[swift-evolution] Subclass Existentials

Matthew Johnson matthew at anandabits.com
Sun Jan 29 16:19:56 CST 2017 

> On Jan 29, 2017, at 4:01 PM, David Hart <david at hartbit.com> wrote:
> 
> Hi Matthew,
> 
> I’ll reply to this post, because it allows me to discuss a few of the points in the discussion, but I’ve read the whole discussion.
> 
>> On 29 Jan 2017, at 18:30, Matthew Johnson <matthew at anandabits.com <mailto:matthew at anandabits.com>> wrote:
>> 
>> Hi David,
>> 
>> This looks like a great start.
>> 
>> One thing we should consider is whether we tie this proposal so tightly to classes or whether it might be better to call these supertype constraints.  The feature might also be useful for value types if / when Swift gets value subtyping.
> 
> This makes sense, especially with the Substring : String discussions going on. When I rework the proposal, I’ll try to make it more general.
> 
>> One enhancement that might be worth considering.  Specifically, allowing protocols to declare a specific supertype requirement in the place where a `class` constraint would usually be found.  After this proposal we could already do something similar using a `class` constraint to define a protocol and a typealias to bind it to the supertype requirement.  It seems like allowing users to state this more directly would be a good idea.
> 
> You lost me there. Can you give examples?

Sure.  In an iOS app you might want to create a protocol that should only be conformed to by `UIViewController` subclasses:

protocol TrackedViewController: UIViewController {
    var analyticsTitle: String { get }
}

This is roughly equivalent to under your proposal:

protocol TrackedObject: class {
    var analyticsTitle: String { get }
}
typealias TrackedViewController = UIViewController & TrackedObject

But in this case we would be able to avoid the TrackedObject protocol which we didn’t care about or want, but were required to create in the process of creating our typealias.


> 
>> As only the first element in the existential composition syntax can be a class type, and by extending this rule to typealias expansions, we can make sure that we only need to read the first element to know if it contains a class requirement.
>> 
>> I think this is unnecessarily limiting in a couple of ways.  I agree that a class should come first if it is mentioned explicitly***.  I am less sure we should require this when the type is part of a typealias combined with other protocol requirements.
> 
> I agree with Chris that I think it’s important to require the class be mentioned first when the class is mentioned explicitly. Even if we lost the Any<Base, Protocol> syntax, there is still enough similarity to a class’s inheritance/conformance clause to keep the consistency and readability.
> 
>> For example, one use case I remember discussing with Austin is refining supertype requirements.  If I have a typealias which requires a superclass `Base` I should be able to form an existential using that typealias that *refines* that requirement to some type *Derived* which is a non-final subtype of `Base`.  This would require syntax that allows us to put a class name in the first position, but also mention a typealias with a supertype requirement in a subsequent position.
> 
> I’ve read the examples in the thread and I think I agree that those cases should be accepted. But just to make sure we are on the same page, what does everyone think of the validity of the following cases? For shorthand, I use parentheses to represent typealias expansion. For example, when I write:
> 
> Protocol1 & (Protocol2 & Protocol3)
> I mean:
> 
> typealias Something = Protocol2 & Protocol3
> Protocol1 & Something
> Questions
> 
> Should class requirements be fixed to first position? I.e., should Protocol & Base be valid and equivalent to Base & Protocol?
> Should repetition of class requirements in the same declaration be allowed? I.e., should Base & Base be valid and equivalent to Base?
> Should repetition of class requirements through typealias expansion be allowed? I.e., should Base & (Base & Protocol) be valid and equivalent to Base & Protocol?
> Should type and sub-type requirements in the same declaration be allowed? I.e., should Base & Derived or Derived & Base be valid and equivalent to Derived?
> Should type and sub-type requirements through typealias expansion be allowed? I.e., should Base & (Derived & Protocol) or Derived & (Base & Protocol) be valid and equivalent to Derived & Protocol?
> My Answers
> 
> No, for the reasons stated above.
> No, because it doesn’t make sense to repeat it in the same declaration.
> Yes, I’m gonna start agreeing with you and think will ease typealias composition.
> No, for the same reasons as 2.
> Yes, for the same reasons as 3.
My answer depends on whether we adopt the perspective of intersection types and allow things like `Base1 & Base2` or not.  If we’re not going to do that (at least not yet) I think your answers are the right ones.  

