Posted Dec 19, 2010 3:39 UTC (Sun) by quotemstr
In reply to: Likely unlikely()s
Parent article: Likely unlikely()s
You're assuming that pcorrect and pincorrect are constant over time, and they're often not. CPUs these days have large global branch prediction buffers and can detect fairly complex patterns (e.g., branch every third time) automatically. The overall result of this dynamic prediction may be superior to a purely static assertion, even if that assertion is correct on average.
Think of a branch that goes one way during the day and another way at night. If we suppose that activity is twice as high during the day, then we'll see a 2:1 branch bias in favor of the daylight path. But over smaller intervals of time (say, an hour), the branch bias will be almost totally in favor of one path or the other. Unlike static assertions, the CPU's automatic prediction can look at the specific circumstances of the system and make predictions based on recent behavior and specific circumstances. Using static branch prediction could actually hurt.
to post comments)