[swift-evolution] [Pitch] Improving unspecified generic usability

Tue Aug 8 04:14:44 CDT 2017

To my understanding, the following is exactly as it should be:

FooStruct<String> as? FooStruct<Any> // Compiles but conversion fails, becomes nil, and that's normal

The reason for this is that FooStruct<String> is not a subtype of FooStruct<Any> (or just FooStruct), while String is of course a subtype of Any, because generic types are not covariant in Swift, and that's the way it should be for a sound static type system. My comments on this were related to what you wrote about  arrays.

In theory a protocol without associated types could be synthesized for all the non generic properties and methods of a generic type, with the ability of casting to it and possibly from it.

It's a useful idea, and I'm all for it (I think literally everyone that uses generics and protocols with associated types encountered these kinds of problems at least once), I'm just saying that I'd rather work on generalized existentials which have already been considered by the core team, at least from a theoretical standpoint, have a greater scope and include cases like this. In general I tend to prefer broader, more generic solutions rooted in type theory when dealing with generics and protocols, but that's just me. 

Elviro

> Il giorno 08 ago 2017, alle ore 10:44, Logan Shire via swift-evolution <swift-evolution at swift.org> ha scritto:
> 
> Thanks for the feedback!
> 
> Félix, sorry about the confusion between FooStruct and FooProtocol - I'll refer to them as such moving forwards.
> 
> David, I don't believe you should be able to cast an [FooStruct<String>] to an [FooStruct<Any>] because those are both valid specifications. If Generalized Existentials <https://github.com/apple/swift/blob/master/docs/GenericsManifesto.md#generalized-existentials> are implemented, that would be another story, but that's outside the scope of this proposal. I do believe you should be able to cast [FooStruct<String>] to [FooStruct], and that you should be able to flatMap [FooStruct] into [FooStruct<Any>] with as?, but all of the casts would fail and you would be left with an empty array.
> 
> In regards to the Named protocol, yes, that is the current idiomatic approach to solving this problem (along with making a function unnecessarily generic and then using the generic type as a constraint). I want to avoid jumping through those hoops. We'd essentially be synthesizing the Named protocol with the same name as the non-generic version of the type. I.e. FooStruct<T>: FooStruct
> 
> Félix, I believe the above answers some of your questions, but in regards to protocols with associated types, I'd imagine it would work the same way. If FooProtocol has an associated type T, there would be another protocol, FooProtocol, without the associated type. (behind the scenes its garbled name would be different)
> 
> Also, an aside, it would be nice if protocols could use the generic syntax for their associated type constraints. I.e. "FooProtocol with T = Int" could be expressed as FooProtocol<T: Int>. It feels strange that we have two different syntaxes for essentially the same language construct. At the very least, I want some way to cast a value to a protocol type with an associated value. E.g. "if let grassEater = any as? Animal where Food = Grass"
> 
> Elviro, yes, the generalized existentials would help a lot here, but that's outside the scope of what I'm proposing. In the near term I'd like to be able to use a generic type's non-generic interface, casting to and from it. See the above discussion regarding the Named protocol. Essentially we'd be synthesizing the Named protocol, but where the type's name is the same as the non-generic version of the type name.
> 
> FooStruct<String> as FooStruct // works
> FooStruct as? FooStruct<String> // works
> FooStruct as? FooStruct<Any> // Compiles but conversion fails, becomes nil
> FooStruct<String> as? FooStruct<Any> // Compiles but conversion fails, becomes nil
> 
> Let me know if you have any other questions!
> 
> Logan
> 
> 
> On Tue, Aug 8, 2017 at 9:43 AM Félix Cloutier <felixcloutier at icloud.com <mailto:felixcloutier at icloud.com>> wrote:
> I'm going to separate your examples into FooStruct and FooProtocol for clarity.
> 
> I agree that generics tend to propagate virally and I remember that at some point I wanted type erasure, though I don't remember for what exactly. The solution for `sayHi`, right now, is to make that one generic too:
> 
>> func sayHi<T>(to foo: T) where T: FooProtocol {
