Cython should be happy with whatever CPython uses (as long as CPython's header files agree with CPython's build ;-) ).

I saw the RasPi benchmarks on the ML. That would have been my suggested trial platform as well.
https://mail.python.org/archives/list/python-dev@python.org/message/5RJGI6THWCDYTTEPXMWXU7CK66RQUTD4/

The results look ok. Maybe the slowdown for pickling is really the increased data size of integers. And it's visible that some compute-heavily benchmarks like pyaes did get a little slower. I doubt that they represent a real use case on such a platform, though. Doing any kind of number crunching on a RasPi without NumPy would appear like a rather strange adventure.

That said, if we decide to keep 15-bit digits in the end, I wonder if "SIZEOF_VOID_P" is the right decision point. It seems more of a "has reasonably fast 64-bit multiply or not" kind of decision – however that translates into code. I'm sure there are 32-bit platforms that would actually benefit from 30-bit digits today.

If we find a platform that would be fine with 30-bits but lacks a fast 64-bit multiply, then we could still try to add a platform specific value size check for smaller numbers. Since those are common case, branch prediction might help us more often than not.

But then, I wonder how much complexity this is even worth, given that the goal is to reduce the complexity. Platform maintainers can still decide to configure the digit size externally for the time being, if it makes a difference for them. Maybe switching off 15-bits by default is just good enough for the next couple of years to come. :)