Vectorize atanh & add a missing definition and unit test for atan.

Eigen/src/Core/arch/SSE/MathFunctions.h

@@ -81,11 +81,6 @@ Packet4f pacos<Packet4f>(const Packet4f& _x)
   return pacos_float(_x);
 }
 
-template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
-Packet2d patan<Packet2d>(const Packet2d& _x) {
-  return patan_double(_x);
-}
-
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
 Packet4f pasin<Packet4f>(const Packet4f& _x)
 {
@@ -98,6 +93,17 @@ Packet4f patan<Packet4f>(const Packet4f& _x)
   return patan_float(_x);
 }
 
+template <> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet2d patan<Packet2d>(const Packet2d& _x) {
+  return patan_double(_x);
+}
+
+template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS
+Packet4f patanh<Packet4f>(const Packet4f& _x)
+{
+  return patanh_float(_x);
+}
+
 // Notice that for newer processors, it is counterproductive to use Newton
 // iteration for square root. In particular, Skylake and Zen2 processors
 // have approximately doubled throughput of the _mm_sqrt_ps instruction

Eigen/src/Core/arch/SSE/PacketMath.h

@@ -142,6 +142,7 @@ struct packet_traits<float> : default_packet_traits {
     HasACos = 1,
     HasASin = 1,
     HasATan = 1,
+    HasATanh = 1,
     HasLog = 1,
     HasLog1p = 1,
     HasExpm1 = 1,