[swift-evolution] Mark protocol methods with their protocol

Xiaodi Wu xiaodi.wu at gmail.com
Tue Sep 20 18:25:01 CDT 2016


This was the conversation we were having in the other thread. Perhaps I'm
still not understanding something, but I'm not convinced that this feature
is an improvement.

Currently, protocols represent a declaration that you have discovered that
your type has certain semantics and guarantees a certain API, and you
declare that fulfillment "by name only," just as you invoke members "by
name only"; unintentional failure to fulfill the contract is a compile-time
error. The status quo is precisely my idea of what protocols should be.
What are the two ideas you have about them?
On Tue, Sep 20, 2016 at 18:02 Karl <razielim at gmail.com> wrote:

> I’m using String as an example of where this issue of conformance
> conflicts crops up in the standard library. Ideally, String (or any data
> type) should be able to conform to protocols whose requirements have
> conflicting names.
> Currently, in order to work around this limitation, you have to delegate
> the conformance to a supporting type. This is more complicated to write and
> maintain, and pollutes your internal API. String gives us an example of
> this, but it’s not the worst example.
> It basically implements what I’m talking about anyway, but manually via a
> supporting type instead of directly inside the data-type (String). As an
> ABI consideration, we should scope all protocol members to their protocols.
> This would resolve all naming conflicts.
>
> I can’t understand how anybody would argue for the status quo - we’re
> currently in-between two ideas of what protocols should be - they are
> explicit but their conformances are resolved by name only and can overlap
> without warning.
>
> On 21 Sep 2016, at 00:48, Xiaodi Wu <xiaodi.wu at gmail.com> wrote:
>
> Sorry, I'm still not sure I understand what you're getting at about this.
> How would String conforming to Collection multiple times simplify or
> improve the implementation of String? Or are you arguing it would be better
> for users of String? If so, how?
>
> On Tue, Sep 20, 2016 at 17:26 Karl <razielim at gmail.com> wrote:
>
>> I’m not saying vital support is missing, just that it is more awkward.
>> String doesn’t conform to collection, String.UTF8View does; so if you
>> change some implementation detail (in StringCore, because that’s where they
>> live for String), you get the error popping up somewhere other than the
>> place you just changed. That’s what I mean when I say “language support”.
>> If you do what StringCore does, and you’re changing stuff which is
>> ultimately going to be used to conform to, say, Collection, you have to
>> build distance between the implementation and the (only) conformance it is
>> used for, and it’s less optimal.
>>
>> Let’s say I have an object MyComplexDataType, it implements
>> “InternalStructureView” and “Collection”:
>>
>> ```
>> protocol InternalStructureView {
>>     associatedtype Index
>>     var count : Int { get } // whatever, just some stuff that will cause
>> a name conflict
>>     func doInternalMagic(at: Index)
>> }
>>
>> struct MyComplexDataType {
>>     var __collection_count : Int {
>>         // This is quite a complex operation which we’d rather leave
>> inside the type, otherwise we’d need to expose a bunch of implementation
>> details internally
>>     }
>>    var __internal_count : Int {
>>        // Again, best left here
>>    }
>>
>>    struct CollectionView  : Collection {
>>        init(parent: MyComplexDataType) { … }
>>         var count { return parent.__collection_count }
>>         // ...etc
>>    }
>>
>>   struct InternalStructure : InternalStructureView {
>>       init(parent: MyComplexDataType) { … }
>>       var count { return parent.__internal_count }
>>       // ...etc
>>   }
>>
>>   var collection : CollectionView { return CollectionView(self) }
>>   var internalStructure : InternalStructure { return
>> InternalStructure(self) }
>> }
>> ```
>>
>> This is basically what String does (except that it wants to conform to
>> Collection multiple times with different indexes and results). It’s a lot
>> of work to maintain, especially if you have evolving protocols that are
>> conformed to in several places.
>> We should have a better solution. We should be able to define:
>>
>> “protocol UTF8Collection : Collection {}
>>  protocol UTF16Collection : Collection {}”
>>
>> and have String conform to both of them at the same time. At the same
>> time, since we’re now being explicit about which protocol requirement is
>> satisfied where - we should also be able to delegate our conformance,
>> telling the compiler to dispatch any unimplemented methods to another
>> object. For example, lets say you want to wrap a Collection and observe
>> mutations to it; you might override replaceSubrange(), but every other
>> method (such as count, index(after:)… all the rest) is just a forwarding
>> function.
