And improve the call machinery in general. This path currently traverses many, many layers and it is not clear which of them are really necessary at this point.

bz, do you have a simple (representative) testcase for this?

The testing patch and testcase in bug 1019081 is one option. Another option is http://mozilla.pettay.fi/moztests/events/event_speed_3.html where if my profile is not lying to me the time under JS::Call is 20% or so, of which only 9% is actually running the jitcode we end up running.