Regression in method resolution through Deref

[sfackler]

This should compile but no longer does:

use std::cell::RefCell;

trait Foo {
    fn foo(&self) {}
}

impl<T> Foo for T where T: Clone {}

struct Bar;

impl Bar {
    fn foo(&self) {}
}

fn main() {
    let b = RefCell::new(Bar);
    b.borrow().foo()
}

test.rs:17:5: 17:21 error: the trait `core::clone::Clone` is not implemented for the type `core::cell::Ref<'_,Bar>`
test.rs:17     b.borrow().foo()
               ^~~~~~~~~~~~~~~~

Note that calling Bar.foo() directly works.

[sfackler]

cc @pcwalton I'm assuming this was introduced in #17464.

[bkoropoff]

This works fine if you change the call to:

    (*b.borrow()).foo()

#17464 seems to indicate that this is intentional.

[sfackler]

I don't think that applies. If Foo was actually implemented for RefMut<Bar>, then I'd expect that version of foo to be called, but it isn't.

[bkoropoff]

So it should back off and continue dereferencing instead? That would certainly be more intuitive and convenient behavior.

[sfackler]

That's what I would expect.

[bkoropoff]

This example seems to have worked before due to the special treatment of inherent methods. It's possible to construct an example that does not use them that fails in a similar way even prior to #17464:

trait Foo {
    fn foo(&self) { fail!() }
}

impl<'a, T> Foo for &'a T where T: Clone {}

trait UsableFoo {
    fn foo(&self) {}
}

struct Bar;

impl UsableFoo for Bar {}

fn main() {
    let b = &&Bar;
    b.foo();
}

test.rs:17:5: 17:12 error: the trait `core::clone::Clone` is not implemented for the type `Bar`
test.rs:17     b.foo();
               ^~~~~~~
error: aborting due to previous error

[sfackler]

Interesting! I wonder if this is a duplicate, then.

[aturon]

@sfackler I think the problem here is partly that we still don't have full trait reform -- in other words, the trait matching code sees the blanket impl and assumes it applies, without first checking the Clone bound. I believe that once @nikomatsakis lands the rest of trait reform, this problem will go away.

[nikomatsakis]

This is fallout (intensional) from the change to out inherent and trait methods on equal footing. However @aturon is correct that once I integrated conditional/multi dispatch into method resolution (next patch...) this should work again.

[sfackler]

Should I close this as a dup of the trait reform issue?

[nikomatsakis]

Closing as dup of #5527

[nikomatsakis]

I'm opening this and flagging as needs test. It works now.

[nikomatsakis]

(Also, PR #18224 adds a test)