genrc

This crate provides alternatives to std::sync::Arc and std::rc::Rc which are (almost) drop-in replacements, but allow refcounted pointers to subobjects, like C++'s shared_ptr.

The main feature, which adds a surprising amount of flexibility: if you have an Rc<T>, and T contains some subobject of type U, then you can construct an Rc<U> that shares ownership with the original object by calling Rc::project().

    use genrc::rc::{Rc, Weak};
    let a: Rc<[i32; 3]> = Rc::new([1, 2, 3]);

    // convert the sized array into a slice
    let b: Rc<[i32]> = Rc::project(a, |x| &x[..]);

    // get a reference to one element of the array
    let c: Rc<i32> = Rc::project(b, |x| &x[1]);

There are also types RcBox<T> (and ArcBox<T>) that are returned from new_unique(), which take advantage of the fact that a newly created refcounted pointer is still unique, so can be used mutably.

Uses

Easier and safer initialization

You can use RcBox<Option<T>> to indicate not-yet-initialized types instead of the the various unsafe MaybeInit-related APIs in std::rc. After the object is initialized, you can use project to convert it to a plain Rc<T>:

    # use genrc::rc::{Rc, RcBox, Weak};

    // construct the object initially uninitialized
    let mut obj : RcBox<Option<i32>> = Rc::new_unique(None);

    // ... later ...
    // initialize the object
    obj.replace(5);

    // project to the inner value that we just created
    let obj : Rc<i32> = RcBox::project(obj, |x| x.as_ref().unwrap());

    assert_eq!(*obj, 5);

You can also create cyclic data structures without needing RefCell or new_cyclic:

    use genrc::rc::{Rc, RcBox, Weak};
    struct Node {
        edges: Vec<Weak<Node>>,
    }

    // Make a graph
    let mut graph: Vec<RcBox<Node>> = (0..5).map(|_| {
        Rc::new_unique(Node { edges: vec![] })
    }).collect();

    // Make some random edges in the graph
    for i in 0..5 {
        for d in 1..3 {
            let j = (i + d) % 5;

            let link = RcBox::downgrade(&graph[j]);
            graph[i].edges.push(link);
        }
    }

    // we still have unique handles on the nodes, so attempting to upgrade
    // weak pointers will fail.
    let p = graph[1].edges[0].clone();
    assert!(p.upgrade().is_none());

    // convert `RcBox` to a normal `Rc` with `into()`.
    let graph: Vec<Rc<Node>> = graph.into_iter().map(Into::into).collect();

    // now the weak pointers are valid - we've made a graph with (weak)
    // cycles, no unsafe or internal mutation required.
    assert!(Rc::ptr_eq(&graph[0].edges[1].upgrade().unwrap(), &graph[2]));

Static data

Unlike std, references can point to static data without copying, again using project():

    # use genrc::rc::Rc;
    static BIGBUF: [u8; 1024] = [1; 1024];

    let p: Rc<()> = Rc::new(());
    let p: Rc<[u8]> = Rc::project(p, |_| &BIGBUF[..]);

    assert!(std::ptr::eq(&BIGBUF[..], &*p));

So you can use Rc to keep track of possibly-owned, possibly-static data, similar to Cow.

Other stuff

Nightly Rust allocator_api Support

The allocator_api feature enables the unstable allocator API, allowing Rc and Arc to use custom allocators.

Rc::new_in returns an Rc<T, A>, with the allocator as part of the type. You can use Rc::erase_allocator() to hide the allocator from the type when that is desirable.

Lifetimes

Somewhat surprisingly,std::rc::Rc allows you to create an Rc pointing to a local variable. E.g. this is legal:

    use std::{cell::Cell, rc::Rc};
    let x = Cell::new(1);
    let y : Rc<&Cell<i32>> = Rc::new(&x);
    x.set(2);
    assert_eq!(y.get(), 2);

The type of such an Rc is Rc<&'a T>, where 'a is the lifetime of the referent, so the Rc can't outlive the referent.

genrc::Rc allows this too. But what if you use project() to turn Rc<&'a T> into an Rc<T> pointing to the same object? The latter type has nowhere for the lifetime 'a to go, so if allowed this would let the reference live too long and be a soundness bug.

To avoid this, the type Rcl<'a, T> adds a lifetime parameter to Rc.

(In fact Rc<T> is just an alias for Rcl<'static, T>, and Arc<T> is an alias for Arcl<'static, T>. And all of them are aliases for genrc::Genrc, which is generic over lifetime, referent type, atomicity, allocator, and uniqueness.)

