bump to 3.2.9

(grafted from 476beed7f8
)

CMakeLists.txt

@@ -1,6 +1,5 @@
 project(Eigen)
-
+cmake_minimum_required(VERSION 2.8.5)
-cmake_minimum_required(VERSION 2.8.2)
 
 # guard against in-source builds
 
@@ -55,6 +54,7 @@ endif(EIGEN_HG_CHANGESET)
 
 
 include(CheckCXXCompilerFlag)
+include(GNUInstallDirs)
 
 set(CMAKE_MODULE_PATH ${PROJECT_SOURCE_DIR}/cmake)
 
@@ -288,25 +288,26 @@ option(EIGEN_TEST_C++0x "Enables all C++0x features." OFF)
 
 include_directories(${CMAKE_CURRENT_SOURCE_DIR} ${CMAKE_CURRENT_BINARY_DIR})
 
-# the user modifiable install path for header files
+# Backward compatibility support for EIGEN_INCLUDE_INSTALL_DIR
-set(EIGEN_INCLUDE_INSTALL_DIR ${EIGEN_INCLUDE_INSTALL_DIR} CACHE PATH "The directory where we install the header files (optional)")
+if(EIGEN_INCLUDE_INSTALL_DIR AND NOT INCLUDE_INSTALL_DIR)
-
+  set(INCLUDE_INSTALL_DIR ${EIGEN_INCLUDE_INSTALL_DIR}
-# set the internal install path for header files which depends on wether the user modifiable
+      CACHE PATH "The directory relative to CMAKE_PREFIX_PATH where Eigen header files are installed")
-# EIGEN_INCLUDE_INSTALL_DIR has been set by the user or not.
-if(EIGEN_INCLUDE_INSTALL_DIR)
-  set(INCLUDE_INSTALL_DIR
-    ${EIGEN_INCLUDE_INSTALL_DIR}
-    CACHE INTERNAL
-    "The directory where we install the header files (internal)"
-  )
 else()
   set(INCLUDE_INSTALL_DIR
-    "${CMAKE_INSTALL_PREFIX}/include/eigen3"
+      "${CMAKE_INSTALL_INCLUDEDIR}/eigen3"
-    CACHE INTERNAL
+      CACHE PATH "The directory relative to CMAKE_PREFIX_PATH where Eigen header files are installed"
-    "The directory where we install the header files (internal)"
+      )
-  )
 endif()
 
+set(CMAKEPACKAGE_INSTALL_DIR
+    "${CMAKE_INSTALL_LIBDIR}/cmake/eigen3"
+    CACHE PATH "The directory relative to CMAKE_PREFIX_PATH where Eigen3Config.cmake is installed"
+    )
+set(PKGCONFIG_INSTALL_DIR
+    "${CMAKE_INSTALL_DATADIR}/pkgconfig"
+    CACHE PATH "The directory relative to CMAKE_PREFIX_PATH where eigen3.pc is installed"
+    )
+
 # similar to set_target_properties but append the property instead of overwriting it
 macro(ei_add_target_property target prop value)
 
@@ -324,21 +325,9 @@ install(FILES
   )
 
 if(EIGEN_BUILD_PKGCONFIG)
-    SET(path_separator ":")
+    configure_file(eigen3.pc.in eigen3.pc @ONLY)
-    STRING(REPLACE ${path_separator} ";" pkg_config_libdir_search "$ENV{PKG_CONFIG_LIBDIR}")
-    message(STATUS "searching for 'pkgconfig' directory in PKG_CONFIG_LIBDIR ( $ENV{PKG_CONFIG_LIBDIR} ), ${CMAKE_INSTALL_PREFIX}/share, and ${CMAKE_INSTALL_PREFIX}/lib")
-    FIND_PATH(pkg_config_libdir pkgconfig ${pkg_config_libdir_search} ${CMAKE_INSTALL_PREFIX}/share ${CMAKE_INSTALL_PREFIX}/lib ${pkg_config_libdir_search})
-    if(pkg_config_libdir)
-        SET(pkg_config_install_dir ${pkg_config_libdir})
-        message(STATUS "found ${pkg_config_libdir}/pkgconfig" )
-    else(pkg_config_libdir)
-        SET(pkg_config_install_dir ${CMAKE_INSTALL_PREFIX}/share)
-        message(STATUS "pkgconfig not found; installing in ${pkg_config_install_dir}" )
-    endif(pkg_config_libdir)
-
-    configure_file(eigen3.pc.in eigen3.pc)
     install(FILES ${CMAKE_CURRENT_BINARY_DIR}/eigen3.pc
-        DESTINATION ${pkg_config_install_dir}/pkgconfig
+        DESTINATION ${PKGCONFIG_INSTALL_DIR}
         )
 endif(EIGEN_BUILD_PKGCONFIG)
 
@@ -401,12 +390,15 @@ if(cmake_generator_tolower MATCHES "makefile")
   message(STATUS "--------------+--------------------------------------------------------------")
   message(STATUS "Command       |   Description")
   message(STATUS "--------------+--------------------------------------------------------------")
-  message(STATUS "make install  | Install to ${CMAKE_INSTALL_PREFIX}. To change that:")
+  message(STATUS "make install  | Install Eigen. Headers will be installed to:")
-  message(STATUS "              |     cmake . -DCMAKE_INSTALL_PREFIX=yourpath")
+  message(STATUS "              |     <CMAKE_INSTALL_PREFIX>/<INCLUDE_INSTALL_DIR>")
-  message(STATUS "              |   Eigen headers will then be installed to:")
+  message(STATUS "              |   Using the following values:")
-  message(STATUS "              |     ${INCLUDE_INSTALL_DIR}")
+  message(STATUS "              |     CMAKE_INSTALL_PREFIX: ${CMAKE_INSTALL_PREFIX}")
-  message(STATUS "              |   To install Eigen headers to a separate location, do:")
+  message(STATUS "              |     INCLUDE_INSTALL_DIR:  ${INCLUDE_INSTALL_DIR}")
-  message(STATUS "              |     cmake . -DEIGEN_INCLUDE_INSTALL_DIR=yourpath")
+  message(STATUS "              |   Change the install location of Eigen headers using:")
+  message(STATUS "              |     cmake . -DCMAKE_INSTALL_PREFIX=yourprefix")
+  message(STATUS "              |   Or:")
+  message(STATUS "              |     cmake . -DINCLUDE_INSTALL_DIR=yourdir")
   message(STATUS "make doc      | Generate the API documentation, requires Doxygen & LaTeX")
   message(STATUS "make check    | Build and run the unit-tests. Read this page:")
   message(STATUS "              |   http://eigen.tuxfamily.org/index.php?title=Tests")

Eigen/CholmodSupport

@@ -12,7 +12,7 @@ extern "C" {
 /** \ingroup Support_modules
   * \defgroup CholmodSupport_Module CholmodSupport module
   *
-  * This module provides an interface to the Cholmod library which is part of the <a href="http://www.cise.ufl.edu/research/sparse/SuiteSparse/">suitesparse</a> package.
+  * This module provides an interface to the Cholmod library which is part of the <a href="http://www.suitesparse.com">suitesparse</a> package.
   * It provides the two following main factorization classes:
   * - class CholmodSupernodalLLT: a supernodal LLT Cholesky factorization.
   * - class CholmodDecomposiiton: a general L(D)LT Cholesky factorization with automatic or explicit runtime selection of the underlying factorization method (supernodal or simplicial).

Eigen/Core

@@ -123,7 +123,7 @@
     #undef bool
     #undef vector
     #undef pixel
-  #elif defined  __ARM_NEON__
+  #elif defined  __ARM_NEON
     #define EIGEN_VECTORIZE
     #define EIGEN_VECTORIZE_NEON
     #include <arm_neon.h>

Eigen/SPQRSupport

@@ -10,7 +10,7 @@
 /** \ingroup Support_modules
   * \defgroup SPQRSupport_Module SuiteSparseQR module
   * 
-  * This module provides an interface to the SPQR library, which is part of the <a href="http://www.cise.ufl.edu/research/sparse/SuiteSparse/">suitesparse</a> package.
+  * This module provides an interface to the SPQR library, which is part of the <a href="http://www.suitesparse.com">suitesparse</a> package.
   *
   * \code
   * #include <Eigen/SPQRSupport>

Eigen/SparseCore

@@ -14,7 +14,7 @@
 /** 
   * \defgroup SparseCore_Module SparseCore module
   *
-  * This module provides a sparse matrix representation, and basic associatd matrix manipulations
+  * This module provides a sparse matrix representation, and basic associated matrix manipulations
   * and operations.
   *
   * See the \ref TutorialSparse "Sparse tutorial"

Eigen/UmfPackSupport

@@ -12,7 +12,7 @@ extern "C" {
 /** \ingroup Support_modules
   * \defgroup UmfPackSupport_Module UmfPackSupport module
   *
-  * This module provides an interface to the UmfPack library which is part of the <a href="http://www.cise.ufl.edu/research/sparse/SuiteSparse/">suitesparse</a> package.
+  * This module provides an interface to the UmfPack library which is part of the <a href="http://www.suitesparse.com">suitesparse</a> package.
   * It provides the following factorization class:
   * - class UmfPackLU: a multifrontal sequential LU factorization.
   *

Eigen/src/Cholesky/LDLT.h

@@ -235,6 +235,11 @@ template<typename _MatrixType, int _UpLo> class LDLT
     }
 
   protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
 
     /** \internal
       * Used to compute and store the Cholesky decomposition A = L D L^* = U^* D U.
@@ -434,6 +439,8 @@ template<typename MatrixType> struct LDLT_Traits<MatrixType,Upper>
 template<typename MatrixType, int _UpLo>
 LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::compute(const MatrixType& a)
 {
+  check_template_parameters();
+  
   eigen_assert(a.rows()==a.cols());
   const Index size = a.rows();
 
@@ -457,7 +464,7 @@ LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::compute(const MatrixType& a)
   */
 template<typename MatrixType, int _UpLo>
 template<typename Derived>
-LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::rankUpdate(const MatrixBase<Derived>& w, const typename NumTraits<typename MatrixType::Scalar>::Real& sigma)
+LDLT<MatrixType,_UpLo>& LDLT<MatrixType,_UpLo>::rankUpdate(const MatrixBase<Derived>& w, const typename LDLT<MatrixType,_UpLo>::RealScalar& sigma)
 {
   const Index size = w.rows();
   if (m_isInitialized)

Eigen/src/Cholesky/LLT.h

@@ -174,6 +174,12 @@ template<typename _MatrixType, int _UpLo> class LLT
     LLT rankUpdate(const VectorType& vec, const RealScalar& sigma = 1);
 
   protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+    
     /** \internal
       * Used to compute and store L
       * The strict upper part is not used and even not initialized.
@@ -283,7 +289,7 @@ template<typename Scalar> struct llt_inplace<Scalar, Lower>
         return k;
       mat.coeffRef(k,k) = x = sqrt(x);
       if (k>0 && rs>0) A21.noalias() -= A20 * A10.adjoint();
-      if (rs>0) A21 *= RealScalar(1)/x;
+      if (rs>0) A21 /= x;
     }
     return -1;
   }
@@ -384,6 +390,8 @@ template<typename MatrixType> struct LLT_Traits<MatrixType,Upper>
 template<typename MatrixType, int _UpLo>
 LLT<MatrixType,_UpLo>& LLT<MatrixType,_UpLo>::compute(const MatrixType& a)
 {
+  check_template_parameters();
+  
   eigen_assert(a.rows()==a.cols());
   const Index size = a.rows();
   m_matrix.resize(size, size);

Eigen/src/Cholesky/LLT_MKL.h

@@ -60,7 +60,7 @@ template<> struct mkl_llt<EIGTYPE> \
     lda = m.outerStride(); \
 \
     info = LAPACKE_##MKLPREFIX##potrf( matrix_order, uplo, size, (MKLTYPE*)a, lda ); \
-    info = (info==0) ? Success : NumericalIssue; \
+    info = (info==0) ? -1 : info>0 ? info-1 : size; \
     return info; \
   } \
 }; \

Eigen/src/CholmodSupport/CholmodSupport.h

@@ -78,7 +78,7 @@ cholmod_sparse viewAsCholmod(SparseMatrix<_Scalar,_Options,_Index>& mat)
   {
     res.itype = CHOLMOD_INT;
   }
-  else if (internal::is_same<_Index,UF_long>::value)
+  else if (internal::is_same<_Index,SuiteSparse_long>::value)
   {
     res.itype = CHOLMOD_LONG;
   }
@@ -395,7 +395,7 @@ class CholmodSimplicialLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimpl
     CholmodSimplicialLLT(const MatrixType& matrix) : Base()
     {
       init();
-      compute(matrix);
+      Base::compute(matrix);
     }
 
     ~CholmodSimplicialLLT() {}
@@ -442,7 +442,7 @@ class CholmodSimplicialLDLT : public CholmodBase<_MatrixType, _UpLo, CholmodSimp
     CholmodSimplicialLDLT(const MatrixType& matrix) : Base()
     {
       init();
-      compute(matrix);
+      Base::compute(matrix);
     }
 
     ~CholmodSimplicialLDLT() {}
@@ -487,7 +487,7 @@ class CholmodSupernodalLLT : public CholmodBase<_MatrixType, _UpLo, CholmodSuper
     CholmodSupernodalLLT(const MatrixType& matrix) : Base()
     {
       init();
-      compute(matrix);
+      Base::compute(matrix);
     }
 
     ~CholmodSupernodalLLT() {}
@@ -534,7 +534,7 @@ class CholmodDecomposition : public CholmodBase<_MatrixType, _UpLo, CholmodDecom
     CholmodDecomposition(const MatrixType& matrix) : Base()
     {
       init();
-      compute(matrix);
+      Base::compute(matrix);
     }
 
     ~CholmodDecomposition() {}

Eigen/src/Core/Array.h

@@ -124,6 +124,21 @@ class Array
     }
 #endif
 
+#ifdef EIGEN_HAVE_RVALUE_REFERENCES
+    Array(Array&& other)
+      : Base(std::move(other))
+    {
+      Base::_check_template_params();
+      if (RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic)
+        Base::_set_noalias(other);
+    }
+    Array& operator=(Array&& other)
+    {
+      other.swap(*this);
+      return *this;
+    }
+#endif
+
     /** Constructs a vector or row-vector with given dimension. \only_for_vectors
       *
       * Note that this is only useful for dynamic-size vectors. For fixed-size vectors,

Eigen/src/Core/ArrayBase.h

@@ -46,9 +46,6 @@ template<typename Derived> class ArrayBase
 
     typedef ArrayBase Eigen_BaseClassForSpecializationOfGlobalMathFuncImpl;
 
-    using internal::special_scalar_op_base<Derived,typename internal::traits<Derived>::Scalar,
-                typename NumTraits<typename internal::traits<Derived>::Scalar>::Real>::operator*;
-
     typedef typename internal::traits<Derived>::StorageKind StorageKind;
     typedef typename internal::traits<Derived>::Index Index;
     typedef typename internal::traits<Derived>::Scalar Scalar;
@@ -56,6 +53,7 @@ template<typename Derived> class ArrayBase
     typedef typename NumTraits<Scalar>::Real RealScalar;
 
     typedef DenseBase<Derived> Base;
+    using Base::operator*;
     using Base::RowsAtCompileTime;
     using Base::ColsAtCompileTime;
     using Base::SizeAtCompileTime;

