RFC 2300: self-in-typedefs

lang (data-types | syntax | resolve)

Summary

The special Self identifier is now permitted in struct, enum, and union type definitions. A simple example struct is:

enum List<T>
where
    Self: PartialOrd<Self> // <-- Notice the `Self` instead of `List<T>`
{
    Nil,
    Cons(T, Box<Self>) // <-- And here.
}

Motivation

Removing exceptions and making the language more uniform

The contextual identifier Self can already be used in type context in cases such as when defining what an associated type is for a particular type as well as for generic parameters in impls as in:

trait Foo<T = Self> {
    type Bar;

    fn wibble<U>() where Self: Sized;
}

struct Quux;

impl Foo<Self> for Quux {
    type Bar = Self;

    fn wibble<U>() where Self: Sized {}
}

But this is not currently possible inside both fields and where clauses of type definitions. This makes the language less consistent with respect to what is allowed in type positions than what it could be.

Principle of least surprise

Users, just new to the language and experts in the language alike, also have a reasonable expectations that using Self inside type definitions is in fact already possible. Users may have and have these expectations because Self already works in other places where a type is expected. If a user attempts to use Self today, that attempt will fail, breaking the users intuition of the languages semantics. Avoiding that breakage will reduce the paper cuts newcomers face when using the language. It will also allow the community to focus on answering more important questions.

Better ergonomics with smaller edit distances

When you have complex recursive enums with many variants and generic types, and want to rename a type parameter or the type itself, it would make renaming and refactoring the type definitions easier if you did not have to make changes in the variant fields which mention the type. This can be helped by IDEs to some extent, but you do not always have such IDEs and even then, the readability of using Self is superior to repeating the type in variants and fields since it is a more visual cue that can be highlighted for specially.

Encouraging descriptively named types, type variables, and more generic code

Making it simpler and more ergonomic to have longer type names and more generic parameters in type definitions can also encourage using more descriptive identifiers for both the type and the type variables used. It may also encourage more generic code altogether.

Guide-level explanation

An Obligatory Public Service Announcement: When reading this RFC, keep in mind that these lists are only examples. Always consider if you really need to use linked lists!

We will now go through a few examples of what you can and can't do with this RFC.

Simple example

Let's look at a simple cons-list of u8s. Before this RFC, you had to write:

enum U8List {
    Nil,
    Cons(u8, Box<U8List>)
}

But with this RFC, you can now instead write:

enum U8List {
    Nil,
    Cons(u8, Box<Self>) // <-- Notice 'Self' here
}

If you had written this example with Self without this RFC, the compiler would have greeted you with:

error[E0411]: cannot find type `Self` in this scope
 --> src/main.rs:3:18
  |
3 |     Cons(u8, Box<Self>) // <-- Notice 'Self' here
  |                  ^^^^ `Self` is only available in traits and impls

With this RFC, the compiler will never do so.

This new way of writing with Self can be thought of as literally desugaring to the way it is written in the example before it. This also extends to generic types (non-nullary type constructors) that are recursive.

With generic type parameters

Continuing with the cons lists, let's take a look at how the canonical linked-list example can be rewritten using this RFC.

We start off with:

enum List<T> {
    Nil,
    Cons(T, Box<List<T>>)
}

With this RFC, the snippet above can be rewritten as:

enum List<T> {
    Nil,
    Cons(T, Box<Self>) // <-- Notice 'Self' here
}

Notice in particular how we used just Self for both U8List and List<T>. This applies to types with any number of parameters, including those that are parameterized by lifetimes.

