[Bug fortran/102596] [11/12 Regression] ICE in gfc_omp_clause_default_ctor, at fortran/trans-openmp.c:713 since r11-4883-ge929ef532ad52cde

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=102596

Tobias Burnus <burnus at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org
           Keywords|                            |ice-on-valid-code

--- Comment #2 from Tobias Burnus <burnus at gcc dot gnu.org> ---
The problem is that gfc_omp_clause_default_ctor requires outer != NULL_TREE
(gcc_assert). 

(C++ does ignore the last argument and C uses the hook_tree_tree_tree_tree_null
fallback.)

That is used in Fortran for:
  /* Allocatable arrays and scalars in PRIVATE clauses need to be set to
     "not currently allocated" allocation status if outer
     array is "not currently allocated", otherwise should be allocated.  */

But omp-low.cc has:
                  x = lang_hooks.decls.omp_clause_default_ctor
                                (c, unshare_expr (new_var),
                                 cond ? NULL_TREE
                                 : build_outer_var_ref (var, ctx));
Note the NULL_TREE.


My impression is that NULL_TREE is fine for reduction - and there is also code
like:


  /* Reduction clause requires allocated ALLOCATABLE.  */
  if (OMP_CLAUSE_CODE (clause) != OMP_CLAUSE_REDUCTION
      && OMP_CLAUSE_CODE (clause) != OMP_CLAUSE_IN_REDUCTION
      && OMP_CLAUSE_CODE (clause) != OMP_CLAUSE_TASK_REDUCTION)
    {
      gfc_init_block (&cond_block);
...
      tree tem = fold_convert (pvoid_type_node,
                               GFC_DESCRIPTOR_TYPE_P (type)
                               ? gfc_conv_descriptor_data_get (outer) :
outer);/*...*/
      cond = fold_build2_loc (input_location, NE_EXPR, logical_type_node,
                              tem, null_pointer_node);


However - there is other code which uses 'outer' like:
      tree tem = gfc_walk_alloc_comps (outer, decl,
or 
      gfc_add_modify (&cond_block, decl, outer);
and those make use of 'outer'.

I don't quickly see whether outer is always required or it can be deduced in
this case. 
(Does using OMP_CLAUSE_REDUCTION_PLACEHOLDER() make sense here?) — Or whether
some is-always-used case (→ 'cond' case) is needed as additional flag or
encoded in outer.


PS: I don't think it is a real regression as 'reduction(task:' wasn't supported
before.