[swift-dev] Conditional conformance: Removing the opt-in flag

Tue Jan 2 13:14:11 CST 2018

On 27 Dec 2017, at 12:41, Douglas Gregor via swift-dev wrote:

> Sent from my iPhone
>
>> On Dec 27, 2017, at 11:05 AM, Karl Wagner <razielim at gmail.com> wrote:
>>
>>
>>
>>> On 22. Dec 2017, at 07:13, Ted Kremenek via swift-dev 
>>> <swift-dev at swift.org> wrote:
>>>
>>>
>>>
>>>> On Dec 19, 2017, at 9:39 PM, Ted Kremenek via swift-dev 
>>>> <swift-dev at swift.org> wrote:
>>>>
>>>>
>>>>
>>>>> On Dec 19, 2017, at 8:57 PM, Douglas Gregor <dgregor at apple.com> 
>>>>> wrote:
>>>>>
>>>>>
>>>>>
>>>>> Sent from my iPhone
>>>>>
>>>>>> On Dec 19, 2017, at 8:31 PM, Ted Kremenek <kremenek at apple.com> 
>>>>>> wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>>> On Dec 19, 2017, at 3:59 PM, Douglas Gregor <dgregor at apple.com> 
>>>>>>> wrote:
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>> On Dec 19, 2017, at 2:26 PM, Ted Kremenek via swift-dev 
>>>>>>>> <swift-dev at swift.org> wrote:
>>>>>>>>
>>>>>>>>
>>>>>>>>> On Dec 18, 2017, 4:53 PM -0800, Douglas Gregor via swift-dev 
>>>>>>>>> <swift-dev at swift.org>, wrote:
>>>>>>>>> Hi all,
>>>>>>>>>
>>>>>>>>> A little while back, I added an error to the Swift 4.1 
>>>>>>>>> compiler that complains if one tries to use conditional 
>>>>>>>>> conformances, along with a flag 
>>>>>>>>> “-enable-experimental-conditional-conformances” to enable 
>>>>>>>>> the feature. We did this because we haven’t implemented the 
>>>>>>>>> complete proposal yet; specifically, we don’t yet handle 
>>>>>>>>> dynamic casting that involves conditional conformances, and 
>>>>>>>>> won’t in Swift 4.1.
>>>>>>>>>
>>>>>>>>> I’d like to take away the 
>>>>>>>>> "-enable-experimental-conditional-conformances” flag and 
>>>>>>>>> always allow conditional conformances in Swift 4.1, because 
>>>>>>>>> the changes in the standard library that make use of 
>>>>>>>>> conditional conformances can force users to change their code 
>>>>>>>>> *to themselves use conditional conformances*. Specifically, if 
>>>>>>>>> they had code like this:
>>>>>>>>>
>>>>>>>>> extension MutableSlice : P { }
>>>>>>>>> extension MutableBidirectionalSlice : P { }
>>>>>>>>> // …
>>>>>>>>>
>>>>>>>>> they’ll get an error about overlapping conformances, and 
>>>>>>>>> need to do something like the following to fix the issue:
>>>>>>>>>
>>>>>>>>> extension Slice: P where Base: MutableCollection { }
>>>>>>>>>
>>>>>>>>> which is way more elegant, but would require passing 
>>>>>>>>> "-enable-experimental-conditional-conformances”. That 
>>>>>>>>> seems… unfortunate… given that we’re forcing them to use 
>>>>>>>>> this feature.
>>>>>>>>>
>>>>>>>>> My proposal is, specifically:
>>>>>>>>>
>>>>>>>>> Allow conditional conformances to be used in Swift 4.1 (no 
>>>>>>>>> flag required)
>>>>>>>>> Drop the -enable-experimental-conditional-conformances flag 
>>>>>>>>> entirely
>>>>>>>>> Add a runtime warning when an attempt to dynamic cast fails 
>>>>>>>>> due to a conditional conformance, so at least users know 
>>>>>>>>> what’s going on
>>>>>>>>>
>>>>>>>>
>>>>>>>> The last bullet doesn’t feel right to me.  It sounds like we 
>>>>>>>> would ship a feature that we know only partially works, but 
>>>>>>>> issue a runtime warning in the case we know isn’t fully 
>>>>>>>> implemented?  I’m I interpretting that point correctly?
>>>>>>>
>>>>>>> Yes, that’s correct. We will fail to match the conformance 
>>>>>>> (i.e., return “nil” from an “as?” cast), which might be 
>>>>>>> correct and might be wrong.
>>>>>>>
>>>>>>> 	- Doug
>>>>>>
>>>>>> Hmm.  I’m concerned that a warning runtime would be to settle. 
>>>>>> Many people would possibly not even notice it.  It’s 
>>>>>> essentially an edge case in a feature that isn’t fully 
