Segfault on O2 optimisation for test case for PR18689

[Tobias Schlüter]

Andrew Pinski wrote:
>>I tried your patch together with Jakub's fix for PR15080, but the
>>scalarization still goes wrong.  So it looks like you found a 
>>different bug
>>somewhere, the .original dump looks fine.  The bug is very fragile: 
>>addding
>>debug code makes the testcase pass.  Maybe we're dealing with a
>>misoptimization for a change.
> 
> 
> Does this fail with -O2 -fno-strict-aliasing.

No, I don't think this says much given the erratic appearance of this bug.
I'll have a look at the .optimized dumps and will try to figure out what goes
wrong if I find time for this before I leave for the weekend.

>  I know there was a bug 
> report
> about wrong aliasing sets for some types (there is a testcase which is 
> xfailed
> because of it too).

Yes, that's the ret_pointer_1.f90 failure.  Aliasing of arrays with
descriptors is broken IIRC.

- Tobi