[swift-evolution] [swift-evolution-announce] [Review] SE-0089: Replace protocol<P1, P2> syntax with Any<P1, P2>

Thorsten Seitz tseitz42 at icloud.com
Sat Jun 25 11:34:27 CDT 2016

Sorry for the late reply — I had hoped to be able to think more deeply about various points, 
but I’m going to delay that instead of delaying the reply even more :-)

> Am 17.06.2016 um 19:04 schrieb Dave Abrahams <dabrahams at apple.com>:
> on Thu Jun 16 2016, Thorsten Seitz <tseitz42-AT-icloud.com> wrote:
>>> Am 13.06.2016 um 04:04 schrieb Dave Abrahams <dabrahams at apple.com>:
>>> on Fri Jun 10 2016, Thorsten Seitz <tseitz42-AT-icloud.com> wrote:
>>>>> Am 09.06.2016 um 19:50 schrieb Thorsten Seitz via swift-evolution <swift-evolution at swift.org>:
>>>>>> Am 09.06.2016 um 18:49 schrieb Dave Abrahams via swift-evolution <swift-evolution at swift.org>:
>>>>>> on Wed Jun 08 2016, Jordan Rose <swift-evolution at swift.org> wrote:
>>>>>>>> On Jun 8, 2016, at 13:16, Dave Abrahams via swift-evolution
>>>>>>>> <swift-evolution at swift.org> wrote:
>>>>>>>> on Wed Jun 08 2016, Thorsten Seitz
>>>>>>>> <swift-evolution at swift.org
>>>>>>>> <mailto:swift-evolution at swift.org>>
>>>>>>>> wrote:
>>>>>>>>> Ah, thanks, I forgot!  I still consider this a bug, though (will have
>>>>>>>>> to read up again what the reasons are for that behavior).
>>>>>>>> Yes, but in the case of the issue we're discussing, the choices are:
>>>>>>>> 1. Omit from the existential's API any protocol requirements that depend
>>>>>>>> on Self or associated types, in which case it *can't* conform to
>>>>>>>> itself because it doesn't fulfill the requirements.
>>>>>>>> 2. Erase type relationships and trap at runtime when they don't line up.
>>>>>>>> Matthew has been arguing against #2, but you can't “fix the bug” without
>>>>>>>> it.
>>>>>>> #1 has been my preference for a while as well, at least as a starting
>>>>>>> point.
>>>>>> I should point out that with the resyntaxing of existentials to
>>>>>> Any<Protocols...>, the idea that Collection's existential doesn't
>>>>>> conform to Collection becomes far less absurd than it was, so maybe this
>>>>>> is not so bad.
>>>>> I think the problem is more that Any<Collection> does not conform to
>>>>> a specific value for a type parameter T: Collection
>>>>> What I mean by this is that `Collection` denotes a type family, a
>>>>> generic parameter `T: Collection` denotes a specific (though
>>>>> unknown) member of that type family and `Any<Collection>` denotes
>>>>> the type family again, so there is really no point in writing
>>>>> Any<Collection> IMO.
>>>>> The type family cannot conform to T because T is just one fixed member of it.
>>>>> It conforms to itself, though, as I can write
>>>>> let c1: Any<Collection> = …
>>>>> let c2: Any<Collection> = c1
>>>>> That’s why I think that we could just drop Any<Collection> and simply write Collection.
>>>> Let me expand that a bit:
>>>> Actually all this talk about existentials vs. generics or protocols
>>>> vs. classes has had me confused somewhat and I think there are still
>>>> some misconceptions present on this list sometimes, so I’ll try to
>>>> clear them up:
>>> There are several objectively incorrect statements here, and several
>>> others with which I disagree.  I was hoping someone else would write
>>> this for me, but since the post has such a tone of authority I feel I
>>> must respond.
>> You are right, the tone of my post was not appropriate, for which I
>> want to apologize sincerely.
> My fundamental disagreement is with the content, not the tone.
>> I still believe my statements to be valid, though, and will respond to
>> your arguments inline. Please don't get me wrong, I'm not trying to
>> have an argument for the argument's sake. All I want is to contribute
>> maybe a tiny bit to make Swift even better than it already is, by
>> sharing ideas and thoughts not only from me but from the designs of
>> other perhaps more obscure programming languages which I happen to
>> have stumbled upon in the past (often with much delight).
> And I want you to know, even though I disagree with what you've written,
> that I very much appreciate the contribution you're making.

