[swift-evolution] [Pitch] Exhaustive pattern matching forprotocols and classes

Matthew Johnson matthew at anandabits.com
Thu May 26 11:01:25 CDT 2016


> On May 26, 2016, at 10:45 AM, Thorsten Seitz <tseitz42 at icloud.com> wrote:
> 
>> 
>> Am 26.05.2016 um 17:29 schrieb Matthew Johnson <matthew at anandabits.com <mailto:matthew at anandabits.com>>:
>> 
>>> 
>>> On May 26, 2016, at 10:13 AM, Thorsten Seitz <tseitz42 at icloud.com <mailto:tseitz42 at icloud.com>> wrote:
>>> 
>>>> 
>>>> Am 26.05.2016 um 16:59 schrieb Matthew Johnson <matthew at anandabits.com <mailto:matthew at anandabits.com>>:
>>>> 
>>>>> 
>>>>> On May 26, 2016, at 9:52 AM, Thorsten Seitz <tseitz42 at icloud.com <mailto:tseitz42 at icloud.com>> wrote:
>>>>> 
>>>>>> 
>>>>>> Am 26.05.2016 um 15:40 schrieb Matthew Johnson <matthew at anandabits.com <mailto:matthew at anandabits.com>>:
>>>>>> 
>>>>>> 
>>>>>> 
>>>>>> Sent from my iPad
>>>>>> 
>>>>>> On May 26, 2016, at 8:25 AM, Thorsten Seitz <tseitz42 at icloud.com <mailto:tseitz42 at icloud.com>> wrote:
>>>>>> 
>>>>>>> Ceylon requires checks whether cases are disjoint, i.e. when one case contains a superclass of another case then this will be a type error „cases are not disjoint“.
>>>>>>> 
>>>>>>> FWIW: Ceylon requires classes with enumerated subclasses to be abstract. 
>>>>>> 
>>>>>> Interesting, thanks for mentioning this.  The abstract requirement is what makes disjointedness at least partly possible (what if a subclass has further descendants though?).  But it still only works for a single level of inheritance:
>>>>>> 
>>>>>> sealed abstract class A {}
>>>>>> class B : A {}
>>>>>> class C : A {}
>>>>>> class D : B {}
>>>>>> class E : B {}
>>>>>> 
>>>>>> With a disjoint requirement I cannot ever match D and E because that would not be exhaustive and I am prohibited from matching them along side B which would be exhaustive but isn't disjoint.
>>>>> 
>>>>> Why?
>>>>> 
>>>>> switch a {
>>>>> case C: …
>>>>> case D: …
>>>>> case E: …
>>>>> }
>>>>> 
>>>>> is exhaustive because B has to be sealed as well (I would require this to be declared explicitly).
>>>>> Abstractness seems not to be necessary for that IMO.
>>>>> Either you match against B *or* against all its subclasses.
>>>> 
>>>> In my example `B` is not abstract.  `a` might have dynamic type of `B` which would not match any of those cases.  
>>> 
>>> Ah, now I realize why the requirement for abstractness makes sense. Of course. Thanks for pointing that out!
>>> 
>>> 
>>>> You didn’t mention that you have to make all non-leaf classes `abstract` (or at least I didn’t understand that from what you wrote).  
>>> 
>>> The requirement in Ceylon is for classes with enumerated subclasses. B would have to declare D and E as enumerated subclasses (like I did for Child2 in the Ceylon example). In Ceylon this is not required, B could have normal subclasses D and E, but then you could not match against D or E at all because these would not be exhaustive (B might have other subclasses). The problem therefore is not stemming from disjointness requirements.
>>> As soon as D and E are declared as disjoint (i.e. if B would be declared `sealed`) matching against them would be possible but B would have to be abstract.
>> 
>> In my example B is implicitly sealed because it has a sealed ancestor.  If it weren’t you would be able to violate the sealed property of the ancestor by inheriting from B.
>> 
>>> 
>>> If I remove the enumerated subclasses from Child2 in the example I am allowed to make Child2 non-abstract. This of course results in the type error that the cases Child1 | Grandchild1 | Grandchild2 are not covering Parent anymore and I have to change the cases to match Child1 | Child2, removing the grandchildren matches because they are not disjoint with Child2.
