New issue 2387 by esei...@chromium.org: Parser/tiny-innerHTML is spending  
17.5% of time in sem_post under v8?
http://code.google.com/p/v8/issues/detail?id=2387

Parser/tiny-innerHTML is spending 17.5% of time in sem_post under v8?

Tools/Scripts/run-perf-tests Parser/tiny-innerHTML.html
or directly:  
http://trac.webkit.org/browser/trunk/PerformanceTests/Parser/tiny-inn...

first 10 lines from pprof --text output:

Running Parser/tiny-innerHTML.html (10 of 12)
      363  17.5%  17.5%      363  17.5% sem_post
       45   2.2%  19.7%      186   9.0% WebCore::HTMLTokenizer::nextToken
       43   2.1%  21.7%       43   2.1% MaskPtr (inline)
       34   1.6%  23.4%       34   1.6% append (inline)
       32   1.5%  24.9%       66   3.2%  
WebCore::SegmentedString::advanceAndUpdateLineNumber
       30   1.4%  26.3%      695  33.5%  
WebCore::HTMLDocumentParser::pumpTokenizer
       29   1.4%  27.7%       34   1.6%  
WebCore::SegmentedString::advanceAndUpdateLineNumber (inline)
       27   1.3%  29.0%       65   3.1% tcmalloc::FL_Next
       25   1.2%  30.3%       36   1.7%  
WebCore::HTMLConstructionSite::reconstructTheActiveFormattingElements
       23   1.1%  31.4%       45   2.2% SegmentedString (inline)

(The sample was generated by using the run-perf-tests --profile patch,  
currently attached to https://bugs.webkit.org/show_bug.cgi?id=99517.)

I suspect we may be tickling some pathological case in v8, but I don't yet  
know how to get v8 symbols to see.

Comment #1 on issue 2387 by esei...@chromium.org: Parser/tiny-innerHTML is  
spending 17.5% of time in sem_post under v8?
http://code.google.com/p/v8/issues/detail?id=2387

(No comment was entered for this change.)

Attachments:
        pprof9657.0.svg  104 KB

Comment #2 on issue 2387 by willc...@chromium.org: Parser/tiny-innerHTML is  
spending 17.5% of time in sem_post under v8?
http://code.google.com/p/v8/issues/detail?id=2387

I looked at the svg output. There are around 2064 samples in this svg  
output (out of the nearly 2100, so there's some slop in other non-core  
areas, but that's not worth examining). When we go from the WebCore  
callback into V8, we still have all 2064 samples, so that part's cheap. By  
the time we exit V8 into WebCore, we're at 1554 samples. Out of the 510  
samples we spent in V8, there's a big chunk of that (360 samples) in  
sem_post. The rest of the 1554 samples is distributed through WebCore, and  
looks to primarily be in tokenizing.