[swift-evolution] [Pitch] SE-0083 revisited: removing bridging behavior from `as`/`is`/`as?` casts

[Joe Groff]

I’d like to investigate separating Objective-C bridging from the behavior of the as/as?/is operator family again for Swift 4. Last year, I proposed SE–0083 <https://github.com/apple/swift-evolution/blob/master/proposals/0083-remove-bridging-from-dynamic-casts.md>, but we deferred the proposal for lack of time to evaluate its impact. As complicating factors, we now have source compatibility with Swift 3 as a requirement, and the id-as-Any work from SE–0116 <https://github.com/apple/swift-evolution/blob/master/proposals/0116-id-as-any.md> more or less requires bridging dynamic casts to work. I think we can nonetheless make important improvements in this area in order to simplify the core language and provide more portable behavior across platforms with and without ObjC interop. In retrospect, submitting SE–0083 as an omnibus “fix casting” proposal was a mistake. We can separate out a few smaller subproblems from the overall concept:

Replacing as for bridging coercion

Swift 0 shipped with implicit conversions between standard library value types and their bridged Cocoa classes in both directions, and as we’ve eased off of the implicit conversions, we still left the as operator with the ability to force the conversions. This complicates the meaning of as: normally, it just provides type context, but it also has the power to force bridging conversions. These meanings are often at odds:

// `NSNumber` is `ExpressibleByIntegerLiteral`, so this gives type context to the literal 0
// and is equivalent to `NSNumber(integerLiteral: 0)`
0 as NSNumber

// `x` already has type `Int`, so this forces the bridging conversion and is equivalent to
// `_bridgeToObjectiveC(x)` (and thereby gives you a different kind of `NSNumber`!)
let x: Int = 0
x as NSNumber
Aside from the complexity and non-portability of this behavior, this is also inconsistent with the Swift naming conventions, which recommend that conversions between related types be presented as initializers. Additionally, the bridging conversions often have specialized behavior for performance or semantic reasons that aren’t intended to be exposed in the normal API of either type (for example, bridging a Swift number type to NSNumber produces a “type-preserving” instance of NSNumber so that the bridge doesn’t lose type information, even though NSNumber’s own API presents a type-agnostic numeric object). Therefore, I propose that we remove the bridging behavior from as, and provide APIs for conversion where they don’t yet exist. as is purely a compile-time construct with no runtime interaction, so the Swift 3 compatibility and ABI issues are much simpler than they are when runtime casting behavior becomes involved.

Warning on is/as?/as! casts that statically induce bridging

Without changing the runtime behavior of casting, we could still discourage users from using dynamic casting to perform bridging conversions when it’s statically evident that a bridging conversion is the only way a cast succeeds. For example:

func abuseBridgingCasts(on object: AnyObject) {
    // warning: dynamic cast requires a bridging conversion; use `Int(bridgedFrom:)` instead
    let _ = object as? Int
}
This wouldn’t be perfect, since we wouldn’t be able to warn about fully dynamic casts, but it could help encourage users to write portable code that doesn’t rely on the Objective-C bridge in common situations.

Limiting when the runtime allows bridging in dynamic casts

Ideally, we would be able to change runtime dynamic casting itself to not involve bridging. However, as I mentioned above, there are at least two situations where bridging dynamic casts are necessary to meet design requirements:

To maintain compatibility with Swift 3 code
For dynamic-casting Anys that were bridged from an Objective-C id, since we don’t know statically whether a bridge to any particular Swift type is needed
However, neither of these is an insurmountable barrier. For Swift 3 compatibility, if nothing else, we could ship a parallel set of runtime entry points for dynamic casting with the Swift 3 behavior that would be used when compiling code in Swift 3 mode (and those entry points could possibly be banished to a separate compatibility dylib to avoid weighing down the ABI-stable Swift runtime forever). For id-as-Any bridging, existentials could potentially carry a bit to indicate whether casting out of that particular value should admit bridging casts. Doing that has complications of its own—having otherwise equivalent Any values have different behavior is undeniably weird—but, if we can keep the behavior isolated to code that directly interfaces with ObjC, I think it’s worth investigating in the interest of making the overall language more predictable.

-Joe
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.swift.org/pipermail/swift-evolution/attachments/20170301/b4aa3049/attachment.html>