Generics: Overview

This document is a high-level description of Carbon’s generics design, with pointers to other design documents that dive deeper into individual topics.

Goals
Summary
What are generics?
Future work
References

Goals

Carbon generics supports generalizing code to apply to more situations by adding compile-time parameters, allowing generic programming. Carbon supports both checked and template generics.

Template generics provide a similar model to C++ templates, to help with interop and migration. They can be more convenient to write, and support some use cases, like metaprogramming, that are difficult with checked generics.

Checked generics are an alternative that has advantages including:

Function calls and bodies are checked independently against the function signatures.
Clearer and earlier error messages.
Fast builds, particularly development builds.
Support for both static and dynamic dispatch.

Checked generics do have some restrictions, but are expected to be more appropriate at public API boundaries than templates.

For more detail, see the detailed discussion of generics goals and generics terminology.

Summary

Summary of how Carbon generics work:

Generics are compile-time parameterized functions, types, and other language constructs. Those parameters allow a single definition to apply more generally. They are used to avoid writing specialized, near-duplicate code for similar situations.
The definition of a checked generic is typechecked once, without having to know the specific argument values of the generic parameters it is instantiated with. Typechecking the definition of a checked generic requires a precise contract specifying the requirements on the argument values.
For parameters that will be used as types, those requirements are written using interfaces. Interfaces have a name and describe methods, functions, and other entities for types to implement.
Types must explicitly implement interfaces to indicate that they support their functionality. A given type may implement an interface at most once.
Implementations may be declared inline in the body of a class definition or out-of-line.
Types may extend an implementation declared inline, in which case you can directly call the interface’s methods on those types.
Out-of-line implementations may be defined in the library defining the interface as an alternative to the type, or the library defining a type argument.
Interfaces may be used as the type of a generic parameter. Interfaces are facet types, whose values are the subset of all types that implement the interface. Facet types in general specify the capabilities and requirements of the type. The value of a interface is called a facet. Facets are not types, but are usable as types.
With a template generic, the concrete argument value used by the caller is used for name lookup and typechecking. With checked generics, that is all done with the declared restrictions expressed as the types of bindings in the declaration. Inside the body of a checked generic with a facet parameter, the API of the facet is just the names defined by the facet type.
Deduced parameters are parameters whose values are determined by the types of the explicit arguments. Generic facet parameters are typically deduced.
A function with a generic parameter can have the same function body as an unparameterized one. Functions can freely mix checked, template, and regular parameters.
Interfaces can require other interfaces be implemented.
Interfaces can extend required interfaces.
The & operation on facet types allows you conveniently combine interfaces. It gives you all the names that don’t conflict.
You may also declare a new facet type directly using “named constraints”. Named constraints can express requirements that multiple interfaces be implemented, and give you control over how name conflicts are handled.
Alternatively, you may resolve name conflicts by using a qualified member access expression to directly call a function from a specific interface using a qualified name.

What are generics?

Generics are a mechanism for writing parameterized code that applies generally instead of making near-duplicates for very similar situations, much like C++ templates. For example, instead of having one function per type-you-can-sort:

fn SortInt32Vector(a: Vector(i32)*) { ... }
fn SortStringVector(a: Vector(String)*) { ... }
...

You might have one generic function that could sort any array with comparable elements:

fn SortVector(T:! Comparable, a: Vector(T)*) { ... }

The syntax above adds a ! to indicate that the parameter named T is compile-time. By default compile-time parameters are checked, the template keyword may be added to make it a template generic.

Given an i32 vector iv, SortVector(i32, &iv) is equivalent to SortInt32Vector(&iv). Similarly for a String vector sv, SortVector(String, &sv) is equivalent to SortStringVector(&sv). Thus, we can sort any vector containing comparable elements using this single SortVector function.

This ability to generalize makes SortVector a generic.

Interfaces

The SortVector function requires a definition of Comparable, with the goal that the compiler can perform checking. This has two pieces:

definition checking: completely type check a checked-generic definition without information from calls;
encapsulation: completely type check a call to a generic with information only from the function’s signature, and not from its body. For Rust, this is called “Rust’s Golden Rule.”

In this example, Comparable is an interface.

Interfaces describe all the requirements needed for the type T. Given that the compiler knows T satisfies those requirements, it can type check the body of the SortVector function. This includes checking that the Comparable requirement covers all of the uses of T inside the function.

Later, when the compiler comes across a call to SortVector, it can type check against the requirements expressed in the function’s signature. Using only the types at the call site, the compiler can check that the member elements of the passed-in array satisfy the function’s requirements. There is no need to look at the body of the SortVector function, since we separately checked that those requirements were sufficient.

