[swift-evolution] Pitch: @required attribute for closures

Matthew Johnson matthew at anandabits.com
Sun Jun 5 21:57:05 CDT 2016


> On Jun 5, 2016, at 9:40 PM, Andrew Bennett <cacoyi at gmail.com> wrote:
> 
> Responses inline
> 
> On Mon, Jun 6, 2016 at 12:23 PM, Matthew Johnson <matthew at anandabits.com <mailto:matthew at anandabits.com>> wrote:
> 
> 
> Sent from my iPad
> 
> On Jun 5, 2016, at 8:43 PM, Andrew Bennett <cacoyi at gmail.com <mailto:cacoyi at gmail.com>> wrote:
> 
>> Thanks Matthew, my responses are inline:
>> 
>> On Mon, Jun 6, 2016 at 10:32 AM, Matthew Johnson <matthew at anandabits.com <mailto:matthew at anandabits.com>> wrote:
>> 
>> 
>> Sent from my iPad
>> 
>> On Jun 5, 2016, at 6:50 PM, Andrew Bennett <cacoyi at gmail.com <mailto:cacoyi at gmail.com>> wrote:
>> 
>>> Perhaps I was unclear, in my explanation. The guarantee I'm enforcing is that the closure is called exactly once before being released.
>>> 
>>> Everything I suggested is a compile-time check.
>>> 
>>> The compile-time warning and runtime `fatalError` I suggested could be replaced with a compile-time error, however even in this case it is still statically checked for the warning.
>>> 
>>> The compiler can statically guarantee exactly one of these things happens in methods using the closure:
>>> the closure is called
>>> the closure is stored
>>> the closure is passed to another method
>>> the program aborts with something like a fatalError
>>> If the closure is stored then there must be a deinit, and those checks apply there as well.
>>> 
>>> I believe this is sufficient to ensure the closure is called once. Please let me know if there are any cases these checks miss.
>> 
>> If the closure is stored in a member it could be called in the implementation of any other member.  Calls to other members could come from arbitrary locations in the surrounding program at arbitrary points in time (unless you have static analysis that can prove a narrower set of possibilities).
>>  
>> This isn't a problem if the member has to have the type annotation, all uses of the member, whether from elsewhere in the program, or other methods, would have to pass the checks.
>> 
>> If you call the closure you must nil/replace the member.
> 
> Ok, so the member must be an optional?  If that is the idea I would suggest considering something similar to 'weak' where it automatically gets set to nil after it is called (which could maybe become a property behavior in the future).
>  
> That's a great idea, auto-nilling would work well. I was also thinking it should be possible to replace an existing non-optional stored closure if you call the old one in the old scope (see the callClosure method in my previous post). However auto-nilling could simplify the concept.
> 
> 
> I don't think you mentioned the case of reassigning the member when it is non-nil.  You would have to require users to verify it is nil before setting it or if it is not nil, calling it before assigning to it.
> 
>> 
>> And if you have a model that relies on behavior in a deinit then storing the closure won't be possible for structs.
>>  
>> This is true, considering you don't want to copy a @once closure you probably don't want value-type semantics anyway.
> 
> That's a good point.  Since we don't have control over copy behavior in Swift it wouldn't make sense at all unless / until we can make structs that have move semantics (maybe if / when we get a Rust-like ownership system?).
> Interesting, thanks, I wasn't aware of Rust's ownership concept - I've been working on a related proposal for Swift.

Rust is really leading the way here in bringing linear types into a mainstream language.  Getting it right isn’t easy.  Rust’s system has evolved quite a bit over the years since they began.  It will be good to learn as much as possible from Rust (both what it gets right and also areas that could maybe be improved). 

If you have a draft of anything and would like feedback feel free to send me a link (on or off list).  I have a lot of interest in this area as well.

