[swift-evolution] [Pitch] Small bit of sugar for enum case with Void associated value

Xiaodi Wu xiaodi.wu at gmail.com
Tue Jul 25 18:54:47 CDT 2017

On Tue, Jul 25, 2017 at 16:35 Robert Bennett <rltbennett at icloud.com> wrote:

> Optional.some(Void) gives me "Expected member name or constructor call
> after type name”.

That would appear to be straightforwardly a compiler bug. For instance, the
following compiles:

let _Void = Void
let foo = Optional.some(_Void)
// I'll discuss the warnings you see in a bit, below.

In addition, the following compiles just fine, which is the crux of your
use case:

enum Foo<T> {
    case bar(T)
    case baz

let x = Foo.bar(Void)

And as I said, the compiler should not be allowed to use the absence of
> parentheses to infer a Void generic type; only when the generic type is
> known to be Void may the parentheses be omitted. Under this proposal,
> Optional.some would not be allowed.

There is no precedent for such a rule in Swift. Inferring Void is allowed
but results in a warning under certain circumstances, but it is not
forbidden. This has been discussed on the list before and it is not an
oversight that it's a warning and not an error, and that the warning is
about storing the initialized result and not about inferring the type.

To be clear, if the rule you propose is permitted, `Optional.some` *would*
be inferred to be `Optional.some(Void)` unless type inference rules are
revisited (which, I would argue, is entirely overkill in terms of the scope
of that undertaking as compared to the magnitude of the issue being
addressed). Instead, there would be a warning if you tried to bind
`Optional.some` to a variable, but you would absolutely be allowed to pass
it to a function: `callSomeOtherFunction(Optional.some!)` would be an
alternative spelling for `callSomeOtherFunction(Void)`; effectively,
`Optional.some!` would become a synonym for `Void` under certain
circumstances but not others.

The scenario I’m working with, which inspired this thought, is roughly:
> enum Result<T> {
> case success(T)
> case failure(Error)
> }
> func doSomething(completion: @escaping (Result<Void>)->()) { /* At some
> point, call completion(.success(())) */ }
> The idea is that in case of success, I merely want to indicate success,
> whereas in case of an error I want to pass the error through to the
> completion. This requires calling `completion(.success(()))`.

This is, as far as I can tell, essentially the same use case that SE-0110
rendered problematic. As the core team wrote in their decision there, it's
clear that Swift doesn't have a good story for how to support certain
coding styles that involve Void arguments. For the moment, a limited and
temporary exception was approved to roll back parts of SE-0110, but the
idea is that all of this needs to be revisited.

I'd agree that your use case here is a good example of something that
should be taken into account as part of that re-examination, but IMO
inventing one or several new rules specific to generic enums cannot be the
way to go--especially since it is certain that SE-0110 and related issues
*will* be revisited and implicit or explicit tuple splatting will be given
great consideration. By that time, implementation of the existing and
approved proposal on enums will mean that the exact same issue will be
addressed automatically for enum cases.

I don’t think I should have to make another enum with the same cases but no
> associated value for success (and no generic type) to make this work nicely
> — that makes things brittle in case I ever want to replace Void with a
> different type.

If I understand Swift's error handling rationale correctly, a Result type
with no associated value for success would be, _by design_, an
Optional<Error>. Why wouldn't you use that in this scenario? (Well,
actually, if I understand Swift's error handling rationale, the intended
way to spell this in Swift is `throws -> Void` (aka `throws`)).

> On Jul 25, 2017, at 5:20 PM, Xiaodi Wu <xiaodi.wu at gmail.com> wrote:
> Yes, I discussed this some time back. It breaks some things with
> overloads, if I recall, but off the top of my head not recalling what.
> At some point, I also suggested regarding any case without associated
> values as equivalent to a case with an associated value of type Void. This
> was never adopted. There _is_ an underlying issue with that idea, as the
> property “foo” is not the same as the method “foo(_: Void)”, and it is even
> permitted to have both of these on the same type. The purpose of the most
> recent Swift proposal on enum cases was to align the syntax with functions
> as closely as possible; if overloading the base name of a case (as proposed
> back then to favorable reviews) is ever allowed, then the same will
> naturally be possible with enums (that is, a case named “foo” and another
> named “foo(_: Void)” in the same type).
> In the case of concrete enums, a case with associated value of type Void
> can be given a default value once they are supported (already approved);
> the trouble here is that in the generic case presented here the same is not
> possible. It is unclear to me whether this special rule would see broad
> use, as it appears to come into play *only* for a generic enum. Moreover,
> in the one scenario I can think of, it reads cryptically: “let foo =
> Optional.some” would then be allowed (while “let bar = Optional.none” still
> would not). The upside is that it avoids having to write four or six
> letters (“Optional.some(Void())” is never necessary, as it’s also just
> “Optional.some(Void)”, and that reads acceptably in my view).
> On Tue, Jul 25, 2017 at 15:26 David Sweeris via swift-evolution <
> swift-evolution at swift.org> wrote:
>> On Jul 25, 2017, at 12:38, Robert Bennett via swift-evolution <
>> swift-evolution at swift.org> wrote:
>> Currently if you have the following enum:
>> enum E<T> {
>> case c(T)
>> }
>> then if T is Void, you have to write one of the following:
>> let x: E<Void> = .c(Void())
>> let y: E<Void> = .c(())
>> Looks awkward, no? In this case you can omit `<Void>` after `E` because
>> it can be inferred, but if writing a (non-generic) function taking an
>> argument of type `E<Void>`, then the `<Void>` cannot be omitted, and you
>> still have to write `.c(())` for the case name.
>> I’m proposing that for enum cases with a single associated value of Void
>> type, or of a generic type that is equal to Void in some instance, you may
>> omit the parentheses altogether and merely write
>> let x: E<Void> = .c
>> The rationale is twofold: first, double parentheses just looks bad;
>> second, there is only a single value of type Void, which means the
>> associated value of `.c` is trivially inferable, and hence should be
>> omissible.
>> I am not proposing that a bare `E.c` imply a type of `E<Void>` — `E.c`
>> should still be illegal in the absence of specification of the generic type
>> — only that when the type is known to be `E<Void>`, `.c` can replace
>> `.c(())`.
>> Thoughts?
>> My first response is +1
>> My second response is to ask just how much would it break things to
>> expand this and allow omitting the argument anywhere its type is known to
>> be Void? Maybe by implicitly providing a default value of `()` wherever
>> there's a `Void` argument? Like make `func foo(x: Void) {...}` implicitly
>> become `func foo(x: Void = ()) {...}`? I have a sneaking suspicion that
>> this question's already been asked, but I'm not sure.
>> - Dave Sweeris
>> _______________________________________________
>> swift-evolution mailing list
>> swift-evolution at swift.org
>> https://lists.swift.org/mailman/listinfo/swift-evolution
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-evolution/attachments/20170725/b26587ea/attachment.html>

More information about the swift-evolution mailing list