There is a IEEE totalOrder predicate (IEEE 754-2008) that defines the sort order for IEEE floats. Rust implements it as total_cmp, C++20 implements it as std::strong_order. It is what you might expect, positive floats (sign bit == 0) are ordered as integers, negative floats (sign bit == 1) as negative integers. The effect is -0 sorts before +0 and positive nans to the end, negative nans to the beginning. NumPy sorts don't follow that convention, so I wonder if we should define a npy_totalOrder predicate somewhere. Maybe try to have a totalOrder predicate as a scientific python standard. I note that total_cmp looks like a slightly modified < and doesn't seem to offer any speed advantage.