Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Generics details 8: interface default and final members #990

Merged
merged 22 commits into from
Jan 21, 2022
Merged
Show file tree
Hide file tree
Changes from 21 commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
164 changes: 152 additions & 12 deletions docs/design/generics/details.md
Original file line number Diff line number Diff line change
Expand Up @@ -92,12 +92,14 @@ SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
- [`final` impls](#final-impls)
- [Libraries that can contain `final` impls](#libraries-that-can-contain-final-impls)
- [Comparison to Rust](#comparison-to-rust)
- [Interface members with definitions](#interface-members-with-definitions)
- [Interface defaults](#interface-defaults)
- [`final` members](#final-members)
- [Future work](#future-work)
- [Dynamic types](#dynamic-types)
- [Runtime type parameters](#runtime-type-parameters)
- [Runtime type fields](#runtime-type-fields)
- [Abstract return types](#abstract-return-types)
- [Interface defaults](#interface-defaults)
- [Evolution](#evolution)
- [Testing](#testing)
- [Operator overloading](#operator-overloading)
Expand Down Expand Up @@ -4270,6 +4272,154 @@ differences between the Carbon and Rust plans:
ordering on type structures, picking one as higher priority even without one
being more specific in the sense of only applying to a subset of types.

## Interface members with definitions

Interfaces may provide definitions for members, such as a function body for an
associated function or method or a value for an associated constant. If these
definitions may be overridden in implementations, they are called "defaults."
Otherwise they are called "final members."

### Interface defaults

An interface may provide a default implementation of methods in terms of other
methods in the interface.

```
interface Vector {
fn Add[me: Self](b: Self) -> Self;
fn Scale[me: Self](v: f64) -> Self;
// Default definition of `Invert` calls `Scale`.
fn Invert[me: Self]() -> Self {
return me.Scale(-1.0);
}
}
```

An impl of that interface for a type may omit a definition of `Invert` to use
the default, or provide a definition to override the default.

Interface defaults are helpful for [evolution](#evolution), as well as reducing
boilerplate. Defaults address the gap between the minimum necessary for a type
to provide the desired functionality of an interface and the breadth of API that
developers desire. As an example, in Rust the
[iterator trait](https://doc.rust-lang.org/std/iter/trait.Iterator.html) only
has one required method but dozens of "provided methods" with defaults.

Defaults may also be provided for associated constants, such as associated
types, and interface parameters, using the `= <default value>` syntax.

```
interface Add(Right:! Type = Self) {
let Result:! Type = Self;
fn DoAdd[me: Self](right: Right) -> Result;
}

impl String as Add() {
// Right == Result == Self == String
fn DoAdd[me: Self](right: Self) -> Self;
}
```

Note that `Self` is a legal default value for an associated type or type
parameter. In this case the value of those names is not determined until `Self`
is, so `Add()` is equivalent to the constraint:

```
// Equivalent to Add()
constraint AddDefault {
extends Add(Self);
}
```

Note also that the parenthesis are required after `Add`, even when all
parameters are left as their default values.

More generally, default expressions may reference other associated types or
`Self` as parameters to type constructors. For example:

```
interface Iterator {
let Element:! Type;
let Pointer:! Type = Element*;
}
```

Carbon does **not** support providing a default implementation of a required
interface.

```
interface TotalOrder {
fn TotalLess[me: Self](right: Self) -> Bool;
// ❌ Illegal: May not provide definition
// for required interface.
impl PartialOrder {
fn PartialLess[me: Self](right: Self) -> Bool {
return me.TotalLess(right);
}
}
}
```

The workaround for this restriction is to use a [blanket impl](#blanket-impls)
instead:

```
interface TotalOrder {
fn TotalLess[me: Self](right: Self) -> Bool;
impl PartialOrder;
}

external impl [T:! TotalOrder] T as PartialOrder {
zygoloid marked this conversation as resolved.
Show resolved Hide resolved
fn PartialLess[me: Self](right: Self) -> Bool {
return me.TotalLess(right);
}
}
```

Note that by the [orphan rule](#orphan-rule), this blanket impl must be defined
in the same library as `PartialOrder`.

