Type promotion rules for comparison operations

[crusaderky]

The page on type promotion rules extensively discusses how arithmetic operations must work. However, it is not clear on how comparisons should behave.

xp.asarray(x, dtype=xp.int64) + xp.asarray(x, dtype=float64) is explicitly disallowed for evident reasons of ambiguity about the output precision. However, I can't think of any cases where __eq__, __gt__, etc. between the two can be ambiguous? Just need to state that inf and nan use the same rules as if both arguments were float.

The latest version of array-api-strict disallows these operations, resulting in unnecessarily cumbersome code.

[rgommers]

However, I can't think of any cases where __eq__, __gt__, etc. between the two can be ambiguous?

That's not the only problem - it's simply not supported by all libraries. And if you look under the hood, these operators are still implemented as kernels with two operands of the same type only - and to get them to be the same type, one has to apply type promotion rules (which don't exist for mixed integer and floating-point arrays).

It is also technically possible for there to be integers that aren't exactly representable as floating-point numbers of a given precision, yielding questions about corner cases.