[swift-evolution] Subclass Existentials

Mon Jan 30 13:36:37 CST 2017

Given that supporting value subtyping or self-conforming existentials in
the future would not be a code-breaking change AFAICT, I don't think it
makes sense to spend much energy discussing it now. If we do end up getting
value subtyping for Swift 4, we can have a compiler engineer estimate how
much work it would require to support value subtyping constraints on
existentials, and then amend the proposal to loosen the restrictions if the
amount of work makes sense for the Swift 4 timeframe. If no significant
design work is required, it should only take a few days to get a revision
through. A formal proposal might not even be required.

Tangentially, I don't think the type system should be any more complicated
than it needs to be, and I think every addition to the type system needs to
rigorously justify its existence (allowing for clever tricks or a slightly
more elegant representation of a type I don't consider sufficient). It
becomes significantly more difficult to reason about the correctness of
something like a type system as the number of features that interact with
each other increases (c.f. http://io.livecode.ch/learn/namin/unsound).

For this reason, I would be a strong (-1) for allowing `Never`-equivalent
bounds to be defined (at least in this iteration of the proposal). If we do
allow multiple class bounds, the rule should be they must all be part of
the same inheritance chain and that the lowest class on the chain should be
the effective bound. This satisfies the need for typealias composition
without introducing special rules that allow some spellings but not other,
equivalent ones.

Best,
Austin

On Mon, Jan 30, 2017 at 7:42 AM, Matthew Johnson via swift-evolution <
swift-evolution at swift.org> wrote:

