* move some compile time "if" to their respective unroller (assign and dot)

* fix a couple of compilation issues when unrolling is disabled
* reduce default unrolling limit to a more reasonable value

7 files changed, 144 insertions, 132 deletions

diff --git a/Eigen/src/Core/Assign.h b/Eigen/src/Core/Assign.h
index 21d5e3d11..98df25235 100644
--- a/Eigen/src/Core/Assign.h
+++ b/Eigen/src/Core/Assign.h
@@ -58,10 +58,30 @@ struct ei_matrix_assignment_unroller<Derived1, Derived2, 0>
   inline static void run(Derived1 &, const Derived2 &) {}
 };
 
+// Dynamic col-major
 template<typename Derived1, typename Derived2>
-struct ei_matrix_assignment_unroller<Derived1, Derived2, Dynamic>
+struct ei_matrix_assignment_unroller<Derived1, Derived2, -1>
 {
-  inline static void run(Derived1 &, const Derived2 &) {}
+  inline static void run(Derived1 &dst, const Derived2 &src)
+  {
+    for(int j = 0; j < dst.cols(); j++)
+      for(int i = 0; i < dst.rows(); i++)
+        dst.coeffRef(i, j) = src.coeff(i, j);
+  }
+};
+
+// Dynamic row-major
+template<typename Derived1, typename Derived2>
+struct ei_matrix_assignment_unroller<Derived1, Derived2, -2>
+{
+  inline static void run(Derived1 &dst, const Derived2 &src)
+  {
+    // traverse in row-major order
+    // in order to allow the compiler to unroll the inner loop
+    for(int i = 0; i < dst.rows(); i++)
+      for(int j = 0; j < dst.cols(); j++)
+        dst.coeffRef(i, j) = src.coeff(i, j);
+  }
 };
 
 //----
@@ -103,10 +123,12 @@ struct ei_matrix_assignment_packet_unroller<Derived1, Derived2, Dynamic>
 template <typename Derived, typename OtherDerived,
 bool Vectorize = (int(Derived::Flags) & int(OtherDerived::Flags) & VectorizableBit)
               && ((int(Derived::Flags)&RowMajorBit)==(int(OtherDerived::Flags)&RowMajorBit))
-              && (  (int(Derived::Flags) & int(OtherDerived::Flags) & Like1DArrayBit)
-                  ||((int(Derived::Flags)&RowMajorBit)
-                    ? int(Derived::ColsAtCompileTime)!=Dynamic && (int(Derived::ColsAtCompileTime)%ei_packet_traits<typename Derived::Scalar>::size==0)
-                    : int(Derived::RowsAtCompileTime)!=Dynamic && (int(Derived::RowsAtCompileTime)%ei_packet_traits<typename Derived::Scalar>::size==0)) ),
+              && (   (int(Derived::Flags) & int(OtherDerived::Flags) & Like1DArrayBit)
+                  || ((int(Derived::Flags) & RowMajorBit)
+                    ?     int(Derived::ColsAtCompileTime)!=Dynamic
+                      && (int(Derived::ColsAtCompileTime)%ei_packet_traits<typename Derived::Scalar>::size==0)
+                    :     int(Derived::RowsAtCompileTime)!=Dynamic
+                      && (int(Derived::RowsAtCompileTime)%ei_packet_traits<typename Derived::Scalar>::size==0)) ),
 bool Unroll = Derived::SizeAtCompileTime * OtherDerived::CoeffReadCost <= EIGEN_UNROLLING_LIMIT>
 struct ei_assignment_impl;
 
@@ -156,39 +178,21 @@ inline Derived& MatrixBase<Derived>
 
 //----
 
-template <typename Derived, typename OtherDerived>
-struct ei_assignment_impl<Derived, OtherDerived, false, true> // no vec + unrolling
+// no vectorization
+template <typename Derived, typename OtherDerived, bool Unroll>
+struct ei_assignment_impl<Derived, OtherDerived, false, Unroll>
 {
   static void run(Derived & dst, const OtherDerived & src)
   {
     ei_matrix_assignment_unroller
-      <Derived, OtherDerived, int(Derived::SizeAtCompileTime)
+      <Derived, OtherDerived,
+      Unroll ? int(Derived::SizeAtCompileTime)
+      : Derived::ColsAtCompileTime == Dynamic || Derived::RowsAtCompileTime != Dynamic ? -1 // col-major
+      : -2 // row-major
       >::run(dst.derived(), src.derived());
   }
 };
 
