`TypeId` documentation should mention gotcha re covariant subtypes

[uazu]

A covariant subtype gets a different type-ID, but can be converted freely from one to the other. This means that taking a type-ID of a generic type argument in a type's new method is not guaranteed to give the same type-ID that the type will eventually be called with. The solution is to force the generic type argument to be invariant using PhantomData. See this bug in qcell crate.

It would be helpful if this were documented to avoid similar bugs occurring in other crates.

use std::any::TypeId;
fn main() {
    println!(
        "{:?}, {:?}",
        TypeId::of::<fn(&'_ ())>(),
        TypeId::of::<fn(&'static ())>(),
    );
}

[clubby789]

@rustbot label +A-docs

[hkBst]

@uazu, could you translate this: "This means that taking a type-ID of a generic type argument in a type's new method is not guaranteed to give the same type-ID that the type will eventually be called with." to code, so I have a better chance of understanding this issue?

Alternatively, since this seems quite technical and no-one has turned up to fix this, maybe propose your own fix?

[uazu]

I am not a type theorist! My understanding of the problem is probably incomplete.

However the way I understand it is as follows: Normally Rust won't let you silently convert types to other types (I'm talking about types themselves, not values). So if you take the type-id of a type, then later take the type-id of the same type specified separately, it will be the same. BUT that doesn't apply to lifetimes. Rust will happily let you change the lifetime on a type, and that will give a different type-id! So you can't count on the type-id staying the same if you have lifetimes involved. This is called covariance, as I understand it.

This caused an actual bug in the qcell crate, and it allowed a crate user to cause undefined behaviour (see the bug report mentioned above). The fix was to have a struct: struct Invariant<T>(fn(T) -> T);, which gets rid of the compiler's freedom to change the lifetimes within T. This would be used for example in type Id<'id> = PhantomData<Invariant<&'id ()>>; to create a type Id whose lifetime can't be changed.

The best code to demonstrate was provided by @steffahn in the bug report, which demonstrates bypassing TCell's type-id based check by converting the lifetime on a generic type using as.

So if I have a struct that records a type-id of its generic type argument in one call, that type-id might not actually correspond to the generic type argument with which the exact same instance of that struct is called later on. Rust allows the caller to change the generic type argument through covariance to another type with a different lifetime and different type-id, without the instance's knowledge (converting TCellOwner<T1> to TCellOwner<T2> in steffahn's example -- the instance still thinks it's a TCellOwner<T1>). It's like casting a Vec<bool> to a Vec<u8> without Vec knowing that anything has happened. Rust doesn't allow that case, but it does allow a similar case where only a lifetime changes. If Vec relied on the type-id to know what is going on, it would be in trouble. Vec doesn't use the type-id, but my TCellOwner does.

The change I would like to see to Rust is simply to document this, so that someone like me coming along in the future won't be tripped up by this. But I do not know how best to phrase it. Essentially the issue is that you cannot assume that the type-id of a generic type argument is always the same from one call to another if lifetimes are involved. Perhaps that would be enough of a warning, and maybe mention the Invariant work-around. But ideally someone who fully understands this should write the text.

I hope this makes things clearer.

[hkBst]

https://doc.rust-lang.org/nomicon/subtyping.html has an explanation about subtyping and variance, but in short: Something with a longer lifetime can always be used somewhere that only a shorter lifetime is necessary, like objects with a 'static lifetime can be used anywhere. Then if a generic type T using the 'static lifetime (T::<'static>) can be used anywhere that that same type with a shorter lifetime (T::<'a>) can be used, then T is covariant in that generic lifetime parameter.

Then I look at https://doc.rust-lang.org/nightly/std/any/struct.TypeId.html and your example code and it strikes me that the TypeId's produced are different and that they indeed compare (==) as different. And I see "A TypeId represents a globally unique identifier for a type." and "A TypeId is currently only available for types which ascribe to 'static, but this limitation may be removed in the future." in the docs. Certainly that "ascribe to 'static" is quite poetic, though it seems to refer to the bound on T in TypeId::of::<T: 'static + ?Sized>.

If I write "fn x<'a> (: &'a i32) -> bool {
TypeId::of::<&'static i32>() == TypeId::of::<&' i32>()
}"
it seems to always return true... sorry still not quite there, but perhaps closer?

