RFC 2532: associated-type-defaults

lang (traits | typesystem | impls | associated-types | defaults)

Summary

Resolve the design of associated type defaults, first introduced in RFC 192, such that provided methods and other items may not assume type defaults. This applies equally to default with respect to specialization. Finally, dyn Trait will assume provided defaults and allow those to be elided.

Motivation

As discussed in the background and mentioned in the summary, associated type defaults were introduced in RFC 192. These defaults are valuable for a few reasons:

  1. You can already provide defaults for consts and fns. Allowing types to have defaults adds consistency and uniformity to the language, thereby reducing surprises for users.

  2. Associated type defaults in traits simplify the grammar, allowing the grammar of traits them to be more in line with the grammar of impls. In addition, this brings traits more in line with type aliases.

The following points were also noted in RFC 192, but we expand upon them here:

  1. Most notably, type defaults allow you to provide more ergonomic APIs.

    For example, we could change proptest's API to be:

    trait Arbitrary: Sized + fmt::Debug {
        type Parameters: Default = ();
    
        fn arbitrary_with(args: Self::Parameters) -> Self::Strategy;
    
        fn arbitrary() -> Self::Strategy {
            Self::arbitrary_with(Default::default())
        }
    
        type Strategy: Strategy<Value = Self>;
    }
    

    Being able to say that the default of Parameters is () means that users, who are not interested in this further detail, may simply ignore specifying Parameters.

    The inability of having defaults results in an inability to provide APIs that are both a) simple to use, and b) flexible / customizable. By allowing defaults, we can have our cake and eat it too, enabling both a) and b) concurrently.

  2. Type defaults also aid in API evolution. Consider a situation such as Arbitrary from above; The API might have originally been:

    trait Arbitrary: Sized + fmt::Debug {
        fn arbitrary() -> Self::Strategy;
    
        type Strategy: Strategy<Value = Self>;
    }
    

    with an implementation:

    impl Arbitrary for usize {
        fn arbitrary() -> Self::Strategy { 0..100 }
    
        type Strategy = Range<usize>;
    }
    

    By allowing defaults, we can transition to this more flexible API without breaking any consumers by simply saying:

    trait Arbitrary: Sized + fmt::Debug {
        type Parameters: Default = ();
    
        fn arbitrary() -> Self::Strategy {
            Self::arbitrary_with(Default::default())
        }
    
        fn arbitrary_with(_: Self::Parameters) -> Self::Strategy {
            Self::arbitrary()
            // This co-recursive definition will blow the stack.
            // However; since we can assume that previous implementors
            // actually provided a definition for `arbitrary` that
            // can't possibly reference `arbitrary_with`, we are OK.
            // You would only run into trouble for new implementations;
            // but that can be dealt with in documentation.
        }
    
        type Strategy: Strategy<Value = Self>;
    }
    

    The implementation Arbitrary for usize remains valid even after the change.

Guide-level explanation

Background and The status quo

Let's consider a simple trait with an associated type and another item (1):

trait Foo {
    type Bar;

    const QUUX: Self::Bar;

    fn wibble(x: Self::Bar) -> u8;
}

Ever since RFC 192, Rust has been capable of assigning default types to associated types as in (2):

#![feature(associated_type_defaults)]

trait Foo {
    type Bar = u8;

    const QUUX: Self::Bar = 42u8;

    fn wibble(x: Self::Bar) -> u8 { x }
}

However, unlike as specified in RFC 192, which would permit (2), the current implementation rejects (2) with the following error messages (3):

error[E0308]: mismatched types
 --> src/lib.rs:6:29
  |
6 |     const QUUX: Self::Bar = 42u8;
  |                             ^^^^ expected associated type, found u8
  |
  = note: expected type `<Self as Foo>::Bar`
             found type `u8`

error[E0308]: mismatched types
 --> src/lib.rs:8:37
  |
8 |     fn wibble(x: Self::Bar) -> u8 { x }
  |                                --   ^ expected u8, found associated type
  |                                |
  |                                expected `u8` because of return type
  |
  = note: expected type `u8`
             found type `<Self as Foo>::Bar`

The compiler rejects snippet (2) to preserve the soundness of the type system. It must be rejected because a user might write (4):

struct Bar { ... }

impl Foo for Bar {
    type Bar = Vec<u8>;
}

Given snippet (4), Self::Bar will evaluate to Vec<u8>, which is therefore the type of <Bar as Foo>::QUUX. However, we have not given a different value for the constant, and so it must be 42u8, which has the type u8. Therefore, we have reached an inconsistency in the type system: <Bar as Foo>::QUUX is of value 42u8, but of type Vec<u8>. So we may accept either impl Foo for Bar as defined in (4), or the definition of Foo as in (2), but not both.

RFC 192 solved this dilemma by rejecting the implementation and insisting that if you override one associated type, then you must override all other defaulted items. Or stated in its own words:

  • If a trait implementor overrides any default associated types, they must also override all default functions and methods.
  • Otherwise, a trait implementor can selectively override individual default methods/functions, as they can today.

Meanwhile, as we saw in the error message above (3), the current implementation takes the alternative approach of accepting impl Foo for Bar (4) but not the definition of Foo as in (2).

Changes in this RFC

In this RFC, we change the approach in RFC 192 to the currently implemented approach. Thus, you will continue to receive the error message above and you will be able to provide associated type defaults.

