[swift-evolution] [Proposal] Sealed classes by default

[Jean-Daniel Dupas]

> Le 14 août 2016 à 20:43, Karl via swift-evolution <swift-evolution at swift.org> a écrit :
> 
> 
>> On 14 Aug 2016, at 11:17, John Holdsworth via swift-evolution <swift-evolution at swift.org> wrote:
>> 
>> Hi folks,
>> 
>> I see from building the latest Swift-3.0 branch that I’m a little behind the times and this proposal has been accepted :(
>> 
>> https://github.com/apple/swift-evolution/blob/master/proposals/0117-non-public-subclassable-by-default.md
>> 
>> Is there no way we can revisit this decision? 60 emails on a mail group not read by the whole community
>> doesn’t quite seem enough validation for such a significant change to me. Of those that expressed an
>> opinion on the thread many were against sealing classes outside the module. 
>> 
>> https://lists.swift.org/pipermail/swift-evolution/Week-of-Mon-20160627/022364.html
>> 
>> I personally feel it's a huge mistake for the following reasons:
>> 
>> 1) it “breaks open source” reducing the reusability of third party software by default.
>> 
>> 2) Contrary to arguments about the likely performance benefits of de-virtualisation of method dispatch it is
>> likely to make Swift programs launch more slowly due to the dynamic linker having to slide large numbers
>> of function pointers for most method calls (itself an expensive operation) whether a call is even made.
>> 
>> On the first point it's an established technique in OOP to be able to subclass and override to adapt a class
>> for a role perhaps it’s author never considered. Call this “monkey-patching” if you like but it is a part of being
>> able to use open source libraries without having to make a fork as would be the case if sealed by default.
>> Are we expecting developers to remember to leave their classes (and methods?) "open”? If they do feel
>> that others shouldn’t subclass or override there was always “final". Distinctions about whether this is possible
>> inside or outside a module seem a mute point to me.
>> 
>> The second point is more of an engineering one. “Virtual” dispatch on Swift is not a significant overhead
>> compared to fully dynamic dispatch on Objective-C which even then was seldom a problem. There is a
>> cost however to de-virtualisation and resolving to a direct call to a method address in that the dynamic
>> linker has to resolve and slide the pointer on load. This type of concern is a bad argument to use in 
>> designing a language anyway even if it is called Swift.
>> 
>> Well, I know the boat has left on this one and I’ll likely have to live with it but this is the proposal I
>> most disagree with on evolution this far. Programmers have morals and just because you shouldn’t
>> do something doesn’t mean the language needs to turn that into can’t by default when the need arises.
>> If this change does go ahead please try to get it into the next Xcode beta as it is a very incompatible one.
>> 
>> Cheers,
>> 
>> John
>> 
>> 
>> _______________________________________________
>> swift-evolution mailing list
>> swift-evolution at swift.org
>> https://lists.swift.org/mailman/listinfo/swift-evolution
> 
> It we are happy with non-open != final, I would be in favour of some kind of monkey-patching import attribute, similar to @testable, which allows you to subclass non-open classes at your own risk.
> 

The non subclassable across module boundary is not just a compiler enforced limitation. Once you compile a module with classes that are ‘final’ in the module, the compiler may devirtualize all call sites in the module, or even inlining some calls, making subclassing unpredictable and pointless.