Eigen/src/Core/Assign.h

@@ -439,19 +439,26 @@ struct assign_impl<Derived1, Derived2, SliceVectorizedTraversal, NoUnrolling, Ve
   typedef typename Derived1::Index Index;
   static inline void run(Derived1 &dst, const Derived2 &src)
   {
-    typedef packet_traits<typename Derived1::Scalar> PacketTraits;
+    typedef typename Derived1::Scalar Scalar;
+    typedef packet_traits<Scalar> PacketTraits;
     enum {
       packetSize = PacketTraits::size,
       alignable = PacketTraits::AlignedOnScalar,
-      dstAlignment = alignable ? Aligned : int(assign_traits<Derived1,Derived2>::DstIsAligned) ,
+      dstIsAligned = assign_traits<Derived1,Derived2>::DstIsAligned,
+      dstAlignment = alignable ? Aligned : int(dstIsAligned),
       srcAlignment = assign_traits<Derived1,Derived2>::JointAlignment
     };
+    const Scalar *dst_ptr = &dst.coeffRef(0,0);
+    if((!bool(dstIsAligned)) && (size_t(dst_ptr) % sizeof(Scalar))>0)
+    {
+      // the pointer is not aligend-on scalar, so alignment is not possible
+      return assign_impl<Derived1,Derived2,DefaultTraversal,NoUnrolling>::run(dst, src);
+    }
     const Index packetAlignedMask = packetSize - 1;
     const Index innerSize = dst.innerSize();
     const Index outerSize = dst.outerSize();
     const Index alignedStep = alignable ? (packetSize - dst.outerStride() % packetSize) & packetAlignedMask : 0;
-    Index alignedStart = ((!alignable) || assign_traits<Derived1,Derived2>::DstIsAligned) ? 0
+    Index alignedStart = ((!alignable) || bool(dstIsAligned)) ? 0 : internal::first_aligned(dst_ptr, innerSize);
-                       : internal::first_aligned(&dst.coeffRef(0,0), innerSize);
 
     for(Index outer = 0; outer < outerSize; ++outer)
     {

Eigen/src/Core/Block.h

@@ -66,8 +66,9 @@ struct traits<Block<XprType, BlockRows, BlockCols, InnerPanel> > : traits<XprTyp
                          : ColsAtCompileTime != Dynamic ? int(ColsAtCompileTime)
                          : int(traits<XprType>::MaxColsAtCompileTime),
     XprTypeIsRowMajor = (int(traits<XprType>::Flags)&RowMajorBit) != 0,
-    IsRowMajor = (MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1
+    IsDense = is_same<StorageKind,Dense>::value,
-               : (MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0
+    IsRowMajor = (IsDense&&MaxRowsAtCompileTime==1&&MaxColsAtCompileTime!=1) ? 1
+               : (IsDense&&MaxColsAtCompileTime==1&&MaxRowsAtCompileTime!=1) ? 0
                : XprTypeIsRowMajor,
     HasSameStorageOrderAsXprType = (IsRowMajor == XprTypeIsRowMajor),
     InnerSize = IsRowMajor ? int(ColsAtCompileTime) : int(RowsAtCompileTime),

Eigen/src/Core/CommaInitializer.h

@@ -76,8 +76,11 @@ struct CommaInitializer
   template<typename OtherDerived>
   CommaInitializer& operator,(const DenseBase<OtherDerived>& other)
   {
-    if(other.cols()==0 || other.rows()==0)
+    if(other.rows()==0)
+    {
+      m_col += other.cols();
       return *this;
+    }
     if (m_col==m_xpr.cols())
     {
       m_row+=m_currentBlockRows;
@@ -86,7 +89,7 @@ struct CommaInitializer
       eigen_assert(m_row+m_currentBlockRows<=m_xpr.rows()
         && "Too many rows passed to comma initializer (operator<<)");
     }
-    eigen_assert(m_col<m_xpr.cols()
+    eigen_assert((m_col<m_xpr.cols() || (m_xpr.cols()==0 && m_col==0))
       && "Too many coefficients passed to comma initializer (operator<<)");
     eigen_assert(m_currentBlockRows==other.rows());
     if (OtherDerived::SizeAtCompileTime != Dynamic)

Eigen/src/Core/CwiseBinaryOp.h

@@ -81,7 +81,8 @@ struct traits<CwiseBinaryOp<BinaryOp, Lhs, Rhs> >
         )
      ),
     Flags = (Flags0 & ~RowMajorBit) | (LhsFlags & RowMajorBit),
-    CoeffReadCost = LhsCoeffReadCost + RhsCoeffReadCost + functor_traits<BinaryOp>::Cost
+    Cost0 = EIGEN_ADD_COST(LhsCoeffReadCost,RhsCoeffReadCost),
+    CoeffReadCost = EIGEN_ADD_COST(Cost0,functor_traits<BinaryOp>::Cost)
   };
 };
 } // end namespace internal

Eigen/src/Core/CwiseUnaryOp.h

@@ -47,7 +47,7 @@ struct traits<CwiseUnaryOp<UnaryOp, XprType> >
     Flags = _XprTypeNested::Flags & (
       HereditaryBits | LinearAccessBit | AlignedBit
       | (functor_traits<UnaryOp>::PacketAccess ? PacketAccessBit : 0)),
-    CoeffReadCost = _XprTypeNested::CoeffReadCost + functor_traits<UnaryOp>::Cost
+    CoeffReadCost = EIGEN_ADD_COST(_XprTypeNested::CoeffReadCost, functor_traits<UnaryOp>::Cost)
   };
 };
 }

Eigen/src/Core/CwiseUnaryView.h

@@ -38,7 +38,7 @@ struct traits<CwiseUnaryView<ViewOp, MatrixType> >
   typedef typename remove_all<MatrixTypeNested>::type _MatrixTypeNested;
   enum {
     Flags = (traits<_MatrixTypeNested>::Flags & (HereditaryBits | LvalueBit | LinearAccessBit | DirectAccessBit)),
-    CoeffReadCost = traits<_MatrixTypeNested>::CoeffReadCost + functor_traits<ViewOp>::Cost,
+    CoeffReadCost = EIGEN_ADD_COST(traits<_MatrixTypeNested>::CoeffReadCost, functor_traits<ViewOp>::Cost),
     MatrixTypeInnerStride =  inner_stride_at_compile_time<MatrixType>::ret,
     // need to cast the sizeof's from size_t to int explicitly, otherwise:
     // "error: no integral type can represent all of the enumerator values

Eigen/src/Core/DenseBase.h

@@ -40,15 +40,14 @@ static inline void check_DenseIndex_is_signed() {
   */
 template<typename Derived> class DenseBase
 #ifndef EIGEN_PARSED_BY_DOXYGEN
-  : public internal::special_scalar_op_base<Derived,typename internal::traits<Derived>::Scalar,
+  : public internal::special_scalar_op_base<Derived, typename internal::traits<Derived>::Scalar,
-                                     typename NumTraits<typename internal::traits<Derived>::Scalar>::Real>
+                                            typename NumTraits<typename internal::traits<Derived>::Scalar>::Real,
+                                            DenseCoeffsBase<Derived> >
 #else
   : public DenseCoeffsBase<Derived>
 #endif // not EIGEN_PARSED_BY_DOXYGEN
 {
   public:
-    using internal::special_scalar_op_base<Derived,typename internal::traits<Derived>::Scalar,
-                typename NumTraits<typename internal::traits<Derived>::Scalar>::Real>::operator*;
 
     class InnerIterator;
 
@@ -63,8 +62,9 @@ template<typename Derived> class DenseBase
     typedef typename internal::traits<Derived>::Scalar Scalar;
     typedef typename internal::packet_traits<Scalar>::type PacketScalar;
     typedef typename NumTraits<Scalar>::Real RealScalar;
+    typedef internal::special_scalar_op_base<Derived,Scalar,RealScalar, DenseCoeffsBase<Derived> > Base;
 
-    typedef DenseCoeffsBase<Derived> Base;
+    using Base::operator*;
     using Base::derived;
     using Base::const_cast_derived;
     using Base::rows;
@@ -183,10 +183,6 @@ template<typename Derived> class DenseBase
     /** \returns the number of nonzero coefficients which is in practice the number
       * of stored coefficients. */
     inline Index nonZeros() const { return size(); }
-    /** \returns true if either the number of rows or the number of columns is equal to 1.
-      * In other words, this function returns
-      * \code rows()==1 || cols()==1 \endcode
-      * \sa rows(), cols(), IsVectorAtCompileTime. */
 
     /** \returns the outer size.
       *
@@ -266,11 +262,13 @@ template<typename Derived> class DenseBase
     template<typename OtherDerived>
     Derived& operator=(const ReturnByValue<OtherDerived>& func);
 
-#ifndef EIGEN_PARSED_BY_DOXYGEN
+    /** \internal Copies \a other into *this without evaluating other. \returns a reference to *this. */
-    /** Copies \a other into *this without evaluating other. \returns a reference to *this. */
     template<typename OtherDerived>
     Derived& lazyAssign(const DenseBase<OtherDerived>& other);
-#endif // not EIGEN_PARSED_BY_DOXYGEN
+
+    /** \internal Evaluates \a other into *this. \returns a reference to *this. */
+    template<typename OtherDerived>
+    Derived& lazyAssign(const ReturnByValue<OtherDerived>& other);
 
     CommaInitializer<Derived> operator<< (const Scalar& s);
 

Eigen/src/Core/DenseStorage.h

@@ -122,33 +122,41 @@ template<typename T, int Size, int _Rows, int _Cols, int _Options> class DenseSt
 {
     internal::plain_array<T,Size,_Options> m_data;
   public:
-    inline DenseStorage() {}
+    DenseStorage() {}
-    inline DenseStorage(internal::constructor_without_unaligned_array_assert)
+    DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()) {}
-    inline DenseStorage(DenseIndex,DenseIndex,DenseIndex) {}
+    DenseStorage(const DenseStorage& other) : m_data(other.m_data) {}
-    inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); }
+    DenseStorage& operator=(const DenseStorage& other)
-    static inline DenseIndex rows(void) {return _Rows;}
+    {
-    static inline DenseIndex cols(void) {return _Cols;}
+      if (this != &other) m_data = other.m_data;
-    inline void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {}
+      return *this;
-    inline void resize(DenseIndex,DenseIndex,DenseIndex) {}
+    }
-    inline const T *data() const { return m_data.array; }
+    DenseStorage(DenseIndex,DenseIndex,DenseIndex) {}
-    inline T *data() { return m_data.array; }
+    void swap(DenseStorage& other) { std::swap(m_data,other.m_data); }
+    static DenseIndex rows(void) {return _Rows;}
+    static DenseIndex cols(void) {return _Cols;}
+    void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {}
+    void resize(DenseIndex,DenseIndex,DenseIndex) {}
+    const T *data() const { return m_data.array; }
+    T *data() { return m_data.array; }
 };
 
 // null matrix
 template<typename T, int _Rows, int _Cols, int _Options> class DenseStorage<T, 0, _Rows, _Cols, _Options>
 {
   public:
-    inline DenseStorage() {}
+    DenseStorage() {}
-    inline DenseStorage(internal::constructor_without_unaligned_array_assert) {}
+    DenseStorage(internal::constructor_without_unaligned_array_assert) {}
-    inline DenseStorage(DenseIndex,DenseIndex,DenseIndex) {}
+    DenseStorage(const DenseStorage&) {}
-    inline void swap(DenseStorage& ) {}
+    DenseStorage& operator=(const DenseStorage&) { return *this; }
-    static inline DenseIndex rows(void) {return _Rows;}
+    DenseStorage(DenseIndex,DenseIndex,DenseIndex) {}
-    static inline DenseIndex cols(void) {return _Cols;}
+    void swap(DenseStorage& ) {}
-    inline void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {}
+    static DenseIndex rows(void) {return _Rows;}
-    inline void resize(DenseIndex,DenseIndex,DenseIndex) {}
+    static DenseIndex cols(void) {return _Cols;}
-    inline const T *data() const { return 0; }
+    void conservativeResize(DenseIndex,DenseIndex,DenseIndex) {}
-    inline T *data() { return 0; }
+    void resize(DenseIndex,DenseIndex,DenseIndex) {}
+    const T *data() const { return 0; }
+    T *data() { return 0; }
 };
 
 // more specializations for null matrices; these are necessary to resolve ambiguities
@@ -168,18 +176,29 @@ template<typename T, int Size, int _Options> class DenseStorage<T, Size, Dynamic
     DenseIndex m_rows;
     DenseIndex m_cols;
   public:
-    inline DenseStorage() : m_rows(0), m_cols(0) {}
+    DenseStorage() : m_rows(0), m_cols(0) {}
-    inline DenseStorage(internal::constructor_without_unaligned_array_assert)
+    DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()), m_rows(0), m_cols(0) {}
-    inline DenseStorage(DenseIndex, DenseIndex nbRows, DenseIndex nbCols) : m_rows(nbRows), m_cols(nbCols) {}
+    DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows), m_cols(other.m_cols) {}
-    inline void swap(DenseStorage& other)
+    DenseStorage& operator=(const DenseStorage& other)
+    {
+      if (this != &other)
+      {
+        m_data = other.m_data;
+        m_rows = other.m_rows;
+        m_cols = other.m_cols;
+      }
+      return *this;
+    }
+    DenseStorage(DenseIndex, DenseIndex nbRows, DenseIndex nbCols) : m_rows(nbRows), m_cols(nbCols) {}
+    void swap(DenseStorage& other)
     { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); }
-    inline DenseIndex rows() const {return m_rows;}
+    DenseIndex rows() const {return m_rows;}
-    inline DenseIndex cols() const {return m_cols;}
+    DenseIndex cols() const {return m_cols;}
-    inline void conservativeResize(DenseIndex, DenseIndex nbRows, DenseIndex nbCols) { m_rows = nbRows; m_cols = nbCols; }
+    void conservativeResize(DenseIndex, DenseIndex nbRows, DenseIndex nbCols) { m_rows = nbRows; m_cols = nbCols; }
-    inline void resize(DenseIndex, DenseIndex nbRows, DenseIndex nbCols) { m_rows = nbRows; m_cols = nbCols; }
+    void resize(DenseIndex, DenseIndex nbRows, DenseIndex nbCols) { m_rows = nbRows; m_cols = nbCols; }
-    inline const T *data() const { return m_data.array; }
+    const T *data() const { return m_data.array; }
-    inline T *data() { return m_data.array; }
+    T *data() { return m_data.array; }
 };
 
 // dynamic-size matrix with fixed-size storage and fixed width
@@ -188,17 +207,27 @@ template<typename T, int Size, int _Cols, int _Options> class DenseStorage<T, Si
     internal::plain_array<T,Size,_Options> m_data;
     DenseIndex m_rows;
   public:
-    inline DenseStorage() : m_rows(0) {}
+    DenseStorage() : m_rows(0) {}
-    inline DenseStorage(internal::constructor_without_unaligned_array_assert)
+    DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()), m_rows(0) {}
-    inline DenseStorage(DenseIndex, DenseIndex nbRows, DenseIndex) : m_rows(nbRows) {}
+    DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_rows(other.m_rows) {}
-    inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
+    DenseStorage& operator=(const DenseStorage& other)
-    inline DenseIndex rows(void) const {return m_rows;}
+    {
-    inline DenseIndex cols(void) const {return _Cols;}
+      if (this != &other)
-    inline void conservativeResize(DenseIndex, DenseIndex nbRows, DenseIndex) { m_rows = nbRows; }
+      {
-    inline void resize(DenseIndex, DenseIndex nbRows, DenseIndex) { m_rows = nbRows; }
+        m_data = other.m_data;
-    inline const T *data() const { return m_data.array; }
+        m_rows = other.m_rows;
-    inline T *data() { return m_data.array; }
+      }
+      return *this;
+    }
+    DenseStorage(DenseIndex, DenseIndex nbRows, DenseIndex) : m_rows(nbRows) {}
+    void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
+    DenseIndex rows(void) const {return m_rows;}
+    DenseIndex cols(void) const {return _Cols;}
+    void conservativeResize(DenseIndex, DenseIndex nbRows, DenseIndex) { m_rows = nbRows; }
+    void resize(DenseIndex, DenseIndex nbRows, DenseIndex) { m_rows = nbRows; }
+    const T *data() const { return m_data.array; }
+    T *data() { return m_data.array; }
 };
 
