Started to model the cost of divisions more accurately.

Eigen/src/Core/NumTraits.h

@@ -60,6 +60,23 @@ template<typename T> struct GenericNumTraits
     MulCost = 1
   };
 
+  // Division is messy but important, because it is expensive and throughput
+  // varies significantly. The following numbers are based on min division
+  // throughput on Haswell.
+  template<bool Vectorized>
+  struct Div {
+    enum {
+#ifdef EIGEN_VECTORIZE_AVX
+      AVX = true,
+#else
+      AVX = false,
+#endif
+      Cost = IsInteger ? (sizeof(T) == 8 ? (IsSigned ? 24 : 21) : (IsSigned ? 8 : 9)):
+          Vectorized ? (sizeof(T) == 8 ? (AVX ? 16 : 8) : (AVX ? 14 : 7)) : 8
+    };
+  };
+
+
   typedef T Real;
   typedef typename internal::conditional<
                      IsInteger,

Eigen/src/Core/functors/BinaryFunctors.h

@@ -238,7 +238,13 @@ template<typename Scalar> struct scalar_hypot_op {
 };
 template<typename Scalar>
 struct functor_traits<scalar_hypot_op<Scalar> > {
-  enum { Cost = 5 * NumTraits<Scalar>::MulCost, PacketAccess=0 };
+  enum
+  {
+    Cost = 3 * NumTraits<Scalar>::AddCost +
+           2 * NumTraits<Scalar>::MulCost +
+           2 * NumTraits<Scalar>::template Div<false>::Cost,
+    PacketAccess = false
+  };
 };
 
 /** \internal
@@ -564,6 +570,10 @@ struct scalar_inverse_mult_op {
   { return internal::pdiv(pset1<Packet>(m_other),a); }
   Scalar m_other;
 };
+template<typename Scalar>
+struct functor_traits<scalar_inverse_mult_op<Scalar> >
+{ enum { PacketAccess = packet_traits<Scalar>::HasDiv, Cost = NumTraits<Scalar>::template Div<PacketAccess>::Cost }; };
+
 
 } // end namespace internal