[swift-evolution] [Draft] open and public protocols

Xiaodi Wu xiaodi.wu at gmail.com
Sun Feb 19 17:18:33 CST 2017 

On Sun, Feb 19, 2017 at 5:10 PM, Adrian Zubarev <
adrian.zubarev at devandartist.com> wrote:

> Because someInstance["key", .string("key1"), .integer(1),
> .string(stringKeyInstance), .integer(intIndexInstance),
> .integer(intIndexInstance), …] is simply ugly and should be hidden. Such
> API looks horrible.
>
I agree. This cries out for improvements to enums. Indeed, I have run into
the same issue with this and have a proposal idea saved up regarding
anonymous enum cases and subtyping relationships. It would not have been
in-scope for phase 1, so I did not write to the list about it. I'm short on
time these days, but eventually I'll propose it unless someone else gets to
it first. However, this problem does not justify `open` vs. `public`.


> Is this discussion really going to end with every statement being
> questioned with “why”?
>
Why not?


> Why is 42 the number that rules the universe?
>
Surely, it is for the same reason that May 25 is Towel Day.


> --
> Adrian Zubarev
> Sent with Airmail
>
> Am 20. Februar 2017 um 00:03:44, Xiaodi Wu (xiaodi.wu at gmail.com) schrieb:
>
> On Sun, Feb 19, 2017 at 4:51 PM, Adrian Zubarev <
> adrian.zubarev at devandartist.com> wrote:
>
>> Matthew has pretty much summed up everything with a single question in
>> his last reply. It makes me tired of repeating myself over and over again.
>>
>> Here’s the challenge for you:
>>
>> Implement a subscript where the first parameter is a String (trivial),
>> but the second to n are String’s and/or Int’s. There is no restriction of
>> order for the Int’s and String’s and there could be as many as you’d like,
>> so no overloads are allowed (a hint: variadics). You should be able to use
>> variables in that subscript and literals. You’re not allowed to use enum
>> constructors
>>
> Why not? You want a parameter that takes either String or Int. In another
> language, you might express this as `String | Int`, but Chris Lattner and
> others have said on this list that enums are the intended way of expressing
> this in Swift.
>
>> and the client should not be able to extend the set of types you can pass
>> to that subscript. That said it can look like this: someInstance["key",
>> "key1", 1, stringKeyInstance, intIndexInstance, intIndexInstance, …]
>>
>> Creating hacks like unreachable types with hidden inits is prohibited.
>>
>> Have fun!
>>
>>
>> --
>> Adrian Zubarev
>> Sent with Airmail
>>
>> Am 19. Februar 2017 um 23:25:11, Xiaodi Wu (xiaodi.wu at gmail.com) schrieb:
>>
>> Sorry, I have read through this thread twice and do not understand the
>> point you are making. Can you explain your example once more?
>>
>> Specifically, I do not understand why it is that your code should have
>> problems with third-party types that conform to your protocol. If your
>> protocol has requirements that third-party types cannot fulfill, then
>> naturally those requirements will prevent a third-party from conforming new
>> types to your protocol. OTOH, if your protocol does not have any
>> requirements that third-party types cannot fulfill, then your code that
>> accepts anything that conforms to your protocol should be indifferent to
>> whether the conforming type is defined by you or someone else. After all, a
>> conforming type would fulfill all the semantic and syntactic requirements
>> of the protocol! If your protocol has requirements that you are unable to
>> state in code, causing third-party types to conform that really shouldn't,
>> then that is a case for additional features that allow you to express those
>> requirements more accurately, not an argument for having protocols that
>> can't be conformed to by a third-party. What am I missing?
>>
>>
>> On Sun, Feb 19, 2017 at 11:53 AM, Adrian Zubarev via swift-evolution <
>> swift-evolution at swift.org> wrote:
>>
>>> That’s the whole point I was making. :) Thank you Matthew. That makes my
>>> example a real world example and not just some bike shedding.
