[swift-evolution] [Proposal] Change Void meaning

Tue Jun 13 12:15:14 CDT 2017

> On Jun 13, 2017, at 4:41 AM, Xiaodi Wu <xiaodi.wu at gmail.com> wrote:
> 
> On Tue, Jun 13, 2017 at 3:06 AM, John McCall <rjmccall at apple.com <mailto:rjmccall at apple.com>> wrote:
>> On Jun 13, 2017, at 3:30 AM, Jérémie Girault <jeremie.girault at gmail.com <mailto:jeremie.girault at gmail.com>> wrote:
>> 
>> Exactly, 
>> The reflexion behind it is: 
>> 
>> - Let's understand that 0110 and other tuple SE are important for the compiler, we do not want them to rollback
>> - However we have number of regressions for generics / functional programmers
>> - Let’s solve this step by step like a typical problem
>> 
>> - Step 0 is adressing this Void tuple of size zero :
>> 	- Zero is in many problems of CS an edge case, so let’s handle this case first
>> 	- The compiler knows what Void is, and its only value (or non-value)
>> 	- It was handled historically by the compiler because of implicit side effects
>> 	- Let’s handle it explicitely with rules in current context
>> 	- one effect of the proposal is source compatibility
>> 	- but the goal is to build atop and strengthen 0110, 0066 and other tuple-related SE
>> 
> 
> There are four difficulties I see with this proposal.
> 
> The first is that it is a first step that quite clearly does not lead to anything.  It resolves a difficulty with exactly one case of function composition, but we would need completely different solutions to handle any of the other compositional regressions of SE-0110.
> 
> The second is that it's a huge source of complexity for the type system.  The type checker would not be able to do even rudimentary type matching, e.g. when checking a call, without having first resolved all of the argument and parameter types to prove that they are not Void.  This would probably render it impossible to type-check many programs without some ad-hoc rule of inferring that certain types are not Void.  It would certainly make type-checking vastly more expensive.
> 
> The third is that it is not possible to prevent values of Void from existing, because (unlike Never, which cannot be constructed) they are always created by returning from a Void-returning function, and a generic function can do anything it likes with that value — turn it into an Any, store it in an Array, whatever.  The proposal seems to only consider using the value as a parameter.
> 
> Hang on, though. If Jérémie is interested only in addressing the issue of Void as a parameter and his idea can be adequately carried out by inferring a default value of Void for every parameter of type Void, this should be a fairly self-contained change, should it not? And would the impact on the cost of type checking really be vastly greater in that case?

If the proposal was phrased in terms of defaults, e.g. "trailing parameters do not require a matching argument if they have Void type", then yes, that would be implementable because it still admits a "local" reduction on call constraints, one which does not need to immediately reason about the actual types of arguments.  It is not clear that this rule allows function compositions of the sort that Jérémie is looking for, though.

Anyway, that is not the proposal; the proposal is that parameters — in any position — are simply removed from the parameter sequence if they have Void type.  In order to allow composition (i.e. f(g(x)), where g: X -> Void), you then need a matching rule that arguments are dropped from the argument sequence (for purposes of type-checking) if they have Void type.  Either of these rules is sufficient to turn the reduction of function-type matches into an extremely messy combinatoric matching problem where e.g. (τ0, Int) can be passed to a function taking (Int, τ1) if we can decide that τ0 == τ1 == Void.

> This idea is now rather intriguing to me because it extends beyond just addressing one symptom of SE-0110. Swift allows us to omit the spelling out of return types that are Void, it allows warning-free discarding of return values that are Void, etc. This could add a nice consistency and rationalize some of the weirdness of passing a value that is stipulated by the parameter type.
> 
> Finally, it would allow a lot of inadvertent errors with the use of generic functions, because any argument of unconstrained type could be accidentally specialized with Void.  For example, if you forgot to pass an argument to this function, it would simply infer T=Void:
>   func append<T>(value: T)
> It seems more likely that this would lead to unexpected, frustrating bugs than that this would actually be desired by the programmer.  You really just want this to kick in in more generic situations.
> 
> Hmm, at first glance, that seemed like it could be bad. But if, say, a particular collection can store an element of type Void, is it so undesirable to allow `append()` to append Void?

append on a Collection is not an unconstrained generic; its parameter type is determined by the type of the collection.  Perhaps this was a poorly-chosen example.

John.