 // dynamic-size matrix with fixed-size storage and fixed height
@@ -207,17 +236,27 @@ template<typename T, int Size, int _Rows, int _Options> class DenseStorage<T, Si
     internal::plain_array<T,Size,_Options> m_data;
     DenseIndex m_cols;
   public:
-    inline DenseStorage() : m_cols(0) {}
+    DenseStorage() : m_cols(0) {}
-    inline DenseStorage(internal::constructor_without_unaligned_array_assert)
+    DenseStorage(internal::constructor_without_unaligned_array_assert)
       : m_data(internal::constructor_without_unaligned_array_assert()), m_cols(0) {}
-    inline DenseStorage(DenseIndex, DenseIndex, DenseIndex nbCols) : m_cols(nbCols) {}
+    DenseStorage(const DenseStorage& other) : m_data(other.m_data), m_cols(other.m_cols) {}
-    inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
+    DenseStorage& operator=(const DenseStorage& other)
-    inline DenseIndex rows(void) const {return _Rows;}
+    {
-    inline DenseIndex cols(void) const {return m_cols;}
+      if (this != &other)
-    inline void conservativeResize(DenseIndex, DenseIndex, DenseIndex nbCols) { m_cols = nbCols; }
+      {
-    inline void resize(DenseIndex, DenseIndex, DenseIndex nbCols) { m_cols = nbCols; }
+        m_data = other.m_data;
-    inline const T *data() const { return m_data.array; }
+        m_cols = other.m_cols;
-    inline T *data() { return m_data.array; }
+      }
+      return *this;
+    }
+    DenseStorage(DenseIndex, DenseIndex, DenseIndex nbCols) : m_cols(nbCols) {}
+    void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
+    DenseIndex rows(void) const {return _Rows;}
+    DenseIndex cols(void) const {return m_cols;}
+    void conservativeResize(DenseIndex, DenseIndex, DenseIndex nbCols) { m_cols = nbCols; }
+    void resize(DenseIndex, DenseIndex, DenseIndex nbCols) { m_cols = nbCols; }
+    const T *data() const { return m_data.array; }
+    T *data() { return m_data.array; }
 };
 
 // purely dynamic matrix.
@@ -227,18 +266,35 @@ template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynam
     DenseIndex m_rows;
     DenseIndex m_cols;
   public:
-    inline DenseStorage() : m_data(0), m_rows(0), m_cols(0) {}
+    DenseStorage() : m_data(0), m_rows(0), m_cols(0) {}
-    inline DenseStorage(internal::constructor_without_unaligned_array_assert)
+    DenseStorage(internal::constructor_without_unaligned_array_assert)
        : m_data(0), m_rows(0), m_cols(0) {}
-    inline DenseStorage(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols)
+    DenseStorage(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols)
       : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(nbRows), m_cols(nbCols)
     { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN }
-    inline ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, m_rows*m_cols); }
+#ifdef EIGEN_HAVE_RVALUE_REFERENCES
-    inline void swap(DenseStorage& other)
+    DenseStorage(DenseStorage&& other)
+      : m_data(std::move(other.m_data))
+      , m_rows(std::move(other.m_rows))
+      , m_cols(std::move(other.m_cols))
+    {
+      other.m_data = nullptr;
+    }
+    DenseStorage& operator=(DenseStorage&& other)
+    {
+      using std::swap;
+      swap(m_data, other.m_data);
+      swap(m_rows, other.m_rows);
+      swap(m_cols, other.m_cols);
+      return *this;
+    }
+#endif
+    ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, m_rows*m_cols); }
+    void swap(DenseStorage& other)
     { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); std::swap(m_cols,other.m_cols); }
-    inline DenseIndex rows(void) const {return m_rows;}
+    DenseIndex rows(void) const {return m_rows;}
-    inline DenseIndex cols(void) const {return m_cols;}
+    DenseIndex cols(void) const {return m_cols;}
-    inline void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols)
+    void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex nbCols)
     {
       m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, m_rows*m_cols);
       m_rows = nbRows;
@@ -258,8 +314,11 @@ template<typename T, int _Options> class DenseStorage<T, Dynamic, Dynamic, Dynam
       m_rows = nbRows;
       m_cols = nbCols;
     }
-    inline const T *data() const { return m_data; }
+    const T *data() const { return m_data; }
-    inline T *data() { return m_data; }
+    T *data() { return m_data; }
+  private:
+    DenseStorage(const DenseStorage&);
+    DenseStorage& operator=(const DenseStorage&);
 };
 
 // matrix with dynamic width and fixed height (so that matrix has dynamic size).
@@ -268,15 +327,30 @@ template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Ro
     T *m_data;
     DenseIndex m_cols;
   public:
-    inline DenseStorage() : m_data(0), m_cols(0) {}
+    DenseStorage() : m_data(0), m_cols(0) {}
-    inline DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_cols(0) {}
+    DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_cols(0) {}
-    inline DenseStorage(DenseIndex size, DenseIndex, DenseIndex nbCols) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_cols(nbCols)
+    DenseStorage(DenseIndex size, DenseIndex, DenseIndex nbCols) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_cols(nbCols)
     { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN }
-    inline ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Rows*m_cols); }
+#ifdef EIGEN_HAVE_RVALUE_REFERENCES
-    inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
+    DenseStorage(DenseStorage&& other)
-    static inline DenseIndex rows(void) {return _Rows;}
+      : m_data(std::move(other.m_data))
-    inline DenseIndex cols(void) const {return m_cols;}
+      , m_cols(std::move(other.m_cols))
-    inline void conservativeResize(DenseIndex size, DenseIndex, DenseIndex nbCols)
+    {
+      other.m_data = nullptr;
+    }
+    DenseStorage& operator=(DenseStorage&& other)
+    {
+      using std::swap;
+      swap(m_data, other.m_data);
+      swap(m_cols, other.m_cols);
+      return *this;
+    }
+#endif
+    ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Rows*m_cols); }
+    void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_cols,other.m_cols); }
+    static DenseIndex rows(void) {return _Rows;}
+    DenseIndex cols(void) const {return m_cols;}
+    void conservativeResize(DenseIndex size, DenseIndex, DenseIndex nbCols)
     {
       m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, _Rows*m_cols);
       m_cols = nbCols;
@@ -294,8 +368,11 @@ template<typename T, int _Rows, int _Options> class DenseStorage<T, Dynamic, _Ro
       }
       m_cols = nbCols;
     }
-    inline const T *data() const { return m_data; }
+    const T *data() const { return m_data; }
-    inline T *data() { return m_data; }
+    T *data() { return m_data; }
+  private:
+    DenseStorage(const DenseStorage&);
+    DenseStorage& operator=(const DenseStorage&);
 };
 
 // matrix with dynamic height and fixed width (so that matrix has dynamic size).
@@ -304,15 +381,30 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dyn
     T *m_data;
     DenseIndex m_rows;
   public:
-    inline DenseStorage() : m_data(0), m_rows(0) {}
+    DenseStorage() : m_data(0), m_rows(0) {}
-    inline DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_rows(0) {}
+    DenseStorage(internal::constructor_without_unaligned_array_assert) : m_data(0), m_rows(0) {}
-    inline DenseStorage(DenseIndex size, DenseIndex nbRows, DenseIndex) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(nbRows)
+    DenseStorage(DenseIndex size, DenseIndex nbRows, DenseIndex) : m_data(internal::conditional_aligned_new_auto<T,(_Options&DontAlign)==0>(size)), m_rows(nbRows)
     { EIGEN_INTERNAL_DENSE_STORAGE_CTOR_PLUGIN }
-    inline ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Cols*m_rows); }
+#ifdef EIGEN_HAVE_RVALUE_REFERENCES
-    inline void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
+    DenseStorage(DenseStorage&& other)
-    inline DenseIndex rows(void) const {return m_rows;}
+      : m_data(std::move(other.m_data))
-    static inline DenseIndex cols(void) {return _Cols;}
+      , m_rows(std::move(other.m_rows))
-    inline void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex)
+    {
+      other.m_data = nullptr;
+    }
+    DenseStorage& operator=(DenseStorage&& other)
+    {
+      using std::swap;
+      swap(m_data, other.m_data);
+      swap(m_rows, other.m_rows);
+      return *this;
+    }
+#endif
+    ~DenseStorage() { internal::conditional_aligned_delete_auto<T,(_Options&DontAlign)==0>(m_data, _Cols*m_rows); }
+    void swap(DenseStorage& other) { std::swap(m_data,other.m_data); std::swap(m_rows,other.m_rows); }
+    DenseIndex rows(void) const {return m_rows;}
+    static DenseIndex cols(void) {return _Cols;}
+    void conservativeResize(DenseIndex size, DenseIndex nbRows, DenseIndex)
     {
       m_data = internal::conditional_aligned_realloc_new_auto<T,(_Options&DontAlign)==0>(m_data, size, m_rows*_Cols);
       m_rows = nbRows;
@@ -330,8 +422,11 @@ template<typename T, int _Cols, int _Options> class DenseStorage<T, Dynamic, Dyn
       }
       m_rows = nbRows;
     }
-    inline const T *data() const { return m_data; }
+    const T *data() const { return m_data; }
-    inline T *data() { return m_data; }
+    T *data() { return m_data; }
+  private:
+    DenseStorage(const DenseStorage&);
+    DenseStorage& operator=(const DenseStorage&);
 };
 
 } // end namespace Eigen

Eigen/src/Core/DiagonalMatrix.h

@@ -44,10 +44,10 @@ class DiagonalBase : public EigenBase<Derived>
     template<typename DenseDerived>
     void evalTo(MatrixBase<DenseDerived> &other) const;
     template<typename DenseDerived>
-    void addTo(MatrixBase<DenseDerived> &other) const
+    inline void addTo(MatrixBase<DenseDerived> &other) const
     { other.diagonal() += diagonal(); }
     template<typename DenseDerived>
-    void subTo(MatrixBase<DenseDerived> &other) const
+    inline void subTo(MatrixBase<DenseDerived> &other) const
     { other.diagonal() -= diagonal(); }
 
     inline const DiagonalVectorType& diagonal() const { return derived().diagonal(); }
@@ -98,7 +98,7 @@ class DiagonalBase : public EigenBase<Derived>
 
 template<typename Derived>
 template<typename DenseDerived>
-void DiagonalBase<Derived>::evalTo(MatrixBase<DenseDerived> &other) const
+inline void DiagonalBase<Derived>::evalTo(MatrixBase<DenseDerived> &other) const
 {
   other.setZero();
   other.diagonal() = diagonal();

Eigen/src/Core/DiagonalProduct.h

@@ -34,8 +34,9 @@ struct traits<DiagonalProduct<MatrixType, DiagonalType, ProductOrder> >
     _Vectorizable = bool(int(MatrixType::Flags)&PacketAccessBit) && _SameTypes && (_ScalarAccessOnDiag || (bool(int(DiagonalType::DiagonalVectorType::Flags)&PacketAccessBit))),
     _LinearAccessMask = (RowsAtCompileTime==1 || ColsAtCompileTime==1) ? LinearAccessBit : 0,
 
-    Flags = ((HereditaryBits|_LinearAccessMask) & (unsigned int)(MatrixType::Flags)) | (_Vectorizable ? PacketAccessBit : 0) | AlignedBit,//(int(MatrixType::Flags)&int(DiagonalType::DiagonalVectorType::Flags)&AlignedBit),
+    Flags = ((HereditaryBits|_LinearAccessMask|AlignedBit) & (unsigned int)(MatrixType::Flags)) | (_Vectorizable ? PacketAccessBit : 0),//(int(MatrixType::Flags)&int(DiagonalType::DiagonalVectorType::Flags)&AlignedBit),
-    CoeffReadCost = NumTraits<Scalar>::MulCost + MatrixType::CoeffReadCost + DiagonalType::DiagonalVectorType::CoeffReadCost
+    Cost0 = EIGEN_ADD_COST(NumTraits<Scalar>::MulCost, MatrixType::CoeffReadCost),
+    CoeffReadCost = EIGEN_ADD_COST(Cost0,DiagonalType::DiagonalVectorType::CoeffReadCost)
   };
 };
 }

Eigen/src/Core/Dot.h

@@ -59,7 +59,7 @@ struct dot_nocheck<T, U, true>
   */
 template<typename Derived>
 template<typename OtherDerived>
-typename internal::scalar_product_traits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType
+inline typename internal::scalar_product_traits<typename internal::traits<Derived>::Scalar,typename internal::traits<OtherDerived>::Scalar>::ReturnType
 MatrixBase<Derived>::dot(const MatrixBase<OtherDerived>& other) const
 {
   EIGEN_STATIC_ASSERT_VECTOR_ONLY(Derived)

Eigen/src/Core/Functors.h

@@ -259,6 +259,47 @@ template<> struct functor_traits<scalar_boolean_or_op> {
   };
 };
 
+/** \internal
+  * \brief Template functors for comparison of two scalars
+  * \todo Implement packet-comparisons
+  */
+template<typename Scalar, ComparisonName cmp> struct scalar_cmp_op;
+
+template<typename Scalar, ComparisonName cmp>
+struct functor_traits<scalar_cmp_op<Scalar, cmp> > {
+  enum {
+    Cost = NumTraits<Scalar>::AddCost,
+    PacketAccess = false
+  };
+};
+
+template<ComparisonName Cmp, typename Scalar>
+struct result_of<scalar_cmp_op<Scalar, Cmp>(Scalar,Scalar)> {
+  typedef bool type;
+};
+
+
+template<typename Scalar> struct scalar_cmp_op<Scalar, cmp_EQ> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_STRONG_INLINE bool operator()(const Scalar& a, const Scalar& b) const {return a==b;}
+};
+template<typename Scalar> struct scalar_cmp_op<Scalar, cmp_LT> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_STRONG_INLINE bool operator()(const Scalar& a, const Scalar& b) const {return a<b;}
+};
+template<typename Scalar> struct scalar_cmp_op<Scalar, cmp_LE> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_STRONG_INLINE bool operator()(const Scalar& a, const Scalar& b) const {return a<=b;}
+};
+template<typename Scalar> struct scalar_cmp_op<Scalar, cmp_UNORD> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_STRONG_INLINE bool operator()(const Scalar& a, const Scalar& b) const {return !(a<=b || b<=a);}
+};
+template<typename Scalar> struct scalar_cmp_op<Scalar, cmp_NEQ> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cmp_op)
+  EIGEN_STRONG_INLINE bool operator()(const Scalar& a, const Scalar& b) const {return a!=b;}
+};
+
 // unary functors:
 
 /** \internal

Eigen/src/Core/GeneralProduct.h

@@ -205,9 +205,6 @@ class GeneralProduct<Lhs, Rhs, InnerProduct>
   public:
     GeneralProduct(const Lhs& lhs, const Rhs& rhs)
     {
-      EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::RealScalar, typename Rhs::RealScalar>::value),
-        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
-
       Base::coeffRef(0,0) = (lhs.transpose().cwiseProduct(rhs)).sum();
     }
 
@@ -257,15 +254,13 @@ template<typename Lhs, typename Rhs>
 class GeneralProduct<Lhs, Rhs, OuterProduct>
   : public ProductBase<GeneralProduct<Lhs,Rhs,OuterProduct>, Lhs, Rhs>
 {
-    template<typename T> struct IsRowMajor : internal::conditional<(int(T::Flags)&RowMajorBit), internal::true_type, internal::false_type>::type {};
+    template<typename T> struct is_row_major : internal::conditional<(int(T::Flags)&RowMajorBit), internal::true_type, internal::false_type>::type {};
     
   public:
     EIGEN_PRODUCT_PUBLIC_INTERFACE(GeneralProduct)
 
     GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs)
     {
-      EIGEN_STATIC_ASSERT((internal::is_same<typename Lhs::RealScalar, typename Rhs::RealScalar>::value),
-        YOU_MIXED_DIFFERENT_NUMERIC_TYPES__YOU_NEED_TO_USE_THE_CAST_METHOD_OF_MATRIXBASE_TO_CAST_NUMERIC_TYPES_EXPLICITLY)
     }
     
     struct set  { template<typename Dst, typename Src> void operator()(const Dst& dst, const Src& src) const { dst.const_cast_derived()  = src; } };
@@ -281,22 +276,22 @@ class GeneralProduct<Lhs, Rhs, OuterProduct>
     
     template<typename Dest>
     inline void evalTo(Dest& dest) const {
-      internal::outer_product_selector_run(*this, dest, set(), IsRowMajor<Dest>());
+      internal::outer_product_selector_run(*this, dest, set(), is_row_major<Dest>());
     }
     
     template<typename Dest>
     inline void addTo(Dest& dest) const {
-      internal::outer_product_selector_run(*this, dest, add(), IsRowMajor<Dest>());
+      internal::outer_product_selector_run(*this, dest, add(), is_row_major<Dest>());
     }
 
     template<typename Dest>
     inline void subTo(Dest& dest) const {
-      internal::outer_product_selector_run(*this, dest, sub(), IsRowMajor<Dest>());
+      internal::outer_product_selector_run(*this, dest, sub(), is_row_major<Dest>());
     }
 
     template<typename Dest> void scaleAndAddTo(Dest& dest, const Scalar& alpha) const
     {
-      internal::outer_product_selector_run(*this, dest, adds(alpha), IsRowMajor<Dest>());
+      internal::outer_product_selector_run(*this, dest, adds(alpha), is_row_major<Dest>());
     }
 };
 