>> 
>> Got it.  The enumerated subclasses are “partial sealing”.  They seal one level of one branch of the inheritance hierarchy.  Interesting.  You can get exhaustive matching without needing to seal the entire hierarchy.  I’m not sure if I like this or not.  It is something to consider though, and it is more flexible.
>> 
>>> 
>>> 
>>>> I don’t like a design that requires non-leaf classes to be abstract.  First, it would require introducing abstract classes into Swift, which is a totally separate conversation and something I am not convinced is a good idea.  
>>> 
>>> Well, I think abstract classes are something that is sorely missing in Swift! Very sorely missing…
>> 
>> I’ll keep my mind open about it.  My concern is that people will gravitate to the familiar where protocols would be a better design.  I haven’t yet missed abstract classes in Swift.
>> 
>>> But as you say that is a separate conversation and an easy fix would be to restrict `sealed` to protocols first.
>> 
>> Why do that?  We can have exhaustive switches without requiring abstract.
> 
> That’s right. So, exhaustiveness in case of non-abstract B would be given by matching either for B or for exactly B and all its subclasses, i.e.

Yep

> 
> Either
> 
> switch a {
> case is C: ...
> case is B: ...
> }
> 
> or
> 
> switch a {
> case is C: …
> case isexactly B: …
> case is D: …
> case is E: …
> }
> 
> In case of abstract B (in case this would be added to Swift one day) we would have:
> 
> switch a {
> case is C: …
> case is D: …
> case is E: …
> }
> 
> 
>> 
>>> 
>>> 
>>>> But it is also overly restrictive.  There are valid cases where you might want an exhaustive switch for a sealed hierarchy that has concrete parent classes.
>>> 
>>> In that case your suggestion of `isExactly` (or something shorter :-) would indeed be the solution.
>> 
>> Yep.  Any ideas for something more concise?
> 
> Maybe `iss` for „is strictly“? Probably too cryptic, though.

Yeah, too cryptic.  I’m continuing to try and think of something better.  But Swift favors clarity over conciseness.  I’m not sure we’ll find something more concise that is also at least as clear.  Let me know if you think of anything else!

> 
> -Thorsten
> 
> 
>> 
>>> 
>>> -Thorsten
>>> 
>>> 
>>> 
>>>> 
>>>> If you want all non-leaf types to be abstract you should probably consider using protocols in Swift.  
>>>> 
>>>>> 
>>>>> Example in Ceylon:
>>>>> abstract class Parent() of Child1 | Child2 {}
>>>>> 
>>>>> class Child1() extends Parent() {}
>>>>> 
>>>>> abstract class Child2() of Grandchild1 | Grandchild2 extends Parent() {}
>>>>> 
>>>>> class Grandchild1() extends Child2() {}
>>>>> 
>>>>> class Grandchild2() extends Child2() {}
>>>>> 
>>>>> void main() {
>>>>>     Parent foo = Child1();
>>>>> 
>>>>>     switch (foo)
>>>>>     case (is Child1) {
>>>>>         print("Child1");
>>>>>     }
>>>>>     case (is Grandchild1) {
>>>>>         print("Grandchild1");
>>>>>     }
>>>>>     case (is Grandchild2) {
>>>>>         print("Grandchild2");
>>>>>     }
>>>>> }
>>>>> 
>>>>> -Thorsten
>>>>> 
>>>>>  
>>>>>> 
>>>>>> I don't think that solution is appropriate to Swift.  
>>>>>> 
>>>>>>> 
>>>>>>> -Thorsten
>>>>>>> 
>>>>>>>  
>>>>>>>> Am 25.05.2016 um 19:49 schrieb Matthew Johnson via swift-evolution <swift-evolution at swift.org <mailto:swift-evolution at swift.org>>:
>>>>>>>> 
>>>>>>>> 
>>>>>>>> 
>>>>>>>> Sent from my iPad
>>>>>>>> 
>>>>>>>> On May 25, 2016, at 12:41 PM, Charlie Monroe <charlie at charliemonroe.net <mailto:charlie at charliemonroe.net>> wrote:
>>>>>>>> 
>>>>>>>>>> Got it.  You could also say it is safer because you can't have a supertype case "swallow" a subtype value accidentally.  An "exact type" cast would prevent this possibility.
