Hello Swift community,
Associated type inference, which is the process by which the Swift compiler attempts to infer typealiases to satisfy associated-type requirements based on other requirements, has caused both implementation problems and user confusion for a long time. Some of you might remember a previous (failed) attempt to remove this feature from the Swift language, in SE-0108 “Remove associated type inference”. <https://github.com/apple/swift-evolution/blob/master/proposals/0108-remove-assoctype-inference.md>
I’m not sure we can remove this feature outright (i.e., the concerns that sank that proposal are still absolutely valid), because it is so very convenient and a lot of user code likely depends on it in some form or other. So, here I’d like to describe the uses of the feature, its current (very problematic) implementation approach, and a half-baked proposal to narrow the scope of associated type inference to something that I think is more tractable. I need help with the design of this feature, because I feel like it’s going to be a delicate balance between implementability and expressiveness.
A Motivating Example
As a motivating example, let’s consider a “minimal” random access collection:
struct MyCollection<T> {
var contents: [T]
}
extension MyCollection: RandomAccessCollection {
var startIndex: Int { return contents.startIndex }
var endIndex: Int { return contents.endIndex }
subscript(index: Int) -> T { return contents[index] }
}
This is actually pretty awesome: by providing just two properties and a subscript, we get the full breadth of the random access collection API! This is relying heavily on associated type inference (for associated type requirements) and default implementations specified on protocol extensions. Without associated type inference, we would have had to write:
extension MyCollection: RandomAccessCollection {
typealias Element = T
typealias Index = Int
typealias Indices = CountableRange<Int>
typealias Iterator = IndexingIterator<MyCollection<T>>
typealias SubSequence = RandomAccessSlice<MyCollection<T>>
var startIndex: Int { return contents.startIndex }
var endIndex: Int { return contents.endIndex }
subscript(index: Int) -> T { return contents[index] }
}
where the bolded typealiases are currently inferred. It was worse back when we reviewed SE-0108, because IIRC there were a few underscored associated types (e.g., _Element) that have since been removed. Still, that’s a bit of additional work to define a “minimal” collection, and requires quite a bit more understanding: how do I know to choose IndexingIterator, and CountableRange, and RandomAccessSlice?
The issue would get worse with, e.g., SE-0174 “Change filter to return an associated type” <https://github.com/apple/swift-evolution/blob/master/proposals/0174-filter-range-replaceable.md>, which adds an associated type Filtered that almost nobody will ever customize, and isn’t really fundamental to the way collections work. Adding Filtered to the standard library would be a source-breaking change, because users would have to write a typealias giving it its default.
Associated Type Defaults
One of the ways in which we avoid having to specify typealiases is to use associated type defaults. For example, the standard library contains associated type defaults for Indices, Iterator, and SubSequence. Let’s focus on Indices:
protocol Collection : Sequence {
associatedtype Indices = DefaultIndices<Self>
// ...
}
protocol BidirectionalCollection : Collection {
associatedtype Indices = DefaultBidirectionalIndices<Self>
// ...
}
protocol RandomAccessCollection : BidirectionalCollection {
associatedtype Indices = DefaultRandomAccessIndices<Self>
// ...
}
The basic idea here is that different protocols in the hierarchy provide different defaults, and you presumably want the default from the most specific protocol. If I define a type and make it conform to BidirectionalCollection, I’d expect to get DefaultBidirectionalIndices<Self> for Indices. If a define a type and make it conform to RandomAccessIterator, I’d expect to get DefaultRandomAccessIndices<Self>.
(Aside: DefaultRandomAccessIndices and DefaultBidirectionalIndices got collapsed into DefaultIndices now that we have conditional conformances for the standard library <https://github.com/apple/swift/pull/12913>, but the issues I’m describing remain).
Associated type defaults seem like a reasonable feature that fits well enough into the design. However, it’s not the only thing in place with our MyCollection example, for which Indices was inferred to CountableRange. How’s that happen?
Associated Type Inference
Associated type inference attempts to look at the requirements of a protocol, and then looks into the conforming type for declarations that might satisfy those requirements, and infers associated types from the types of those declarations. Let’s look at some examples. RandomAccessCollection has some requirements that mention the Index and Element types:
protocol RandomAccessCollection : BidirectionalCollection {
associatedtype Element
associatedtype Index
var startIndex: Index { get }
var endIndex: Index { get }
subscript (i: Index) -> Element { get }
}
Associated type inference wil try to satisfy those requirements for MyCollection, and will find these declarations:
extension MyCollection: RandomAccessCollection {
var startIndex: Int { return contents.startIndex }
var endIndex: Int { return contents.endIndex }
subscript(index: Int) -> T { return contents[index] }
}
and match them up. From startIndex, we infer Index := Int. From endIndex, we infer Index := Int. From subscript, we infer Index := Int and Element := T. Those results are all consistent, so we’ve properly inferred Index and Element. Yay.