However, if we decide to allow uninhabiatable existentials like`Base1 & Base2` I think it would be best to just relax the rules such that they follow the logic of intersection types.

The good news is that (as far as I can tell) it would be an additive change to start with your rules and later adopt the intersection types approach.


> David.
> 
>> Matthew
>> 
>> *** One argument against requiring a class to come first is that we could conceptualize `&` as a type operator with a handful of overloads.  This would include both lhs and rhs are “protocol only kinds” as well as overloads with a “protocol only kind” in either lhs or rhs and a “supertype kind” in the other position.  The tricky part of pulling this off would be including an overload where both lhs and rhs have a “supertype kind”, but only when the operands have a subtype / supertype relationship with each other.  
>> 
>> I suspect this conceptualization isn’t worth the complexity it brings, but it is tempting to try and view `&` as a type operator.  As long as this only involves a loosening of restrictions it could probably be introduced as an additive change down the road.
>> 
>>> On Jan 29, 2017, at 10:39 AM, David Hart <david at hartbit.com <mailto:david at hartbit.com>> wrote:
>>> 
>>> Hello,
>>> 
>>> As promised, I wrote the first draft of a proposal to add class requirements to the existential syntax. Please let me know what you think.
>>> 
>>> https://github.com/hartbit/swift-evolution/blob/subclass-existentials/proposals/XXXX-subclass-existentials.md <https://github.com/hartbit/swift-evolution/blob/subclass-existentials/proposals/XXXX-subclass-existentials.md>
>>> 
>>> Regards,
>>> David.
>>> 
>>> Existentials for classes conforming to protocols
>>> 
>>> Proposal: SE-XXXX <https://github.com/hartbit/swift-evolution/blob/subclass-existentials/proposals/XXXX-subclass-existentials.md>
>>> Authors: David Hart <http://github.com/hartbit/>, Austin Zheng <http://github.com/austinzheng>
>>> Review Manager: TBD
>>> Status: TBD
>>>  <https://github.com/hartbit/swift-evolution/tree/subclass-existentials/proposals#introduction>Introduction
>>> 
>>> This proposal brings more expressive power to the type system by allowing Swift to represent existentials of classes and subclasses which conform to protocols.
>>> 
>>>  <https://github.com/hartbit/swift-evolution/tree/subclass-existentials/proposals#motivation>Motivation
>>> 
>>> Currently, the only existentials which can be represented in Swift are conformances to a set of protocols, using the &syntax:
>>> 
>>> let existential: Hashable & CustomStringConvertible
>>> On the other hand, Objective-C is capable of expressing existentials of subclasses conforming to protocols with the following syntax:
>>> 
>>> UIViewController<UITableViewDataSource, UITableViewDelegate>* existential;
>>> We propose to provide similar expressive power to Swift, which will also improve the bridging of those types from Objective-C.
>>> 
>>>  <https://github.com/hartbit/swift-evolution/tree/subclass-existentials/proposals#proposed-solution>Proposed solution
>>> 
>>> The proposal keeps the existing & syntax but allows the first element, and only the first, to be of class type. The equivalent declaration to the above Objective-C declaration would look like this:
>>> 
>>> let existential: UIViewController & UITableViewDataSource & UITableViewDelegate
>>> As in Objective-C, this existential represents classes which have UIViewController in their parent inheritance hierarchy and which also conform to the UITableViewDataSource and UITableViewDelegate protocols.
>>> 
>>> As only the first element in the existential composition syntax can be a class type, and by extending this rule to typealias expansions, we can make sure that we only need to read the first element to know if it contains a class requirement. As a consequence, here is a list of valid and invalid code and the reasons for them:
>>> 
>>> let a: Hashable & CustomStringConvertible
>>> // VALID: This is still valid, as before
