Add map function to Ref and MutRef of RefCell

[tbu-]

This enables one to return references to things inside the primary RefCell
content, like this code:

use std::cell::RefCell;
use std::cell::Ref;
struct ContainsVector {
vec: RefCell<Vec<u32>>,
}
impl ContainsVector {
fn new() -> ContainsVector {
ContainsVector { vec: RefCell::new(vec![0]) }
}
fn index<'a>(&'a self, i: uint) -> Ref<'a, &'a u32> {
self.vec.borrow().map(|v| &v[i])
}
fn index_mut<'a>(&'a self, i: uint) -> RefMut<'a, &'a mut u32> {
self.vec.borrow_mut().map(|v| &mut v[i])
}
}
fn main() {
let cv = ContainsVector::new();
**cv.index_mut(0) = 1;
cv.index(0);
}

[rust-highfive]

Warning

• These commits modify unsafe code. Please review it carefully!

[tbu-]

This pull request would change the semantics of RefCell, so I included the fixes for the standard library and rustc as an example.

Feedback would be appreciated! :)

[alexcrichton]

The Cell and RefCell types are quite core abstractions in the standard library, so modifying them shouldn't be taken lightly. Could you please elaborate as to why you have modified the semantics, as well as precisely how the semantics have changed? It's difficult to see what has changed with this new map function as well as what this means for users of RefCell without deeply understanding the patch itself.

I would recommend a description in lines with our breaking changes policy or perhaps a mini-RFC-style description.

[tbu-]

Motivation

Previously, you were unable to return a reference to something non-trivial inside of a RefCell. Consider you have this struct:

struct ContainsVector {
vec: RefCell<Vec<u32>>,
}

Then you're currently unable to return a reference an element of the vector, despite this having somewhat clear semantics (it would keep the RefCell borrowed).

Proposed changes

This proposal changes the direct usage semantics of the Refs and RefMuts into a RefCell, although it doesn't change the actual representation: If you borrow a RefCell, you now get a Ref<&T> out of it, instead of a Ref<T>, which means that it needs a double-dereference (**) to access the actual value. This enables changing the interior of such a references via the map function, which takes a closure |T| -> U (both types bound by lifetime) and transforms a Ref<T> to a Ref<U>.

Advantages

This should be the most simple change in order to properly return references to inner members of RefCells, by reusing the existing Ref type and changing the usage semantics slightly. The changed usage semantics do not change the assembly output:

Before (approximation):

struct Ref<T> { val: &T }
let ref: Ref<T>;
ref = ref;
*ref = *ref.val;

After:

struct Ref<T> { val: &T }
let ref: Ref<T>;
ref = ref;
*ref = ref.val;
**ref = *ref.val;

(So the assembly output still just contains one dereference.)

Disadvantages

Usage sites that do not benefit from auto-derefencing need to be changed. This includes e. g.

*refcell.borrow_mut() = value;

to

**refcell.borrow_mut() = value;

[tbu-]

@alexcrichton I believe the biggest argument against this change is that it changes a core type – however since we're still pre-1.0 it might be a bigger concern to get it right than to maintain backward-compatiblity.

[alexcrichton]

I don't think that this is safe, I checked this out and this code compiles (when it shouldn't)

fnmain(){let rc = RefCell::new(1u);let b:&uint = {let a = rc.borrow();*a };letmut c = rc.borrow_mut();**c = 3;println!("{}",*b);}

[tbu-]

You're right, this is wrong. I'll look into this.

[SimonSapin]

I’m making another attempt at this in #25747.