[swift-evolution] [Proposal] Guarded self in closures

Matthew Johnson matthew at anandabits.com
Wed Feb 22 18:19:02 CST 2017 

> On Feb 22, 2017, at 6:06 PM, Anton Mironov <antonvmironov at gmail.com> wrote:
> 
> 
>> On Feb 23, 2017, at 01:18, Matthew Johnson <matthew at anandabits.com <mailto:matthew at anandabits.com>> wrote:
>> 
>>> 
>>> On Feb 22, 2017, at 5:06 PM, Anton Mironov <antonvmironov at gmail.com <mailto:antonvmironov at gmail.com>> wrote:
>>> 
>>> -1
>>> I support improvements in this area but I do not think that adding guarded closures will fix the case.
>>> It raises multiple concerns:
>>> - prepending ? to the closure declaration is as forgettable as `[weak self]`
>> 
>> No, this is why I included the `@guarded` parameter annotation.  This allows an API to require its callers to use a guarded closure.  Strong references would have to be explicit in the capture list.
>> 
>>> - reactive programming often assumes chaining of operations. How guarded closures affect next operations in the chain?
>> 
>> Can you provide a concrete example of real world code you wrote manually?  I will convert it to use guarded closures to show how it is affected.
> 
> You can use the second piece of code I’ve provided before.

How do `map` and `onUpdate` store the reference to the context?  Is it weak?  If so, what happens when the context is released?  After you answer that I will be able to show how it would look under this proposal.

> 
>> 
>>> - the closure must exist until either the control deallocates (source of actions) or self deallocates (destination of actions). Guarded closure will not provide an expected behavior
>> 
>> Yes they will.  The guarded closure lives until the control releases it.  But it becomes a no-op if any of the references captured with a guard are released before that happens.  This is much like the behavior of the target / action pattern but generalized to support closures.
> 
> I doubt that turning closure into no-op is a simple thing to do. It will require having a registry of closures that depend on an instance. A runtime will have to go through the registry and turn closures into no-op. Or there is an another solution that I do not see.

What I mean when I say no-op is that the code the user places in the closure would be prefixed by a `guard` clause with an early return.  The no-op is when the guard clause is triggered, just as if you had written it manually.

I imagine the compiler might also place an additional check inside the `else` of the `guard` which would release any context it was still hanging on to as it would know that the context was no longer needed.  Imagine this as if the compiler synthesized code had access to an optional strong reference to the context and set it to nil in the else clause.  It would also be possible for this closure to expose an `isAlive: Bool` property that would check whether or not the context was still around.  This is a bit of hand waving - I’m sure the real implementation would be more sophisticated.  But I think that conveys the basic idea.

Here is some pseudocode:

let foo: Foo = Foo()
let bar: Bar = Bar()

// hand out references to the owner that controls the lifetime of foo and bar

?{ 
         // compiler synthesized code
         // this is a property on the closure object itself, not visible within the scope of the user code in the closure
         // it is used by libraries to detect when they can safely discard their reference to the closure because it has become a no-op
         // context is a super secret compiler reference to the context of the closure
         var isActive: Bool { return context != nil }

       // compiler synthesized code that prefixes the user code
       guard let foo = foo, let bar = bar else {
           context = nil
           return
       }
       // end compiler synthesized code
       
