[llvm] InferAddressSpaces: Handle llvm.fake.use (PR #109567)

[Stephen Tozer via llvm-commits]

SLTozer wrote:

> I do not think this is a reasonable restriction to have.

That might be fine - I'm not the original author so I don't the justification for the restriction when it was written, but it's likely simply that we never produce fake use intrinsics with more than one argument, and so code that handles it uses the hardcoded assumption that there is one argument; now that this intrinsic is part of upstream LLVM, allowing it to be used as a full variadic may be reasonable unless there are other technical constraints.

https://github.com/llvm/llvm-project/pull/109567