Skip to content

Safe-to-use proc-macro-free self-referential structs in stable Rust.

License

Notifications You must be signed in to change notification settings

Voultapher/self_cell

Repository files navigation

github crates.io docs.rs

self_cell!

Use the macro-rules macro: self_cell! to create safe-to-use self-referential structs in stable Rust, without leaking the struct internal lifetime.

In a nutshell, the API looks roughly like this:

// User code:

self_cell!(
    struct NewStructName {
        owner: Owner,

        #[covariant]
        dependent: Dependent,
    }
    
    impl {Debug}
);

// Generated by macro:

struct NewStructName(...);

impl NewStructName {
    fn new(
        owner: Owner,
        dependent_builder: impl for<'a> FnOnce(&'a Owner) -> Dependent<'a>
    ) -> NewStructName { ... }
    fn borrow_owner<'a>(&'a self) -> &'a Owner { ... }
    fn borrow_dependent<'a>(&'a self) -> &'a Dependent<'a> { ... }
    [...]
    // See the macro level documentation for a list of all generated functions
    // https://docs.rs/self_cell/latest/self_cell/macro.self_cell.html#generated-api.
}

impl Debug for NewStructName { ... }

Self-referential structs are currently not supported with safe vanilla Rust. The only reasonable safe alternative is to expect the user to juggle 2 separate data structures which is a mess. The library solution ouroboros is expensive to compile due to its use of procedural macros.

This alternative is no_std, uses no proc-macros, some self contained unsafe and works on stable Rust, and is miri tested. With a total of less than 300 lines of implementation code, which consists mostly of type and trait implementations, this crate aims to be a good minimal solution to the problem of self-referential structs.

It has undergone community code review from experienced Rust users.

Fast compile times

$ rm -rf target && cargo +nightly build -Z timings

Compiling self_cell v0.9.0
Completed self_cell v0.9.0 in 0.2s

Because it does not use proc-macros, and has 0 dependencies compile-times are fast.

Measurements done on a slow laptop.

A motivating use case

use self_cell::self_cell;

#[derive(Debug, Eq, PartialEq)]
struct Ast<'a>(pub Vec<&'a str>);

self_cell!(
    struct AstCell {
        owner: String,

        #[covariant]
        dependent: Ast,
    }

    impl {Debug, Eq, PartialEq}
);

fn build_ast_cell(code: &str) -> AstCell {
    // Create owning String on stack.
    let pre_processed_code = code.trim().to_string();

    // Move String into AstCell, then build Ast inplace.
    AstCell::new(
        pre_processed_code,
        |code| Ast(code.split(' ').filter(|word| word.len() > 1).collect())
    )
}

fn main() {
    let ast_cell = build_ast_cell("fox = cat + dog");

    println!("ast_cell -> {:?}", &ast_cell);
    println!("ast_cell.borrow_owner() -> {:?}", ast_cell.borrow_owner());
    println!("ast_cell.borrow_dependent().0[1] -> {:?}", ast_cell.borrow_dependent().0[1]);
}
$ cargo run

ast_cell -> AstCell { owner: "fox = cat + dog", dependent: Ast(["fox", "cat", "dog"]) }
ast_cell.borrow_owner() -> "fox = cat + dog"
ast_cell.borrow_dependent().0[1] -> "cat"

There is no way in safe Rust to have an API like build_ast_cell, as soon as Ast depends on stack variables like pre_processed_code you can't return the value out of the function anymore. You could move the pre-processing into the caller but that gets ugly quickly because you can't encapsulate things anymore. Note this is a somewhat niche use case, self-referential structs should only be used when there is no good alternative.

Under the hood, it heap allocates a struct which it initializes first by moving the owner value to it and then using the reference to this now Pin/Immovable owner to construct the dependent inplace next to it. This makes it safe to move the generated SelfCell but you have to pay for the heap allocation.

See the documentation for a more in-depth API overview and advanced examples: https://docs.rs/self_cell

Installing

See cargo docs.

Running the tests

cargo test

cargo miri test

Related projects

Min required rustc version

By default the minimum required rustc version is 1.51.

There is an optional feature you can enable called "old_rust" that enables support down to rustc version 1.36. However this requires polyfilling std library functionality for older rustc with technically UB versions. Testing does not show older rustc versions (ab)using this. Use at your own risk.

The minimum versions are best-effort and may change with any new major release.

Contributing

Please respect the CODE_OF_CONDUCT.md when contributing.

Versioning

We use SemVer for versioning. For the versions available, see the tags on this repository.

Authors

See also the list of contributors who participated in this project.

License

This project is licensed under the Apache License, Version 2.0 - see the LICENSE.md file for details.