Recently I've found that uncontended atomics are less expensive than what back-to-back tests would suggest, and so the real-world performance of TLS vs unconten

Recently I've found that uncontended atomics are less expensive than what back-to-back tests would suggest, and so the real-world performance of TLS vs uncontended atomics might be close than the numbers in the article would imply (assuming you aren't actually in the back-to-back [1] scenario). If you really want to gold plate things, you could imagine a hybrid strategy that initially uses a shared CAS-based counter for new threads, but moves some threads to a private TLS counter if they are using it heavily for some definition of heavily. Revocation, where you decide that a thread no longer deserves a dedicated counter, is more tricky! --- [1] Roughly, "not back to back" means that your atomics are spaced out by more than 20 cycles or so.