-template <typename Derived, typename OtherDerived>
-struct ei_assignment_impl<Derived, OtherDerived, false, false> // no vec + no unrolling + col major order
-{
-  static void run(Derived & dst, const OtherDerived & src)
-  {
-    if(Derived::ColsAtCompileTime == Dynamic || Derived::RowsAtCompileTime != Dynamic)
-    {
-      for(int j = 0; j < dst.cols(); j++)
-        for(int i = 0; i < dst.rows(); i++)
-          dst.coeffRef(i, j) = src.coeff(i, j);
-    }
-    else
-    {
-      // traverse in row-major order
-      // in order to allow the compiler to unroll the inner loop
-      for(int i = 0; i < dst.rows(); i++)
-        for(int j = 0; j < dst.cols(); j++)
-          dst.coeffRef(i, j) = src.coeff(i, j);
-    }
-  }
-};
-
 //----
 
 template <typename Derived, typename OtherDerived>
@@ -224,7 +228,7 @@ struct ei_assignment_impl<Derived, OtherDerived, true, false> // vec + no-unroll
 };
 
 template <typename Derived, typename OtherDerived>
-struct ei_packet_assignment_seclector<Derived, OtherDerived, true, true> // row-major + complex 1D array
+struct ei_packet_assignment_seclector<Derived, OtherDerived, true, true> // row-major + complex 1D array like
 {
   static void run(Derived & dst, const OtherDerived & src)
   {
diff --git a/Eigen/src/Core/Dot.h b/Eigen/src/Core/Dot.h
index c85d9e3fb..a1c1ffce8 100644
--- a/Eigen/src/Core/Dot.h
+++ b/Eigen/src/Core/Dot.h
@@ -26,17 +26,17 @@
 #define EIGEN_DOT_H
 
 template<int Index, int Size, typename Derived1, typename Derived2>
-struct ei_dot_unroller
+struct ei_dot_impl
 {
   inline static void run(const Derived1 &v1, const Derived2& v2, typename Derived1::Scalar &dot)
   {
-    ei_dot_unroller<Index-1, Size, Derived1, Derived2>::run(v1, v2, dot);
+    ei_dot_impl<Index-1, Size, Derived1, Derived2>::run(v1, v2, dot);
     dot += v1.coeff(Index) * ei_conj(v2.coeff(Index));
   }
 };
 
 template<int Size, typename Derived1, typename Derived2>
-struct ei_dot_unroller<0, Size, Derived1, Derived2>
+struct ei_dot_impl<0, Size, Derived1, Derived2>
 {
   inline static void run(const Derived1 &v1, const Derived2& v2, typename Derived1::Scalar &dot)
   {
@@ -44,15 +44,20 @@ struct ei_dot_unroller<0, Size, Derived1, Derived2>
   }
 };
 
-template<int Index, typename Derived1, typename Derived2>
-struct ei_dot_unroller<Index, Dynamic, Derived1, Derived2>
+template<typename Derived1, typename Derived2>
+struct ei_dot_impl<Dynamic, Dynamic, Derived1, Derived2>
 {
-  inline static void run(const Derived1&, const Derived2&, typename Derived1::Scalar&) {}
+  inline static void run(const Derived1& v1, const Derived2& v2, typename Derived1::Scalar& dot)
+  {
+    dot = v1.coeff(0) * ei_conj(v2.coeff(0));
+    for(int i = 1; i < v1.size(); i++)
+      dot += v1.coeff(i)* ei_conj(v2.coeff(i));
+  }
 };
 
 // prevent buggy user code from causing an infinite recursion
 template<int Index, typename Derived1, typename Derived2>
-struct ei_dot_unroller<Index, 0, Derived1, Derived2>
+struct ei_dot_impl<Index, 0, Derived1, Derived2>
 {
   inline static void run(const Derived1&, const Derived2&, typename Derived1::Scalar&) {}
 };
@@ -83,22 +88,16 @@ MatrixBase<Derived>::dot(const MatrixBase<OtherDerived>& other) const
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(_OtherNested);
   EIGEN_STATIC_ASSERT_SAME_VECTOR_SIZE(_Nested,_OtherNested);
   ei_assert(nested.size() == otherNested.size());
-  Scalar res;
   const bool unroll = SizeAtCompileTime
                       * (_Nested::CoeffReadCost + _OtherNested::CoeffReadCost + NumTraits<Scalar>::MulCost)
                       + (int(SizeAtCompileTime) - 1) * NumTraits<Scalar>::AddCost
                       <= EIGEN_UNROLLING_LIMIT;
-  if(unroll)
-    ei_dot_unroller<int(SizeAtCompileTime)-1,
-                unroll ? int(SizeAtCompileTime) : Dynamic,
-                _Nested, _OtherNested>
-      ::run(nested, otherNested, res);
-  else
-  {
-    res = nested.coeff(0) * ei_conj(otherNested.coeff(0));
-    for(int i = 1; i < size(); i++)
-      res += nested.coeff(i)* ei_conj(otherNested.coeff(i));
-  }
+
+  Scalar res;
+  ei_dot_impl<unroll ? int(SizeAtCompileTime)-1 : Dynamic,
+              unroll ? int(SizeAtCompileTime) : Dynamic,
+              _Nested, _OtherNested>
+    ::run(nested, otherNested, res);
   return res;
 }
 
