[SR-1779] Generic Objective-C classes cannot conform to protocols with associated types

[swift-ci]

Previous ID	SR-1779
Radar	rdar://problem/27652431
Original Reporter	austin (JIRA User)
Type	Bug

Attachment: Download

Environment

OS X 10.11.5, Xcode 8S128d

Additional Detail from JIRA

Votes	0
Component/s	Compiler
Labels	Bug, 3.0Regression, CompilerCrash
Assignee	@jckarter
Priority	Medium

md5: e97d74f501ffd9cfb45f8b3fe956579f

Issue Description:

I've attached an Xcode 8b1 project that contains a reproduction case. You can trigger the compiler segfault by attempting to run the unit tests (building the framework target alone isn't sufficient).

This bug is a little complicated. Here is the setup:

1. We have a framework target written in Objective-C. We've defined a generic Objective-C class `C` that conforms to an Objective-C protocol `P`.

2. To make things easier for our users, we provide a separate Swift file that is intended to provide additional functionality by making `C` conform to Swift's `Sequence` protocol.

2b. That file is not part of the framework target. In order to stand in for our users' applications, the file belongs to the unit test target, which acts like an application. The unit test target has a bridging header that pulls in our framework's Objective-C sources, and it also builds the aforementioned Swift file to provide the enhanced functionality.

3. That Swift file contains two things: a non-generic Swift class `Z` that is supposed to act as the iterator, and an extension to `C` to conform it to `Sequence` (by using `Z` as its iterator).

4. `Z` initializer takes in an argument of type `P`. The extension on `C` attempts to construct a `Z` iterator by passing in `self` to the initializer (since `C : P`, this should work).

5. Instead, the compiler crashes. I've attached the output as well in the text file.

[belkadan]

@slavapestov, look familiar?

[slavapestov]

Yeah, I think this is the issue where allocating constructors cannot be called from an extension of an Obj-C generic class.

[swift-ci]

Comment by Austin Zheng (JIRA)

I think so too; the following is enough to trigger the same issue:

final class TestClass {
    var a : Int = 0
}

extension SomeObjCGenericClass {
    func doSomething() {
        let _ = TestClass()
    }
}

[swift-ci]

Comment by Austin Zheng (JIRA)

If there's another Jira ticket tracking the issue, I'm happy to close this one as a duplicate; otherwise I can rename it to be more correct.

[bdash]

This is still an issue with Swift 3 in Xcode 8 beta 3.

[bdash]

And with Xcode 8 beta 4, too.

[slavapestov]

No longer crashes on master.

[slavapestov]

Just verifying this again. With beta 4, I get this error:

LLVM ERROR: unsupported relocation with subtraction expression, symbol '__TMnCSo24RLMZeroElementCollection' can not be undefined in a subtraction expression
2016-08-08 12:01:15.916 xcodebuild[37128:7937400]  DVTAssertions: Warning in /Library/Caches/com.apple.xbs/Sources/IDEFrameworks/IDEFrameworks-11222.4/IDEFoundation/Playgrounds/IDEPlaygroundAuxiliarySourceCompilerOperation.m:383
Details:  Unable to read diagnostics from file "/Users/slava/Downloads/CrashReproCase/build/CrashReproCase.build/Release/CrashReproCaseTests.build/Objects-normal/x86_64/Extensions.dia" (Invalid File): Invalid diagnostics signature
Function: void XCGenerateDiagnosticsFromFile(NSString *__strong, NSString *__strong, NSDictionary *__strong, NSDictionary *__strong, IDEActivityLogSectionRecorder *__strong, BOOL (^__strong)(IDEActivityLogMessage *__strong))
Thread:   <NSThread: 0x7fdf4d3913b0>{number = 6, name = (null)}

On master, I get this:

/Users/slava/Downloads/CrashReproCase/CrashReproCase/Extensions.swift:27:15: error: extension of a generic Objective-C class cannot access the class's generic parameters at runtime
  public func makeIterator() -> RLMUniversalIterator {
              ^
/Users/slava/Downloads/CrashReproCase/CrashReproCase/Extensions.swift:29:33: note: generic parameter used here
    return RLMUniversalIterator(self)
                                ^

I'm assuming this is an unsupported case where the associated type depends on Objective-C generic parameters, which is something we'd like to allow eventually, but not yet.

[bdash]

Xcode 8b5 still crashes as well.

Slava, is there a workaround for the unsupported case you describe? I'm not clear how else we can make an Objective-C collection that uses Objective-C generics for its element type conform to Swift's Sequence protocol.

[jckarter]

8b5 is the same compiler as 8b4. You'd need to verify using a snapshot for now.

[slavapestov]

@jckarter any ideas re: workaround?

[jckarter]

It looks like we're overly aggressive with that error now. I noticed the same while investigating another crasher. Does it help if you cast self as! ClassName<AnyObject> first thing?

[bdash]

It does not appear to make any difference:

CrashReproCase/Extensions.swift:27:15: error: extension of a generic Objective-C class cannot access the class's generic parameters at runtime
  public func makeIterator() -> RLMUniversalIterator {
              ^
CrashReproCase/Extensions.swift:29:21: note: generic parameter used here
    let typedSelf = self as! RLMZeroElementCollection<NSObject>
                    ^

I used NSObject as that's necessary to satisfy the generic constraint of the Objective-C class. Removing the constraint and using AnyObject in the cast also fails in the same way.

[bdash]

I should note that I was testing that change with swift-DEVELOPMENT-SNAPSHOT-2016-08-07-a-osx.pkg.

[swift-ci]

Comment by Xuan (JIRA)

may I ask the progress? As Apple is holding the event on Sep 7th, iOS 10 / swift 3.0 should come in Sep or early Oct, however this bug is still open, and seems not getting much attention.

[jckarter]

We don't have time to fully implement this before 3.0. However, the workarounds, such as using AnyObject as your associated type and introducing as! casts to change the generic parameters, should be functioning in the latest toolchains.

[jadengeller]

I ran into this same issue conforming a generic Objective-C type to a Swift protocol with no associated types (and no requirements either). Adding an `@objc` annotation to the protocol fixed it, but it seems that the problem is broader than what is mentioned here.