numpy.uint64 + int = float64

[nschloe]

When adding int 1 to a numpy {u}int array, usually a type conversion to int64 happens. This is normal. If, however, the type is uint64, the array gets converted to float64. I had not expected that. Perhaps it's correct though. Opinions?

MWE:

import numpy

a = numpy.array(0, dtype=int)
print((a + 1).dtype)  # int64

a = numpy.array(0, dtype=numpy.uint16)
print((a + 1).dtype)  # int64

a = numpy.array(0, dtype=numpy.uint32)
print((a + 1).dtype)  # int64

a = numpy.array(0, dtype=numpy.uint64)
print((a + 1).dtype)  # float64!

[nschloe]

I think casting it to float64 is a sensible choice after all. The problem is that 1 is cast into a numpy.in64 first. The sum of a uint64 and an int64 does not fit into either of the two data types, so out comes a float64.

[seberg]

Closing as duplicate, e.g. of gh-12525. But yes it is an issue. The problem is that the only way to solve such issues would probably to revamp promotion rules within quite a lot. In most languages categories of types (float, int) are well ordered. If they are not things get weird for example.

The int + uint casting for 64bit integers is a bit separate to the dtype ordering though.