[swift-dev] Should non-final classes be allowed to conform to Collection?

[Slava Pestov]

Actually I think I figured out a solution. If the witness has ‘Self’ in a non-covariant position, we use the concrete type for Self. Otherwise, we do it the new way. This makes both examples in https://gist.github.com/slavapestov/f455c6184b402f3ffc656f904f1e5d59 <https://gist.github.com/slavapestov/f455c6184b402f3ffc656f904f1e5d59> work.

Sorry for the noise!

Slava

> On Oct 6, 2017, at 12:46 AM, Slava Pestov <spestov at apple.com> wrote:
> 
> I just tried this and it causes major breakage with associated type inference and the expression checker so its probably more trouble than its worth.
> 
> So back to the three options:
> 
> 1) Do nothing, and give up on fixing https://bugs.swift.org/browse/SR-617 <https://bugs.swift.org/browse/SR-617> for now
> 2) Finish https://github.com/apple/swift/pull/12174 <https://github.com/apple/swift/pull/12174> and make the new semantics only take effect in -swift-version 5
> 3) ??? Magic
> 
> Slava
> 
>> On Oct 6, 2017, at 12:37 AM, Slava Pestov <spestov at apple.com <mailto:spestov at apple.com>> wrote:
>> 
>> Another solution is to change the Collection protocol as follows,
>> 
>> protocol Collection {
>>   associatedtype ConformingType = Self
>>   associatedtype Iterator = IndexingIterator<ConformingType>
>> 
>>   …
>> }
>> 
>> extension Collection where Iterator == IndexingIterator<ConformingType> {
>>   func makeIterator() -> IndexingIterator<ConformingType> { … }
>> }
>> 
>> I believe this will fix the source compatibility issue and also make ‘for x in Derived()’ type check. The downside is that the witness table for a Collection conformance now stores an additional associated type for the static conforming class type. However that’s exactly what you need to store somewhere to make this work for non-final classes.
>> 
>> Slava
>> 
>>> On Oct 6, 2017, at 12:25 AM, Slava Pestov via swift-dev <swift-dev at swift.org <mailto:swift-dev at swift.org>> wrote:
>>> 
>>> Hi all,
>>> 
>>> Consider this code,
>>> 
>>> class Base : Collection {
>>>   var startIndex: Int { return 0 }
>>> 
>>>   var endIndex: Int { return 10 }
>>> 
>>>   func index(after i: Int) -> Int { return i + 1 }
>>> 
>>>   subscript(index: Int) -> Int { return index }
>>> }
>>> 
>>> We infer the associated type ‘Iterator’ as ‘IndexingIterator<Base>’. I can use an instance of Base as a sequence just fine:
>>> 
>>> for x in Base() {} // OK
>>> 
>>> Now if I subclass Base, the associated type is still ‘IndexingIterator<Base>’:
>>> 
>>> class Derived : Base {}
>>> 
>>> However the implementation of makeIterator is defined in a constrained extension by the standard library,
>>> 
>>> extension Collection where Self.Iterator == IndexingIterator<Self> {
>>>   func makeIterator() -> IndexingIterator<Self> { … }
>>> }
>>> 
>>> So I cannot call it on a subclass:
>>> 
>>> for x in Derived() {} // fails
>>> 
>>> The error is bizarre, "'IndexingIterator<Base>' is not convertible to 'IndexingIterator<Derived>’” — I’m not doing a conversion here.
>>> 
>>> If you try to call makeIterator() directly, you get an ambiguity error instead:
>>> 
>>> col.swift:17:5: error: ambiguous reference to member 'makeIterator()'
>>> _ = Derived().makeIterator()
>>>     ^~~~~~~~~
>>> Swift.Collection:6:17: note: found this candidate
>>>     public func makeIterator() -> IndexingIterator<Self>
>>>                 ^
>>> Swift.Sequence:5:17: note: found this candidate
>>>     public func makeIterator() -> Self
>>>                 ^
>>> 
>>> Now I couldn’t come up with an example where the code compiles but crashes at runtime because of a type mismatch, but it’s not outside the realm of possibility.
>>> 
>>> With my PR here the conformance itself no longer type checks: https://github.com/apple/swift/pull/12174 <https://github.com/apple/swift/pull/12174>
>>> 
>>> col.swift:1:7: error: type 'Base' does not conform to protocol 'Collection'
>>> class Base : Collection {
>>>       ^
>>> Swift.Sequence:5:17: note: candidate has non-matching type '<Self> () -> Self' [with Element = Int, Index = Int, IndexDistance = Int, Iterator = IndexingIterator<Base>, SubSequence = Slice<Base>, Indices = DefaultIndices<Base>]
>>>     public func makeIterator() -> Self
>>>                 ^
>>> Swift.Collection:6:17: note: candidate has non-matching type '<Self> () -> IndexingIterator<Self>' [with Element = Int, Index = Int, IndexDistance = Int, Iterator = IndexingIterator<Base>, SubSequence = Slice<Base>, Indices = DefaultIndices<Base>]
>>>     public func makeIterator() -> IndexingIterator<Self>
>>> 
>>> I found one example in our code base where this pattern comes up, and that’s SyntaxCollection in tools/SwiftSyntax/SyntaxCollection.swift. It has no subclasses so making it final works there.
>>> 
>>> This was reported externally as https://bugs.swift.org/browse/SR-1863 <https://bugs.swift.org/browse/SR-1863>. I’m not sure if the user expects it to work or just to produce a reasonable diagnostic instructing them to make the class final.
>>> 
>>> What does everyone think of this?
>>> 
>>> 1) Can anyone suggest a way to make it work, so that ‘for x in Derived()’ type checks and the correct Self type (Base, not Derived) for the substitution?
>>> 
>>> 2) Should we just ban such ’non-covariant’ conformances? There is precedent for this — in Swift 3, we used to allow non-final classes to conform to protocols whose requirements had same-type constraints with the right hand side equal to ‘Self’, and Doug closed this hole in Swift 4. My PR is essentially a more comprehensive fix for this hole.
>>> 
>>> 3) Should we allow the hole to remain in place, admitting non-final classes that model Collection, at the cost of not being able to ever fix https://bugs.swift.org/browse/SR-617 <https://bugs.swift.org/browse/SR-617>?
>>> 
>>> Slava
>>> 
>>> _______________________________________________
>>> swift-dev mailing list
>>> swift-dev at swift.org <mailto:swift-dev at swift.org>
>>> https://lists.swift.org/mailman/listinfo/swift-dev <https://lists.swift.org/mailman/listinfo/swift-dev>
>> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-dev/attachments/20171006/e647f966/attachment.html>