https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121185
kargls at comcast dot net changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |mikael at gcc dot gnu.org
--- Comment #25 from kargls at comcast dot net ---
(In reply to anlauf from comment #18)
> (In reply to Jürgen Reuter from comment #17)
> > (In reply to anlauf from comment #16)
> > > (In reply to Jürgen Reuter from comment #15)
> > > > Created attachment 61950 [details]
> > > > Close-to-minimal reproducer, 186 lines
> > > >
> > > > The division of the array pmatrix%value by trace is not performed
> > > > elemental
> > > > but just on the first array element.
> > >
> > > Great reproducer!
> > >
> > > Changing the line
> > >
> > > pmatrix%value = pmatrix%value / trace * pmatrix%degree
> > >
> > > to
> > >
> > > pmatrix%value(:) = pmatrix%value / trace * pmatrix%degree
> > >
> > > to suppress the reallocation on assignment fixes the testcase for me.
> > >
> > > Can you try to revert Mikael's recent changes and see if it fixes the
> > > regression?
> >
> > Which commit by Mikael is that? must have been after Monday 0:01 or so.
> > I am still checking the second example with the VEGAS/VAMP integration
> > whether this is a similar problem or something else.
>
> The following commits come to mind:
>
> r16-2247, r16-2248, r16-2249
Yes, I suspect one of Mikael's patch is hitting a condition that was not
anticipated and not covered in the testsuite. I haven't had time to bisect,
yet.
