All graphs for this bug:
  https://chromeperf.appspot.com/group_report?bug_id=794312

(For debugging:) Original alerts at time of bug-filing:
  https://chromeperf.appspot.com/group_report?sid=cbdaf1f44feec100244b6ee0c3e5cf6efd16e015a96d3d1423c1c4380a6b636a


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

chromium-rel-mac11

📍 Pinpoint job started.
https://pinpoint-dot-chromeperf.appspot.com/job/14944eca040000

😿 Pinpoint job stopped with an error.
https://pinpoint-dot-chromeperf.appspot.com/job/14944eca040000

Looks like https://github.com/catapult-project/catapult/issues/3921 , a failure to resolve the skia DEPS roll.

Luckily, the roll only has one commit.

Remove instanced rendering
By bsalomon@google.com · Fri Dec 08 13:41:10 2017 +0000
skia @ f06c3582677270c741a5e4fb366ec8b5d4ec4c21

This change delete code that was never called from Chrome. If it really is due to this change then it'd have to be something like code or class layout, vtable size, or something along those lines.

Who would be the right person to track down any of your hypotheses?

I suppose one could use profiling tools to look at cache behavior, etc on this or similarly configured machine any possibly identify what changed. But whether any one of those conjectures is true or not, this doesn't seem actionable as whatever the cause, we aren't likely to restore >4K lines of dead code to get this performance back. AFAICT a small number of tests on one Mac are affected. If this was more widespread or a larger performance change I'd think it'd be worth more investigation as maybe we could get this performance back shy of reverting this change. However, given the size of the potential win and the nebulousness of figuring out the root cause I'm inclined to think it isn't worth the opportunity cost. I did look into Skia's standalone performance metrics for this change and found a mix of negligible performance increases and decreases (https://perf.skia.org/e/?begin=1512682143&end=1512753192&keys=5662757143707648&xbaroffset=31590 and https://perf.skia.org/e/?begin=1512682143&end=1512753192&keys=5745745709760512&xbaroffset=31590).

Fair enough.