>>>>>>>>> 
>>>>>>>>> This still can be an issue since you still need to do the switch in init(instance:), but it's just one place within the entire module, so it can be more easily managed...
>>>>>>>> 
>>>>>>>> Yes, agree.  That's why your enum is safer.  I think we do need an exact type cast to prevent this problem.  'isExaclty' and 'asExactly' seem are a bit verbose but are very clear.  I can't think of anything I like that is more concise.
>>>>>>>> 
>>>>>>>>> 
>>>>>>>>>> 
>>>>>>>>>>> 
>>>>>>>>>>> enum AnimalSubclasses {
>>>>>>>>>>> 	
>>>>>>>>>>> 	case Dog
>>>>>>>>>>> 	case Cat
>>>>>>>>>>> 
>>>>>>>>>>> 	init(instance: Animal) {
>>>>>>>>>>> 		switch instance {
>>>>>>>>>>> 		case is Dog: self = .Dog
>>>>>>>>>>> 		case is Cat: self = .Cat
>>>>>>>>>>> 		default: fatalError("Unhandled instance \(instance)!")
>>>>>>>>>>> 	}
>>>>>>>>>>> 
>>>>>>>>>>> }
>>>>>>>>>>> 
>>>>>>>>>>>> One thing I have considered that might also be worth introducing is an exact match cast.  This would prevent the possibility of putting a superclass case first and having it “steal” subclasses which were intended to be covered by a case later in the switch.  If we introduce exact match you would be able to write a switch that must always cover every concrete type, including all subclasses.  
>>>>>>>>>>>> 
>>>>>>>>>>>>> 
>>>>>>>>>>>>> Charlie
>>>>>>>>>>>>> 
>>>>>>>>>>>>> 
>>>>>>>>>>>>>> On May 25, 2016, at 4:41 AM, Leonardo Pessoa via swift-evolution <swift-evolution at swift.org <mailto:swift-evolution at swift.org>> wrote:
>>>>>>>>>>>>>> 
>>>>>>>>>>>>>> Limiting the amount of subclasses is not really a good idea as you would need to introduce another mechanism in the language while the proposed feature requires much less. And you're thinking only about the restrictive set (internal and private) and forgetting the more open end (public). Why is it so bad for this proposal to support requiring the default case? If its possible for the compiler to discover you covered all possible cases it would be fine not having default but IMHO in most cases it will find out there are more not explicitly covered.
>>>>>>>>>>>>>> From: David Sweeris <mailto:davesweeris at mac.com>
>>>>>>>>>>>>>> Sent: ‎24/‎05/‎2016 11:01 PM
>>>>>>>>>>>>>> To: Austin Zheng <mailto:austinzheng at gmail.com>
>>>>>>>>>>>>>> Cc: Leonardo Pessoa <mailto:me at lmpessoa.com>; swift-evolution <mailto:swift-evolution at swift.org>
>>>>>>>>>>>>>> Subject: Re: [swift-evolution] [Pitch] Exhaustive pattern matching forprotocols and classes
>>>>>>>>>>>>>> 
>>>>>>>>>>>>>> Or if there was a way to declare that a class/protocol can only have a defined set of subclasses/conforming types.
>>>>>>>>>>>>>> 
>>>>>>>>>>>>>> Sent from my iPhone
>>>>>>>>>>>>>> 
>>>>>>>>>>>>>> On May 24, 2016, at 15:35, Austin Zheng via swift-evolution <swift-evolution at swift.org <mailto:swift-evolution at swift.org>> wrote:
>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> If you pattern match on a type that is declared internal or private, it is impossible for the compiler to not have an exhaustive list of subclasses that it can check against.
