cc @rudkx

The behavior you're seeing is a result of the change for https://bugs.swift.org/browse/SR-4248.

The behavior now matches what you would get if you conditionally cast to the same type in a non-generic context.

It would have probably been best to put this under -swift-version 5, but apparently the implementor and reviewer did not catch the fact that it would change behavior in this way.

Ah yup, this one's my bad. I didn't realise the (now obvious with hindsight) compatibility implications. Would it be worth restoring the old behaviour for Swift 5 under Swift 4 compat mode?

Under the new behaviour, the compiler will assume that a generic placeholder could be a type with an arbitrary level of optionality, and so therefore won't eagerly unwrap the value before performing the cast – it will instead leave any unwrapping to the runtime. And because you're casting an Any? to Any, the runtime will happily box the value in Any and then promote that to Any? as the successful result of the conditional cast. As Mark says, this matches the behaviour of performing the cast in a non-generic context.

To get the desired behaviour for your function, you can either cast to the optional type and then coalesce nil:

func foo<Value>(_ x: Any?, as type: Value.Type) -> Value? {
  return (x as? Value?) ?? nil
}

Or unwrap x before performing the cast – for example, using flatMap:

func foo<Value>(_ x: Any?, as type: Value.Type) -> Value? {
  return x.flatMap { $0 as? Value }
}

Given that this affects Xcode 10 beta, which is Swift 4.2, not Swift 5, I would say the question should be whether old behavior should be restored in Swift 4 mode and new behavior used in Swift 4.2 mode (and I think the answer should be "yes" because this is a rather surprising behavior change to get just by upgrading Xcode 10 while keeping Swift 4 mode).

Unfortunately, I believe the ship has sailed for making such changes to the 4.2 compiler, so the most we can do is restore the behaviour for the Swift 5 compiler under Swift 4 compat mode.

It's a compatibility regression, so I'm not sure how Ted and co would feel about taking it for a 4.2.1 release.

@belkadan That would be great if we could! Would it be possible to also consider #18488 & #18952 for such a release?

It depends how risky you and your reviewers think the changes are. If you think it's worth doing, you can follow the same process as before to get changes into the 4.2 branch, and Ted will either approve or deny the change.

(See https://forums.swift.org/t/swift-4-2-in-final-convergence-swift-4-2-branch-open-for-post-4-2-0-changes/15128.)

Opened a PR to limit the new casting behaviour to Swift 5 mode: #19217

Comment by Evgeny Kaz (JIRA)

@hamishknight But what the philosophy of Optionals in swift says about it? Is this new behaviour right or not? Shall we rely on it?

I think the new behavior is perfectly reasonable, it's just the fact that it's a potentially-breaking change that wasn't good.

Comment by Evgeny Kaz (JIRA)

Eridius (JIRA User) I feel like it should not be "reasonable" - it is kinda crucial for the language concept of optional and nil coalescing. It should be described very clearly in the documentation. Look an the next code:

import UIKit

protocol Factory {
    
    associatedtype ViewController: UIViewController
    
    associatedtype Context

    func build(with context: Context) -> ViewController

}

struct StringFactory: Factory {

    func build(with context: String?) -> UINavigationController {
        return UINavigationController(nibName: nil, bundle: nil)
    }

}

struct FactoryBox<F: Factory> {
    
    let factory: F
    
    init(_ factory: F) {
        self.factory = factory
    }
    
    func build(with context: Any?) -> UIViewController? {
        guard let typedContext = context as? F.Context else {
            print("Context is not supported")
            return nil
        }
        return factory.build(with: typedContext)
    }
}

StringFactory().build(with: nil) // Returns UINavigationController in both XCode 9 and 10
FactoryBox(StringFactory()).build(with: nil) // Returns nil in XCode 9 printing "Context is not supported", returns UINavigationController in XCode 10

As for me - the new behaviour is reasonable, but on other hand a construction like `guard let b = a as? Whatever else ...` becomes a bit confusing. What exactly is b?

Resolved by #19217 which preserves the old behaviour in Swift 5 under Swift 4/4.2 compat mode.

I have also opened #19562 to hopefully get this into 4.2.1.

I confirmed that the behavior described in this issue has been changed on Swift 5 with -swift-version 4.2

$ pbpaste
func foo<Value>(_ x: Any?, as type: Value.Type) -> Value? {
    return x as? Value
}

$ pbpaste|xcrun --toolchain org.swift.42320190228a swift -
▿ Optional(nil)
  - some: nil
$ pbpaste|xcrun --toolchain org.swift.5020190306a swift -
▿ Optional(nil)
  - some: nil
$ pbpaste|xcrun --toolchain org.swift.5020190306a swift -swift-version 4.2 -
- nil

But is it correct behavior that using `-swift-version 4.2` makes Swift 5 to incompatible with Swift 4.2?

@norio-nomura That is correct, albeit unfortunate. It's worth noting that there would be still be incompatibility between Swift 4.2 and Swift 5 with `-swift-version 4` even if we preserved the behavioural change in `-swift-version 4.2`.