NFC: better CanType perf

[davezarzycki]

These two commits improve the Release+Assert build performance of Swift.o by about 5.4% on my machine. This matters less without asserts enabled.

[davezarzycki]

@swift-ci please smoke test

[CodaFi]

How much of the perf improvement is from dropping llvm::PointerUnion?

[jrose-apple]

I imagine the win is that you no longer have to mask off the low bit to dereference the pointer, not anything else in PointerUnion.

[jrose-apple]

…oh, right, assertions. Never mind, question still valid.

[davezarzycki]

Dropping this specific llvm::PointerUnion yields:

• 0.83% faster on average (wall clock).
• 2.36% fewer branches

Please also note that PointerUnion isn't bad per se, we just have spare bits elsewhere and this is a critical path.

[CodaFi]

Excellent work. Gonna run full tests first because this is a core abstraction.

[CodaFi]

@swift-ci please test

[davezarzycki]

I always run ninja check-swift-validation on my Linux box before pushing. I'd be surprised if anything breaks.