Tracking issue for `arbitrary_self_types`

[arielb1]

Tracking issue for #![feature(arbitrary_self_types)].

This needs an RFC before stabilization, and also requires the following issues to be handled:

• figure out the object safety situation
• figure out the handling of inference variables behind raw pointers
• decide whether we want safe virtual raw pointer methods

Object Safety

See #27941 (comment)

Handling of inference variables

Calling a method on *const _ could now pick impls of the form

impl RandomType {
    fn foo(*const Self) {}
}

Because method dispatch wants to be "limited", this won't really work, and as with the existing situation on &_ we should be emitting an "the type of this value must be known in this context" error.

This feels like fairly standard inference breakage, but we need to check the impact of this before proceeding.

Safe virtual raw pointer methods

e.g. this is UB, so we might want to force the call <dyn Foo as Foo>::bar to be unsafe somehow - e.g. by not allowing dyn Foo to be object safe unless bar was an unsafe fn

trait Foo {
    fn bar(self: *const Self);
}

fn main() {
    // creates a raw pointer with a garbage vtable
    let foo: *const dyn Foo = unsafe { mem::transmute([0usize, 0x1000usize]) };
    // and call it
    foo.bar(); // this is UB
}

However, even today you could UB in safe code with mem::size_of_val(foo) on the above code, so this might not be actually a problem.

More information

There's no reason the self syntax has to be restricted to &T, &mut T and Box<T>, we should allow for more types there, e.g.

trait MyStuff {
    fn do_async_task(self: Rc<Self>);
}

impl MyStuff for () {
    fn do_async_task(self: Rc<Self>) {
        // ...
    }
}

Rc::new(()).do_async_stuff();

This doesn't have an RFC, but we want to experiment on this without one.

See #27941.

[porky11]

Why would you need this?
Why wouldn't you write an impl like this:

impl MyStuff for Rc<()> {
    fn do_async_task(self) {
        // ...
    }
}

I'd rather define the trait different. Maybe like this:

trait MyStuff: Rc {
    fn do_async_task(self);
}

In this case, Rc would be a trait type. If every generic type implemented a specific trait (this could be implemented automatically for generic types) this seems more understandable to me.

[cuviper]

This could only be allowed for trait methods, right?

For inherent methods, I can't impl Rc<MyType>, but if impl MyType can add methods with self: Rc<Self>, it seems like that would enable weird method shadowing.

[arielb1]

@cuviper

This is still pending lang team decisions (I hope there will be at least 1 RFC) but I think it will only be allowed for trait method impls.

[arielb1]

@porky11

You can't implement anything for Rc<YourType> from a crate that does not own the trait.

[arielb1]

So changes needed:

• remove the current error message for trait methods only, but still have a feature gate.
• make sure fn(self: Rc<Self>) doesn't accidentally become object-safe
• make sure method dispatch woks for Rc<Self> methods
• add tests

[SimonSapin]

I’ll look into this.

[arielb1]

Note that this is only supported to work with trait methods (and trait impl methods), aka

trait Foo {
    fn foo(self: Rc<Self>);
}
impl Foo for () {
    fn foo(self: Rc<Self>) {}
}

and is NOT supposed to work for inherent impl methods:

struct Foo;
impl Foo {
    fn foo(self: Rc<Self>) {}
}

[SimonSapin]

I got caught in some more Stylo work that's gonna take a while, so if someone else wants to work on this in the meantime feel free.

[kennytm]

Is this supposed to allow any type as long as it involves Self? Or must it impl Deref<Target=Self>?

trait MyStuff {
    fn a(self: Option<Self>);
    fn b(self: Result<Self, Self>);
    fn c(self: (Self, Self, Self));
    fn d(self: Box<Box<Self>>);
}

impl MyStuff for i32 {
   ...
}

Some(1).a();  // ok?
Ok(2).b();  // ok?
(3, 4, 5).c(); // ok?
(box box 6).d(); // ok?

[mikeyhew]

I've started working on this issue. You can see my progress on this branch

[mikeyhew]

@arielb1 You seem adamant that this should only be allowed for traits and not structs. Aside from method shadowing, are there other concerns?

[arielb1]

@mikeyhew

inherent impl methods are loaded based on the type. You shouldn't be able to add a method to Rc<YourType> that is usable without any use statement.

[arielb1]

That's it, if you write something like

trait Foo {
    fn bar(self: Rc<Self>);
}

Then it can only be used if the trait Foo is in-scope. Even if you do something like fn baz(self: u32); that only works for modules that use the trait.