>>>
>>> --
>>> Adrian Zubarev
>>> Sent with Airmail
>>>
>>> Am 19. Februar 2017 um 18:50:31, Matthew Johnson (matthew at anandabits.com)
>>> schrieb:
>>>
>>>
>>>
>>> Sent from my iPad
>>>
>>> On Feb 19, 2017, at 11:29 AM, David Waite via swift-evolution <
>>> swift-evolution at swift.org> wrote:
>>>
>>> Just FYI, I solved this issue in my own library (which included a json
>>> jpointer implementation) via:
>>>
>>> public enum SubscriptParameter {
>>>   case string(String)
>>>   case int(Int)
>>> }
>>>
>>> extension SubscriptParameter : ExpressibleByIntegerLiteral {
>>>   public init(integerLiteral value: Int) {
>>>     self = .int(value)
>>>   }
>>> }
>>>
>>> extension SubscriptParameter : ExpressibleByStringLiteral {
>>>   public init(stringLiteral value: String) {
>>>     self = .string(value)
>>>   }
>>>   public init(extendedGraphemeClusterLiteral value: String) {
>>>     self.init(stringLiteral: value)
>>>   }
>>>
>>>   public init(unicodeScalarLiteral value: String) {
>>>     self.init(stringLiteral: value)
>>>   }
>>> }
>>>
>>> extension SubscriptParameter : CustomStringConvertible {
>>>   public var description: String {
>>>     switch self {
>>>     case .string(let str):
>>>       return "\"\(str)\""
>>>     case .int(let i):
>>>       return String(i)
>>>     }
>>>   }
>>> }
>>>
>>> func debug(_ path:SubscriptParameter...) {
>>>   print("path is \(path)")
>>> }
>>>
>>> debug(1, "foo", 2, "bar”) // path is [1, “foo”, 2, “bar”]
>>>
>>>
>>> Can you make this work with variables - not just literals - and still
>>> prevent users from extending the set of types and without requiring an enum
>>> case constructor to be used by clients?
>>>
>>>
>>>
>>> On Feb 19, 2017, at 1:14 AM, Adrian Zubarev <
>>> adrian.zubarev at devandartist.com> wrote:
>>>
>>> If you haven’t followed the other thread Matthew previously opened than
>>> you have missed the example I showed there.
>>>
>>> Here it is again:
>>>
>>> public protocol SubscriptParameterType {
>>>
>>>     // This property was needed to prevent the client from breaking
>>>     // the library by conforming to the protocol, but I'd like to
>>>     // keep it invisible for the client, or even better prevent the
>>>     // client from conforming to the protocol.
>>>     var parameter: Document.SubscriptParameter { get }
>>> }
>>>
>>> extension Document {
>>>
>>>     public enum SubscriptParameter {
>>>
>>>         case string(String)
>>>         case integer(Int)
>>>     }
>>> }
>>>
>>> extension String : SubscriptParameterType {
>>>
>>>     public var parameter: Document.SubscriptParameter {
>>>
>>>         return .string(self)
>>>     }
>>> }
>>>
>>> extension Int : SubscriptParameterType {
>>>
>>>     public var parameter: Document.SubscriptParameter {
>>>
>>>         return .integer(self)
>>>     }
>>> }
>>>
>>> // Somewhere inside the `Document` type
>>> public subscript(firstKey: String, parameters: SubscriptParameterType...) -> Value? { … }
>>>
>>> The absence of closed protocols forced me to create a special
>>> requirement on that protocol to prevent the client from conforming to that
>>> protocol and passing instances of other types my API wouldn’t want to deal
>>> with. That creates unnecessary copies and I need to unpack the enum payload
>>> to find out which type the user passed. Instead I could simply close the
>>> protocol, wouldn’t need the requirement to exist and I could simply cast
>>> the type to String or Int when needed.
>>>
>>> That implementation enables more safe queries of my Document type like
>>>
>>> document["key1", intIndexInstance, stringKeyInstance, 10, "key"]
>>>
>>> rather than
>>>
>>> document["key1/\(intIndexInstance)/\(stringKeyInstance)/10/key"].
