On Sun, Oct 02, 2016 at 06:15:18PM +0800, Chung-Lin Tang wrote:
> This patch fixes the two ICEs listed on PR77371.
> One is due to the Fortran omp_privatize_by_reference hook returning true
> for types like 'character(kind=1)[1:XX] *', causing them to be processed
> by the path intended for C++ reference types.
The path isn't something intended for C++ reference types, but for all the
vars where whatever they point to should be privatized rather than just
their value. Consider
program p
integer, allocatable :: n
integer :: m
allocate (n)
n = 6
!$acc parallel firstprivate(n) private(m)
m = n
!$acc end parallel
end
testcase which with -fopenacc ICEs the same way, and then look carefully
what is done on
program p
integer, allocatable :: n
integer :: m
allocate (n)
n = 6
!$omp parallel firstprivate(n) private(m)
m = n
!$omp end parallel
end
with -fopenmp. The var is actually properly allocatable in the latter case,
while it is not with your patch on the first testcase, you just copy over the
host pointer, that
is definitely not going to work on non-shared memory offloading.
There is nothing special about references that use POINTER_TYPE as opposed
to REFERENCE_TYPE. So, please first get this working with firstprivate on
allocatables and only then start to play with reductions.
> The other one is simply not setting 'remove = true' while error_at() was
> already called.
The gimplify.c change is ok for trunk.
> Tested without regressions, committed on gomp-4_0-branch,
> is this okay for trunk as well?
>
> Thanks,
> Chung-Lin
>
> PR fortran/77371
> * omp-low.c (lower_omp_target): Avoid reference-type processing
> on pointers for firstprivate clause.
> * gimplify.c (gimplify_adjust_omp_clauses): Add 'remove = true'
> when emitting error on private/firstprivate reductions.
>
> testsuite/
> * gfortran.dg/goacc/pr77371-1.f90: New test.
> * gfortran.dg/goacc/pr77371-2.f90: New test.
Jakub