diff --git a/Eigen/src/Core/Part.h b/Eigen/src/Core/Part.h
index 647e6cf2d..eb8dcbba7 100644
--- a/Eigen/src/Core/Part.h
+++ b/Eigen/src/Core/Part.h
@@ -92,7 +92,7 @@ inline void Part<MatrixType, Mode>::operator=(const Other& other)
 }
 
 template<typename Derived1, typename Derived2, unsigned int Mode, int UnrollCount>
-struct ei_part_assignment_unroller
+struct ei_part_assignment_impl
 {
   enum {
     col = (UnrollCount-1) / Derived1::RowsAtCompileTime,
@@ -101,7 +101,7 @@ struct ei_part_assignment_unroller
 
   inline static void run(Derived1 &dst, const Derived2 &src)
   {
-    ei_part_assignment_unroller<Derived1, Derived2, Mode, UnrollCount-1>::run(dst, src);
+    ei_part_assignment_impl<Derived1, Derived2, Mode, UnrollCount-1>::run(dst, src);
 
     if(Mode == SelfAdjoint)
     {
@@ -122,7 +122,7 @@ struct ei_part_assignment_unroller
 };
 
 template<typename Derived1, typename Derived2, unsigned int Mode>
-struct ei_part_assignment_unroller<Derived1, Derived2, Mode, 1>
+struct ei_part_assignment_impl<Derived1, Derived2, Mode, 1>
 {
   inline static void run(Derived1 &dst, const Derived2 &src)
   {
@@ -133,65 +133,78 @@ struct ei_part_assignment_unroller<Derived1, Derived2, Mode, 1>
 
 // prevent buggy user code from causing an infinite recursion
 template<typename Derived1, typename Derived2, unsigned int Mode>
-struct ei_part_assignment_unroller<Derived1, Derived2, Mode, 0>
+struct ei_part_assignment_impl<Derived1, Derived2, Mode, 0>
 {
   inline static void run(Derived1 &, const Derived2 &) {}
 };
 
-template<typename Derived1, typename Derived2, unsigned int Mode>
-struct ei_part_assignment_unroller<Derived1, Derived2, Mode, Dynamic>
+template<typename Derived1, typename Derived2>
+struct ei_part_assignment_impl<Derived1, Derived2, Upper, Dynamic>
 {
-  inline static void run(Derived1 &, const Derived2 &) {}
+  inline static void run(Derived1 &dst, const Derived2 &src)
+  {
+    for(int j = 0; j < dst.cols(); j++)
+      for(int i = 0; i <= j; i++)
+        dst.coeffRef(i, j) = src.coeff(i, j);
+  }
 };
 
+template<typename Derived1, typename Derived2>
+struct ei_part_assignment_impl<Derived1, Derived2, Lower, Dynamic>
+{
+  inline static void run(Derived1 &dst, const Derived2 &src)
+  {
+    for(int j = 0; j < dst.cols(); j++)
+      for(int i = j; i < dst.rows(); i++)
+        dst.coeffRef(i, j) = src.coeff(i, j);
+  }
+};
 
