Post #1643525

LLVM is so afraid of optimizing around atomics that it is more pessimistic about an inlined atomic than an extern function: https://discourse.llvm.org/t/missed-store-forwarding-in-the-presence-of-atomic-operations/90653