Examples with lifetimes

An example of this can be seen in the following cons list:

enum StackList<'a, T: 'a> {
    Nil,
    Cons(T, &'a StackList<'a, T>)
}

which is rewritten with this RFC as:

enum StackList<'a, T: 'a> {
    Nil,
    Cons(T, &'a Self) // <-- Still using just 'Self'
}

Structs and unions

You can also use Self in structs as in:

struct NonEmptyList<T> {
    head: T,
    tail: Option<Box<NonEmptyList<T>>>,
}

which is written with this RFC as:

struct NonEmptyList<T> {
    head: T,
    tail: Option<Box<Self>>,
}

This also extends to unions.

where-clauses

In today's Rust, it is possible to define a type such as:

struct Foo<T>
where
    Foo<T>: SomeTrait
{
    // Some fields..
}

and with some impls:

trait SomeTrait { ... }

impl SomeTrait for Foo<u32> { ... }
impl SomeTrait for Foo<String> { ... }

this idiom bounds the types that the type parameter T can be of but also avoids defining an Auxiliary trait which one bound T with as in:

struct Foo<T: Auxiliary> {
    // Some fields..
}

You could also have the type on the right hand side of the bound in the where clause as in:

struct Bar<T>
where
    T: PartialEq<Bar<T>>
{
    // Some fields..
}

with this RFC, you can now redefine Foo<T> and Bar<T> as:

struct Foo<T>
where
    Self: SomeTrait // <-- Notice `Self`!
{
    // Some fields..
}

struct Bar<T>
where
    T: PartialEq<Self> // <-- Notice `Self`!
{
    // Some fields..
}

This makes the bound involving Self slightly more clear.

When Self can not be used

Consider the following small expression language:

trait Ty { type Repr: ::std::fmt::Debug; }

#[derive(Debug)]
struct Int;
impl Ty for Int { type Repr = usize; }

#[derive(Debug)]
struct Bool;
impl Ty for Bool { type Repr = bool; }

#[derive(Debug)]
enum Expr<T: Ty> {
    Lit(T::Repr),
    Add(Box<Expr<Int>>, Box<Expr<Int>>),
    If(Box<Expr<Bool>>, Box<Expr<T>>, Box<Expr<T>>),
}

fn main() {
    let expr: Expr<Int> =
        Expr::If(
            Box::new(Expr::Lit(true)),
            Box::new(Expr::Lit(1)),
            Box::new(Expr::Add(
                Box::new(Expr::Lit(1)),
                Box::new(Expr::Lit(1))
            ))
        );
    println!("{:#?}", expr);
}

You may perhaps reach for this:

#[derive(Debug)]
enum Expr<T: Ty> {
    Lit(T::Repr),
    Add(Box<Self>, Box<Self>),
    If(Box<Self>, Box<Self>, Box<Self>),
}

But you have now changed the definition of Expr semantically. The changed semantics are due to the fact that Self in this context is not the same type as Expr<Int> or Expr<Bool>. The compiler, when desugaring Self in this context, will simply substitute Self with what it sees in Expr<T: Ty> (with any bounds removed).

You may at most use Self by changing the definition of Expr<T> to:

#[derive(Debug)]
enum Expr<T: Ty> {
    Lit(T::Repr),
    Add(Box<Expr<Int>>, Box<Expr<Int>>),
    If(Box<Expr<Bool>>, Box<Self>, Box<Self>),
}

Types of infinite size

Consider the following example:

enum List<T> {
    Nil,
    Cons(T, List<T>)
}

If you try to compile it this today, the compiler will greet you with:

error[E0072]: recursive type `List` has infinite size
 --> src/main.rs:1:1
  |
1 | enum List<T> {
  | ^^^^^^^^^^^^ recursive type has infinite size
2 |     Nil,
3 |     Cons(T, List<T>)
  |             -------- recursive without indirection
  |
  = help: insert indirection (e.g., a `Box`, `Rc`, or `&`) at some point to make `List` representable

If we use the syntax introduced by this RFC as in:

enum List<T> {
    Nil,
    Cons(T, Self)
}

you will still get an error since it is fundamentally impossible to construct a type of infinite size. The error message would however use Self as you wrote it instead of List<T> as seen in this snippet:

error[E0072]: recursive type `List` has infinite size
 --> src/main.rs:1:1
  |
1 | enum List<T> {
  | ^^^^^^^^^^^^ recursive type has infinite size
2 |     Nil,
3 |     Cons(T, Self)
  |             ----- recursive without indirection
  |
  = help: insert indirection (e.g., a `Box`, `Rc`, or `&`) at some point to make `List` representable

Teaching the contents of this RFC

When talking about and teaching recursive types in Rust, since it is now possible to use Self, the ability to use Self in this context should be taught along side those types. An example of where this can be introduced is the "Learning Rust With Entirely Too Many Linked Lists" guide.

