> I seem not to be able to verify the promised increased 
> performance of the alternative callout mechanism compared
> to the old one using libffi. I get about the same timings
> using the code below with a 3.4 vm (with libffi) and a 
> 3.6 (with x86-sysv). Somehow I was expecting that the time spent in 
> primitives had to be significant lower in the 3.6 case. Maybe 
> it's just my naive profiling method?

Most likely XDrawArc is an extremely bad method to call for profiling this.
Note that drawing an arc is very unlikely to be a highly optimized method
for zero area arcs (why would you do this to begin with?) and it also
possibly requires a trip through the X server. If you want to measure the
difference you should use a method which you know returns pretty much
immediately. For examples, the ffi test methods (such as ffiTestInts,
ffiTestFloats or ffiPrintString) might be better. Here, you know what will
be happening.

Cheers,
  - Andreas