With respect to specialization, the behaviour is the same. That is, if you write (5):

#![feature(specialization)]

trait Foo {
    type Bar;

    fn quux(x: Self::Bar) -> u8;
}

struct Wibble<T>;

impl<T> Foo for Wibble<T> {
    default type Bar = u8;

    default fn quux(x: Self::Bar) -> u8 { x }
}

The compiler will reject this because you are not allowed to assume, just like before, that x: u8. The reason why is much the same as we have previously discussed in the background.

One place where this proposal diverges from what is currently implemented is with respect to the following example (6):

#![feature(associated_type_defaults)]

trait Foo {
    type Bar = usize;

    fn baz(x: Self::Bar) -> usize;
}

impl<T> Foo for Vec<T> {
    fn baz(x: Self::Bar) -> usize { x }
}

In the current implementation, (6) is rejected because the compiler will not let you assume that x is of type usize. But in this proposal, you would be allowed to assume this. To permit this is not a problem because Foo for Vec<T> is not further specializable since Bar in the implementation has not been marked as default.

Trait objects

Another divergence in this RFC as compared to the current implementation is with respect to trait objects. Currently, if you write (7):

trait Foo {
    type Bar = u8;
    fn method(&self) -> Self::Bar;
}

type Alpha = Box<dyn Foo>;

the compiler will reject it with (8):

error[E0191]: the value of the associated type `Bar` (from the trait `Foo`) must be specified
 --> src/lib.rs:8:17
  |
4 |     type Bar = u8;
  |     -------------- `Bar` defined here
...
8 | type Alpha = Box<dyn Foo>;
  |                  ^^^^^^^ associated type `Bar` must be specified

With this RFC however, the error in (8) will disappear and (7) will be accepted. That is, Box<dyn Foo> is taken as equivalent as Box<dyn Foo<Bar = u8>>.

If we complicate the situation slightly and introduce another associated type Baz which refers to Bar in its default, the compiler will still let us elide specifying the defaults (9):

trait Foo {
    type Bar = u8;
    type Baz = Vec<Self::Bar>;

    fn method(&self) -> (Self::Bar, Self::Baz);
}

type Alpha = Box<dyn Foo>;
//               -------
//               Same as: `dyn Foo<Bar = u8, Baz = Vec<u8>>`.

type Beta = Box<dyn Foo<Bar = u16>>;
//              ------------------
//              Same as: `dyn Foo<Bar = u16, Baz = Vec<u16>>`.

Note that in Beta, Bar was specified but Baz was not. The compiler can infer that Baz is Vec<u16> since Self::Bar = u16 and Baz = Vec<Self::Bar>.

With these changes, we consider the design of associated type defaults to be finalized.

Reference-level explanation

The proposal makes no changes to the dynamic semantics and the grammar of Rust.

Static semantics

This section supersedes RFC 192 with respect to associated type defaults.

Associated types can be assigned a default type in a trait definition:

trait Foo {
    type Bar = $default_type;

    $other_items
}

Any item in $other_items, which have any provided definitions, may only assume that the type of Self::Bar is Self::Bar. They may not assume that the underlying type of Self::Bar is $default_type. This property is essential for the soundness of the type system.

When an associated type default exists in a trait definition, it need not be specified in the implementations of that trait. If implementations of that trait do not make that associated type available for specialization, the $default_type may be assumed in other items specified in the implementation. If an implementation does make the associated type available for further specialization, then other definitions in the implementation may not assume the given underlying specified type of the associated type and may only assume that it is Self::TheAssociatedType.

This applies generally to any item inside a trait. You may only assume the signature of an item, but not any provided definition, in provided definitions of other items. For example, this means that you may not assume the value of an associated const item in other items with provided definition in a trait definition.

Interaction with dyn Trait<...>

Then, in a type of form (where m ≤ n ≤ l):

dyn X<
  L0, .., Lk,
  σ0, .. σn,
  A0 = σ_{n + 1}, .., Ao = σ_{n + o}
>

the associated types in E must be bound in A0, .., Ao whereas those in D may be omitted selectively (i.e. omit zero, some, or all).

When inferring the types of the omitted projections in D, projections in the assigned defaults of types in D will use the types in A0, .., Ao instead of the defaults specified in D. For example, if given:

trait X {
    type A0 = u8;
    type A1 = Vec<Self::A0>;
}

then the type dyn X<A0 = u16> is inferred to dyn X<A0 = u16, A1 = Vec<u16>> as opposed to dyn X<A0 = u16, A1 = Vec<u8>>.

Interaction with existential type

RFC 2071 defines a construct existential type Foo: Bar; which is permitted in associated types and results in an opaque type. This means that the nominal type identity is hidden from certain contexts and only Bar is extensionally known about the type wherefore only the operations of Bar is afforded. This construct is sometimes written as type Foo = impl Bar; in conversation instead.

With respect to this RFC, the semantics of type Assoc = impl Bar; inside a trait definition, where Assoc is the name of the associated type, is understood as what it means in terms of default impl .. as discussed in RFC 1210. What this means in concrete terms is that given:

trait Foo {
    type Assoc = impl Bar;

    ...
}

the underlying type of Assoc stays the same for all implementations which do not change the default of Assoc. The same applies to specializations. With respect to type opacity, it is the same as that of existential type.