If you write an inherent impl, then it can be called without having the trait in-scope, which means we have to be more careful to not allow these sorts of things.

[mikeyhew]

@arielb1 Can you give an example of what we want to avoid? I'm afraid I don't really see what the issue is. A method you define to take &self will still be callable on Rc<Self>, the same as if you define it to take self: Rc<Self>. And the latter only affectsRc<MyStruct>, not Rc<T> in general.

[mikeyhew]

I've been trying to figure out how we can support dynamic dispatch with arbitrary self types. Basically we need a way to take a CustomPointer<Trait>, and do two things: (1) extract the vtable, so we can call the method, and (2) turn it into a CustomPointer<T> without knowing T.

(1) is pretty straightforward: call Deref::deref and extract the vtable from that. For (2), we'll effectively need to do the opposite of how unsized coercions are implemented for ADTs. We don't know T, but we can can coerce to CustomPointer<()>, assuming CustomPointer<()> has the same layout as CustomPointer<T> for all T: Sized. (Is that true?)

The tough question is, how do we get the type CustomPointer<()>? It looks simple in this case, but what if CustomPointer had multiple type parameters and we had a CustomPointer<Trait, Trait>? Which type parameter do we switch with ()? In the case of unsized coercions, it's easy, because the type to coerce to is given to us. Here, though, we're on our own.

@arielb1 @nikomatsakis any thoughts?

[nikomatsakis]

@arielb1

and is NOT supposed to work for inherent impl methods:

Wait, why do you not want it work for inherent impl methods? Because of scoping? I'm confused. =)

[nikomatsakis]

@mikeyhew

I've been trying to figure out how we can support dynamic dispatch with arbitrary self types.

I do want to support that, but I expected it to be out of scope for this first cut. That is, I expected that if a trait uses anything other than self, &self, &mut self, or self: Box<Self> it would be considered no longer object safe.

[mikeyhew]

@nikomatsakis

I do want to support that, but I expected it to be out of scope for this first cut.

I know, but I couldn't help looking into it, it's all very interesting to me :)

[arielb1]

Wait, why do you not want it work for inherent impl methods? Because of scoping? I'm confused. =)

We need some sort of "orphan rule" to at least prevent people from doing things like this:

struct Foo;
impl Foo {
    fn frobnicate<T>(self: Vec<T>, x: Self) { /* ... */ }
}

Because then every crate in the world can call my_vec.frobnicate(...); without importing anything, so if 2 crates do this there's a conflict when we link them together.

Maybe the best way to solve this would be to require self to be a "thin pointer to Self" in some way (we can't use Deref alone because it doesn't allow for raw pointers - but Deref + deref of raw pointers, or eventually an UnsafeDeref trait that reifies that - would be fine).

I think that if we have the deref-back requirement, there's no problem with allowing inherent methods - we just need to change inherent method search a bit to also look at defids of derefs. So that's probably a better idea than restricting to trait methods only.

Note that the CoerceSized restriction for object safety is orthogonal if we want allocators:

struct Foo;
impl Tr for Foo {
    fn frobnicate<A: Allocator+?Sized>(self: RcWithAllocator<Self, A>) { /* ... */ }
}

Where an RcWithAllocator<Self, A> can be converted to a doubly-fat RcWithAllocator<Tr, Allocator>.

[mikeyhew]

@arielb1

Because then every crate in the world can call my_vec.frobnicate(...); without importing anything, so if 2 crates do this there's a conflict when we link them together.

Are saying is that there would be a "conflicting symbols for architechture x86_64..." linker error?

Maybe the best way to solve this would be to require self to be a "thin pointer to Self" in some way (we can't use Deref alone because it doesn't allow for raw pointers - but Deref + deref of raw pointers, or eventually an UnsafeDeref trait that reifies that - would be fine).

I'm confused, are you still talking about frobnicate here, or have you moved on to the vtable stuff?

[arielb1]

I'm confused, are you still talking about frobnicate here, or have you moved on to the vtable stuff?

The deref-back requirement is supposed to be for everything, not only object-safety. It prevents the problem when one person does

struct MyType;
impl MyType {
    fn foo<T>(self: Vec<(MyType, T)>) { /* ... */ }
}   

While another person does

struct OurType;
impl OurType {
    fn foo<T>(self: Vec<(T, OurType)>) {/* ... */ }
}   

And now you have a conflict on Vec<(MyType, OurType)>. If you include the deref-back requirement, there is no problem with allowing inherent impls.

