breakage due to a bugfix in rustc

[lcnr]

Your crate will unfortunately stop compiling once rust-lang/rust#121848 is merged. It unintentionally relies on an unsound bug of the type system which we are now fixing: rust-lang/rust#114061.

derive-visitor/derive-visitor/src/lib.rs

Lines 538 to 544 in 39dc342

	impl<T> Drive for T
	where
	T: 'static,
	// HRTB, because this needs to be true for late-bound lifetimes,
	// as both &self and &visitor have anonymous lifetimes.
	for<'a> &'a T: IntoIterator,
	for<'a> <&'a T as IntoIterator>::Item: DerefAndDrive,

This can overlap with the following impl

derive-visitor/derive-visitor/src/lib.rs

Lines 566 to 568 in 39dc342

	impl<T> Drive for Box<T>
	where
	T: Drive,

causing Box<Local> to implement Drive using two separate overlapping impls in the following example:

use std::iter;
struct Local;
impl<'a> IntoIterator for &'a Box<Local> {
    type IntoIter = iter::Once<&'a ()>;
    type Item = &'a ();
    fn into_iter(self) -> Self::IntoIter {
        iter::once(&())
    }
}

impl derive_visitor::Drive for Local {
    fn drive<V>(&self, _: &mut V) {}
}

fn impls_drive<T: derive_visitor::Drive>() {}

fn main() {
    impls_drive::<Box<Local>>();
}

The underlying issue is that, during the coherence overlap check, we consider for<'a> <&'a Box<_> as IntoIterator>::Item: DerefAndDrive to never apply, even though as shown above, it is possible to writean impl so that it does.

I think that there is unfortunately no way to fix the new error except by remove the blanket impl. I apologize for the this.

[lovasoa]

Do you mean that this library will be forever broken without the possibility to implement a fix ?

[lcnr]

The impl as is cannot be supported given that it overlaps with other impls. I expect there to be a way to rewrite this to work mostly the same though.

[lovasoa]

Maybe you could open a PR, then?

[nikis05]

@lcnr thanks for reporting this.

I think we should go back to macro-based impls for concrete collection types. @lovasoa @badicsalex thoughts?