Defining interfaces

Interfaces, then, have a name and describe methods, functions, and other entities for types to implement.

Example:

interface Comparable {
  // `Less` is an associated method.
  fn Less[self: Self](rhs: Self) -> bool;
}

Functions and methods may be given a default implementation by prefixing the declaration with default and putting the function body in curly braces {…} in place of the terminating ; of the function declaration. To prevent that implementation from being overridden, use final instead of default.

Interfaces describe functionality, but not data; no variables may be declared in an interface.

Contrast with templates

Contrast these checked generics with a Carbon or C++ template, where the compiler may be able to do some checking given a function definition, but more checking of the definition is required after seeing the call sites once all the instantiations are known.

Note: Generics terminology goes into more detail about the differences between checked and template generic parameters.

Implementing interfaces

Interfaces themselves only describe functionality by way of method descriptions. A type needs to implement an interface to indicate that it supports its functionality. A given type may implement an interface at most once.

Consider this interface:

interface Printable {
  fn Print[self: Self]();
}

The interface keyword is used to define a nominal interface. That means that types need to explicitly implement them, using an impl block, such as here:

class Song {
  // ...

  // Implementing `Printable` for `Song` inside the definition of `Song`
  // with the keyword `extend` means all names of `Printable`, such
  // as `F`, are included as a part of the `Song` API.
  extend impl as Printable {
    // Could use `Self` in place of `Song` here.
    fn Print[self: Song]() { ... }
  }
}

// Implement `Comparable` for `Song` without changing the API of `Song`
// using an `impl` declaration without `extend`. This may be defined in
// either the library defining `Song` or `Comparable`.
impl Song as Comparable {
  // Could use either `Self` or `Song` here.
  fn Less[self: Self](rhs: Self) -> bool { ... }
}

Implementations may be defined within the class definition itself or out-of-line. Implementations may optionally start with the extend keyword to say the members of the interface are also members of the class, which may only be used in a class scope. Out-of-line implementations may be defined in the library defining the class, the interface, or a type argument.

Accessing members of interfaces

Methods from an interface that a class extends may be called with the simple member access syntax. Methods of all implemented interfaces may be called with a qualified member access expression, whether the class extends them or not.

var song: Song;
// `song.Print()` is allowed, unlike `song.Play()`.
song.Print();
// `Less` is defined in `Comparable`, which `Song`
// does not extend the implementation of.
song.(Comparable.Less)(song);
// Can also call `Print` using a qualified member
// access expression, using the compound member access
// syntax with the qualified name `Printable.Print`:
song.(Printable.Print)();

To type check a function, the compiler needs to be able to verify that uses of a value match the capabilities of the value’s type. In SortVector, the parameter T is a type, but that type is a checked-generic, or symbolic, parameter. That means that the specific type value assigned to T is not known when type checking the SortVector function. Instead it is the constraints on T that let the compiler know what operations may be performed on values of type T. Those constraints are represented by the type of T, a facet type.

In general, a facet type describes the capabilities of a type, while a type defines specific implementations of those capabilities. An interface, like Comparable, may be used as a facet type. In that case, the constraint on the type is that it must implement the interface Comparable.

A facet type also defines a set of names and a mapping to corresponding qualified names. Those names are used for simple member lookup in scopes where the value of the type is not known, such as when the type is a generic parameter.

You may combine interfaces into new facet types using the & operator or named constraints.

Generic functions

We want to be able to call generic functions just like ordinary functions, and write generic function bodies like ordinary functions. There are only a few differences, like that you can’t take the address of generic functions.

Deduced parameters

This SortVector function is explicitly providing type information that is already included in the type of the second argument. To eliminate the argument at the call site, use a deduced parameter.

fn SortVectorDeduced[T:! Comparable](a: Vector(T)*) { ... }

The T parameter is defined in square brackets before the explicit parameter list in parenthesis to indicate it should be deduced. This means you may call the function without the type argument, just like the ordinary functions SortInt32Vector or SortStringVector:

SortVectorDeduced(&anIntVector);
// or
SortVectorDeduced(&aStringVector);

and the compiler deduces that the T argument should be set to i32 or String from the type of the argument.

Deduced arguments are always determined from the call and its explicit arguments. There is no syntax for specifying deduced arguments directly at the call site.

