Fix crash when optimizing protocol composition

[zoecarver]

This patch fixes a compiler crash when trying to optimize a protocol composition between a class and protocol. This is a super simple fix where I simply add a condition to check if the nominal arg type is null. Here's an example program:

class ClassA<T> { }
protocol ProtocolA { }

class MainClass<H> {
    init(x: ClassA<H> & ProtocolA) { }
}

final class ClassB: ClassA<String> { }
extension ClassB: ProtocolA { }

_ = MainClass(x: ClassB())

Resolves #54035

Once this has been reviewed, can someone with commit access trigger swift-ci, please?

[CodaFi]

Does WMO make a difference for this test?

[zoecarver]

Yes, it only crashes with WMO.

[theblixguy]

Crashes for me with just -O as well. Also, I think the above invocation is not correct - it should be not %target-swift-frontend -O -whole-module-optimization %s -emit-sil.

[CodaFi]

This is equivalent to dyn_cast_or_null on the next line. But it sounds like this code needs to handle protocol compositions in a more principled way. It seems like there’s a missing case here.

[zoecarver]

What missing case is that?

[CodaFi]

Not sure @atrick is more familiar with this code.

[atrick]

LGTM. @CodaFi is right, dyn_cast_or_null is cleaner.

There are various scenarios related to protocol composition that we should be able to optimize. This optimization is not really equipped to handle them though so it's best to just bail here.

I filed a bug for these issues:
SR-11695: SILOptimizer should devirtualize calls to a protocol composition type

[atrick]

@swift-ci test

[atrick]

Review comments addressed in #28032

[zoecarver]

@atrick, thank you for the review and for fixing the review comments!