[swift-evolution] Renaming for Protocol Conformance

[Xiaodi Wu]

On Wed, Aug 24, 2016 at 3:39 AM, Jonathan Hull <jhull at gbis.com> wrote:

>
> On Aug 23, 2016, at 8:35 AM, Xiaodi Wu <xiaodi.wu at gmail.com> wrote:
>
> On Tue, Aug 23, 2016 at 3:02 AM, Jonathan Hull <jhull at gbis.com> wrote:
>
>>
>> On Aug 22, 2016, at 11:32 PM, Xiaodi Wu <xiaodi.wu at gmail.com> wrote:
>>
>> On Mon, Aug 22, 2016 at 11:59 PM, Jonathan Hull via swift-evolution <
>> swift-evolution at swift.org> wrote:
>>
>>> Hi everyone,
>>>
>>> We talked about this before when we were discussing mixins, and there
>>> seemed to be generally positive feelings towards it as a feature for the
>>> future.
>>
>>
>> It's been some time now since the original discussion, so perhaps you
>> could refresh our collective memory (or at least, mine): although it
>> *seems* like this feature might be useful, I can't recall a concrete use
>> case where I've felt like I needed this feature--do you have some examples?
>>
>>
>> Ideally, the biggest use is that it helps to (partially) solve the
>> diamond problem (and similar issues) by forcing/allowing disambiguation
>> when there are multiple protocols being conformed to.  This will become
>> more of an issue if we allow protocols or extensions to add storage.  Your
>> proposed syntax actually does a better job of it than mine because mine was
>> always shown as attached to some sort of implementation, whereas yours
>> could potentially allow access to a default implementation under a new name.
>>
>> Other than that, it generally allows us to bypass/mitigate conflicts
>> between protocols.  In the current version, you are unable to conform to
>> both protocols (either because it won’t compile or because you can’t
>> satisfy the semantics of both protocols) without designing the protocols
>> together to avoid conflicts.  (I have definitely had to go back and
>> rename/refactor properties on a protocol for this reason… which I couldn’t
>> have done if I didn’t control both protocols).
>>
>
> I understand something of the difficulty of confronting the diamond
> problem. As I wrote above, I'm inclined to believe that this proposed
> feature would help solve a real issue. However, the point I'm trying to
> make is that, on reflection, I have never actually been hampered by the
> lack of this feature, and so I'd like to continue the discussion to get a
> fuller sense of just how impactful this proposal would be, both positive
> and negative.
>
> It's true, of course, that if you control at least one of two protocols
> (you don't need to control both protocols), it is trivially easy to cause
> this problem to occur, but as you point out it is also possible to resolve
> the problem by re-designing the protocol you control. I'm inclined to think
> (without evidence, admittedly) that re-designing to remove the conflict,
> where possible, would actually be the superior option in most cases.
>
> My question was: have you actually run into a scenario that necessitates
> the feature you propose because you controlled neither conflicting
> protocol? I think it would strengthen the proposal greatly to have a
> concrete, uncontrived example.
>
>
> Right now I commonly have to hand-namespace protocol methods/properties to
> avoid conflicts.  So instead of having ‘var image:UIImage’ (which is the
> name which makes the most sense in the protocol’s context), I have ‘var
> protocolNameImage:UIImage’.  There are lots of things which have common
> properties like ‘count’ which have to be called ‘somethingCount’ or
> ‘countOfSomething’.  In the context of the protocol, these names are full
> of redundant words (especially when measured against the new naming
> guidelines).  We are all used to doing this for Objective C, but it feels
> out of place in Swift.
>
> This will become a much more serious issue as the third-party code
> ecosystem grows.  Without some capability like this, you will have
> frameworks which can’t be used together (or at least with the same
> object).  I would hate to see a ‘best practice’ emerge of adding 3 letter
> prefixes to all protocol methods to get around compatibility issues.
>

Ah, well this isn't exactly the diamond problem you're talking about here.
Instead, I think, we have a fundamental disagreement. I think I've been
told that this opinion of mine is 'insane'--but I hold to it:

Protocols are not merely a vehicle for delivering a reusable bag of code.
One of its most essential purposes is to constrain the shape or API of its
conforming types. Therefore, it is a feature, not a bug, that with every
choice of name in a protocol you foreclose the possibility of composing
that protocol with others that might have colliding names.

Currently, if you the protocol vendor have made the decision that `image`
"makes the most sense in the protocol's context", you must have considered
whether it would be absurd for a conforming type to have another use for
`image`. If it would be absurd, then `image` is the appropriate name for
your protocol requirement and any other word would truly be redundant. But,
if this is only one of many plausible images, then `somethingImage` or
`imageOfSomething` *is* the appropriate name, and trying to shorten the
name isn't at all consistent with Swift guidelines but rather an incorrect
attempt to prioritize brevity over clarity.