@@ -425,15 +420,18 @@ template<> struct gemv_selector<OnTheRight,ColMajor,true>
     ResScalar actualAlpha = alpha * LhsBlasTraits::extractScalarFactor(prod.lhs())
                                   * RhsBlasTraits::extractScalarFactor(prod.rhs());
 
+    // make sure Dest is a compile-time vector type (bug 1166)
+    typedef typename conditional<Dest::IsVectorAtCompileTime, Dest, typename Dest::ColXpr>::type ActualDest;
+
     enum {
       // FIXME find a way to allow an inner stride on the result if packet_traits<Scalar>::size==1
       // on, the other hand it is good for the cache to pack the vector anyways...
-      EvalToDestAtCompileTime = Dest::InnerStrideAtCompileTime==1,
+      EvalToDestAtCompileTime = (ActualDest::InnerStrideAtCompileTime==1),
       ComplexByReal = (NumTraits<LhsScalar>::IsComplex) && (!NumTraits<RhsScalar>::IsComplex),
-      MightCannotUseDest = (Dest::InnerStrideAtCompileTime!=1) || ComplexByReal
+      MightCannotUseDest = (ActualDest::InnerStrideAtCompileTime!=1) || ComplexByReal
     };
 
-    gemv_static_vector_if<ResScalar,Dest::SizeAtCompileTime,Dest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest;
+    gemv_static_vector_if<ResScalar,ActualDest::SizeAtCompileTime,ActualDest::MaxSizeAtCompileTime,MightCannotUseDest> static_dest;
 
     bool alphaIsCompatible = (!ComplexByReal) || (numext::imag(actualAlpha)==RealScalar(0));
     bool evalToDest = EvalToDestAtCompileTime && alphaIsCompatible;
@@ -522,7 +520,7 @@ template<> struct gemv_selector<OnTheRight,RowMajor,true>
         actualLhs.rows(), actualLhs.cols(),
         actualLhs.data(), actualLhs.outerStride(),
         actualRhsPtr, 1,
-        dest.data(), dest.innerStride(),
+        dest.data(), dest.col(0).innerStride(), //NOTE  if dest is not a vector at compile-time, then dest.innerStride() might be wrong. (bug 1166)
         actualAlpha);
   }
 };

Eigen/src/Core/MapBase.h

@@ -123,7 +123,7 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
       return internal::ploadt<PacketScalar, LoadMode>(m_data + index * innerStride());
     }
 
-    inline MapBase(PointerType dataPtr) : m_data(dataPtr), m_rows(RowsAtCompileTime), m_cols(ColsAtCompileTime)
+    explicit inline MapBase(PointerType dataPtr) : m_data(dataPtr), m_rows(RowsAtCompileTime), m_cols(ColsAtCompileTime)
     {
       EIGEN_STATIC_ASSERT_FIXED_SIZE(Derived)
       checkSanity();
@@ -149,6 +149,10 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
       checkSanity();
     }
 
+    #ifdef EIGEN_MAPBASE_PLUGIN
+    #include EIGEN_MAPBASE_PLUGIN
+    #endif
+
   protected:
 
     void checkSanity() const
@@ -157,7 +161,7 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
                                         internal::inner_stride_at_compile_time<Derived>::ret==1),
                           PACKET_ACCESS_REQUIRES_TO_HAVE_INNER_STRIDE_FIXED_TO_1);
       eigen_assert(EIGEN_IMPLIES(internal::traits<Derived>::Flags&AlignedBit, (size_t(m_data) % 16) == 0)
-                   && "data is not aligned");
+                   && "input pointer is not aligned on a 16 byte boundary");
     }
 
     PointerType m_data;
@@ -168,6 +172,7 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
 template<typename Derived> class MapBase<Derived, WriteAccessors>
   : public MapBase<Derived, ReadOnlyAccessors>
 {
+    typedef MapBase<Derived, ReadOnlyAccessors> ReadOnlyMapBase;
   public:
 
     typedef MapBase<Derived, ReadOnlyAccessors> Base;
@@ -230,15 +235,13 @@ template<typename Derived> class MapBase<Derived, WriteAccessors>
 
     Derived& operator=(const MapBase& other)
     {
-      Base::Base::operator=(other);
+      ReadOnlyMapBase::Base::operator=(other);
       return derived();
     }
 
-    // In theory MapBase<Derived, ReadOnlyAccessors> should not make a using Base::operator=,
+    // In theory we could simply refer to Base:Base::operator=, but MSVC does not like Base::Base,
-    // and thus we should directly do: using Base::Base::operator=;
+    // see bugs 821 and 920.
-    // However, this would confuse recent MSVC 2013 (bug 821), and since MapBase<Derived, ReadOnlyAccessors>
+    using ReadOnlyMapBase::Base::operator=;
-    // has operator= to make ICC 11 happy, we can also make MSVC 2013 happy as follow:
-    using Base::operator=;
 };
 
 #undef EIGEN_STATIC_ASSERT_INDEX_BASED_ACCESS

Eigen/src/Core/MathFunctions.h

@@ -218,8 +218,8 @@ struct conj_retval
 * Implementation of abs2                                                 *
 ****************************************************************************/
 
-template<typename Scalar>
+template<typename Scalar,bool IsComplex>
-struct abs2_impl
+struct abs2_impl_default
 {
   typedef typename NumTraits<Scalar>::Real RealScalar;
   static inline RealScalar run(const Scalar& x)
@@ -228,15 +228,26 @@ struct abs2_impl
   }
 };
 
-template<typename RealScalar>
+template<typename Scalar>
-struct abs2_impl<std::complex<RealScalar> >
+struct abs2_impl_default<Scalar, true> // IsComplex
 {
-  static inline RealScalar run(const std::complex<RealScalar>& x)
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  static inline RealScalar run(const Scalar& x)
   {
     return real(x)*real(x) + imag(x)*imag(x);
   }
 };
 
+template<typename Scalar>
+struct abs2_impl
+{
+  typedef typename NumTraits<Scalar>::Real RealScalar;
+  static inline RealScalar run(const Scalar& x)
+  {
+    return abs2_impl_default<Scalar,NumTraits<Scalar>::IsComplex>::run(x);
+  }
+};
+
 template<typename Scalar>
 struct abs2_retval
 {
@@ -294,7 +305,7 @@ struct hypot_impl
     RealScalar _x = abs(x);
     RealScalar _y = abs(y);
     RealScalar p = (max)(_x, _y);
-    if(p==RealScalar(0)) return 0;
+    if(p==RealScalar(0)) return RealScalar(0);
     RealScalar q = (min)(_x, _y);
     RealScalar qp = q/p;
     return p * sqrt(RealScalar(1) + qp*qp);
@@ -707,21 +718,21 @@ struct scalar_fuzzy_impl : scalar_fuzzy_default_impl<Scalar, NumTraits<Scalar>::
 
 template<typename Scalar, typename OtherScalar>
 inline bool isMuchSmallerThan(const Scalar& x, const OtherScalar& y,
-                                   typename NumTraits<Scalar>::Real precision = NumTraits<Scalar>::dummy_precision())
+                              const typename NumTraits<Scalar>::Real &precision = NumTraits<Scalar>::dummy_precision())
 {
   return scalar_fuzzy_impl<Scalar>::template isMuchSmallerThan<OtherScalar>(x, y, precision);
 }
 
 template<typename Scalar>
 inline bool isApprox(const Scalar& x, const Scalar& y,
-                          typename NumTraits<Scalar>::Real precision = NumTraits<Scalar>::dummy_precision())
+                     const typename NumTraits<Scalar>::Real &precision = NumTraits<Scalar>::dummy_precision())
 {
   return scalar_fuzzy_impl<Scalar>::isApprox(x, y, precision);
 }
 
 template<typename Scalar>
 inline bool isApproxOrLessThan(const Scalar& x, const Scalar& y,
-                                    typename NumTraits<Scalar>::Real precision = NumTraits<Scalar>::dummy_precision())
+                               const typename NumTraits<Scalar>::Real &precision = NumTraits<Scalar>::dummy_precision())
 {
   return scalar_fuzzy_impl<Scalar>::isApproxOrLessThan(x, y, precision);
 }

Eigen/src/Core/Matrix.h

@@ -211,6 +211,21 @@ class Matrix
       : Base(internal::constructor_without_unaligned_array_assert())
     { Base::_check_template_params(); EIGEN_INITIALIZE_COEFFS_IF_THAT_OPTION_IS_ENABLED }
 
+#ifdef EIGEN_HAVE_RVALUE_REFERENCES
+    Matrix(Matrix&& other)
+      : Base(std::move(other))
+    {
+      Base::_check_template_params();
+      if (RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic)
+        Base::_set_noalias(other);
+    }
+    Matrix& operator=(Matrix&& other)
+    {
+      other.swap(*this);
+      return *this;
+    }
+#endif
+
     /** \brief Constructs a vector or row-vector with given dimension. \only_for_vectors
       *
       * Note that this is only useful for dynamic-size vectors. For fixed-size vectors,

Eigen/src/Core/MatrixBase.h

@@ -159,13 +159,11 @@ template<typename Derived> class MatrixBase
     template<typename OtherDerived>
     Derived& operator=(const ReturnByValue<OtherDerived>& other);
 
-#ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename ProductDerived, typename Lhs, typename Rhs>
     Derived& lazyAssign(const ProductBase<ProductDerived, Lhs,Rhs>& other);
 
     template<typename MatrixPower, typename Lhs, typename Rhs>
     Derived& lazyAssign(const MatrixPowerProduct<MatrixPower, Lhs,Rhs>& other);
-#endif // not EIGEN_PARSED_BY_DOXYGEN
 
     template<typename OtherDerived>
     Derived& operator+=(const MatrixBase<OtherDerived>& other);
@@ -442,6 +440,15 @@ template<typename Derived> class MatrixBase
     template<typename OtherScalar>
     void applyOnTheRight(Index p, Index q, const JacobiRotation<OtherScalar>& j);
 
+///////// SparseCore module /////////
+
+    template<typename OtherDerived>
+    EIGEN_STRONG_INLINE const typename SparseMatrixBase<OtherDerived>::template CwiseProductDenseReturnType<Derived>::Type
+    cwiseProduct(const SparseMatrixBase<OtherDerived> &other) const
+    {
+      return other.cwiseProduct(derived());
+    }
+
 ///////// MatrixFunctions module /////////
 
     typedef typename internal::stem_function<Scalar>::type StemFunction;

Eigen/src/Core/PermutationMatrix.h

@@ -250,6 +250,35 @@ class PermutationBase : public EigenBase<Derived>
     template<typename Other> friend
     inline PlainPermutationType operator*(const Transpose<PermutationBase<Other> >& other, const PermutationBase& perm)
     { return PlainPermutationType(internal::PermPermProduct, other.eval(), perm); }
+    
+    /** \returns the determinant of the permutation matrix, which is either 1 or -1 depending on the parity of the permutation.
+      *
+      * This function is O(\c n) procedure allocating a buffer of \c n booleans.
+      */
+    Index determinant() const
+    {
+      Index res = 1;
+      Index n = size();
+      Matrix<bool,RowsAtCompileTime,1,0,MaxRowsAtCompileTime> mask(n);
+      mask.fill(false);
+      Index r = 0;
+      while(r < n)
+      {
+        // search for the next seed
+        while(r<n && mask[r]) r++;
+        if(r>=n)
+          break;
+        // we got one, let's follow it until we are back to the seed
+        Index k0 = r++;
+        mask.coeffRef(k0) = true;
+        for(Index k=indices().coeff(k0); k!=k0; k=indices().coeff(k))
+        {
+          mask.coeffRef(k) = true;
+          res = -res;
+        }
+      }
+      return res;
+    }
 
   protected:
 
@@ -555,10 +584,11 @@ struct permut_matrix_product_retval
       const Index n = Side==OnTheLeft ? rows() : cols();
       // FIXME we need an is_same for expression that is not sensitive to constness. For instance
       // is_same_xpr<Block<const Matrix>, Block<Matrix> >::value should be true.
+      const typename Dest::Scalar *dst_data = internal::extract_data(dst);
       if(    is_same<MatrixTypeNestedCleaned,Dest>::value
           && blas_traits<MatrixTypeNestedCleaned>::HasUsableDirectAccess
           && blas_traits<Dest>::HasUsableDirectAccess
-          && extract_data(dst) == extract_data(m_matrix))
+          && dst_data!=0 && dst_data == extract_data(m_matrix))
       {
         // apply the permutation inplace
         Matrix<bool,PermutationType::RowsAtCompileTime,1,0,PermutationType::MaxRowsAtCompileTime> mask(m_permutation.size());

Eigen/src/Core/PlainObjectBase.h

@@ -315,8 +315,8 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     EIGEN_STRONG_INLINE void resizeLike(const EigenBase<OtherDerived>& _other)
     {
       const OtherDerived& other = _other.derived();
-      internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(other.rows(), other.cols());
+      internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(Index(other.rows()), Index(other.cols()));
-      const Index othersize = other.rows()*other.cols();
+      const Index othersize = Index(other.rows())*Index(other.cols());
       if(RowsAtCompileTime == 1)
       {
         eigen_assert(other.rows() == 1 || other.cols() == 1);
@@ -437,6 +437,36 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     }
 #endif
 
+#ifdef EIGEN_HAVE_RVALUE_REFERENCES
+    PlainObjectBase(PlainObjectBase&& other)
+      : m_storage( std::move(other.m_storage) )
+    {
+    }
+
+    PlainObjectBase& operator=(PlainObjectBase&& other)
+    {
+      using std::swap;
+      swap(m_storage, other.m_storage);
+      return *this;
+    }
+#endif
+
+    /** Copy constructor */
+    EIGEN_STRONG_INLINE PlainObjectBase(const PlainObjectBase& other)
+      : m_storage()
+    {
+      _check_template_params();
+      lazyAssign(other);
+    }
+
+    template<typename OtherDerived>
+    EIGEN_STRONG_INLINE PlainObjectBase(const DenseBase<OtherDerived> &other)
+      : m_storage()
+    {
+      _check_template_params();
+      lazyAssign(other);
+    }
+
     EIGEN_STRONG_INLINE PlainObjectBase(Index a_size, Index nbRows, Index nbCols)
       : m_storage(a_size, nbRows, nbCols)
     {
@@ -457,7 +487,7 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     /** \sa MatrixBase::operator=(const EigenBase<OtherDerived>&) */
     template<typename OtherDerived>
     EIGEN_STRONG_INLINE PlainObjectBase(const EigenBase<OtherDerived> &other)
-      : m_storage(other.derived().rows() * other.derived().cols(), other.derived().rows(), other.derived().cols())
+      : m_storage(Index(other.derived().rows()) * Index(other.derived().cols()), other.derived().rows(), other.derived().cols())
     {
       _check_template_params();
       internal::check_rows_cols_for_overflow<MaxSizeAtCompileTime>::run(other.derived().rows(), other.derived().cols());
@@ -573,6 +603,8 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
                  : (rows() == other.rows() && cols() == other.cols())))
         && "Size mismatch. Automatic resizing is disabled because EIGEN_NO_AUTOMATIC_RESIZING is defined");
       EIGEN_ONLY_USED_FOR_DEBUG(other);
+      if(this->size()==0)
+        resizeLike(other);
       #else
       resizeLike(other);
       #endif

Eigen/src/Core/Redux.h

@@ -247,8 +247,9 @@ struct redux_impl<Func, Derived, LinearVectorizedTraversal, NoUnrolling>
   }
 };
 
-template<typename Func, typename Derived>
+// NOTE: for SliceVectorizedTraversal we simply bypass unrolling
-struct redux_impl<Func, Derived, SliceVectorizedTraversal, NoUnrolling>
+template<typename Func, typename Derived, int Unrolling>
+struct redux_impl<Func, Derived, SliceVectorizedTraversal, Unrolling>
 {
   typedef typename Derived::Scalar Scalar;
   typedef typename packet_traits<Scalar>::type PacketScalar;

