Tiny bit more progress on IRGen support for subclass existentials

[slavapestov]

The next step after #8770. Make sure we use the right reference counting style for subclass existentials and class-constrained archetypes. Also, finally remove the getExistentialTypeProtocols() utility method.

[slavapestov]

@swift-ci Please smoke test

[slavapestov]

Hey @jckarter and @rjmccall, why do we lower unowned(unsafe) references as "NativeObject.Type"? Does this have the right spare bits and extra inhabitants given that unowned(unsafe) can contain Objective-C tagged pointers?

[jckarter]

Good catch. It would need to be more conservative.

[slavapestov]

@jckarter @rjmccall This hack doesn't make sense -- we should just use the spare bits from the superclass constraint, if any, otherwise the spare bits of an unknown pointer, right?

[jckarter]

Yeah, that would work and be a bit more optimal. If we don't already, we should also check the @_unsafe_no_tagged_pointer attribute and whether ObjC interop is enabled at all.

[jrose-apple]

This assertion isn't valid; this code runs even when type-checking fails.

[slavapestov]

Do you have a test case in mind I can try? I'll address this in an upcoming patch along with the other stuff above about spare bits.

[jrose-apple]

I think even just the obvious thing would crash:

// compile with -emit-reference-dependencies and check the resulting swiftdeps file
class MyClass {}
extension Int: (MyClass & Equatable) {}

[slavapestov]

@swift-ci Please smoke test