I've been playing around with AsyncSequence which has now been updated to have a Failure associatedtype, with availability restricted to the latest platform versions. But I encountered something curious. The flatMap operator looks like this:
@available(macOS 10.15, iOS 13.0, watchOS 6.0, tvOS 13.0, *)
extension AsyncSequence {
...
@preconcurrency @inlinable public func flatMap<SegmentOfResult>(_ transform: @escaping @Sendable (Self.Element) async -> SegmentOfResult) -> AsyncFlatMapSequence<Self, SegmentOfResult> where SegmentOfResult : AsyncSequence, Self.Failure == SegmentOfResult.Failure
}
Notice there is no restriction to latest platform versions (it's only restricted to versions that support concurrency), but it does make use of Failure, which if I follow to its definition, is restricted to the latest platform versions.
I discovered this after I upgraded Xcode (I can't remember if the problem surfaced before or after changing the language version to 6), in a library where I wrote a flatten operator (i.e. flatMap { $0 }). The compiler complained that there were no matching overloads, and that the correct one isn't a candidate because it requires the constraint on Failure. So I added that constraint, and got an error that Failure isn't available on the target platform version. I eventually had to change the return type to AsyncThrowingFlatMapSequence, and write another version of flatten that is restricted to latest platform versions.
It was possible previously for it to return AsyncFlatMapSequence because AsyncSequence was @rethrows, but now this seems to not be possible anymore, which means it's a breaking change.
How is it possible that this extension is written so that it is available on older platforms but uses Failure? Is that a magic capability available only to the standard or "core" modules like _Concurrency?
Xcode let slip while showing possible overloads that the function is actually marked with @_alwaysEmitIntoClient, so I thought maybe that is why it works (that it would defer the restriction up to whoever imports the module), but I tried adding that and it didn't get rid of the error.