Ok. Reproducible in x86 linux at home in a clean profile, sooo, that's not it.
Almost all the time (39000 of 41000 samples) is spent under nsViewManager::Refresh when things get slow. Of that Refresh time, 92% is under the nsRegion::Sub call that mozilla::layers::ThebesLayouerBuffer::BeginPaint makes (doing SubRect, which calls Copy and InsertInPlace, etc). Sounds like we end up with a more and more complex region, persisting over time, and we don't simplify it until a whole-window invalidate or something? That would be consistent with what nemo and I (him on Linux, me on Mac) see: blurring the Mozilla window drops the CPU usage.
blocking2.0: --- → ?
Assignee: nobody → roc
blocking2.0: ? → final+
Is this a regression?
D3D9/D3D10 ThebesLayers just assign to mValidRegion so don't need this fix.
Attachment #503028 - Flags: review?(jones.chris.g)
Whiteboard: [softblocker] → [softblocker][needs review]
Attachment #503028 - Flags: review?(jones.chris.g) → review+
Whiteboard: [softblocker][needs review] → [softblocker][needs landing]
Status: NEW → RESOLVED
Closed: 9 years ago
Resolution: --- → FIXED
Whiteboard: [softblocker][needs landing] → [softblocker]
You need to log in before you can comment on or make changes to this bug.