Reference-level explanation

The identifier Self is (now) allowed in type contexts in fields of structs, unions, and the variants of enums. The identifier Self is also allowed as the left hand side of a bound in a where clause and as a type argument to a trait bound on the right hand side of a where clause.

Desugaring

When the compiler encounters Self in type contexts inside the places described above, it will substitute them with the header of the type definition but remove any bounds on generic parameters prior.

An example: the following cons list:

enum StackList<'a, T: 'a + InterestingTrait> {
    Nil,
    Cons(T, &'a Self)
}

desugars into:

enum StackList<'a, T: 'a + InterestingTrait> {
    Nil,
    Cons(T, &'a StackList<'a, T>)
}

Note in particular that the source code is not desugared into:

enum StackList<'a, T: 'a + InterestingTrait> {
    Nil,
    Cons(T, &'a StackList<'a, T: 'a + InterestingTrait>)
}

An example of Self in where bounds is:

struct Foo<T>
where
    Self: PartialEq<Self>
{
    // Some fields..
}

which desugars into:

struct Foo<T>
where
    Foo<T>: PartialEq<Foo<T>>
{
    // Some fields..
}

In relation to RFC 2102 and what Self refers to.

It should be noted that Self always refers to the top level type and not the inner unnamed struct or union because those are unnamed. Specifically, Self always applies to the innermost nameable type. In type definitions in particular, this is equivalent: Self always applies to the top level type.

Error messages

When Self is used to construct an infinite type as in:

enum List<T> {
    Nil,
    Cons(T, Self)
}

The compiler will emit error E0072 as in:

error[E0072]: recursive type `List` has infinite size
 --> src/main.rs:1:1
  |
1 | enum List<T> {
  | ^^^^^^^^^^^^ recursive type has infinite size
2 |     Nil,
3 |     Cons(T, Self)
  |             ----- recursive without indirection
  |
  = help: insert indirection (e.g., a `Box`, `Rc`, or `&`) at some point to make `List` representable

Note in particular that Self is used and not List<T> on line 3.

In relation to other RFCs

This RFC expands on RFC 593 and RFC 1647 with respect to where the keyword Self is allowed.

Drawbacks

Some may argue that we shouldn't have many ways to do the same thing and that it introduces new syntax whereby making the surface language more complex. However, the RFC may equally be said to simplify the surface language since it removes exceptional cases especially in the users mental model.

Using Self in a type definition makes it harder to search for all positions in which a pattern can appear in an AST.

Rationale and alternatives

The rationale for this particular design is straightforward as it would be uneconomic, confusing, and inconsistent to use other keywords.

The consistency of what Self refers to

As explained in the [reference-level explanation], we said that:

Self always applies to the innermost nameable type.

We arrive at this conclusion by examining a few different cases and what they have in common.

Current Rust - Shadowing in impls

First, let's take a look at shadowing in impls.

fn main() { Foo {}.foo(); }

#[derive(Debug)]
struct Foo;

impl Foo {
    fn foo(&self) {
        // Prints "Foo", which is the innermost type.
        println!("{:?}", Self {});

        #[derive(Debug)]
        struct Bar;

        impl Bar {
            fn bar(&self) {
                // Prints "Bar", also the innermost type in this context.
                println!("{:?}", Self {});
            }
        }
        Bar {}.bar();
    }
}

Let's also consider trait impls instead of inherent impls:

impl Trait for Foo {
    fn foo(&self) {
        impl Trait for Bar {
            // Self is shadowed here...
        }
    }
}

We see that the conclusion holds for both examples.

