https://treeherder.mozilla.org/perf.html#/graphs?series=autoland,1646251,1,1&series=autoland,1683323,1,1&series=autoland,1651466,1,1&series=autoland,1684581,1,1 sessionrestore performance is worse with WR enabled on both Linux (slightly) and Windows (more). Marking P1 for windows regression.

This got worse on linux64-qr with bug 1471962.

Bug 1476876 helped on windows, but it's still worse with WR.

We can't release this to the field, but we can let this ride to beta. However, we want to investigate and understand the bad perf numbers asap.

Profile: https://perfht.ml/2Mma8PW It looks like this is also shader compiling causing a long first composite. The Renderer thread has a long period of being idle before it receives the first paint and starts compiling shaders. Once we have shared shaders between windows (bug 1416650), then we should also be able to compile those shared shaders earlier, during the period that is currently idle. The idle period is longer than the long-composite, so moving all that time into the gap should mean we're fully ready to go when the composite request arrives. Interestingly we're also compiling shaders for our D3D11 Non-WR compositor (though on a different thread), and it would be nice to stop doing that.

(In reply to Matt Woodrow (:mattwoodrow) from comment #4) > > Interestingly we're also compiling shaders for our D3D11 Non-WR compositor > (though on a different thread), and it would be nice to stop doing that. We need to make sure that fallback from WR to D3D11 Non-WR compositor works if we stop doing it.

Bug 1490528 is created for comment 5.

With the shader compilation stuff and other stuff that landed recently, this is fixed for windows.