Associated type inference often has to deal with ambiguities. For example, there is an extension of Collection that provides a range subscript operator:
extension Collection {
subscript (bounds: Range<Index>) -> Slice<Self> { … }
}
When we look at that and match it to the subscript requirement in RandomAccessCollection (remembering that argument labels aren’t significant for subscripts by default!), we infer Index := Range<Index> (admittedly weird) and Element := Slice<Self> (could be legitimate). We have to discard this candidate because the deduction from startIndex/endIndex (Index := Int) collides with this deduction (Index := Range<Index>).
In our initial example, we saw that Indices was inferred to CountableRange<Int>. Let’s look at another slice of the RandomAccessCollection protocol that’s relevant to this associated type:
protocol RandomAccessCollection : BidirectionalCollection {
associatedtype Index
associatedtype Indices: RandomAccessCollection where Indices.Element == Index
var indices: Indices { get }
}
We will match that requirement for an “indices” property against a member of a constrained protocol extension of RandomAccessCollection:
extension RandomAccessCollection where Index : Strideable, Index.Stride == IndexDistance, Indices == CountableRange<Index> {
public var indices: CountableRange<Index> {
return startIndex..<endIndex
}
}
Associated type inference can determine here that Indices := CountableRange<Index>, but only when Index: Strideable and Index.Stride == IndexDistance. In other words, there are a whole bunch of other constraints to verify before we can accept this inference for Indices.
What’s a Good Solution Look Like?
Our current system for associated type inference and associated type defaults is buggy and complicated. The compiler gets it right often enough that people depend on it, but I don’t think anyone can reasonably be expected to puzzle out what’s going to happen, and this area is rife with bugs. If we were to design a new solution from scratch, what properties should it have?
It should allow the author of a protocol to provide reasonable defaults, so the user doesn’t have to write them
It shouldn’t require users to write typealiases for “obvious” cases, even when they aren’t due to defaults
It shouldn’t infer an inconsistent set of typealiases
It should be something that a competent Swift programmer could reason about when it will succeed, when and why it will fail, and what the resulting inferred typealiases would be
It should admit a reasonable implementation in the compiler that is performant and robust
A Rough Proposal
I’ve been thinking about this for a bit, and I think there are three ways in which we should be able to infer an associated type witness:
Associated type defaults, which are specified with the associated type itself, e.g.,
associatedtype Indices = DefaultIndices<Self>
These are easy to reason about for both the programmer and the compiler.
Typealiases in (possibly constrained) protocol extensions, e.g.,
extension RandomAccessCollection where Index : Strideable, Index.Stride == IndexDistance {
typealias RandomAccessCollection.Indices = CountableRange<Index>
}
I’m intentionally using some odd ‘.’ syntax here to indicate that this typealias is intended only to be found when trying to satisfy an associated type requirement, and is not a general typealias that could be found by normal name lookup. Let’s set the syntax bike shed aside for the moment. The primary advantage of this approach (vs. inferring Indices from “var Indices: CountableRange<Index>” in a constrained protocol extension) is that there’s a real typealias declaration that compiler and programmer alike can look at and reason about based just on the name “Indices”.
Note that this mechanism technically obviates the need for (1), in the same sense that default implementations in protocols <swift/GenericsManifesto.md at main · apple/swift · GitHub; are merely syntactic sugar.
Declarations within the nominal type declaration or extension that declares conformance to the protocol in question. This is generally the same approach as described in “associated type inference” above, where we match requirements of the protocol against declarations that could satisfy those requirements and infer associated types from there. However, I want to turn it around: instead of starting with the requirements of the protocol any looking basically anywhere in the type or any protocol to which it conforms (for implementations in protocol extensions), start with the declarations that the user explicitly wrote at the point of the conformance and look for requirements they might satisfy. For example, consider our initial example:
extension MyCollection: RandomAccessCollection {
var startIndex: Int { return contents.startIndex }
var endIndex: Int { return contents.endIndex }
subscript(index: Int) -> T { return contents[index] }
}
Since startIndex, endIndex, and subscript(_:) are declared in the same extension that declares conformance to RandomAccessIterator, we should look for requirements with the same name as these properties and subscript within RandomAccessCollection (or any protocol it inherits) and infer Index := Int and Element := T by matching the type signatures. This is still the most magical inference rule, because there is no declaration named “Index” or “Element” to look at. However, it is much narrower in scope than the current implementation, because it’s only going to reason from the (probably small) set of declarations that the user wrote alongside the conformance, so it’s more likely to be intentional. Note that this is again nudging programmers toward the style of programming where one puts one protocol conformance per extension, which is admittedly my personal preference.
Thoughts?
I think this approach is more predictable and more implementable than the current model. I’m curious whether the above makes sense to someone other than me, and whether it covers existing use cases well enough. Thoughts?
- Doug
. The only requirement I'd have is that a use case as described in SR-6516 [1] will work. That is