"conflicting associated type bounds when expanding trait alias", but there is no actual conflict

[Jules-Bertholet]

I tried this code:

#![feature(trait_alias)]

trait Foo {
    type Assoc;
}

trait Alias = Foo<Assoc = fn(&())>;

type Dyn = dyn Alias<Assoc = fn(&())>;

I expected to see this happen: It compiles, just as it would if fn(&()) were replaced with i32

Instead, this happened:

error: conflicting associated type bounds for `Assoc` when expanding trait alias
 --> src/lib.rs:9:12
  |
7 | trait Alias = Foo<Assoc = fn(&())>;
  |                   --------------- `Assoc` is specified to be `for<'a> fn(&'a ())` here
8 |
9 | type Dyn = dyn Alias<Assoc = fn(&())>;
  |            ^^^^^^^^^^---------------^
  |                      |
  |                      `Assoc` is specified to be `for<'a> fn(&'a ())` here

AIUI this is a consequence of #136458?

@rustbot label T-types F-trait_alias A-dyn-trait A-associated-items requires-nightly

Meta

rustc --version --verbose:

rustc 1.91.0-nightly (02c7b1a7a 2025-09-13)
binary: rustc
commit-hash: 02c7b1a7ac1d739663878030510508372e46f254
commit-date: 2025-09-13
host: x86_64-unknown-linux-gnu
release: 1.91.0-nightly
LLVM version: 21.1.1

[lukaslueg]

bisects to 2025-02-03, which includes #137406, which includes #136458

[fmease]

Semi-related: Since we don't normalize+unify but use ==, trait Alias = Foo<Assoc = ()>; type Dyn = dyn Alias<Assoc = <() as Identity>::Output>; (with trait Identity { type Output; } impl<T> Identity for T { type Output = T; }) also leads to a conflict. That might be a wontfix tho (unless we remove the == check entirely which is probably unsound) since we don't want to normalize during HIR ty lowering (since it would lead to certain classes of ill-formed types not getting detected).