> I did some spelunking.  Guido committed the similar optimization in r8306.    The diff is at:
> http://svn.python.org/view/python/trunk/Python/ceval.c?r1=8087&r2=8306
> 
> His commit message was:
> 
>     Huge speedup by inlining some common integer operations:
>     int+int, int-int, int <compareop> int, and list[int].
>     (Unfortunately, int*int is way too much code to inline.)
> 
>     Also corrected a NULL that should have been a zero.
> 
> It's possible that these kinds of optimizations were worthwhile with
> PyInt but aren't worthwhile with PyLong.

It also doesn't say the individual contribution of each optimization,
and it also doesn't say on which kind of workloads the "huge speedup"
was witnessed (I would say that pystone is a possibility, or perhaps
even some timeit micro-benchmark).