Drawbacks

The main drawbacks of this proposal are that:

  1. if you have implementations where you commonly would have needed to write default { .. } because you need to assume the type of an associated type default in a provided method, then the solution proposed in this RFC is less ergonomic.

    However, it is the contention of this RFC that such needs will be less common and that the nesting mechanism or other similar ideas will be sufficiently ergonomic for such cases. This is discussed below.

Rationale and alternatives

Alternatives

The main alternative is to retain the behaviour in RFC 192 such that you may assume the type of associated type defaults in provided methods. As noted in the drawbacks section, this would be useful for certain types of APIs. However, it is more likely than not that associated type defaults will be used as a mechanism for code reuse than for other constructs. As such, we consider the approach in this RFC to be more ergonomic.

Another alternative to the mechanism proposed in this RFC is to somehow track which methods rely on which associated types as well as constants. However, we have historically had a strong bias toward being explicit in signatures about such things, avoiding to infer them. With respect to semantic versioning, such an approach may also cause surprises for crate authors and their dependents alike because it may be difficult at glance to decide what the dependencies are. This in turn reduces the maintainability and readability of code.

Consistency with associated consts

Consider the following valid example from stable Rust:

trait Foo {
    const BAR: usize = 1;

    fn baz() { println!("Hi I'm baz."); }
}

impl Foo for () {
    fn baz() { println!("Hi I'm () baz."); }
}

As we can see, you are permitted to override baz but leave BAR defaulted. This is consistent with the behaviour in this RFC in that it has the same property: "you don't need to override all items if you override one".

Consistency and uniformity of any programming language is vital to make its learning easy and to rid users of surprising corner cases and caveats. By staying consistent, as shown above, we can reduce the cost to our complexity budget that associated type defaults incur.

Overriding everything is less ergonomic

We have already discussed this to some extent. Another point to consider is that Rust code frequently sports traits such as Iterator and Future that have many provided methods and few associated types. While these particular traits may not benefit from associated type defaults, many other traits, such as Arbitrary defined in the motivation, would.

True API evolution by inferring in dyn Trait

While impl Trait will not take associated type defaults into account, dyn trait will. This may seem inconsistent. However, it is justified by the inherent difference in semantics between these constructs and by the goal set out in the motivation to facilitate API evolution.

As an illustration, consider Iterator:

trait Iterator {
    type Item;

    ...
}

Currently, you may write:

fn foo() -> impl Iterator { 0..1 }

and when foo is called, you will know nothing about Item.

However, you cannot write:

fn bar() -> Box<dyn Iterator> { Box::new(0..1) }

since the associated type Item is not specified.

In bar, the type of Item is unknown and so the compiler does not know how to generate the vtable. As a result, an error is emitted:

L | fn bar() -> Box<dyn Iterator> { Box::new(0..1) }
    |                 ^^^^^^^^^^^^ missing associated type `Item` value

If we introduced a default for Item:

    type Item = ();

then bar would become legal under this RFC and so strictly more code than today would be accepted.

Meanwhile, if impl Iterator meant impl Iterator<Item = ()>, this would impose a stronger requirement on existing code where impl Iterator is used and thus it would be a breaking change to the users of Iterator.

For Iterator, it would not be helpful to introduce a default for Item. However, for the purposes of API evolution, the value is not in assigning defaults to the existing associated types of a trait. Rather, the value comes from being able to add associated types without breaking dependent crates.

Due to the possible breakage of dyn Trait<..> when adding an associated type to Trait, to truly achieve API evolution, defaults must be taken into account and be inferable for dyn Trait. The opposite is true for impl Trait. To facilitate API evolution, stronger requirements must not be placed on impl Trait and therefore defaults should not be taken into account.

Prior art

Haskell

As Rust traits are a form of type classes, we naturally look for prior art from were they first were introduced. That language, being Haskell, permits a user to specify associated type defaults. For example, we may write the following legal program:

{-# LANGUAGE TypeFamilies #-}

class Foo x where
  type Bar x :: *
  -- A default:
  type Bar x = Int

  -- Provided method:
  baz :: x -> Bar x -> Int
  baz _ _ = 0

data Quux = Quux

instance Foo Quux where
  baz _ y = y

As in this proposal, we may assume that y :: Int in the above snippet.

In this case, we are not assuming that Bar x unifies with Int in the class. Let's try to assume that now:

{-# LANGUAGE TypeFamilies #-}

class Foo x where
  type Bar x :: *
  -- A default:
  type Bar x = Int

  -- Provided method:
  baz :: x -> Bar x -> Int
  baz _ barX = barX

This snippet results in a type checking error (tested on GHC 8.0.1):

main.hs:11:16: error:
    • Couldn't match expected type ‘Int’ with actual type ‘Bar x’
    • In the expression: barX
      In an equation for ‘baz’: baz _ barX = barX
    • Relevant bindings include
        barX :: Bar x (bound at main.hs:11:9)
        baz :: x -> Bar x -> Int (bound at main.hs:11:3)
<interactive>:3:1: error:

The thing to pay attention to here is:

Couldn't match expected type ‘Int’ with actual type ‘Bar x

We can clearly see that the type checker is not allowing us to assume that Int and Bar x are the same type. This is consistent with the approach this RFC proposes.

To our knowledge, Haskell does not have any means such as default { .. } to change this behaviour. Presumably, this is the case because Haskell preserves parametricity thus lacking specialization, wherefore default { .. }, as suggested in the future possibilities, might not carry its weight.

