wikimedia / parsoid / 1818

85%

push

travis-ci

Don't process token attributes unnecessarily

* The async flag is false in AttributeTransformManager only when the
  attributes of the token being processed won't get updated or expanded.

* Given that, there is no reason going async and then rebuilding
  new attributes. Based on some logging, I found that > 60% of tokens
  went through buildExpandedAttrs even though ~1% of all tokens
  actually got their attributes updated.

  Accordingly, juggled around some yield/async logic to bypass the
  buildExpandedAttrs method altogether where possible.

* Eliminated unnecessary promise creation from the common case path
  where the attributes come through unchanged. So, I moved back from
  async/yield to promises in AttributeExpander.onToken
  (and TemplateHandler.onTemplate just to be consistent).

* Together, the previous two optimizations lead to significant perf
  improvement.

  On enwiki:Berlin, went from 8.3s -> 8.1s -> 7.9s on my laptop

  parserTests complete faster as well.

* There is another source of inefficiency lurking around still that will be
  addressed later. While the 60% of tokens now bypass buildExpandedAttrs,
  they still go through AttributeTransformManager.process and build a new
  key-value array that is not actually used. It would be useful to skip
  that as well in a subsequent patch.

Change-Id: I2e8d90dcc

8914 of 10955 branches covered (81.37%)

14253 of 16555 relevant lines covered (86.09%)

102048.48 hits per line