REGRESSION (r132941): attribute modification 10% performance regression

Regression range: http://trac.webkit.org/log/?verbose=on&rev=132949&stop_rev=132924

http://trac.webkit.org/changeset/132941/ seems like the most likely culprit to me.

Most clear graph. Click on the part in the graph where the performance drops then click on the webkit link on the right to see the webkit blamelist:
http://build.chromium.org/f/chromium/perf/chromium-rel-win7-webkit/dromaeo_jslibattrjquery/report.html?rev=165231&graph=jslib_attr_jquery_jQuery___removeClass&trace=score&history=150

Some less clear graphs that point at the same webkit blamelist:
http://build.chromium.org/f/chromium/perf/chromium-rel-win7-webkit/dromaeo_jslibattrprototype/report.html?rev=165213&graph=jslib_attr_prototype_Prototype___removeClassName&trace=score&history=150
http://build.chromium.org/f/chromium/perf/chromium-rel-win7-webkit/dromaeo_jslibattrprototype/report.html?rev=165213&graph=jslib_attr_prototype_Prototype___writeAttribute&trace=score&history=150
http://build.chromium.org/f/chromium/perf/chromium-rel-win7-webkit/dromaeo_jslibattrjquery/report.html?rev=165213&graph=jslib_attr_jquery_jQuery___attr_class_test_&trace=score&history=150

Smaller regression range from the Chromium linux bot: http://trac.webkit.org/log/?verbose=on&rev=132944&stop_rev=132938

http://build.chromium.org/f/chromium/perf/linux-release-webkit-latest/dromaeo_jslibattrjquery/report.html?rev=165139&graph=jslib_attr_jquery_jQuery___attr_class_test_&trace=score&history=150
http://build.chromium.org/f/chromium/perf/linux-release-webkit-latest/dromaeo_jslibattrprototype/report.html?rev=165139&graph=jslib_attr_prototype_Prototype___removeClassName&trace=score&history=150

If it was http://trac.webkit.org/changeset/132941, I think http://trac.webkit.org/changeset/133021 should fix it.

(In reply to comment #2)
> If it was http://trac.webkit.org/changeset/132941, I think http://trac.webkit.org/changeset/133021 should fix it.

133021 did not fix it.

Not sure it will. We'll see. I think a sufficiently artificial test case might see a regression from that change. There are ways to improve if needed.

Turns out the difference finding algorithm using SpaceSplitString mutation is too slow:

Running Time    Self        Symbol Name
1714.0ms    5.6%    302,0          WebCore::Element::classAttributeChanged(WTF::AtomicString const&)
532.0ms    1.7%    174,0           WebCore::SpaceSplitString::remove(WTF::AtomicString const&)
277.0ms    0.9%    66,0            WebCore::SpaceSplitString::ensureUnique()
136.0ms    0.4%    136,0             WTF::fastMalloc(unsigned long)
75.0ms    0.2%    75,0             WTF::Vector<WTF::AtomicString, 4ul>::Vector(WTF::Vector<WTF::AtomicString, 4ul> const&)

Should be easy to optimize this away.

Created an attachment (id=171893) [details]
patch

(From update of attachment 171893 [details])
View in context: https://bugs.webkit.org/attachment.cgi?id=171893&action=review

> Source/WebCore/dom/Element.cpp:723
> +    StyleResolver* styleResolver = document()->styleResolverIfExists();
> +    bool testShouldInvalidateStyle = attached() && styleResolver && styleChangeType() < FullStyleChange;

I almost wonder if it's worth adding a static helper function here since this code is duplicated with Element::classAttributeChanged.

static StyleResolver* styleResolverForTestingStyleInvalidation(Document* document, Element* element)
{
    if (!element->attached() || element->needsStyleRecalc() || styleTypeChange >= FullStyleChange)
        return 0;
    return document()->styleResolverIfExists();
}

Then you wouldn't need the testShouldInvalidateStyle at all and could just null check the returned style resolver.

It's not clearly better to me, but it does avoid some duplication. Up to you.

> Source/WebCore/dom/Element.cpp:733
> +            if (shouldInvalidateStyle)
> +                testShouldInvalidateStyle = false;

I think you could avoid needing to do this if you put the check below into an else clause. That'd simplify the logic some I think.

} else {
    if (name == HTMLNames::nameAttr)
        setHasName(!newValue.isNull());
    shouldInvalidateStyle = ...;
}

> Source/WebCore/dom/Element.cpp:738
> +    shouldInvalidateStyle = testShouldInvalidateStyle && styleResolver->hasSelectorForAttribute(name.localName());

Not sure it matters, but we currently only do this if (!needsStyleRecalc()). Maybe testShouldInvalidateStyle should also && !needsStyleRecalc()? Obviously the same would apply in Element::classAttributeChanged. Saves a hashtable lookup. Not sure it matters in practice.

(In reply to comment #7)
> Not sure it matters, but we currently only do this if (!needsStyleRecalc()). Maybe testShouldInvalidateStyle should also && !needsStyleRecalc()? Obviously the same would apply in Element::classAttributeChanged. Saves a hashtable lookup. Not sure it matters in practice.

styleChangeType() < FullStyleChange is the equivalent test (more correct in fact).

(In reply to comment #8)
> styleChangeType() < FullStyleChange is the equivalent test (more correct in fact).

Ah. Got it.

(From update of attachment 171893 [details])
Attachment 171893 [details] did not pass chromium-ews (chromium-xvfb):
Output: http://queues.webkit.org/results/14686598

New failing tests:
css3/viewport-percentage-lengths/css3-viewport-percentage-lengths-getStyle.html
fast/dynamic/004.html

http://trac.webkit.org/changeset/133214 (with the test breaking bug fixed)

Lets see what the bots say.

Looks like it fixed it.

http://build.chromium.org/f/chromium/perf/chromium-rel-win7-webkit/dromaeo_jslibattrjquery/report.html?graph=jslib_attr_jquery_jQuery___removeClass&trace=score&history=100&rev=165526

Yup. Looks great!