Idris

Idris has a concept it calls interfaces. These resemble type classes in Haskell, and by extension traits in Rust. However, unlike Haskell and Rust, these interfaces do not have the property of coherence and will permit multiple implementations of the same interface.

Since Idris is language with full spectrum dependent types, it does not distinguish between terms and types, instead, types are terms. Therefore, there is really not a distinct concept called "associated type". However, an interface may require certain definitions to be provided and this includes types. For example, we may write:

interface Iterator self where
    item : Type
    next : self -> Maybe (self, item)

implementation Iterator (List a) where
    item = a
    next [] = Nothing
    next (x :: xs) = Just (xs, x)

Like in Haskell, in Idris, a function or value in an interface may be given a default definition. For example, the following is a valid program:

interface Foo x where
    bar : Type
    bar = Bool

    baz : x -> bar

implementation Foo Int where
    baz x = x == 0

However, if we provide a default for baz in the interface which assumes the default value Bool of bar, as with the following example:

interface Foo x where
    bar : Type
    bar = Bool

    baz : x -> bar
    baz _ = True

then we run into an error:

Type checking .\foo.idr
foo.idr:6:13-16:
  |
6 |     baz _ = True
  |             ~~~~
When checking right hand side of Main.default#baz with expected type
        bar x _

Type mismatch between
        Bool (Type of True)
and
        bar x _ (Expected type)

The behaviour here is exactly as in Haskell and as proposed in this RFC.

C++

In C++, it is possible to provide associated types and specialize them as well. This is shown in the following example:

#include <iostream>
#include <string>

template<typename T> struct wrap {};

template<typename T> struct foo { // Unspecialized.
    using bar = int;

    bar make_a_bar() { return 0; };
};

template<typename T> struct foo<wrap<T>> { // Partial specialization.
    using bar = std::string;

    bar make_a_bar() { return std::string("hello world"); };
};

int main() {
    foo<void> a_foo;
    std::cout << a_foo.make_a_bar() << std::endl;

    foo<wrap<void>> b_foo;
    std::cout << b_foo.make_a_bar() << std::endl;
}

You will note that C++ allows us to assume in both the base template class, as well as the specialization, that bar is equal to the underlying type. This is because one cannot specialize any part of a class without specializing the whole of it. It's equivalent to one atomic default { .. } block.

Swift

One language which does have associated types and defaults but which does not have provided definitions for methods is Swift. As an example, we may write:

protocol Foo {
    associatedtype Bar = Int

    func append() -> Bar
}

struct Quux: Foo {
    func baz() -> Bar {
        return 1
    }
}

However, we may not write:

protocol Foo {
    associatedtype Bar = Int

    func append() -> Bar { return 0 }
}

This would result in:

main.swift:4:23: error: protocol methods may not have bodies
    func baz() -> Bar { return 0 }

Scala

Another language which allows for these kinds of type projections and defaults for them is Scala. While Scala does not have type classes like Rust and Haskell does, it does have a concept of trait which can be likened to a sort of incoherent "type class" system. For example, we may write:

trait Foo {
    type Bar = Int

    def baz(x: Bar): Int = x
}

class Quux extends Foo {
    override type Bar = Int
    override def baz(x: Bar): Int = x
}

There are a few interesting things to note here:

  1. We are allowed to specify a default type Int for Bar.

  2. A default definition for baz may be provided.

  3. This default definition may assume the default given for Bar.

  4. However, we must explicitly state that we are overriding baz.

  5. If we change the definition of of override type Bar to Double, the Scala compiler will reject it.

Unresolved questions

1. When do suitability of defaults need to be proven?

Consider a trait Foo<T> defined as:

trait Foo<T> {
    type Bar: Clone = Vec<T>;
}

Let's also assume the following implementation of Clone:

impl<T: Clone> Clone for Vec<T> { ... }

To prove that Vec<T>: Clone, we must prove that T: Clone. However, Foo<T> does not say that T: Clone so is its definition valid? If the suitability of Vec<T> is checked where Foo<T> is defined (1), then we don't know that T: Clone and so the definition must be rejected. To make the compiler admit Foo<T>, we would have to write:

trait Foo<T: Clone> {
    type Bar: Clone = Vec<T>;
}

Now it is provable that T: Clone so Vec<T>: Clone which is what was required.

If instead the suitability of defaults are checked in implementations (2), then proving Vec<T>: Clone would not be required in Foo<T>'s definition and so then Foo<T> would type-check. As a result, it would be admissible to write:

#[derive(Copy, Clone)]
struct A;

struct B;

impl Foo<A> for B {}

since Vec<A>: Clone holds.

With condition (2), strictly more programs are accepted than with (1). It may be that useful programs are rejected if we enforce (1) rather than (2). However, it would also be the more conservative choice, allowing us to move towards (2) when necessary. As it is currently unclear what solution is best, this question is left unresolved.

2. Where are cycles checked?

Consider a program (playground):

#![feature(associated_type_defaults)]

trait A {
    type B = Self::C; // B defaults to C,
    type C = Self::B; // C defaults to B, and we have a cycle!
}

impl A for () {}

fn _foo() {
    let _x: <() as A>::B;
}

// Removing this function will make the example compile.
fn main() {
    let _x: <() as A>::B;
}

Currently, this results in a crash. This will need to be fixed. At the very latest, impl A for () {} should have been an error.

trait A {
    type B = Self::C;
    type C = Self::B;
}

impl A for () {} // This OK but shouldn't be.

