Generic associated impl trait type with bounds on self formatted to broken code

[tim3z]

Describe the bug

When specifying an impl trait type alias for a generic associated type with a lifetime bound on Self, rustfmt will remove the bound, causing a compiler error, i.e.:

impl SomeTrait for SomeType {
    type SomeGAT<'a> where Self: 'a = impl SomeOtherTrait;
}

will be changed to

impl SomeTrait for SomeType {
    type SomeGAT<'a> = impl SomeOtherTrait;
}

To Reproduce

The following code compiles, but when running rustfmt on it it will remove the Self bound where Self" 'a in the implementation for Foo leading to a compiler error:

#![allow(incomplete_features)]
#![feature(generic_associated_types)]
#![feature(min_type_alias_impl_trait)]

struct Foo<'b>(&'b ());

trait Iterable {
    type Iter<'a>
    where
        Self: 'a;
    fn iter(&self) -> Self::Iter<'_>;
}

impl<'b> Iterable for Foo<'b> {
    type Iter<'a> where Self: 'a = impl Iterator + 'a;

    fn iter(&self) -> Self::Iter<'_> {
        std::iter::once(self.0)
    }
}

Expected behavior

The above code compiles and should not be formatted so it doesn't compile.

Meta

• rustfmt 1.4.37-nightly (014026d 2021-07-19)
• From where did you install rustfmt?: rustup
• How do you run rustfmt: rustfmt, sublime RustFmt package

[calebcartwright]

Thanks for the report but I'm going to close this as a duplicate of various prior issues. It's been fixed in source but not pulled into a nightly release yet.

[tim3z]

Thanks for the response. I looked into other related closed issues (#4257 #4315 #4322) but they are from July 2020 and the examples from these issues are all formatted correctly with the current rustfmt version. So while this seems very similar, I don't think that this is a duplicate. The combination with the impl trait type alias seems to cause another problem.

[calebcartwright]

Ah I see now. The way the snippets were structured in the issue description was a bit of a misdirection. Would be helpful to update extend that initial snippet to a minimal repro itself (within the impl) since it's not reproducible as is.

[tim3z]

Updated initial example