[swift-evolution] Pitch: really_is and really_as operators

Xiaodi Wu xiaodi.wu at gmail.com
Thu Aug 25 15:56:57 CDT 2016 

On Thu, Aug 25, 2016 at 3:24 PM, Matthew Johnson <matthew at anandabits.com>
wrote:

>
> On Aug 25, 2016, at 11:56 AM, Xiaodi Wu <xiaodi.wu at gmail.com> wrote:
>
> On Thu, Aug 25, 2016 at 11:28 AM, Matthew Johnson <matthew at anandabits.com>
> wrote:
>
>>
>> On Aug 25, 2016, at 11:15 AM, Xiaodi Wu <xiaodi.wu at gmail.com> wrote:
>>
>> On Thu, Aug 25, 2016 at 10:07 AM, Matthew Johnson <matthew at anandabits.com
>> > wrote:
>>
>>>
>>> On Aug 25, 2016, at 9:37 AM, Xiaodi Wu <xiaodi.wu at gmail.com> wrote:
>>>
>>> Charles clarified that indeed he was pitching casting operators that
>>> match subclasses.
>>>
>>>
>>> Ok, I missed that.
>>>
>>>
>>> If the compiler knows that a class is sealed, why do you think there's a
>>> new keyword needed for the compiler to prove exhaustiveness?
>>>
>>>
>>> I said I wasn’t sure if there was a different / better way to do it.
>>> Just that this *would* do it.
>>>
>>>
>>> First, you can already match exact types by writing `type(of: instance)
>>> == Base.self` (which doesn't evaluate to true if instance is of a subclass
>>> of Base).
>>>
>>>
>>> This might be an alternative if the compiler adds special knowledge of
>>> this syntax to prove exhaustiveness.
>>>
>>
>> I might be in favor of that. As it is, I can write this:
>>
>> ```
>> func ~= <T, U>(lhs: T.Type, rhs: U.Type) -> Bool {
>>   return lhs == rhs
>> }
>>
>> class Base {
>>   init() { }
>> }
>> class A1 : Base { }
>> class A2 : Base { }
>>
>> let a = A1()
>>
>> switch type(of: a) {
>> case A1.self:
>>   print(1)
>> case A2.self:
>>   print(2)
>> case Base.self:
>>   print(0)
>> default:
>>   fatalError()
>> }
>> ```
>>
>> It'd be nice if the compiler would know about exhaustiveness (and if I
>> didn't have to define my own `~=`). This is, afaict, doable without any
>> additional syntax in the language.
>>
>> Second, if your class hierarchy is Base > A > B > C, then _even if_ there
>>> existed no way to match exact types (which there is), you have the option
>>> of switching over the type of an instance, providing cases that match in
>>> the order C, B, A, Base in order to perform a different action for each.
>>> This requires no additional knowledge at compile time beyond what you
>>> already stipulated for your use case, namely that the entire class
>>> hierarchy must be known at compile time.
>>>
>>>
>>> This order requirement is fragile.  If you put Base first it will always
>>> match, which probably isn’t the intent.  I would like to see a solution
>>> that requires you to match each type in the hierarchy without being subject
>>> to bugs related to ordering of the cases.
>>>
>>
>> Given that the hierarchy is known at compile-time, a solution that would
>> meet your criteria (not being subject to bugs related to ordering) would be
>> diagnostics for unreachable cases (i.e., if Base is matched before A, `case
>> is A` should be flagged as unreachable).
>>
>>
>> That isn’t an adequate solution IMO.  First, it forces an ordering of the
>> cases.  But more importantly, it will still allow accidentally matching
>> superclass and omitting subclass cases (so long as there are no unreachable
>> cases).
>>
>
> You cannot accidentally match superclasses by incorrect ordering, because
> it will always result in unreachable cases in a language without multiple
> inheritance. As to omitting subclasses, what's the harm in that? Suppose
> you had a hierarchy Base > A > B > C, but Joe the Programmer doesn't know
> about C. So he switches over the cases exhaustively and uses Base, A, and B
> exclusively. What problems will he encounter?
>
>
> Maybe in that case he won’t encounter any problems.  However, if you have:
>
>  Base
>   |    |
> C1 C2
>  |     |
> D1 D2
>
> And you just switch over Base, C1, and C2 you may have a match for all
> possible runtime values but you don’t have cases for all of the possible
> type.  Imagine if you have code that is initially written before D1 and D2
> are introduced and your intent is to include a case for each possible
> runtime type.  You want the code to fail to compile when D1 and D2 are
> introduced until you revisit the switch statement
>

