As step one of flipping on downscale-during-decode, let's flip on decode-on-draw on all platforms with APZ enabled. ImageFactory refuses to use decode-on-draw even if the pref is on if APZ is *disabled*, so actually this just means flipping the decode-on-draw pref on everywhere. It will then automatically be enabled on APZ platforms (including OS X when the appropriate pref is turned on). I want to flip this on one day before I flip on downscale-during-decode so there is a Nightly that has decode-on-draw but not downscale-during-decode. This should help clarify where any perf regressions actually come from.

Here's the patch. Changes nothing but prefs.

Try job here: https://tbpl.mozilla.org/?tree=Try&rev=c6e7518fca58

Comment on attachment 8552264 [details] [diff] [review] Flip on decode-on-draw on APZ platforms This patch is confusing to review because: 1) that it affects only APZC is hidden 2) we've had something that we've called "decode on draw" for a while, the new wrinkle here is that we try to avoid decoding except for on draw.

That try job was not based on current trunk. Cancelled and pushed a new one: https://tbpl.mozilla.org/?tree=Try&rev=17c0fcec4b5a

Thanks for the review! I *think* this try job looks OK. There are a couple of odd failures there, but nothing that seemed logically connected to this patch or that I could get to repeat with retriggers.

Pushed: https://hg.mozilla.org/integration/mozilla-inbound/rev/bed02f045941

We've been hitting frequent B2G reftest-20 failures today like the log below shows. Looking back around the time it started, this seems like the most-likely candidate. Rather than backing out the entire push, I'm reverting only the B2G change for now. Maybe better to just spin-off B2G to a new bug at this point. Otherwise, feel free to set the leave-open keyword. https://treeherder.mozilla.org/logviewer.html#?job_id=5715834&repo=mozilla-inbound https://hg.mozilla.org/integration/mozilla-inbound/rev/27648f940644

Confirmed that the reftest failures went away on B2G.

Let's see how difficult it is to fix that issue. For now, I'll go with leave-open. Thanks for not backing out on Android; I pushed this partly to get feedback about performance effects, and Android is the platform I expect to see that feedback on.

https://hg.mozilla.org/mozilla-central/rev/bed02f045941 https://hg.mozilla.org/mozilla-central/rev/27648f940644

It turned out to be hard to fix, so I will indeed spin off a new bug, where I'll re-push the B2G pref change.

Removing leave-open keyword from resolved bugs, per :sylvestre.