Eigen/src/Core/Ref.h

@@ -108,7 +108,8 @@ struct traits<Ref<_PlainObjectType, _Options, _StrideType> >
       OuterStrideMatch = Derived::IsVectorAtCompileTime
                       || int(StrideType::OuterStrideAtCompileTime)==int(Dynamic) || int(StrideType::OuterStrideAtCompileTime)==int(Derived::OuterStrideAtCompileTime),
       AlignmentMatch = (_Options!=Aligned) || ((PlainObjectType::Flags&AlignedBit)==0) || ((traits<Derived>::Flags&AlignedBit)==AlignedBit),
-      MatchAtCompileTime = HasDirectAccess && StorageOrderMatch && InnerStrideMatch && OuterStrideMatch && AlignmentMatch
+      ScalarTypeMatch = internal::is_same<typename PlainObjectType::Scalar, typename Derived::Scalar>::value,
+      MatchAtCompileTime = HasDirectAccess && StorageOrderMatch && InnerStrideMatch && OuterStrideMatch && AlignmentMatch && ScalarTypeMatch
     };
     typedef typename internal::conditional<MatchAtCompileTime,internal::true_type,internal::false_type>::type type;
   };
@@ -187,9 +188,11 @@ protected:
 template<typename PlainObjectType, int Options, typename StrideType> class Ref
   : public RefBase<Ref<PlainObjectType, Options, StrideType> >
 {
+  private:
     typedef internal::traits<Ref> Traits;
     template<typename Derived>
-    inline Ref(const PlainObjectBase<Derived>& expr);
+    inline Ref(const PlainObjectBase<Derived>& expr,
+               typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0);
   public:
 
     typedef RefBase<Ref> Base;
@@ -198,13 +201,15 @@ template<typename PlainObjectType, int Options, typename StrideType> class Ref
 
     #ifndef EIGEN_PARSED_BY_DOXYGEN
     template<typename Derived>
-    inline Ref(PlainObjectBase<Derived>& expr)
+    inline Ref(PlainObjectBase<Derived>& expr,
+               typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0)
     {
       EIGEN_STATIC_ASSERT(static_cast<bool>(Traits::template match<Derived>::MatchAtCompileTime), STORAGE_LAYOUT_DOES_NOT_MATCH);
       Base::construct(expr.derived());
     }
     template<typename Derived>
-    inline Ref(const DenseBase<Derived>& expr)
+    inline Ref(const DenseBase<Derived>& expr,
+               typename internal::enable_if<bool(Traits::template match<Derived>::MatchAtCompileTime),Derived>::type* = 0)
     #else
     template<typename Derived>
     inline Ref(DenseBase<Derived>& expr)
@@ -231,13 +236,23 @@ template<typename TPlainObjectType, int Options, typename StrideType> class Ref<
     EIGEN_DENSE_PUBLIC_INTERFACE(Ref)
 
     template<typename Derived>
-    inline Ref(const DenseBase<Derived>& expr)
+    inline Ref(const DenseBase<Derived>& expr,
+               typename internal::enable_if<bool(Traits::template match<Derived>::ScalarTypeMatch),Derived>::type* = 0)
     {
 //      std::cout << match_helper<Derived>::HasDirectAccess << "," << match_helper<Derived>::OuterStrideMatch << "," << match_helper<Derived>::InnerStrideMatch << "\n";
 //      std::cout << int(StrideType::OuterStrideAtCompileTime) << " - " << int(Derived::OuterStrideAtCompileTime) << "\n";
 //      std::cout << int(StrideType::InnerStrideAtCompileTime) << " - " << int(Derived::InnerStrideAtCompileTime) << "\n";
       construct(expr.derived(), typename Traits::template match<Derived>::type());
     }
+    
+    inline Ref(const Ref& other) : Base(other) {
+      // copy constructor shall not copy the m_object, to avoid unnecessary malloc and copy
+    }
+
+    template<typename OtherRef>
+    inline Ref(const RefBase<OtherRef>& other) {
+      construct(other.derived(), typename Traits::template match<OtherRef>::type());
+    }
 
   protected:
 

Eigen/src/Core/ReturnByValue.h

@@ -72,6 +72,8 @@ template<typename Derived> class ReturnByValue
     const Unusable& coeff(Index,Index) const { return *reinterpret_cast<const Unusable*>(this); }
     Unusable& coeffRef(Index) { return *reinterpret_cast<Unusable*>(this); }
     Unusable& coeffRef(Index,Index) { return *reinterpret_cast<Unusable*>(this); }
+    template<int LoadMode>  Unusable& packet(Index) const;
+    template<int LoadMode>  Unusable& packet(Index, Index) const;
 #endif
 };
 
@@ -83,6 +85,15 @@ Derived& DenseBase<Derived>::operator=(const ReturnByValue<OtherDerived>& other)
   return derived();
 }
 
+template<typename Derived>
+template<typename OtherDerived>
+Derived& DenseBase<Derived>::lazyAssign(const ReturnByValue<OtherDerived>& other)
+{
+  other.evalTo(derived());
+  return derived();
+}
+
+
 } // end namespace Eigen
 
 #endif // EIGEN_RETURNBYVALUE_H

Eigen/src/Core/SelfCwiseBinaryOp.h

@@ -180,15 +180,9 @@ inline Derived& DenseBase<Derived>::operator*=(const Scalar& other)
 template<typename Derived>
 inline Derived& DenseBase<Derived>::operator/=(const Scalar& other)
 {
-  typedef typename internal::conditional<NumTraits<Scalar>::IsInteger,
-                                        internal::scalar_quotient_op<Scalar>,
-                                        internal::scalar_product_op<Scalar> >::type BinOp;
   typedef typename Derived::PlainObject PlainObject;
-  SelfCwiseBinaryOp<BinOp, Derived, typename PlainObject::ConstantReturnType> tmp(derived());
+  SelfCwiseBinaryOp<internal::scalar_quotient_op<Scalar>, Derived, typename PlainObject::ConstantReturnType> tmp(derived());
-  Scalar actual_other;
+  tmp = PlainObject::Constant(rows(),cols(), other);
-  if(NumTraits<Scalar>::IsInteger)  actual_other = other;
-  else                              actual_other = Scalar(1)/other;
-  tmp = PlainObject::Constant(rows(),cols(), actual_other);
   return derived();
 }
 

Eigen/src/Core/SolveTriangular.h

@@ -116,17 +116,17 @@ template<typename Lhs, typename Rhs, int Mode, int Index, int Size>
 struct triangular_solver_unroller<Lhs,Rhs,Mode,Index,Size,false> {
   enum {
     IsLower = ((Mode&Lower)==Lower),
-    I = IsLower ? Index : Size - Index - 1,
+    RowIndex = IsLower ? Index : Size - Index - 1,
-    S = IsLower ? 0     : I+1
+    S = IsLower ? 0     : RowIndex+1
   };
   static void run(const Lhs& lhs, Rhs& rhs)
   {
     if (Index>0)
-      rhs.coeffRef(I) -= lhs.row(I).template segment<Index>(S).transpose()
+      rhs.coeffRef(RowIndex) -= lhs.row(RowIndex).template segment<Index>(S).transpose()
                          .cwiseProduct(rhs.template segment<Index>(S)).sum();
 
     if(!(Mode & UnitDiag))
-      rhs.coeffRef(I) /= lhs.coeff(I,I);
+      rhs.coeffRef(RowIndex) /= lhs.coeff(RowIndex,RowIndex);
 
     triangular_solver_unroller<Lhs,Rhs,Mode,Index+1,Size>::run(lhs,rhs);
   }
@@ -243,7 +243,8 @@ template<int Side, typename TriangularType, typename Rhs> struct triangular_solv
 
   template<typename Dest> inline void evalTo(Dest& dst) const
   {
-    if(!(is_same<RhsNestedCleaned,Dest>::value && extract_data(dst) == extract_data(m_rhs)))
+    const typename Dest::Scalar *dst_data = internal::extract_data(dst);
+    if(!(is_same<RhsNestedCleaned,Dest>::value && dst_data!=0 && extract_data(dst) == extract_data(m_rhs)))
       dst = m_rhs;
     m_triangularMatrix.template solveInPlace<Side>(dst);
   }

Eigen/src/Core/Transpose.h

@@ -331,11 +331,11 @@ inline void MatrixBase<Derived>::adjointInPlace()
 
 namespace internal {
 
-template<typename BinOp,typename NestedXpr,typename Rhs>
+template<typename BinOp,typename Xpr,typename Rhs>
-struct blas_traits<SelfCwiseBinaryOp<BinOp,NestedXpr,Rhs> >
+struct blas_traits<SelfCwiseBinaryOp<BinOp,Xpr,Rhs> >
- : blas_traits<NestedXpr>
+ : blas_traits<typename internal::remove_all<typename Xpr::Nested>::type>
 {
-  typedef SelfCwiseBinaryOp<BinOp,NestedXpr,Rhs> XprType;
+  typedef SelfCwiseBinaryOp<BinOp,Xpr,Rhs> XprType;
   static inline const XprType extract(const XprType& x) { return x; }
 };
 
@@ -392,7 +392,6 @@ struct checkTransposeAliasing_impl
                       ::run(extract_data(dst), other))
           && "aliasing detected during transposition, use transposeInPlace() "
              "or evaluate the rhs into a temporary using .eval()");
-
     }
 };
 

Eigen/src/Core/Transpositions.h

@@ -376,7 +376,8 @@ struct transposition_matrix_product_retval
       const int size = m_transpositions.size();
       Index j = 0;
 
-      if(!(is_same<MatrixTypeNestedCleaned,Dest>::value && extract_data(dst) == extract_data(m_matrix)))
+      const typename Dest::Scalar *dst_data = internal::extract_data(dst);
+      if(!(is_same<MatrixTypeNestedCleaned,Dest>::value && dst_data!=0 && dst_data == extract_data(m_matrix)))
         dst = m_matrix;
 
       for(int k=(Transposed?size-1:0) ; Transposed?k>=0:k<size ; Transposed?--k:++k)

Eigen/src/Core/Visitor.h

@@ -76,14 +76,17 @@ template<typename Derived>
 template<typename Visitor>
 void DenseBase<Derived>::visit(Visitor& visitor) const
 {
+  typedef typename internal::remove_all<typename Derived::Nested>::type ThisNested;
+  typename Derived::Nested thisNested(derived());
+
   enum { unroll = SizeAtCompileTime != Dynamic
                    && CoeffReadCost != Dynamic
                    && (SizeAtCompileTime == 1 || internal::functor_traits<Visitor>::Cost != Dynamic)
                    && SizeAtCompileTime * CoeffReadCost + (SizeAtCompileTime-1) * internal::functor_traits<Visitor>::Cost
                       <= EIGEN_UNROLLING_LIMIT };
-  return internal::visitor_impl<Visitor, Derived,
+  return internal::visitor_impl<Visitor, ThisNested,
       unroll ? int(SizeAtCompileTime) : Dynamic
-    >::run(derived(), visitor);
+    >::run(thisNested, visitor);
 }
 
 namespace internal {

Eigen/src/Core/arch/NEON/Complex.h

@@ -110,7 +110,7 @@ template<> EIGEN_STRONG_INLINE Packet2cf ploaddup<Packet2cf>(const std::complex<
 template<> EIGEN_STRONG_INLINE void pstore <std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_ALIGNED_STORE pstore((float*)to, from.v); }
 template<> EIGEN_STRONG_INLINE void pstoreu<std::complex<float> >(std::complex<float> *   to, const Packet2cf& from) { EIGEN_DEBUG_UNALIGNED_STORE pstoreu((float*)to, from.v); }
 
-template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *   addr) { __pld((float *)addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<std::complex<float> >(const std::complex<float> *   addr) { EIGEN_ARM_PREFETCH((float *)addr); }
 
 template<> EIGEN_STRONG_INLINE std::complex<float>  pfirst<Packet2cf>(const Packet2cf& a)
 {

Eigen/src/Core/arch/NEON/PacketMath.h

@@ -218,8 +218,8 @@ template<> EIGEN_STRONG_INLINE void pstore<int>(int*       to, const Packet4i& f
 template<> EIGEN_STRONG_INLINE void pstoreu<float>(float*  to, const Packet4f& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_f32(to, from); }
 template<> EIGEN_STRONG_INLINE void pstoreu<int>(int*      to, const Packet4i& from) { EIGEN_DEBUG_UNALIGNED_STORE vst1q_s32(to, from); }
 
-template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { __pld(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<float>(const float* addr) { EIGEN_ARM_PREFETCH(addr); }
-template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*     addr) { __pld(addr); }
+template<> EIGEN_STRONG_INLINE void prefetch<int>(const int*     addr) { EIGEN_ARM_PREFETCH(addr); }
 
 // FIXME only store the 2 first elements ?
 template<> EIGEN_STRONG_INLINE float  pfirst<Packet4f>(const Packet4f& a) { float EIGEN_ALIGN16 x[4]; vst1q_f32(x, a); return x[0]; }
@@ -384,6 +384,7 @@ template<> EIGEN_STRONG_INLINE int predux_max<Packet4i>(const Packet4i& a)
   a_lo = vget_low_s32(a);
   a_hi = vget_high_s32(a);
   max = vpmax_s32(a_lo, a_hi);
+  max = vpmax_s32(max, max);
 
   return vget_lane_s32(max, 0);
 }

Eigen/src/Core/arch/SSE/MathFunctions.h

@@ -126,7 +126,7 @@ Packet4f pexp<Packet4f>(const Packet4f& _x)
   _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p4, 1.6666665459E-1f);
   _EIGEN_DECLARE_CONST_Packet4f(cephes_exp_p5, 5.0000001201E-1f);
 
-  Packet4f tmp = _mm_setzero_ps(), fx;
+  Packet4f tmp, fx;
   Packet4i emm0;
 
   // clamp x
@@ -195,7 +195,7 @@ Packet2d pexp<Packet2d>(const Packet2d& _x)
   _EIGEN_DECLARE_CONST_Packet2d(cephes_exp_C2, 1.42860682030941723212e-6);
   static const __m128i p4i_1023_0 = _mm_setr_epi32(1023, 1023, 0, 0);
 
-  Packet2d tmp = _mm_setzero_pd(), fx;
+  Packet2d tmp, fx;
   Packet4i emm0;
 
   // clamp x
@@ -279,7 +279,7 @@ Packet4f psin<Packet4f>(const Packet4f& _x)
   _EIGEN_DECLARE_CONST_Packet4f(coscof_p2,  4.166664568298827E-002f);
   _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4 / M_PI
 
-  Packet4f xmm1, xmm2 = _mm_setzero_ps(), xmm3, sign_bit, y;
+  Packet4f xmm1, xmm2, xmm3, sign_bit, y;
 
   Packet4i emm0, emm2;
   sign_bit = x;
@@ -378,7 +378,7 @@ Packet4f pcos<Packet4f>(const Packet4f& _x)
   _EIGEN_DECLARE_CONST_Packet4f(coscof_p2,  4.166664568298827E-002f);
   _EIGEN_DECLARE_CONST_Packet4f(cephes_FOPI, 1.27323954473516f); // 4 / M_PI
 
-  Packet4f xmm1, xmm2 = _mm_setzero_ps(), xmm3, y;
+  Packet4f xmm1, xmm2, xmm3, y;
   Packet4i emm0, emm2;
 
   x = pabs(x);

Eigen/src/Core/arch/SSE/PacketMath.h

@@ -235,63 +235,27 @@ template<> EIGEN_STRONG_INLINE Packet4i pload<Packet4i>(const int*     from) { E
     return _mm_loadu_ps(from);
     #endif
   }
-  template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_pd(from); }
-  template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int*    from) { EIGEN_DEBUG_UNALIGNED_LOAD return _mm_loadu_si128(reinterpret_cast<const Packet4i*>(from)); }
 #else
-// Fast unaligned loads. Note that here we cannot directly use intrinsics: this would
-// require pointer casting to incompatible pointer types and leads to invalid code
-// because of the strict aliasing rule. The "dummy" stuff are required to enforce
-// a correct instruction dependency.
-// TODO: do the same for MSVC (ICC is compatible)
 // NOTE: with the code below, MSVC's compiler crashes!
 