>
> On Jan 29, 2017, at 10:47 PM, Slava Pestov <spestov at apple.com> wrote:
>
>
> On Jan 29, 2017, at 2:05 PM, Matthew Johnson via swift-evolution <
> swift-evolution at swift.org> wrote:
>
>
> On Jan 29, 2017, at 3:52 PM, Xiaodi Wu <xiaodi.wu at gmail.com> wrote:
>
> On Sun, Jan 29, 2017 at 3:35 PM, Matthew Johnson <matthew at anandabits.com>
> wrote:
>
>>
>> On Jan 29, 2017, at 3:24 PM, Xiaodi Wu <xiaodi.wu at gmail.com> wrote:
>>
>> On Sun, Jan 29, 2017 at 3:11 PM, Matthew Johnson <matthew at anandabits.com>
>>  wrote:
>>
>>>
>>> On Jan 29, 2017, at 3:05 PM, Xiaodi Wu <xiaodi.wu at gmail.com> wrote:
>>>
>>> On Sun, Jan 29, 2017 at 2:40 PM, Matthew Johnson <matthew at anandabits.com
>>> > wrote:
>>>
>>>>
>>>> On Jan 29, 2017, at 2:25 PM, Xiaodi Wu <xiaodi.wu at gmail.com> wrote:
>>>>
>>>> On Sun, Jan 29, 2017 at 2:16 PM, Matthew Johnson <matthew at anandabits.co
>>>> m> wrote:
>>>>
>>>>>
>>>>> On Jan 29, 2017, at 2:01 PM, Xiaodi Wu <xiaodi.wu at gmail.com> wrote:
>>>>>
>>>>> On Sun, Jan 29, 2017 at 1:37 PM, Matthew Johnson <
>>>>> matthew at anandabits.com> wrote:
>>>>>
>>>>>>
>>>>>>
>>>>>> Sent from my iPad
>>>>>>
>>>>>> On Jan 29, 2017, at 12:58 PM, Xiaodi Wu via swift-evolution <
>>>>>> swift-evolution at swift.org> wrote:
>>>>>>
>>>>>> Cool. Another avenue of improvement here is relaxing the single-class
>>>>>> spelling rule for the sake of composing typealiases.
>>>>>>
>>>>>> As Matthew mentioned, if I have class Base and typealiases Foo = Base
>>>>>> & Protocol1 and Bar = Base & Protocol2, it'd be nice to allow Foo & Bar.
>>>>>>
>>>>>> It'd be nice to go one step further: given class Derived : Base, if I
>>>>>> have typealiases Foo2 = Base & Protocol1 and Bar2 = Derived & Protocol2,
>>>>>> then it could be permitted to write Foo2 & Bar2, since there is effectively
>>>>>> only one subclass requirement (Derived).
>>>>>>
>>>>>> As I understand it, the rationale for allowing only one subclass
>>>>>> requirement is that Swift supports only single inheritance. Thus, two
>>>>>> disparate subclass requirements Base1 & Base2 would make your existential
>>>>>> type essentially equivalent to Never. But Base1 & Base1 & Base1 is fine for
>>>>>> the type system, the implementation burden (though greater) shouldn't be
>>>>>> too awful, and you would measurably improve composition of typealiases.
>>>>>>
>>>>>>
>>>>>> Yes, this is what I was indicating in my post as well.
>>>>>>
>>>>>> Are you suggesting that Base1 & Base2 compose to a type that is
>>>>>> treated identically to Never do you think it should be an immediate
>>>>>> compiler error?  I remember having some discussion about this last year and
>>>>>> think somebody came up with a very interesting example of where the former
>>>>>> might be useful.
>>>>>>
>>>>>
>>>>> Last year's discussion totally eludes me for some reason. But sure, if
>>>>> deferring the error until runtime is actually useful then why not? In the
>>>>> absence of an interesting use case, though, I think it'd be nice for the
>>>>> compiler to warn you that Base1 & Base2 is not going to be what you want.
>>>>>
>>>>>
>>>>> Deferring to runtime isn’t what I mean.  If you try to actually *do*
>>>>> anything that requires an instance of `Base1 & Based` (which you almost
>>>>> always would) you would still get a compile time error.
>>>>>
>>>>> I managed to dig up the example from last year’s thread and it is
>>>>> definitely a good one:
>>>>>
>>>>> func intersection<T, U>(ts; Set<T>, us: Set<U>) -> Set<T & U>
>>>>>
>>>>> The desire is that we are always able to produce a result set.  When T
>>>>> & U is uninhabitable it will simply be an empty set just like Set<Never>
>>>>> has a single value which is the empty set.
>>>>>
>>>>
>>>> Currently, Set<Never> is impossible because Never is not Hashable :)
>>>>
>>>>
>>>> Ahh, good point.  I hadn’t tried it.  It can easily be made Hashable
>>>> with a simple extension though - this code compiles today:
>>>>
>>>> extension Never: Hashable {
>>>>     public var hashValue: Int { return 0 }
>>>> }
>>>> public func ==(lhs: Never, rhs: Never) -> Bool { return false }
>>>> let s = Set<Never>()
>>>>
>>>> Since concrete types *can't* be used, this example seems like it'd be
>>>> of little use currently. How widely useful would it be to have an
>>>> intersection facility such as this when T != U even if that restriction
>>>> were lifted, though? Seems like the only real useful thing you can do with
>>>> generic Set<T & U> is based on the fact that it'd be Set<Hashable>. Other
>>>> than those immediate thoughts, I'll have to think harder on this.
>>>>
>>>>
>>>> Sure, it’s possible that this is the only interesting example and may
>>>> not have enough value to be worthwhile.  But I found it interesting enough
>>>> that it stuck around in the back of my mind for 8 months! :)
>>>>
>>>
>>> Hmm, it had not occurred to me: instantiating a Set<Hashable> is not
>>> supported (and you can substitute for Hashable any protocol you want).
>>> Thus, for any Set<T> and Set<U> that you can actually instantiate, unless T
>>> and U are both classes and one inherits from the other (in which case the
>>> generic `intersection<X>(a: Set<X>, b: Set<X>) -> Set<X>` already
>>> suffices), Set<T & U> must be the empty set. This is not a very interesting
>>> result.
>>>
>>>
>>> Yes, but this is a limitation due to the fact the existentials for a
>>> protocol do not conform to the protocol.  In some cases the existential
>>> *cannot* conform to the protocol but in many cases (especially common
>>> cases) it *can*.  It just doesn’t today.  There is widespread desire to see
>>> this situation improve.
>>>
>>
>> Sure, but when will be the day that existentials conform to their own
>> protocol when they can do so, *and* we extend `&` to value types (probably
>> not until they can express some sort of meaningful subtyping relationship
>> to each other)?
>>
>>
>> I hope it isn’t *too* long before existentials conform to their own