>>> ------------------------------
>>>
>>> Here is a list of hidden and semi-hidden protocols from the standard
>>> library that could be closed. Formatted version: https://gist.github.c
>>> om/DevAndArtist/168c800d784829be536c407311953ab7
>>> Path Protocol
>>> /swift/stdlib/public/core/AnyHashable.swift:16
>>> _HasCustomAnyHashableRepresentation
>>> /swift/stdlib/public/core/BidirectionalCollection.swift:21
>>> _BidirectionalIndexable
>>> /swift/stdlib/public/core/BridgeObjectiveC.swift:19
>>> _ObjectiveCBridgeable
>>> /swift/stdlib/public/core/Collection.swift:20 _IndexableBase
>>> /swift/stdlib/public/core/Collection.swift:176 _Indexable
>>> /swift/stdlib/public/core/CompilerProtocols.swift:193
>>> _ExpressibleByBuiltinIntegerLiteral
>>> /swift/stdlib/public/core/CompilerProtocols.swift:240
>>> _ExpressibleByBuiltinFloatLiteral
>>> /swift/stdlib/public/core/CompilerProtocols.swift:283
>>> _ExpressibleByBuiltinBooleanLiteral
>>> /swift/stdlib/public/core/CompilerProtocols.swift:316
>>> _ExpressibleByBuiltinUnicodeScalarLiteral
>>> /swift/stdlib/public/core/CompilerProtocols.swift:350
>>> _ExpressibleByBuiltinExtendedGraphemeClusterLiteral
>>> /swift/stdlib/public/core/CompilerProtocols.swift:398
>>> _ExpressibleByBuiltinStringLiteral
>>> /swift/stdlib/public/core/CompilerProtocols.swift:407
>>> _ExpressibleByBuiltinUTF16StringLiteral
>>> /swift/stdlib/public/core/CompilerProtocols.swift:670
>>> _ExpressibleByStringInterpolation
>>> /swift/stdlib/public/core/CompilerProtocols.swift:709
>>> _ExpressibleByColorLiteral
>>> /swift/stdlib/public/core/CompilerProtocols.swift:720
>>> _ExpressibleByImageLiteral
>>> /swift/stdlib/public/core/CompilerProtocols.swift:730
>>> _ExpressibleByFileReferenceLiteral
>>> /swift/stdlib/public/core/CompilerProtocols.swift:750
>>> _DestructorSafeContainer
>>> /swift/stdlib/public/core/FixedPoint.swift.gyb:53 _Integer
>>> /swift/stdlib/public/core/FixedPoint.swift.gyb:70 _SignedInteger
>>> /swift/stdlib/public/core/FixedPoint.swift.gyb:108
>>> _DisallowMixedSignArithmetic
>>> /swift/stdlib/public/core/Hashable.swift:16 _Hashable
>>> /swift/stdlib/public/core/Index.swift:16 _Incrementable
>>> /swift/stdlib/public/core/IntegerArithmetic.swift.gyb:33
>>> _IntegerArithmetic
>>> /swift/stdlib/public/core/Mirror.swift:721
>>> _DefaultCustomPlaygroundQuickLookable
>>> /swift/stdlib/public/core/MutableCollection.swift:20 _MutableIndexable
>>> /swift/stdlib/public/core/NewtypeWrapper.swift.gyb:16
>>> _SwiftNewtypeWrapper
>>> /swift/stdlib/public/core/Pointer.swift:16 _Pointer
>>> /swift/stdlib/public/core/RandomAccessCollection.swift:20
>>> _RandomAccessIndexable
>>> /swift/stdlib/public/core/RangeReplaceableCollection.swift.gyb:27
>>> _RangeReplaceableIndexable
>>> /swift/stdlib/public/core/ReflectionLegacy.swift:41 _Mirror
>>> /swift/stdlib/public/core/ShadowProtocols.swift:27 _ShadowProtocol
>>> /swift/stdlib/public/core/ShadowProtocols.swift:31 _NSFastEnumeration
>>> /swift/stdlib/public/core/ShadowProtocols.swift:41 _NSEnumerator
>>> /swift/stdlib/public/core/ShadowProtocols.swift:51 _NSCopying
>>> /swift/stdlib/public/core/ShadowProtocols.swift:61 _NSArrayCore
>>> /swift/stdlib/public/core/ShadowProtocols.swift:83 _NSDictionaryCore
>>> /swift/stdlib/public/core/ShadowProtocols.swift:125 _NSDictionary
>>> /swift/stdlib/public/core/ShadowProtocols.swift:137 _NSSetCore
>>> /swift/stdlib/public/core/ShadowProtocols.swift:171 _NSSet
>>> /swift/stdlib/public/core/ShadowProtocols.swift:177 _NSNumber
>>> /swift/stdlib/public/core/ShadowProtocols.swift:187 _NSArrayCore
>>> /swift/stdlib/public/core/ShadowProtocols.swift:188 _NSDictionaryCore
>>> /swift/stdlib/public/core/ShadowProtocols.swift:189 _NSSetCore
>>> /swift/stdlib/public/core/StringBridge.swift:194 _NSStringCore
>>> /swift/stdlib/public/SDK/Foundation/NSError.swift:353
>>> _ObjectiveCBridgeableError
>>> /swift/stdlib/public/SDK/Foundation/NSError.swift:379 __BridgedNSError
>>> /swift/stdlib/public/SDK/Foundation/NSError.swift:446 _BridgedNSError
>>> /swift/stdlib/public/SDK/Foundation/NSError.swift:456
>>> _BridgedStoredNSError
>>> /swift/stdlib/public/SDK/Foundation/NSError.swift:564 _ErrorCodeProtocol
>>>
>>>
>>>
>>> --
>>> Adrian Zubarev
>>> Sent with Airmail
>>>
>>> Am 19. Februar 2017 um 07:59:45, David Waite via swift-evolution (
>>> swift-evolution at swift.org) schrieb:
>>>
>>> I am unsure if this feature is a good idea. Does someone have a
>>> real-world use for this which isn’t just hiding strong implementation
>>> coupling behind a protocol?
