Bug 129065 – Render tree construction is O(N^2) in number of siblings

• Component: Layout and Rendering
• Status: RESOLVED FIXED

[Maciej Stachowiak]

Render tree construction is O(N^2) in the number of sibling nodes.

Here is a test case and JSFiddle link demonstrating the issue; if you create 10x as many nodes, the time growth is approximately quadratic, rather than linear.

var t = Date.now();
for (var i = 0; i < 5000; ++i)
  document.body.appendChild(document.createElement('div'));
document.body.offsetTop;
document.body.textContent = Date.now() - t;

http://jsfiddle.net/vwG2J/2/


Blink fixed this by assembling the Render tree (approximately) backwards:
https://codereview.chromium.org/24350009


I am not sure offhand if this is the best or only solution.

[Maciej Stachowiak]

<rdar://problem/16114572>

[Antti Koivisto]

Created attachment 227760 [details]
patch

[Darin Adler]

Comment on attachment 227760 [details]
patch

View in context: https://bugs.webkit.org/attachment.cgi?id=227760&action=review

The approach here looks good. The details are a little confusing.

> Source/WebCore/style/StyleResolveTree.cpp:7
> + * Copyright (C) 2004, 2005, 2006, 2007, 2008, 2009, 2010, 2012, 2013, 2014 Apple Inc. All rights reserved.

We learned recently that year ranges are OK. So this should say:

    2004-2010, 2012-2014

> Source/WebCore/style/StyleResolveTree.cpp:75
> +    RenderObject* before;

The name “before” here is not so great. We name the other member “parent”, not “above” or “containing”, despite the fact that it’s not really the parent of anything. I think this might be clearer as “nextSibling”.

> Source/WebCore/style/StyleResolveTree.cpp:76
> +    bool isValid;

I don’t think that from this name it’s clear that isValid is specifically about the before field. I think it should probably be named nextSiblingPointerIsValid.

I find that the rules for when a RenderTreePosition is valid and when it might not be are confusing.

> Source/WebCore/style/StyleResolveTree.cpp:78
> +    unsigned assertCount;

I would call this assertionLimitCounter.

> Source/WebCore/style/StyleResolveTree.cpp:237
> +static void computeRenderTreePositionIfNeeded(RenderTreePosition& renderTreePosition, const Node& node, ContainerNode& renderingParentNode)

I don’t think the IfNeeded suffix in these two function names helps. I would just take it out.

> Source/WebCore/style/StyleResolveTree.cpp:249
> +static void invalidateRenderTreePositionIfNeeded(RenderTreePosition& renderTreePosition, const Node& node)

I don’t think the IfNeeded suffix in these two function names helps. I would just take it out.

> Source/WebCore/style/StyleResolveTree.cpp:280
> +    ASSERT(renderTreePosition.isValid);

Making the position a little more of a class would be good. Then the assertion that the pointer is valid could be a part of the getter rather than something we assert here, but use elsewhere.

> Source/WebCore/style/StyleResolveTree.cpp:285
> +        insertionPosition.before = parentFlowRenderer->nextRendererForElement(element);

It’s strange and subtle that we rely on the fact that isValid is already set here.

> LayoutTests/fast/css/sibling-renderer-On2.html:11
> +shouldBe("timeRenderTreeCreation(5000) < 20 * timeRenderTreeCreation(500)", "true");

I’m surprised you made a custom test here. I would have suggested instead doing a LayoutTests/perf test; those are designed to check that things are linear.

But perhaps the number "20" here is related to the the number "20" in the assertion limit count for debug builds?

[Antti Koivisto]

> I’m surprised you made a custom test here. I would have suggested instead doing a LayoutTests/perf test; those are designed to check that things are linear.

I didn't know about LayoutTests/perf. It is not immediately obvious to me how they are better but I'll move/refactor the test.

> But perhaps the number "20" here is related to the the number "20" in the assertion limit count for debug builds?

No it isn't.

[Antti Koivisto]

https://trac.webkit.org/r166303

Factored RenderTreePosition to be more of a class.
Moved the test to LayoutTests/perf but didn't use magnitude-perf.js as it seemed unsuitable for a something that requires a relatively long testing function (it calls the function many times).