In relation to RFC 2102

Let's consider a modified example from RFC 2102:

#[repr(C)]
struct S {
    a: u32,
    _: union {
        b: Box<Self>,
        c: f32,
    },
    d: u64,
}

In this example, the inner union is not nameable, and so Self refers to the only nameable introduced type S. Therefore, the conclusion holds.

Type definitions inside impls

If in the future we decide to permit type definitions inside impls as in:

impl Trait for Foo {
    struct Bar {
        head: u8,
        tail: Option<Box<Self>>,
    }
}

as sugar for:

enum _Bar {
    head: u8,
    tail: Option<Box<Self>>,
}
impl Trait for Foo {
    type Bar = _Bar;
}

In the desugared example, we see that the only possible meaning of Self is that it refers to _Bar and not Foo. To be consistent with the desugared form, the sugared variant should have the same meaning and so Self refers to Bar there.

Let's now consider an alternative possible syntax:

impl Trait for Foo {
    type Bar = struct /* there is no ident here */ {
        outer: Option<Box<Self>>,
        inner: Option<Box<Self::Item>>,
    }
}

Notice here in particular that there is no identifier after the keyword struct. Because of this, it is reasonable to say that the struct assigned to the associated type Bar is not directly nameable as Bar. Instead, a user must qualify Bar with Self::Bar. With this in mind, we arrive at the following interpretation:

impl Trait for Foo {
    type Bar = struct /* there is no ident here */ {
        outer: Option<Box<Foo>>,
        inner: Option<Box<Foo::Bar>>,
    }
}

Conclusion

We've now examined a few cases and seen that indeed, the meaning of Self is consistent in all of them as well as with what the meaning in in today's Rust.

Doing nothing

One alternative to the changes proposed in this RFC is to simply not implement those changes. However, this has the downsides of not increasing the ergonomics and keeping the language less consistent than what it could be. Not improving the ergonomics here may be especially problematic when dealing with "recursive" types that have long names and/or many generic parameters and may encourage developers to use type names which are less descriptive and keep their code less generic than what is appropriate.

Internal scoped type aliases

Another alternative is to allow users to specify type aliases inside type definitions and use any generic parameters specified in that definition. An example is:

enum Tree<T> {
    type S = Box<Tree<T>>;

    Nil,
    Node(T, S, S),
}

instead of:

enum Tree<T> {
    Nil,
    Node(T, Box<Self>, Box<Self>),
}

When dealing with generic associated types (GATs), we can then write:

enum Tree<T, P: PointerFamily> {
    type S = P::Pointer<Tree<T>>;

    Nil,
    Node(T, S, S),
}

instead of:

enum Tree<T, P: PointerFamily> {
    Nil,
    Node(T, P::Pointer<Tree<T>>, P::Pointer<Tree<T>>),
}

As we can see, this approach and alternative is more flexible compared to what is proposed in this RFC, particularly in the case of GATs. However, this alternative requires introducing and teaching more concepts compared to this RFC, which comparatively builds more on what users already know. Mixing ; and , has also proven to be controversial in the past. The alternative also opens up questions such as if the type alias should be permitted before the variants, or after the variants.

For simpler cases such as the first tree-example, using Self is also more readable as it is a special construct that you can easily syntax-highlight for in a more noticeable way. Further, while there is an expectation from some users that Self already works, as discussed in the motivation, the expectation that this alternative already works has not been brought forth by anyone as far as this RFC's author is aware.

It is also unclear how internal scoped type aliases would syntactically work with where bounds.

Strictly speaking, this particular alternative is not in conflict with this RFC in that both can be supported technically. The alternative should be considered interesting future work, but for now, a more conservative approach is preferred.

Unresolved questions