 doesnt segfault */; i2++)
>                 for (int i1 = 0; i1 < 1; i1++)
>                         for (int i0 = 0; i0 < 1; i0++)
>                                 u(i0,i1,i2) = 0;
>         return 0;
> }
>
> and still get the segfault, but it is fixed with either running the first
> loop only to i2 < 3, or removing any of the loops.
>
> In fact, looking at the assembler, there are two calls to
> _ZNK21UniformGridLayoutDataILi3EE8globalIDEiii inside the loop, and the
> loop itself doesnt change with the noreturn attribute on toss_cookies, but
> the called function does, as it calls toss_cookies several times.

But if I remove the toss_cookies calls from the offending function the
segfault persists. Ugh. Feels like chasing a Heisenbug.

Going to bed now,
Richard.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=13653