Implementation of Dual for `NaNMath.pow` does not use `NaNMath.log`

[jClugstor]

I would expect this:

using ForwardDiff
using NaNMath

function new_pow(x)
    NaNMath.pow(x[1],x[2])
end

ForwardDiff.gradient(new_pow, [-1.0, 1.0])

to return a NaN I think?

The issue is right here,

for f in (:(Base.:^), :(NaNMath.pow))
    @eval begin
        @define_binary_dual_op(
            $f,
            begin
                vx, vy = value(x), value(y)
                expv = ($f)(vx, vy)
                powval = vy * ($f)(vx, vy - 1)
                if isconstant(y)
                    logval = one(expv)
                elseif iszero(vx) && vy > 0
                    logval = zero(vx)
                else
                    logval = expv * log(vx)
                end
                new_partials = _mul_partials(partials(x), partials(y), powval, logval)
                return Dual{Txy}(expv, new_partials)
            end,

in logval = expv*log(vx) , I think NaNMath.log should be used if f is NaNMath.pow.

[KristofferC]

Makes sense to me.