[raphaelcohn]

Just played around with this a bit and found that whilst a receiver of self: *mut Self is allowed, self: NonNull<Self> isn't (in a trait) and produces a compile error.

error[E0307]: invalid method receiver type: std::ptr::NonNull<Self>
  --> src/main.rs:38:29
   |
38 |     unsafe fn initialize(self: NonNull<Self>, arguments: Self::InitializeArguments);
   |                                ^^^^^^^^^^^^^
   |
   = note: type of `self` must be `Self` or a type that dereferences to it
   = help: consider changing to `self`, `&self`, `&mut self`, or `self: Box<Self>`

To my mind, where I know the receiver is a non-null pointer, but not necessarily initialized properly, this is a useful way of writing code as documentation; particularly for a trait, where the implementor can be certain he hasn't been given null.

It might also be worth considering Option<NonNull<Self>>, too, even if only to warn that it would be better expressed as *mut Self.

[shepmaster]

self: NonNull<Self> isn't

Because NonNull doesn't implement Deref. See also #48277.

considering Option<NonNull<Self>>

Option also doesn't implement Deref. AFAICT it cannot, at least not without panicking when it's None, which seems like a very bad ergonomic idea.

[SimonSapin]

*mut Self doesn’t implement Deref either.

[shepmaster]

*mut Self doesn’t implement Deref either.

Sure, but that was a specific addition (See also #46664):

Maybe the best way to solve this would be to require self to be a "thin pointer to Self" in some way (we can't use Deref alone because it doesn't allow for raw pointers - but Deref + deref of raw pointers, or eventually an UnsafeDeref trait that reifies that - would be fine).

Raw pointers are definitely special-cased in the compiler, while NonNull is a regular struct and not a language item.

[mikeyhew]

Raw pointers are definitely special-cased in the compiler

Indeed. NonNull<Self> is something that really should be supported at some point (at least if raw pointer receivers are), but it would require a separate trait from Deref that returns *const Self::Target, and we would need an RFC for that

[raphaelcohn]

@mikeyhew Another RFC? That's disappointing to learn. I haven't got the desire to go down that route. At least the thought's there.

@shepmaster I the code I am now commonly writing when dealing with low-level pointer stuff, Option<NonNull<T>> comes up quite a bit. It'd just be nice to have the compiler either deal with it by assuming *mut Self.

[mikeyhew]

This is the trait that I was thinking of. Both NonNull<T> and Option<NonNull<T>> could implement it.

trait DerefRaw {
    type Target;

    fn deref_raw(&self) -> *const Self::Target;
}

[SimonSapin]

I’m really not sure about doing this for Option<NonNull<T>>. Isn’t the whole point of it that you must check for None/null before accessing the pointer?

[raphaelcohn]

Option<NonNull<T>> is, in many cases, much easier to work with than *mut T and contains far more expressive power. It also more 'naturally' inevitably gets created by code working with NonNull<T>; suddenly moving to *mut T is weird, and, less-Rusty; one can't simply do Some(non_null) as a result of a function, say. It also allows 'balance' in match clauses, which are shorter to read and more ergonomic than if with is_null(); it does not require the casual reader of code to suddenly understand as much of raw pointers - knowledge of Rust's Option type is enough; and lastly, it makes up for the absence of is_not_null(), for which !x.something.is_null() reads appalling and requires more visual processing (and is easier to miss).

[SimonSapin]

Sorry, I should have clarified. Using Option<NonNull<T>> in general sounds good. What I’m not sure about is implementing DerefRaw as proposed above for Option<NonNull<T>>. Sure, it has an obvious implementation, but converting to a nullable *const T seems to defeat the point of using Option<NonNull<T>> in the first place.

[withoutboats]

the big question around raw pointers is the requirements as to the validity of vtables for invalid pointers. I’m not sure what we should do, but if we made Option<NonNull<T>> impl DerefRaw we’d be assuming the vtable could be invalid; calling a dynamic method on None would just always be UB, I believe.

…

On Friday, January 25, 2019, Simon Sapin ***@***.***> wrote: Sorry, I should have clarified. Using Option<NonNull<T>> in general sounds good. What I’m not sure about is implementing DerefRaw as proposed above for Option<NonNull<T>>. Sure, it has an obvious implementation, but converting to a nullable *const T seems to defeat the point of using Option<NonNull<T>> in the first place. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#44874 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AIpL0Jj1TskgEDDbOJ_vjVIk9U258spjks5vGwougaJpZM4Pk3fK> .

