Created attachment 549374 [details] [diff] [review] v1 We could have done this ever since we dropped 10.4 compatibility. I don't know why I didn't think of this sooner. The current mouse move event targeting setup has two drawbacks: 1. It's unnecessarily CPU-intensive because we traverse all windows on the screen on every mouse move, even if we're not the active application, and even if all our own windows are hidden behind other apps' windows. 2. The system wide mouse move event monitor might be the reason for bug 611068. (If it's not, the mouse down event tap is.) We're taking these measures because [window setAcceptsMouseMoveEvents:YES] will always send the event to the current first responder, no matter which window the mouse is actually over. It also doesn't by default send events when our app is inactive. Well, NSTrackingArea changes all that and actually behaves like you'd expect it to: Mouse move events are always and only sent to the tracking area under the mouse, and you can even specify that you'd like to get events regardless of app activation status (NSTrackingActiveAlways). NSTrackingAreas have one imperfection: You don't get a mouse enter event if the mouse is already inside the area that's being created. But that's not so bad because we can correct the mouse enter state as soon as the mouse moves. Most of this patch is very straightforward: There's one tracking area per window, and it gets recreated whenever the window size changes. As for the rest: - The sendEvent override on PopupWindow has been unnecessary ever since we dropped 10.4 compatibility. I've extensively tested this ages ago but never got around to filing a bug about it. However, now it's even actively harmful because it routes mouse events to the first responder, so they don't get to the tracking area. That's why I'm removing it. - The testing story is tricky. Synthesizing mouse moves and getting tracking area events seems impossible, so the test simulates the expected events. I've also removed those parts of the test that sent mouse events to windows that are not under the mouse because those cases don't occur with tracking areas. I've tested this patch on 10.6 and 10.7 and couldn't find any regressions. I've gone through all the bugs in the blame range of the code I'm removing. This build contains the patch: https://email@example.com/try-macosx64/firefox-8.0a1.en-US.mac.dmg
This patch replaces some of the most unfortunate code in Cocoa widgets. Thanks for doing this!
I've backed this out to see whether it's responsible for the 60% Trace Malloc MaxHeap increase that can be seen on http://mzl.la/paIasC http://hg.mozilla.org/integration/mozilla-inbound/rev/6b943c8d7c5e http://hg.mozilla.org/integration/mozilla-inbound/rev/ca13b9114ce6
http://hg.mozilla.org/mozilla-central/rev/fee47b64b378 The backout changeset has not made it to m-c yet.
And relanded on inbound because it wasn't responsible. (Bug 668195 is.) http://hg.mozilla.org/integration/mozilla-inbound/rev/9c4561248f89
The backout made it to m-c: http://hg.mozilla.org/mozilla-central/rev/6b943c8d7c5e