> On 17 Aug 2017, at 18:04, Chris Lattner <[email protected]> wrote:
>> On Aug 17, 2017, at 5:00 AM, Haravikk via swift-evolution
>> <[email protected] <mailto:[email protected]>> wrote:
>>> On 17 Aug 2017, at 11:42, Robert Bennett <[email protected]
>>> <mailto:[email protected]>> wrote:
>>>
>>> Chris mentions that the intent was to mimic a default implementation in a
>>> constrained protocol extension, with one extension per type that doesn’t
>>> define its own ==/hashValue while conforming to Equatable/Hashable.
>>> Constrained extensions are allowed to use type information from the
>>> constraint, so I don’t think there is an issue here.
>>
>> And I disagree; this isn't a constraint extension either, not even close,
>> we're talking here about automatic behaviour based upon variables the
>> protocol knows literally nothing about, in a way that can result in new
>> errors that are currently impossible (as you can't currently conform to
>> Equatable without providing some kind of code to implement it).
>
> I understand and recognize your concern. Your points are apply equally to
> the default implementation of the Codable protocols as well, and the core
> team specifically discussed this.
>
> Also, if I were to nitpick your argument a bit, it isn’t true that the
> protocol knows “nothing" about the type anyway, because the protocol has
> access to self. The default implementation could conceptually use reflection
> to access all the state in the type: we’re producing the same effect with
> more efficient code.
Parts that the protocol specifically defines are fine, I've said as much; the
part I object to is everything else as the issue here is that the protocol must
make assumptions about the concrete type and I find that dangerous, especially
on an existing protocol. I would argue that any default implementation using
reflection to do the same is likewise flawed unless the behaviour is very
carefully and very clearly defined as part of a contract that developers are
explicitly opting in to.
But that's not what's happening here; Equatable is an existing and well
understood protocol and you are proposing to change it arbitrarily to suddenly
imply functionality that doesn't currently exist and which has the potential to
introduce bugs.
I have seen not one shred of justification why this feature must be implicit
through Equatable except that Codable does it, and frankly I don't find that
even close to acceptable as a precedent at all as Codable is not preexisting,
and personally I don't like it in Codable's case either, and wasn't aware of
the discussion around it. But for Equatable we're talking about a preexisting
protocol that will today catch 100% of missing conformance bugs, being changed
such that that is no longer the case because of a default behaviour that will
hide such bugs by making potentially flawed assumptions about a concrete type.
Frankly, setting a precedent for this kind of automated background reflective
guesswork on basic protocols is a horrifying prospect to me, even more so if it
is being introduced arbitrarily on existing protocols. At least with things
like unit testing that reflect test methods the rules are clearly known from
the outset (i.e- there's a clear naming convention for what is reflected,
everything else is the developer's domain, you opt-in by following that naming
convention), this is not the case here.
And what exactly is the burden from opting in explicitly? A different protocol
name, a keyword or an attribute are not going to trouble developers, and will
clarify exactly what they intended without hiding bugs. This is why I feel this
hasn't been considered sufficiently at all, as no justification is given why
"AutoEquatable" or whatever is somehow an unacceptable burden to clarify what a
developer actually wanted.
For me the whole point of a basic protocol is that it forces me to implement
some requirements in order to conform; I can throw a bunch of protocols onto a
type and know that it won't compile until I've finished it, developers get
distracted, leave things unfinished to go back to later, make typos etc. etc.
To me declaring a conformance is a declaration of "my type will meet the
requirements for this make, sure I do it", not "please, please use some magic
to do this for me"; there needs to be a clear difference between the two.
This is an overreach, plain and simple. While it may seem small, I would
frankly rather go back to coding everything in C++ than continue using Swift if
this is to be the direction of travel, as there's a big a difference between
convenience and trying to guess what a developer wanted. Cutting down on
boilerplate is a fine goal, but it should be because I as the developer wanted
it, not someone else who thinks they're doing me a favour despite knowing
nothing about my code.
_______________________________________________
swift-evolution mailing list
[email protected]
https://lists.swift.org/mailman/listinfo/swift-evolution