[steffahn]

The relevant detail here is that subtyping also supports coercions from HRTB-like types (i.e. trait objects with higher-ranked bounds like dyn for<'a> Trait<'a>, or function pointer types like for<'a> fn(Type<'a>)) down to more concrete versions of the same type (e.g. dyn Trait<'b> or fn(Type<'b>) for concrete lifetimes 'b). This creates a subtyping relationship between multiple types T that each fulfill T: 'static. Finally, the fn pointer types used here contain some implicit higher-ranked-ness, because fn(&()) or fn(&'_ ()) are effectively syntactic sugar for for<'a> fn(&'a ()).

For a more illustrative and explanation, check out (the top part of) this comment by @danielhenrymantilla in a different issue:

• iter::Fuse is unsound with how specialization currently behaves around HRTB fn pointers #85863 (comment)

(it comes with a nice diagram 😻)

---

As @hkBst correctly observes, merely differentiating &'static i32 from &'_ i32 doesn’t make a difference in run-time type identity; though the better way towards this conclusion is to argue about lifetimes being erased during compilation, and the code example provided, with TypeId::of::<&'_ i32>, doesn’t demonstrate anything, as the T: 'static bound of TypeId::of makes the compiler infer the &'_ i32 here as &'static i32, anyway.

As a TL;DR, the choice of some lifetime 'a vs 'b doesn’t really matter (and for creating a TypeId, you’d usually need to use 'static everywhere anyway, as far as concrete lifetimes are concerned), but the choice between different versions of higher-ranked lifetimes, (the thing you can get with for<'a> … syntax) can make a difference in TypeId.

---

As a fun additional fact: in the context of higher-ranked lifetimes, you can even have different types that are mutual subtypes of each other. E.g. these two types:

type T1 = for<'a, 'b> fn(&'a (), &'b ());
type T2 = for<'c> fn(&'c (), &'c ());

(playground, showing the subtyping and TypeIds)

[hkBst]

So if I understand correctly, using two types where S is a subtype of T and so can be used in its place, if you have a covariant parameter in some generic type G, then TypeId::of::<G<S>> would be different from TypeId::of::<G<T>>, but in other code you could freely take a G<S> and treat it as a G<T>. This may then lead to method confusion, where changing the type would lead to method calls being resolved to different impl blocks, and code from two different impl blocks being mixed. I can see how that is a problem of you rely on other methods in your impl block to behave in a certain way to uphold correctness or safety invariants.

[uazu]

For my case, it is the same impl block, but I'm relying (for safety invariants) on the type-id staying the same through all calls to the struct methods. That was obviously an error on my part, since the caller can freely change the type (within the subtype restrictions), and hence the type-id. This was quite unexpected, and lead to the bug. So some form of warning in the docs would be very helpful.

[hkBst]

Different instantiations of the same generic impl block?

I think we should add an example like this, to showcase the danger.

[uazu]

Yes, it must be different instantiations of the same generic impl block. I tried it in the playground to verify that that was occurring, and it is (in playground):

use std::any::TypeId;
use std::marker::PhantomData;

// T1 subtype of T2, both 'static
type T1 = fn(&());
type T2 = fn(&'static ());

struct Test<T> {
    phantomdata: PhantomData<T>,
}

impl<T: 'static> Test<T> {
    fn new() -> Self {
        println!("Created with type-id {:?}", TypeId::of::<T>());
        Self { phantomdata: PhantomData }
    }

    fn test(&self) {
        println!("Type-id is now {:?}", TypeId::of::<T>());
    }
}

fn main() {
    let test = Test::<T1>::new() as Test<T2>;
    test.test();
}

I think an example along these lines would definitely help explain the danger.

(In my case the type-id check occurs only once, and it's a lifetime comparison done by the compiler on another method that causes the issue, but if I had repeated the type-id check on that method, I would have been able to see the divergence between my first check and the second.)

[hkBst]

@uazu I think this is an excellent example showing that you cannot rely on the TypeId for safety. I think I have enough information now to attempt some doc fix, if you still prefer that, probably using this example mostly verbatim. But since you have now done a lot of the preparatory work already, perhaps you feel you might as well take the final step? I'm happy to go either way!

[uazu]

Please go ahead with making the doc changes with this example or whatever you feel is appropriate. I think it's better done by someone with more familiarity with the internal Rust organization processes. I'd consider it a "win" if it leads to people avoiding this class of bugs in the future.

[hkBst]

@uazu Alright. I'm looking with fresh eyes at the bug report and code in qcell and it is making a lot more sense suddenly :D. I'm considering upgrading your example to something that is perhaps a bit closer to one of the cell types in qcell. Which cell type do you think would be most suited for that, in that the result would be the simplest to understand?

[hkBst]

Hmm, seems QCell is the simplest, but it uses an allocation for uniqueness, so it does not seem useful here, so maybe TCell.