Omp workshare error with 21.3 (f90) with -Mipa=inline

Accelerated Computing HPC Compilers nvc, nvc++ and nvfortran
1

I’m seeing the following regression with 21.3.
This worked with some older versions, including pgi/19.10 on the same machine.

Removing -mp or inline from -Mipa solves the problem.

$ cat test_self_conf.F90; nvfortran -mp -Mipa=fast,inline test_self_conf.F90; nvfortran -V
program test_simple
  implicit none
  real, allocatable, dimension(:,:,:) :: arr1, arr2

  contains
  subroutine test_ss

#ifdef _OPENMP
    print *, 'Using OpenMP'
#endif

    allocate(arr1(100,100,100))
    allocate(arr2(100,100,100))


    arr1(:,:,:) = 11.0
    !$omp parallel workshare
    arr2(:,:,:) = arr1(:,:,:)
    !$omp end parallel workshare

    print *, arr2(50,50,50)
  end subroutine test_ss
end program test_simple
NVFORTRAN-S-0000-Internal compiler error. queue_dtype: illegal type     115  (test_self_conf.F90: 6)
NVFORTRAN-S-0000-Internal compiler error. queue_dtype: illegal type     118  (test_self_conf.F90: 6)
NVFORTRAN-S-0000-Internal compiler error. queue_dtype: illegal type     121  (test_self_conf.F90: 6)
  0 inform,   0 warnings,   3 severes, 0 fatal for __nv_test_simple_test_ss_F1L17_1

nvfortran 21.3-0 linuxpower target on Linuxpower
NVIDIA Compilers and Tools
Copyright (c) 2020, NVIDIA CORPORATION.  All rights reserved.
2

Thanks Thomas,

I’ve filed a problem report, TPR #29968, and sent it to our compiler engineers for further investigation.

-Mat