Sema: Don't complain about internal type aliases referenced from @inlinable functions in Swift < 4.2

[slavapestov]

No description provided.

[slavapestov]

@swift-ci Please smoke test

[jrose-apple]

Hm. :-/ What happens in Swift 4.0 mode when Swift 5 come out?

[slavapestov]

I’m not sure I follow?

[jrose-apple]

I thought the reason we were adding these warnings is because they could lead to linker errors in Swift 5. Maybe I'm misremembering.

[slavapestov]

They can lead to linker errors in any release. Here is a test case:

struct Internal {}
typealias Alias = Internal
@inlinable func f() {
  print(Alias.self)
}

In 4.1, this produced no warning and failed to link when another module inlined f(). In Swift 5 mode this is an error. In 4.2 it's just a warning.

The reason it's just a warning is because there are valid code patterns you can write too, eg the underlying type of the Alias is @usableFromInline:

@usableFromInline struct Internal {}
typealias Alias = Internal
@inlinable func f() {
  print(Alias.self)
}

The correct way to write the above is to write, the following, which is the only variant that compiles without errors in Swift 5 mode:

@usableFromInline struct Internal {}
@usableFromInline typealias Alias = Internal
@inlinable func f() {
  print(Alias.self)
}

However, in Swift 4.1, @usableFromInline was not allowed on type aliases at all. SwiftNIO wants to have the same source compile warning-free on both 4.1 and 4.2. So they're just not declaring their type aliases as @usableFromInline and promising to ensure the underlying type has the right visibility.

Once 4.2 is GM, I expect they will drop support for 4.1, and I will go back to emitting these warnings unconditionally. (And in Swift 5 mode it's an error as always).

[jrose-apple]

Thanks for the explanation.