If cycles are checked for in impl A for (), then it would be valid to write:

trait A {
    type B = Self::C;
    type C = Self::B;
}

impl A for () {
    type B = u8; // The cycle is broken!
}

Alternatively, cycles could be checked for in A's definition. This is similar to the previous question in (1).

Future possibilities

This section in the RFC used to be part of the proposal. To provide context for considerations made in the proposal, it is recorded here.

Summary

Introduce the concept of default { .. } groups in traits and their implementations which may be used to introduce atomic units of specialization (if anything in the group is specialized, everything must be). These groups may be nested and form a tree of cliques.

Motivation

For default { .. } groups

Finally, because we are making changes to how associated type defaults work in this RFC, a new mechanism is required to regain the loss of expressive power due to these changes. This mechanism is described in the section on default { .. } groups as alluded to in the summary.

These groups not only retain the expressive power due to RFC 192 but extend power such that users get fine grained control over what things may and may not be overridden together. In addition, these groups allow users to assume the definition of type defaults in other items in a way that preserves soundness.

Examples where it is useful for other items to assume the default of an associated type include:

  1. A default method whose return type is an associated type:

    /// "Callbacks" for a push-based parser
    trait Sink {
        fn handle_foo(&mut self, ...);
    
        default {
            type Output = Self;
    
            // OK to assume what `Output` really is because any overriding
            // must override both `Outout` and `finish`.
            fn finish(self) -> Self::Output { self }
        }
    }
    
  2. There are plenty of other examples in rust-lang/rust#29661.

  1. Other examples where default { .. } would have been useful can be found in the tracking issue for specialization:

  1. Encoding a more powerful std::remove_reference

    We can encode a more powerful version of C++'s remove_reference construct, which allows you to get the base type of a reference type recursively. Without default groups, we can get access to the base type like so:

    trait RemoveRef {
        type WithoutRef;
    }
    
    impl<T> RemoveRef for T {
        default type WithoutRef = T;
    }
    
    impl<'a, T: RemoveRef> RemoveRef for &'a T {
        type WithoutRef = T::WithoutRef;
    }
    

    However, we don't have any way to transitively dereference to &Self::WithoutRef. With default groups we can gain that ability with:

    trait RemoveRef {
        type WithoutRef;
        fn single_ref(&self) -> &Self::WithoutRef;
    }
    
    impl<T> RemoveRef for T {
        default {
            type WithoutRef = T;
    
            fn single_ref(&self) -> &Self::WithoutRef {
                // We can assume that `T == Self::WithoutRef`.
                self
            }
        }
    }
    
    impl<'a, T: RemoveRef> RemoveRef for &'a T {
        type WithoutRef = T::WithoutRef;
    
        fn single_ref(&self) -> &Self::WithoutRef {
            // We can assume that `T::WithoutRef == Self::WithoutRef`.
            T::single_ref(*self)
        }
    }
    

    We can then proceed to writing things such as:

    fn do_stuff(recv: impl RemoveRef<WithoutRef: MyTrait>) {
        recv.single_ref().my_method();
    }
    

Guide-level explanation

default specialization groups

Note: Overlapping implementations, where one is more specific than the other, requires actual support for specialization.

Now, you might be thinking: - "Well, what if I do need to assume that my defaulted associated type is what I said in a provided method, what do I do then?". Don't worry; We've got you covered.

To be able to assume that Self::Bar is truly u8 in snippets (2) and (5), you may henceforth use default { .. } to group associated items into atomic units of specialization. This means that if one item in default { .. } is overridden in an implementation, then all all the items must be. An example (7):

struct Country(&'static str);

struct LangSec { papers: usize }
struct CategoryTheory { papers: usize }

trait ComputerScientist {
    default {
        type Details = Country;
        const THE_DETAILS: Self::Details = Country("Scotland"); // OK!
        fn papers(details: Self::Details) -> u8 { 19 } // OK!
    }
}

// https://en.wikipedia.org/wiki/Emily_Riehl
struct EmilyRiehl;

// https://www.cis.upenn.edu/~sweirich/
struct StephanieWeirich;

// http://www.cse.chalmers.se/~andrei/
struct AndreiSabelfeld;

// https://en.wikipedia.org/wiki/Conor_McBride
struct ConorMcBride;

impl ComputerScientist for EmilyRiehl {
    type Details = CategoryTheory;

    // ERROR! You must override THE_DETAILS and papers.
}

impl ComputerScientist for StephanieWeirich {
    const THE_DETAILS: Country = Country("USA");
    fn papers(details: Self::Details) -> u8 { 86 }

    // ERROR! You must override Details.
}

impl ComputerScientist for AndreiSabelfeld {
    type Details = LangSec;
    const THE_DETAILS: Self::Details = LangSec { papers: 90 };
    fn papers(details: Self::Details) -> u8 { details.papers }

    // OK! We have overridden all items in the group.
}

impl ComputerScientist for ConorMcBride {
    // OK! We have not overridden anything in the group.
}

