Miss Dada  🏳️‍⚧️

rustc only inferred the second two steps. The rest was explicitly provided to it; we *told* the compiler "T: Send => Mutex<T>: Sync+Send"

relevant line: unsafe impl<T: ?Sized + Send> Sync for Mutex<T> { }

It really does seem like using `PhantomData` would have made more sense than opting out of the autotrait impl

PhantomData<&mut T> I suppose? Perhaps. I wonder how the variance of that works out. Actually, there's a chance variance is wrong there.

Yes. The variance shouldn't change since it'll still have the `&Mutex<T>` reference with the same lifetime.

No, but.. If the T itself has a lifetime, we need MutexGuard to be semantically equivalent to &mut T, not &T, so that it's invariant.

Wouldn't that imply that the current variance is probably broken?

Yes. I'm looking into it.