What you're arguing is that protocol designers should be able to design
protocols without regard for how they will compose with others in
conforming types, relying on a new member-renaming feature instead. But, as
you point out, you can already use a protocol as a mere bag of code by
naming all members with unique, prefixed names, then have conforming types
forward their own choice of names to these.

This member-renaming feature you propose would enhance the aesthetic
pleasure of the protocol designer, allowing simple names that don't ever
have to appear in the public API of a concrete type to be used for a
protocol member without placing any restrictions on the API of conforming
types. However, I don't see anything wrong with the current hand-prefixing
method being enshrined as "best practice" for the bag-of-code approach to
protocols. If, as you predict, a growing third-party code ecosystem makes
name collisions worse, then in fact having uniquely distinguishable
prefixed members would be less confusing than having conforming types
renaming protocol members as a matter of course.

Take a look at Eiffel’s ‘rename’ & ’select’ features for similar
>> functionality and use-cases.
>>
>> Ultimately, this is a step in the direction of having true mixins.
>>
>>
> Sure, maybe. I couldn't evaluate that claim. I'm inclined to favor the
> proposal, but it'd have to stand on its own merits, not as a step to an
> as-yet undesigned feature.
>
>
> This would be part of that design.  We have to start somewhere.  A journey
> of 1000 miles begins with a single step.
>
>
>
>    I am fairly certain this affects the ABI though, so I thought I would
>> bring it up now.
>>
>>>
>>> If two protocols have methods/properties with the same name, but
>>> different signatures, we need a way to distinguish between them when
>>> attempting to conform to both.
>>>
>>>         protocol A {
>>>                 var x:Int {get set}
>>>         }
>>>
>>>         protocol B {
>>>                 var x:Double {get set}
>>>         }
>>>
>>
>> Methods can be overloaded that differ in arguments or return type, so it
>> seems like this problem mainly exists with *properties* that differ in
>> type--am I wrong?
>>
>>
>> There is also the case of functions with the same name and signature, but
>> different semantics.  There may be no single implementation which
>> simultaneously satisfies the semantics for both protocols. By renaming one
>> of the functions, we are able to provide separate implementations for each
>> requirement (which allows both protocols to function as intended).
>>
>
> True. However, putting on my critical hat, this seems like we're
> stretching to provide support for an anti-pattern. It'd be nice to have an
> example where one runs into the motivating problem *and* where the proposed
> feature promotes a _better_ design than is currently possible, rather than
> making a bad design compile.
>
> At this point, I'm imagining scenarios where a user is trying to conform a
> type MyAnimal to both Biped and Quadruped, then worrying that `walk()` has
> two semantics: something has already gone deeply wrong IMO.
>
> [Yes, I know there are animals that can sometimes walk on two or four
> legs. The point here is that the protocols were clearly designed to model
> animals at a certain level of detail, while it appears that the user
> writing `MyAnimal` wants to model the animal at a different level of detail
> than either protocol was designed to handle. You might have specific qualms
> about this particular hypothetical, but I think you can pick out the
> general point that there is a much larger problem inherent in the design
> than the specific problem regarding two colliding method signatures.]
>
>
> To take your example of walk().  Perhaps we have a protocol ‘Walkable’
> which refers to any data structure where the nodes can be walked using the
> ‘walk()’ function.  It is easy to imagine two different protocols A & B
> which specialize on this in different ways (say LinearWalkable &
> RandomWalkable), and both add some methods/properties and use those to
> provide efficient default implementations.  At some point, you may run into
> a data structure which could easily be walked in both ways.
>
> As things are right now, you couldn’t inherit from both protocols.  While
> you could add new ‘linearWalk()’ & ‘randomWalk()’ to the protocols
> respectively (cluttering their interface), there is still the issue of what
> to do when 'walk()’ is called.  You can’t rename walk() in the originating
> protocols because it comes from their common ancestor.  Much better to
> force one (or both) of the methods to be renamed on the conforming data
> structure.  That keeps the interfaces of the protocols clean and makes the
> options available on the data structure clearer (e.g. ‘walk()’ &
> ‘randomWalk()’ )
>
> What I have had to do in the current version is inherit from the original
> protocol and then copy and paste the default implementations from the
> specialized versions.  Now my code has been duplicated and is harder to
> maintain.  We can do better.
>
>
I think Charles's solution is pretty nice, but he's right that the API
surface area will have to grow. I don't know his original use case, so I
don't know how ugly I'd find the final solution to be in that scenario. In
this particular example, I'd say that having `linearWalk()` and
`randomWalk()` distinguished seems pretty sensible and an overall win for
clarity. If the same vendor controlled all three protocols, then `Walkable`
could have the `walk()` requirement removed altogether for even more
clarity.

