Bug 179230 – Web Tooling Benchmark: fix RegExp pathologies

• Status: RESOLVED
• Resolution: FIXED
• Component: JavaScriptCore

[JF Bastien]

We seem pathologically slow on the coffeescript part of Web Tooling Benchmark:
  https://v8.github.io/web-tooling-benchmark/
Source:
  https://github.com/bmeurer/web-tooling-benchmark

I get 0.22 runs / sec on STP 42, versus 8.24 runs / sec on Chrome 61.

[JF Bastien]

Non-fork source is actually: https://github.com/v8/web-tooling-benchmark

[Saam Barati]

We're spending all our time interpreting these RegExps:

^\s*###([^#][\s\S]*?)(?:###[^\n\S]*|###$)|^(?:\s*#(?!##[^#]).*)+ 8014.546220ms
^(?!\d)((?:(?!\s)[$\w\x7f-\uffff])+)([^\n\S]*:(?!:))? 8179.496043ms
^\/(?!\/)((?:[^[\/\n\\]|\\[^\n]|\[(?:\\[^\n]|[^\]\n\\])*\])*)(\/)? 2951.752059ms
^(?:[-=]>|[-+*\/%<>&|^!?=]=|>>>=?|([-+:])\1|([&|<>*\/%])\2=?|\?(\.|::)|\.{2,3}) 2888.876650ms

Michael believes we will get much faster if we JIT these patterns.

[Radar WebKit Bug Importer]

<rdar://problem/35343582>

[Michael Saboff]

Originally, the main RegExp issue was with the coffeescript test which spent most of its time in the aforementioned 4 regular expressions.  Adding JIT support for greedy sub patterns (https://bugs.webkit.org/show_bug.cgi?id=180538) addressed some of these RegExp's.  There is still the need to add JIT support for back references (https://bugs.webkit.org/show_bug.cgi?id=180874) to address the remaining Yarr interpreter use in coffeescript.

Since this bug was created, the Web Tooling benchmark has changed.  A few new tests have been added.  Two of those new tests, acorn and espree, spend significant time in the Yarr interpreter evaluating non-greedy sub patterns.  That work is tracked in https://bugs.webkit.org/show_bug.cgi?id=180876.

[Michael Saboff]

We now have JIT support for greedy and non-greedy nested sub patterns, and back references including case ignore for 8 bit strings.