Sema: Explicit stricter minimization of protocol composition types for redeclaration checking

[AnthonyLatsis]

Enable explicit stronger minimization of protocol compositions via generic signature computation for purposes like redeclaration checking.

Note: This is a source-breaking change, because the following is now a redeclaration error:

protocol P: class {}
func foo(_: P) {}
func foo(_: P & AnyObject) {}

[AnthonyLatsis]

@swift-ci please smoke test

[AnthonyLatsis]

@swift-ci please smoke test macOS

[AnthonyLatsis]

@swift-ci please smoke test macOS

[AnthonyLatsis]

@swift-ci please smoke test macOS

[AnthonyLatsis]

@swift-ci please smoke test macOS

[AnthonyLatsis]

@swift-ci please test source compatibility debug

[AnthonyLatsis]

@swift-ci please smoke test macOS

[slavapestov]

Is this also ABI-breaking potentially, if you previously had a redundant AnyObject constraint in a composition that is now minimized away?

[AnthonyLatsis]

Is this also ABI-breaking potentially, if you previously had a redundant AnyObject constraint in a composition that is now minimized away?

Yes. Does a smoke test not exclude any mangling differences?

[slavapestov]

It's pretty unlikely that someone has a binary framework that relies on this, but it might be safer if minimization was an explicit step and not part of getCanonicalType()?

[AnthonyLatsis]

You mean, to minimize selectively instead of having canonical types be truly minimal?

[slavapestov]

Yeah. I'm assuming this is for the type erasure when you use an existential member that returns an associated type -- you could minimize inside the member access path when computing the erased type of the member.

[AnthonyLatsis]

I'm assuming this is for the type erasure when you use an existential member that returns an associated type

Not anymore. That was the original plan, but it looks like using canonical types is enough to make the erasure work (we never retain sugar in generic signatures anyway, so there's no reason to separate minimization from canonicalization in this particular case).

If the ABI break is a concern, we can drop that part of the code and still benefit from the absence of "premature canonicalization".

[AnthonyLatsis]

@slavapestov What if we special-case protocol composition types in ASTMangler::getDeclTypeForMangling to preserve the mangling and make this ABI-compatible?

[slavapestov]

I'd prefer not to add any special cases and avoid changing getCanonicalType() to minimize protocol compositions. If you need to minimize protocol compositions for something, you can add a new method to ProtocolCompositionType (or ExistentialLayout or whatever) which performs it explicitly.

[AnthonyLatsis]

To be clear, are we ready to take on the source break if I make stricter minimization explicit and use it in redeclaration checking, as well as the other part of this PR (4602bf1)?

[AnthonyLatsis]

@swift-ci please smoke test macOS

[AnthonyLatsis]

@swift-ci please smoke test macOS

[AnthonyLatsis]

@swift-ci please smoke test

[AnthonyLatsis]

@slavapestov Could I ask for another opinion on this? I kept it narrow as per your suggestion.

[slavapestov]

Cool! Can this use ASTContext::getOpenedArchetypeSignature()?

[AnthonyLatsis]

@swift-ci please smoke test

[AnthonyLatsis]

@swift-ci please test source compatibility

[AnthonyLatsis]

@slavapestov Much obliged