Fix Map-with-Stride and cover it by new unit tests.

5 files changed, 39 insertions, 41 deletions

diff --git a/Eigen/src/Core/DenseStorageBase.h b/Eigen/src/Core/DenseStorageBase.h
index 89e6e7112..12ffd2e43 100644
--- a/Eigen/src/Core/DenseStorageBase.h
+++ b/Eigen/src/Core/DenseStorageBase.h
@@ -362,6 +362,9 @@ class DenseStorageBase : public _Base<Derived>
       * while the AlignedMap() functions return aligned Map objects and thus should be called only with 16-byte-aligned
       * \a data pointers.
       *
+      * These methods do not allow to specify strides. If you need to specify strides, you have to
+      * use the Map class directly.
+      *
       * \see class Map
       */
     //@{
diff --git a/Eigen/src/Core/Map.h b/Eigen/src/Core/Map.h
index d9ccb1b20..f8b70b866 100644
--- a/Eigen/src/Core/Map.h
+++ b/Eigen/src/Core/Map.h
@@ -52,11 +52,23 @@ template<typename MatrixType, int Options, typename StrideType>
 struct ei_traits<Map<MatrixType, Options, StrideType> >
   : public ei_traits<MatrixType>
 {
+  typedef typename MatrixType::Scalar Scalar;
   enum {
+    InnerStride = StrideType::InnerStrideAtCompileTime,
+    OuterStride = StrideType::OuterStrideAtCompileTime,
+    HasNoInnerStride = InnerStride <= 1,
+    HasNoOuterStride = OuterStride == 0,
+    HasNoStride = HasNoInnerStride && HasNoOuterStride,
+    IsAligned = int(int(Options)&Aligned)==Aligned,
+    IsDynamicSize = MatrixType::SizeAtCompileTime==Dynamic,
+    KeepsPacketAccess = bool(HasNoInnerStride)
+                        && ( bool(IsDynamicSize)
+                           || HasNoOuterStride
+                           || ( OuterStride!=Dynamic && ((int(OuterStride)*sizeof(Scalar))%16)==0 ) ),
     Flags0 = ei_traits<MatrixType>::Flags,
-    Flags1 = ((Options&Aligned)==Aligned ? Flags0 |  AlignedBit
-                                         : Flags0 & ~AlignedBit),
-    Flags = int(StrideType::InnerStrideAtCompileTime)==1 ? Flags1 : (Flags1 & ~PacketAccessBit)
+    Flags1 = IsAligned ? int(Flags0) |  AlignedBit : int(Flags0) & ~AlignedBit,
+    Flags2 = HasNoStride ? int(Flags1) : int(Flags1 & ~LinearAccessBit),
+    Flags = KeepsPacketAccess ? int(Flags2) : (int(Flags2) & ~PacketAccessBit)
   };
 };
 
@@ -94,23 +106,6 @@ template<typename MatrixType, int Options, typename StrideType> class Map
     inline Map(const Scalar* data, int rows, int cols, const StrideType& stride = StrideType())
       : Base(data, rows, cols), m_stride(stride) {}
 
-/*
-    inline void resize(int rows, int cols)
-    {
-      EIGEN_ONLY_USED_FOR_DEBUG(rows);
-      EIGEN_ONLY_USED_FOR_DEBUG(cols);
-      ei_assert(rows == this->rows());
-      ei_assert(cols == this->cols());
-    }
-
-    inline void resize(int size)
-    {
-      EIGEN_STATIC_ASSERT_VECTOR_ONLY(MatrixType)
-      EIGEN_ONLY_USED_FOR_DEBUG(size);
-      ei_assert(size == this->size());
-    }
-*/
-
     EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Map)
 
   protected:
diff --git a/Eigen/src/Core/Stride.h b/Eigen/src/Core/Stride.h
index 7982035fd..f04039e7d 100644
--- a/Eigen/src/Core/Stride.h
+++ b/Eigen/src/Core/Stride.h
@@ -25,7 +25,7 @@
 #ifndef EIGEN_STRIDE_H
 #define EIGEN_STRIDE_H
 
-template<int _InnerStrideAtCompileTime, int _OuterStrideAtCompileTime>
+template<int _OuterStrideAtCompileTime, int _InnerStrideAtCompileTime>
 class Stride
 {
   public:
@@ -36,45 +36,45 @@ class Stride
     };
 