-template<typename MatrixType, unsigned int Mode>
-template<typename Other>
-void Part<MatrixType, Mode>::lazyAssign(const Other& other)
+template<typename Derived1, typename Derived2>
+struct ei_part_assignment_impl<Derived1, Derived2, StrictlyUpper, Dynamic>
 {
-  const bool unroll = MatrixType::SizeAtCompileTime * Other::CoeffReadCost / 2 <= EIGEN_UNROLLING_LIMIT;
-  ei_assert(m_matrix.rows() == other.rows() && m_matrix.cols() == other.cols());
-  if(unroll)
+  inline static void run(Derived1 &dst, const Derived2 &src)
   {
-    ei_part_assignment_unroller
-      <MatrixType, Other, Mode,
-      unroll ? int(MatrixType::SizeAtCompileTime) : Dynamic
-      >::run(m_matrix, other.derived());
+    for(int j = 0; j < dst.cols(); j++)
+      for(int i = 0; i < j; i++)
+        dst.coeffRef(i, j) = src.coeff(i, j);
   }
-  else
+};
+template<typename Derived1, typename Derived2>
+struct ei_part_assignment_impl<Derived1, Derived2, StrictlyLower, Dynamic>
+{
+  inline static void run(Derived1 &dst, const Derived2 &src)
   {
-    switch(Mode)
+    for(int j = 0; j < dst.cols(); j++)
+      for(int i = j+1; i < dst.rows(); i++)
+        dst.coeffRef(i, j) = src.coeff(i, j);
+  }
+};
+template<typename Derived1, typename Derived2>
+struct ei_part_assignment_impl<Derived1, Derived2, SelfAdjoint, Dynamic>
+{
+  inline static void run(Derived1 &dst, const Derived2 &src)
+  {
+    for(int j = 0; j < dst.cols(); j++)
     {
-      case Upper:
-        for(int j = 0; j < m_matrix.cols(); j++)
-          for(int i = 0; i <= j; i++)
-            m_matrix.coeffRef(i, j) = other.coeff(i, j);
-        break;
-      case Lower:
-        for(int j = 0; j < m_matrix.cols(); j++)
-          for(int i = j; i < m_matrix.rows(); i++)
-            m_matrix.coeffRef(i, j) = other.coeff(i, j);
-        break;
-      case StrictlyUpper:
-        for(int j = 0; j < m_matrix.cols(); j++)
-          for(int i = 0; i < j; i++)
-            m_matrix.coeffRef(i, j) = other.coeff(i, j);
-        break;
-      case StrictlyLower:
-        for(int j = 0; j < m_matrix.cols(); j++)
-          for(int i = j+1; i < m_matrix.rows(); i++)
-            m_matrix.coeffRef(i, j) = other.coeff(i, j);
-        break;
-      case SelfAdjoint:
-        for(int j = 0; j < m_matrix.cols(); j++)
-        {
-          for(int i = 0; i < j; i++)
-            m_matrix.coeffRef(j, i) = ei_conj(m_matrix.coeffRef(i, j) = other.coeff(i, j));
-          m_matrix.coeffRef(j, j) = ei_real(other.coeff(j, j));
-        }
-        break;
+      for(int i = 0; i < j; i++)
+        dst.coeffRef(j, i) = ei_conj(dst.coeffRef(i, j) = src.coeff(i, j));
+      dst.coeffRef(j, j) = ei_real(src.coeff(j, j));
     }
   }
+};
+
+template<typename MatrixType, unsigned int Mode>
+template<typename Other>
+void Part<MatrixType, Mode>::lazyAssign(const Other& other)
+{
+  const bool unroll = MatrixType::SizeAtCompileTime * Other::CoeffReadCost / 2 <= EIGEN_UNROLLING_LIMIT;
+  ei_assert(m_matrix.rows() == other.rows() && m_matrix.cols() == other.cols());
+
+  ei_part_assignment_impl
+    <MatrixType, Other, Mode,
+    unroll ? int(MatrixType::SizeAtCompileTime) : Dynamic
+    >::run(m_matrix, other.derived());
 }
 
 template<typename MatrixType, unsigned int Mode>
diff --git a/Eigen/src/Core/Product.h b/Eigen/src/Core/Product.h
index 3709ebf64..7af1bb3ed 100644
--- a/Eigen/src/Core/Product.h
+++ b/Eigen/src/Core/Product.h
@@ -47,8 +47,8 @@ struct ei_product_impl<0, Size, Lhs, Rhs>
   }
 };
 
-template<int Index, typename Lhs, typename Rhs>
-struct ei_product_impl<Index, Dynamic, Lhs, Rhs>
+template<typename Lhs, typename Rhs>
+struct ei_product_impl<Dynamic, Dynamic, Lhs, Rhs>
 {
   inline static void run(int row, int col, const Lhs& lhs, const Rhs& rhs, typename Lhs::Scalar& res)
   {
@@ -268,7 +268,7 @@ template<typename Lhs, typename Rhs, int EvalMode> class Product : ei_no_assignm
       {
         Scalar res;
         const bool unroll = CoeffReadCost <= EIGEN_UNROLLING_LIMIT;
-        ei_product_impl<Lhs::ColsAtCompileTime-1,
+        ei_product_impl<unroll ? Lhs::ColsAtCompileTime-1 : Dynamic,
                             unroll ? Lhs::ColsAtCompileTime : Dynamic,
                             _LhsNested, _RhsNested>
           ::run(row, col, m_lhs, m_rhs, res);
diff --git a/Eigen/src/Core/Redux.h b/Eigen/src/Core/Redux.h
index 7c5534154..821d3813d 100644
--- a/Eigen/src/Core/Redux.h
+++ b/Eigen/src/Core/Redux.h
@@ -63,7 +63,17 @@ template<typename BinaryOp, typename Derived, int Start>
 struct ei_redux_unroller<BinaryOp, Derived, Start, Dynamic>
 {
   typedef typename ei_result_of<BinaryOp(typename Derived::Scalar)>::type Scalar;
-  static Scalar run(const Derived&, const BinaryOp&) { return Scalar(); }
+  static Scalar run(const Derived& mat, const BinaryOp& func)
+  {
+    Scalar res;
+    res = mat.coeff(0,0);
+    for(int i = 1; i < mat.rows(); i++)
+      res = func(res, mat.coeff(i, 0));
+    for(int j = 1; j < mat.cols(); j++)
+      for(int i = 0; i < mat.rows(); i++)
+        res = func(res, mat.coeff(i, j));
+    return res;
+  }
 };
 
