Make a^b return Float64 when b is signed integer

[PallHaraldsson]

This seems sensible, and consistent with division (which is implied when b is negative). There needs to be an exception when a or b is a BigInt (or then return BigFloat, while BigInt result needs to also be an option). I'm not sure this would break a lot of code, but a way to get old behavior is I guess needed (and already existing).

If this is too breaking just add the 2.0 milestone...

Plan B would be overflow check (I'm not sure there's a separate open issue on it). It's a less good option IMHO, while it wouldn't be breaking, and would also apply for unsigned.

[JeffBezanson]

We can't really do this, because signed integers are the default and the common case, not something you can reasonably have a separate case for.

[PallHaraldsson]

It's because signed Ints are the default that I proposed this, as a breaking 2.0 change.

It would make the language more consistent, now:

julia> 2^-2
0.25

julia> b = -2;

julia> 2^b
ERROR: DomainError with -2:

With this change we could rather special-case the positive literal pow (also evalpoly?), possibly allowing some polynominals to keep status quo. I still doubt we should do that, except to get equivalent Float64 code.

Integer unsigned power would still be an option (or calling literal_pow), for those that really intended it, and can follow with casting to signed.