// ERROR: can't determine `U` from explicit parameters
fn Illegal[T:! type, U:! type](x: T) -> U { ... }

A function with a facet parameter can have the same function body as an unparameterized one.

fn PrintIt[T:! Printable](p: T*) {
  p->Print();
}

fn PrintIt(p: Song*) {
  p->Print();
}

Inside the function body, you can treat the facet parameter just like any other type. There is no need to refer to or access generic parameters differently because they are defined as generic, as long as you only refer to the names defined by facet type for the facet parameter.

You may also refer to any of the methods of interfaces required by the facet type using a qualified member access expression.

A function can have a mix of checked, template, and regular parameters. Each kind of parameter is defined using a different syntax: a checked parameter is uses a symbolic binding pattern, a template parameter uses a template binding pattern, and a regular parameter uses a runtime binding pattern. Likewise, it’s allowed to pass a symbolic or template constant value to a checked or regular parameter. We have decided to support passing a symbolic value to a template parameter, see leads issue #2153: Checked generics calling templates, but incorporating it into the design is future work.

Requiring or extending another interface

Interfaces can require other interfaces be implemented:

interface Equatable {
  fn IsEqual[self: Self](rhs: Self) -> bool;
}

// `Iterable` requires that `Equatable` is implemented.
interface Iterable {
  require Self impls Equatable;
  fn Advance[addr self: Self*]();
}

The extend keyword is used to extend another interface. If interface Derived extends interface Base, Base’s interface is both required and all its methods are included in Derived’s interface.

// `Hashable` extends `Equatable`.
interface Hashable {
  extend Equatable;
  fn Hash[self: Self]() -> u64;
}
// `Hashable` is equivalent to:
interface Hashable {
  require Self impls Equatable;
  alias IsEqual = Equatable.IsEqual;
  fn Hash[self: Self]() -> u64;
}

A type may implement the base interface implicitly by implementing all the methods in the implementation of the derived interface.

class Key {
  // ...
  extend impl as Hashable {
    fn IsEqual[self: Key](rhs: Key) -> bool { ... }
    fn Hash[self: Key]() -> u64 { ... }
  }
  // No need to separately implement `Equatable`.
}
var k: Key = ...;
k.Hash();
k.IsEqual(k);

Combining interfaces

The & operation on facet types allows you conveniently combine interfaces. It gives you all the names that don’t conflict.

interface Renderable {
  fn GetCenter[self: Self]() -> (i32, i32);
  // Draw the object to the screen
  fn Draw[self: Self]();
}
interface EndOfGame {
  fn SetWinner[addr self: Self*](player: i32);
  // Indicate the game was a draw
  fn Draw[addr self: Self*]();
}

fn F[T:! Renderable & EndOfGame](game_state: T*) -> (i32, i32) {
  game_state->SetWinner(1);
  return game_state->Center();
}

Names with conflicts can be accessed using a qualified member access expression.

fn BothDraws[T:! Renderable & EndOfGame](game_state: T*) {
  game_state->(Renderable.Draw)();
  game_state->(GameState.Draw)();
}

Named constraints

You may also declare a new facet type directly using “named constraints”. Named constraints can express requirements that multiple interfaces be implemented, and give you control over how name conflicts are handled. Named constraints have other applications and capabilities not covered here.

constraint Combined {
  require Self impls Renderable;
  require Self impls EndOfGame;
  alias Draw_Renderable = Renderable.Draw;
  alias Draw_EndOfGame = EndOfGame.Draw;
  alias SetWinner = EndOfGame.SetWinner;
}

fn CallItAll[T:! Combined](game_state: T*, int winner) {
  if (winner > 0) {
    game_state->SetWinner(winner);
  } else {
    game_state->Draw_EndOfGame();
  }
  game_state->Draw_Renderable();
  // Can still use a qualified member access expression
  // for names not defined in the named constraint.
  return game_state->(Renderable.Center)();
}

Type erasure

Inside a generic function, the API of a facet argument is erased except for the names defined in the facet type. An equivalent model is to say an archetype is used for type checking and name lookup when the actual type is not known in that scope. The archetype has members dictated by the facet type.

For example: If there were a class CDCover defined this way:

class CDCover  {
  extend impl as Printable {
    ...
  }
}

it can be passed to this PrintIt function:

fn PrintIt[T:! Printable](p: T*) {
  p->Print();
}

Inside PrintIt, T is an archetype with the API of Printable. A call to PrintIt with a value of type CDCover erases everything except the members or Printable. This includes the type connection to CDCover, so it is illegal to cast from T to CDCover.

Adapting types

Carbon has a mechanism called adapting types to create new types that are compatible with existing types but with different interface implementations. This could be used to add or replace implementations, or define implementations for reuse.

