Bug 55626

Created attachment 84477 [details] A shark sample from earlier today showing about 40% in these two functions. Later samples showed the higher 50% which I quoted in the bug title. Either way, seems like this is the hot spot. :)

Maciej pointed out over IRC that replacing children with innerHTML (which is exactly what peacekeeper does) subverts the nextRenderer check, because the parent remains attached and all the kids have no renderers. Meaning this is then O(N^2) the number of dom nodes added via innerHTML. :(

I suspect that http://trac.webkit.org/changeset/75277 (from bug 49481) may have caused us to call nextRenderer() more often, and may have caused a regression on this benchmark, but having not tested I can't say for sure.

Preparing to fix the createRendererIfNeeded case with bug 55720

Maciej suggested computing nextRenderer from the previous renderer for the createRendererIfNeeded() case. However after a few attempts I'm not sure it's that easy. nextRenderer() avoids anonymous renderers and continuations, and gives you the next renderer relating to the position in the DOM (not necessarily relating to where something would be rendered. My most recent attempt: static RenderObject* fastNextRendererForCreateRenderer(Node* node) { // It is more common to have a previousRenderer than a nextRenderer since // we attach from first to last child in the list. RenderObject* previousRenderer = node->previousRenderer(); RenderObject* nextRenderer = 0; if (previousRenderer) { // previousRenderer may be wrapped in an anonymous renderer. if (previousRenderer->parent() != node->parentOrHostNode()->renderer()) return node->nextRenderer(); nextRenderer = previousRenderer->nextSibling(); } else nextRenderer = node->parentOrHostNode()->renderer()->firstChild(); // If we found an anonymous renderer fall back to a (slow) walk of the node list. if (nextRenderer && nextRenderer->isAnonymous()) return node->nextRenderer(); ASSERT(nextRenderer == node->nextRenderer()); if (!nextRenderer) return 0; ASSERT(nextRenderer != node->renderer()); ASSERT(nextRenderer->node()->parentOrHostNode() == node->parentOrHostNode()); return nextRenderer; } Fails editing/deleting/pruning-after-merge-1.html (And a few other editing tests) still.

It might be possible to re-write addChild to compute nextRenderer itself in the few places it may be needed. (It certainly could reach through renderer->node()->nextRenderer() in all non-anonymous cases). Or we could create a RendererAddRequest object to hold nextRenderer in the few places it's actually used. Interestingly enough, when I first did a niave implementation of this fastNextRenderer, it disappeared off of the peacekeeper shark sample, but the peacekeeper numbers didn't get any better. Making me wonder if shark is leading me astray.

(In reply to comment #6) > Interestingly enough, when I first did a niave implementation of this fastNextRenderer, it disappeared off of the peacekeeper shark sample, but the peacekeeper numbers didn't get any better. Making me wonder if shark is leading me astray. Someone who wanted to work on this bug could and *should* confirm that, by simply replacing the nextRenderer() call in createRendererIfNecessary with "0" and re-running peacekeeper. I saw no change in time on my machine, *despite* the previously 25% of the sample which was in nextRenderer() disappearing. However it's possible I misread the peacekeeper numbers. My current theory is that Shark's 25% attribution to nextRenderer() is a bug.

Took test case from Mozilla bug - https://bug922016.bmoattachments.org/attachment.cgi?id=811946 *** Firefox Nightly 126 *** 161 ms *** Safari Technology Preview 191 *** 118 ms *** Chrome Canary 125 *** 191 ms ____ @Rysouke - do we need to squeeze anymore performance from this old benchmark or Speedometer would cover these and this can be closed?

We're no longer tracking this benchmark.