Re: [PATCH 22/26] rust: qom: add casting functionality

[Paolo Bonzini]

On Mon, Dec 16, 2024 at 1:35 PM Zhao Liu <zhao1.liu@intel.com> wrote:
>
> > +/// Macro to mark superclasses of QOM classes.  This enables type-safe
> > +/// up- and downcasting.
> > +///
> > +/// # Safety
> > +///
> > +/// This macro is a thin wrapper around the [`IsA`] trait and performs
> > +/// no checking whatsoever of what is declared.  It is the caller's
> > +/// responsibility to have $struct begin, directly or indirectly, with
> > +/// a field of type `$parent`.
> > +#[macro_export]
> > +macro_rules! qom_isa {
> > +    ($struct:ty : $($parent:ty),* ) => {
>
> This macro is quite good, but it requires specifying all the parents...
> So I am thinking if it is possible to move ParentType to ObjectType, and
> then try to traverse the ParentType in the macro, implementing IsA for
> each ParentType... However, the first difficulty has already stopped me:
> I cannot define ParentType for Object itself.

I am not sure how that could be done, but I've seen people
implementing boolean logic purely with types, like

pub struct True;
pub struct False;
pub trait Boolean {}
impl Boolean for True;
impl Boolean for False;
pub trait Or<T: Boolean>: Boolean {
    type Result: Boolean;
}
impl<T: Boolean> Or<T> for True {
    type Result = True;
}
impl<T: Boolean> Or<T> for False {
    type Result = T;
}

and I think that can be used to implement recursive IsA, but that is a
bit too magic for this first step. QEMU class hierarchies are
relatively shallow.

> > @@ -94,8 +147,224 @@ pub unsafe trait ObjectType: Sized {
> >      /// The name of the type, which can be passed to `object_new()` to
> >      /// generate an instance of this type.
> >      const TYPE_NAME: &'static CStr;
> > +
> > +    /// Return the receiver as an Object.  This is always safe, even
> > +    /// if this type represents an interface.
>
> This comment is a bit confusing to me... EMM, why mention the interface?
> I understand if something implements ObjectType, then it should be an
> Object, so deref/cast here would be valid. And if it is an interface,
> it would need to implement the corresponding trait.

What I meant is that interfaces do (will) not implement IsA<Object>,
but they are ObjectTypes. So if you have let's say an "&impl
UserCreatable" argument, you could use as_object() to obtain an
&Object.

> This cast idea is nice! In the future, class might also need to implement=
> similar cast support (e.g., class_init in virtio).

Ok, good!

Paolo