[Bug fortran/78865] ICE in create_tmp_var, at gimple-expr.c:473

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78865

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
The problem is that gfc_conv_procedure_call changes the type of the sub
function
from one returning void to one returning int:
  /* If there are alternate return labels, function type should be
     integer.  Can't modify the type in place though, since it can be shared
     with other functions.  For dummy arguments, the typing is done to
     this result, even if it has to be repeated for each call.  */
  if (has_alternate_specifier
      && TREE_TYPE (TREE_TYPE (TREE_TYPE (se->expr))) != integer_type_node)
    {
      if (!sym->attr.dummy)
        {
          TREE_TYPE (sym->backend_decl)
                = build_function_type (integer_type_node,
                      TYPE_ARG_TYPES (TREE_TYPE (sym->backend_decl)));
          se->expr = gfc_build_addr_expr (NULL_TREE, sym->backend_decl);
        }
      else
        TREE_TYPE (TREE_TYPE (TREE_TYPE (se->expr))) = integer_type_node;
    }

While this can work for implicitly defined subroutine calls, if
sym->backend_decl
is actually merged with the subroutine definition, changing it afterwards just
doesn't work.  Either it should be diagnosed as bug earlier (noting that caller
expects alternate return, but callee does not), or at least it shouldn't change
the type of the function, but do something different (even just generate
__builtin_trap instead of the call).