Tracking Issue for `#![feature(phantom_variance_markers)]`

[jhpratt]

View all comments

Feature gate: #![feature(phantom_variance_markers)]

This is a tracking issue for phantom variance markers, which are identical to PhantomData but provide a self-documenting variance rather than contrived types such as PhantomData<fn() -> T> to indicate contravariance.

Public API

// in core::marker

pub struct PhantomCovariant<T: ?Sized>(/* ... */);
pub struct PhantomInvariant<T: ?Sized>(/* ... */);
pub struct PhantomContravariant<T: ?Sized>(/* ... */);
pub struct PhantomCovariantLifetime<'a>(/* ... */);
pub struct PhantomInvariantLifetime<'a>(/* ... */);
pub struct PhantomContravariantLifetime<'a>(/* ... */);

pub impl(self) trait Variance {}
impl<T: ?Sized> Variance for PhantomCovariant<T> {}
impl<T: ?Sized> Variance for PhantomInvariant<T> {}
impl<T: ?Sized> Variance for PhantomContravaiant<T> {}
impl<'a> Variance for PhantomCovariantLifetime<'a> {}
impl<'a> Variance for PhantomInvariantLifetime<'a> {}
impl<'a> Variance for PhantomContravaiantLifetime<'a> {}

pub fn variance<T: Variance>() -> T {}

// also the trait impls you would expect; identical to PhantomData

Steps / History

• ACP: Phantom variance ZSTs libs-team#488
• Implementation: Implement phantom variance markers #135807
• Final comment period (FCP)¹
• Stabilization PR

Unresolved Questions

• Should there be a single Lifetime type or one for each possible variance? The former would mean writing PhantomCovariant<Lifetime<'a>> instead of PhantomCovariantLifetime<'a>, with the API otherwise being identical.
• Should the values be directly constructable, eliminating the need for ::new()? This would presumably mean adding 4–6 lang items, depending on the previous bullet point.
• Are we happy exposing the (unstable) #[rustc_pub_transparent] on these, allowing indirect inclusion in #[repr(transparent)] types?

Footnotes

1. https://std-dev-guide.rust-lang.org/feature-lifecycle/stabilization.html ↩

[QuineDot]

As a note for future consideration: since lifetimes are not types, it's somewhat arbitrary to say whether e.g. &T is covariant in the lifetime or contravariant in the lifetime. Today saying covariant is the norm, but it used to be contravariant, and some people feel contravariant is the more sensible terminology.

Stabilizing all of the current names would lock in the choice. (Which may be fine, but should be a conscious decision.)

((The Invariant versions don't lock in the choice.))

[ChrisDenton]

Maybe we should try to come up with some new terminology if it's arbitrary anyway? I have heard people say they have to look up what covariant and contravariant mean (in terms of lifetime) every single time because it can be hard to keep in ones head.

[Sky9x]

Just ran into a slight hiccup with this feature: using these marker types in a #[repr(transparent)] struct throws the future-incompatibility lint repr_transparent_external_private_fields.

Example (playground):

#![feature(phantom_variance_markers)]

use std::marker::PhantomCovariantLifetime;
use std::ptr::NonNull;

#[repr(transparent)]
pub struct Foo<'a, T> {
    inner: NonNull<T>,
    _marker: PhantomCovariantLifetime<'a>,
}

warning: zero-sized fields in `repr(transparent)` cannot contain external non-exhaustive types
 --> src/lib.rs:9:5
  |
9 |     _marker: PhantomCovariantLifetime<'a>,
  |     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  |
  = warning: this was previously accepted by the compiler but is being phased out; it will become a hard error in a future release!
  = note: for more information, see issue #78586 <https://github.com/rust-lang/rust/issues/78586>
  = note: this struct contains `PhantomCovariantLifetime<'a>`, which contains private fields, and makes it not a breaking change to become non-zero-sized in the future.
  = note: `#[warn(repr_transparent_external_private_fields)]` on by default

warning: `playground` (lib) generated 1 warning

[jhpratt]

😔 some day we'll get #[repr(pub transparent)] or something to that effect. I didn't realize it was future incompat warning, though.

[lukas-code]

some day we'll get #[repr(pub transparent)]

We do have #[rustc_pub_transparent] and putting that on the Phantom* types should silence the warning.

[jhpratt]

Thanks @lukas-code — I'll get to that tonight 👍

[jhpratt]

Apparently I forgot to mark them #[repr(transparent)] at all, let alone publicly. Either way, #137054 is now open.

[RalfJung]

I think there's a possible confusion here that at the least should be discussed in the docs, or ideally avoided entirely by making this a language feature: even if I write PhantomCovariant<T>, the argument T might still be invariant! For instance:

struct S<T> {
  a: PhantomCovariant<T>,
  b: fn(T) -> T,
}

PhantomCovariant is really more like PhantomNotContravariant.