> 
> 
>> 
>> You have also missed the case that the closure is captured by another closure (maybe it is a completion block and you call it in a completion block of a method your method calls).
>>  
>> This is correct. I forgot to mention that I'm sorry, thanks for pointing it out!
>> 
>> I was thinking that a @once closure can only be captured by another @once closure. We can add that as another dot-point:
>> 
>> the closure is captured by another @once closure, this is the only time it can be captured.
> 
> I like the idea behind this proposal in theory.  However, it really seems to cry out for linear types.  I have a feeling we would end up with a better (and more general) solution if Swift goes down that path in the future.  At minimum, I would like to hear input from those who work on the type system.  If a more robust, less ad-hoc solution will be possible in the future it might be best to wait.
> 
> On the other hand, completion callbacks that require this guarantee are pretty common.  The semantic is part of the contract whether we have language support for it or not.  Maybe we can do something now that could be subsumed by a more general feature in the future...
> 
> I have the same thinking, I'd like the semantics now, and as as strongly as the language can allow. If the language gets better type support then we get better or less restrictive assurances.
> 
>>  
>> 
>>> 
>>> On Sun, Jun 5, 2016 at 11:59 PM, Matthew Johnson <matthew at anandabits.com <mailto:matthew at anandabits.com>> wrote:
>>> 
>>> 
>>> Sent from my iPad
>>> 
>>> On Jun 5, 2016, at 8:52 AM, Andrew Bennett <cacoyi at gmail.com <mailto:cacoyi at gmail.com>> wrote:
>>> 
>>>> Storing into a member would be fine, as long as it must keep @once as a type annotation and the compiler makes sure you maintain:
>>>>     sum(callCount, storeCount, passCount) == 1
>>>> 
>>>> For example:
>>>>   class Example {
>>>>     private var closure: (@once (T) -> Void)?
>>>> 
>>>>     func callClosure(value: T, replace: (@once (T) -> Void)? = nil) {
>>>>       // the compiler should error if it detects the closure:
>>>>       //  * escaping more than once, while still being stored,
>>>>       //  * or being called while still being stored or escaping,
>>>>       //  * or being overwritten without being called
>>>>       if let closure = self.closure {
>>>>         self.closure = replace
>>>>         closure(value)
>>>>       }
>>>>     }
>>>> 
>>>>     deinit {
>>>>       // compiler warning: that closure is potentially un-called
>>>>       // runtime fatalError if it's .Some(Closure) after deinit
>>>>     }
>>>>   }
>>>> 
>>>> There could be a standard library type with those guarantees built in.
>>> 
>>> I don't consider this compiler verification.  It is runtime verification.  The best the compiler can do is enforce constraints that allow for guaranteed runtime verification.  You can argue that is better than nothing but it is not a static guarantee of correct behavior.
>>> 
>>>> 
>>>> 
>>>> On Sun, Jun 5, 2016 at 10:12 PM, Matthew Johnson <matthew at anandabits.com <mailto:matthew at anandabits.com>> wrote:
>>>> 
>>>> 
>>>> Sent from my iPad
>>>> 
>>>> On Jun 5, 2016, at 6:56 AM, Andrew Bennett <cacoyi at gmail.com <mailto:cacoyi at gmail.com>> wrote:
>>>> 
>>>>> I like this.
>>>>> 
>>>>> One of the suggestions on @noescape(once) was that it just becomes @once and works with escaping closures too. It might be possible if compile time checks verified that the closure isn't copied, and that it is called before being deinit-ialized. Failing that I'm happy with a runtime circumstance in the cases the compiler can't check.
>>>> 
>>>> Yeah, maybe if it is only used asynchronously and never stored in a member or global it could be verified and that is a pretty common case.  That would certainly be easier than the general case.
>>>> 
>>>> I prefer @once over @required if the guarantee is single execution.  If the guarantee is *at least once* obviously @once is not the right attribute, but I'm not convinced @required is either.  Maybe @invoked.
>>>> 
>>>>> 
>>>>> It would be great if @required took into the account the feedback from that proposal and considered the synchronous case too.
>>>>> 
>>>>> As an aside, you can get some of the guarantees you want like this:
>>>>> 
>>>>> func doSomething(completionHandler: (SomeEnum) -> ()) {
>>>>>   dispatch_async(someQueue) {
>>>>>     let result: SomeEnum
>>>>>     // the compiler ensures 'result' is set
>>>>>     defer { completionHandler(result) }
>>>>> 
>>>>>     if aCondition {
>>>>>       if bCondition {
>>>>>         result = .Foo
>>>>>       } else {
>>>>>         result = .Bar
>>>>>       }
>>>>>       // the compiler ensures you do this, because it is 'let'
>>>>>       return
>>>>>     }
>>>>> 
