All graphs for this bug:
  https://chromeperf.appspot.com/group_report?bug_id=615027

Original alerts at time of bug-filing:
  https://chromeperf.appspot.com/group_report?keys=agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgICgrMPluAkM,agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgICgrOjpvgoM,agxzfmNocm9tZXBlcmZyFAsSB0Fub21hbHkYgICgrP3wqAoM


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

webrtc-mac-large-tests

Nisse, one of your patches is also pretty close to this one.

It's not totally unreasonable that https://codereview.webrtc.org/2002523002 would improve performance. As I note in the commit message, "Also, the new [WebRtcSpl_AddSatW32 and WebRtcSpl_SubSatW32] implementations, unlike the old ones, result in branch-free code (tested with clang 3.9 with -O2)."

That said, unless the tests are spending a large fraction of their time doing fixed-point audio math, I would certainly not expect that CL to have such a large impact.

My cl above should affect android only, so that seems unrelated.

Hmm. How can I reproduce this locally? I'd like to test with and without my CL.

Closing this, since it seems unlikely that investigating a somewhat unexpected minor performance increase is more important than fixing other bugs. Anyone who disagrees, feel free to re-open...