     Stride()
-      : m_inner(InnerStrideAtCompileTime), m_outer(OuterStrideAtCompileTime)
+      : m_outer(OuterStrideAtCompileTime), m_inner(InnerStrideAtCompileTime)
     {
       ei_assert(InnerStrideAtCompileTime != Dynamic && OuterStrideAtCompileTime != Dynamic);
     }
 
-    Stride(int innerStride, int outerStride)
-      : m_inner(innerStride), m_outer(outerStride)
+    Stride(int outerStride, int innerStride)
+      : m_outer(outerStride), m_inner(innerStride)
     {
       ei_assert(innerStride>=0 && outerStride>=0);
     }
 
     Stride(const Stride& other)
-      : m_inner(other.inner()), m_outer(other.outer())
+      : m_outer(other.outer()), m_inner(other.inner())
     {}
 
-    inline int inner() const { return m_inner.value(); }
     inline int outer() const { return m_outer.value(); }
+    inline int inner() const { return m_inner.value(); }
 
   protected:
-    ei_int_if_dynamic<InnerStrideAtCompileTime> m_inner;
     ei_int_if_dynamic<OuterStrideAtCompileTime> m_outer;
+    ei_int_if_dynamic<InnerStrideAtCompileTime> m_inner;
 };
 
-template<int Value = Dynamic>
-class InnerStride : public Stride<Value, 0>
+template<int Value>
+class InnerStride : public Stride<0, Value>
 {
-    typedef Stride<Value,0> Base;
+    typedef Stride<0, Value> Base;
   public:
     InnerStride() : Base() {}
-    InnerStride(int v) : Base(v,0) {}
+    InnerStride(int v) : Base(0, v) {}
 };
 
-template<int Value = Dynamic>
-class OuterStride : public Stride<0, Value>
+template<int Value>
+class OuterStride : public Stride<Value, 0>
 {
-    typedef Stride<0,Value> Base;
+    typedef Stride<Value, 0> Base;
   public:
     OuterStride() : Base() {}
-    OuterStride(int v) : Base(0,v) {}
+    OuterStride(int v) : Base(v,0) {}
 };
 
 #endif // EIGEN_STRIDE_H
diff --git a/Eigen/src/Core/util/Constants.h b/Eigen/src/Core/util/Constants.h
index c27c979a6..c167df697 100644
--- a/Eigen/src/Core/util/Constants.h
+++ b/Eigen/src/Core/util/Constants.h
@@ -86,11 +86,11 @@ const unsigned int EvalBeforeAssigningBit = 0x4;
   * Long version: means that the coefficients can be handled by packets
   * and start at a memory location whose alignment meets the requirements
   * of the present CPU architecture for optimized packet access. In the fixed-size
-  * case, there is the additional condition that the total size of the coefficients
-  * array is a multiple of the packet size, so that it is possible to access all the
-  * coefficients by packets. In the dynamic-size case, there is no such condition
-  * on the total size, so it might not be possible to access the few last coeffs
-  * by packets.
+  * case, there is the additional condition that it be possible to access all the
+  * coefficients by packets (this implies the requirement that the size be a multiple of 16 bytes,
+  * and that any nontrivial strides don't break the alignment). In the dynamic-size case,
+  * there is no such condition on the total size and strides, so it might not be possible to access
+  * all coeffs by packets.
   *
   * \note This bit can be set regardless of whether vectorization is actually enabled.
   *       To check for actual vectorizability, see \a ActualPacketAccessBit.
diff --git a/Eigen/src/Core/util/ForwardDeclarations.h b/Eigen/src/Core/util/ForwardDeclarations.h
index 6096272fa..8451d0ebe 100644
--- a/Eigen/src/Core/util/ForwardDeclarations.h
+++ b/Eigen/src/Core/util/ForwardDeclarations.h
@@ -61,7 +61,7 @@ template<typename _Scalar, int SizeAtCompileTime, int MaxSizeAtCompileTime=SizeA
 template<typename MatrixType, typename DiagonalType, int ProductOrder> class DiagonalProduct;
 template<typename MatrixType, int Index> class Diagonal;
 
-template<int InnerStrideAtCompileTime = Dynamic, int OuterStrideAtCompileTime = Dynamic> class Stride;
+template<int InnerStrideAtCompileTime, int OuterStrideAtCompileTime> class Stride;
 template<typename MatrixType, int Options=Unaligned, typename StrideType = Stride<0,0> > class Map;
 
 template<typename Derived> class TriangularBase;