RFC 2570: Linked list cursors

libs (types-libstd | collections)

Summary

Many of the benefits of linked lists rely on the fact that most operations (insert, remove, split, splice etc.) can be performed in constant time once one reaches the desired element. To take advantage of this, a Cursor interface can be created to efficiently edit linked lists. Furthermore, unstable extensions like the IterMut changes will be removed.

Motivation

From Programming Rust:

As of Rust 1.12, Rust’s LinkedList type has no methods for removing a range of elements from a list or inserting elements at specific locations in a list. The API seems incomplete.

Both of these issues have been fixed, but in different and incompatible ways. Removing a range of elements is possible though the unstable drain_filter API, and inserting elements in at specific locations in a list is possible through the linked_list_extras extensions to IterMut.

This motivates the need for a standard interface for insertion and deletion of elements in a linked list. An efficient way to implement this is through the use of "cursors". A cursor represents a position in a collection that can be moved back and forth, somewhat like a DoubleEndedIterator. However, mutable cursors can also edit the collection at their position.

A mutable cursor would allow for constant time insertion and deletion of elements and insertion and splitting of lists at its position. This would allow for simplification of the IterMut API and a complete LinkedList implementation.

Guide-level explanation

The cursor interface would provides two new types: Cursor and CursorMut. These are created in the same way as iterators.

With a Cursor one can seek back and forth through a list and get the current element. With a CursorMut One can seek back and forth and get mutable references to elements, and it can insert and delete elements before and behind the current element (along with performing several list operations such as splitting and splicing).

Lets look at where these might be useful.

Examples

This interface is helpful most times insertion and deletion are used together.

For example, consider you had a linked list and wanted to remove all elements which satisfy a certain predicate, and replace them with another element. With the old interface, one would have to insert and delete separately, or split the list many times. With the cursor interface, one can do the following:

fn remove_replace<T, P, F>(list: &mut LinkedList<T>, p: P, f: F)
    where P: Fn(&T) -> bool, F: Fn(T) -> T
{
    let mut cursor = list.cursor_front_mut();
    // move to the first element, if it exists
    loop {
        let should_replace = match cursor.peek_next() {
            Some(element) => p(element),
            None => break,
        };
        if should_replace {
            let old_element = cursor.remove_current().unwrap();
            cursor.insert_after(f(old_element));
        }
        cursor.move_next();
    }
}

This could also be done using iterators. One could transform the list into an iterator, perform operations on it and collect. This is easier, however it still requires much needless allocation.

For another example, consider code that was previously using IterMut extensions.

fn main() {
    let mut list: LinkedList<_> = (0..10).collect();
    let mut iter = list.iter_mut();
    while let Some(x) = iter.next() {
        if x >= 5 {
            break;
        }
    }
    iter.insert_next(12);
}

This can be changed almost verbatim to CursorMut:

fn main() {
    let mut list: LinkedList<_> = (0..10).collect();
    let mut cursor = list.cursor_front_mut() {
    while let Some(x) = cursor.peek_next() {
        if x >= 5 {
            break;
        }
        cursor.move_next();
    }
    cursor.insert_after(12);
}

In general, the cursor interface is not the easiest way to do something. However, it provides a basic API that can be built on to perform more complicated tasks.

Reference-level explanation

One gets a cursor the exact same way as one would get an iterator. The returned cursor would point to the "empty" element, i.e. if you got an element and called current you would receive None.

/// Provides a cursor to the first element of the list.
pub fn cursor_front(&self) -> Cursor<T>;

/// Provides a mutable cursor to the first element of the list.
pub fn cursor_front_mut(&mut self) -> CursorMut<T>;

/// Provides a cursor to the last element of the list.
pub fn cursor_back(&self) -> Cursor<T>;

/// Provides a mutable cursor to the last element of the list.
pub fn cursor_back_mut(&mut self) -> CursorMut<T>;

These would provide the following interface:

impl<'list, T> Cursor<'list, T> {
    /// Returns the cursor position index within the `LinkedList`.
    pub fn index(&self) -> Option<usize>;

    /// Move to the subsequent element of the list if it exists or the empty
    /// element
    pub fn move_next(&mut self);
    /// Move to the previous element of the list
    pub fn move_prev(&mut self);

    /// Get the current element
    pub fn current(&self) -> Option<&'list T>;
    /// Get the following element
    pub fn peek_next(&self) -> Option<&'list T>;
    /// Get the previous element
    pub fn peek_prev(&self) -> Option<&'list T>;
}

impl<'list T> CursorMut<'list, T> {
    /// Returns the cursor position index within the `LinkedList`.
    pub fn index(&self) -> Option<usize>;

    /// Move to the subsequent element of the list if it exists or the empty
    /// element
    pub fn move_next(&mut self);
    /// Move to the previous element of the list
    pub fn move_prev(&mut self);