You may also use default { .. } in implementations. When you do so, everything in the group is automatically overridable. For any items outside the group, you may assume their signatures, but not the default definitions given. An example:

trait Fruit {
    type Details;
    fn foo();
    fn bar();
    fn baz();
}

struct Citrus<S> { species: S }
struct Orange<V> { variety: V }
struct Blood;
struct Common;

impl<S> Fruit for Citrus<S> {
    default {
        type Details = bool;
        fn foo() {
            let _: Self::Details = true; // OK!
        }
        fn bar() {
            let _: Self::Details = true; // OK!
        }
    }

    fn baz() { // Removing this item here causes an error.
        let _: Self::Details = true;
        // ERROR! You may not assume that `Self::Details == bool` here.
    }
}

impl<V> Fruit for Citrus<Orange<V>> {
    default {
        type Details = u8;
        fn foo() {
            let _: Self::Details = 42u8; // OK!
        }
    }

    fn bar() { // Removing this item here causes an error.
        let _: Self::Details = true;
        // ERROR! You may not assume that `Self::Details == bool` here,
        // even tho we specified that in `Fruit for Citrus<S>`.
        let _: Self::Details = 22u8;
        // ERROR! Can't assume that it's u8 either!
    }
}

impl Fruit for Citrus<Orange<Common>> {
    default {
        type Details = f32;
        fn foo() {
            let _: Self::Details = 1.0f32; // OK!
        }
    }
}

impl Fruit for Citrus<Orange<Blood>> {
    default {
        type Details = f32;
    }

    fn foo() {
        let _: Self::Details = 1.0f32;
        // ERROR! Can't assume it is f32.
    }
}

So far our examples have always included an associated type. However, this is not a requirement. We can also group associated consts and fns together or just fns. An example:

trait Foo {
    default {
        const BAR: usize = 3;

        fn baz() -> [u8; Self::BAR] {
            [1, 2, 3]
        }
    }
}

trait Quux {
    default {
        fn wibble() {
            ...
        }

        fn wobble() {
            ...
        }

        // For whatever reason; The crate author has found it imperative
        // that `wibble` and `wobble` always be defined together.
    }
}

Case study

One instance where default groups could be useful to provide a more ergonomic API is to improve upon RFC 2500. The RFC proposes the following API:

trait Needle<H: Haystack>: Sized {
    type Searcher: Searcher<H::Target>;
    fn into_searcher(self) -> Self::Searcher;

    type Consumer: Consumer<H::Target>;
    fn into_consumer(self) -> Self::Consumer;
}

However, it turns out that usually, Consumer and Searcher are the same underlying type. Therefore, we would like to save the user from some unnecessary work by letting them elide parts of the required definitions in implementations.

One might imagine that we'd write:

trait Needle<H: Haystack>: Sized {
    type Searcher: Searcher<H::Target>;
    fn into_searcher(self) -> Self::Searcher;

    default {
        type Consumer: Consumer<H::Target> = Self::Searcher;
        fn into_consumer(self) -> Self::Consumer { self.into_searcher() }
    }
}

However, the associated type Searcher does not necessarily implement Consumer<H::Target>. Therefore, the above definition would not type check.

However, we can encode the above construct by rewriting it slightly, using the concept of partial implementations from RFC 1210:

default impl<H: Haystack> Needle for T
where Self::Searcher: Consumer<H::Target> {
    default {
        type Consumer = Self::Searcher;
        fn into_consumer(self) -> Self::Consumer { self.into_searcher() }
    }
}

Now we have ensured that Self::Searcher is a Consumer<H::Target> and therefore, the above definition will type check. Having done this, the API has become more ergonomic because we can let users define instances of Needle<H> with half as many requirements.

default fn foo() { .. } is syntactic sugar

In the section of changes to associated type defaults, snippet (5) actually indirectly introduced default groups of a special form, namely "singleton groups". That is, when we wrote:

impl<T> Foo for Wibble<T> {
    default type Bar = u8;

    default fn quux(x: Self::Bar) -> u8 { x }
}

this was actually sugar for:

impl<T> Foo for Wibble<T> {
    default {
        type Bar = u8;
    }

    default {
        fn quux(x: Self::Bar) -> u8 { x }
    }
}

We can see that these are equivalent since in the specialization RFC, the semantics of default fn were that fn may be overridden in more specific implementations. With these singleton groups, you may assume the body of Bar in all other items in the same group; but it just happens to be the case that there are no other items in the group.

Nesting and a tree of cliques

In the summary, we alluded to the notion of groups being nested. However, thus far we have seen no examples of such nesting. This RFC does permit you do that. For example, you may write:

trait Foo {
    default {
        type Bar = usize;

        fn alpha() -> Self::Bar {
            0 // OK! In the same group, so we may assume `Self::Bar == usize`.
        }

        // OK; we can rely on `Self::Bar == usize`.
        default const BETA: Self::Bar = 3;

        default fn gamma() -> [Self::Bar; 4] {
            // OK; we can depend on the underlying type of `Self::Bar`.
            [9usize, 8, 7, 6]
        }

        /// This is rejected:
        default fn delta() -> [Self::Bar; Self::BETA] {
            // ERROR! we may not rely on not on `Self::BETA`'s value because
            // `Self::BETA` is a sibling of `Self::gamma` which is not in the
            // same group and is not an ancestor either.
            [9usize, 8, 7]
        }

