Correct epicsMin/epicsMax behaviour for floating-point NaN arguments.

src/libCom/cppStd/epicsAlgorithm.h

@@ -13,18 +13,61 @@
 #ifndef __EPICS_ALGORITHM_H__
 #define __EPICS_ALGORITHM_H__
 
-template <class T>
-inline const T& epicsMax (const T& a, const T& b)
-{
-    return (a<b) ? b : a;
-}
+#include "epicsMath.h"
+
+// The C++ standard only requires types to be less-than comparable, so
+// the epicsMin and epicsMax templates only use operator <
+
+// epicsMin
 
 template <class T>
 inline const T& epicsMin (const T& a, const T& b)
 {
-    return (a<b) ? a : b;
+    return (b < a) ? b : a;
 }
 
+// If b is a NaN the above template returns a, but should return NaN.
+// These specializations ensure that epicsMin(x,NaN) == NaN
+
+template <>
+inline const float& epicsMin (const float& a, const float& b)
+{
+    return (b < a) || isnan(b) ? b : a;
+}
+
+template <>
+inline const double& epicsMin (const double& a, const double& b)
+{
+    return (b < a) || isnan(b) ? b : a;
+}
+
+
+// epicsMax
+
+template <class T>
+inline const T& epicsMax (const T& a, const T& b)
+{
+    return (a < b) ? b : a;
+}
+
+// If b is a NaN the above template returns a, but should return NaN.
+// These specializations ensure that epicsMax(x,NaN) == NaN
+
+template <>
+inline const float& epicsMax (const float& a, const float& b)
+{
+    return (a < b) || isnan(b) ? b : a;
+}
+
+template <>
+inline const double& epicsMax (const double& a, const double& b)
+{
+    return (a < b) || isnan(b) ? b : a;
+}
+
+
+// epicsSwap
+
 template <class T>
 inline void epicsSwap(T& a, T& b)
 {