> I reliably see gains from unrolling relatively simple loops on very recent cpus. Maybe you didn't have multiple of 4 uops in the loop? Or hit some other (?)

> I reliably see gains from unrolling relatively simple loops on very recent cpus. Maybe you didn't have multiple of 4 uops in the loop? Or hit some other (?) loop buffer limitation. Dependency chain splitting is also indeed a good reason to "unroll". Although it's a bit more complex transformation than simple unrolling.