Thanks! I’m very glad about that!

>>>> (1) misconception: protocols with associated types are somehow very
>>>> different from generics
>>>> I don’t think they are and I will explain why. The only difference is
>>>> the way the type parameters are bound: generics use explicit parameter
>>>> lists whereas protocols use inheritance. That has some advantages
>>>> (think long parameter lists of generics) and some disadvantages.
>>>> These ways are dual in a notation sense: generic types have to have
>>>> all parameters bound whereas protocols cannot bind any of them.
>>>> The „existential“ notation `Any<>` being discussed on this list is
>>>> nothing more than adding the ability to protocols to bind the
>>>> parameters to be used just like Java’s wildcards are adding the
>>>> opposite feature to generics, namely not having to bind all
>>>> parameters.
>>> Protocols and generics fulfill completely different roles in Swift, and
>>> so, **especially in a language design context like the one we're in
>>> here**, must be thought of differently.  The former are an abstraction
>>> mechanism for APIs, and the latter a mechanism for generalizing
>>> implementations.  
>> That's not what I was talking about. Of course, protocols are a
>> mechanism for deriving types from each other whereas generics are a
>> way to parameterize types. My point was that Swift's other way to
>> parameterize types, namely by associated types, is very similar to
>> generics with wildcards when looking a the existentials of such
>> protocols. In addition I was talking about generics in general, not
>> just about generics in Swift which restricts them to implementations
>> and does not support wildcards.
> I'm aware of these other systems.  One of the problems with the way
> you're writing about this is that we're speaking in the context of Swift
> and you're assuming a completely open design space, as though Swift's
> choice to sharply distinguish classes from protocols was not a conscious
> one... but it was.  

I never had assumed that this had been decided lightly ;-)
And I have been favorably impressed by the rationales put forth so far by the Swift 
team, so it would definitely be interesting to learn a bit more about the rationale
being put into that decision and the advantages and disadvantages discussed back then.
Is there something written down somewhere? 

> Yes, Swift could have been designed differently, so
> that a single language construct, a kind of generic class, was stretched
> so it could express almost everything.  Personally, I don't believe that
> results in a better language.

I still believe it would have advantages but I’ll concede that this discussion 
will probably not help advancing Swift as this decision has been made.
Still, it might be of interest to keep in mind for further design considerations.

>> Other languages like Java offer generics for interfaces as well and
>> support wildcards (adding generic types parameters to protocols in
>> Swift is currently discussed on the mailing list as well).  FWIW my
>> arguments were not about whether we should have wildcards in Swift or
>> not, but simply to relate one parametrization feature (associated
>> types) to a more well known parametrization feature (generics with
>> wildcards) in order to understand them better.
>>> The only place you could argue that they intersect is
>>> in generic non-final classes, because a class fills the dual role of
>>> abstraction and implementation mechanism (and some might say that's a
>>> weakness).  But even accounting for generic classes, protocols with
>>> associated types are very different from generics.  Two utterly
>>> different types (an enum and a struct, for example) can conform to any
>>> given protocol P, but generic types always share a common basis
>>> implementation.  
>> The latter is not the case for generic interfaces in Java, for
>> example, so it is just an artificial restriction present in Swift.
> It's not an artificial restriction, it's a design choice.  Sure, if by

I meant artifical in the sense that a different design choice would have been possible.

> “generic type” you just mean anything that encodes a static type
> relationship, lots of things fall into that bucket.

Well, I think Java’s generics are not that advanced, so the bucket does not 
have to be very big :-)

>>> There is no way to produce distinct instances of a generic type with
>>> all its type parameters bound,
>> That is true in Swift (except for generic classes) due to the
>> restriction just mentioned.
>>> but for any protocol P I can make infinitely many instances of P with
>>> P.AssociatedType == Int.
>> This likewise applies to generic interfaces and for generic types in
>> general if taking inheritance into account - just like you do here for
>> protocols.
>>> Back to the my original point: while protocols and generic types have
>>> some similarities, the idea that they are fundamentally the same thing
>>> (I know you didn't say *exactly* that, but I think it will be read that
>>> way) would be wrong and a very unproductive way to approach language
>>> evolution.
>> I said that protocols *with associated types* are much like generics
>> *with wildcards* and tried to show why.
> If all you're trying to do is say that there's an analogy there, then we
> have no argument.