>>
>> Interestingly, we could do this safely (from a code legibility
>> perspective) if we say that every scope which adds a conformance must
>> completely satisfy its requirements, which we would more reasonably be able
>> to require if we made this change (so your internal functions can still be
>> wherever you like, but they won’t automatically satisfy a protocol
>> requirement if they happen to have the same name). Then we could reasonably
>> say that if you add an extension which adds a conformance to, say,
>> Collection, you have to tell us where to find every one of its
>> requirements. That’s where we could put the forwarding syntax for
>> retroactive modelling. Stored properties can’t be defined in extensions, so
>> if you want to back an implementation with one, you’ll need to make its
>> conformance explicit in the main body (or we loosen that to at least
>> extensions in the same file).
>>
>> ```
>> // generates thunks to members of this extension in the base type;
>> // so MyComplexDataType.count —>
>> MyComplexDataType.InternalStructure.count,
>> // to account for conformance being added in later version. Can also be
>> used for renamed protocols, and be tagged on individual members.
>>
>> @makeAvailable(as: _)
>> extension MyComplexDataType : InternalStructure {
>>
>>    typealias Index = InternalIndexType
>>     var count : Int {
>>         // We have access to all of the private members because we’re
>> inside MyComplexDataType
>>         // No need to pollute internal API with conformance
>> implementation details.
>>         // Also, we get errors about non-conformance where we want them —
>> where the implementation is.
>>     }
>>     func doInternalMagic(at: Index) {
>>        ...
>>     }
>> }
>> ```
>>
>> On 20 Sep 2016, at 23:46, Xiaodi Wu <xiaodi.wu at gmail.com> wrote:
>>
>> I'm not sure I understand. What compiler or language support is missing
>> for StringCore?
>> On Tue, Sep 20, 2016 at 16:42 Karl via swift-evolution <
>> swift-evolution at swift.org> wrote:
>>
>>> On 20 Sep 2016, at 23:28, Karl <raziel.im+swift-evo at gmail.com> wrote:
>>>
>>>
>>> On 20 Sep 2016, at 18:43, Nevin Brackett-Rozinsky via swift-evolution <
>>> swift-evolution at swift.org> wrote:
>>>
>>> I have been following this discussion (as well as similar threads
>>> earlier this year) and listening to the ideas put forth by all sides.
>>>
>>> It seems to me that the fundamental difference between classes and
>>> protocols is that classes inherit implementation whereas protocol
>>> conformance is a promise about interface.
>>>
>>> When a class or struct or enum declares itself as conforming to a
>>> protocol, that means it has all the members specified in the protocol. The
>>> protocol conformance simply codifies a fact about the type itself: namely
>>> that all those members are present.
>>>
>>> In this model, any keyword such as `implements` on each conforming
>>> member would introduce substantial boilerplate for negligible gain. The
>>> purpose of a protocol is to communicate that certain members are available,
>>> not to make declaring those members more onerous.
>>>
>>> However, default implementations for protocols blur the line. Now there
>>> is actual implementation being inherited. A conforming type may choose to
>>> roll its own version of a method, or to utilize the default provided by the
>>> protocol. This is closer to the situation with subclassing.
>>>
>>> Moreover, a protocol which conforms to another protocol may itself
>>> define (or redefine!) default implementations for members of that other
>>> protocol. This can create “inheritance chains” of protocol default
>>> implementations. I think there is value in being able to refer to (and
>>> call) the inherited default implementation through some sort of `super`
>>> functionality.
>>>
>>> On the other hand, the existence of a default implementation in a
>>> protocol is in large part merely a convenience: a courtesy so that each
>>> conforming type need not rewrite the same boilerplate code.
>>>
>>> A type which conforms to a protocol may accept the default or it may
>>> provide its own implementation, but it is not “overriding” anything. The
>>> default implementation was offered as a convenience, to be taken or left as
>>> needed. Thus I do not think any keyword (neither `override` nor
>>> `implements`) should be required in that case either.
>>>
>>> The frequently-raised point regarding near-miss member names deserves
>>> some attention. Several people have expressed a desire for the compiler to
>>> assist them in determining whether a given member does or does not meet a
>>> protocol requirement. Specifically, when a type conforms to a protocol with
>>> a default implementation, and the type defines a member with a similar
>>> signature, it is not obvious at glance if that member matches the protocol.
>>>
>>> I think this is a job for linters and IDEs. For example, syntax
>>> highlighting could distinguish members which satisfy a protocol
>>> requirement, thereby providing immediate visual confirmation of success.
>>>
>>> Having followed the lengthy discussion and weighed the numerous ideas
>>> put forth, I come down firmly on the side of no keyword for protocol
>>> conformance.
>>>
>>> A protocol describes an interface and provides a set of customization
>>> points. It may also, as a convenience, offer default implementations. The
>>> protocol simply describes the capabilities of its conforming types, and any
>>> default implementations are there to make things easier for them.