> 
> 
> John.
> 
>> 
>> —
>> very short reply expected - vsre.info <http://vsre.info/>
>> Jérémie Girault
>> 
>> On 13 juin 2017 at 00:44:52, Xiaodi Wu (xiaodi.wu at gmail.com <mailto:xiaodi.wu at gmail.com>) wrote:
>> 
>>> On Mon, Jun 12, 2017 at 5:38 PM, Xiaodi Wu <xiaodi.wu at gmail.com <mailto:xiaodi.wu at gmail.com>> wrote:
>>> On Mon, Jun 12, 2017 at 5:25 PM, Jérémie Girault <jeremie.girault at gmail.com <mailto:jeremie.girault at gmail.com>> wrote:
>>> 
>>> 
>>> —
>>> very short reply expected - vsre.info <http://vsre.info/>
>>> Jérémie Girault
>>> 
>>> On 12 juin 2017 at 23:56:37, Xiaodi Wu (xiaodi.wu at gmail.com <mailto:xiaodi.wu at gmail.com>) wrote:
>>> 
>>>> On Mon, Jun 12, 2017 at 4:47 PM, Jérémie Girault <jeremie.girault at gmail.com <mailto:jeremie.girault at gmail.com>> wrote:
>>>> - Void as arguments is pretty common when using generics, that’s a core point of this proposal. An maybe that’s why we misunderstood ourselves (around 0110 / 0066). This proposal addresses arguments.
>>>> - maybe it should be revised around this ? Simple example : 
>>>> 
>>>> `typealias Callback<T> = (T) -> Void` -> `Callback<Void>` will give `(Void) => Void`. 
>>>> 
>>>> It was acceptable before swift4 but no more. However nobody cares about this `Void` argument and actually we know it’s value. So why let the developer type it ?
>>>> 
>>>> Ah, I see. The purpose of SE-0029...SE-0110 was to make it possible to distinguish an argument list `(Void)` from an argument list `()`. This does cause some verbosity where previously users relied on implicit tuple splatting. Ideally, we would bring back some syntactic sugar to make this more ergonomic. But, whether or not the spelling is made more user-friendly, the point here is that _everybody_ should care about this `Void` argument.
>>> 
>>> It is still be typechecked and appropriate errors should be reported to the user so _nobody_ will ignore it.
>>> 
>>> But with the proposal the code will be striped out of Void arguments at compile-time. I think it's a win for the developer on a lot of grounds. The fact that this proposal integrates with the type-system is also important.
>>> 
>>> If you are not comfortable about Void being stripped, we can also discuss alternatives: someone was suggesting me that it would be possible to replace :
>>> 
>>> ```
>>> 
>>> func foo<T, U, V>(t: T, u: U) -> V {
>>> 
>>>   // do something with t and u
>>> 
>>>   // return some V
>>> 
>>> }
>>> 
>>> ```
>>> 
>>> with
>>> 
>>> ```
>>> 
>>> func foo<Void, Int, String>(u: Int) -> String { let t = ()
>>> 
>>>   // do something with t and u
>>> 
>>>   // return some V
>>> 
>>> }
>>> 
>>> ```
>>> 
>>> or
>>> 
>>> ```
>>> 
>>> func foo<Void, Int, String>(t: Void = (), u: Int) -> String {
>>> 
>>>   // do something with t and u
>>> 
>>>   // return some V
>>> 
>>> }
>>> 
>>> ```
>>> 
>>> I don’t know what you would consider more effective or elegant (at an implementation level) but it’s the same result for the developper.
>>> 
>>> 
>>> Ah, but I think I catch your drift with the last example. Is this a more general point that the compiler should treat every parameter of type Void as having an implied default value of Void? That would be an interesting idea.
>>> 
>>> What is the goal of such changes? Is it to allow you to write `foo()` instead of `foo(())` for a function `foo` of type `(T) -> Void`?
>>> 
>>> If so, then I think what you're seeking to do is reverse SE-0066 (as Vladimir points out), which explicits details how it's is an intentional change to require such a spelling. I think you're starting from the premise that this is unintended or undesirable, when in fact it is deliberate and approved.
>>> 
>>> It is also, unless I'm mistaken, not the issue that was raised initially with respect to SE-0110, which had to do with the extra boilerplate of destructuring a tuple inside a closure, something that was not so obvious before implementation.
>>> 
>>>> My point here is that `Void` should be “striped” by “reducing” argument list signatures.
>>>> 
>>>> —
>>>> very short reply expected - vsre.info <http://vsre.info/>
>>>> Jérémie Girault
>>>> 
>>>> On 12 juin 2017 at 19:15:18, John McCall (rjmccall at apple.com <mailto:rjmccall at apple.com>) wrote:
>>>> 
>>>>> 
>>>>>> On Jun 12, 2017, at 4:48 AM, Jérémie Girault via swift-evolution <swift-evolution at swift.org <mailto:swift-evolution at swift.org>> wrote:
>>>>>> 
>>>>>> Hi here,
>>>>>> 
>>>>>> As I tested swift4 in xcode9b1 I noticed a lot of regressions about tuples usage.
>>>>>> 
>>>>>> After documenting myself about the changes which happened, I thought that they could be improved. Instead of fighting these propositions (which make sense), I wanted create a few proposal which would improve these recent changes with a few simple rules.
>>>>>> 
>>>>>> My propositions are based on the recent decisions and in the continuation of SE-0110. The first one is about Void.
>>>>>> Void is historically defined as the type of the empty tuple. The reason of this is that arguments were initially considered as tuple.
>>>>> 
>>>>> The dominant consideration here was always return types, not parameters.  I'm not sure there was ever much point in writing Void in a parameter list, but whatever reasons there were surely vanished with SE-0066.
>>>>> 
>>>>> Note that 'void' in C was originally exclusively a return type.  ANSI gave it a new purpose it with void*, but the meaning is totally unrelated.
>>>>> 
>>>>> John.
> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-evolution/attachments/20170613/6ab77b1c/attachment.html>