>>>> Essentially `Any<Collection>` in Swift is just the same as
>>>> `Collection<?>` in Java (assuming for comparability’s sake that
>>>> Swift’s Collection had no additional associated types; otherwise I
>>>> would just have to introduce a Collection<Element, Index> in Java).
>>> I don't see how you can use an example that requires *assuming away*
>>> assoociated types to justify an argument that protocols *with associated
>>> types* are the same as generics.
>> Note, that I said *additional* associated types, i.e. in addition to
>> .Element, even giving an example how the Java interface had to be
>> extended by a type parameter `Index` if this assumption was not
>> applied (still simplifying because Generator would have been more
>> correct which would have to be added as type parameter in addition to
>> `Index`).
>> So, in essence the comparison is between the following (I'm using Foo
>> now instead of Collection to avoid the differences mentioned. Note
>> that this has no impact on the argument at all):
>> protocol Foo {
>>    associatedtype T
>>    ...
>> }
>> interface Foo<T> {
>>    ...
>> }
> Yes, those correspond.
>> My argument is that existentials of protocols with associated types
>> are just like generic types with wildcards, i.e. `Any<Foo>` in Swift
>> is just the same as `Foo<?>` in Java.
>> Likewise `Any<Foo where .T: Number>` is just the same as `Foo<?
>> extends Number>` in Java. For me that was an insight I wanted to
>> share.
> It's a good one.


>>>> And just like Collection<?> does not conform to a type parameter `T
>>>> extends Collection<?>` because Collection<?> is the type `forall
>>>> E. Collection<E>` whereas `T extends Collection<?>` is the type
>>>> `T. Collection<T>` for a given T.
>>>> In essence protocols with associated types are like generics with
>>>> wildcards.
>>> It is true that generics with wildcards in Java *are* (not just “like”)
>>> existential types but I don't agree with the statement above.  Because
>>> Java tries to create an “everything is a class” world, generic classes
>>> with bound type parameters end up playing the role of existential type.
>>> But protocols in Swift are not, fundamentally, just existential types,
>>> and the resyntaxing of ProtocolName to Any<ProtocolName> for use in type
>>> context is a huge leap forward in making that distinction clear... when
>>> that's done (unless we leave Array<ProtocolName> around as a synonym for
>>> Array<Any<ProtocolName>>—I really hope we won't!)  protocols indeed
>>> *won't* be types at all, existential or otherwise.
>> I fully agree that protocols are not types, their existentials
>> are. But I haven't seen yet what we really *gain* from making that
>> distinction explicit (except an ugly type syntax :-).
> For me, it helps distinguish static from dynamic polymorphism.

Hmm, I’ll have to think more about that.

>> And like I already wrote in this or another thread we would have to
>> apply the same logic to non-final classes, which are existentials,
>> too.
>>>> Coming back to the questions whether (a) allowing existentials to be
>>>> used as types is useful
>>> That's the only use existentials have.  They *are* types.  Of course
>>> they're useful, and I don't think anyone was arguing otherwise.
>> I'm pretty sure that there was a discussion about whether being able
>> to write something like Any<Collection> is useful. My wording was
>> certainly imprecise, though, and didn't make sense as written. I
>> should have said something like "whether adding the ability to use
>> existential types of protocols with unbound associated types is
>> useful".
>>>> and (b) whether sacrificing type safety would somehow be necessary for
>>>> that, I think we can safely answer (a) yes, it *is* useful to be able
>>>> to use existentials like Any<Collection> as types, because wildcards
>>>> are quite often needed and very useful in Java (they haven’t been
>>>> added without a reason) (b) no, sacrificing type safety does not make
>>>> sense, as the experience with Java’s wildcards shows that this is not
>>>> needed.
>>> I would call this “interesting information,” but hardly conclusive.
>>> Java's generics are almost exactly the same thing as Objective-C
>>> lightweight generics, which are less capable and less expressive in
>>> many ways than Swift's generics.  
>> I agree that Java does not have something like `Self` or associated
>> types (which are really useful for not having to bind all type
>> parameters explicitly, especially when binding type parameters to
>> other generics which makes for long type parameter lists in Java where
>> I have to repeat everything over and over again), but do you mean
>> something else here?
>> Especially in the context of sacrificing type safety?
> I do, but it will take some research for me to recover my memory of
> where the holes are.  It has been years since I thought about Java
> generics.  It's also possible that I'm wrong ;-)

If you happen to remember, I’d be interested in hearing about the problems you meant.

