Coherence problem when a generic has a default value from an associated type

[ohadravid]

I ran into a coherence problem where I think that the compiler is too strict, even though the code should be accepted according to the coherence rules (as far as I can tell).

For the code below, the compiler will report

 error[E0210]: type parameter `<MyString as somelib::ConvertTo>::Output` must be used as the type parameter for some local type ...`,

but passing O explicitly (for ItemOrConverted<MyString, String>) works.

The code:

/// In some other crate.
trait ConvertTo {
    type Output;
}


enum ItemOrConverted<T: ConvertTo, O = <T as ConvertTo>::Output> {
    Item(T),
    Converted(O),
}

/// In our crate.
struct MyString { }

impl ConvertTo for MyString {
    type Output = String;
}

impl PartialEq<MyString> for ItemOrConverted<MyString> {
    fn eq(&self, other: &MyString) -> bool {
        unimplemented!();
    }
}

I have a trait with an associated type (ConvertTo::Output, in the example below), and an enum which is generic over T: ConvertTo and generic over O with a default value for O based on the associated type from T's ConvertTo, all in one crate.

In another crate I have MyString, and I would like to implement PartialEq<MyString> for this enum over MyString, but without changing the default value for O. (Think impl PartialEq<MyString> for Cow<MyString>).

Is this really allowed "in theory" by the coherence rules or am I missing something?

(edit: Compiling with #![feature(re_rebalance_coherence)] works!)

Meta

rustc 1.37.0-nightly (37d001e 2019-05-29)
binary: rustc
commit-hash: 37d001e
commit-date: 2019-05-29
host: x86_64-pc-windows-msvc
release: 1.37.0-nightly
LLVM version: 8.0

[jonas-schievink]

(edit: Compiling with #![feature(re_rebalance_coherence)] works!)

In that case, maybe this should be closed in favor of #55437, which tracks the stabilization of that feature?

[ohadravid]

Yeah, sounds good.
Maybe I'll add a test for this? the implementation PR of #55437 doesn't seem to have one.

[jonas-schievink]

There's a few tests in #56145, but having a more diverse test set is always good, so feel free to add this!