[flang-commits] [flang] [Flang][OpenMP] Fix copyin allocatable lowering to MLIR (PR #122097)

[via flang-commits]

https://github.com/jeanPerier commented:

The patch makes sense for that direction (original item allocated, thread copy unallocated), but I think the logic is probably broken in the other direction too (original item unallocated, copy allocated).

The requirement for copyin in OpenMP spec says that _"If the original list item does not
have the POINTER attribute but has the allocation status of unallocated, each copy will have the same status."_

So the copy should be unallocated in the else case if it allocated.

Example:
```
subroutine allocatable4()
  use omp_lib, only: omp_get_thread_num
  integer, allocatable, save :: a
  !$omp threadprivate(a)
  !$omp parallel
    allocate(a)
  !$omp end parallel

  deallocate(a)

  !$omp parallel copyin(a)
    print *, omp_get_thread_num(), allocated(a)
  !$omp end parallel
end subroutine

  call allocatable4()
end
```

This should print (in any order):
```
           0 F
           3 F
           1 F
           2 F
```

But I think it will print `T` with flang outside of the main thread after this patch (and would crash before the patch).

https://github.com/llvm/llvm-project/pull/122097