[swift-evolution] [Proposal draft] Limiting @objc inference

Rod Brown rodney.brown6 at icloud.com
Fri Jan 6 05:46:29 CST 2017


> On 5 Jan 2017, at 11:50 am, Douglas Gregor via swift-evolution <swift-evolution at swift.org> wrote:
> 
> Hi all,
> 
> Here’s a draft proposal to limit inference of @objc to only those places where we need it for consistency of the semantic model. It’s in the realm of things that isn’t *needed* for ABI stability, but if we’re going to make the source-breaking change here we’d much rather do it in the Swift 4 time-frame than later. Proposal is at:
> 
> 	https://github.com/DougGregor/swift-evolution/blob/objc-inference/proposals/NNNN-objc-inference.md <https://github.com/DougGregor/swift-evolution/blob/objc-inference/proposals/NNNN-objc-inference.md>
> 
> Introduction
> 
> One can explicitly write @objc on any Swift declaration that can be expressed in Objective-C. As a convenience, Swift also infers @objc in a number of places to improve interoperability with Objective-C and eliminate boilerplate. This proposal scales back the inference of @objc to only those cases where the declaration must be available to Objective-C to maintain semantic coherence of the model, e.g., when overriding an @objc method or implementing a requirement of an @objcprotocol. Other cases currently supported (e.g., a method declared in a subclass of NSObject) would no longer infer @objc, but one could continue to write it explicitly to produce Objective-C entry points.
> 
> Swift-evolution thread: here <https://lists.swift.org/pipermail/swift-evolution/Week-of-Mon-20160509/017308.html>
>  <https://github.com/DougGregor/swift-evolution/tree/objc-inference#motivation>Motivation
> 
> There are several observations motivating this proposal. The first is that Swift's rules for inference of @objc are fairly baroque, and it is often unclear to users when @objc will be inferred. This proposal seeks to make the inference rules more straightforward. The second observation is that it is fairly easy to write Swift classes that inadvertently cause Objective-C selector collisions due to overloading, e.g.,
> 
> class MyNumber : NSObject {
>   init(_ int: Int) { }
>   init(_ double: Double) { } // error: initializer 'init' with Objective-C selector 'init:' 
>       // conflicts with previous declaration with the same Objective-C selector
> }
> The example above also illustrates the third observation, which is that code following the Swift API Design Guidelines <https://swift.org/documentation/api-design-guidelines/> will use Swift names that often translate into very poor Objective-C names that violate the Objective-C Coding Guidelines for Cocoa <https://developer.apple.com/library/content/documentation/Cocoa/Conceptual/CodingGuidelines/CodingGuidelines.html>. Specifically, the Objective-C selectors for the initializers above should include a noun describing the first argument, e.g., initWithInteger: and initWithDouble:, which requires explicit @objc annotations anyway:
> 
> class MyNumber : NSObject {
>   @objc(initWithInteger:) init(_ int: Int) { }
>   @objc(initWithDouble:) init(_ double: Double) { }
> }
> The final observation is that there is a cost for each Objective-C entry point, because the Swift compiler must create a "thunk" method that maps from the Objective-C calling convention to the Swift calling convention and is recorded within Objective-C metadata. This increases the size of the binary (preliminary tests on some Cocoa[Touch] apps found that 6-8% of binary size was in these thunks alone, some of which are undoubtedly unused), and can have some impact on load time (the dynamic linker has to sort through the Objective-C metadata for these thunks).
> 
>  <https://github.com/DougGregor/swift-evolution/tree/objc-inference#proposed-solution>Proposed solution
> 
> The proposed solution is to limit the inference of @objc to only those places where it is required for semantic consistency of the programming model. 
> 
>  <https://github.com/DougGregor/swift-evolution/tree/objc-inference#constructs-that-still-infer-objc>Constructs that (still) infer @objc
> 
> Specifically, @objc will continue to be inferred for a declaration when:
> 
> The declaration is an override of an @objc declaration, e.g.,
> 
> class Super {
>   @objc func foo() { }
> }
> 
> class Sub : Super {
>   /* inferred @objc */
>   override func foo() { }
> }
> This inference is required so that Objective-C callers to the method Super.foo() will appropriately invoke the overriding method Sub.foo().
> 
> The declaration satisfies a requirement of an @objc protocol, e.g.,
> 
> @objc protocol MyDelegate {
>   func bar()
> }
> 
> class MyClass : MyDelegate {
>   /* inferred @objc */
>   func bar() { }
> }
> This inference is required because anyone calling MyDelegate.bar(), whether from Objective-C or Swift, will do so via an Objective-C message send, so conforming to the protocol requires an Objective-C entry point.