        // But this is accepted:
        default fn delta() -> [Self::Bar; 3] {
            // OK; we can depend on `Self::Bar == usize`.
            [9, 8, 7]
        }

        default {
            // OK; we can still depend on `Self::Bar == usize`.
            const EPSILON: Self::Bar = 2;

            fn zeta() -> [Self::Bar; Self::Epsilon] {
                // OK; We can assume the value of `Self::EPSILON` because it
                // is a sibling in the same group. We may also assume that
                // `Self::Bar == usize` because it is an ancestor.
                [42usize, 24]
            }
        }
    }
}

struct Eta;
struct Theta;
struct Iota;

impl Foo for Eta {
    // We can override `gamma` without overriding anything else because
    // `gamma` is the sole member of its sub-group. Note in particular
    // that we don't have to override `alpha`.
    fn gamma() -> [Self::Bar; 4] {
        [43, 42, 41, 40]
    }
}

impl Bar for Theta {
    // Since `EPSILON` and `zeta` are in the same group; we must override
    // them together. However, we still don't have to override anything
    // in ancestral groups.
    const EPSILON: Self::Bar = 0;

    fn zeta() -> [Self::Bar; Self::Epsilon] {
        []
    }
}

impl Bar for Iota {
    // We have overridden `Bar` which is in the root group.
    // Since all other items are decendants of the same group as `Bar` is in,
    // they are allowed to depend on what `Bar` is.
    type Bar = u8;

    ... // Definitions for all the other items elided for brevity.
}

In graph theory, a set of a vertices, in a graph, for which each distinct pair of vertices is connected by a unique edge is said to form a clique. What the snippet above encodes is a tree of such cliques. In other words, we can visualize the snippet as:

                            ┏━━━━━━━━━━━━━━━━━┓
                            ┃ + type Bar      ┃
              ┏━━━━━━━━━━━━━┃ + fn alpha      ┃━━━━━━━━━━━━━━┓
              ┃             ┗━━━━━━━━━━━━━━━━━┛              ┃
              ┃               ┃             ┃                ┃
              ┃               ┃             ┃                ┃
              ▼               ▼             ▼                ▼
┏━━━━━━━━━━━━━━━┓    ┏━━━━━━━━━━━━━┓    ┏━━━━━━━━━━━━━┓    ┏━━━━━━━━━━━━━━━━━┓
┃ + const Beta  ┃    ┃ + fn gamma  ┃    ┃ + fn delta  ┃    ┃ + const EPSILON ┃
┗━━━━━━━━━━━━━━━┛    ┗━━━━━━━━━━━━━┛    ┗━━━━━━━━━━━━━┛    ┃ + fn zeta       ┃
                                                           ┗━━━━━━━━━━━━━━━━━┛

Please pay extra attention to the fact that items in the same group may depend on each other's definitions as well as definitions of items that are ancestors (up the tree). The inverse implication holds for what you must override: if you override one item in a group, you must override all items in that groups and all items in sub-groups (recursively). As before, these limitations exist to preserve the soundness of the type system.

Nested groups are intended primarily expected to be used when there is one associated type, for which you want to define a default, coupled with a bunch of functions which need to rely on the definition of the associated type. This is a good mechanism for API evolution in the sense that you can introduce a new associated type, rely on it in provided methods, but still perform no breaking change.

Reference-level explanation

Grammar

Productions in this section which are not defined here are taken from parser-lalr.y.

Given:

trait_item : maybe_outer_attrs trait_item_leaf ;

trait_item_leaf
: trait_const
| trait_type
| trait_method
| item_macro
;

trait_const
: CONST ident maybe_ty_ascription maybe_const_default ';'
;

trait_type : TYPE ty_param ';' ;

trait_method : method_prefix method_common ';' | method_prefix method_provided ;
method_prefix : maybe_unsafe | CONST maybe_unsafe | maybe_unsafe EXTERN maybe_abi ;
method_provided : method_common inner_attrs_and_block ;
method_common
: FN ident generic_params fn_decl_with_self_allow_anon_params maybe_where_clause
;

The production trait_item is changed into:

trait_item : maybe_outer_attrs trait_item_def ;

trait_item_def
: trait_default_group
| trait_default_singleton
| trait_const
| trait_type
| trait_method
| item_macro
;

trait_default_singleton : DEFAULT trait_item ;
trait_default_group : DEFAULT '{' trait_item* '}' ;

trait_type : TYPE ty_param ('=' ty_sum)? ';' ;

Given:

impl_item : attrs_and_vis impl_item_leaf ;
impl_item_leaf
: item_macro
| maybe_default impl_method
| maybe_default impl_const
| maybe_default impl_type
;

impl_const : item_const ;
impl_type : TYPE ident generic_params '=' ty_sum ';' ;
impl_method : method_prefix method_common ;

method_common
: FN ident generic_params fn_decl_with_self maybe_where_clause inner_attrs_and_block
;

The production impl_item is changed into:

impl_item : attrs_and_vis impl_item_def ;
impl_item_def
: impl_default_singleton
| impl_default_group
| item_macro
| impl_method
| impl_const
| impl_type
;

impl_default_singleton : DEFAULT impl_item ;
impl_default_group : DEFAULT '{' impl_item* '}' ;

Note that associated type defaults are already in the grammar due to RFC 192 but we have specified them in the grammar here nonetheless.

Note also that default default fn .. as well as default default { .. } are intentionally recognized by the grammar to make life easier for macro authors even though writing default default .. should never be written directly.

