regionck fails to link lifetime of ref bindings in certain situations

[nikomatsakis]

The recent attempts to improve hashmap seem to have exposed a bug in regionck where it fails to link the lifetimes of ref bindings in some situations.

A mostly minimized test case is here:

fn cb<'a,T>(x: |(&'a uint, &'a (Vec<&'static uint>, bool))| -> T) -> T {fail!()}

fn main() {
    cb(|(k, &(ref v, b))| (*k, v.clone(), b));
}

This test fails to compile but (I believe) should succeed. The problem is that the lifetime of the ref v binding inside the callback from map is not "linked" (in regionck terminology) to the lifetime 'a, and hence region inference infers a lifetime for v that is too small. It was not immediately obvious what the best way is to fix this -- the code around handling ref bindings seems a bit ... off in regionck, so I need to think about it some.

Perturbing the code in various ways can workaround the issue, which is bad for making minimized test cases but good for landing the hasmap patch in the meantime!

[sfackler]

I've run into (I think) another manifestation of this issue that forces the destructuring of a reference to a tuple from the arguments to the closure body: https://github.com/sfackler/rust-phf/blob/master/phf/src/lib.rs#L63

This works, but

        self.get_entry(key, |k| key == k).map(|e| {
            let &(_, ref v) = e;
            v
        })

this does not

        self.get_entry(key, |k| key == k).map(|&(_, ref v)| v)

[steveklabnik]

I think this is the original problem, updated to current syntax:

fn cb<'a,T>(x: Box<Fn((&'a i32, &'a (Vec<&'static i32>, bool))) -> T>) -> T {panic!()}

fn main() {
    cb(|(k, &(ref v, b))| (*k, v.clone(), b));
}

error:

hello.rs:4:28: 4:30 error: the type of this value must be known in this context
hello.rs:4     cb(|(k, &(ref v, b))| (*k, v.clone(), b));
                                      ^~
hello.rs:4:34: 4:39 error: the type of this value must be known in this context
hello.rs:4     cb(|(k, &(ref v, b))| (*k, v.clone(), b));
                                            ^~~~~
hello.rs:4:8: 4:45 error: mismatched types:
 expected `Box<core::ops::Fn((&i32, &(collections::vec::Vec<&'static i32>, bool))) -> _ + 'static>`,
    found `[closure@hello.rs:4:8: 4:45]`
(expected box,
    found closure) [E0308]
hello.rs:4     cb(|(k, &(ref v, b))| (*k, v.clone(), b));
                  ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
hello.rs:4:8: 4:45 help: run `rustc --explain E0308` to see a detailed explanation
error: aborting due to 3 previous errors

Actually, maybe this should be using something other than Box... ugh old closures.

[arielb1]

If you use a boxed closure, remember to box it:

fn cb<'a,T>(x: Box<Fn((&'a i32, &'a (Vec<&'static i32>, bool))) -> T>) -> T {panic!()}

fn main() {
    cb(Box::new(|(k, &(ref v, b))| (*k, v.clone(), b)));
}

This runs on modern rustc.

[steveklabnik]

Ah duh. Thanks!