>>>>>> implemented and thus that part of the feature should not be used 
>>>>>> yet.
>>>>>>
>>>>>> What do you think about making this a hard runtime error instead, 
>>>>>> similar to how we are approaching runtime issues for exclusivity 
>>>>>> checking?  That would be impossible to miss and would convey the 
>>>>>> optics that this runtime aspect of the feature is not yet 
>>>>>> supported and thus should not be used.
>>>>>
>>>>> I’d rather not make it a runtime error, because code that’s 
>>>>> doing dynamic casting to a protocol is generally already handling 
>>>>> the “nil” case (“as?” syntax), so aborting the program 
>>>>> feels far too strong.
>>>>>
>>>>>   - Doug
>>>>
>>>> For me I think the part I’m struggling with is that making it a 
>>>> warning conflates two things together: expected failure in the 
>>>> dynamic cast because the value you are casting doesn’t have that 
>>>> type or — in this case — failure because the cast can never 
>>>> succeed because it is not supported yet.  I feel like we would be 
>>>> silently swallowing an unsupported condition.  If that didn’t 
>>>> matter, why bother issuing a warning?  Clearly were trying to send 
>>>> some kind of message here about this not being supported.
>>>
>>> Doug and I chatted a bit offline.
>>>
>>> I’m now more on the side of thinking a warning is a reasonable 
>>> approach.  I’m still concerned that it will be unnoticed by some 
>>> developers, and I am mixed on conflating failure the cast of “this 
>>> doesn’t work at all for this specific type because it has a 
>>> conditional conformance” versus “this didn’t work because the 
>>> type didn’t conform to the protocol”.  That said, I think the 
>>> cases impacted here are likely very, very small — and a crash in 
>>> the program is probably excessive.
>>> _______________________________________________
>>> swift-dev mailing list
>>> swift-dev at swift.org
>>> https://lists.swift.org/mailman/listinfo/swift-dev
>>
>>
>> What about if we disabled conditional conformances for non-generic 
>> protocols (or keep that part behind the flag)? It seems a bit 
>> arbitrary, but IIRC, the standard library uses conditional 
>> conformances for things like Equatable and the various faces of 
>> Collection, which are not runtime-castable anyway.
>
> This is a reasonable approach. To make it consistent, we would want to 
> make the Codable conformances of Array, Set, Dictionary, and Optional 
> unconditional again—otherwise, user code couldn’t have Codable 
> conformances for some types without adding the flag.
>
> I’m hesitant to do this because the unconditional conformances to 
> Codable are wrong, and fixing them is going to cause some (legitimate, 
> necessary) source breakage. It feels better overall to get that out of 
> the way sooner... before more wrong Codable conformances get layered 
> on top.
>
>   - Doug

Just to be clear, what is the current impact of leaving those `Codable` 
conformances conditional? Having casts like `[1,2,3] as? Codable` fail?
For what it’s worth, I’m not entirely certain how useful type-erased 
`Encodable` and `Decodable` values are in practice (aside from the very 
workarounds we used because conditional conformances weren’t ready 
yet)… Can we try to get some metrics about whether people do this at 
all today and for what purposes?

I don’t feel too strongly either way (these conformances have been 
unconditional so far, so we’re not regressing anything by not doing 
this for `Codable`), but I do think it would be nice to have the actual 
conditional conformances in place if we’re not breaking anyone.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-dev/attachments/20180102/caf1535c/attachment.html>