finish making use of CoeffReadCost and the new XprCopy everywhere

seems appropriate to me.

4 files changed, 75 insertions, 38 deletions

diff --git a/Eigen/src/Core/Dot.h b/Eigen/src/Core/Dot.h
index ba45d5192..1d768b259 100644
--- a/Eigen/src/Core/Dot.h
+++ b/Eigen/src/Core/Dot.h
@@ -72,18 +72,25 @@ template<typename OtherDerived>
 typename ei_traits<Derived>::Scalar
 MatrixBase<Derived>::dot(const MatrixBase<OtherDerived>& other) const
 {
-  typename Derived::XprCopy xprCopy(derived());
-  typename OtherDerived::XprCopy otherXprCopy(other.derived());
+  typedef typename Derived::XprCopy XprCopy;
+  typedef typename OtherDerived::XprCopy OtherXprCopy;
+  typedef typename ei_unref<XprCopy>::type _XprCopy;
+  typedef typename ei_unref<OtherXprCopy>::type _OtherXprCopy;
+  XprCopy xprCopy(derived());
+  OtherXprCopy otherXprCopy(other.derived());
 
-  ei_assert(IsVectorAtCompileTime
-      && OtherDerived::IsVectorAtCompileTime
-      && xprCopy.size() == otherXprCopy.size());
+  ei_assert(_XprCopy::IsVectorAtCompileTime
+         && _OtherXprCopy::IsVectorAtCompileTime
+         && xprCopy.size() == otherXprCopy.size());
   Scalar res;
-  if(SizeAtCompileTime <= EIGEN_UNROLLING_LIMIT)
+  const bool unroll = SizeAtCompileTime
+                      * (_XprCopy::CoeffReadCost + _OtherXprCopy::CoeffReadCost + NumTraits<Scalar>::MulCost)
+                      + (SizeAtCompileTime - 1) * NumTraits<Scalar>::AddCost
+                      <= EIGEN_UNROLLING_LIMIT;
+  if(unroll)
     ei_dot_unroller<SizeAtCompileTime-1,
-                SizeAtCompileTime <= EIGEN_UNROLLING_LIMIT ? SizeAtCompileTime : Dynamic,
-                typename ei_unref<typename Derived::XprCopy>::type,
-                typename ei_unref<typename OtherDerived::XprCopy>::type>
+                unroll ? SizeAtCompileTime : Dynamic,
+                _XprCopy, _OtherXprCopy>
       ::run(xprCopy, otherXprCopy, res);
   else
   {
@@ -142,8 +149,8 @@ template<typename OtherDerived>
 bool MatrixBase<Derived>::isOrtho
 (const MatrixBase<OtherDerived>& other, RealScalar prec) const
 {
-  typename Derived::XprCopy xprCopy(derived());
-  typename OtherDerived::XprCopy otherXprCopy(other.derived());
+  typename ei_xpr_copy<Derived,2>::type xprCopy(derived());
+  typename ei_xpr_copy<OtherDerived,2>::type otherXprCopy(other.derived());
   return ei_abs2(xprCopy.dot(otherXprCopy)) <= prec * prec * xprCopy.norm2() * otherXprCopy.norm2();
 }
 
diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h
index d303cbdb7..7f149075b 100644
--- a/Eigen/src/Core/Product.h
+++ b/Eigen/src/Core/Product.h
@@ -86,13 +86,13 @@ struct ei_traits<Product<Lhs, Rhs, EvalMode> >
   typedef typename Lhs::Scalar Scalar;
   typedef typename ei_xpr_copy<Lhs,Rhs::ColsAtCompileTime>::type LhsXprCopy;
   typedef typename ei_xpr_copy<Rhs,Lhs::RowsAtCompileTime>::type RhsXprCopy;
-  typedef typename ei_unref<LhsXprCopy>::type ActualLhs;
-  typedef typename ei_unref<RhsXprCopy>::type ActualRhs;
+  typedef typename ei_unref<LhsXprCopy>::type _LhsXprCopy;
+  typedef typename ei_unref<RhsXprCopy>::type _RhsXprCopy;
   enum {
-    LhsCoeffReadCost = ActualLhs::CoeffReadCost,
-    RhsCoeffReadCost = ActualRhs::CoeffReadCost,
-    LhsFlags = ActualLhs::Flags,
-    RhsFlags = ActualRhs::Flags,
+    LhsCoeffReadCost = _LhsXprCopy::CoeffReadCost,
+    RhsCoeffReadCost = _RhsXprCopy::CoeffReadCost,
+    LhsFlags = _LhsXprCopy::Flags,
+    RhsFlags = _RhsXprCopy::Flags,
     RowsAtCompileTime = Lhs::RowsAtCompileTime,
     ColsAtCompileTime = Rhs::ColsAtCompileTime,
     MaxRowsAtCompileTime = Lhs::MaxRowsAtCompileTime,
@@ -117,9 +117,10 @@ template<typename Lhs, typename Rhs, int EvalMode> class Product : ei_no_assignm
   public:
 
