Re: [PATCH] Always create a new value object in valpy_do_cast

[Hannes Domani via Gdb-patches <gdb-patches@sourceware.org>]

 Am Sonntag, 22. Januar 2023, 18:49:55 MEZ hat Tom Tromey <tom@tromey.com> Folgendes geschrieben:

> >>>>> "Hannes" == Hannes Domani via Gdb-patches <gdb-patches@sourceware.org> writes:
>
> Hannes> In case a pretty printer casts a value to the same type, so value_cast()
> Hannes> returns the same object again, you get this simplified situation:
>
> Hannes> {
> Hannes>  struct value *val = ((value_object *) self)->value;
> Hannes>  scoped_value_mark free_values;
> Hannes>  res_val = value_cast (type, val); // this returns val
> Hannes>  result = value_to_value_object (res_val);
> Hannes> }
>
> Hannes> So value_to_value_object() removes a value at or before the free_values
> Hannes> marker.
>
> Something seems off about this to me.
>
> value_to_value_object does this:
>
>       val_obj->value = release_value (val).release ();
>
> release_value is like an "incref" but its actual semantics are: if the
> value is on all_values, remove it and return it; otherwise incref.  That
> way the caller always (1) is assured that the value isn't on all_values,
> and (2) always gets a new reference.
>
> So, I would expect that call to have released it from all_values and
> therefore it would not be destroyed.

The problem isn't the value that's reference by gdb.Value, instead
one of the other values in all_values before it.

But release_value removes it from all_values, and it was the exact value
that the scoped_value_mark free_values instance was using as the mark point,
and since it was then missing, all_values was cleared completely.
And one of those earlier values is still used by the printing later on.

> Furthermore, no matter what, I'd expect a gdb.Value to hold an owning
> reference to the underlying value, so it still shouldn't be destroyed by
> value_free_to_mark.
>
> I'm not doubting there's a bug here, but I don't understand how it comes
> about.  Also, I would rather not fix it the way it is done in this
> patch, because I think it is preferable for users of values not to have
> to know about whether or not a given API might return the same value.

It's also weird for me that users of these APIs need to know this, that's
why the way scoped_value_mark works was very surprising for me.


Hannes