Why is this type invariant?

help
1

I've been debugging some gnarly lifetime issues and have narrowed down the problem to a type being invariant when it should be covariant. I've reproduced it with a minimal example program:

// type TestType<'a> = Covariant<&'a u8>; // works if uncommented
type TestType<'a> = Invariant<&'a u8>; // doesn't work

// This struct is covariant.
struct Covariant<T>(T);

// This struct is invariant, but I don't know why.
struct Invariant<T: Item>(T::Type);
trait Item { type Type; }
impl<T> Item for T { type Type = T; }

// Verify that `TestType` is covariant.
fn test_covariance<'short, 'long: 'short>() {
    let long: TestType<'long> = make();
    let short: TestType<'short> = long;
}

fn make<T>() -> T { unimplemented!() }

Error:

error: lifetime may not live long enough
  --> src/main.rs:16:15
   |
15 | fn test_covariance<'short, 'long: 'short>() {
   |                    ------  ----- lifetime `'long` defined here
   |                    |
   |                    lifetime `'short` defined here
16 |     let long: TestType<'long> = make();
   |               ^^^^^^^^^^^^^^^ type annotation requires that `'short` must outlive `'long`
   |
   = help: consider adding the following bound: `'short: 'long`
   = note: requirement occurs because of the type `Invariant<&u8>`, which makes the generic argument `&u8` invariant
   = note: the struct `Invariant<T>` is invariant over the parameter `T`
   = help: see <https://doc.rust-lang.org/nomicon/subtyping.html> for more information about variance

(Explorer)

For reasons I don't understand, Invariant is invariant over T, even though it is intended to be covariant. Why does passing T through the Item trait make it invariant?

2

Associated types are invariant. Since there's no way to constrain the variance of associated types, the compiler has to be conservative and assume only invariance is safe.

6 Likes
3

The compiler could in theory be smarter than it currently is here. In the OP example, the blanket impl of Item constrains Type's variance for all T: Sized, but that information isn't used in variance calculation. (I asked about this on Zulip a few days ago)

1 Like