[swift-evolution] Pitch: Automatically deriving Equatable/Hashable for more value types

Fri May 5 14:16:09 CDT 2017

On Fri, May 5, 2017 at 11:51 AM Matthew Johnson <matthew at anandabits.com>
wrote:

> On May 5, 2017, at 1:33 PM, Tony Allevato <tony.allevato at gmail.com> wrote:
>
>
>
> On Fri, May 5, 2017 at 11:07 AM Matthew Johnson <matthew at anandabits.com>
> wrote:
>
>> On May 5, 2017, at 10:45 AM, Tony Allevato via swift-evolution <
>> swift-evolution at swift.org> wrote:
>>
>> Thanks for your feedback, everybody!
>>
>>
>> Thanks for continuing to drive this forward!
>>
>>
>> I've updated the gist
>> <https://gist.github.com/allevato/2fd10290bfa84accfbe977d8ac07daad> to
>> reflect what seems to be a consensus here:
>>
>> * Derived conformances are now opt-in (this makes the recursive case
>> *much* cleaner, and the complexity involved in that section has been
>> completely removed)
>>
>>
>> Can the opt-in conformance be declared in an extension?  If so, can the
>> extension be in a different module than the original declaration?  If so,
>> do you intend any restrictions, such as requiring all members of the type
>> declared in a different module to be public?  My initial thought is that
>> this should be possible as long as all members are visible.
>>
>
> Declaring the conformance in an extension in the same module should
> definitely be allowed; I believe this would currently be the only way to
> support conditional conformances (such as the `Optional: Hashable where
> Wrapped: Hashable` example in the updated draft), without requiring deeper
> syntactic changes.
>
> I'm less sure about conformances being added in other modules, but I'm
> inclined to agree with your assessment. I could see two ways of
> interpreting it:
>
> * E/H can only be derived in an extension in an external module if all the
> members are accessible (and the other conditions are met).
> * E/H can be derived in an extension in an external module using only the
> subset of accessible members (if the other conditions are met).
>
> These are subtly different. The argument for the first would be "if you
> want to add E/H to a type in a different module, you must *consciously*
> decide which members you want to use in those computations". The argument
> for the second would be "you can already make a type in a different module
> conform to E/H and you'd be restricted to the accessible members there, so
> let's make that path easier for users too."
>
> The first case is probably the safer choice. I'm not sure about the
> implementation difficulty of each.
>
>
>> * Classes are supported now as well
>>
>> Please take a look at the updated version and let me know if there are
>> any concerns! If folks like it, I'll prepare a pull request.
>>
>>
>> Will the synthesis for classes dispatch through a non-final method which
>> is expected to be overridden by subclasses?  You don’t explicitly state
>> this but it seems implied.  If so, what if  the subclass requires a custom
>> implementation?  This would require the signature of the non-final method
>> to be part of the synthesis contract.
>>
>
>> Supporting non-final classes introduces enough complexity (especially
>> when multiple modules are involved).  I would hate to see it get
>> sidetracked in discussions regarding non-final classes and miss the Swift 4
>> window because of that.  Given the limited time left for Swift 4 it might
>> be better to keep the initial proposal simpler and consider a followup in
>> the Swift 5 timeframe to build on the initial proposal.
>>
>
> For ==, the operator must already be "class final" or "static" regardless
> of this proposal, and it can't be "overridden" as such in subclasses
> because the arguments would be different (lhs and rhs would be the
> subclass, not the superclass). So the compiler should be able to generate
> the correct implementation for subclasses in all cases, right?
>
>
> This won’t work because Equatable has a `Self` requirement so the `==`
> defined by the initial conforming class would be called.  In order to
> support non-final classes you would need to have that dispatch through
> something like an `isEqual` method which *can* be overridden.
>

Ah crap, you're right. I was basing my experimentation on this example:

class Base: Equatable {
  let x: Int
  init(x: Int) { self.x = x }
  static func == (lhs: Base, rhs: Base) -> Bool { return lhs.x == rhs.x }
}

class Sub: Base {
  let y: Int
  init(x: Int, y: Int) { self.y = y; super.init(x: x) }

  static func == (lhs: Sub, rhs: Sub) -> Bool {
    guard lhs as Base == rhs as Base else {
      return false
    }
    return lhs.y == rhs.y
  }
}

let s1 = Sub(x: 1, y: 2)
let s2 = Sub(x: 1, y: 3)
let s3 = Sub(x: 1, y: 2)

func usesBase(_ lhs: Base, _ rhs: Base) -> Bool { return lhs == rhs }
func usesSub(_ lhs: Sub, _ rhs: Sub) -> Bool { return lhs == rhs }

print(usesSub(s1, s2)) // false (expected)
print(usesSub(s1, s3)) // true

print(usesBase(s1, s2)) // true (problematic)
print(usesBase(s1, s3)) // true

In that case the problematic behavior is "expected", but when you make it
generic, you're right that you still get the problematic behavior which
kind of makes the conformance useless:

func usesGeneric<T: Equatable>(_ lhs: T, _ rhs: T) -> Bool {
  return lhs == rhs
}

print(usesGeneric(s1, s2)) // true (problematic)
print(usesGeneric(s1, s3)) // true

Sadly, I don't work with classes/inheritance enough in Swift to run into
these problems on a regular basis.

