Is there any mechanism to mark a property as not participating in derived conformances? One instance might be that I have a memoization/cache related property that is stored but should not be considered when equating two instances.
TJ On Tue, May 16, 2017 at 3:51 AM, Xiaodi Wu via swift-evolution < swift-evolution@swift.org> wrote: > On Mon, May 15, 2017 at 7:18 PM, Tony Allevato <tony.allev...@gmail.com> > wrote: > >> >> >> On Mon, May 15, 2017 at 4:38 PM Itai Ferber <ifer...@apple.com> wrote: >> >>> On May 15, 2017, at 4:03 PM, Xiaodi Wu via swift-evolution < >>> swift-evolution@swift.org> wrote: >>> >>> This is nice. Thanks for taking the time to write it up. I do have some >>> concerns/questions: >>> >>> Do the rules you spell out align with those for Codable? I think it is >>> very important that these are paralleled as closely as possible, and that >>> any deviations are explicitly called out in the text with reasoning as to >>> why it must deviate. Knowing when something is synthesized is difficult >>> enough with one set of rules--two is certainly one too many. >>> >>> To spell out the rules of Codable conformance clearly, for reference: >>> >>> For example, is it permitted to extend a type in the same module in >>> order to obtain synthesized Codable conformance? How about for a type in a >>> different module? The same rules should then apply for Equatable and >>> Hashable synthesis. >>> >>> Yes, Codable conformance can be added in an extension both intra-module, >>> and inter-module (i.e. you can add Codable conformance via extensions in >>> your own module, or to types in other modules). If there is a situation >>> where this is not possible, that’s likely a bug. >>> [For reference, it is actually easier to allow this than to prevent it. >>> I had to do very little extra work to support this because of how this is >>> organized in the compiler.] >>> >> >> To the best of my knowledge, the Equatable/Hashable synthesis I added >> uses the same rules as Codable, since I based my implementation on it. >> >> This is slightly different than what we initially discussed in this >> thread, which was that we should not support synthesized conformance in >> extensions in other modules. But after implementing it, my feeling is that >> if it falls out naturally and prohibiting it would be more work, then we >> shouldn't do that unless we have good reason to, and we should do it >> consistently with other derivations. So I'm using the same model. >> >> >> >>> Furthermore, does Codable ignore computed properties? If not, then >>> neither should Equatable and Hashable. >>> >>> Yes. Derived conformance for Codable ignores all computed properties >>> (including lazy properties and their associated storage). This is also some >>> relatively easy default behavior; you can iterate all properties matching >>> this requirement via `NominalTypeDecl.getStoredProperties` >>> (getStoredProperties(/*skipInaccessible=*/true) will skip the storage >>> associated with lazy vars). >>> [The thought process here is that accessing computed vars (and more so >>> lazy vars) will generally have side effects. We don’t want to trigger side >>> effects on encoding/checking for equality/hashing, and in general, those >>> types of properties will not affect equality/hash value/encoded >>> representation.] >>> >> >> Yes, I'm using the same getStoredProperties call to find the struct >> members to apply it to (thanks Itai for the early pointers!), so Eq/Hash >> should be synthesized for structs under the same conditions as Codable. >> >> >>> >>> There are also some complicated rules with generics, if I recall, that >>> may force something to be a computed property. It would be worth exploring >>> if such rules make ignoring computed properties counterintuitive. For >>> instance, if a user has to redesign the type, changing a stored property to >>> a computed property just to satisfy certain rules of the language, and all >>> of a sudden the definition of equality has silently changed as a >>> consequence, then that could end up being very hard to debug. If we find >>> that this is a plausible issue, then it might be worth considering refusing >>> to synthesize Equatable conformance for a type with any computed >>> properties--obviously limiting, but better limiting than surprising. To be >>> clear, I'm not suggesting that we do make this limitation, just that I >>> don't know that the consequences have been adequately explored for not >>> including computed properties. >>> >>> I’m not sure about this — someone else will have to weigh in. I don’t >>> think I’ve ever encountered a situation like this while working on Codable. >>> That being said, if there’s a limiting factor here that we encounter, we >>> should absolutely be consistent between all implementations of derived >>> conformance. >>> >> >> The concern that changing a stored property to a computed property would >> silently change the behavior of Eq/Hash is definitely something we should >> be aware of and we should see if it's something that people run into >> frequently once they start using the synthesis. Nothing obvious comes to >> mind as a way of preventing or warning about it, though—I'd have to think >> more on it. >> >> >>> It would be helpful to document these rules somewhere, so noted. >>> >> >> +1. >> > > Highly agree with all your responses; also, delighted to hear that the > implementation work has fallen into place so naturally. > > > > _______________________________________________ > swift-evolution mailing list > swift-evolution@swift.org > https://lists.swift.org/mailman/listinfo/swift-evolution > >
_______________________________________________ swift-evolution mailing list swift-evolution@swift.org https://lists.swift.org/mailman/listinfo/swift-evolution