[Bug target/85246] [og7, nvptx, openacc] gemm.f90 fails with -mlong-vector-in-workers

[vries at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85246

--- Comment #2 from Tom de Vries <vries at gcc dot gnu.org> ---
(In reply to Tom de Vries from comment #1)
> I went through a couple of cycles of minimizing the failure, seeing
> something suspicious, modifying by hand or writing a tentative patch, but
> every time I went back to the original non-minimized example I got the
> failure again.
> 
> Anyway, things that may be causing this fail:
> 
> 1.
> 
> The og7 fix for PR85204 introduces a unified jump (bra.uni) for a jump
> conditional consisting of a test for vector id == 0 && worker id == 0. The
> fact that we're going a different direction in worker id 0 for vector id 0
> and vector id 1 means the branch diverges, and is _not_ unified. It seems
> prudent to fix this by reverting the og7 fix and backporting the trunk fix.

Fixed here: https://gcc.gnu.org/ml/gcc-patches/2018-04/msg00525.html