       // begin user code
       // do something with foo and bar
}

> 
>> 
>>> - managing lifecycle of nested guarded closures could be complex to understand and implement into the language
>> 
>> I’m glad you brought this up.  I’ll give it some thought.  If there does turn out to be complexity involved I wouldn’t have a problem prohibiting that.
>> 
>>> - why would you consider using @escaping instead of @guarded?
>> 
>> Because sometimes the right default for a function taking an escaping closure is a strong reference.  I wouldn't want `DispatchQueue.async` to take a guarded closure.  That API doesn’t contain any semantic content around *why* you dispatched async.  It’s not a callback, but instead a way of moving work around.
>> 
>>> 
>>> I personally prefer doing something like this:
>>> 
>>> ```swift
>>> self.button.onAction(forEvents: [.touchUpInside], context: self) { (self, sender, event) in
>>> 	self.performSearch(query: self.searchField.text)
>>> }
>>> ```
>>> 
>>> or
>>> 
>>> ```swift
>>> self.button.actions(forEvents: [.touchUpInside])
>>>     .debounce(interval: 3.0)
>>>     .map(context: self) { (self, _) in
>>>         return self.searchField.text
>>>     }
>>>     .distinct()
>>>     .onUpdate(context: self) { (self, searchQuery) in
>>>         self.performSearch(query: searchQuery)
>>>     }
>>> ```
>>> 
>>> This code neither requires an addition of language features nor contains retain cycles. All closures will be released as soon as source or destination deallocates.
>> 
>> This isn’t too bad but it does require manually threading the context.  This is more work for both the library and the client than necessary.  It also does not help users avoid an accidental strong reference in the closure.  It nudges them not to by offering to thread the context but it doesn’t do anything to prevent it.  You can still create a strong reference (event to self) without specifying it in the capture list.
> 
> You are correct. This will code will not help to avoid accidental strong reference. But it gives an opportunity to do things without explicit weak references just as guarded closures do. It also adds an ability to avoid an execution of pure (or just not bound to context) operations you depends on. Deallocation of context will lead to cancellation of full chain of operations and unsubscription from button event.
> 
>> I think there is a place for a language solution here.
> 
> The only language solution I expect is a static analyzer warning about retain cycle (like in ObjC).
> 
> I’m starting to think that my solution is similar to yours. I’ve done these things with a library rather than with language support. I will definitely take advantage of guarded self in closures as soon as the proposal will be accepted. But I would prefer and suggest using my solution for now.
> 
>> 
>>> 
>>>> On Feb 22, 2017, at 22:57, Matthew Johnson via swift-evolution <swift-evolution at swift.org <mailto:swift-evolution at swift.org>> wrote:
>>>> 
>>>> Hi David,
>>>> 
>>>> I just shared a draft proposal to introduce guarded closures last week: https://lists.swift.org/pipermail/swift-evolution/Week-of-Mon-20170213/032478.html <https://lists.swift.org/pipermail/swift-evolution/Week-of-Mon-20170213/032478.html>.  I think you would find it very interesting.
>>>> 
>>>> I considered including a new capture list specifier `guard` in this proposal but decided against it.  Guarded behavior requires prefixing the contents of the closure with a guard clause that returns immediately if the guard is tripped.  This is a property of the closure as a whole, not of an individual capture.  For that reason, I decided that allowing a `guard` specifier for an individual capture would be inappropriate.  
>>>> 
>>>> Instead, a guarded closure has a guarded by default capture behavior which can be overridden with `weak`, `unowned` or `strong` in the capture list.  The thread on this proposal was relatively brief.  I plan to open a PR soon after making a few minor modifications.
>>>> 
>>>> Matthew
>>>> 
>>>>> On Feb 22, 2017, at 2:48 PM, David Hedbor via swift-evolution <swift-evolution at swift.org <mailto:swift-evolution at swift.org>> wrote:
>>>>> 
>>>>> Hello,
>>>>> 
>>>>> (apologies if this got sent twice - gmail and Apple mail seems to confused as to what account the first mail was sent from)
>>>>> 
>>>>> I’m new to this mailing list, but have read some archived messages, and felt that this would be a reasonable subject to discuss. It’s somewhat related to the recent posts about @selfsafae/@guarded but distinctly different regardless.
>>>>> 
>>>>> 
>>>>> Problem:
>>>>> 
>>>>> It’s often desirable not to capture self in closures, but the syntax for doing so adds significant boilerplate code for [weak self] or us unsafe when used with [unowned self]. Typically you’d do something like this:
>>>>> 
>>>>>   { [weak self] in    self?.execute() }
>>>>> 
>>>>> This is simple enough but often doesn’t work:
>>>>> 
>>>>> { [weak self] in self?.boolean = self?.calculateBoolean() ]
>>>>> 
>>>>> This fails because boolean is not an optional. This in turn leads to code like this:
>>>>> 
>>>>> { [weak self] in
>>>>>    guard let strongSelf = self else { return }
>>>>>    strongSelf.boolean = self.calculateBoolean()  }
>>>>> 
>>>>> And this is the boilerplate code. My suggestion is to add a syntax that works the same as the third syntax, yet doesn’t require the boilerplate code.
>>>>> 
>>>>> 
>>>>> Solution:
>>>>> 
>>>>> Instead of using unowned or weak, let’s use guard/guarded syntax:
>>>>> 
>>>>> 
>>>>> { [guard self] in
>>>>>    self.isExecuted = self.onlyIfWeakSelfWasCaptured()
>>>>> }
>>>>> 
>>>>> In essence, guarded self is equivalent to a weak self, that’s captured when the closure is executed. If it was already released at that point, the closure is simply not executed. It’s equivalent to:
>>>>> 
>>>>> { [weak self] in
>>>>>    guard let strongSelf = self else { return }
>>>>>    strongSelf.isExecuted = strongSelf.onlyIfWeakSelfWasCaptured()
>>>>> }
>>>>> 
>>>>> Except with a lot less boilerplate code, while not losing any clarify in what it does.
>>>>> 
>>>>> Impact / compatibility:
>>>>> 
>>>>> This is simply additive syntax, and wouldn’t affect any existing code.
>>>>> _______________________________________________
>>>>> 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/20170222/28d6273c/attachment.html>

More information about the swift-evolution mailing list