**Comparison with other languages:** Rust supports specifying defaults for
[methods](https://doc.rust-lang.org/book/ch10-02-traits.html#default-implementations),
[interface parameters](https://doc.rust-lang.org/book/ch19-03-advanced-traits.html#default-generic-type-parameters-and-operator-overloading),
and
[associated constants](https://doc.rust-lang.org/reference/items/associated-items.html#associated-constants-examples).
Rust has found them valuable.

### `final` members
zygoloid marked this conversation as resolved.
Show resolved Hide resolved
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Super minor (and fine for a follow-up): there isn't much motivation here in the design for this feature compared to others. The motivation in the proposal makes sense to me, but maybe three is a way to briefly summarize that so it doesn't read like an omission here?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Done in 620d6de


As an alternative to providing a definition of an interface member as a default,
members marked with the `final` keyword will not allow that definition to be
overridden in impls.

```
interface TotalOrder {
fn TotalLess[me: Self](right: Self) -> Bool;
final fn TotalGreater[me: Self](right: Self) -> Bool {
return right.TotalLess(me);
}
}

class String {
impl as TotalOrder {
fn TotalLess[me: Self](right: Self) -> Bool { ... }
// ❌ Illegal: May not provide definition of final
// method `TotalGreater`.
fn TotalGreater[me: Self](right: Self) -> Bool { ... }
}
}

interface Add(T:! Type = Self) {
// `AddWith` *always* equals `T`
final let AddWith:! Type = T;
// Has a *default* of `Self`
let Result:! Type = Self;
fn DoAdd[me: Self](right: AddWith) -> Result;
}
```

Note that this applies to associated entities, not interface parameters.

## Future work

### Dynamic types
Expand Down Expand Up @@ -4305,17 +4455,6 @@ In Swift, there are discussions about implementing this feature under the name
[4](https://forums.swift.org/t/se-0244-opaque-result-types-reopened/22942),
Swift is considering spelling this `<V: Collection> V` or `some Collection`.

### Interface defaults

Rust supports specifying defaults for
[interface parameters](https://doc.rust-lang.org/book/ch19-03-advanced-traits.html#default-generic-type-parameters-and-operator-overloading),
[methods](https://doc.rust-lang.org/book/ch10-02-traits.html#default-implementations),
[associated constants](https://doc.rust-lang.org/reference/items/associated-items.html#associated-constants-examples).
We should support this too. It is helpful for evolution, as well as reducing
boilerplate. Defaults address the gap between the minimum necessary for a type
to provide the desired functionality of an interface and the breadth of API that
user's desire.

### Evolution

There are a collection of use cases for making different changes to interfaces
Expand Down Expand Up @@ -4401,3 +4540,4 @@ parameter, as opposed to an associated type, as in `N:! u32 where ___ >= 2`.
- [#920: Generic parameterized impls (details 5)](https://github.com/carbon-language/carbon-lang/pull/920)
- [#950: Generic details 6: remove facets](https://github.com/carbon-language/carbon-lang/pull/950)
- [#983: Generic details 7: final impls](https://github.com/carbon-language/carbon-lang/pull/983)
- [#990: Generics details 8: interface default and final members](https://github.com/carbon-language/carbon-lang/pull/990)
1 change: 1 addition & 0 deletions docs/design/lexical_conventions/words.md
Original file line number Diff line number Diff line change
Expand Up @@ -49,6 +49,7 @@ The following words are interpreted as keywords:
- `else`
- `extends`
- `external`
- `final`
- `fn`
- `for`
- `friend`
Expand Down
155 changes: 155 additions & 0 deletions proposals/p0990.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,155 @@
# Generics details 8: interface default and final members

<!--
Part of the Carbon Language project, under the Apache License v2.0 with LLVM
Exceptions. See /LICENSE for license information.
SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
-->

[Pull request](https://github.com/carbon-language/carbon-lang/pull/990)

<!-- toc -->

## Table of contents

- [Problem](#problem)
- [Background](#background)
- [Proposal](#proposal)
- [Rationale based on Carbon's goals](#rationale-based-on-carbons-goals)
- [Alternatives considered](#alternatives-considered)
- [Defaulting to less specialized impls](#defaulting-to-less-specialized-impls)
- [Allow default implementations of required interfaces](#allow-default-implementations-of-required-interfaces)
- [Don't support `final`](#dont-support-final)

<!-- tocstop -->

## Problem

Rust has found that allowing interfaces to define default values for its
associated entities is valuable:

- Helps with evolution by reducing the changes needed to add new members to an
interface.
- Reduces boilerplate when some value is more common than others.
- Addresses the gap between the minimum necessary for a type to provide the
desired functionality of an interface and the breadth of API that user's
desire.