>>> 
>>> let b: MyObject & Hashable
>>> // VALID: This is the new rule which allows an object type in first position
>>> 
>>> let c: CustomStringConvertible & MyObject
>>> // INVALID: MyObject is not allowed in second position. A fix-it should help transform it to:
>>> // let c: MyObject & CustomStringConvertible
>>> 
>>> typealias MyObjectStringConvertible = MyObject & CustomStringConvertible
>>> let d: Hashable & MyObjectStringConvertible
>>> // INVALID: The typealias expansion means that the type of d expands to Hashable & MyObject & CustomStringConvertible, which has the class in the wrong position. A fix-it should help transform it to:
>>> // let d: MyObjectStringConvertible & Hashable
>>> 
>>> typealias MyObjectStringConvertible = MyObject & CustomStringConvertible
>>> let e: MyOtherObject & MyObjectStringConvertible
>>> // INVALID: The typealias expansion would allow an existential with two class requirements, which is invalid
>>> The following examples could technically be legal, but we believe we should keep them invalid to keep the rules simple:
>>> 
>>> let a: MyObject & MyObject & CustomStringConvertible
>>> // This is equivalent to MyObject & CustomStringConvertible
>>> 
>>> let b: MyObjectSubclass & MyObject & Hashable
>>> // This is equivalent to MyObjectSubclass & Hashable
>>> 
>>> typealias MyObjectStringConvertible = MyObject & CustomStringConvertible
>>> let d: MyObject & MyObjectStringConvertible
>>> // This is equivalent to MyObject & CustomStringConvertible
>>>  <https://github.com/hartbit/swift-evolution/tree/subclass-existentials/proposals#source-compatibility>Source compatibility
>>> 
>>> This is a source breaking change. All types bridged from Objective-C which use the equivalent Objective-C feature import without the protocol conformances in Swift 3. This change would increase the existential's requirement and break on code which does not meet the new protocol requirements. For example, the following Objective-C code:
>>> 
>>> @interface MyViewController
>>> - (void)setup:(nonnull UIViewController<UITableViewDataSource,UITableViewDelegate>*)tableViewController;
>>> @end
>>> is imported into Swift 3 as:
>>> 
>>> class MyViewController {
>>>     func setup(tableViewController: UIViewController) {}
>>> }
>>> which allows calling the function with an invalid parameter:
>>> 
>>> let myViewController: MyViewController()
>>> myViewController.setup(UIViewController())
>>> The previous code would have worked as long as the Objective-C code did not call any method of UITableViewDataSource or UITableViewDelegate. But if this proposal is accepted and implemented as-is, the Objective-C code would now be imported as:
>>> 
>>> class MyViewController {
>>>     func setup(tableViewController: UIViewController & UITableViewDataSource & UITableViewDelegate) {}
>>> }
>>> That would then cause the Swift code to fail to compile with an error which states that UIViewController does not conform to the UITableViewDataSource and UITableViewDelegate protocols.
>>> 
>>> It is a source-breaking change, but should have a minimal impact for the following reasons:
>>> 
>>> Not many Objective-C code used the existential syntax in practice.
>>> There generated errors are a good thing because they point out potential crashes which would have gone un-noticed.
>>>  <https://github.com/hartbit/swift-evolution/tree/subclass-existentials/proposals#alternatives-considered>Alternatives considered
>>> 
>>> None.
>>> 
>>>  <https://github.com/hartbit/swift-evolution/tree/subclass-existentials/proposals#acknowledgements>Acknowledgements
>>> 
>>> Thanks to Austin Zheng <http://github.com/austinzheng> and Matthew Johnson <https://github.com/anandabits> who brought a lot of attention to existentials in this mailing-list and from whom most of the ideas in the proposal come from.
>> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-evolution/attachments/20170129/f6508ad0/attachment.html>

More information about the swift-evolution mailing list