807408 - laggy text entry

Status: RESOLVED INACTIVE

(Core :: JavaScript Engine, defect)

Description

[(dormant account)]

Holdin down a key on repeat in bugzilla(while in balanced perf profile in win7) results in laggy text input. See http://people.mozilla.com/~bgirard/cleopatra/?report=9dd651fb9b9623449feae2e07cf9445b1a39464a

See http://people.mozilla.com/~bgirard/cleopatra/?report=deecf01f076502ca7529bd083d5f4ab0cd984fcc for similar lag in gmail

Comment 1

[(dormant account)]

Attachment: memchaser log for bill

Comment 2

[Jeff Muizelaar [:jrmuizel]]

Here's my thoughts on what's happening here.

Each key press is causing two paints:
1. from the refresh driver
2. from nsWindow::DispatchStarvedPaints

I'm not sure why we're getting two paints per keypress and that seems wrong. Why are we getting WM_PAINTs right after we've just painted? Where are they coming from?

The real badness it seems is in GC though. Every time we go into DispatchStarvedPaints we do two GCs and both are more expensive than the paint.

Comment 3

[Jeff Muizelaar [:jrmuizel]]

To be clear, about 86% of the time in starvedpaint is in GC

Comment 4

[Matt Woodrow (:mattwoodrow)]

Attachment: Cache the dummy target for no-composite transactions

Untested, but this should remove around 9% of the profile.

Comment 5

[Jeff Muizelaar [:jrmuizel]]

Comment on attachment 677163 [details] [diff] [review]
Cache the dummy target for no-composite transactions

Review of attachment 677163 [details] [diff] [review]:
-----------------------------------------------------------------

::: gfx/layers/basic/BasicLayerManager.cpp
@@ +541,5 @@
> +      // TODO: We should really just set mTarget to null and make sure we can handle that further down the call chain
> +      nsRefPtr<gfxASurface> surf = gfxPlatform::GetPlatform()->CreateOffscreenSurface(gfxIntSize(1, 1), gfxASurface::CONTENT_COLOR);
> +      mDummyTarget = new gfxContext(surf);
> +    }
> +    mTarget = mDummyTarget;

Is this missing a header change? If so it might be good to give mDummyTarget a more descriptive name like m1x1DummyTarget. You should also add a comment about why it's important enough to cache this surface.

Comment 6

[Matt Woodrow (:mattwoodrow)]

https://hg.mozilla.org/integration/mozilla-inbound/rev/7e2d29d8f825

Need to evaluate if this is still an issue with the above patch, and bug 807472.

Comment 7

[Ed Morley [:emorley]]

https://hg.mozilla.org/mozilla-central/rev/7e2d29d8f825