Sema: Infer 'dynamic' for overrides of imported methods

[slavapestov]

Fixes rdar://problem/31104529.

[slavapestov]

@swift-ci Please smoke test

[jrose-apple]

Is it worth limiting this to open methods?

[jckarter]

The "infer as dynamic" hack has had a bunch of annoying weird side effects when we used it for extensions to ObjC classes. Is that the best approach here? I don't think we want the full effects of late binding, we only want to funnel dispatch through the ObjC method dispatch interface.

[jrose-apple]

I think we do need the full effects of late binding—the point is that there may be overrides that cannot be seen by the program, possibly even in the base class.

[jckarter]

Is there any reason to prevent devirtualization, though, in situations where you can see the overrides?

[jrose-apple]

Devirtualizing the super call would not be okay. Devirtualizing a direct call probably is.

I don't suspect this will be a major hit since this is how Objective-C works all the time.

[jrose-apple]

(To be clear, I fully sympathize with the "bunch of weird side effects". This will trigger diagnostics about things being marked dynamic, for example. I would have preferred to do this later as well. But a lot of the semantics of dynamic, such as automatically being inherited, do match what we want here, and it's conservative.)

[jckarter]

Will final or static overrides gets the mystery "declaration cannot be both final and dynamic" error?

[jrose-apple]

That should already be checked for in this method, but adding a test case is good.