>>>
>>> When I consume a protocol, it is under the assumption that the protocol
>>> is documented such that I would be able to work against *any*
>>> implementation of the protocol. With a closed protocol, I would have to
>>> assume that there are significant side effects, either undocumented or
>>> difficult for a third party to duplicate. To my experience, that sounds
>>> brittle.
>>>
>>> Assuming you aren’t switching on the implementing type of a protocol
>>> (which itself can be a sign that your design isn’t properly using
>>> polymorphism), one could get this design by creating a struct with the
>>> interface desired, and passing invocations through to an internal protocol
>>> reference.
>>>
>>> -DW
>>>
>>> > On Feb 18, 2017, at 1:41 PM, Matthew Johnson via swift-evolution <
>>> swift-evolution at swift.org> wrote:
>>> >
>>> > Now that we’re in phase 2 I’d like to officially propose we introduce
>>> `open` protocols and require conformances to `public` protocols be inside
>>> the declaring module. Let’s use this thread for feedback on the official
>>> proposal. After a healthy round of discussion I’ll open a PR to submit it
>>> for review.
>>> >
>>> >
>>> > # Feature name
>>> >
>>> > * Proposal: [SE-NNNN](NNNN-open-public-protocols.md)
>>> > * Authors: [Matthew Johnson](https://github.com/anandabits)
>>> > * Review Manager: TBD
>>> > * Status: **Awaiting review**
>>> >
>>> > ## Introduction
>>> >
>>> > This proposal introduces `open protocol` and changes the meaning of
>>> `public protocol` to match the meaning of `public class` (in this case,
>>> conformances are only allowed inside the declaring module).
>>> >
>>> > The pitch thread leading up to this proposal was: [consistent public
>>> access modifiers](https://lists.swift.org/pipermail/swift-evolution
>>> /Week-of-Mon-20170206/031653.html)
>>> >
>>> > ## Motivation
>>> >
>>> > A general principle the Swift community has adopted for access control
>>> is that defaults should reserve maximum flexibility for a library. The
>>> ensures that any capabilities beyond mere visibility are not available
>>> unless the author of the library has explicitly declared their intent that
>>> the capabilities be made available. Finally, when it is possible to switch
>>> from one semantic to another without breaking clients (but not vice-versa)
>>> we should prefer the more forgiving (i.e. fixable) semantic as the (soft)
>>> default.
>>> >
>>> > `public` is considered a "soft default" in the sense that it is the
>>> first access modifier a user will reach for when exposing a declaration
>>> outside of the module. In the case of protocols the current meaning of
>>> `public` does not meet the principle of preserving maximum flexibility for
>>> the author of the library. It allows users of the library to conform to the
>>> protocol.
>>> >
>>> > There are good reasons a library may not wish to allow users to add
>>> conformances to a protocol. For example, it may not wish to expose the
>>> conforming concrete types. While similar behavior could be accomplished
>>> with an enum if cases could be private, that requires an implementation to
>>> use switch statements rather than polymorphism.
>>> >
>>> > Even if all the conforming types are also public there are cases where
>>> polymorphism is the preferred implementation. For example, if the set of
>>> conforming types is not expected to be fixed (despite all being inside the
>>> library) the authors may not want to have to maintain switch statements
>>> every time they need to add or remove a confroming type which would be
>>> necessary if an enum were used instead. Polymorphism allows us to avoid
>>> this, giving us the ability to add and remove conforming types within the
>>> implementation of the library without the burden of maintaining switch
>>> statements.
>>> >
>>> > Aligning the access modifiers for protocols and classes allows us to
>>> specify both conformable and non-conformable protocols, provides a soft
>>> default that is consistent with the principle of (soft) defaults reserving
>>> maximum flexibility for the library, and increases the overall consistency
>>> of the language by aligning the semantics of access control for protocols
>>> and classes.
>>> >
>>> > The standard library currently has at least one protocol
>>> (`MirrorPath`) that is documented as disallowing client conformances. If
>>> this proposal is adopted it is likely that `MirrorPath` would be declared
>>> `public protocol` and not `open protocol`.
