This roughly halves stattimer overhead, from +80% to +40%
(on a very small benchmark; more complex benchmarks probably
have lower overhead)