Re: [patch,gfortran] PR 17740

[Tobias Schlüter <tobias dot schlueter at physik dot uni-muenchen dot de>]

Erik Edelmann wrote:
> I've been looking at PR 17740, where we get an ICE when calling
> elemental functions via a generic names.  The problem seems to be
> the check if the called function is elemental in
> gfc_trans_arrayfunc_assign():
> 
>   /* Elemental functions don't need a temporary anyway.  */
>     if (expr2->symtree->n.sym->attr.elemental)
>         return NULL;
> 
> The problem is that expr2->symtree->n.sym is (appearantly) the
> symbol of the generic name, not the specific function called. I
> wrote a patch (attached) that checks the symbol
> expr2->value.function.esym instead. With the patch, the testcase
> in the PR compiles, and the patch passes reg. testing (Linux/x86,
> 4.0 and mainline).
> 
> There is however, one thing that I am still a bit unsure about (I
> hope this question doesn't sound too stupid); I learned (the hard
> way) that expr2->value.function.esym is NULL sometimes (which is
> why the patch checks that first).  Can I trust
> expr2->value.function.esym to allways be non-NULL for elemental
> functions?  I browsed the gfortran source code, but found no
> definite answer to that question.

It can apparently, see resolve.c:767.
  if (sym->attr.generic)
    {
      s =
	gfc_search_interface (sym->generic, 0, &expr->value.function.actual);
      if (s != NULL)
	{
	  expr->value.function.name = s->name;
	  expr->value.function.esym = s;
	  expr->ts = s->ts;
	  if (s->as != NULL)
	    expr->rank = s->as->rank;
	  return MATCH_YES;
	}

      /* TODO: Need to search for elemental references in generic interface */
    }
I assume what's missing is that array arguments are matched against elemental
function in interfaces, but I'm not sure.

Furthermore, trans-expr.c:1550 says this:
  /* expr.value.function.esym is the resolved (specific) function symbol for
     most functions.  However this isn't set for dummy procedures.  */
  sym = expr->value.function.esym;
  if (!sym)
    sym = expr->symtree->n.sym;
  gfc_conv_function_call (se, sym, expr->value.function.actual);

It looks like your patch is correct, but I'd rather we understand when esym
should be set, first.

- Tobi