maths: use __builtin_signbit to handle +- zeros

maths.hpp

@@ -104,9 +104,9 @@ digits [[gnu::pure]] (const T& value);
 
 
 //-----------------------------------------------------------------------------
-template <typename T>
+constexpr int sign (int);
-typename try_signed<T>::type
+constexpr float sign (float);
-sign [[gnu::pure]] (T val);
+constexpr double sign (double);
 
 
 //-----------------------------------------------------------------------------

maths.ipp

@@ -50,10 +50,23 @@ pow (T x, unsigned y)
 ///
 /// This really needs to be inline for performance as it's used in a few inner
 /// loops where the possibility of a function call is a deal breaker.
-template <typename T>
+constexpr int
-typename try_signed<T>::type
+sign (int v)
-sign (T val) {
+{
-    return val >= 0 ? T{1} : T{-1};
+    return __builtin_signbitl (v) ? -1 : 1;
+}
+
+constexpr float
+sign (float v)
+{
+    return __builtin_signbitf (v) ? -1.f : 1.f;
+}
+
+
+constexpr double
+sign (double v)
+{
+    return __builtin_signbit (v) ? -1. : 1.f;
 }