1098561 - 2.34% WinXP PGO Canvasmark regression on inbound (v.36) Nov 11th from push

Status: RESOLVED WORKSFORME

(Core :: Audio/Video, defect)

Description

[Joel Maher ( :jmaher ) (UTC -8)]

graph server shows a regression:
http://graphs.mozilla.org/graph.html#tests=%5B%5B289,63,37%5D%5D&sel=none&displayrange=7&datatype=running

I did a lot of retriggers on tbpl:
https://tbpl.mozilla.org/?tree=Mozilla-Inbound&fromchange=a02407bfde3b&tochange=540b81dfaf97&jobname=Windows%20XP.*talos%20chromez

This is the push where we stay right around or below 7000:
https://hg.mozilla.org/integration/mozilla-inbound/pushloghtml?changeset=7b6182fd38fc

Here is more information about the test:
https://wiki.mozilla.org/Buildbot/Talos/Tests#CanvasMark

This is interesting as this is a PGO only regression.  The opt version of the test yields no regression (and oddly enough opt produces better results on this specific test than PGO).

As this is a single test/single platform and a really low percentage regression, I wouldn't treat this as a top priority.  I would recommend seeing if there is something obvious that might be causing this though :)

Comment 1

[Matt Woodrow (:mattwoodrow)]

I can't see any way this code change could affect CanvasMark. The changed code only gets hit when playing videos using Media Source Extension, which definitely isn't happen during this test.

Is it possible that we just got unlucky and the PGO compiler is generating worse code in an unrelated area?

Comment 2

[Joel Maher ( :jmaher ) (UTC -8)]

that is very possible.  It has happened a few times in the past, shall we close this bug?