> 
> The declaration has the @IBAction or @IBOutlet attribute. This inference is required because the interaction with Interface Builder occurs entirely through the Objective-C runtime, and therefore depends on the existence of an Objective-C entrypoint.
> 
> The declaration has the @NSManaged attribute. This inference is required because the interaction with CoreData occurs entirely through the Objective-C runtime, and therefore depends on the existence of an Objective-C entrypoint.
> 
> 
+1.

> The list above describes cases where Swift 3 already performs inference of @objc and will continue to do so if this proposal is accepted.
> 
>  <https://github.com/DougGregor/swift-evolution/tree/objc-inference#dynamic-no-longer-infers-objc>dynamic no longer infers @objc
> 
> A declaration that is dynamic will no longer infer @objc. For example:
> 
> class MyClass {
>   dynamic func foo() { }       // error: 'dynamic' method must be '@objc'
>   @objc dynamic func bar() { } // okay
> }
> This change is intended to separate current implementation limitations from future language evolution: the current implementation supports dynamic by always using the Objective-C message send mechanism, allowing replacement of dynamic implementations via the Objective-C runtime (e.g., class_addMethod and class_replaceMethod). In the future, it is plausible that the Swift language and runtime will evolve to support dynamic without relying on the Objective-C runtime, and it's important that we leave the door open for that language evolution.
> 
> This change therefore does two things. First, it makes it clear that the dynamic behavior is tied to the Objective-C runtime. Second, it means that well-formed Swift 4 code will continue to work in the same way should Swift gain the ability to provide dynamic without relying on Objective-C: at that point, the method foo() above will become well-formed, and the method bar() will continue to work as it does today through the Objective-C runtime. Indeed, this change is the right way forward even if Swift never supports dynamic in its own runtime, following the precedent of SE-0070 <https://github.com/apple/swift-evolution/blob/master/proposals/0070-optional-requirements.md>, which required the Objective-C-only protocol feature "optional requirements" to be explicitly marked with @objc.
> 
>  <https://github.com/DougGregor/swift-evolution/tree/objc-inference#nsobject-derived-classes-no-longer-infer-objc>NSObject-derived classes no longer infer @objc
> 
> A declaration within an NSObject-derived class will no longer infer objc`. For example:
> 
> class MyClass : NSObject {
>   func foo() { } // not exposed to Objective-C in Swift 4
> }
> This is the only major change of this proposal, because it means that a large number of methods that Swift 3 would have exposed to Objective-C (and would, therefore, be callable from Objective-C code in a mixed project) will no longer be exposed. On the other hand, this is the most unpredictable part of the Swift 3 model, because such methods infer @objconly when the method can be expressed in Objective-C. For example:
> 
> extension MyClass {
>   func bar(param: ObjCClass) { } // exposed to Objective-C in Swift 3; not exposed by this proposal
>   func baz(param: SwiftStruct) { } // not exposed to Objective-C
> }
> With this proposal, neither method specifies @objc nor is either required by the semantic model to expose an Objective-C entrypoint, so they don't infer @objc: there is no need to reason about the type of the parameter's suitability in Objective-C.
> 
>  <https://github.com/DougGregor/swift-evolution/tree/objc-inference#side-benefit-more-reasonable-expectations-for-objc-protocol-extensions>Side benefit: more reasonable expectations for @objc protocol extensions
> 
> Users are often surprised to realize that extensions of @objc protocols do not, in fact, produce Objective-C entrypoints:
> 
> @objc protocol P { }
> 
> extension P {
>   func bar() { }
> }
> 
> class C : NSObject { }
> 
> let c = C()
> print(c.respondsToSelector("bar")) // prints "false"
> The expectation that P.bar() has an Objective-C entry point is set by the fact that NSObject-derived Swift classes do implicitly create Objective-C entry points for declarations within class extensions when possible, but Swift does not (and, practically speaking, cannot) do the same for protocol extensions.
> 
> A previous mini-proposal discussed here <https://lists.swift.org/pipermail/swift-evolution/Week-of-Mon-20160104/005312.html> suggested requiring @nonobjc for members of @objc protocol extensions. However, limiting inference of @objc eliminates the expectation itself, addressing the problem from a different angle.
> 
> 
+1. I like this cleanup.