In this example, we have multiple ways of sorting a collection of Song values.

class Song { ... }

class SongByArtist {
  extend adapt Song;
  extend impl as Comparable { ... }
}

class SongByTitle {
  extend adapt Song;
  extend impl as Comparable { ... }
}

Values of type Song may be cast to SongByArtist or SongByTitle to get a specific sort order.

Interface inputs and outputs

Associated constants and interface parameters allow function signatures to vary with the implementing type. The biggest difference between these is that associated constants (“outputs”) may be deduced from a type, and types can implement the same interface multiple times with different interface parameters (“inputs”).

Associated constants

Expect parts of function signatures that vary in an interface to be associated constants by default. Since associated constants may be deduced, they are more convenient to use. Imagine a Stack interface. Different types implementing Stack will have different element types:

interface Stack {
  let ElementType:! Movable;
  fn Push[addr self: Self*](value: ElementType);
  fn Pop[addr self: Self*]() -> ElementType;
  fn IsEmpty[addr self: Self*]() -> bool;
}

ElementType is an associated constant of the interface Stack. Types that implement Stack give ElementType a specific value that is some type (really, facet) implementing Movable. Functions that accept a type implementing Stack can deduce the ElementType from the stack type.

// ✅ This is allowed, since the type of the stack will determine
// `ElementType`.
fn PeekAtTopOfStack[StackType:! Stack](s: StackType*)
    -> StackType.ElementType;

Parameterized interfaces

Parameterized interfaces are commonly associated with overloaded operators. Imagine an interface for determining if two values are equivalent that allows those types to be different. An element in a hash map might have type Pair(String, i64) that implements both Equatable(String) and Equatable(Pair(String, i64)).

interface Equatable(T:! type) {
  fn IsEqual[self: Self](compare_to: T) -> bool;
}

T is a parameter to interface Equatable. A type can implement Equatable multiple times as long as each time it is with a different value of the T parameter. Functions may accept types implementing Equatable(i32) or Equatable(f32). Functions can’t accept types implementing Equatable(T) in general, unless some other parameter determines T.

// ✅ This is allowed, since the value of `T` is determined by the
// `v` parameter.
fn FindInVector[T:! type, U:! Equatable(T)](v: Vector(T), needle: U)
    -> Optional(i32);

// ❌ This is forbidden. Since `U` could implement `Equatable`
// multiple times, there is no way to determine the value for `T`.
// Contrast with `PeekAtTopOfStack` in the associated constant
// example.
fn CompileError[T:! type, U:! Equatable(T)](x: U) -> T;

Constraints

Facet types can be further constrained using a where clause:

fn FindFirstPrime[T:! Container where .Element = i32]
    (c: T, i: i32) -> Optional(i32) {
  // The elements of `c` have type `T.Element`, which is `i32`.
  ...
}

fn PrintContainer[T:! Container where .Element impls Printable](c: T) {
  // The type of the elements of `c` is not known, but we do know
  // that type satisfies the `Printable` interface.
  ...
}

Constraints limit the types that the generic function can operate on, but increase the knowledge that may be used in the body of the function to operate on values of those types.

Constraints are also used when implementing an interface to specify the values of associated constants.

class Vector(T:! Movable) {
  extend impl as Stack where .ElementType = T { ... }
}

Parameterized impl declarations

Implementations can be parameterized to apply to multiple types. Those parameters can have constraints to restrict when the implementation applies. When multiple implementations apply, there is a rule to pick which one is considered the most specific:

All parameters in each impl declaration are replaced with question marks ?. This is called the type structure of the impl declaration.
Given two type structures, find the first difference when read from left-to-right. The one with a ? is less specific, the one with a concrete type name in that position is more specific.
If there is more than one impl declaration with the most specific type structure, pick the one listed first in the priority ordering.

To ensure coherence, an impl may only be declared in a library defining some name from its type structure. If a library defines multiple implementations with the same type structure, they must be listed in priority order in a prioritization block.

Operator overloading

To overload an operator, implement the corresponding interface from the standard library. For example, to define how the unary - operator behaves for a type, implement the Negatable interface for that type. The interfaces and rewrites used for a given operator may be found in the expressions design.

As a convenience, there is a shortcut for defining an implementation that supports any type implicitly convertible to a specified type, using like:

// Support multiplying values of type `Distance` with
// values of type `f64` or any type implicitly
// convertible to `f64`.
impl Distance as MultipliableWith(like f64) ...

Future work

Functions should have a way to accept types that vary at runtime.
You should have the ability to mark entities as upcoming or deprecated to support evolution.
There should be a way to define generic associated and higher-ranked/kinded types.