Returning an `impl Trait` in associated type position results in error at point-of-use instead of point-of-definition

[Nemo157]

Example (playground):

pub trait Foo {
    type Bar: std::ops::Add<Self::Bar>;
    fn bar(self) -> Self::Bar;
}

impl<T: std::ops::Add<T>> Foo for T {
    type Bar = T;
    fn bar(self) -> Self::Bar {
        self
    }
}

pub fn foo() -> impl Foo<Bar = impl std::ops::Sub<usize>> {
    5usize
}

fn main() {
    foo().bar() - 4usize;
}

This should error at the definition of foo since impl std::ops::Sub<usize> does not satisfy the std::ops::Add bound required by Foo::Bar. Instead it errors at the use-site in main, and if that line is deleted (e.g. if this is a pub fn in a library) there is no error.

error[E0277]: cannot add `impl std::ops::Sub<usize>` to `impl std::ops::Sub<usize>`
  --> src/main.rs:18:11
   |
18 |     foo().bar() - 4usize;
   |           ^^^ no implementation for `impl std::ops::Sub<usize> + impl std::ops::Sub<usize>`
   |
   = help: the trait `std::ops::Add` is not implemented for `impl std::ops::Sub<usize>`

[Nemo157]

@rustbot modify labels to +A-impl-trait

[aliemjay]

Fixed in 1.49.0

@rustbot label E-needs-test

[JohnTitor]

So the problem here was diagnostics, right? Since it now compiles fine, I think it's more of "hidden" than "fixed". Given that, I'm not sure if it's really worth adding a regression test, any thoughts?

[Nemo157]

Wait, how does this compile now? fn foo should be invalid, it's returning a type that doesn't implement the trait it claims to implement.

A slight modification shows that it's now somehow leaking additional trait impls that weren't declared on the impl Sub type (playground):

pub trait Foo {
    type Bar: std::ops::Add<Self::Bar, Output = Self::Bar> + std::fmt::Debug;
    fn bar(self) -> Self::Bar;
}

impl<T: std::ops::Add<T, Output = T> + std::fmt::Debug> Foo for T {
    type Bar = T;
    fn bar(self) -> Self::Bar {
        self
    }
}

pub fn foo() -> impl Foo<Bar = impl std::ops::Sub<usize>> {
    5usize
}

fn baz<T: Foo>(foo1: T, foo2: T) { dbg!(foo1.bar() + foo2.bar()); }

fn main() {
    baz(foo(), foo());
}

[oli-obk]

Yea, this is pretty much that one backcompat hack that I want to eliminate:

rust/compiler/rustc_trait_selection/src/traits/project.rs

Line 260 in 4f372b1

selcx.infcx().replace_opaque_types_with_inference_vars(

should get removed and cratered.

[JohnTitor]

@oli-obk thanks for the pointer! Are we ready to run crater or do you want to add some more patches before it?

[oli-obk]

Unless something comes up during the removal PR itself, this should be ready to go