>>>
>>> Conforming types should not be afflicted with extraneous keywords: that
>>> would run contrary to the purpose of having protocols in the first place.
>>>
>>> Nevin
>>>
>>>
>>> On Tue, Sep 20, 2016 at 11:16 AM, Xiaodi Wu via swift-evolution <
>>> swift-evolution at swift.org> wrote:
>>>
>>>> As I mentioned above, I agree that better diagnostics for near-misses
>>>> are necessary, but they are possible without new syntax. There is no win in
>>>> avoiding unintentional behavior because, without a default implementation,
>>>> these issues are caught at compile time already.
>>>>
>>>> On Tue, Sep 20, 2016 at 10:14 Vladimir.S via swift-evolution <
>>>> swift-evolution at swift.org> wrote:
>>>>
>>>>>
>>>>>  > extension P {
>>>>>  > implement func foo() -> [String : String] { return [:] }
>>>>>  > }
>>>>>
>>>>> Yes, it seems like we need `implement` (or `override` as another
>>>>> suggestion) in protocol extension also just for the same reasons - be
>>>>> clear
>>>>> about our intention regarding implementing the requirement, to show
>>>>> that
>>>>> this func *depends* on the previous definition of P protocol and to
>>>>> avoid
>>>>> possible mistakes related to protocol conformance.
>>>>>
>>>>> On 20.09.2016 17:38, Charles Srstka wrote:
>>>>> >> On Sep 20, 2016, at 8:17 AM, Vladimir.S via swift-evolution
>>>>> >> <swift-evolution at swift.org <mailto:swift-evolution at swift.org>>
>>>>> wrote:
>>>>> >>
>>>>> >> On 20.09.2016 3:03, Xiaodi Wu via swift-evolution wrote:
>>>>> >>> I definitely think Vladimir's suggestion is a great starting
>>>>> point, IMO.
>>>>> >>>
>>>>> >>> However, I think it could be improved in one key respect where
>>>>> previous
>>>>> >>> proposals using `override` are superior. Namely, the proposed
>>>>> `implement`
>>>>> >>> keyword adds no additional safety when a type implements a protocol
>>>>> >>> requirement that doesn't have a default implementation. This is
>>>>> because, if
>>>>> >>
>>>>> >> Yes, *at the moment of writing* the type's code there could be no
>>>>> default
>>>>> >> implementation for protocol requirement. But, *at the moment of
>>>>> >> compilation* such default implementation could appear.
>>>>> >>
>>>>> >> Let's discuss such scenario in case we'll take your suggestion:
>>>>> >>
>>>>> >> You got SomeClass.swift file, 3rd party file you don't want to
>>>>> change or
>>>>> >> changes are not allowed. Content:
>>>>> >>
>>>>> >> public protocol SomeProtocol {
>>>>> >> func foo()
>>>>> >> }
>>>>> >>
>>>>> >> public class SomeClass : SomeProtocol {
>>>>> >> func foo() {...} // no default implementation *at the moment of
>>>>> writing*,
>>>>> >> no need in `overload`
>>>>> >> }
>>>>> >>
>>>>> >> Now, you adds SomeClass.swift file to your project and in some
>>>>> *other*
>>>>> >> file you write:
>>>>> >>
>>>>> >> extension SomeProtocol {
>>>>> >> func foo() {...}
>>>>> >> }
>>>>> >>
>>>>> >> As you see, you don't control the SomeClass.swift but you suggest
>>>>> in this
>>>>> >> case SomeClass.foo() should be defined with `override`.
>>>>> >>
>>>>> >> With 'implement' SomeClass.foo() will be marked initially and will
>>>>> save
>>>>> >> us if protocol's requirement PLUS default implementation changed.
>>>>> >
>>>>> > Requiring the ‘implement’ keyword can help us even if no default
>>>>> > implementation is involved. Consider:
>>>>> >
>>>>> > protocol P {
>>>>> > func foo() -> [String : Any]
>>>>> > }
>>>>> >
>>>>> > struct S : P {
>>>>> > func foo() -> [String : String] { return [:] }
>>>>> > }
>>>>> >
>>>>> > We will get an error here that S does not conform to P. However,
>>>>> this is
>>>>> > not the correct error, since S in fact *tries* to conform to P, but
>>>>> it has
>>>>> > a mistake in a method signature. This misleads us as to the true
>>>>> nature of
>>>>> > the problem, and if S has enough members in it that we fail to spot
>>>>> the
>>>>> > existing foo(), we might solve the problem by reimplementing foo(),
>>>>> and
>>>>> > leaving the original foo() as dangling dead code. Having an
>>>>> ‘implement’
>>>>> > keyword on the existing foo() function would change the compiler
>>>>> error to
>>>>> > let us know that we have an existing foo() that is incorrectly
>>>>> declared.