I understand that this is what you want, but what is your use case for
this? Why should you want your code to fail to compile?

To be sure, it is currently *possible* to get this behavior (at runtime,
admittedly) by having a default case that calls fatalError() after a series
of cases that explicitly compare dynamic type using type(of:) and ==, but
this is a code smell at the very least, and certainly not something we
should want to make easier with a keyword (even if that allows the failure
to manifest at compile time). For one, it breaks our understanding of
library versioning:

Library Foo defines the following class hierarchy, which is entirely public:

  Base (sealed) > A > B > C

Sally uses library Foo and uses your proposed keyword in an exhaustive
switch.

The next minor release of library Foo (and it's a minor release because the
public API is unchanged) adds an internal subclass of C called __C (this
would be analogous to how NSString and __NSCFString are related, but of
course those are Objective-C classes in Darwin Foundation).

Various methods that formerly returned a value of type C now return a value
of type __C upcast as C. Sally's code ceases to compile, and through no
fault of the library developer. There is nothing she can do using your
proposed keyword to fix this defect, because __C is not public. She *must*
use casting operators that match subclasses.

In the next minor version of library Foo, __C is no longer necessary and is
removed. Sally's code doesn't break this time, because she stopped using
your proposed keyword.


> , just as it does when you add a new case to an enum.  However, given the
> current tools there is nothing you can do to force this.  This is what I
> meant by “accidentally”.
>
> Having a tool for preventing these kinds of mistakes seems like a useful
> thing and puts classes (and protocols if it works with them as well when
> all conforming types are known) on equal footing with enums in terms of
> pattern matching.  This gives us more options for structuring our code.
>

I'd have to be convinced that this tool is at all wise with a compelling
use case. IMO, it encourages far, far more fragile code than a poorly
ordered series of cases would do currently. Therefore, at the moment, it
seems to me that this tool would be the greater mistake.

To be clear, your pitch has nothing to do with exhaustiveness on par with
enums, because it is possible to write switch statements over class
hierarchies either exhaustively or non-exhaustively using the existing
casting operators; the compiler can be given enough smarts to emit the
appropriate errors when the cases are non-exhaustive and to stop insisting
on a `default` case when the cases are exhaustive. I'd be absolutely in
favor of that change.

