unsoundness relating to WF requirements on trait object types

[nikomatsakis]

So @scalexm pointed me at this curious example:

trait Animal<X>: 'static { }

fn foo<Y, X>() where Y: Animal<X> + ?Sized {
    // `Y` implements `Animal<X>` so `Y` is 'static.
    baz::<Y>()
}

fn bar<'a>(arg: &'a i32) {
    foo::<Animal<&'a i32>, &'a i32>()
    
    // If you uncomment this line, you will get an error.
    //baz::<Animal<&'a i32>>()
}

fn baz<T: 'static + ?Sized>() {
}

fn main() {
    let a = 5;
    bar(&a);
}

There's a lot going on here, but the key point is that, in the call to foo::<Animal<&'a i32>, &'a i32>, we assume that Animal<&'a i32>: Animal<&'a i32>. This is reasonable, since it's a trait object type -- however, it is a malformed trait object type, and hence one where no impl could ever exist (i.e., we could never make an instance of this type). Interestingly, we are not smart enough to extend this to the case where the 'static bound is applied directly (hence the commented out call to baz::<Animal<&'a i32>>.

At the very least, this is inconsistent behavior, but I feel like there is an unsoundness lurking here. What makes me nervous is that we say that a projection type <P0 as Trait<P1...Pn>>::Foo outlives 'a if Pi: 'a for all i. I am worried that you could use an impl like this:

trait Projector { type Foo; }

impl<X> Projector for Animal<X> {
  type Foo = X;
}

to prove then that <Animal<&'a i32> as Projector>::Foo outlives 'static (say). I have not yet quite managed to weaponize this, but I got pretty close -- that's a gist in which we invoke a function is_static<U: 'static>(u: U) where U will be of type &'a i32 when monomophized. If rustc were as smart as chalk, I suspect we could easily craft a transmute of some kind to &'static i32. There may be a way to do it within current rustc too.

I'm not sure what's the best fix here. I suspect we should require that trait object types meet the WF requirements declared on the trait; and yet, I think the reason we did not is because we don't know the Self type, which means we can't fully validate that. =) This also feels a mite inconsistent with implied bounds. (Perhaps most likely is that we could tweak the WF rules for trait objects to specifically target lifetime bounds, but I have to think about that, feels .. hacky?)

We could tweak the notion of what type parameters are constrained -- or at least which may appear in a projection. e.g., we could disallow that impl of Projector for Animal<X>, because X would not be considered "sufficiently" constrained to appear in type Foo = .... Backwards incompatible, obviously, and feels unfortunate.

---

This causes actual unsoundness, as discovered by @scalexm in #44454 (comment);

use std::any::Any;

trait Animal<X>: 'static { }

trait Projector {
  type Foo;
}

impl<X> Projector for dyn Animal<X> {
  type Foo = X;
}

fn make_static<'a, T>(t: &'a T) -> &'static T {
    let x: <dyn Animal<&'a T> as Projector>::Foo = t;
    let any = generic::<dyn Animal<&'a T>, &'a T>(x);
    any.downcast_ref::<&'static T>().unwrap()
}

fn generic<T: Projector + Animal<U> + ?Sized, U>(x: <T as Projector>::Foo)
    -> Box<dyn Any>
{
    make_static_any(x)
}

fn make_static_any<U: 'static>(u: U) -> Box<dyn Any> {
    Box::new(u)
}

fn main() {
    let a = make_static(&"salut".to_string());
    println!("{}", *a);
}

cc @arielb1 @aturon

[nikomatsakis]

Actually, thinking about this more, the problem is maybe not quite how I was thinking of it. That is, it seems like the problem lies in the definition of object safety. That is, if you imagine that you had to write the impl for Animal<X> by hand, it would look like:

impl<X> Animal<X> for Animal<X> { ... }

and this would fail the well-formedness checks because Animal<X>: 'static does not hold, at least not without relying on the impl we are writing =).

The goal of the object safety rules are to ensure such an impl could be written.

Maybe the rules of object safety have to be more conditional -- i.e., Animal<X> is object safe if X: 'static.

[scalexm]

Here is an actual unsoundness, using std::any::Any:
https://play.rust-lang.org/?gist=1385ed4e051354b305ea2d2413f8724f&version=stable

[arielb1]

This looks similar to #27675 - in both cases we have a trait object that is not WF.

I think a reasonable way to solve this would be to have a trait object predicate depend on a WF(Self) predicate, so we basically have the rule

'λ₀...'λₙ lifetimes
τ₀...τₙ types
Trait object-safe
~~WF(for<...> Trait<'λ₀...'λₙ, τ₀...τₙ>) - this part is new~~
WF(for<...> Trait<'λ₀...'λₙ, τ₀...τₙ>: Trait<'λ₀...'λₙ, τ₀...τₙ>) - this part is new
----------------
for<...> Trait<'λ₀...'λₙ, τ₀...τₙ>: for<...> Trait<'λ₀...'λₙ, τ₀...τₙ>

[arielb1]

The "this part is new" part is actually more subtle than I realized.

Object types (aka dyn Trait<'λ₀...'λₙ, τ₀...τₙ>, to avoid confusion) are actually a funny way of writing

exists<σ> (σ, for<...> WF(dyn Trait<'λ₀...'λₙ, τ₀...τₙ>: Trait<'λ₀...'λₙ, τ₀...τₙ>) → σ: Trait<'λ₀...'λₙ, τ₀...τₙ>)

And the idea is that if the trait is object-safe, we can write a "compiler impl"

impl<'λ₀...'λₙ, τ₀...τₙ> Traitl<'λ₀...'λₙ, τ₀...τₙ> for dyn Trait<'λ₀...'λₙ, τ₀...τₙ> {
    // ...
}

If we want to write such an impl, we need the where-clauses on the trait (both associated types and others) to hold. For supertraits, this follows by obvious induction. For other constraints, this is something that won't hold by itself.

Therefore, we probably need to add these other predicates as conditions on the impl, and have

impl<'λ₀...'λₙ, τ₀...τₙ> Trait<'λ₀...'λₙ, τ₀...τₙ> for dyn Trait<'λ₀...'λₙ, τ₀...τₙ>
    where WF(Trait<'λ₀...'λₙ, τ₀...τₙ>: dyn Trait<'λ₀...'λₙ, τ₀...τₙ>)
{
    // ...
}

[nikomatsakis]

I agree that it is the same as #27675. As we said on IRC, I think that the current fix is roughly as I described here -- in particular, when we check that dyn Trait<P...>: Trait<P...>, we need to also check that the where-clauses declared on Trait hold for dyn Trait<P...>. This subsumes some of the object safety checks (e.g. the search for where Self: Sized).

[bstrie]

I agree that it is the same as #27675.

Shall we close this as a dupe, then?