To use project() such on a short-lived reference, you must use Rcl::project(), which returns an Rcl with a non-static lifetime.

    use genrc::rc::Rcl;

    // Imagine we have some JSON data that we loaded from a file
    // (or data allocated in an arena, etc)
    let bigdata : Vec<u8> = b"Not really json, use your imagination".to_vec();

    // buf points directly into `bigdata`, not a copy
    let buf : Rcl<[u8]> = Rcl::from_ref(&bigdata[..]);
    assert!(std::ptr::eq(&*buf, &bigdata[..]));

    let word : Rcl<[u8]> = Rcl::project(buf, |x| &x[4..10]);
    assert!(std::ptr::eq(&*word, &bigdata[4..10]));

Since the lifetime is usually inferred, in most cases Rcl works exactly like Rc. The main exception is in data types, where you may need it to explicitly specify a lifetime. E.g. if you want a field that's an Rc<str> where the string might be short-lived, you could write:

    use genrc::rc::Rcl;

    // Token in a parser where the buffer is an `Rc<str>`, and `text` can point
    // directly into the buffer. (Or `text` can point to owned data, e.g. for
    // unescaped strings, and callers generally don't have to care.)
    struct Token<'a> {
      some_data: u32,
      text: Rcl<'a, str>
    }

The lifetime parameter is also needed when type-erasing a custom allocator that has a lifetime, since Rc<T> also hides the allocator.

Other differences from std::sync::Arc and std::rc::Rc

Rc::from_box does not copy the object from the original box. Instead it takes ownership of the box as-is, with the counts in a separate allocation.

If you leak so many Rc objects that the refcount overflows, the std pointers will abort. genrc does not, because there is no abort() function in no_std.

Implicit conversion from Rc<T> to Rc<dyn Trait> is not supported, because that requires some unstable traits. However you can do the conversion explicitly with Rc::project. [TODO: support this behind a nightly-requiring feature.]

The std pointers have various MaybeUninit-related methods for initializing objects after allocation. That API isn't provided in Genrc, because you can accomplish the same thing entirely in safe code using Option and project:

    # use genrc::rc::{Rc, RcBox, Weak};
    // construct the object uninitialized
    let mut obj : RcBox<Option<i32>> = Rc::new_unique(None);

    // ... later ...
    // initialize the object
    obj.replace(5);

    // project to the inner value that we just created
    let obj : Rc<i32> = RcBox::project(obj, |x| x.as_ref().unwrap());

    assert_eq!(*obj, 5);

Unlike in std, Rc and Arc (and RcBox and ArcBox) share a single generic implementation. Rc<T> is an alias Genrc<'static, T, Nonatomic> and Arc<T> is an alias for Genrc<'static, T, Atomic>. This does make the documentation a little uglier, since it's all on struct Genrc instead of the actual types you normally care about.

std::rc::Rc::ptr_eq(a,b) returns true if a and b share the same allocation, which is the same as asking if they're equal pointers. But in genrc, these are two different questions: you can have pointers to two different subobjects from the same allocation, or pointers that came from two different allocations that are pointing to the same object! (E.g. they may have been projected to a static object). So here we have Rc::ptr_eq which is equivalent to std::ptr::eq(&*a, &*b), and Rc::root_ptr_eq which checks if the counts are shared.

from_raw and into_raw are not available because there may be no relationship between the returned pointer and the original allocation.

Differences from shared-rc

shared-rc is a very similar crate to this one; I would not have written this if I'd known that shared-rc already existed. That said, there are some differences:

• shared-rc uses the std versions of Arc and Rc under the hood, so it cannot support zero-alloc usage.

• shared-rc includes an Owner type param, with an explicit erase_owner method to hide it. genrc::arc::Arc always type-erases the owner. This saves one word of overhead in the pointer when a type-erased shared-rc is pointing to an unsized type. (e.g. shared_rc::rc::[u8] is 32 bytes, but genrc::rc::[u8] is 24.)

• genrc is generic over atomic vs. shared. shared-rc uses macros for that, which makes the rustdocs harder to read but "go to definition" easier to read.

Differences from rc-box

The rc-box crate adds a nice API around std Arc/Rc: immediately after creating one, you know you have the unique pointer to it, so put that in a wrapper type that implements DerefMut. This crate copies that API.

• Since rc-box is built on top of the std types, it would be unsafe to allow weak pointers to its RcBox types, so it cannot replace new_cyclic as in the graph example above.

• The implementation in genrc is generic over whether the pointer is unique or not (the UNIQ parameter to GenRc). This allows writing code generic over the uniqueness of the pointer, which may be useful for initialization (like the graph-creating example above, where the graph is a Vec<RcBox<Node>> during initialization, then gets converted to a Vec<Rc<Node>>.)

Related Crates

• shared-rc: Similar to this crate, but wraps the std versions of Arc and Rc rather than reimplementing them.
• rc-box: Known unique versions of Rc and Arc.
• erasable: Erase pointers of their concrete type.
• rc-borrow: Borrowed forms of Rc and Arc.

Todo

Implement the various Unsize traits behind a feature. (They require nightly even though they've been unchanged since 1.0, and are required to fully implement smart ptrs.)

Make behavior match std if count overflows

Richer custom allocator APIs. Currently only new_in and from_box are provided; should have try_* and support no_global_oom_handling.

More doc examples.

License

genrc is licensed under either the MIT or Apache 2.0 license, whichever you prefer.