When a MethodInst is called as a function in the VM, it needs to push
its `self_binding` member to the stack. Previously, we were downcasting
(if possible) to MethodInst in the VM, but really, we are calling
`MethodInst::call` anyway, so it makes more sense to do
MethodInst-specific stuff in the MethodInst-specific function.
Signed-off-by: Alek Ratzloff <alekratz@gmail.com>
Types now can have vtable elements which are used by instances to bind
themselves to methods. When Op::GetAttr is executed, it calls a new
function, Obj::get_attr_lazy. This will search:
* attributes on the object
* vtable on the object's type
* vtable on the object's type's type,
* etc.
This searches up the type tree for a named value. If it exists as an
attribute, it will be returned immediately. If it exists in the type's
vtable, then it will be inserted as an attribute. If the vtable value is
a function, the object that it is being called on will be bound to that
method as the `self` parameter.
Signed-off-by: Alek Ratzloff <alekratz@gmail.com>
Wow, what a ride. I think everything should be working now. In short:
* Objects use the `gc` crate, which as a `Gc` garbage-collected pointer
type. I may choose to implement my own in contiguous memory in the
future. We will see.
* The type system is no longer global. This is a bit of a burden,
because now, whenever you want to create a new object, you need to
pass its type object into the `Obj::instantiate` method, as well as
its `::create` static method.
* This burden is somewhat alleviated by the `ObjFactory` trait, which
helps create new objects as long as you have access to a `builtins`
hashmap. So something that would normally look like this:
fn init_builtins(builtins: &mut HashMap<String, ObjP>) {
let print_builtin = upcast_obj(BuiltinFunctionInst::create(
ObjP::clone(&builtins.get("BuiltinFunction").unwrap()),
"print",
print,
1
);
builtins.insert("print".to_string(), print_builtin)
// other builtins inserted here...
}
now looks like this:
fn init_builtins(builtins: &mut HashMap<String, ObjP>) {
let print_builtin = builtins.create_builtin_function("print", print, 1);
builtins.insert("print".to_string(), print_builtin);
}
(turns out, if all you need is a HashMap<String, ObjP>, you can
implement ObjFactory for HashMap<String, ObjP> itself(!))
Overall, I'm happier with this design, and I think this is what is going
to get merged. It's a little weird to be querying type names that are
used in the language itself to get those type objects, but whatever
works, I guess.
Next up is vtables.
Signed-off-by: Alek Ratzloff <alekratz@gmail.com>
I'm not super happy with this. But, the RwLock has been moved to the
`BaseObjInst::attrs` member. Although this is not exactly how it appears
in code, it basically does this:
type Ptr<T> = Arc<RwLock<T>>;
struct BaseObjInst {
attr: HashMap<String, Ptr<dyn Obj>>,
// etc
}
becomes
type Ptr<T> = Arc<T>;
struct BaseObjInst {
attr: RwLock<HashMap<String, ObjP>>,
// etc
}
This makes things a lot more ergonomic (don't have to use try_read() and
try_write() everywhere), but it also eliminates compile-time errors that
would catch mutability errors. This is currently rearing its ugly head
when initializing the typesystem, since `Type` needs to hold a circular
reference itself (which it already shouldn't be doing since it's a
reference-counted pointer!). Currently, all tests are failing because of
this limitation.
There are a couple of ways around this limitation.
The first solution would be just copying all of the object
instantiation code into the `init_types` function and avoid calling
`some_base_type.instantiate()`. This would probably be literal
copy-pasting, or maybe an (ugly) macro, and probably a nightmare to
maintain long-term. I don't like this option, but it would make
everything "just work" with reference-counted pointers.
The second solution would be to write our own garbage collector, which
would allow for circular references and (hypothetically) mutably
updating these references. This is something that I am looking into,
because I really want a RefCell that you can pass around in a more
ergonomic way.
I think the fundamental error that I'm running into is trying to borrow
the same value multiple times mutably, which you *really* shouldn't be
doing. I believe I need to write better code and does the same thing.
The only unsolved problem is circular references. This is not a problem
right now because I'm not writing code that has circular references
besides the base typesystem (which is not a problem because they need to
live the entire lifetime of the program), but it will be a latent
problem until it gets fixed.
Signed-off-by: Alek Ratzloff <alekratz@gmail.com>
