Re: Does anyone grok this one ?

[Tobias Burnus <burnus at net-b dot de>]

Hi Toon,

Toon Moene wrote:
> $ /usr/snp/bin/gfortran -c globals.f90
> globals.f90:37.24:
>
>   TYPE(weight_t) g_winfo ! weights info
>                        1
> Error: Object 'g_winfo' at (1) must have the SAVE attribute for
> default initialization of a component

> *I* don't see any "default initialization of a component" ...
Neither do I ...

I could reproduce it with gfortran 4.2 and 4.3.

Well, the solution is simple: Either fill a PR or better provide a patch ;-)

Looking at  resolve_fl_variable and gfc_default_initializer, I think the
problem is that you have allocatable variables which are treated for
this test wrongly as default initializer.

I don't see at the moment why allocatables need a special treatment, but
gfc_default_initializer is used in:
- decl.c: variable_decl
- resolve.c: resolve_allocate_expr and apply_default_init and
resolve_fl_variable (twice)

Maybe someone who knows the allocate extension should have a look at
this as he might remember why gfc_default_initializer returns also for
allocatables non-NULL.

Tobias

  ------------

resolve.c: resolve_fl_variable

  constructor_expr = NULL;
  if (sym->ts.type == BT_DERIVED && !(sym->value || flag))
    constructor_expr = gfc_default_initializer (&sym->ts);

  if (sym->ns->proc_name
      && sym->ns->proc_name->attr.flavor == FL_MODULE
      && constructor_expr
      && !sym->ns->save_all && !sym->attr.save
      && !sym->attr.pointer && !sym->attr.allocatable)
    {
      gfc_error("Object '%s' at %L must have the SAVE attribute %s",


And in expr.c:
gfc_default_initializer (gfc_typespec *ts)
  init = NULL;
  /* See if we have a default initializer.  */
  for (c = ts->derived->components; c; c = c->next)
    {
      if ((c->initializer || c->allocatable) && init == NULL)
        init = gfc_get_expr ();
    }