Incorrect borrow checking

[strega-nil]

UPDATE: This now works in NLL land. We just need someone to add a test.

---

https://is.gd/O6RlWi

left.1.0 and left.0.0 should be separate borrows into the same structure.

Note that changing **left to *&mut**left makes it borrow check, and changing Box<Foo> to &'static mut Foo in the Foo struct also makes it borrow check. This makes me think it's something to do with the special casing of Box.

[arielb1]

This is intentional. rust-lang/rfcs#130.

[strega-nil]

@arielb1 no... it isn't. This is incorrect behavior. It should treat left.1.0 and left.0.0 as separate borrows.

(as seen by the fact that it works in the outer match statement)

[eddyb]

@ubsan I'm afraid @arielb1 is right, we really dumbed down Box.
Proof: box ref mut makes the error go away while matching on **(f: &mut Box<Foo>) reproduces it.

[strega-nil]

@eddyb then it's a bug with the specification (and the issue should be left open).

[strega-nil]

given that this works: https://is.gd/clzryh, rust should be able to figure out the difference between a moving deref and a non-moving; it can already figure out mutable vs. non-mutable.

[arielb1]

Sure, the equivalent of your code with a DerefMut type instead of a Box works:

#[derive(Clone)]
struct BoxFoo;
impl std::ops::Deref for BoxFoo {
    type Target = Foo;
    fn deref(&self) -> &Foo { panic!() }
}

impl std::ops::DerefMut for BoxFoo {
    fn deref_mut(&mut self) -> &mut Foo { panic!() }
}


#[derive(Clone)]
struct Foo(Option<BoxFoo>, Option<BoxFoo>);

fn test(f: &mut Foo) {
  match *f {
    Foo(Some(ref mut left), Some(ref mut right)) => match **left {
      Foo(Some(ref mut left), Some(ref mut right)) => panic!(),
      _ => panic!(),
    },
    _ => panic!(),
  }
}

fn main() {
}

The current borrowck is sort-of in maintenance mode, so this is unlikely to be fixed before MIR borrowck.

[pnkfelix]

Nominating. I suggest we retag this as C-enhancement rather than C-bug. I separately suggest that we prioritize it as P-medium. (It does not need to land as part of the initial MIR-borrowck / NLL work.)

[pnkfelix]

(looking at #23610, I am wondering: what's the difference between C-feature-request and C-enhancement? Should we collapse those two labels into one...?)

[nikomatsakis]

In the NLL / MIR borrow checker, we decided to just be smarter around Box. Therefore, the code now type-checks.

[nikomatsakis]

I think I would be happy to just add a test for this (with #![feature(nll)]) and close this issue.