>>>> Especially if something like path dependent types is used like
>>>> proposed and some notation to open an existential’s type is added,
>>>> which is both something that Java does not have.
>>>> (2) misconception: POP is different from OOP
>>>> It is not. Protocols are just interfaces using subtyping like OOP has
>>>> always done. They just use associated types instead of explicit type
>>>> parameters for generics (see above).
>>> They are not the same thing at all (see above ;->).  To add to the list
>>> above, protocols can express fundamental relationships—like Self
>>> requirements—that OOP simply can't handle.
>> Eiffel has something like Self, it is called anchoring and allows
>> binding the type of a variable to that of another one or self (which
>> is called `Current` in Eiffel). And Eiffel does model everything with
>> classes which may be abstract and allow for real multiple inheritance
>> with abilities to resolve all conflicts including those concerning
>> state (which is what other languages introduce interfaces for to avoid
>> conflicts concerning state while still failing to solve *semantic*
>> conflicts with the same diamond pattern).
>> No protocols or interfaces needed. Why do you say this is not OOP? The
>> book which describes Eiffel is called "Object-Oriented Software
>> Construction" (and is now about 20 years old).
> It's not *incompatible* with OOP, but it is not part of the essence of
> OOP either.  If you survey object-oriented languages, what you find in
> common is inheritance-based dynamic polymorphism and reference
> semantics.  Those are the defining characteristics of OOP, and taking an
> object-oriented approach to a given problem means reaching for those
> features.

Agreed, it is not part of most OOP *implementations* while being compatible with OOP.
There have been lots of papers and research languages about typing problems like 
binary methods, null pointers etc., though, so taking the mainstream OO languages
as the yardstick for OOP is jumping a little bit too short IMO. 

>>> There's a reason Java can't
>>> express Comparable without losing static type-safety.  
>> You are certainly right that Java is not the best language out there
>> especially when talking about type systems (I often enough rant about
>> it :-) but I'm not sure what you mean here. Java's Comparable<T> seems
>> quite typesafe to me. Or do you mean that one could write `class A
>> implements Comparable<B>` by mistake? That's certainly a weak point
>> but doesn't compromise type safety, does it?
> Java has cleverly avoided compromising type safety here by failing to
> express the constraint that comparable conformance means a type can be
> compared to itself ;-)

Indeed :-)

>> Ceylon has an elegant solution for that without using Self types:
>> interface Comparable<in Other> of Other given Other satisfies Comparable<Other> {...}
>> Note the variance annotation (which Swift currently has not) and the
>> `of` which ensures that the only subtype of Comparable<T> is T. This
>> is a nice feature that I haven't seen often in programming languages
>> (only Cecil comes to mind IIRC) and which is used for enumerations as
>> well in Ceylon. In Swift I cannot do this but can use Self which
>> solves this problem differently, albeit with some drawbacks compared
>> to Ceylon's solution (having to redefine the compare method in all
>> subtypes, 
> That sounds interesting but is a bit vague.  A concise example of how
> this plays out in Swift and in Ceylon would be instructive here.

Sorry, the difficulty with Self I was thinking of only occurs when Self is in a covariant position
which is not the case in Comparable, of course. Let’s take a modified example instead with Self 
in a covariant position:


protocol Minimizable {
    func min(from other: Self) -> Self

final class A : Minimizable { // has to be final
    let x: Int
    init(x: Int) {
        self.x = x
    func min(from other: A) -> A {
        return x < other.x ? self : other


interface Minimizable<Other> of Other given Other satisfies Minimizable<Other> {
    shared formal Other min(Other other);

class A() satisfies Minimizable<A> {

    Integer x = 0;

    shared actual default A min(A other) {
        if (x < other.x) {
            return this;
        } else {
            return other;

In Ceylon class A does not have to be final and choosing the minimum of two values would be available for values from the whole subtree of types rooted in A (the `of` ensures that such a declaration cannot „cross“ into other subtrees) whereas `Self` enforces that there is no subtree below class A.