-#if defined(__GNUC__) && defined(__i386__)
-  // bug 195: gcc/i386 emits weird x87 fldl/fstpl instructions for _mm_load_sd
-  #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 1
-#elif defined(__clang__)
-  // bug 201: Segfaults in __mm_loadh_pd with clang 2.8
-  #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 1
-#else
-  #define EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS 0
-#endif
-
 template<> EIGEN_STRONG_INLINE Packet4f ploadu<Packet4f>(const float* from)
 {
   EIGEN_DEBUG_UNALIGNED_LOAD
-#if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS
   return _mm_loadu_ps(from);
-#else
-  __m128d res;
-  res =  _mm_load_sd((const double*)(from)) ;
-  res =  _mm_loadh_pd(res, (const double*)(from+2)) ;
-  return _mm_castpd_ps(res);
-#endif
 }
+#endif
+
 template<> EIGEN_STRONG_INLINE Packet2d ploadu<Packet2d>(const double* from)
 {
   EIGEN_DEBUG_UNALIGNED_LOAD
-#if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS
   return _mm_loadu_pd(from);
-#else
-  __m128d res;
-  res = _mm_load_sd(from) ;
-  res = _mm_loadh_pd(res,from+1);
-  return res;
-#endif
 }
 template<> EIGEN_STRONG_INLINE Packet4i ploadu<Packet4i>(const int* from)
 {
   EIGEN_DEBUG_UNALIGNED_LOAD
-#if EIGEN_AVOID_CUSTOM_UNALIGNED_LOADS
+  return _mm_loadu_si128(reinterpret_cast<const __m128i*>(from));
-  return _mm_loadu_si128(reinterpret_cast<const Packet4i*>(from));
-#else
-  __m128d res;
-  res =  _mm_load_sd((const double*)(from)) ;
-  res =  _mm_loadh_pd(res, (const double*)(from+2)) ;
-  return _mm_castpd_si128(res);
-#endif
 }
-#endif
+
 
 template<> EIGEN_STRONG_INLINE Packet4f ploaddup<Packet4f>(const float*   from)
 {

Eigen/src/Core/products/CoeffBasedProduct.h

@@ -134,7 +134,7 @@ class CoeffBasedProduct
     };
 
     typedef internal::product_coeff_impl<CanVectorizeInner ? InnerVectorizedTraversal : DefaultTraversal,
-                                   Unroll ? (InnerSize==0 ? 0 : InnerSize-1) : Dynamic,
+                                   Unroll ? InnerSize : Dynamic,
                                    _LhsNested, _RhsNested, Scalar> ScalarCoeffImpl;
 
     typedef CoeffBasedProduct<LhsNested,RhsNested,NestByRefBit> LazyCoeffBasedProductType;
@@ -185,7 +185,7 @@ class CoeffBasedProduct
     {
       PacketScalar res;
       internal::product_packet_impl<Flags&RowMajorBit ? RowMajor : ColMajor,
-                              Unroll ? (InnerSize==0 ? 0 : InnerSize-1) : Dynamic,
+                              Unroll ? InnerSize : Dynamic,
                               _LhsNested, _RhsNested, PacketScalar, LoadMode>
         ::run(row, col, m_lhs, m_rhs, res);
       return res;
@@ -243,7 +243,17 @@ struct product_coeff_impl<DefaultTraversal, UnrollingIndex, Lhs, Rhs, RetScalar>
   static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res)
   {
     product_coeff_impl<DefaultTraversal, UnrollingIndex-1, Lhs, Rhs, RetScalar>::run(row, col, lhs, rhs, res);
-    res += lhs.coeff(row, UnrollingIndex) * rhs.coeff(UnrollingIndex, col);
+    res += lhs.coeff(row, UnrollingIndex-1) * rhs.coeff(UnrollingIndex-1, col);
+  }
+};
+
+template<typename Lhs, typename Rhs, typename RetScalar>
+struct product_coeff_impl<DefaultTraversal, 1, Lhs, Rhs, RetScalar>
+{
+  typedef typename Lhs::Index Index;
+  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res)
+  {
+    res = lhs.coeff(row, 0) * rhs.coeff(0, col);
   }
 };
 
@@ -251,9 +261,9 @@ template<typename Lhs, typename Rhs, typename RetScalar>
 struct product_coeff_impl<DefaultTraversal, 0, Lhs, Rhs, RetScalar>
 {
   typedef typename Lhs::Index Index;
-  static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res)
+  static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, RetScalar &res)
   {
-    res = lhs.coeff(row, 0) * rhs.coeff(0, col);
+    res = RetScalar(0);
   }
 };
 
@@ -293,6 +303,16 @@ struct product_coeff_vectorized_unroller<0, Lhs, Rhs, Packet>
   }
 };
 
+template<typename Lhs, typename Rhs, typename RetScalar>
+struct product_coeff_impl<InnerVectorizedTraversal, 0, Lhs, Rhs, RetScalar>
+{
+  typedef typename Lhs::Index Index;
+  static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, RetScalar &res)
+  {
+    res = 0;
+  }
+};
+
 template<int UnrollingIndex, typename Lhs, typename Rhs, typename RetScalar>
 struct product_coeff_impl<InnerVectorizedTraversal, UnrollingIndex, Lhs, Rhs, RetScalar>
 {
@@ -302,8 +322,7 @@ struct product_coeff_impl<InnerVectorizedTraversal, UnrollingIndex, Lhs, Rhs, Re
   static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, RetScalar &res)
   {
     Packet pres;
-    product_coeff_vectorized_unroller<UnrollingIndex+1-PacketSize, Lhs, Rhs, Packet>::run(row, col, lhs, rhs, pres);
+    product_coeff_vectorized_unroller<UnrollingIndex-PacketSize, Lhs, Rhs, Packet>::run(row, col, lhs, rhs, pres);
-    product_coeff_impl<DefaultTraversal,UnrollingIndex,Lhs,Rhs,RetScalar>::run(row, col, lhs, rhs, res);
     res = predux(pres);
   }
 };
@@ -371,7 +390,7 @@ struct product_packet_impl<RowMajor, UnrollingIndex, Lhs, Rhs, Packet, LoadMode>
   static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res)
   {
     product_packet_impl<RowMajor, UnrollingIndex-1, Lhs, Rhs, Packet, LoadMode>::run(row, col, lhs, rhs, res);
-    res =  pmadd(pset1<Packet>(lhs.coeff(row, UnrollingIndex)), rhs.template packet<LoadMode>(UnrollingIndex, col), res);
+    res =  pmadd(pset1<Packet>(lhs.coeff(row, UnrollingIndex-1)), rhs.template packet<LoadMode>(UnrollingIndex-1, col), res);
   }
 };
 
@@ -382,12 +401,12 @@ struct product_packet_impl<ColMajor, UnrollingIndex, Lhs, Rhs, Packet, LoadMode>
   static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res)
   {
     product_packet_impl<ColMajor, UnrollingIndex-1, Lhs, Rhs, Packet, LoadMode>::run(row, col, lhs, rhs, res);
-    res =  pmadd(lhs.template packet<LoadMode>(row, UnrollingIndex), pset1<Packet>(rhs.coeff(UnrollingIndex, col)), res);
+    res =  pmadd(lhs.template packet<LoadMode>(row, UnrollingIndex-1), pset1<Packet>(rhs.coeff(UnrollingIndex-1, col)), res);
   }
 };
 
 template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
-struct product_packet_impl<RowMajor, 0, Lhs, Rhs, Packet, LoadMode>
+struct product_packet_impl<RowMajor, 1, Lhs, Rhs, Packet, LoadMode>
 {
   typedef typename Lhs::Index Index;
   static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res)
@@ -397,7 +416,7 @@ struct product_packet_impl<RowMajor, 0, Lhs, Rhs, Packet, LoadMode>
 };
 
 template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
-struct product_packet_impl<ColMajor, 0, Lhs, Rhs, Packet, LoadMode>
+struct product_packet_impl<ColMajor, 1, Lhs, Rhs, Packet, LoadMode>
 {
   typedef typename Lhs::Index Index;
   static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet &res)
@@ -406,16 +425,35 @@ struct product_packet_impl<ColMajor, 0, Lhs, Rhs, Packet, LoadMode>
   }
 };
 
+template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct product_packet_impl<RowMajor, 0, Lhs, Rhs, Packet, LoadMode>
+{
+  typedef typename Lhs::Index Index;
+  static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, Packet &res)
+  {
+    res = pset1<Packet>(0);
+  }
+};
+
+template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
+struct product_packet_impl<ColMajor, 0, Lhs, Rhs, Packet, LoadMode>
+{
+  typedef typename Lhs::Index Index;
+  static EIGEN_STRONG_INLINE void run(Index /*row*/, Index /*col*/, const Lhs& /*lhs*/, const Rhs& /*rhs*/, Packet &res)
+  {
+    res = pset1<Packet>(0);
+  }
+};
+
 template<typename Lhs, typename Rhs, typename Packet, int LoadMode>
 struct product_packet_impl<RowMajor, Dynamic, Lhs, Rhs, Packet, LoadMode>
 {
   typedef typename Lhs::Index Index;
   static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet& res)
   {
-    eigen_assert(lhs.cols()>0 && "you are using a non initialized matrix");
+    res = pset1<Packet>(0);
-    res = pmul(pset1<Packet>(lhs.coeff(row, 0)),rhs.template packet<LoadMode>(0, col));
+    for(Index i = 0; i < lhs.cols(); ++i)
-      for(Index i = 1; i < lhs.cols(); ++i)
+      res =  pmadd(pset1<Packet>(lhs.coeff(row, i)), rhs.template packet<LoadMode>(i, col), res);
-        res =  pmadd(pset1<Packet>(lhs.coeff(row, i)), rhs.template packet<LoadMode>(i, col), res);
   }
 };
 
@@ -425,10 +463,9 @@ struct product_packet_impl<ColMajor, Dynamic, Lhs, Rhs, Packet, LoadMode>
   typedef typename Lhs::Index Index;
   static EIGEN_STRONG_INLINE void run(Index row, Index col, const Lhs& lhs, const Rhs& rhs, Packet& res)
   {
-    eigen_assert(lhs.cols()>0 && "you are using a non initialized matrix");
+    res = pset1<Packet>(0);
-    res = pmul(lhs.template packet<LoadMode>(row, 0), pset1<Packet>(rhs.coeff(0, col)));
+    for(Index i = 0; i < lhs.cols(); ++i)
-      for(Index i = 1; i < lhs.cols(); ++i)
+      res =  pmadd(lhs.template packet<LoadMode>(row, i), pset1<Packet>(rhs.coeff(i, col)), res);
-        res =  pmadd(lhs.template packet<LoadMode>(row, i), pset1<Packet>(rhs.coeff(i, col)), res);
   }
 };
 

Eigen/src/Core/products/GeneralMatrixMatrix.h

@@ -140,8 +140,10 @@ static void run(Index rows, Index cols, Index depth,
       // Release all the sub blocks B'_j of B' for the current thread,
       // i.e., we simply decrement the number of users by 1
       for(Index j=0; j<threads; ++j)
+      {
         #pragma omp atomic
-        --(info[j].users);
+        info[j].users -= 1;
+      }
     }
   }
   else
@@ -390,13 +392,17 @@ class GeneralProduct<Lhs, Rhs, GemmProduct>
 
     GeneralProduct(const Lhs& lhs, const Rhs& rhs) : Base(lhs,rhs)
     {
+#if !(defined(EIGEN_NO_STATIC_ASSERT) && defined(EIGEN_NO_DEBUG))
       typedef internal::scalar_product_op<LhsScalar,RhsScalar> BinOp;
       EIGEN_CHECK_BINARY_COMPATIBILIY(BinOp,LhsScalar,RhsScalar);
+#endif
     }
 
     template<typename Dest> void scaleAndAddTo(Dest& dst, const Scalar& alpha) const
     {
       eigen_assert(dst.rows()==m_lhs.rows() && dst.cols()==m_rhs.cols());
+      if(m_lhs.cols()==0 || m_lhs.rows()==0 || m_rhs.cols()==0)
+        return;
 
       typename internal::add_const_on_value_type<ActualLhsType>::type lhs = LhsBlasTraits::extract(m_lhs);
       typename internal::add_const_on_value_type<ActualRhsType>::type rhs = RhsBlasTraits::extract(m_rhs);

Eigen/src/Core/products/Parallelizer.h

@@ -125,19 +125,22 @@ void parallelize_gemm(const Functor& func, Index rows, Index cols, bool transpos
   if(transpose)
     std::swap(rows,cols);
 
-  Index blockCols = (cols / threads) & ~Index(0x3);
-  Index blockRows = (rows / threads) & ~Index(0x7);
-  
   GemmParallelInfo<Index>* info = new GemmParallelInfo<Index>[threads];
 
-  #pragma omp parallel for schedule(static,1) num_threads(threads)
+  #pragma omp parallel num_threads(threads)
-  for(Index i=0; i<threads; ++i)
   {
+    Index i = omp_get_thread_num();
+    // Note that the actual number of threads might be lower than the number of request ones.
+    Index actual_threads = omp_get_num_threads();
+    
+    Index blockCols = (cols / actual_threads) & ~Index(0x3);
+    Index blockRows = (rows / actual_threads) & ~Index(0x7);
+    
     Index r0 = i*blockRows;
-    Index actualBlockRows = (i+1==threads) ? rows-r0 : blockRows;
+    Index actualBlockRows = (i+1==actual_threads) ? rows-r0 : blockRows;
 
     Index c0 = i*blockCols;
-    Index actualBlockCols = (i+1==threads) ? cols-c0 : blockCols;
+    Index actualBlockCols = (i+1==actual_threads) ? cols-c0 : blockCols;
 
     info[i].rhs_start = c0;
     info[i].rhs_length = actualBlockCols;

Eigen/src/Core/products/TriangularMatrixMatrix_MKL.h

@@ -109,7 +109,7 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,true, \
 /* Non-square case - doesn't fit to MKL ?TRMM. Fall to default triangular product or call MKL ?GEMM*/ \
    if (rows != depth) { \
 \
-     int nthr = mkl_domain_get_max_threads(MKL_BLAS); \
+     int nthr = mkl_domain_get_max_threads(EIGEN_MKL_DOMAIN_BLAS); \
 \
      if (((nthr==1) && (((std::max)(rows,depth)-diagSize)/(double)diagSize < 0.5))) { \
      /* Most likely no benefit to call TRMM or GEMM from MKL*/ \
@@ -223,7 +223,7 @@ struct product_triangular_matrix_matrix_trmm<EIGTYPE,Index,Mode,false, \
 /* Non-square case - doesn't fit to MKL ?TRMM. Fall to default triangular product or call MKL ?GEMM*/ \
    if (cols != depth) { \
 \
-     int nthr = mkl_domain_get_max_threads(MKL_BLAS); \
+     int nthr = mkl_domain_get_max_threads(EIGEN_MKL_DOMAIN_BLAS); \
 \
      if ((nthr==1) && (((std::max)(cols,depth)-diagSize)/(double)diagSize < 0.5)) { \
      /* Most likely no benefit to call TRMM or GEMM from MKL*/ \

Eigen/src/Core/products/TriangularSolverMatrix.h

@@ -81,7 +81,7 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conju
     // coherence when accessing the rhs elements
     std::ptrdiff_t l1, l2;
     manage_caching_sizes(GetAction, &l1, &l2);
-    Index subcols = cols>0 ? l2/(4 * sizeof(Scalar) * otherStride) : 0;
+    Index subcols = cols>0 ? l2/(4 * sizeof(Scalar) *  std::max<Index>(otherStride,size)) : 0;
     subcols = std::max<Index>((subcols/Traits::nr)*Traits::nr, Traits::nr);
 
     for(Index k2=IsLower ? 0 : size;
@@ -115,8 +115,9 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conju
           {
             // TODO write a small kernel handling this (can be shared with trsv)
             Index i  = IsLower ? k2+k1+k : k2-k1-k-1;
-            Index s  = IsLower ? k2+k1 : i+1;
             Index rs = actualPanelWidth - k - 1; // remaining size
+            Index s  = TriStorageOrder==RowMajor ? (IsLower ? k2+k1 : i+1)
+                                                 :  IsLower ? i+1 : i-rs;
 
