> On Jan 12, 2017, at 11:23 PM, Slava Pestov via swift-evolution
> <[email protected]> wrote:
>> On Jan 12, 2017, at 7:05 PM, Karl Wagner <[email protected]
>> <mailto:[email protected]>> wrote:
>>
>>
>>> On 12 Jan 2017, at 22:37, Slava Pestov via swift-evolution
>>> <[email protected] <mailto:[email protected]>> wrote:
>>>
>>>>
>>>> On Jan 12, 2017, at 9:53 AM, Chris Eidhof via swift-evolution
>>>> <[email protected] <mailto:[email protected]>> wrote:
>>>>
>>>> Ok, I've got a draft up as a gist:
>>>> https://gist.github.com/chriseidhof/6c681677d44903045587bf75fb17eb25
>>>> <https://gist.github.com/chriseidhof/6c681677d44903045587bf75fb17eb25>
>>>>
>>>> Before I submit it, could someone let me know if adding generics to
>>>> subscripts would influence the ABI? ( still feel pretty clueless in that
>>>> area).
>>>
>>> It won’t change the ABI of existing subscript calls, but if the standard
>>> library introduces new generic subscripts that replace older non-generic
>>> subscripts, it will impact ABI.
>>>
>>> Slava
>>
>> Why are subscripts so different, anyway? One would think they are basically
>> sugared functions, but they don’t support so many things that regular
>> functions support. Not just syntax stuff, either - they also don’t support
>> @inline(__always) for some reason…
>
> Nice catch with @inline(__always). Please file a JIRA issue, since I’m
> actively working on this stuff now.
>
> Subscripts are a bit different from ordinary functions because they’re
> lvalues. You can call mutating members on the result of a subscript, or chain
> it with another property access or subscript. So the code path in SILGen is a
> little different than ordinary function calls.
I would put it a little differently. Subscripts are used to refer to a logical
component of a value, the same way that a property is. Subscripts are
obviously different from properties because they're parameterized by some sort
of index, but otherwise they're very similar in the core sense that the
expression by itself just abstractly refers to a component of another value,
and it's *how that expression is used* that really determines its semantics.
That is, "a[i]" would always be a very odd thing to have as a statement on its
own, just like "pi" would be a very odd thing to have as a statement on its own
— generally you're reading from it or writing to it, and which one exactly
you're doing can result in very different code being executed. It's that
contextual sensitivity that makes subscripts quite different from just "sugared
functions".
> This is the reason subscripts cannot have default arguments also. With a bit
> of refactoring we can unify the code path for forming call arguments in
> ordinary calls and subscripts, and hopefully the default argument and generic
> cases will fall out naturally also.
Right. There isn't any particular reason that subscripts don't support default
arguments; it's just a little extra work that we've never done.
>> Where generic subscripts are concerned, there are a couple of different
>> things to express:
>> - Generic parameter (I can understand various co-ordinates for the data)
>> - Generic return type (I can construct your preferred representation of the
>> data)
>> - Generic setter type (I can set the data using various compatible types):
>
> I think all of these should be expressed with a single generic signature on
> the subscript itself. The element type passed to the setter and returned from
> the getter should be the same IMO, otherwise it’s not clear how it will work.
Yes. It's quite important that any particular subscript reference is still a
single consistent entity, even if generic; we would not want, say, a
read-modify-write access to be able to somehow invoke the getter and setter at
different generic arguments, or to traffic in different element types.
I'm also not sure we'd ever want the element type to be inferred from context
like this. Generic subscripts as I see it are about being generic over
*indexes*, not somehow about presenting a polymorphic value.
John.
>
>>
>> protocol MeaningfulToFoo {}
>> protocol ConstructableFromFoo {}
>>
>> struct Foo {
>> subscript<Index>(index: Index) where Index: SignedInteger {
>> get<T> where T: ConstructableFromFoo { return T(self) }
>> set<U> where T: MeaningfulToFoo { self.someProperty =
>> newValue.someData }
>> }
>> }
>>
>> The syntax looks a bit awkward, though, IMO. I’m wondering if it might be
>> better to have some kind of combined subscript + property behaviours
>> (remember those?) and allow those to be generic instead. Subscripts and
>> properties are very similar anyway - they are both bundles of functions to
>> represent getting and setting data (not just regular-old “get” and “set”,
>> either, but also magic stuff like “mutableAddressWithPinnedNativeOwner”).
>> The only difference is that property getters can’t have parameters — which
>> is something I would also like to see lifted one day (I believe I’ve even
>> seen people asking for “named subscripts” due to the lack of this =P)
>>
>> - Karl
>>
>
> _______________________________________________
> swift-evolution mailing list
> [email protected]
> https://lists.swift.org/mailman/listinfo/swift-evolution
_______________________________________________
swift-evolution mailing list
[email protected]
https://lists.swift.org/mailman/listinfo/swift-evolution