> I would prefer a solution that requires *all classes* in the hierarchy to
>> be matched exactly.  The existing casts would still work for cases when you
>> don’t care about that and *do* want to use a superclass to match subclasses.
>>
>
> I don't understand the use case. This seems esoteric and goes beyond
> exhaustive pattern matching.
>
>
>>> Third, your motivating example in the previous thread already works.
>>> Slightly renamed to match the examples above, the following compiles:
>>>
>>>
>>> ```
>>> class Base { init() { } } class A1 : Base { } class A2 : Base { } func
>>> foo(_ b: Base) -> Int { switch b { case is A1: return 1 case is A2: return
>>> 2 case is Base: return 0 } }
>>>
>>> let a = A1() let b = A2() foo(a) // 1 foo(b) // 2
>>> ```
>>>
>>> There is a warning that `case is Base` is always true. Perhaps something
>>> could be done about that diagnostic, since that is after all what you want
>>> in a switch statement without a default case.
>>>
>>> I'm sure you were aware of all of these points, so I guess I'm asking,
>>> what exactly are you pitching?
>>>
>>>
>>> See above.  I am looking for a solution that avoids this warning
>>> precisely because it will *not* always be true.  The compiler gaining
>>> special knowledge of the `type(of: instance) == Base.self` pattern
>>> could be *part* of a solution but it still doesn’t bind a name the correct
>>> type. For example, with the Base > A > B > C hierarchy when I match
>>> `type(of: instance) == B.self` I also want a variable bound with a type of
>>> `B`. This gets pretty verbose and requires the compiler to have special
>>> knowledge of pretty specific pattern:
>>>
>>> func foo(_ b: Base) -> Int {
>>>   switch b {
>>>   case let base as Base where type(of: instance) == Base.self: return 1
>>>   case let a as A where type(of: instance) == A.self: return 2
>>>   case let b as B where type(of: instance) == B.self: return 3
>>>   case let c as C where type(of: instance) == C.self: return 4
>>>   }
>>> }
>>>
>>
>>
>>> If the compiler could prove exhaustiveness here I would accept that
>>> solution.  But it seems like an exact match cast operator would be much
>>> more elegant.
>>>
>>> In any case, anything that requires matching every type in a hierarchy
>>> without being subject to case ordering bugs and doesn’t require a default
>>> clause would be acceptable to me.  That is the problem I would like to see
>>> solved.
>>>
>>
>> Looking back, it seems like diagnostics for unreachable cases would meet
>> your criteria exactly and would be the most straightforward. I don't think
>> it would even require an evolution proposal. I would love to see type(of:)
>> work with switch statements out-of-the-box, but that seems more esoteric.
>> None of this requires additional syntax, IMHO.
>>
>>
>> Diagnostics would be an improvement for sure, but aren’t a complete
>> solution IMO for the reasons noted above.  We don’t necessarily need new
>> syntax but I don’t have any great alternatives at the moment.
>>
>>
>>
>>>
>>> On Thu, Aug 25, 2016 at 08:40 Matthew Johnson <matthew at anandabits.com>
>>> wrote:
>>>
>>>> On Aug 24, 2016, at 9:33 PM, Xiaodi Wu <xiaodi.wu at gmail.com> wrote:
>>>>
>>>> On Wed, Aug 24, 2016 at 9:25 PM, Matthew Johnson via swift-evolution <
>>>> swift-evolution at swift.org> wrote:
>>>>
>>>>>
>>>>> On Aug 24, 2016, at 9:09 PM, Robert Widmann via swift-evolution <
>>>>> swift-evolution at swift.org> wrote:
>>>>>
>>>>> I have 3 qualms with this proposal as it stands:
>>>>>
>>>>> - type(of:) will never lie to you.
>>>>>
>>>>> The only question it won’t answer to your satisfaction is the dynamic
>>>>> type of the NSString you boxed up as an Any.
>>>>>
>>>>> - No more keywords without significant justification.
>>>>>
>>>>> I don’t buy the performance use case at all - if you were properly
>>>>> concerned about performance you would try to use as many of Swift’s static
>>>>> features as possible.
>>>>>
>>>>> - Especially no more keywords that look like they belong in Rust or
>>>>> PHP!
>>>>>
>>>>> There is no precedent for the spelling of these operations other than
>>>>> the suffixed punctuation. Given that they’re domain-specific, will
>>>>> definitely be hard to use (given that NSString(string: "Bar”) may not
>>>>> “really” given you an NSString yet that’s what you asked us to check for “
>>>>> *really*"), and will be obviated by the implementation of SE-0083, I
