Thanks for the update.

I'm not too worried about performance: I suspect that any additional overhead would be lost in the overhead of the Python machinery. It would be worth profiling to check, of course, before any change went in. I'd expect that most performance critical code of this type would be using NumPy/SciPy anyway.

I'm still unsure about whether the code is worth changing: I do agree that the behaviour suggested by C99 Annex G is (in the abstract) an improvement on Python's current behaviour, and compatibility with C would be a plus. On the other side, introducing an incompatibility with other versions of Python and with NumPy isn't ideal. For me, the potential benefits don't really overcome the drawbacks here, but I'd like to hear other opinions.