Carbon would benefit in the same ways.

## Background

Rust supports specifying defaults for
[methods](https://doc.rust-lang.org/book/ch10-02-traits.html#default-implementations),
[interface parameters](https://doc.rust-lang.org/book/ch19-03-advanced-traits.html#default-generic-type-parameters-and-operator-overloading),
and
[associated constants](https://doc.rust-lang.org/reference/items/associated-items.html#associated-constants-examples).

## Proposal

This proposal defines both how defaults for interface members are specified in
Carbon code as well as final interface members in the
[generics details design doc](/docs/design/generics/details.md#interface-defaults).

## Rationale based on Carbon's goals

This proposal advances these goals of Carbon:

- [Performance-critical software](/docs/project/goals.md#performance-critical-software):
Final members of interfaces can avoid some dynamic dispatch overhead.
- [Software and language evolution](/docs/project/goals.md#software-and-language-evolution):
Defaults simplify adding new members to an interface without having to
simultaneously update all impls of that interface.
- [Code that is easy to read, understand, and write](/docs/project/goals.md#code-that-is-easy-to-read-understand-and-write):
Defaults both reduce boilerplate, making code easier to read and write.
Marking interface members as `final` makes the code more predictable to
users of that member.

## Alternatives considered

### Defaulting to less specialized impls

Rust has observed
([1](https://rust-lang.github.io/rfcs/1210-impl-specialization.html#default-impls),
[2](http://aturon.github.io/tech/2015/09/18/reuse/)) that interface defaults
could be generalized into a feature for reusing definitions between impls. This
would involve allowing more specific implementations to be incomplete and reuse
more general implementations for anything unspecified.

However,
[they also observed](http://smallcultfollowing.com/babysteps/blog/2016/09/29/distinguishing-reuse-from-override/):

> [To be sound,] if an impl A wants to reuse some items from impl B, then impl A
> must apply to a subset of impl B's types. ... This implies we will have to
> separate the concept of "when you can reuse" (which requires subset of types)
> from "when you can override" (which can be more general).

This is a source of complexity that we don't want in Carbon. If we do eventually
support inheritance of implementation between impls in Carbon, it will do this
by explicitly identifying the impl being reused instead of having it be
determined by their specialization relationship.

### Allow default implementations of required interfaces

Here are the reasons we considered for not allowing interfaces to provide
default implementations of interfaces they require:

- This feature would lead to incoherence unless types implementing
`TotalOrder` also must explicitly implement `PartialOrder`, possibly with an
empty definition. The problem arises since querying whether `PartialOrder`
is implemented for a type does not require that an implementation of
`TotalOrder` be visible.
- It would be unclear how to resolve the ambiguity of which default to use
when two different interfaces provide different defaults for a common
interface requirement.
- It would be ambiguous whether the required interface should be
[external](/docs/design/generics/terminology.md#external-impl) or
[internal](/docs/design/generics/terminology.md#internal-impl) unless
`PartialOrder` is implemented explicitly.
- There would be a lot of overlap between default impls and blanket impls.
Eliminating default impls keeps the language smaller and simpler.

The rules for blanket impls already provide resolution of the questions about
coherence and priority and make it clear that the provided definition of the
required interface will be external.

### Don't support `final`

There are a few reasons to support `final` on associated entities in the
interface:

- Clarity of intent when default methods are just to provide an expanded API
for the convenience of callers, reducing the uncertainty about what code is
called.
- Matches the functionality available to base classes in C++, namely
non-virtual functions.
- Could reduce the amount of dynamic dispatch needed when using an interface
in a `DynPtr`.

The main counter-argument is that you could achieve something similar using a
`final` impl:

```
interface I {
fn F();
final fn CallF() { F(); }
}
```

could be replaced by:

```
interface IImpl {
fn F();
}
interface I {
extends IImpl;
fn CallF();
}
final impl (T:! IImpl) as I {
fn CallF() { F(); }
}
```

This is both verbose and a bit awkward to use since you would need to
`impl as IImpl` but use `I` in constraints.