Confusing `impl Trait` limitation

[c410-f3r]

For some reason, impl Trait is allowed in a type's first mention but not in a where clause.

use core::ops::Add;

// Allowed
fn foo<T: Add<Output = impl Default>>(t: T) {}
fn bar<T: AsRef<impl Default>>(t: T) {}

// Not allowed
fn another_foo<T>(t: T)
where
    T: Add<Output = impl Default>,
{
}
fn another_bar<T>(t: T)
where
    T: AsRef<impl Default>,
{
}

Besides, the error message isn't very helpful.

`impl Trait` only allowed in function and inherent method return types, not in *****bound***** (which bound?)

[Rageking8]

@rustbot label +T-compiler +A-diagnostics D-confusing

[cjgillot]

There are 2 stages to fixing this issue. Everything happens in rustc_ast_lowering.

1. Make the diagnostic more explicit. The diagnostic is named MisplacedImplTrait. The term "bound" comes from the Display impl for ImplTraitPosition.
2. Allow impl-trait in where clauses. We'd need to investigate a bit why lowering marks impl-trait as disallowed in one case and not the other.

[jonathanCogan]

@rustbot claim

[estebank]

A wording change and an applicable suggestion to write something like this would be great.

[bombless]

What does "only allowed in function and inherent method return types" actually mean?
In the original RFC, it says "may only be written within the return type of a freestanding or inherent-impl function, not in trait definitions or any non-return type position", but it is not the case now per RFC1951 and related PR.
So I think this "inherent method return types" phrase should be removed entirely.
Further more what does "only allowed in function" mean anyway?

According to RFC1951 the major compilation error we deal with here should be "ERROR impl Trait not allowed within Fn trait sugar".

All RFC1522-related tests seems to be outdated, e.g. ui/impl-trait/where-allowed.rs.

I'm trying to address above-mentioned issues, so discussion above is what I have thought of so far, I'm writing it down in case I miss anything.

[jonathanCogan]

I think that the current message means “only allowed in function [return types] and inherent method return types” but I agree that it is no longer correct because impl Trait is allowed in the argument position, and there are plans to expand it to more locations.

[estebank]

TBH, it should suggest adding a type param.

[bombless]

https://github.com/rust-lang/rust/blob/master/src/test/ui/impl-trait/where-allowed.rs#L13
we are already allowing this form, it's just inconsistent.
So suggesting adding a type param might be a regression