Fallback math methods for ::Real are likely to call themselves

[tkoolen]

On 0.6.2:

julia> using ForwardDiff

julia> ForwardDiff.derivative(sinpi, 1.)
0.0

on b571283:

julia> using ForwardDiff

julia> ForwardDiff.derivative(sinpi, 1.)
ERROR: StackOverflowError:
Stacktrace:
 [1] sinpi(::ForwardDiff.Dual{ForwardDiff.Tag{typeof(sinpi),Float64},Float64,1}) at ./special/trig.jl:865 (repeats 80000 times)

This is because of

julia/base/special/trig.jl

Line 865 in 7674aca

sinpi(x::Real) = sinpi(float(x))

which seems unfortunate. Similarly, log1p required a workaround in DiffRules. I wholeheartedly agree with JuliaDiff/ForwardDiff.jl#261 (comment):

Looks like Base.Math.JuliaLibm.log1p(x::Real) just calls Base.Math.JuliaLibm.log1p(float(x)), which basically guarantees infinite recursion in the general case

log and cospi also have this issue. Maybe others as well?

[StefanKarpinski]

Do you have a different way to express this in mind that wouldn't have this issue?

[tkoolen]

How about turning

sinpi(x::T) where T<:AbstractFloat

into

_sinpi(x::T) where T<:AbstractFloat

and have

sinpi(x::Real) = _sinpi(float(x)) 

[mbauman]

This is similarly handled in other places with a generic AbstractFloat specialization.

[tkoolen]

That assumes that float always returns an AbstractFloat. ForwardDiff.Dual doesn't adhere to that, and it probably shouldn't be an AbstractFloat, since the backing type can also be non-AbstractFloat.

[martinholters]

We could make that assumption more explicit by using convert(AbstractFloat, x) instead of float (x).

[tkoolen]

That would certainly be an improvement, MethodError is better than StackOverflowError. But the sinpi/log1p/... changes from 0.6.2 to 0.7 would still be breaking, and it's not like the old implementation returned wrong results (as far as I understand). Restoring the old behavior using something like #26552 (comment) seems to me like it would save a lot of hassle for package authors.

[chriscoey]

any path forward on this? I am struggling with this issue: JuliaDiff/ForwardDiff.jl#324