don't propagate NaN/Inf if Partials are zero

Author: jrevels

src/partials.jl

@@ -75,9 +75,13 @@ Base.convert{N,T}(::Type{Partials{N,T}}, partials::Partials{N,T}) = partials
 @inline @compat(Base.:+){N}(a::Partials{N}, b::Partials{N}) = Partials(add_tuples(a.values, b.values))
 @inline @compat(Base.:-){N}(a::Partials{N}, b::Partials{N}) = Partials(sub_tuples(a.values, b.values))
 @inline @compat(Base.:-)(partials::Partials) = Partials(minus_tuple(partials.values))
-@inline @compat(Base.:*)(partials::Partials, x::Real) = Partials(scale_tuple(partials.values, x))
-@inline @compat(Base.:*)(x::Real, partials::Partials) = partials*x
 @inline @compat(Base.:/)(partials::Partials, x::Real) = Partials(div_tuple_by_scalar(partials.values, x))
+@inline @compat(Base.:*)(x::Real, partials::Partials) = partials*x
+
+@inline function @compat(Base.:*)(partials::Partials, x::Real)
+    y = ifelse((isnan(x) || isinf(x)) && iszero(partials), one(x), x)
+    return Partials(scale_tuple(partials.values, y))
+end
 
 @inline function _mul_partials{N}(a::Partials{N}, b::Partials{N}, afactor, bfactor)
     return Partials(mul_tuples(a.values, b.values, afactor, bfactor))