>>>>> >
>>>>> > In addition, ‘implement’ can help us when the declaration in
>>>>> question *is*
>>>>> > the default implementation:
>>>>> >
>>>>> > protocol P {
>>>>> > func foo() -> [String : Any]
>>>>> > }
>>>>> >
>>>>> > extension P {
>>>>> > implement func foo() -> [String : String] { return [:] }
>>>>> > }
>>>>> >
>>>>> > Here we will get an error with the proposed ‘implement’ keyword,
>>>>> because
>>>>> > foo() does not have a signature matching anything in the protocol,
>>>>> whereas
>>>>> > without ‘implement’ we would happily and silently generate a useless
>>>>> > dangling function that would never be used, and then pass the buck
>>>>> to the
>>>>> > concrete type that implements P:
>>>>> >
>>>>> > protocol P {
>>>>> > func foo() -> [String : Any]
>>>>> > }
>>>>> >
>>>>> > extension P {
>>>>> > func foo() -> [String : String] { return [:] } // The error is here:
>>>>> > }
>>>>> >
>>>>> > struct S : P {} // But it gets reported here.
>>>>> >
>>>>> > Charles
>>>>> >
>>>>> _______________________________________________
>>>>> 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
>>>
>>>
>>>
>>> I agree that a new keyword is unwanted. Conforming to protocols is quite
>>> a common thing, so you want it to be easy to remember.
>>>
>>> I think the best way is to prefix the member name with the protocol, e.g:
>>>
>>> protocol MyProto {
>>>     var aVariable : Int
>>>     func aFunction()
>>> }
>>> class MyClass : MyProto {
>>>     var MyProto.aVariable : Int
>>>     func MyProto.aFunction() { … }
>>> }
>>>
>>> This is consistent with how we refer to other members of types (e.g.
>>> “extension MyClass.MyInternalClass”). It will be easy for autocompletion to
>>> provide good suggestions, too.
>>> As I see it, the only problem is what if `MyClass` wants its own
>>> function called `aFunction()`? What if the same name satisfies 2 protocols,
>>> which do you write?
>>>
>>> The way to solve all of the problems in a consistent way is to make the
>>> function actually called “MyProto.aFunction”, and for it to be a separate
>>> function from plain “aFunction()” or from “SomeotherProto.aFunction”.
>>>
>>> I believe it is crucial to protocols that we can do this. Maybe I have
>>> some complex data structure and it has its own API, but I want people to be
>>> able to view it as a Collection. By conforming to Collection, I reserve
>>> lots of keywords and indexing operations which I now can’t use in my own
>>> API. Maybe I’m just providing Collection as a convenience to work with
>>> generic algorithms, but my own API has more efficient semantics for some
>>> operations. We’re relegated to using less-obvious and legible names in
>>> order to avoid conflicts.
>>>
>>> We have a way to work around this, which String uses - create a struct
>>> which references your object and calls internal methods such as
>>> “_collection_count” so you can have separate interfaces. This adds up to
>>> quite a lot of boilerplate and maintenance overhead.
>>>
>>>
>>> Also to add here: you’re basically implementing what I’m proposing
>>> manually if you do this; only you don’t get language/compiler support.
>>> String basically does this - it shares StringCore with UTF8View and
>>> defines some internal functions to support it.
>>>
>>> The String views could then be made in to protocols on String, turning
>>> “UTF8View” in to “UTF8Representable”, and opening up algorithms which can
>>> work on generic sequences of UTF8 bytes. I think that’s pretty cool, and
>>> could open up better integration with other types which are (for example)
>>> UTF8Representable — for example a stream of UTF8 bytes (depending on how
>>> flexible implementation allows us to make the protocol).
>>>
>>>
>>> I don’t agree that Protocol conformances are kind-of incidental, as
>>> others here have written. This isn’t like Objective-C where anything that
>>> has the correctly-named methods conforms. Protocol conformances are
>>> completely explicit, and in fact we have empty protocols (“marker
>>> protocols”) for exactly that purpose. I think it is consistent that we make
>>> every member of a conformance specify which protocol it belongs to, and to
>>> have its name scoped to that protocol.
>>>
>>> Karl
>>>
>>>
>>> CC-ing Dave A, to understand better if this fits with the vision of
>>> protocols
>>>
>>> _______________________________________________
>>> 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/20160920/97264b2d/attachment.html>


More information about the swift-evolution mailing list