[kennytm]

Unless we have decided a representation of a dynamic-sized enum I'm opposed to Option<NonNull<dyn Stuff>>.

[mikeyhew]

@withoutboats Ohh, good point, I forgot about that. That means there's potentially a real difference between *mut T and Option<NonNull<T>>.

@kennytm I think you're confusing Option<NonNull<dyn Trait>> with Option<dyn Trait>

[kennytm]

@mikeyhew Oh right, sorry.

[arielb1]

@mikeyhew

Are you talking about Option<NonNull<*mut dyn Trait>> (which is a type that makes sense) or about NonNull<dyn Trait> (which I don't think makes much sense)?

[arielb1]

BTW, from an API design standpoint, one thing that would work would be to have a special ArbitrarySelfTypePtr trait (bikeshed on name):

trait ArbitrarySelfTypePtr {
    type Next: ?Sized;
}

impl<T: Deref + ?Sized> ArbitrarySelfTypePtr for T {
    type Next = T::Target;
}

impl<T: ?Sized> ArbitrarySelfTypePtr for *mut T {
    type Next = T;
}

impl<T: ?Sized> ArbitrarySelfTypePtr for *const T {
    type Next = T;
}

impl<T: ?Sized> ArbitrarySelfTypePtr for NonNull<T> {
    type Next = T;
}

// Maybe, if we really want that?
impl<T: ?Sized> ArbitrarySelfTypePtr for Option<T> {
    type Next = T;
}

[arielb1]

The "ArbitrarySelfTypePtr" exists solely so we can have a "sane" deref chain for inherent methods etc.

[raphaelcohn]

@arielb1 I like your suggested ArbitrarySelfTypePtr trait because it logically documents what is and isn't possible for a self type. It would make it easy to refer to in compiler errors, eg an arbitrary self type must implement the trait ArbitrarySelfTypePtr.

BTW, for the avoidance of doubt, I'm not absolutely wedded to the Option<NonNull<T>> requirement, it's just that it logically seemed it should exist.

[kellerkindt]

Did anyone consider f(self: Weak<Self>)?
In my point of view this seems reasonable since Arcand Rc are already allowed.

[mark-i-m]

That seems unsafe since the self pointer could then be dropped while yhe method is executing, leaving self to dangle.

[kellerkindt]

That seems unsafe since the self pointer could then be dropped while yhe method is executing, leaving self to dangle.

Hmm? Am I missing something? (probably)
Isn't it just an ergonomic 'shortcut', being able to write f(self: Weak<Self>) instead of i.e. f(me: Weak<Self>)?
Thus it is still required to call self.upgrade() to access Self which isn't unsafe.

[mikeyhew]

Weak<Self> is in the same boat as *const/mut Self and NonNull<Self>: it's safe to call a method with it, but it doesn't implement Deref. Right now the unstable arbitrary_self_types feature only supports receiver types that implement Deref, plus *const/mut Self as a special case.

[kellerkindt]

Weak<Self> is in the same boat as *const/mut Self and NonNull<Self>: it's safe to call a method with it, but it doesn't implement Deref. Right now the unstable arbitrary_self_types feature only supports receiver types that implement Deref, plus *const/mut Self as a special case.

Thank you for the clarification!

Any chance Weak<Self> could be added to the additional special cases? (caution rust-compiler-newbie speaking) Looking at the current state, it seems that one would need to handle another speicialised case in receiver_is_valid or more precisely in Autoderef and it would need to be called like include_raw_pointers. From the naming scheme that feels odd(?) to me, because Weak<Self> can fail to "Deref" to Self (well, as you mentioned, like *const/mut Self ).

Maybe another reason for an ArbitrarySelfType-likish trait?

[mikeyhew]

Yeah, I think using a separate trait from Deref as @arielb1 suggested would be a good idea. There is a running list of types that would benefit from being method receivers, but can't implement Deref:

• *const/mut Self
• Weak<Self> (both Rc and Arc versions)
• NonNull<Self>
• RefCell<Self>, and by composition, Rc<RefCell<Self>>
• similarly, Mutex<Self>
• Option<Self> and Result<E, Self> (these ones are questionable)

Of the above list, all but Option and Result could produce a valid *const Self and therefore could implement a hypothetical UnsafeDeref trait. However, I'm starting to think that it would be better to have a trait that is specifically used for method lookup, whether or not it is decided thatOption<Self> should be usable as a method receiver or not.

Also, here is a potential name for the trait: MethodReceiver