>>     print("hi \(foo.name <http://foo.name/>)")
>> }
> 
> The "let foos: [FooStruct] = [FooStruct(name: "Int", value: 2), FooStruct(name: "Double", value: 2.0)]" part can't work for structs because arrays require each element to have the same size (but it could work for classes).
> 
> Even then, you couldn't infer the type to [FooClass<Any>] because contravariance isn't permissible in that situation: doing so would allow you to assign any Any to a FooClass's value.
> 
> Another problem that this would have to solve is that once you lose the associatedtype that came with the protocol, there is nothing you can do to recover it; you currently can't express "FooProtocol with T = Int" as a type that you can cast to, so you would only be able to pass the instance to functions that don't have constraints on T.
> 
> But all in all, with my current understanding of the issue, I think that I'm favorable to the idea.
> 
> Félix
> 
> 
>> Le 7 août 2017 à 19:35, David Sweeris via swift-evolution <swift-evolution at swift.org <mailto:swift-evolution at swift.org>> a écrit :
>> 
> 
>>> 
>>> On Aug 7, 2017, at 3:00 PM, Logan Shire via swift-evolution <swift-evolution at swift.org <mailto:swift-evolution at swift.org>> wrote:
>>> 
>>> One of my longstanding frustrations with generic types and protocols has been how hard it is to work with them when their type is unspecified.
>>> Often I find myself wishing that I could write a function that takes a generic type or protocol as a parameter, but doesn’t care what its generic type is.
>>> 
>>> For example, if I have a type:
>>> 
>>> struct Foo<T> {
>>>     let name: String
>>>     let value: T
>>> }
>>> 
>>> or:
>>> 
>>> protocol Foo {
>>>     associatedtype T
>>>     var name: String { get }
>>>     var value: T { get }
>>> }
>>> 
>>> And I want to write a function that only cares about Foo.name, I’d like to be able to:
>>> 
>>> func sayHi(to foo: Foo) {
>>>     print("hi \(foo.name <http://foo.name/>)")
>>> }
>>> 
>>> But instead I get the error, “Reference to generic type Foo requires arguments in <…>”
>>> 
>>> Also, when you want to have a polymorphic array of generic types, you can’t:
>>> 
>>> let foos: [Foo] = [Foo(name: "Int", value: 2), Foo(name: "Double", value: 2.0)]
>>> 
>>> And if you remove the explicit type coercion, you just get [Any]
>>> 
>>> let foos = [Foo(name: "Int", value: 2), Foo(name: "Double", value: 2.0)]
>>> 
>>> I wish that could be inferred to be [Foo].
>> 
>> What happens if you try to say "foos: [Foo<Any>] = ..."? 
>> 
>> 
>> 
>>> I’d like to propose being able to use the non-generic interface of a type normally. 
>>> I.e. if you have a type Foo<T>, it is implicitly of type Foo as well. The type Foo could be used like any other type.
>>> It could be a parameter in a function, a variable, or even the generic type of another type (like a Dictionary<String, Foo>)
>>> 
>>> The only restriction is that if you want to call or access, directly or indirectly, a function or member that requires the generic type,
>>> the generic type would have to be known at that point.
>>> 
>>> Foo<T> should be able to be implicitly casted to Foo wherever you want, and Foo could be cast to Foo<T> conditionally.
>>> Initializers would still obviously have to know the generic type, but given the above example, you should be able to:
>>> 
>>> let names = foos.map { $0.name }
>>> 
>>> However, you could not do the following:
>>> 
>>> let foos = [Foo]()
>>> 
>>> Because the initializer would need to know the generic type in order to allocate the memory.
>>> 
>>> Let me know what you think!
>> 
>> 
>> The idiomatic solution would be to create a `Named` protocol with a `var name: String {get}` property, and write your function like `func sayHi(to foo:Named) {...}`. However, this `Named`protocol is really pretty trivial -- its purpose is simply to "degenericify" a generic type, not to provide any semantic meaning. Perhaps an analogy could be drawn between such "trivial protocols" and how we sometimes view tuples as "trivial structs"? Dunno, maybe I'm just trying to turn two trees into a forest, but this kinda smells like it might be part of a bigger issue, and if it is I'd rather tackle that and then see if we still need to address anything here.
>> 
>> +1, either way, though.
>> 
>> - Dave Sweeris
> 
>> _______________________________________________
>> 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
> https://lists.swift.org/mailman/listinfo/swift-evolution

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-evolution/attachments/20170808/2474ec80/attachment.html>