Also, just a hunch, but I suspect your hypothetical would never hold. Could
you envision how the requirements for RandomWalkable might be such that
it's possible to implement an efficient _default_ implementation of a
random walk for any of several conforming data structures, but only one of
these data structures is LinearWalkable, _and_ such a linear walk is
efficient using another _default_implementation for an overlapping but not
identical set of data structures? It's not mere trivia here, because the
crux of your argument is that there exist default implementations that
require copying and pasting into conforming types (and sufficiently
efficient default implementations so that copying and pasting is
appropriate rather than implementing a more efficient version). More likely
in diamond problem scenarios, I think, colliding members are going to be
properties or methods either without default implementations or than need
to supply more efficient versions of default implementations anyway.

There may also be functions in different protocols with different names but
>> the same semantics and signature.  This will allow a single implementation
>> to satisfy both protocols without duplication.
>>
>
> This is a poor argument IMO. You can already implement foo() and then have
> bar() forward to foo() with trivial effort and really minimal boilerplate.
> It's an existing solution, and a more general solution because it doesn't
> require matching signatures. Also, it's a better solution IMO because it
> preserves the notion that a type T : Fooable, Barrable provides the full
> API guaranteed by Fooable and Barrable.
>
> Finally, we may want to rename an inherited default implementation to
>> avoid conflicting with another protocol's default implementation in cases
>> where we don’t want to override it.
>>
>
> Yes, I think this would be handy. I can't think of an existing way to do
> this, and I expect it might make a big difference in designing good
> protocols. So here, I think we have a strong argument.
>
> Again, though, could we find a concrete example of how this feature would
> promoter a better design of an actual type and/or protocol?
>
>  One possibility is to allow a struct/class/enum to conform to the
>> protocol while renaming one (or both) of the clashing methods:
>>
>>>
>>>         struct C: A,B {
>>>                 var x:Int
>>>                 var y:Double implements B.x
>>>         }
>>>
>>> The conforming method/property would still have to have the same
>>> signature, but could have a different name (and parameter labels).  It
>>> would also allow protocol methods which have identical signatures and
>>> semantics, but different names to be implemented using the same method (i.e
>>> ‘implements D.z & E.w’).
>>>
>>> When something is cast to the protocol (say ‘as B’), then calling the
>>> property (e.g. ‘x’) would end up calling the implementation of the renamed
>>> property ( ‘y’ in this example) on the conforming type.
>>>
>>
>> Reflecting on this proposed change, it occurs to me that something of
>> value would be lost, and I think that this something is actually rather
>> valuable:
>>
>> Today, when I see that a type conforms to (for example) Sequence, I know
>> that certain methods and/or properties exist on that type. Protocol
>> conformance guarantees a certain API, not just certain semantics.
>>
>>
>> It isn’t actually lost, however.  When working with it as a Sequence (for
>> example), that API would be intact using the original names.  It is only
>> when working with it as its own type that the renaming would have an effect.
>>
>
> That is not my point. With this proposal, knowing that MyGreatType
> conforms to Sequence would no longer yield any information as to the
> MyGreatType API. That is definitely something lost and we should
> acknowledge that.
>
> Also, recall that Sequence has Self or associated type requirements. So:
>
> ```
> let m = MyGreatType()
> // There is nothing I can write here to use the Sequence API with `m`,
> IIUC;
> // however, depending on how this feature is designed, I *might* be able to
> // call a generic function that operates on a type `T : Sequence` and work
> // with `m` that way.
> ```
>
>
> I get what you are saying, but this is really asking for an Xcode feature
> to better show the generated interfaces.  Code completion will still work
> properly.
>
>
>
>
> Perhaps one way to mitigate this loss would be to have any renamed members
>> listed *in the declaration of conformance*, something like this (with some
>> additional bikeshedding):
>>
>> ```
>> struct MyGreatType : Sequence (count => length) {
>>   // MyGreatType conforms to Sequence but renames `count` to `length`
>> }
>> ```
>>
>>
>> Yes, putting it in the conformance declaration is a definite possibility
>> we should consider.
>>
>>
>> I think we would also want a way to retroactively conform using existing
>>> properties/methods in an extension declaring conformance.  Not sure what
>>> the best syntax for that would be.  Off the top of my head (though I would
>>> love to have something with less cruft):
>>>
>>>         extension D:B {
>>>                 @conform(to: B.x, with: D.y)
>>>         }
>>>
>>> or maybe just:
>>>
>>>         extension D:B {
>>>                 D.y implements B.x
>>>         }
>>>
>>
>> If renamed members are declared along with protocol conformance, then the
>> syntax for retroactive modeling follows naturally:
>>
>> ```
>> extension D : B (x => y) { }
>> // again, the actual notation here is ugly
>> // but the underlying idea, I think, is worth considering
>> ```
>>
>>
>> Yup
>>
>> One thing I like about this is that it helps to solve the diamond
>> problem.  ‘x’ could be a default implementation in B which D does not
>> override.  I think this is an important case which my original proposal
>> didn’t address fully.
>>
>> We should keep bikeshedding the syntax though...
>>
>> Thanks,
>> Jon
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-evolution/attachments/20160824/aaab6c74/attachment.html>