allow vectorization of mat44.col() by adding a InnerPanel boolean

template parameter to Block

5 files changed, 18 insertions, 17 deletions

diff --git a/Eigen/src/Core/Block.h b/Eigen/src/Core/Block.h
index a16aa22d4..2a28ea7cd 100644
--- a/Eigen/src/Core/Block.h
+++ b/Eigen/src/Core/Block.h
@@ -58,8 +58,8 @@
   *
   * \sa DenseBase::block(Index,Index,Index,Index), DenseBase::block(Index,Index), class VectorBlock
   */
-template<typename XprType, int BlockRows, int BlockCols, bool HasDirectAccess>
-struct ei_traits<Block<XprType, BlockRows, BlockCols, HasDirectAccess> > : ei_traits<XprType>
+template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool HasDirectAccess>
+struct ei_traits<Block<XprType, BlockRows, BlockCols, InnerPanel, HasDirectAccess> > : ei_traits<XprType>
 {
   typedef typename ei_traits<XprType>::Scalar Scalar;
   typedef typename ei_traits<XprType>::StorageKind StorageKind;
@@ -92,15 +92,16 @@ struct ei_traits<Block<XprType, BlockRows, BlockCols, HasDirectAccess> > : ei_tr
     MaskPacketAccessBit = (InnerSize == Dynamic || (InnerSize % ei_packet_traits<Scalar>::size) == 0)
                        && (InnerStrideAtCompileTime == 1)
                         ? PacketAccessBit : 0,
+    MaskAlignedBit = (InnerPanel && (OuterStrideAtCompileTime!=Dynamic) && ((OuterStrideAtCompileTime % ei_packet_traits<Scalar>::size) == 0)) ? AlignedBit : 0,
     FlagsLinearAccessBit = (RowsAtCompileTime == 1 || ColsAtCompileTime == 1) ? LinearAccessBit : 0,
-    Flags0 = ei_traits<XprType>::Flags & (HereditaryBits | MaskPacketAccessBit | LvalueBit | DirectAccessBit),
+    Flags0 = ei_traits<XprType>::Flags & (HereditaryBits | MaskPacketAccessBit | LvalueBit | DirectAccessBit | MaskAlignedBit),
     Flags1 = Flags0 | FlagsLinearAccessBit,
     Flags = (Flags1 & ~RowMajorBit) | (IsRowMajor ? RowMajorBit : 0)
   };
 };
 
-template<typename XprType, int BlockRows, int BlockCols, bool HasDirectAccess> class Block
-  : public ei_dense_xpr_base<Block<XprType, BlockRows, BlockCols, HasDirectAccess> >::type
+template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel, bool HasDirectAccess> class Block
+  : public ei_dense_xpr_base<Block<XprType, BlockRows, BlockCols, InnerPanel, HasDirectAccess> >::type
 {
   public:
 
@@ -229,9 +230,9 @@ template<typename XprType, int BlockRows, int BlockCols, bool HasDirectAccess> c
 };
 
 /** \internal */
-template<typename XprType, int BlockRows, int BlockCols>
-class Block<XprType,BlockRows,BlockCols,true>
-  : public MapBase<Block<XprType, BlockRows, BlockCols,true> >
+template<typename XprType, int BlockRows, int BlockCols, bool InnerPanel>
+class Block<XprType,BlockRows,BlockCols, InnerPanel,true>
+  : public MapBase<Block<XprType, BlockRows, BlockCols, InnerPanel, true> >
 {
   public:
 