     EIGEN_GENERIC_PUBLIC_INTERFACE(Product)
-
     typedef typename ei_traits<Product>::LhsXprCopy LhsXprCopy;
     typedef typename ei_traits<Product>::RhsXprCopy RhsXprCopy;
+    typedef typename ei_traits<Product>::_LhsXprCopy _LhsXprCopy;
+    typedef typename ei_traits<Product>::_RhsXprCopy _RhsXprCopy;
 
     Product(const Lhs& lhs, const Rhs& rhs)
       : m_lhs(lhs), m_rhs(rhs)
@@ -144,8 +145,7 @@ template<typename Lhs, typename Rhs, int EvalMode> class Product : ei_no_assignm
       {
         ei_product_unroller<Lhs::ColsAtCompileTime-1,
                             unroll ? Lhs::ColsAtCompileTime : Dynamic,
-                            typename ei_unref<LhsXprCopy>::type,
-                            typename ei_unref<RhsXprCopy>::type>
+                            _LhsXprCopy, _RhsXprCopy>
           ::run(row, col, m_lhs, m_rhs, res);
       }
       else
diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h
index bccfe42b4..f7fdbc077 100644
--- a/Eigen/src/Core/Redux.h
+++ b/Eigen/src/Core/Redux.h
@@ -87,15 +87,19 @@ struct ei_traits<PartialRedux<Direction, BinaryOp, MatrixType> >
   typedef typename ei_result_of<
                      BinaryOp(typename MatrixType::Scalar)
                    >::type Scalar;
+  typedef typename ei_xpr_copy<MatrixType>::type MatrixTypeXprCopy;
+  typedef typename ei_unref<MatrixTypeXprCopy>::type _MatrixTypeXprCopy;
   enum {
     RowsAtCompileTime = Direction==Vertical   ? 1 : MatrixType::RowsAtCompileTime,
     ColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::ColsAtCompileTime,
-    MaxRowsAtCompileTime = MatrixType::MaxRowsAtCompileTime,
-    MaxColsAtCompileTime = MatrixType::MaxColsAtCompileTime,
+    MaxRowsAtCompileTime = Direction==Vertical   ? 1 : MatrixType::MaxRowsAtCompileTime,
+    MaxColsAtCompileTime = Direction==Horizontal ? 1 : MatrixType::MaxColsAtCompileTime,
     Flags = (RowsAtCompileTime == Dynamic || ColsAtCompileTime == Dynamic)
-          ? (unsigned int)MatrixType::Flags
-          : (unsigned int)MatrixType::Flags & ~LargeBit,
-    CoeffReadCost = 1 //FIXME -- unimplemented!
+          ? (unsigned int)_MatrixTypeXprCopy::Flags
+          : (unsigned int)_MatrixTypeXprCopy::Flags & ~LargeBit,
+    TraversalSize = Direction==Vertical ? RowsAtCompileTime : ColsAtCompileTime,
+    CoeffReadCost = TraversalSize * _MatrixTypeXprCopy::CoeffReadCost
+                  + (TraversalSize - 1) * ei_functor_traits<BinaryOp>::Cost
   };
 };
 
@@ -106,6 +110,8 @@ class PartialRedux : ei_no_assignment_operator,
   public:
 
     EIGEN_GENERIC_PUBLIC_INTERFACE(PartialRedux)
+    typedef typename ei_traits<PartialRedux>::MatrixTypeXprCopy MatrixTypeXprCopy;
+    typedef typename ei_traits<PartialRedux>::_MatrixTypeXprCopy _MatrixTypeXprCopy;
 
     PartialRedux(const MatrixType& mat, const BinaryOp& func = BinaryOp())
       : m_matrix(mat), m_functor(func) {}
@@ -124,7 +130,7 @@ class PartialRedux : ei_no_assignment_operator,
     }
 
   protected:
-    const typename MatrixType::XprCopy m_matrix;
+    const MatrixTypeXprCopy m_matrix;
     const BinaryOp m_functor;
 };
 
