Re: [patch 2/3] Implement support for PowerPC BookE ranged watchpoints

[Pedro Alves <pedro@codesourcery.com>]

On Tuesday 28 December 2010 05:00:42, Joel Brobecker wrote:
> > +/* Return true if TYPE is scalar.  */
> > +
> > +int
> > +is_scalar_type (struct type *type)
> > +{
> > +  CHECK_TYPEDEF (type);
> > +
> > +  while (TYPE_CODE (type) == TYPE_CODE_REF)
> > +    {
> > +      type = TYPE_TARGET_TYPE (type);
> > +      CHECK_TYPEDEF (type);
> > +    }
> 
> I suggest you make it explicit how TYPE_CODE_REF types are handled.
> In certain cases, I think it might be making a difference (a REF
> to a struct might have a different classification than the struct
> in terms of argument passing, for instance).

Yeah.  For watching purposes, you want to consider the
reference, or really, its underlying pointer representation
as a scalar, nomatter what the reference is bound to.  Let me
try to clarify:

 int i;
 int &ref = i;

Currently, a watchpoint on "ref" only monitors changes
for the address stored in the reference-as-pointer.
So, for exact-watchpoint's purposes, when considering
individual pieces/locations of an expression to watch, the
TYPE_CODE_REF alone is a scalar, even if its target type
is a structure.
I've just tried "watch ref" per above, and I'm surprised, since
I thought a watchpoint on "ref" actually monitored two locations:
the underlying address stored in "ref"; and its target
value, similarly to watching '*ptr' in "int i; int *ptr= &i;".
I think we can actually call this a bug...

Contrast with valprint's purposes, where what matters
is the reference's target type, what the user sees.

-- 
Pedro Alves