white space atomization during parsing is expensive

• Component: DOM
• Status: RESOLVED FIXED

[Cameron McCormack (:heycam)]

We have a memory optimization where the HTML parser will atomize any text node string that is all white space.  The process for this is a bit expensive, since we must loop over all the characters in the string three times:

* First, to check that all the characters are white space
* Second, to hash the string when looking up the atom hash table
* Third, to check the string for equality with any existing atom hash table entry

Most white space strings we encounter have a limited form -- they have at most three or four runs of consecutive equal white space characters, e.g. it's common to see a newline followed by a number of space characters.  We can take advantage of this by compressing the white space string into a simple run-length encoded form while we check that the string is entirely white space.  If we keep a cache of recently atomized white space strings that can be quickly looked up, keyed off the encoded form, we can re-use a previous result of atomizing an identical string and avoid the hashing and hash entry equality checks.

I have a WIP patch for this that is showing a 1% improvement on Speedometer 2 overall (due to 2-4% improvements on a few of the subtests) and no change to PLT5 on my local machine.

[Radar WebKit Bug Importer]

<rdar://problem/81601565>

[Cameron McCormack (:heycam)]

Created attachment 435057 [details]
Patch

[Cameron McCormack (:heycam)]

Created attachment 435138 [details]
Patch

[Yusuke Suzuki]

Comment on attachment 435138 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=435138&action=review

r=me

> Source/WebCore/html/parser/HTMLConstructionSite.cpp:858
> +        code |= (end - startOfRun);

It should be `character - startOfRun`.

[Cameron McCormack (:heycam)]

Created attachment 435169 [details]
Patch

[EWS]

Committed r280772 (240356@main): <https://commits.webkit.org/240356@main>

All reviewed patches have been landed. Closing bug and clearing flags on attachment 435169 [details].

[Yusuke Suzuki]

Committed r280773 (240357@main): <https://commits.webkit.org/240357@main>

[Per Arne Vollan]

Comment on attachment 435169 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=435169&action=review

> Source/WebCore/html/parser/HTMLConstructionSite.cpp:896
> +        WTFLogAlways("reuse code %llx", code);

Will this create a lot of logging?

> Source/WebCore/html/parser/HTMLConstructionSite.cpp:901
> +        WTFLogAlways("override");

Ditto.

> Source/WebCore/html/parser/HTMLConstructionSite.cpp:906
> +        WTFLogAlways("replace code %llx", code);

Ditto.

> Source/WebCore/html/parser/HTMLConstructionSite.cpp:913
> +    WTFLogAlways("new code %llx", code);

Ditto.

[Cameron McCormack (:heycam)]

(In reply to Per Arne Vollan from comment #8)
> Comment on attachment 435169 [details]
> Patch
> 
> View in context:
> https://bugs.webkit.org/attachment.cgi?id=435169&action=review
> 
> > Source/WebCore/html/parser/HTMLConstructionSite.cpp:896
> > +        WTFLogAlways("reuse code %llx", code);
> 
> Will this create a lot of logging?

Yes, Yusuke removed them shortly after this patch landed.