[Bug middle-end/103416] [12 Regression][OpenMP] Bogus firstprivate(n) map(to:n [len: 4][implicit])

["burnus at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103416

--- Comment #4 from Tobias Burnus <burnus at gcc dot gnu.org> ---
Created attachment 51872
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51872&action=edit
RFC Patch to avoid the pointless evaluation, see comment 4

(In reply to Tobias Burnus from comment #3)

> * Why is the var generated in the parent scope of 'omp target' instead of
> inside 'omp target'?

The problem is a forced evaluation of the array bounds, which I regard as
pointless if the variable is just a plain variable - no array ref, not struct
ref no ...

Cf. attachment. (The question is when 'force=true' is needed and whether the
DECL_P check is the right one or whether more or less should be permitted.)

This is indeed the same as issue as PR80330 (8...)

 * * *

The
  libgomp: cuCtxSynchronize error: misaligned address
is a regression – see comment 1 for a patch which fixes it. This is PR90030
(9...)


 * * *

> * Why is 'map(to:' and not 'firstprivate' used?

Because of:

gfc_omp_predetermined_mapping (tree decl)
{
  if (DECL_ARTIFICIAL (decl)
      && ! GFC_DECL_RESULT (decl)
      && ! (DECL_LANG_SPECIFIC (decl)
            && GFC_DECL_SAVED_DESCRIPTOR (decl)))
    return OMP_CLAUSE_DEFAULTMAP_TO;

I wonder whether OMP_DEFAULTMAP_FIRSTPRIVATE  wouldn't make more sense in this
case – at least for gfc_omp_scalar_target_p ?

Which is also related to PR80330 (8...)