             Scalar a = (Mode & UnitDiag) ? Scalar(1) : Scalar(1)/conj(tri(i,i));
             for (Index j=j2; j<j2+actual_cols; ++j)
@@ -133,7 +134,6 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheLeft,Mode,Conju
               }
               else
               {
-                Index s = IsLower ? i+1 : i-rs;
                 Scalar b = (other(i,j) *= a);
                 Scalar* r = &other(s,j);
                 const Scalar* l = &tri(s,i);
@@ -302,9 +302,12 @@ EIGEN_DONT_INLINE void triangular_solve_matrix<Scalar,Index,OnTheRight,Mode,Conj
                 for (Index i=0; i<actual_mc; ++i)
                   r[i] -= a[i] * b;
               }
-              Scalar b = (Mode & UnitDiag) ? Scalar(1) : Scalar(1)/conj(rhs(j,j));
+              if((Mode & UnitDiag)==0)
-              for (Index i=0; i<actual_mc; ++i)
+              {
-                r[i] *= b;
+                Scalar b = conj(rhs(j,j));
+                for (Index i=0; i<actual_mc; ++i)
+                  r[i] /= b;
+              }
             }
 
             // pack the just computed part of lhs to A

Eigen/src/Core/util/BlasUtil.h

@@ -171,12 +171,13 @@ template<typename XprType> struct blas_traits
 };
 
 // pop conjugate
-template<typename Scalar, typename NestedXpr>
+template<typename Scalar, typename Xpr>
-struct blas_traits<CwiseUnaryOp<scalar_conjugate_op<Scalar>, NestedXpr> >
+struct blas_traits<CwiseUnaryOp<scalar_conjugate_op<Scalar>, Xpr> >
- : blas_traits<NestedXpr>
+ : blas_traits<typename internal::remove_all<typename Xpr::Nested>::type>
 {
+  typedef typename internal::remove_all<typename Xpr::Nested>::type NestedXpr;
   typedef blas_traits<NestedXpr> Base;
-  typedef CwiseUnaryOp<scalar_conjugate_op<Scalar>, NestedXpr> XprType;
+  typedef CwiseUnaryOp<scalar_conjugate_op<Scalar>, Xpr> XprType;
   typedef typename Base::ExtractType ExtractType;
 
   enum {
@@ -188,12 +189,13 @@ struct blas_traits<CwiseUnaryOp<scalar_conjugate_op<Scalar>, NestedXpr> >
 };
 
 // pop scalar multiple
-template<typename Scalar, typename NestedXpr>
+template<typename Scalar, typename Xpr>
-struct blas_traits<CwiseUnaryOp<scalar_multiple_op<Scalar>, NestedXpr> >
+struct blas_traits<CwiseUnaryOp<scalar_multiple_op<Scalar>, Xpr> >
- : blas_traits<NestedXpr>
+ : blas_traits<typename internal::remove_all<typename Xpr::Nested>::type>
 {
+  typedef typename internal::remove_all<typename Xpr::Nested>::type NestedXpr;
   typedef blas_traits<NestedXpr> Base;
-  typedef CwiseUnaryOp<scalar_multiple_op<Scalar>, NestedXpr> XprType;
+  typedef CwiseUnaryOp<scalar_multiple_op<Scalar>, Xpr> XprType;
   typedef typename Base::ExtractType ExtractType;
   static inline ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); }
   static inline Scalar extractScalarFactor(const XprType& x)
@@ -201,12 +203,13 @@ struct blas_traits<CwiseUnaryOp<scalar_multiple_op<Scalar>, NestedXpr> >
 };
 
 // pop opposite
-template<typename Scalar, typename NestedXpr>
+template<typename Scalar, typename Xpr>
-struct blas_traits<CwiseUnaryOp<scalar_opposite_op<Scalar>, NestedXpr> >
+struct blas_traits<CwiseUnaryOp<scalar_opposite_op<Scalar>, Xpr> >
- : blas_traits<NestedXpr>
+ : blas_traits<typename internal::remove_all<typename Xpr::Nested>::type>
 {
+  typedef typename internal::remove_all<typename Xpr::Nested>::type NestedXpr;
   typedef blas_traits<NestedXpr> Base;
-  typedef CwiseUnaryOp<scalar_opposite_op<Scalar>, NestedXpr> XprType;
+  typedef CwiseUnaryOp<scalar_opposite_op<Scalar>, Xpr> XprType;
   typedef typename Base::ExtractType ExtractType;
   static inline ExtractType extract(const XprType& x) { return Base::extract(x.nestedExpression()); }
   static inline Scalar extractScalarFactor(const XprType& x)
@@ -214,13 +217,14 @@ struct blas_traits<CwiseUnaryOp<scalar_opposite_op<Scalar>, NestedXpr> >
 };
 
 // pop/push transpose
-template<typename NestedXpr>
+template<typename Xpr>
-struct blas_traits<Transpose<NestedXpr> >
+struct blas_traits<Transpose<Xpr> >
- : blas_traits<NestedXpr>
+ : blas_traits<typename internal::remove_all<typename Xpr::Nested>::type>
 {
+  typedef typename internal::remove_all<typename Xpr::Nested>::type NestedXpr;
   typedef typename NestedXpr::Scalar Scalar;
   typedef blas_traits<NestedXpr> Base;
-  typedef Transpose<NestedXpr> XprType;
+  typedef Transpose<Xpr> XprType;
   typedef Transpose<const typename Base::_ExtractType>  ExtractType; // const to get rid of a compile error; anyway blas traits are only used on the RHS
   typedef Transpose<const typename Base::_ExtractType> _ExtractType;
   typedef typename conditional<bool(Base::HasUsableDirectAccess),

Eigen/src/Core/util/Constants.h

@@ -162,7 +162,7 @@ const unsigned int HereditaryBits = RowMajorBit
 /** \ingroup enums
   * Enum containing possible values for the \p Mode parameter of 
   * MatrixBase::selfadjointView() and MatrixBase::triangularView(). */
-enum {
+enum UpLoType {
   /** View matrix as a lower triangular matrix. */
   Lower=0x1,                      
   /** View matrix as an upper triangular matrix. */
@@ -187,7 +187,7 @@ enum {
 
 /** \ingroup enums
   * Enum for indicating whether an object is aligned or not. */
-enum { 
+enum AlignmentType {
   /** Object is not correctly aligned for vectorization. */
   Unaligned=0, 
   /** Object is aligned for vectorization. */
@@ -217,7 +217,7 @@ enum DirectionType {
 
 /** \internal \ingroup enums
   * Enum to specify how to traverse the entries of a matrix. */
-enum {
+enum TraversalType {
   /** \internal Default traversal, no vectorization, no index-based access */
   DefaultTraversal,
   /** \internal No vectorization, use index-based access to have only one for loop instead of 2 nested loops */
@@ -239,7 +239,7 @@ enum {
 
 /** \internal \ingroup enums
   * Enum to specify whether to unroll loops when traversing over the entries of a matrix. */
-enum {
+enum UnrollingType {
   /** \internal Do not unroll loops. */
   NoUnrolling,
   /** \internal Unroll only the inner loop, but not the outer loop. */
@@ -251,7 +251,7 @@ enum {
 
 /** \internal \ingroup enums
   * Enum to specify whether to use the default (built-in) implementation or the specialization. */
-enum {
+enum SpecializedType {
   Specialized,
   BuiltIn
 };
@@ -259,7 +259,7 @@ enum {
 /** \ingroup enums
   * Enum containing possible values for the \p _Options template parameter of
   * Matrix, Array and BandMatrix. */
-enum {
+enum StorageOptions {
   /** Storage order is column major (see \ref TopicStorageOrders). */
   ColMajor = 0,
   /** Storage order is row major (see \ref TopicStorageOrders). */
@@ -272,7 +272,7 @@ enum {
 
 /** \ingroup enums
   * Enum for specifying whether to apply or solve on the left or right. */
-enum {
+enum SideType {
   /** Apply transformation on the left. */
   OnTheLeft = 1,  
   /** Apply transformation on the right. */
@@ -418,7 +418,7 @@ namespace Architecture
 
 /** \internal \ingroup enums
   * Enum used as template parameter in GeneralProduct. */
-enum { CoeffBasedProductMode, LazyCoeffBasedProductMode, OuterProduct, InnerProduct, GemvProduct, GemmProduct };
+enum ProductImplType { CoeffBasedProductMode, LazyCoeffBasedProductMode, OuterProduct, InnerProduct, GemvProduct, GemmProduct };
 
 /** \internal \ingroup enums
   * Enum used in experimental parallel implementation. */
@@ -433,6 +433,19 @@ struct MatrixXpr {};
 /** The type used to identify an array expression */
 struct ArrayXpr {};
 
+namespace internal {
+  /** \internal
+  * Constants for comparison functors
+  */
+  enum ComparisonName {
+    cmp_EQ = 0,
+    cmp_LT = 1,
+    cmp_LE = 2,
+    cmp_UNORD = 3,
+    cmp_NEQ = 4
+  };
+}
+
 } // end namespace Eigen
 
 #endif // EIGEN_CONSTANTS_H

Eigen/src/Core/util/DisableStupidWarnings.h

@@ -35,6 +35,14 @@
     #pragma clang diagnostic push
   #endif
   #pragma clang diagnostic ignored "-Wconstant-logical-operand"
+
+#elif defined __GNUC__ && __GNUC__>=6
+
+  #ifndef EIGEN_PERMANENTLY_DISABLE_STUPID_WARNINGS
+    #pragma GCC diagnostic push
+  #endif
+  #pragma GCC diagnostic ignored "-Wignored-attributes"
+
 #endif
 
 #endif // not EIGEN_WARNINGS_DISABLED

Eigen/src/Core/util/ForwardDeclarations.h

@@ -235,6 +235,9 @@ template<typename Scalar> class Rotation2D;
 template<typename Scalar> class AngleAxis;
 template<typename Scalar,int Dim> class Translation;
 
+// Sparse module:
+template<typename Derived> class SparseMatrixBase;
+
 #ifdef EIGEN2_SUPPORT
 template<typename Derived, int _Dim> class eigen2_RotationBase;
 template<typename Lhs, typename Rhs> class eigen2_Cross;

Eigen/src/Core/util/MKL_support.h

@@ -76,6 +76,38 @@
 #include <mkl_lapacke.h>
 #define EIGEN_MKL_VML_THRESHOLD 128
 
+/* MKL_DOMAIN_BLAS, etc are defined only in 10.3 update 7 */
+/* MKL_BLAS, etc are not defined in 11.2 */
+#ifdef MKL_DOMAIN_ALL
+#define EIGEN_MKL_DOMAIN_ALL MKL_DOMAIN_ALL
+#else
+#define EIGEN_MKL_DOMAIN_ALL MKL_ALL
+#endif
+
+#ifdef MKL_DOMAIN_BLAS
+#define EIGEN_MKL_DOMAIN_BLAS MKL_DOMAIN_BLAS
+#else
+#define EIGEN_MKL_DOMAIN_BLAS MKL_BLAS
+#endif
+
+#ifdef MKL_DOMAIN_FFT
+#define EIGEN_MKL_DOMAIN_FFT MKL_DOMAIN_FFT
+#else
+#define EIGEN_MKL_DOMAIN_FFT MKL_FFT
+#endif
+
+#ifdef MKL_DOMAIN_VML
+#define EIGEN_MKL_DOMAIN_VML MKL_DOMAIN_VML
+#else
+#define EIGEN_MKL_DOMAIN_VML MKL_VML
+#endif
+
+#ifdef MKL_DOMAIN_PARDISO
+#define EIGEN_MKL_DOMAIN_PARDISO MKL_DOMAIN_PARDISO
+#else
+#define EIGEN_MKL_DOMAIN_PARDISO MKL_PARDISO
+#endif
+
 namespace Eigen {
 
 typedef std::complex<double> dcomplex;

Eigen/src/Core/util/Macros.h

@@ -13,23 +13,292 @@
 
 #define EIGEN_WORLD_VERSION 3
 #define EIGEN_MAJOR_VERSION 2
-#define EIGEN_MINOR_VERSION 3
+#define EIGEN_MINOR_VERSION 9
 
 #define EIGEN_VERSION_AT_LEAST(x,y,z) (EIGEN_WORLD_VERSION>x || (EIGEN_WORLD_VERSION>=x && \
                                       (EIGEN_MAJOR_VERSION>y || (EIGEN_MAJOR_VERSION>=y && \
                                                                  EIGEN_MINOR_VERSION>=z))))
+
+
+// Compiler identification, EIGEN_COMP_*
+
+/// \internal EIGEN_COMP_GNUC set to 1 for all compilers compatible with GCC
 #ifdef __GNUC__
+  #define EIGEN_COMP_GNUC 1
+#else
+  #define EIGEN_COMP_GNUC 0
+#endif
+
+/// \internal EIGEN_COMP_CLANG set to 1 if the compiler is clang (alias for __clang__)
+#if defined(__clang__)
+  #define EIGEN_COMP_CLANG 1
+#else
+  #define EIGEN_COMP_CLANG 0
+#endif
+
+
+/// \internal EIGEN_COMP_LLVM set to 1 if the compiler backend is llvm
+#if defined(__llvm__)
+  #define EIGEN_COMP_LLVM 1
+#else
+  #define EIGEN_COMP_LLVM 0
+#endif
+
+/// \internal EIGEN_COMP_ICC set to __INTEL_COMPILER if the compiler is Intel compiler, 0 otherwise
+#if defined(__INTEL_COMPILER)
+  #define EIGEN_COMP_ICC __INTEL_COMPILER
+#else
+  #define EIGEN_COMP_ICC 0
+#endif
+
+/// \internal EIGEN_COMP_MINGW set to 1 if the compiler is mingw
+#if defined(__MINGW32__)
+  #define EIGEN_COMP_MINGW 1
+#else
+  #define EIGEN_COMP_MINGW 0
+#endif
+
+/// \internal EIGEN_COMP_SUNCC set to 1 if the compiler is Solaris Studio
+#if defined(__SUNPRO_CC)
+  #define EIGEN_COMP_SUNCC 1
+#else
+  #define EIGEN_COMP_SUNCC 0
+#endif
+
+/// \internal EIGEN_COMP_MSVC set to _MSC_VER if the compiler is Microsoft Visual C++, 0 otherwise.
+#if defined(_MSC_VER)
+  #define EIGEN_COMP_MSVC _MSC_VER
+#else
+  #define EIGEN_COMP_MSVC 0
+#endif
+
+/// \internal EIGEN_COMP_MSVC_STRICT set to 1 if the compiler is really Microsoft Visual C++ and not ,e.g., ICC
+#if EIGEN_COMP_MSVC && !(EIGEN_COMP_ICC)
+  #define EIGEN_COMP_MSVC_STRICT _MSC_VER
+#else
+  #define EIGEN_COMP_MSVC_STRICT 0
+#endif
+
+/// \internal EIGEN_COMP_IBM set to 1 if the compiler is IBM XL C++
+#if defined(__IBMCPP__) || defined(__xlc__)
+  #define EIGEN_COMP_IBM 1
+#else
+  #define EIGEN_COMP_IBM 0
+#endif
+
+/// \internal EIGEN_COMP_PGI set to 1 if the compiler is Portland Group Compiler
+#if defined(__PGI)
+  #define EIGEN_COMP_PGI 1
+#else
+  #define EIGEN_COMP_PGI 0
+#endif
+
+/// \internal EIGEN_COMP_ARM set to 1 if the compiler is ARM Compiler
+#if defined(__CC_ARM) || defined(__ARMCC_VERSION)
+  #define EIGEN_COMP_ARM 1
+#else
+  #define EIGEN_COMP_ARM 0
+#endif
+
+
+/// \internal EIGEN_GNUC_STRICT set to 1 if the compiler is really GCC and not a compatible compiler (e.g., ICC, clang, mingw, etc.)
+#if EIGEN_COMP_GNUC && !(EIGEN_COMP_CLANG || EIGEN_COMP_ICC || EIGEN_COMP_MINGW || EIGEN_COMP_PGI || EIGEN_COMP_IBM || EIGEN_COMP_ARM )
+  #define EIGEN_COMP_GNUC_STRICT 1
+#else
+  #define EIGEN_COMP_GNUC_STRICT 0
+#endif
+
+
+#if EIGEN_COMP_GNUC
   #define EIGEN_GNUC_AT_LEAST(x,y) ((__GNUC__==x && __GNUC_MINOR__>=y) || __GNUC__>x)
+  #define EIGEN_GNUC_AT_MOST(x,y)  ((__GNUC__==x && __GNUC_MINOR__<=y) || __GNUC__<x)
+  #define EIGEN_GNUC_AT(x,y)       ( __GNUC__==x && __GNUC_MINOR__==y )
 #else
   #define EIGEN_GNUC_AT_LEAST(x,y) 0
+  #define EIGEN_GNUC_AT_MOST(x,y)  0
+  #define EIGEN_GNUC_AT(x,y)       0
 #endif
- 
+
-#ifdef __GNUC__
+// FIXME: could probably be removed as we do not support gcc 3.x anymore
-  #define EIGEN_GNUC_AT_MOST(x,y) ((__GNUC__==x && __GNUC_MINOR__<=y) || __GNUC__<x)
+#if EIGEN_COMP_GNUC && (__GNUC__ <= 3)
+#define EIGEN_GCC3_OR_OLDER 1
 #else
-  #define EIGEN_GNUC_AT_MOST(x,y) 0
+#define EIGEN_GCC3_OR_OLDER 0
 #endif
 