>  <https://github.com/DougGregor/swift-evolution/tree/objc-inference#source-compatibility>Source compatibility
> 
> The two changes that remove inference of @objc are both source-breaking in different ways. The dynamic change mostly straightforward:
> 
> In Swift 4 mode, introduce an error that when a dynamic declaration does not explicitly state @objc, with a Fix-It to add the @objc.
> 
> In Swift 3 compatibility mode, continue to infer @objc for dynamic methods. However, introduce a warning that such code will be ill-formed in Swift 4, along with a Fix-It to add the @objc. This
> 
> A Swift 3-to-4 migrator could employ the same logic as Swift 3 compatibility mode to update dynamic declarations appropriately.
> 
> The elimination of inference of @objc for declarations in NSObject subclasses is more complicated. Considering again the three cases:
> 
> In Swift 4 mode, do not infer @objc for such declarations. Source-breaking changes that will be introduced include:
> 
> If #selector or #keyPath refers to one such declaration, an error will be produced on previously-valid code that the declaration is not @objc. In most cases, a Fix-It will suggest the addition of @objc.
> 
> The lack of @objc means that Objective-C code in mixed-source projects won't be able to call these declarations. Most problems caused by this will result in warnings or errors from the Objective-C compiler (due to unrecognized selectors), but some might only be detected at runtime. These latter cases will be hard-to-detect.
> 
> Other tools and frameworks that rely on the presence of Objective-C entrypoints but do not make use of Swift's facilities for referring to them will fail. This case is particularly hard to diagnose well, and failures of this sort are likely to cause runtime failures that only the developer can diagnose and correct.
> 
> In Swift 3 compatibility mode, continue to infer @objc for these declarations. When @objc is inferred based on this rule, modify the generated header (i.e., the header used by Objective-C code to call into Swift code) so that the declaration contains a "deprecated" attribute indicating that the Swift declaration should be explicitly marked with @objc. For example:
> 
> class MyClass : NSObject {
>   func foo() { }
> }
> will produce a generated header that includes:
> 
> @interface MyClass : NSObject
> -(void)foo NS_DEPRECATED("MyClass.foo() requires an explicit `@objc` in Swift 4");
> @end
> This way, any reference to that declaration from Objective-C code will produce a warning about the deprecation. Users can silence the warning by adding an explicit @objc.
> 
> A Swift 3-to-4 migrator is the hardest part of the story. Ideally, the migrator to only add @objc in places where it is needed, so that we see some of the expected benefits of code-size reduction. However, there are two problems with doing so:
> 
> Some of the uses that imply the need to add @objc come from Objective-C code, so a Swift 3-to-4 migrator would also need to compile the Objective-C code (possibly with a modified version of the Objective-C compiler) and match up the "deprecated" warnings mentioned in the Swift 3 compatibility mode bullet with Swift declarations.
> 
> The migrator can't reason about dynamically-constructed selectors or the behavior of other tools that might directly use the Objective-C runtime, so failing to add a @objc will lead to migrated programs that compile but fail to execute correctly.
> 
> 
Yeah, this is definitely a point of concern. I’m not sure that the first part deals with stuff like frameworks and libraries that currently expose functionality to external users in Obj-C. How could you be sure which API was designed to be used by Obj-C when it comes to libraries?

As others have mentioned, perhaps we could have options in the converter? I’m not generally a fan of this type of preference selection but in this case it seems reasonable.

The dynamically constructed selector case seems rare, especially considering that using string-based selectors has been deprecated since Swift 2.2.
>  <https://github.com/DougGregor/swift-evolution/tree/objc-inference#overriding-of-declarations-introduced-in-class-extensions>Overriding of declarations introduced in class extensions
> 
> Swift's class model doesn't support overriding of declarations introduced in class extensions. For example, the following code produces an amusing error message on the override:
> 
> class MySuperclass { }
> 
> extension MySuperclass {
>   func extMethod() { }
> }
> 
> class MySubclass : MySuperclass { }
> 
> extension MySubclass {
>   override func extMethod() { }   // error: declarations in extensions cannot override yet
> }
> However, this does work in Swift 3 when the method is @objc, e.g.,
> 
> class MySuperclass { }
> 
> extension MySuperclass {
>   @objc func extMethod() { }
> }
> 
> class MySubclass : MySuperclass { }
> 
> extension MySubclass {
>   override func extMethod() { }   // okay! Objective-C message dispatch allows this
> }
> Removing @objc inference for NSObject subclasses will therefore break this correct Swift 3 code:
> 
> class MySuperclass { }
> 
> extension MySuperclass : NSObject {
>   func extMethod() { } // implicitly @objc in Swift 3, not @objc in Swift 4
> }
> 
> class MySubclass : MySuperclass { }
> 
> extension MySubclass {
>   override func extMethod() { }   // okay in Swift 3, error in Swift 4: declarations in extensions cannot override yet
> }
> There are several potential solutions to this problem, but both are out-of-scope for this particular proposal:
> 
> Require that a non- at objc declaration in a class extension by explicitly declared final so that it is clear from the source that this declaration cannot be overridden.