>>>>>     if cCondition {
>>>>>       result = .Baz
>>>>>     }
>>>>>   }
>>>>> }
>>>>> 
>>>>> On Sun, Jun 5, 2016 at 9:42 PM, Matthew Johnson via swift-evolution <swift-evolution at swift.org <mailto:swift-evolution at swift.org>> wrote:
>>>>> 
>>>>> 
>>>>> Sent from my iPad
>>>>> 
>>>>> On Jun 5, 2016, at 5:02 AM, Patrick Pijnappel via swift-evolution <swift-evolution at swift.org <mailto:swift-evolution at swift.org>> wrote:
>>>>> 
>>>>>> This has actually been proposed before, see SE-0073: https://github.com/apple/swift-evolution/blob/master/proposals/0073-noescape-once.md <https://github.com/apple/swift-evolution/blob/master/proposals/0073-noescape-once.md>
>>>>> Actually that proposal was for noescape closures and this suggestion is for escaping closures.  I don't think the compiler can verify this for noescape closures.  If it is possible it would be far more complicated.
>>>>> 
>>>>>> 
>>>>>> 
>>>>>> On Sun, Jun 5, 2016 at 11:37 AM, Charles Srstka via swift-evolution <swift-evolution at swift.org <mailto:swift-evolution at swift.org>> wrote:
>>>>>> MOTIVATION:
>>>>>> 
>>>>>> As per the current situation, there is a pitfall when writing asynchronous APIs that does not occur when writing synchronous APIs. Consider the following synchronous API:
>>>>>> 
>>>>>> func doSomething() -> SomeEnum {
>>>>>>         if aCondition {
>>>>>>                 if bCondition {
>>>>>>                         return .Foo
>>>>>>                 } else {
>>>>>>                         return .Bar
>>>>>>                 }
>>>>>>         } else {
>>>>>>                 if cCondition {
>>>>>>                         return .Baz
>>>>>>                 }
>>>>>>         }
>>>>>> }
>>>>>> 
>>>>>> The compiler will give an error here, since if both aCondition and cCondition are false, the function will not return anything.
>>>>>> 
>>>>>> However, consider the equivalent async API:
>>>>>> 
>>>>>> func doSomething(completionHandler: (SomeEnum) -> ()) {
>>>>>>         dispatch_async(someQueue) {
>>>>>>                 if aCondition {
>>>>>>                         if bCondition {
>>>>>>                                 completionHandler(.Foo)
>>>>>>                         } else {
>>>>>>                                 completionHandler(.Bar)
>>>>>>                         }
>>>>>>                 } else {
>>>>>>                         if cCondition {
>>>>>>                                 completionHandler(.Baz)
>>>>>>                         }
>>>>>>                 }
>>>>>>         }
>>>>>> }
>>>>>> 
>>>>>> Whoops, now the function can return without ever firing its completion handler, and the problem might not be discovered until runtime (and, depending on the complexity of the function, may be hard to find).
>>>>>> 
>>>>>> PROPOSED SOLUTION:
>>>>>> 
>>>>>> Add a @required attribute that can be applied to closure arguments. This attribute simply states that the given closure will always be eventually called, and the compiler can enforce this.
>>>>>> 
>>>>>> DETAILED DESIGN:
>>>>>> 
>>>>>> - The @required attribute states in our API contract that a given closure *must* be called at some point after the function is called.
>>>>>> 
>>>>>> - Standard API calls like dispatch_async that contractually promise to execute a closure or block get @required added to their signatures.
>>>>>> 
>>>>>> - When the compiler sees a @required closure in a function declaration, it checks to make sure that every execution path either calls the closure at some point, or sends a @required closure to another API that eventually ends up calling the closure.
>>>>>> 
>>>>>> - If there’s a way for a @required closure not to be called, the compiler emits an error letting the developer know about the bug in his/her code.
>>>>>> 
>>>>>> IMPACT ON EXISTING CODE:
>>>>>> 
>>>>>> None. This is purely additive.
>>>>>> 
>>>>>> ALTERNATIVES CONSIDERED:
>>>>>> 
>>>>>> I got nothin’.
>>>>>> 
>>>>>> Charles
>>>>>> 
>>>>>> _______________________________________________
>>>>>> swift-evolution mailing list
>>>>>> swift-evolution at swift.org <mailto:swift-evolution at swift.org>
>>>>>> https://lists.swift.org/mailman/listinfo/swift-evolution <https://lists.swift.org/mailman/listinfo/swift-evolution>
>>>>>> 
>>>>>> _______________________________________________
>>>>>> swift-evolution mailing list
>>>>>> swift-evolution at swift.org <mailto:swift-evolution at swift.org>
>>>>>> https://lists.swift.org/mailman/listinfo/swift-evolution <https://lists.swift.org/mailman/listinfo/swift-evolution>
>>>>> 
>>>>> _______________________________________________
>>>>> swift-evolution mailing list
>>>>> swift-evolution at swift.org <mailto: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/20160605/a5cad4cb/attachment.html>


More information about the swift-evolution mailing list