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> Austin
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> On Tue, May 24, 2016 at 1:29 PM, Leonardo Pessoa <me at lmpessoa.com <mailto:me at lmpessoa.com>> wrote:
>>>>>>>>>>>>>>> I like this but I think it would be a lot hard to ensure you have all
>>>>>>>>>>>>>>> subclasses covered. Think of frameworks that could provide many
>>>>>>>>>>>>>>> unsealed classes. You could also have an object that would have to
>>>>>>>>>>>>>>> handle a large subtree (NSObject?) and the order in which the cases
>>>>>>>>>>>>>>> are evaluated would matter just as in exception handling in languages
>>>>>>>>>>>>>>> such as Java (or require some evaluation from the compiler to raise
>>>>>>>>>>>>>>> warnings). I'm +1 for this but these should be open-ended like strings
>>>>>>>>>>>>>>> and require the default case.
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> On 24 May 2016 at 17:08, Austin Zheng via swift-evolution
>>>>>>>>>>>>>>> <swift-evolution at swift.org <mailto:swift-evolution at swift.org>> wrote:
>>>>>>>>>>>>>>> > I have been hoping for the exhaustive pattern matching feature for a while
>>>>>>>>>>>>>>> > now, and would love to see a proposal.
>>>>>>>>>>>>>>> >
>>>>>>>>>>>>>>> > Austin
>>>>>>>>>>>>>>> >
>>>>>>>>>>>>>>> > On Tue, May 24, 2016 at 1:01 PM, Matthew Johnson via swift-evolution
>>>>>>>>>>>>>>> > <swift-evolution at swift.org <mailto:swift-evolution at swift.org>> wrote:
>>>>>>>>>>>>>>> >>
>>>>>>>>>>>>>>> >> Swift currently requires a default pattern matching clause when you switch
>>>>>>>>>>>>>>> >> on an existential or a non-final class even if the protocol or class is
>>>>>>>>>>>>>>> >> non-public and all cases are covered.  It would be really nice if the
>>>>>>>>>>>>>>> >> default clause were not necessary in this case.  The compiler has the
>>>>>>>>>>>>>>> >> necessary information to prove exhaustiveness.
>>>>>>>>>>>>>>> >>
>>>>>>>>>>>>>>> >> Related to this is the idea of introducing something like a `sealed`
>>>>>>>>>>>>>>> >> modifier that could be applied to public protocols and classes.  The
>>>>>>>>>>>>>>> >> protocol or class would be visible when the module is imported, but
>>>>>>>>>>>>>>> >> conformances or subclasses outside the declaring module would be prohibited.
>>>>>>>>>>>>>>> >> Internal and private protocols and classes would implicitly be sealed since
>>>>>>>>>>>>>>> >> they are not visible outside the module.  Any protocols that inherit from a
>>>>>>>>>>>>>>> >> sealed protocol or classes that inherit from a sealed class would also be
>>>>>>>>>>>>>>> >> implicitly sealed (if we didn’t do this the sealing of the superprotocol /
>>>>>>>>>>>>>>> >> superclass could be violated by conforming to or inheriting from a
>>>>>>>>>>>>>>> >> subprotocol / subclass).