>>>>> can’t see a reason why we need any of this in the language proper.
>>>>>
>>>>>
>>>>> One related topic to consider is exhaustive pattern matching for
>>>>> classes.  Now that SE-0117 has been accepted it will be possible to do this
>>>>> for many classes (I would say most if it weren’t for Objective-C classes
>>>>> being so common in Swift and are imported as `open`).  Supporting
>>>>> exhaustive pattern matching well would require some kind of syntax for
>>>>> matching the runtime type exactly.  I have imagined this as being “exact
>>>>> match” cast operators, which is what the `really_*` operators are.
>>>>>
>>>>
>>>> I don't understand. As pitched, these operators remove bridging magic,
>>>> but `Subclass really_is Superclass == true`. How would you use this for
>>>> classes?
>>>>
>>>>
>>>> Bridging is the use case motivating the pitch.  I am bringing up a
>>>> related use case.
>>>>
>>>> The pitch does not specify `Subclass really_is Superclass == true` and
>>>> I would argue that this is not the semantics we would want.  My
>>>> interpretation of the proposed solution is:
>>>>
>>>> "I propose the following operators: really_is, really_as, really_as?,
>>>> and really_as!. These operators would only return a positive result if the
>>>> type actually was what was being asked for, instead of something that might
>>>> be able to bridge to that type *or a superclass of that type*."
>>>>
>>>> We discussed the exhaustive pattern matching previously in this thread:
>>>> https://lists.swift.org/pipermail/swift-evolution/We
>>>> ek-of-Mon-20160523/018799.html where the “exact match” cast operators
>>>> were called `isExactly` and `asExactly`.
>>>>
>>>> I think the exhaustive pattern matching use case for classes (and
>>>> protocols if / when we get sealed protocols) is an important one.  I also
>>>> think doing it right requires the ability to match exact types (i.e. not
>>>> match subclasses).  Maybe there is a better mechanism than a new operators
>>>> but they would certainly do the job well.
>>>>
>>>>
>>>>
>>>>> Do you have an alternative in mind for exhaustive pattern matching if
>>>>> we do not introduce exact match cast operators?
>>>>>
>>>>>
>>>>> ~Robert Widmann
>>>>>
>>>>> On Aug 24, 2016, at 5:08 PM, Charles Srstka via swift-evolution <
>>>>> swift-evolution at swift.org> wrote:
>>>>>
>>>>> MOTIVATION:
>>>>>
>>>>> SE-0083 appears to be dead in the water, having been deferred until
>>>>> later in Swift 3 back in May and not having been heard from since then,
>>>>> with the Swift 3 release looming closer and closer. However, the
>>>>> predictability gains that would have been provided by this change remain
>>>>> desirable for cases where one needs to know the actual dynamic type of an
>>>>> entity before any bridging magic is involved. Additionally,
>>>>> performance-critical code may desire the ability to check something’s type
>>>>> quickly without incurring the overhead of Objective-C bridging code.
>>>>>
>>>>> PROPOSED SOLUTION:
>>>>>
>>>>> I propose the following operators: really_is, really_as, really_as?,
>>>>> and really_as!. These operators would only return a positive result if the
>>>>> type actually was what was being asked for, instead of something that might
>>>>> be able to bridge to that type.
>>>>>
>>>>> DETAILED DESIGN:
>>>>>
>>>>> let foo: Any = "Foo"
>>>>> let bar: Any = NSString(string: "Bar")
>>>>>
>>>>> let fooIsString = foo is String                  // true
>>>>> let fooReallyIsString = foo really_is String     // true
>>>>>
>>>>> let fooIsNSString = foo is NSString              // true
>>>>> let fooReallyIsNSString = foo really_is NSString // false
>>>>>
>>>>> let barIsString = bar is String                  // true
>>>>> let barReallyIsString = bar really_is String     // false
>>>>>
>>>>> let barIsNSString = bar is NSString              // true
>>>>> let barReallyIsNSString = bar really_is NSString // true
>>>>>
>>>>> ALTERNATIVES CONSIDERED:
>>>>>
>>>>> Stick with using an unholy combination of Mirror and unsafeBitCast
>>>>> when you need to know what you’ve actually got.
>>>>>
>>>>> Charles
>>>>>
>>>>> _______________________________________________
>>>>> 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
>>>>>
>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> 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/20160825/ef8dd24c/attachment.html>

More information about the swift-evolution mailing list