Enh: Improve handling of RegExp in the form of /.*blah.*/

• Component: JavaScriptCore
• Status: RESOLVED FIXED

[Michael Saboff]

The current regular expression code handles REs in the form of /.*blah.*/ by greedily matching all characters with the first ".*" and then back tracking to find the "blah". Without capturing parens anywhere in the RE, we can look for the "blah" and then process the leading and trailing ".*".  This should provide a big win for these types of REs.

[Michael Saboff]

Created attachment 100311 [details]
Proposed patch

Proposed enhancement to both the Yarr interpreter and JIT to handle expressions in the form of /[^].*[?]<sub-expr>.*[$]/[m].  The terms in between the leading and trailing .*'s cannot capture and also this enhancement is limited to single alternative expressions.  Process the inner terms and then look for the beginning of the string and end of the string.

[Michael Saboff]

Created attachment 100374 [details]
Updated patch to fix win compilation error.

[Gavin Barraclough]

Comment on attachment 100374 [details]
Updated patch to fix win compilation error.

View in context: https://bugs.webkit.org/attachment.cgi?id=100374&action=review

> Source/JavaScriptCore/yarr/YarrInterpreter.cpp:1074
> +        if (pattern->m_multiline) {

These two if statements could be merged & save a repeating the match assign & return true.

if (!pattern->m_multiline && ((term.anchors.m_bol && matchBegin) || (term.anchors.m_eol && matchEnd != input.end())))
    return false;

> Source/JavaScriptCore/yarr/YarrInterpreter.cpp:1379
> +            BACKTRACK();

I don't think we should ever be able to backtrack into here; I think this should be ASSERT_NOT_REACHED()?

> Source/JavaScriptCore/yarr/YarrJIT.cpp:1011
> +        saveStartIndex.append(branch32(Equal, matchPos, Imm32(m_checked)));

Please modify per our discussion, /.*\n\d+.*/ "abc\n123".

> Source/JavaScriptCore/yarr/YarrPattern.cpp:716
> +        for (size_t alt = 0; alt < alternatives.size(); ++alt) {

We shouldn't need a for-loop here, given the above guard!   possibly should switch the check to "if (alternatives.size() != 1) return;" (I don't think we ever have 0 alternatives, but just to be safe!), then drop the loop & just work over alternatives[0].

[Michael Saboff]

Created attachment 100690 [details]
Patch with suggested updates

[Michael Saboff]

Created attachment 100694 [details]
Previous patch with regex JIT turned back on.

[Michael Saboff]

Committed r90962: <http://trac.webkit.org/changeset/90962>