    /// Get the current element
    pub fn current(&mut self) -> Option<&mut T>;
    /// Get the next element
    pub fn peek_next(&mut self) -> Option<&mut T>;
    /// Get the previous element
    pub fn peek_prev(&mut self) -> Option<&mut T>;

    /// Get an immutable cursor at the current element
    pub fn as_cursor<'cm>(&'cm self) -> Cursor<'cm, T>;

    // Now the list editing operations

    /// Insert `item` after the cursor
    pub fn insert_after(&mut self, item: T);
    /// Insert `item` before the cursor
    pub fn insert_before(&mut self, item: T);

    /// Remove the current item. The new current item is the item following the
    /// removed one.
    pub fn remove_current(&mut self) -> Option<T>;

    /// Insert `list` between the current element and the next
    pub fn splice_after(&mut self, list: LinkedList<T>);
    /// Insert `list` between the previous element and current
    pub fn splice_before(&mut self, list: LinkedList<T>);

    /// Split the list in two after the current element
    /// The returned list consists of all elements following the current one.
    pub fn split_after(&mut self) -> LinkedList<T>;
    /// Split the list in two before the current element
    pub fn split_before(&mut self) -> LinkedList<T>;
}

One should closely consider the lifetimes in this interface. Both Cursor and CursorMut operate on data in their LinkedList. This is why, they both hold the annotation of 'list.

The lifetime elision for their constructors is correct as

pub fn cursor_front(&self) -> Cursor<T>

becomes

pub fn cursor_front<'list>(&'list self) -> Cursor<'list, T>

which is what we would expect. (the same goes for CursorMut).

Since Cursor cannot mutate its list, current, peek_next and peek_prev all live as long as 'list. However, in CursorMut we must be careful to make these methods borrow. Otherwise, one could produce multiple mutable references to the same element.

The only other lifetime annotation is with as_cursor. In this case, the returned Cursor must borrow its generating CursorMut. Otherwise, it would be possible to achieve a mutable and immutable reference to the same element at once.

One question that arises from this interface is what happens if move_next is called when a cursor is on the last element of the list, or is empty (or move_prev and the beginning). A simple way to solve this is to make cursors wrap around this list back to the empty element. One could complicate the interface by having move return a bool, however this is unnecessary since current is sufficient to know whether the iterator is at the end of the list.

A large consequence of this new interface is that it is a complete superset of the already existing Iter and IterMut API. Therefore, the following two methods added to IterMut in the linked_list_extras features should be removed or depreciated:

Drawbacks

The cursor interface is rather clunky, and while it allows for efficient code, it is probably not useful outside of many use-cases.

One of the largest issues with the cursor interface is that it exposes the exact same interface of iterators (and more), which leads to unnecessary code duplication. However, the purpose of iterators seems to be simple, abstract and easy to use rather than efficient mutation, so cursors and iterators should be used in different places.

Rationale and alternatives

There are several alternatives to this:

  1. Implement cursors as a trait extending Iterator (see the cursors pseudo-rfc below)

Since the cursors are just an extension of iterators, it makes some sense to create them as a trait. However, I see several reasons why this is not the best.

First, cursors work differently than the existing Iterator extensions like DoubleEndedIterator. In a DoubleEndedIterator, if one calls next_back and then next, it should not return the same value, so unlike a cursor, a DoubleEndedIterator does not move back and forth throughout a collection.

Furthermore, while Iterator is a general interface for many collections, Cursor is very much specific to linked lists. In other collections such as Vec a cursor does not make sense. So it makes little sense to make a trait when it will only be used in one place.

  1. Using the IterMut linked list extensions

Insertion was added to IterMut in the linked_list_extras feature. Many of these features could be added to it just as well. But, this overcrowds IterMut with many methods that have nothing to do with iteration (such as deletion, splitting etc.) It makes sense to put these explicitly in their own type, and this can be CursorMut.

  1. Do not create cursors at all

Everything that cursors do can already be done, albeit in sometimes a less efficient way. Efficient code can be written by splitting linked lists often, and while this is a complicated way to do things, the rarity of the use case may justify keeping things how they are.

Prior art

This rust internals post describes an early attempt at making cursors. The language was in a different state when it was written (pre-1.0), so details have changed since then. But this describes several different approaches to making cursors and where they led.

Java (and other languages) tried to fix this by adding a remove function to their iterators. However, I feel this method would not be the best choice for Rust (even for specific IterMuts like those in LinkedList) since it diverges from the expected behaviour of iterators.

Discussion on the issue tracker about how this is currently managed with modifications to IterMut. The consensus seems to be that it is incomplete, and it is suggested to create a new Cursor and CursorMut types.

Unresolved questions

Will we keep both Iter and Cursor types? Implement one with another? I feel like they should be different things, but there is reason to consolidate them.

Should we implement this for more collections? It could make sense for other collections, such as trees and arrays, but the design would have to be reworked.