Add classdb_construct_object3 and classdb_register_extension_class6 (refcount-aware inits)

[Ivorforce]

• Follow-up to Make memnew(RefCounted) return Ref, to improve ownership safety #111965
• Works towards addressing Issues related to initializing new instance in GDExtension/Godot modules #111075

Brings the gdextension interface on the same level as core, by offering functions that create objects with a refcount already established (and expecting callers to assume ownership).

This can be used to address #111075, and may in the future lead to the ability to remove the refcount_init hack, which is a significant performance cost.

Hoping to establish this new interface before a 4.7 release, such that gdextensions can start adopting it already.

[dsnopek]

Thanks! I'm super excited to get this into GDExtension :-)

[Yarwin]

Doesn't make any difference – but should it really be set as an alias for GDExtensionClassCreationInfo4?

GDExtensionClassCreationInfo4 defines create_instance_func as:

                {
                    "name": "create_instance_func",
                    "type": "GDExtensionClassCreateInstance2",
                    "description": [
                        "(Default) constructor; mandatory. If the class is not instantiable, consider making it virtual or abstract."
                    ]
                },

while GDExtensionClassCreationInfo6 uses GDExtensionClassCreateInstance3.
As I said it makes no difference - same signature, different semantics, you set it accordingly ( extension->gdextension.create_instance3 = p_extension_funcs->create_instance_func;) if caller uses classdb_register_extension_class6, so effectively it is a nitpick

[Yarwin]

I adjusted rust bindings to this branch (https://github.com/Yarwin/gdext/tree/feature/support-new-initialization, WIP ofc) and ran it against simple project:

rust part

#[derive(GodotClass)]
#[class(base = RefCounted)]
pub struct MyClass {
    base: Base<RefCounted>,
}

#[godot_api]
impl IRefCounted for MyClass {
    fn init(base: Base<Self::Base>) -> Self {
        godot_print!("henlo from myclass constructor!");
        let me = Self { base };
        // Cloning the base increases the refcount and drops it afterwards.
        // Note that it is not possible with safeguards enabled on current master; we prohibit such operation.
        godot_print!("I am: {}", me.base().clone());
        me
    }
}

Gdscript part

extends Node2D

var weak: WeakRef
var other_weak: WeakRef

func other_fn():
	print("some time passed!")
	print("my class weak: ", weak.get_ref())
	print("other class weak: ", other_weak.get_ref())
	pass

func _ready() -> void:
	var other_refcounted = RefCounted.new()
	self.other_weak = weakref(other_refcounted)
	var my_class = MyClass.new()
	self.weak = weakref(my_class)
	
	get_tree().create_timer(0.5).timeout.connect(other_fn)

Godot 4.6: panic/(UB without safeguards) after print ("I am ...").

This branch + gdext with classdb_register_extension_class6 (and classdb_construct_object3):

henlo from myclass constructor!
I am: <RefCounted#-9223372009004595715>
some time passed!
my class weak: <null>
other class weak: <null>

---

Additionally I ran current version of godot-rust test suite against Godot built from this branch (godot-rust (API v4.6.stable.official, runtime v4.7.dev.custom_build, safeguards strict)) and everything seems to be fine 👍 (in other words - we are probably good on backward compatibility front).

Finally I ran full adjusted godot-rust test suite against this branch ((API v4.7.dev.custom_build, runtime v4.7.dev.custom_build, safeguards strict)) and it works as well (well, after addressing one UB due to missing type check 😁).

[dsnopek]

Thanks! This looks good to me

[Repiteo]

Thanks!