All graphs for this bug:
  https://chromeperf.appspot.com/group_report?bug_id=618239

Original alerts at time of bug-filing:
  https://chromeperf.appspot.com/group_report?keys=agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgICg3NCmuwoM


Bot(s) for this bug's original alert(s):

webrtc-mac-large-tests

kwiberg@, peah@, FYI: Slight improvement in APM performance due to "Add clz functions".

This looks expected, right?

As I see it:
1) The __builtin_clz?? intrinsics are available on mac, right? In that case, those should reduce the complexity.
2) The new code includes short-circuits when the input is 0, which should lower the complexity significantly for this case. If zero values dominates where the code is used, that is likely to give a lower complexity.

kwiberg@: WDYT? Does this make sense?

Yes, __builtin_clz?? is implemented by clang and gcc everywhere. They use supposedly good implementations for each platform; on at least ARM and x86, they're implemented with a single instruction.

I don't think that (2) helps much when we already have (1).

Since the old implementation was significantly longer than a single instruction and IIRC even had branches, I'm not surprised that the new one is faster. I didn't think we used it enough that it would be visible in an entire benchmark run, though.

See also  issue 617123 .