diff --git a/Eigen/src/Core/DenseBase.h b/Eigen/src/Core/DenseBase.h
index 02eb31d76..a00b791cd 100644
--- a/Eigen/src/Core/DenseBase.h
+++ b/Eigen/src/Core/DenseBase.h
@@ -233,17 +233,17 @@ template<typename Derived> class DenseBase
     /** \internal the return type of MatrixBase::eigenvalues() */
     typedef Matrix<typename NumTraits<typename ei_traits<Derived>::Scalar>::Real, ei_traits<Derived>::ColsAtCompileTime, 1> EigenvaluesReturnType;
     /** \internal expression type of a column */
-    typedef Block<Derived, ei_traits<Derived>::RowsAtCompileTime, 1> ColXpr;
+    typedef Block<Derived, ei_traits<Derived>::RowsAtCompileTime, 1, !IsRowMajor> ColXpr;
     /** \internal expression type of a row */
-    typedef Block<Derived, 1, ei_traits<Derived>::ColsAtCompileTime> RowXpr;
+    typedef Block<Derived, 1, ei_traits<Derived>::ColsAtCompileTime, IsRowMajor> RowXpr;
     /** \internal expression type of a block of whole columns */
-    typedef Block<Derived, ei_traits<Derived>::RowsAtCompileTime, Dynamic> ColsBlockXpr;
+    typedef Block<Derived, ei_traits<Derived>::RowsAtCompileTime, Dynamic, !IsRowMajor> ColsBlockXpr;
     /** \internal expression type of a block of whole rows */
-    typedef Block<Derived, Dynamic, ei_traits<Derived>::ColsAtCompileTime> RowsBlockXpr;
+    typedef Block<Derived, Dynamic, ei_traits<Derived>::ColsAtCompileTime, IsRowMajor> RowsBlockXpr;
     /** \internal expression type of a block of whole columns */
-    template<int N> struct NColsBlockXpr { typedef Block<Derived, ei_traits<Derived>::RowsAtCompileTime, N> Type; };
+    template<int N> struct NColsBlockXpr { typedef Block<Derived, ei_traits<Derived>::RowsAtCompileTime, N, !IsRowMajor> Type; };
     /** \internal expression type of a block of whole rows */
-    template<int N> struct NRowsBlockXpr { typedef Block<Derived, N, ei_traits<Derived>::ColsAtCompileTime> Type; };
+    template<int N> struct NRowsBlockXpr { typedef Block<Derived, N, ei_traits<Derived>::ColsAtCompileTime, IsRowMajor> Type; };
 
 
 #endif // not EIGEN_PARSED_BY_DOXYGEN
diff --git a/Eigen/src/Core/Map.h b/Eigen/src/Core/Map.h
index 3386c6d69..763948453 100644
--- a/Eigen/src/Core/Map.h
+++ b/Eigen/src/Core/Map.h
@@ -100,7 +100,7 @@ struct ei_traits<Map<PlainObjectType, MapOptions, StrideType> >
                            || ( OuterStrideAtCompileTime!=Dynamic
                            && ((static_cast<int>(sizeof(Scalar))*OuterStrideAtCompileTime)%16)==0 ) ),
     Flags0 = ei_traits<PlainObjectType>::Flags,
-    Flags1 = IsAligned ? int(Flags0) | AlignedBit : int(Flags0) & ~AlignedBit,
+    Flags1 = IsAligned ? (int(Flags0) | AlignedBit) : (int(Flags0) & ~AlignedBit),
     Flags2 = HasNoStride ? int(Flags1) : int(Flags1 & ~LinearAccessBit),
     Flags = KeepsPacketAccess ? int(Flags2) : (int(Flags2) & ~PacketAccessBit)
   };
diff --git a/Eigen/src/Core/arch/SSE/MathFunctions.h b/Eigen/src/Core/arch/SSE/MathFunctions.h
index e4ca82985..cb73fd205 100644
--- a/Eigen/src/Core/arch/SSE/MathFunctions.h
+++ b/Eigen/src/Core/arch/SSE/MathFunctions.h
@@ -373,7 +373,7 @@ Packet4f ei_pcos<Packet4f>(const Packet4f& _x)
   return _mm_xor_ps(y, sign_bit);
 }
 
-// This is Quake3's fast inverse square root.
+// This is based on Quake3's fast inverse square root.
 // For detail see here: http://www.beyond3d.com/content/articles/8/
 template<> EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS EIGEN_UNUSED
 Packet4f ei_psqrt<Packet4f>(const Packet4f& _x)
diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h
index e4466db5b..423aa110e 100644
--- a/Eigen/src/Core/util/ForwardDeclarations.h
+++ b/Eigen/src/Core/util/ForwardDeclarations.h
@@ -60,7 +60,7 @@ template<typename ExpressionType> class NestByValue;
 template<typename ExpressionType> class ForceAlignedAccess;
 template<typename ExpressionType> class SwapWrapper;
 
-template<typename XprType, int BlockRows=Dynamic, int BlockCols=Dynamic,
+template<typename XprType, int BlockRows=Dynamic, int BlockCols=Dynamic, bool InnerPanel = false,
          bool HasDirectAccess = ei_has_direct_access<XprType>::ret> class Block;
 
 template<typename MatrixType, int Size=Dynamic> class VectorBlock;