> 
> Extend Swift's class model to permit overriding of declarations introduced in extensions.
> 
> 
I will definitely be running into this issue if/when this occurs. A lot of users use extensions to break up their code into clear elements for organisational reasons, and also adding functionality to external libraries. Both here would end up with odd behaviour and I really hope we go with Option 2 here.
>  <https://github.com/DougGregor/swift-evolution/tree/objc-inference#alternatives-considered>Alternatives considered
> 
> Aside from the obvious alternative of "do nothing", there are ways to address some of the problems called out in theMotivation <https://github.com/DougGregor/swift-evolution/tree/objc-inference#motivation> section without eliminating inference in the cases we're talking about, or to soften the requirements on some constructs.
> 
>  <https://github.com/DougGregor/swift-evolution/tree/objc-inference#mangling-objective-c-selectors>Mangling Objective-C selectors
> 
> Some of the problems with Objective-C selector collisions could be addressed by using "mangled" selector names for Swift-defined declarations. For example, given:
> 
> class MyClass : NSObject {
>   func print(_ value: Int) { }
> }
> Instead of choosing the Objective-C selector "print:" by default, which is likely to conflict, we could use a mangled selector name like "__MyModule__MyClass__print__Int:" that is unlikely to conflict with anything else in the program. However, this change would also be source-breaking for the same reasons that restricting @objc inference is: dynamic behavior that constructs Objective-C selectors or tools outside of Swift that expect certain selectors will break at run-time.
> 
>  <https://github.com/DougGregor/swift-evolution/tree/objc-inference#completely-eliminating-objc-inference>Completely eliminating @objc inference
> 
> Another alternative to this proposal is to go further and completely eliminate @objc inference. This would simplify the programming model further---it's exposed to Objective-C only if it's marked @objc---but at the cost of significantly more boilerplate for applications that use Objective-C frameworks. For example:
> 
> class Sub : Super {
>   @objc override func foo() { }  // @objc is now required
> }
> 
> class MyClass : MyDelegate {
>   @objc func bar() { }  // @objc is now required
> }
> I believe that this proposal strikes the right balance already, where @objc is inferred when it's needed to maintain the semantic model, and can be explicitly added to document those places where the user is intentionally exposing an Objective-C entrypoint for some reason. Thus, explicitly writing @objc indicates intent without creating boilerplate.
> 
>  <https://github.com/DougGregor/swift-evolution/tree/objc-inference#proposal-add-on-objc-annotations-on-class-definitions-and-extensions>Proposal add-on: @objc annotations on class definitions and extensions
> 
> If the annotation burden of @objc introduced by this proposal is too high, we could make @objc on a class definition or extension thereof imply @objc on those members that can be exposed to Objective-C. For example:
> 
> @objc extension MyClass {
>   // implicitly @objc
>   func f() { }
> 
>   // Cannot be exposed to Objective-C because tuples aren't representable in Objective-C
>   func g() -> (Int, Int) { return (1, 2) }
> }
> This would reduce (but not eliminate) the annotation burden introduced by this proposal, allowing developers to group Objective-C-exposed declarations together under a single @objc annotation. This reduces the amount of boilerplate. With such a change, we'd need to decide what to do with MyClass.g(), which could be either:
> 
> Make it an error because the context implies that it is @objc, but it cannot be. The error would be suppressed with an explicit @nonobjc annotation.
> 
> Make it implicitly @nonobjc.
> 
> Option (1) seems more in line with the rest of the proposal, because it maintains a predictable model. Regardless, this add-on makes sense if we like the benefits of the proposal to limit @objc inference but the annotation burden turns out to be annoyingly high.
> 
> 
> 	- Doug
> 
> 
> _______________________________________________
> swift-evolution mailing list
> 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/20170106/516a20aa/attachment.html>


More information about the swift-evolution mailing list