>>> >
>>> > Jordan Rose has indicated that the Apple frameworks also include a
>>> number of protocols documented with the intent that users do not add
>>> conformances. Perhaps an importer annotation would allow the compiler to
>>> enforce these semantics in Swift code as well.
>>> >
>>> > ## Proposed solution
>>> >
>>> > The proposed solution is to change the meaning of `public protocol` to
>>> disallow conformances outside the declaring module and introduce `open
>>> protocol` to allow conformances outside the decalring module (equivalent to
>>> the current meaning of `public protocol`).
>>> >
>>> > ## Detailed design
>>> >
>>> > The detailed design is relatively straightforward but there are three
>>> important wrinkles to consider.
>>> >
>>> > ### User refinement of public protocols
>>> >
>>> > Consider the following example:
>>> >
>>> > ```swift
>>> > // Library module:
>>> > public protocol P {}
>>> > public class C: P {}
>>> >
>>> > // User module:
>>> > protocol User: P {}
>>> > extension C: User {}
>>> > ```
>>> >
>>> > The user module is allowed to add a refinement to `P` because this
>>> does not have any impact on the impelementation of the library or its
>>> possible evolution. It simply allows the user to write code that is generic
>>> over a subset of the conforming types provided by the library.
>>> >
>>> > ### Public protocols with open conforming classes
>>> >
>>> > Consider the following example:
>>> >
>>> > ```swift
>>> > public protocol P P{}
>>> > open class C: P {}
>>> > ```
>>> >
>>> > Users of this module will be able to add subclasses of `C` that have a
>>> conformance to `P`. This is allowed becuase the client of the module did
>>> not need to explicitly declare a conformance and the module has explicitly
>>> stated its intent to allow subclasses of `C` with the `open` access
>>> modifier.
>>> >
>>> > ### Open protocols that refine public protocols
>>> >
>>> > Consider the following example:
>>> >
>>> > ```swift
>>> > // library module:
>>> > public protocol P {}
>>> > open protocol Q: P {}
>>> > open protocol R: P {}
>>> >
>>> > // user module:
>>> > struct S: P {} // error `P` is not `open`
>>> > struct T: Q {} // ok
>>> > struct U: R {} // ok
>>> > ```
>>> >
>>> > The user module is allowed to introudce a conformance to `P`, but only
>>> indirectly by also conforming to `Q`. The meaning we have ascribed to the
>>> keywords implies that this should be allowed and it offers libraries a very
>>> wide design space from which to choose. The library is able to have types
>>> that conform directly to `P`, while placing additional requirements on user
>>> types if necessary.
>>> >
>>> > ## Source compatibility
>>> >
>>> > This proposal breaks source compatibility, but in a way that allows
>>> for a simple mechanical migration. A multi-release stratgegy will be used
>>> to roll out this proposal to provide maximum possible source compatibility
>>> from one release to the next.
>>> >
>>> > 1. In Swift 4, introduce the `open` keyword and the `@nonopen`
>>> attribute (which can be applied to `public protocol` to give it the new
>>> semantics of `public`).
>>> > 2. In Swift 4 (or 4.1 if necessary) start warning for `public
>>> protocol` with no annotation.
>>> > 3. In the subsequent release `public protocol` without annotation
>>> becomes an error.
>>> > 4. In the subsequent relase `public protocol` without annotation takes
>>> on the new semantics.
>>> > 5. `@nonopen` becomes a warning, and evenutally an erro as soon as we
>>> are comfortable making those changes.
>>> >
>>> > ## Effect on ABI stability
>>> >
>>> > I would appreciate it if others can offer input regarding this
>>> section. I believe this proposal has ABI consequences, but it's possible
>>> that it could be an additivie ABI change where the ABI for conformable
>>> protocols remains the same and we add ABI for non-conformable protocols
>>> later. If that is possible, the primary impact would be the ABI of any
>>> standard library protocols that would prefer to be non-conformable.
>>> >
>>> > ## Effect on API resilience
>>> >
>>> > This proposal would may impact one or more protocols in the standard
>>> library, such as `MirrorPath`, which would likely choose to remain `public`
>>> rather than adopt `open`.
>>> >
>>> > ## Alternatives considered
>>> >
>>> > The primary alternatives are to either make no change, or to add
>>> something like `closed protocol`. The issues motivating the current
>>> proposal as a better alternative than either of these options are covered
>>> in the motivation section.
>>> >
>>> > _______________________________________________
>>> > 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/20170219/c3743eb9/attachment.html>

More information about the swift-evolution mailing list