>> protocol at least in simple cases - Swift 5 if it doesn’t make it into
>> Swift 4.
>>
>
> I will bet you two virtual alcoholic beverages that it won't happen before
> Swift 7 or one that it won't happen before Swift 9.
>
>
>> I am suggesting this proposal be generalized such that it discusses
>> concrete subtype / supertype relationships rather than restricting it’s
>> scope to classes.  If that approach is adopted then `&` would allow value
>> types as soon as this proposal is implemented.
>>
>> It seems arbitrary and unnecessary to restrict it to classes even if that
>> is where it would be most useful when it is first implemented.
>>
>
> One can naturally relax the rules in tandem with the design and
> implementation of prerequisite features that make the more relaxed rules
> useful. The point I'm trying to make is: there is currently no value of X
> for which the following statement holds true--
>
> "If only it weren't for the restriction against writing `Base1 & Base2`,
> I'd be able to implement the interesting algorithm X."
>
> So it seems reasonable to error on `Base1 & Base2`.
>
>
> If you’re right about the timeline for existentials conforming to their
> protocols I would agree with you wholeheartedly.  I’m not sure why we have
> such different perspectives on when that might happen.  It has certainly
> received plenty of demand from the community and caused some degree of
> confusion, usually in relatively simple cases (since we only have
> relatively simple existentials today).
>
>
> Here are a couple of e-mails I sent recently explaining why
> self-conforming existentials are tricky. Someone with a deep understanding
> of the compiler needs to put some thought into how an implementation could
> work…
>
> https://lists.swift.org/pipermail/swift-evolution/
> Week-of-Mon-20160815/026349.html
> https://lists.swift.org/pipermail/swift-users/Week-of-
> Mon-20161226/004292.html
>
>
> Thanks for the links, I hadn’t seen these previously.  For some reason I
> thought the implementation complexity was focused on protocols with Self
> and associated type requirements and so basic support might be easier to
> implement as a first step.
>
>
>
>
>
>> At that point, I'd advocate for using compiler magic to make uninhabited
>> types like Never a subtype of all types conforming to all protocols. Then,
>> we could actually write Set<Never> without having to implement conformance
>> to Hashable by writing a bogus `==` function. And we could replace
>> EmptyCollection with Collection<Never> and simplify the standard library
>> API surface that way (since Array<Never>() would then be a value of type
>> Array<T>, etc.). And, your demonstrated use case would become interesting.
>> Since there is a pretty good chance that you and I won't be alive by then,
>> I'm happy to punt on the ideation process for this :)
>>
>>
>>> It generalizes easily to any cases where you have a generic type that is
>>>> useful despite not necessarily having access to instances of the
>>>> parameterized type.
>>>>
>>>> If we allow this, I *think* all uninhabitable types could be unified
>>>> semantically by making `Never` a protocol and giving them implicit
>>>> conformance.
>>>>
>>>>
>>>> This example points even more strongly in the direction of allowing
>>>>> *any* concrete type to be used, not just classes - even today we could
>>>>> produce uninhabitable existentials like this using value types.
>>>>>
>>>>> Here’s the link to the thread: https://lists.swift.or
>>>>> g/pipermail/swift-evolution/Week-of-Mon-20160523/019463.html
>>>>>
>>>>>
>>>>>
>>>>> On Sun, Jan 29, 2017 at 12:41 Austin Zheng <austinzheng at gmail.com>
>>>>>> wrote:
>>>>>>
>>>>>>> The "class comes first" requirement made more sense when the
>>>>>>> proposed syntax was still "Any<T, U, V>", intentionally mirroring how the
>>>>>>> superclass and conformances are declared on a class declaration (the
>>>>>>> archives contain more detailed arguments, both pro and con). Now that the
>>>>>>> syntax is "T & U & V", I agree that privileging the class requirement is
>>>>>>> counterintuitive and probably unhelpful.
>>>>>>>
>>>>>>> Austin
>>>>>>>
>>>>>>> > On Jan 29, 2017, at 10:37 AM, Matt Whiteside via swift-evolution <
>>>>>>> swift-evolution at swift.org> wrote:
>>>>>>> >
>>>>>>> > Thanks for writing this proposal David.
>>>>>>> >
>>>>>>> >> On Jan 29, 2017, at 10:13, Xiaodi Wu via swift-evolution <
>>>>>>> swift-evolution at swift.org> wrote:
>>>>>>> >>
>>>>>>> >> As Matthew mentioned, the rules can certainly later be relaxed,
>>>>>>> but given that this proposal has the compiler generating fix-its for
>>>>>>> subclasses in second position, is there a reason other than stylistic for
>>>>>>> demanding MyClass & MyProtocol instead of MyProtocol & MyClass?
>>>>>>> >>
>>>>>>> >> From a naive perspective, it seems that if the compiler
>>>>>>> understands my meaning perfectly, it should just accept that spelling
>>>>>>> rather than complain.
>>>>>>> >
>>>>>>> > I had that thought too.  Since ‘and’ is a symmetric operation,
>>>>>>> requiring the class to be in the first position seems counter-intuitive.
>>>>>>> >
>>>>>>> > -Matt
>>>>>>> >
>>>>>>> > _______________________________________________
>>>>>>> > swift-evolution mailing list
>>>>>>> > swift-evolution at swift.org
>>>>>>> > https://lists.swift.org/mailman/listinfo/swift-evolution
>>>>>>>
>>>>>>> _______________________________________________
>>>>>> swift-evolution mailing list
>>>>>> swift-evolution at swift.org
>>>>>> https://lists.swift.org/mailman/listinfo/swift-evolution
>>>>>>
>>>>>>
>>>
>>
>>
>
> _______________________________________________
> swift-evolution mailing list
> swift-evolution at swift.org
> https://lists.swift.org/mailman/listinfo/swift-evolution
>
>
>
>
> _______________________________________________
> 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/20170130/2f7a8ed0/attachment.html>