I have to admit that I am not well versed using `Self`, yet, so maybe I’m wrong here. In addition I am sure that `Self` allows designs
that are not possible with Ceylon’s `of`.

The usage of Ceylon’s `of` for enumeration types is as follows (example taken from http://ceylon-lang.org/documentation/tour/types/):

abstract class Node() of Leaf | Branch {}

class Leaf(shared Object element) 
        extends Node() {}

class Branch(shared Node left, shared Node right) 
        extends Node() {}
void printTree(Node node) {
    switch (node)
    case (is Leaf) {
        print("Found a leaf: ``node.element``!");
    case (is Branch) {

>> which has lead to lengthy discussion threads about Self, StaticSelf,
>> #Self etc.).
>>> Finally, in a
>>> language with first-class value types, taking a protocol-oriented
>>> approach to abstraction leads to *fundamentally* different designs from
>>> what you get using OOP.
>> Eiffel has expanded types which are value types with copy semantics
>> quite like structs in Swift. These expanded types are pretty much
>> integrated into Eiffel's class-only type system. Just define a class
>> as `expanded` and you are done. 
> Unless this part of the language has changed since 1996, or unless I've
> misread https://www.cs.kent.ac.uk/pubs/1996/798/content.pdf, you can't
> make an efficient array with value semantics in Eiffel.  That, IMO,
> cannot be considered a language with first-class value types.

I haven’t had time yet to really evaluate that paper, but if you are right, then I agree
with you that Eiffel cannot be considered as having first-calss value types.

At least one of the deficiencies listed in the paper does not exist anymore AFAIU 
(expanded types not having constructors), though, so maybe things actually do have 
changed since then.

>> Eiffel seems to have no need to introduce interfaces or protocols to
>> the language to support value types.  
> No, of course not.  By saying that everything from abstract interfaces
> to static constraints and even value types is to be expressed a kind of
> possibly-generic class, you can eliminate distinctions in the language
> that IMO help to clarify design intent.  This is a language design
> choice one could make, but not one I'd want to.  In LISP, everything is
> an S-expression.  That has certain upsides, but for me it fails the
> expressivity test.

That’s certainly a valid point.

Furthermore I do understand (and fully support) that being interoperable with Objective-C is an 
important restriction on Swift’s design space and I think it is absolutely awesome how
that has been achieved!

>> You can even derive from expanded classes which is currently not
>> possible in Swift but has already been discussed several times on this
>> mailing list.  Polymorphic usage is only possible for non expanded
>> super types, which means as far as I understood that a reference is
>> used in that case. Variables with an expanded type do not use refences
>> and therefore may not be used polymorphically in Eiffel.  This should
>> be similar in Swift, at least as far as I did understand it. The
>> question whether variables with a value type can be used
>> polymorphically currently does not arise in Swift as structs cannot
>> inherit from each other (yet?).
>>>> The more important distinction of Swift is emphasizing value types and
>>>> making mutation safely available by enforcing copy semantics for value
>>>> types.  
>>> We don't, in fact, enforce copy semantics for value types.  That's
>>> something I'd like to change.  But regardless, value types would be a
>>> *lot* less useful if they couldn't conform to protocols, and so they
>>> would be a lot less used.  Heck, before we got protocol extensions in
>>> Swift 2, there was basically *no way* to share implementation among
>>> value types.  So you can't take protocols out of the picture without
>>> making value types, and the argument for value semantics, far weaker.
>> Why? Like I said, Eiffel *has* value types without needing
>> protocols. They just have a unified mechanism built around classes.
> Because I'm speaking about Swift, not some other world where Protocol ==
> Generic Class ;-)

Ah, ok, I took your statement of protocols being needed for strong value semantics
to be of general validity, not confined to Swift :-) 
Within Swift that is certainly true!


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-evolution/attachments/20160625/2f3bd705/attachment.html>

More information about the swift-evolution mailing list