Automated tests showed an increase in JS time and pose prediction latency after change r507688 that improved framerate: 

https://chromeperf.appspot.com/report?sid=56bac5ac977a6eefbb6967b1345380d91bd6a937421af4b211d60f700e25ed3a

The difference seems to be that the JS VSync for the rAF callback seems to be triggered consistently a bit earlier in the frame lifecycle, so the JS side spends more time in the "wait for previous render to complete" phase.

It should be possible to reduce latency by triggering the rAF callback a bit later so that it can work with a fresher pose.

Also, the "wait for previous render to complete" step shouldn't count as JS processing time, it's only used to delay asking for the next VSync and this could be done on the VrShellGl side.