@@ -171,10 +177,13 @@ template<typename BinaryOp>
 typename ei_result_of<BinaryOp(typename ei_traits<Derived>::Scalar)>::type
 MatrixBase<Derived>::redux(const BinaryOp& func) const
 {
-  if(SizeAtCompileTime <= EIGEN_UNROLLING_LIMIT)
+  const bool unroll = SizeAtCompileTime * CoeffReadCost
+                    + (SizeAtCompileTime-1) * ei_functor_traits<BinaryOp>::Cost
+                    <= EIGEN_UNROLLING_LIMIT;
+  if(unroll)
     return ei_redux_unroller<BinaryOp, Derived, 0,
-       (SizeAtCompileTime>0 && SizeAtCompileTime <= EIGEN_UNROLLING_LIMIT) ?
-        SizeAtCompileTime : Dynamic>::run(derived(), func);
+                             unroll ? SizeAtCompileTime : Dynamic>
+           ::run(derived(), func);
   else
   {
     Scalar res;
@@ -291,10 +300,12 @@ struct ei_any_unroller<Derived, Dynamic>
 template<typename Derived>
 bool MatrixBase<Derived>::all(void) const
 {
-  if(SizeAtCompileTime <= EIGEN_UNROLLING_LIMIT)
+  const bool unroll = SizeAtCompileTime * (CoeffReadCost + NumTraits<Scalar>::AddCost)
+                      <= EIGEN_UNROLLING_LIMIT;
+  if(unroll)
     return ei_all_unroller<Derived,
-       (SizeAtCompileTime>0 && SizeAtCompileTime <= EIGEN_UNROLLING_LIMIT) ?
-        SizeAtCompileTime : Dynamic>::run(derived());
+                           unroll ? SizeAtCompileTime : Dynamic
+     >::run(derived());
   else
   {
     for(int j = 0; j < cols(); j++)
@@ -311,10 +322,12 @@ bool MatrixBase<Derived>::all(void) const
 template<typename Derived>
 bool MatrixBase<Derived>::any(void) const
 {
-  if(SizeAtCompileTime <= EIGEN_UNROLLING_LIMIT)
+  const bool unroll = SizeAtCompileTime * (CoeffReadCost + NumTraits<Scalar>::AddCost)
+                      <= EIGEN_UNROLLING_LIMIT;
+  if(unroll)
     return ei_any_unroller<Derived,
-       (SizeAtCompileTime>0 && SizeAtCompileTime <= EIGEN_UNROLLING_LIMIT) ?
-        SizeAtCompileTime : Dynamic>::run(derived());
+                           unroll ? SizeAtCompileTime : Dynamic
+           >::run(derived());
   else
   {
     for(int j = 0; j < cols(); j++)
diff --git a/Eigen/src/Core/Visitor.h b/Eigen/src/Core/Visitor.h
index 2cff3a576..0aeac0cfc 100644
--- a/Eigen/src/Core/Visitor.h
+++ b/Eigen/src/Core/Visitor.h
@@ -74,10 +74,13 @@ template<typename Derived>
 template<typename Visitor>
 void MatrixBase<Derived>::visit(Visitor& visitor) const
 {
-  if(SizeAtCompileTime <= EIGEN_UNROLLING_LIMIT)
+  const bool unroll = SizeAtCompileTime * CoeffReadCost
+                    + (SizeAtCompileTime-1) * ei_functor_traits<Visitor>::Cost
+                    <= EIGEN_UNROLLING_LIMIT;
+  if(unroll)
     return ei_visitor_unroller<Visitor, Derived,
-        (SizeAtCompileTime>0 && SizeAtCompileTime <= EIGEN_UNROLLING_LIMIT) ?
-        SizeAtCompileTime : Dynamic>::run(derived(), visitor);
+        unroll ? SizeAtCompileTime : Dynamic
+      >::run(derived(), visitor);
   else
   {
     visitor.init(coeff(0,0), 0, 0);
@@ -124,6 +127,13 @@ struct ei_min_coeff_visitor : ei_coeff_visitor<Scalar>
   }
 };
 
+template<typename Scalar>
+struct ei_functor_traits<ei_min_coeff_visitor<Scalar> > {
+  enum {
+    Cost = NumTraits<Scalar>::AddCost
+  };
+};
+
 /** \internal
   * \brief Visitor computing the max coefficient with its value and coordinates
   *
@@ -143,6 +153,13 @@ struct ei_max_coeff_visitor : ei_coeff_visitor<Scalar>
   }
 };
 
+template<typename Scalar>
+struct ei_functor_traits<ei_max_coeff_visitor<Scalar> > {
+  enum {
+    Cost = NumTraits<Scalar>::AddCost
+  };
+};
+
 /** \returns the minimum of all coefficients of *this
   * and puts in *row and *col its location.
   *