Math.pow returns different results between Int and Float exponents

[HertzDevil]

On non-Windows platforms, Float64#**(Int32) and Float64#**(Float64) are backed by llvm.powi and llvm.pow respectively. The two can sometimes return different results, in particular when dealing with certain subnormal results:

2.0 ** -1024.0 # => 5.562684646268003e-309
2.0 ** -1024   # => 0.0
0.5 ** 1024.0  # => 5.562684646268003e-309
0.5 ** 1024    # => 5.562684646268003e-309

Reduced: (it doesn't matter whether the base value is a constant or the variable x)

class String
  def to_unsafe
    pointerof(@c)
  end
end

lib LibC
  fun printf(fmt : UInt8*, ...) : Int32
end

@[Link("m")]
lib LibM
  fun pow_f64 = "llvm.pow.f64"(value : Float64, power : Float64) : Float64
  fun powi_f64 = "llvm.powi.f64.i32"(value : Float64, power : Int32) : Float64
end

x = 2.0
LibC.printf("2.0 ** -1024.0 # => %e\n", LibM.pow_f64(x, -1024.0))
LibC.printf("2.0 ** -1024   # => %e\n", LibM.powi_f64(x, -1024)) # prints zero
x = 0.5
LibC.printf("0.5 ** 1024.0  # => %e\n", LibM.pow_f64(x, 1024.0))
LibC.printf("0.5 ** 1024    # => %e\n", LibM.powi_f64(x, 1024))

Although there are no guarantees about either form's precision, this could still be a problem since only Windows delegates #**(Int32) to #**(Float64), thus the same code will not print zero when compiled using MSVC. Maybe we should simply drop powi on the other platforms as well?

See also: llvm/llvm-project#65088