>>>>>>>>>>>>>>> >>
>>>>>>>>>>>>>>> >> Here are examples that I would like to see be valid:
>>>>>>>>>>>>>>> >>
>>>>>>>>>>>>>>> >> protocol P {}
>>>>>>>>>>>>>>> >> // alternatively public sealed protocol P {}
>>>>>>>>>>>>>>> >> struct P1: P {}
>>>>>>>>>>>>>>> >> struct P2: P {}
>>>>>>>>>>>>>>> >>
>>>>>>>>>>>>>>> >> func p(p: P) -> Int {
>>>>>>>>>>>>>>> >>     switch p {
>>>>>>>>>>>>>>> >>     case is P1: return 1 // alternatively an `as` cast
>>>>>>>>>>>>>>> >>     case is P2: return 2 // alternatively an `as` cast
>>>>>>>>>>>>>>> >>     }
>>>>>>>>>>>>>>> >> }
>>>>>>>>>>>>>>> >>
>>>>>>>>>>>>>>> >> class C {}
>>>>>>>>>>>>>>> >> // alternatively public sealed class C {}
>>>>>>>>>>>>>>> >> class C1: C {}
>>>>>>>>>>>>>>> >> class C2: C {}
>>>>>>>>>>>>>>> >>
>>>>>>>>>>>>>>> >> func c(c: C) -> Int {
>>>>>>>>>>>>>>> >>     switch c {
>>>>>>>>>>>>>>> >>     case is C1: return 1 // alternatively an `as` cast
>>>>>>>>>>>>>>> >>     case is C2: return 2 // alternatively an `as` cast
>>>>>>>>>>>>>>> >>     case is C: return 0   // alternatively an `as` cast
>>>>>>>>>>>>>>> >>     }
>>>>>>>>>>>>>>> >> }
>>>>>>>>>>>>>>> >>
>>>>>>>>>>>>>>> >> I am wondering if this is something the community is interested in.  If
>>>>>>>>>>>>>>> >> so, I am wondering if this is something that might be possible in the Swift
>>>>>>>>>>>>>>> >> 3 timeframe (maybe just for private and internal protocols and classes) or
>>>>>>>>>>>>>>> >> if it should wait for Swift 4 (this is likely the case).
>>>>>>>>>>>>>>> >>
>>>>>>>>>>>>>>> >> -Matthew
>>>>>>>>>>>>>>> >> _______________________________________________
>>>>>>>>>>>>>>> >> swift-evolution mailing list
>>>>>>>>>>>>>>> >> swift-evolution at swift.org <mailto:swift-evolution at swift.org>
>>>>>>>>>>>>>>> >> https://lists.swift.org/mailman/listinfo/swift-evolution <https://lists.swift.org/mailman/listinfo/swift-evolution>
>>>>>>>>>>>>>>> >
>>>>>>>>>>>>>>> >
>>>>>>>>>>>>>>> >
>>>>>>>>>>>>>>> > _______________________________________________
>>>>>>>>>>>>>>> > swift-evolution mailing list
>>>>>>>>>>>>>>> > swift-evolution at swift.org <mailto:swift-evolution at swift.org>
>>>>>>>>>>>>>>> > https://lists.swift.org/mailman/listinfo/swift-evolution <https://lists.swift.org/mailman/listinfo/swift-evolution>
>>>>>>>>>>>>>>> >
>>>>>>>>>>>>>>> 
>>>>>>>>>>>>>>> _______________________________________________
>>>>>>>>>>>>>>> swift-evolution mailing list
>>>>>>>>>>>>>>> swift-evolution at swift.org <mailto:swift-evolution at swift.org>
>>>>>>>>>>>>>>> https://lists.swift.org/mailman/listinfo/swift-evolution <https://lists.swift.org/mailman/listinfo/swift-evolution>
>>>>>>>>>>>>>> _______________________________________________
>>>>>>>>>>>>>> swift-evolution mailing list
>>>>>>>>>>>>>> swift-evolution at swift.org <mailto:swift-evolution at swift.org>
>>>>>>>>>>>>>> https://lists.swift.org/mailman/listinfo/swift-evolution <https://lists.swift.org/mailman/listinfo/swift-evolution>
>>>>>>>>>>>>> 
>>>>>>>>>>>>> _______________________________________________
>>>>>>>>>>>>> swift-evolution mailing list
>>>>>>>>>>>>> swift-evolution at swift.org <mailto:swift-evolution at swift.org>
>>>>>>>>>>>>> https://lists.swift.org/mailman/listinfo/swift-evolution <https://lists.swift.org/mailman/listinfo/swift-evolution>
>>>>>>>>> 
>>>>>>>> _______________________________________________
>>>>>>>> swift-evolution mailing list
>>>>>>>> swift-evolution at swift.org <mailto:swift-evolution at swift.org>
>>>>>>>> https://lists.swift.org/mailman/listinfo/swift-evolution <https://lists.swift.org/mailman/listinfo/swift-evolution>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-evolution/attachments/20160526/771c2269/attachment.html>


More information about the swift-evolution mailing list