 /** \returns the result of a full redux operation on the whole matrix or vector using \a func
@@ -81,21 +91,9 @@ MatrixBase<Derived>::redux(const BinaryOp& func) const
   const bool unroll = SizeAtCompileTime * CoeffReadCost
                     + (SizeAtCompileTime-1) * ei_functor_traits<BinaryOp>::Cost
                     <= EIGEN_UNROLLING_LIMIT;
-  if(unroll)
-    return ei_redux_unroller<BinaryOp, Derived, 0,
-                             unroll ? int(SizeAtCompileTime) : Dynamic>
-           ::run(derived(), func);
-  else
-  {
-    Scalar res;
-    res = coeff(0,0);
-    for(int i = 1; i < rows(); i++)
-      res = func(res, coeff(i, 0));
-    for(int j = 1; j < cols(); j++)
-      for(int i = 0; i < rows(); i++)
-        res = func(res, coeff(i, j));
-    return res;
-  }
+  return ei_redux_unroller<BinaryOp, Derived, 0,
+                            unroll ? int(SizeAtCompileTime) : Dynamic>
+          ::run(derived(), func);
 }
 
 /** \returns the sum of all coefficients of *this
diff --git a/Eigen/src/Core/Visitor.h b/Eigen/src/Core/Visitor.h
index 6e00e4142..bd469f747 100644
--- a/Eigen/src/Core/Visitor.h
+++ b/Eigen/src/Core/Visitor.h
@@ -52,7 +52,15 @@ struct ei_visitor_unroller<Visitor, Derived, 1>
 template<typename Visitor, typename Derived>
 struct ei_visitor_unroller<Visitor, Derived, Dynamic>
 {
-  inline static void run(const Derived &, Visitor&) {}
+  inline static void run(const Derived& mat, Visitor& visitor)
+  {
+    visitor.init(mat.coeff(0,0), 0, 0);
+    for(int i = 1; i < mat.rows(); i++)
+      visitor(mat.coeff(i, 0), i, 0);
+    for(int j = 1; j < mat.cols(); j++)
+      for(int i = 0; i < mat.rows(); i++)
+        visitor(mat.coeff(i, j), i, j);
+  }
 };
 
 
@@ -77,19 +85,9 @@ void MatrixBase<Derived>::visit(Visitor& visitor) const
   const bool unroll = SizeAtCompileTime * CoeffReadCost
                     + (SizeAtCompileTime-1) * ei_functor_traits<Visitor>::Cost
                     <= EIGEN_UNROLLING_LIMIT;
-  if(unroll)
-    return ei_visitor_unroller<Visitor, Derived,
-        unroll ? int(SizeAtCompileTime) : Dynamic
-      >::run(derived(), visitor);
-  else
-  {
-    visitor.init(coeff(0,0), 0, 0);
-    for(int i = 1; i < rows(); i++)
-      visitor(coeff(i, 0), i, 0);
-    for(int j = 1; j < cols(); j++)
-      for(int i = 0; i < rows(); i++)
-        visitor(coeff(i, j), i, j);
-  }
+  return ei_visitor_unroller<Visitor, Derived,
+      unroll ? int(SizeAtCompileTime) : Dynamic
+    >::run(derived(), visitor);
 }
 
 /** \internal
diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 717e3e072..dff4d7885 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -34,7 +34,7 @@
 
 /** Defines the maximal loop size to enable meta unrolling of loops */
 #ifndef EIGEN_UNROLLING_LIMIT
-#define EIGEN_UNROLLING_LIMIT 400
+#define EIGEN_UNROLLING_LIMIT 100
 #endif
 
 #ifdef EIGEN_DEFAULT_TO_ROW_MAJOR