`std::marker::Freeze` disallows blanket impl since it "could be implemented on `Cell` and `UnsafeCell` in the future"

[ds84182]

I tried this code:

#![feature(freeze)]
use std::{cell::Cell, marker::Freeze};

unsafe trait IntoFreeze {
  type Frozen: Freeze;
}

unsafe impl<T: Freeze> IntoFreeze for T {
  type Frozen = T;
}

unsafe impl<T: IntoFreeze> IntoFreeze for Cell<T> {
  type Frozen = T::Frozen;
}

I expected to see this happen: Code should compile since Freeze will never be implemented for Cell and UnsafeCell.

Instead, this happened:

error[E0119]: conflicting implementations of trait `IntoFreeze` for type `Cell<_>`
  --> src/lib.rs:12:1
   |
8  | unsafe impl<T: Freeze> IntoFreeze for T {
   | --------------------------------------- first implementation here
...
12 | unsafe impl<T: IntoFreeze> IntoFreeze for Cell<T> {
   | ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ conflicting implementation for `Cell<_>`
   |
   = note: upstream crates may add a new impl of trait `std::marker::Freeze` for type `std::cell::Cell<_>` in future versions

For more information about this error, try `rustc --explain E0119`.

Meta

rustc --version --verbose:

1.85.0-nightly (2024-11-30 7442931d49b199ad0a1c)

[jieyouxu]

I think this is intentional, but cc @oli-obk just in case.

[oli-obk]

We explicitly impl<T: ?Sized> !Freeze for UnsafeCell<T> {} which I assumed should allow at least "overlapping" impls for UnsafeCell, but that may not be true for auto traits? Or does reasoning about negative impls require another feature and only work on nightly?

[zachs18]

Note that with #![feature(with_negative_coherence)], the UnsafeCell impl is allowed (due to the explicit impl !Freeze for UnsafeCell<_>), but the Cell impl is still disallowed (IIUC since it gets its "!Freeze-ness" via auto trait impl instead of an explicit negative impl): playground link.

---

I'm not sure if the original code not being allowed is related to Freeze being an auto trait, and don't know of any non-auto traits with negative impls to compare to off the top of my head. Maybe the impl !Fn for &str? But Fn is #[fundamental] which may also affect things. I misremembered, that is unrelated to negative impls; Fn being #[fundamental] is why there doesn't need to be an impl !Fn for &str source

Hmm, though that would indicate that Freeze being #[fundamental] would allow the impls in the OP? but would also make adding or removing interior mutability from a type be a semver-breaking change, but IIRC thas already somewhat of an issue due to consts?