Calling Math.max() from a parallel section causes many bailouts with both 2 and 4 arguments. Using a custom max() function gets rid of the bailouts. What concerns me, in a run on the attached program, is that there are many fewer bailouts than calls to mapPar, but (unsubstantiated speculation) perhaps the code gets marked as parallel-unfriendly at some point and execution falls back to sequential. Also, there are fewer bailouts when running in a debug shell than in a release shell. AMD FX-4100, Ubuntu 13.10 64-bit, mozilla-central 171972:714c8927d6af.

There were some fixes to Math.max previously, I forget the bug number, perhaps those were enough: with current code and Math.max I see no bailouts, and there's plenty of parallel work.