[Chromium][Performance] Optimize innerText and outerText in Chromium/Mac

• Component: HTML Editing
• Status: RESOLVED FIXED

[Kentaro Hara]

Created attachment 131989 [details]
Performance test

In Mac, HTMLElement.innerText in Chromium/V8 is 3~ times slower than HTMLElement.innerText in AppleWebKit/JavaScriptCore. We should optimize it.

The results of the attached performance test in my local Mac environemnt are as follows:

- Chromium/V8: 10250.8 ms
- AppleWebKit/JavaScriptCore: 3075.4 ms

[Kentaro Hara]

Created attachment 131990 [details]
Patch

[Kentaro Hara]

Created attachment 131994 [details]
Updated performance test

The change improves the performance of outerText too. Here are the results in my local Mac:

- AppleWebKit/JavaScriptCore
div.innerText : 2978.4ms
div.outerText : 2944.4ms

- Chromium/V8 without the patch
div.innerText : 10050.8ms
div.outerText : 10072.2ms

- Chromium/V8 with the patch
div.innerText: 2536.4ms
div.outerText: 2714ms

[Kentaro Hara]

Created attachment 131995 [details]
Patch

[Antti Koivisto]

Comment on attachment 131995 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=131995&action=review

I'll leave reviewing to Chromium people.

> Source/JavaScriptCore/wtf/Platform.h:1195
> +#if PLATFORM(CHROMIUM)
> +#define WTF_TEXT_ITERATOR_BUFFER_INITIAL_CAPACITY (1 << 15)
> +#else
> +#define WTF_TEXT_ITERATOR_BUFFER_INITIAL_CAPACITY (1 << 16)
> +#endif /* PLATFORM(CHROMIUM) */

This is too small for Platform.h. It should be in the function itself.

[Kentaro Hara]

Created attachment 132038 [details]
Patch

[Kentaro Hara]

(In reply to comment #4)
> > Source/JavaScriptCore/wtf/Platform.h:1195
> > +#if PLATFORM(CHROMIUM)
> > +#define WTF_TEXT_ITERATOR_BUFFER_INITIAL_CAPACITY (1 << 15)
> > +#else
> > +#define WTF_TEXT_ITERATOR_BUFFER_INITIAL_CAPACITY (1 << 16)
> > +#endif /* PLATFORM(CHROMIUM) */
> 
> This is too small for Platform.h. It should be in the function itself.

Fixed. Thanks.

[Dimitri Glazkov (Google)]

Comment on attachment 132038 [details]
Patch

View in context: https://bugs.webkit.org/attachment.cgi?id=132038&action=review

Whoa, this is an awesome patch. For a second there, I was excited that you were speeding up the setters. But even speeding up the getters so much is kick-ass!

> Source/WebCore/editing/TextIterator.cpp:2515
> +#if PLATFORM(CHROMIUM)

Why CHROMIUM? That is way too wide and completely mysterious. Can we narrow down the specific bit that's required?

[Dai Mikurube]

(In reply to comment #7)
This issue may depend on the fact that Chromium/Mac doesn't use tcmalloc.  So I think we can narrow it down to Mac.

Great work, haraken.

[Kentaro Hara]

Comment on attachment 132038 [details]
Patch

Thanks for comments. Let me invalidate r? for now to investigate more. A new patch is coming soon.

[Kentaro Hara]

Created attachment 132190 [details]
Patch

[Kentaro Hara]

(In reply to comment #7)
> > Source/WebCore/editing/TextIterator.cpp:2515
> > +#if PLATFORM(CHROMIUM)
> 
> Why CHROMIUM? That is way too wide and completely mysterious. Can we narrow down the specific bit that's required?

(In reply to comment #8)
> (In reply to comment #7)
> This issue may depend on the fact that Chromium/Mac doesn't use tcmalloc.  So I think we can narrow it down to Mac.

Fixed it.

Another thing I tried is to remove plainTextToMallocAllocatedBuffer() and replace the logic with StringBuilder. While it cleaned up the code, it degraded performance by 10%.

Consequently, I just changed the initial buffer size if "PLATFORM(CHROMIUM) && PLATFORM(MAC)".

[WebKit Review Bot]

Comment on attachment 132190 [details]
Patch

Clearing flags on attachment: 132190

Committed r111109: <http://trac.webkit.org/changeset/111109>

[WebKit Review Bot]

All reviewed patches have been landed.  Closing bug.

[Maciej Stachowiak]

Why was this even made a platform ifdef at all? Is there any platform where 1 << 15 is worse than 1 << 16? If not, we should just keep the value consistent.

[Dimitri Glazkov (Google)]

(In reply to comment #14)
> Why was this even made a platform ifdef at all? Is there any platform where 1 << 15 is worse than 1 << 16? If not, we should just keep the value consistent.

Even better.

[Kentaro Hara]

(In reply to comment #14)
> Why was this even made a platform ifdef at all? Is there any platform where 1 << 15 is worse than 1 << 16?

- As far as I experimented locally, 1<<15 is better than 1<<16 in Chromium/Mac. 1<<15 and 1<<16 are the same in AppleWebKit/Mac and Chromium/Linux (their performance "gap" exists at between 1<<17 and 1<<18).
- According to anttik (who wrote 1<<16 a long time ago), there was no strong reason for 1<<16.
- The reason why I changed 1<<16 to 1<<15 on Chromium/Mac only is that it seems the value highly depends on malloc systems and I was afraid that changing to 1<<15 _might_ cause performance regression in some platform.

But as you pointed, I think 1<<15 would not be worse than 1<<16 in all platforms. I'll change it to 1<<15 in a following patch.

[Kentaro Hara]

Reopening to attach new patch.

[Kentaro Hara]

Created attachment 132490 [details]
Patch

[WebKit Review Bot]

Comment on attachment 132490 [details]
Patch

Clearing flags on attachment: 132490

Committed r111133: <http://trac.webkit.org/changeset/111133>

[WebKit Review Bot]

All reviewed patches have been landed.  Closing bug.