FWIW, the same issue came up in SE-0091
<https://github.com/apple/swift-evolution/blob/master/proposals/0091-improving-operators-in-protocols.md>
and
the consensus seemed to be that since Equatable already didn't work the
"right way" for inheritance anyway, that would need to be fixed separately,
so I'd be interested to know if the core team feels the same way here.

>
>
> For hashValue, I think the possibilities are:
>
> * Sub is a subclass of Super. Super conforms to Hashable and implements
> non-final hashValue. The compiler can derive it for Sub and call
> super.hashValue in its implementation.
>
>
> Yes, this makes sense.  The primary difficulty with Hashable is that it
> refines Equatable.  Refining a non-final implementation of `hashValue` is
> relatively straightforward.
>

Right... there could be situations where a function that takes a generic
<T: Hashable> gets inconsistent results for == and hashValue because the
latter is dynamically dispatched but the former isn't. That's already the
situation we have today, but it could be argued that deriving
implementations would "make the bleeding worse".

>
> * Sub is a subclass of Super. Super conforms to Hashable and implements a
> final hashValue. The compiler cannot derive one for Super and would
> silently not do so.
>
>
> Do you mean “the compiler cannot derive one for Sub”?
>

Yes, sorry.

>
> * Sub is a subclass of Super. Super does not conform to Hashable, but Sub
> asks to derive it. This can either (1) not be allowed, telling the user
> that they need to write it manually in this case, or (2) be allowed and use
> all accessible members to compute the hashValue (including those from the
> superclass).
>
> What do Encodable/Decodable do in these situations? It seems similar
> solutions there would apply here.
>
>
> That’s a good question.  I don’t recall whether this was addressed
> explicitly or not.
>
>
> But after writing this all out, I'm inclined to agree that I'd rather see
> structs/enums make it into Swift 4 even if it meant pushing classes to
> Swift 4+x.
>
>
> That is reasonable.
>
> On the other hand, I think you could come up with straightforward
> semantics for synthesizing conformance for final classes as well.  Final
> classes with no superclass should be straightforward.  Final classes that
> do have a superclass would be similarly straightforward if we decide to
> allow this as described in option (2) above regarding hashValue.
>
> I’m on the fence on this - if we can include final classes using option
> (2) without jeopardizing getting this in for Swift 4 I would support that.
> If it’s going to put support for value types in Swift 4 at risk then I
> would not.
>

I'm on the fence as well. It sounds like we could scratch non-final classes
off the list for now without losing much because they're already very
problematic. The question then is, how much worse does supporting final
classes make the implementation.

Since I only did the work for enums in my initial experiment, I don't have
enough insight yet to know for sure. Hopefully I'll have some time in the
next couple days to try it out.

>
>
>
>>
>>
>>
>> On Fri, May 5, 2017 at 8:16 AM Nevin Brackett-Rozinsky via
>> swift-evolution <swift-evolution at swift.org> wrote:
>>
>>> On Fri, May 5, 2017 at 1:47 AM, Xiaodi Wu via swift-evolution <
>>> swift-evolution at swift.org> wrote:
>>>
>>>> On Fri, May 5, 2017 at 12:41 AM, Brent Royal-Gordon <
>>>> brent at architechies.com> wrote:
>>>>
>>>>> I would think only final classes could participate in this, since a
>>>>> subclassable class would need to allow subclasses to override equality, and
>>>>> you can't override a static `==` operator method.
>>>>>
>>>>
>>>> I work so rarely with classes that I'm embarrassed to have to ask this
>>>> question: can classes not satisfy Equatable with a `public class func ==`?
>>>>
>>>
>>> Currently:
>>>
>>> class C: Equatable {
>>>     class func == (lhs: C, rhs: C) -> Bool {
>>>         return lhs === rhs
>>>     }
>>> }
>>>
>>> Yields an error, “Operator '==' declared in non-final class 'C' must be
>>> 'final'”.
>>>
>>> Nevin
>>> _______________________________________________
>>> swift-evolution mailing list
>>> swift-evolution at swift.org
>>> https://lists.swift.org/mailman/listinfo/swift-evolution
>>>
>> _______________________________________________
>> swift-evolution mailing list
>> swift-evolution at swift.org
>> https://lists.swift.org/mailman/listinfo/swift-evolution
>>
>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-evolution/attachments/20170505/878540bb/attachment.html>