`Base.literal_pow` undesired floating-point conversion for base of `±1`?

[oscardssmith]

julia> Base.literal_pow(^, -1, Val(-1))
-1.0

julia> y = -1
-1

julia> (-1)^y
-1

This has been broken since at least 1.6

[nhz2]

I don't think this is a bug because the docs for ^ says.

If y is a negative integer literal, then
Base.literal_pow transforms the operation to inv(x)^-y by default, where -y is
positive.

and inv(-1) is -1.0

I think the reason for this is so that functions like foo(x::Int) = x^-1 are type stable.

[oscardssmith]

the problem here is that one of the fundamental properties of literal pow is that it doesn't change the results for inputs where x^y returns an answer.

[KlausC]

IMO the problem is, that literal_pow(x, Val{-1}) cannot be type stable and at the same time behave like x^Int(-1), because the latter throws DomainError for x = 2 while the first is supposed to return 0.5.
In order to be type stable, it has to return 1.0 for x = 1 then.

[stevengj]

I'm removing the "bug" label since this is the intended behavior for negative literal powers, regardless of the base, and it's not clear whether anything can or should be changed here in a type-stable way.