Desugaring

After macro expansion, wherever the production trait_default_singleton occurs, it is treated in all respects as, except for error reporting -- which is left up to implementations of Rust, and is desugared to DEFAULT '{' trait_item '}'. The same applies to impl_default_singleton. In other words: default fn f() {} is desugared to default { fn f() {} }.

Semantics and type checking

Semantic restrictions on the syntax

According to the grammar, the parser will accept items inside default { .. } without a body. However, such an item will later be rejected during type checking. The parser will also accept visibility modifiers on default { .. } (e.g. pub default { .. }). However, such a visibility modifier will also be rejected by the type checker.

Specialization groups

Implementations of a trait as well as traits themselves may now contain "specialization default groups" (henceforth: "group(s)") as defined by the grammar.

A group forms a clique and is considered an atomic unit of specialization wherein each item can be specialized / overridden.

Groups may contain other groups - such groups are referred to as "nested groups" and may be nested arbitrarily deeply. Items which are not in any group are referred to as 0-deep. An item directly defined in a group which occurs at the top level of a trait or an impl definition is referred to as being 1-deep. An item in a group which is contained in a 1-deep group is 2-deep. If an item is nested in k groups it is k-deep.

A group and its sub-groups form a tree of cliques. Given a group $g with items $x_1, .. $x_n, an item $x_j in $g can assume the definitions of $x_i, ∀ i ∈ { 1..n } as well as any definitions of items in $f where $f is an ancestor of $g (up the tree). Conversely, items in $g may not assume the definitions of items in descendant groups $h_i of $g as well as items which are grouped at all or which are in groups which are not ancestors of $g.

If an impl block overrides one item $x_j in $g, it also has to override all $x_i in $g where i ≠ j as well as all items in groups $h_i which are descendants of $g (down the tree). Otherwise, items do not need to be overridden.

For example, you may write:

trait Foo {
    default {
        type Bar = u8;
        fn baz() {
            let _: Self::Bar = 1u8;
        }

        default {
            const SIZE: usize = 3;
            fn quux() {
                let_: [Self::Bar; Self::SIZE] = [1u8, 2u8, 3u8];
            }
        }
    }
}

impl Foo for () {
    type Bar = Vec<u8>;
    fn baz() {}
    const SIZE: usize = 5;
    fn quux() {}
}

Linting redundant defaults

When in source code (but not as a consequence of macro expansion), any of the following occurs, a warn-by-default lint (redundant_default) will be emitted:

    default default $item
//  ^^^^^^^ warning: Redundant `default`
//          hint: remove `default`.

    default default {
//  ^^^^^^^ warning: Redundant `default`
//          hint: remove `default`.
        ...
    }

    default {
        ...

        default $item
//      ^^^^^^^ warning: Redundant `default`
//              hint: remove `default`.

        ...
    }

Drawbacks

The main drawbacks of this proposal are that:

  1. default { .. } is introduced, adding to the complexity of the language.

    However, it should be noted that token default is already accepted for use by specialization and for default impl. Therefore, the syntax is only partially new.

Rationale and alternatives

Alternatives

One may consider mechanisms such as default(Bar, BAZ) { .. } to give more freedom as to which dependency graphs may be encoded. However, in practice, we believe that the tree of cliques approach proposed in this RFC should be more than enough for practical applications.

default { .. } is syntactically light-weight

When you actually do need to assume the underlying default of an associated type in a provided method, default { .. } provides a syntax that is comparatively not that heavy weight.

In addition, when you want to say that multiple items are overridable, default { .. } provides less repetition than specifying default on each item would. Thus, we believe the syntax is ergonomic.

Finally, default { .. } works well and allows the user a good deal of control over what can and can't be assumed and what must be specialized together. The grouping mechanism also composes well as seen in the section where it is discussed.

Tree of cliques is familiar

The "can depend on" rule is similar to the rule used to determine whether a non-pub item in a module tree is accessible or not. Familiarity is a good tool to limit complexity costs.

Non-special treatment for methods

In this RFC we haven't given methods any special treatment. We could do so by allowing methods to assume the underlying type of an associated type and still be overridable without having to override the type. However, this might lead to semantic breakage in the sense that the details of an fn may be tied to the definition of an associated type. When those details change, it may also be prudent to change the associated type. Default groups give users a mechanism to enforce such decisions.

Future work

where clauses on default { .. } groups

From our case study, we noticed that we had to depart from our trait definition into a separate default impl.. to handle the conditionality of Self::Searcher: Consumer<H::Target>. However, one method to regain the locality provided by having default { .. } inside the trait definition is to realize that we could attach an optional where clause to the group. This would allow us to write:

trait Needle<H: Haystack>: Sized {
    type Searcher: Searcher<H::Target>;
    fn into_searcher(self) -> Self::Searcher;

    default where
        Self::Searcher: Consume<H::Target>
    {
        type Consumer: Consumer<H::Target> = Self::Searcher;
        fn into_consumer(self) -> Self::Consumer { self.into_searcher() }
    }
}

The defaults in this snippet would then be equivalent to the default impl.. snippet noted in the case study.

This default where $bounds construct should be able to subsume common cases where you only have a single default impl.. but provide comparatively better local reasoning.

However, we do not propose this at this stage because it is unclear how common default impl.. will be in practice.