+
+// Architecture identification, EIGEN_ARCH_*
+
+#if defined(__x86_64__) || defined(_M_X64) || defined(__amd64)
+  #define EIGEN_ARCH_x86_64 1
+#else
+  #define EIGEN_ARCH_x86_64 0
+#endif
+
+#if defined(__i386__) || defined(_M_IX86) || defined(_X86_) || defined(__i386)
+  #define EIGEN_ARCH_i386 1
+#else
+  #define EIGEN_ARCH_i386 0
+#endif
+
+#if EIGEN_ARCH_x86_64 || EIGEN_ARCH_i386
+  #define EIGEN_ARCH_i386_OR_x86_64 1
+#else
+  #define EIGEN_ARCH_i386_OR_x86_64 0
+#endif
+
+/// \internal EIGEN_ARCH_ARM set to 1 if the architecture is ARM
+#if defined(__arm__)
+  #define EIGEN_ARCH_ARM 1
+#else
+  #define EIGEN_ARCH_ARM 0
+#endif
+
+/// \internal EIGEN_ARCH_ARM64 set to 1 if the architecture is ARM64
+#if defined(__aarch64__)
+  #define EIGEN_ARCH_ARM64 1
+#else
+  #define EIGEN_ARCH_ARM64 0
+#endif
+
+#if EIGEN_ARCH_ARM || EIGEN_ARCH_ARM64
+  #define EIGEN_ARCH_ARM_OR_ARM64 1
+#else
+  #define EIGEN_ARCH_ARM_OR_ARM64 0
+#endif
+
+/// \internal EIGEN_ARCH_MIPS set to 1 if the architecture is MIPS
+#if defined(__mips__) || defined(__mips)
+  #define EIGEN_ARCH_MIPS 1
+#else
+  #define EIGEN_ARCH_MIPS 0
+#endif
+
+/// \internal EIGEN_ARCH_SPARC set to 1 if the architecture is SPARC
+#if defined(__sparc__) || defined(__sparc)
+  #define EIGEN_ARCH_SPARC 1
+#else
+  #define EIGEN_ARCH_SPARC 0
+#endif
+
+/// \internal EIGEN_ARCH_IA64 set to 1 if the architecture is Intel Itanium
+#if defined(__ia64__)
+  #define EIGEN_ARCH_IA64 1
+#else
+  #define EIGEN_ARCH_IA64 0
+#endif
+
+/// \internal EIGEN_ARCH_PPC set to 1 if the architecture is PowerPC
+#if defined(__powerpc__) || defined(__ppc__) || defined(_M_PPC)
+  #define EIGEN_ARCH_PPC 1
+#else
+  #define EIGEN_ARCH_PPC 0
+#endif
+
+
+
+// Operating system identification, EIGEN_OS_*
+
+/// \internal EIGEN_OS_UNIX set to 1 if the OS is a unix variant
+#if defined(__unix__) || defined(__unix)
+  #define EIGEN_OS_UNIX 1
+#else
+  #define EIGEN_OS_UNIX 0
+#endif
+
+/// \internal EIGEN_OS_LINUX set to 1 if the OS is based on Linux kernel
+#if defined(__linux__)
+  #define EIGEN_OS_LINUX 1
+#else
+  #define EIGEN_OS_LINUX 0
+#endif
+
+/// \internal EIGEN_OS_ANDROID set to 1 if the OS is Android
+// note: ANDROID is defined when using ndk_build, __ANDROID__ is defined when using a standalone toolchain.
+#if defined(__ANDROID__) || defined(ANDROID)
+  #define EIGEN_OS_ANDROID 1
+#else
+  #define EIGEN_OS_ANDROID 0
+#endif
+
+/// \internal EIGEN_OS_GNULINUX set to 1 if the OS is GNU Linux and not Linux-based OS (e.g., not android)
+#if defined(__gnu_linux__) && !(EIGEN_OS_ANDROID)
+  #define EIGEN_OS_GNULINUX 1
+#else
+  #define EIGEN_OS_GNULINUX 0
+#endif
+
+/// \internal EIGEN_OS_BSD set to 1 if the OS is a BSD variant
+#if defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || defined(__bsdi__) || defined(__DragonFly__)
+  #define EIGEN_OS_BSD 1
+#else
+  #define EIGEN_OS_BSD 0
+#endif
+
+/// \internal EIGEN_OS_MAC set to 1 if the OS is MacOS
+#if defined(__APPLE__)
+  #define EIGEN_OS_MAC 1
+#else
+  #define EIGEN_OS_MAC 0
+#endif
+
+/// \internal EIGEN_OS_QNX set to 1 if the OS is QNX
+#if defined(__QNX__)
+  #define EIGEN_OS_QNX 1
+#else
+  #define EIGEN_OS_QNX 0
+#endif
+
+/// \internal EIGEN_OS_WIN set to 1 if the OS is Windows based
+#if defined(_WIN32)
+  #define EIGEN_OS_WIN 1
+#else
+  #define EIGEN_OS_WIN 0
+#endif
+
+/// \internal EIGEN_OS_WIN64 set to 1 if the OS is Windows 64bits
+#if defined(_WIN64)
+  #define EIGEN_OS_WIN64 1
+#else
+  #define EIGEN_OS_WIN64 0
+#endif
+
+/// \internal EIGEN_OS_WINCE set to 1 if the OS is Windows CE
+#if defined(_WIN32_WCE)
+  #define EIGEN_OS_WINCE 1
+#else
+  #define EIGEN_OS_WINCE 0
+#endif
+
+/// \internal EIGEN_OS_CYGWIN set to 1 if the OS is Windows/Cygwin
+#if defined(__CYGWIN__)
+  #define EIGEN_OS_CYGWIN 1
+#else
+  #define EIGEN_OS_CYGWIN 0
+#endif
+
+/// \internal EIGEN_OS_WIN_STRICT set to 1 if the OS is really Windows and not some variants
+#if EIGEN_OS_WIN && !( EIGEN_OS_WINCE || EIGEN_OS_CYGWIN )
+  #define EIGEN_OS_WIN_STRICT 1
+#else
+  #define EIGEN_OS_WIN_STRICT 0
+#endif
+
+/// \internal EIGEN_OS_SUN set to 1 if the OS is SUN
+#if (defined(sun) || defined(__sun)) && !(defined(__SVR4) || defined(__svr4__))
+  #define EIGEN_OS_SUN 1
+#else
+  #define EIGEN_OS_SUN 0
+#endif
+
+/// \internal EIGEN_OS_SOLARIS set to 1 if the OS is Solaris
+#if (defined(sun) || defined(__sun)) && (defined(__SVR4) || defined(__svr4__))
+  #define EIGEN_OS_SOLARIS 1
+#else
+  #define EIGEN_OS_SOLARIS 0
+#endif
+
+
 #if EIGEN_GNUC_AT_MOST(4,3) && !defined(__clang__)
   // see bug 89
   #define EIGEN_SAFE_TO_USE_STANDARD_ASSERT_MACRO 0
@@ -37,12 +306,6 @@
   #define EIGEN_SAFE_TO_USE_STANDARD_ASSERT_MACRO 1
 #endif
 
-#if defined(__GNUC__) && (__GNUC__ <= 3)
-#define EIGEN_GCC3_OR_OLDER 1
-#else
-#define EIGEN_GCC3_OR_OLDER 0
-#endif
-
 // 16 byte alignment is only useful for vectorization. Since it affects the ABI, we need to enable
 // 16 byte alignment on all platforms where vectorization might be enabled. In theory we could always
 // enable alignment, but it can be a cause of problems on some platforms, so we just disable it in
@@ -96,6 +359,20 @@
 #define EIGEN_DEFAULT_DENSE_INDEX_TYPE std::ptrdiff_t
 #endif
 
+// A Clang feature extension to determine compiler features.
+// We use it to determine 'cxx_rvalue_references'
+#ifndef __has_feature
+# define __has_feature(x) 0
+#endif
+
+// Do we support r-value references?
+#if (__has_feature(cxx_rvalue_references) || \
+     (defined(__cplusplus) && __cplusplus >= 201103L) || \
+     (defined(_MSC_VER) && _MSC_VER >= 1600))
+  #define EIGEN_HAVE_RVALUE_REFERENCES
+#endif
+
+
 // Cross compiler wrapper around LLVM's __has_builtin
 #ifdef __has_builtin
 #  define EIGEN_HAS_BUILTIN(x) __has_builtin(x)
@@ -278,6 +555,7 @@ namespace Eigen {
   #error Please tell me what is the equivalent of __attribute__((aligned(n))) for your compiler
 #endif
 
+#define EIGEN_ALIGN8  EIGEN_ALIGN_TO_BOUNDARY(8)
 #define EIGEN_ALIGN16 EIGEN_ALIGN_TO_BOUNDARY(16)
 
 #if EIGEN_ALIGN_STATICALLY
@@ -313,7 +591,7 @@ namespace Eigen {
 // just an empty macro !
 #define EIGEN_EMPTY
 
-#if defined(_MSC_VER) && (!defined(__INTEL_COMPILER))
+#if defined(_MSC_VER) && (_MSC_VER < 1900) && (!defined(__INTEL_COMPILER))
 #define EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived) \
   using Base::operator =;
 #elif defined(__clang__) // workaround clang bug (see http://forum.kde.org/viewtopic.php?f=74&t=102653)
@@ -332,8 +610,11 @@ namespace Eigen {
   }
 #endif
 
-#define EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Derived) \
+/** \internal
-  EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived)
+ * \brief Macro to manually inherit assignment operators.
+ * This is necessary, because the implicitly defined assignment operator gets deleted when a custom operator= is defined.
+ */
+#define EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Derived) EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Derived)
 
 /**
 * Just a side note. Commenting within defines works only by documenting
@@ -405,6 +686,8 @@ namespace Eigen {
 #define EIGEN_SIZE_MAX(a,b) (((int)a == Dynamic || (int)b == Dynamic) ? Dynamic \
                            : ((int)a >= (int)b) ? (int)a : (int)b)
 
+#define EIGEN_ADD_COST(a,b) int(a)==Dynamic || int(b)==Dynamic ? Dynamic : int(a)+int(b)
+
 #define EIGEN_LOGICAL_XOR(a,b) (((a) || (b)) && !((a) && (b)))
 
 #define EIGEN_IMPLIES(a,b) (!(a) || (b))

Eigen/src/Core/util/Memory.h

@@ -466,9 +466,8 @@ template<typename T, bool Align> inline void conditional_aligned_delete_auto(T *
 template<typename Scalar, typename Index>
 static inline Index first_aligned(const Scalar* array, Index size)
 {
-  enum { PacketSize = packet_traits<Scalar>::size,
+  static const Index PacketSize = packet_traits<Scalar>::size;
-         PacketAlignedMask = PacketSize-1
+  static const Index PacketAlignedMask = PacketSize-1;
-  };
 
   if(PacketSize==1)
   {
@@ -508,7 +507,12 @@ template<typename T> void smart_copy(const T* start, const T* end, T* target)
 
 template<typename T> struct smart_copy_helper<T,true> {
   static inline void run(const T* start, const T* end, T* target)
-  { memcpy(target, start, std::ptrdiff_t(end)-std::ptrdiff_t(start)); }
+  {
+    std::ptrdiff_t size = std::ptrdiff_t(end)-std::ptrdiff_t(start);
+    if(size==0) return;
+    eigen_internal_assert(start!=0 && end!=0 && target!=0);
+    memcpy(target, start, size);
+  }
 };
 
 template<typename T> struct smart_copy_helper<T,false> {
@@ -516,7 +520,6 @@ template<typename T> struct smart_copy_helper<T,false> {
   { std::copy(start, end, target); }
 };
 
-
 /*****************************************************************************
 *** Implementation of runtime stack allocation (falling back to malloc)    ***
 *****************************************************************************/
@@ -524,7 +527,7 @@ template<typename T> struct smart_copy_helper<T,false> {
 // you can overwrite Eigen's default behavior regarding alloca by defining EIGEN_ALLOCA
 // to the appropriate stack allocation function
 #ifndef EIGEN_ALLOCA
-  #if (defined __linux__)
+  #if (defined __linux__) || (defined __APPLE__) || (defined alloca)
     #define EIGEN_ALLOCA alloca
   #elif defined(_MSC_VER)
     #define EIGEN_ALLOCA _alloca
@@ -631,6 +634,8 @@ template<typename T> class aligned_stack_memory_handler
       } \
       void operator delete(void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
       void operator delete[](void * ptr) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
+      void operator delete(void * ptr, std::size_t /* sz */) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
+      void operator delete[](void * ptr, std::size_t /* sz */) throw() { Eigen::internal::conditional_aligned_free<NeedsToAlign>(ptr); } \
       /* in-place new and delete. since (at least afaik) there is no actual   */ \
       /* memory allocated we can safely let the default implementation handle */ \
       /* this particular case. */ \
@@ -654,99 +659,60 @@ template<typename T> class aligned_stack_memory_handler
 
 /****************************************************************************/
 
+
 /** \class aligned_allocator
-* \ingroup Core_Module
+  * \ingroup Core_Module
-*
+  *
-* \brief STL compatible allocator to use with with 16 byte aligned types
+  * \brief STL compatible allocator to use with with 16 byte aligned types
-*
+  *
-* Example:
+  * Example:
-* \code
+  * \code
-* // Matrix4f requires 16 bytes alignment:
+  * // Matrix4f requires 16 bytes alignment:
-* std::map< int, Matrix4f, std::less<int>, 
+  * std::map< int, Matrix4f, std::less<int>,
-*           aligned_allocator<std::pair<const int, Matrix4f> > > my_map_mat4;
+  *           aligned_allocator<std::pair<const int, Matrix4f> > > my_map_mat4;
-* // Vector3f does not require 16 bytes alignment, no need to use Eigen's allocator:
+  * // Vector3f does not require 16 bytes alignment, no need to use Eigen's allocator:
-* std::map< int, Vector3f > my_map_vec3;
+  * std::map< int, Vector3f > my_map_vec3;
-* \endcode
+  * \endcode
-*
+  *
-* \sa \ref TopicStlContainers.
+  * \sa \blank \ref TopicStlContainers.
-*/
+  */
 template<class T>
-class aligned_allocator
+class aligned_allocator : public std::allocator<T>
 {
 public:
-    typedef size_t    size_type;
+  typedef size_t          size_type;
-    typedef std::ptrdiff_t difference_type;
+  typedef std::ptrdiff_t  difference_type;
-    typedef T*        pointer;
+  typedef T*              pointer;
-    typedef const T*  const_pointer;
+  typedef const T*        const_pointer;
-    typedef T&        reference;
+  typedef T&              reference;
-    typedef const T&  const_reference;
+  typedef const T&        const_reference;
-    typedef T         value_type;
+  typedef T               value_type;
 
-    template<class U>
+  template<class U>
-    struct rebind
+  struct rebind
-    {
+  {
-        typedef aligned_allocator<U> other;
+    typedef aligned_allocator<U> other;
-    };
+  };
 
-    pointer address( reference value ) const
+  aligned_allocator() : std::allocator<T>() {}
-    {
-        return &value;
-    }
 
-    const_pointer address( const_reference value ) const
+  aligned_allocator(const aligned_allocator& other) : std::allocator<T>(other) {}
-    {
-        return &value;
-    }
 
-    aligned_allocator()
+  template<class U>
-    {
+  aligned_allocator(const aligned_allocator<U>& other) : std::allocator<T>(other) {}
-    }
 
-    aligned_allocator( const aligned_allocator& )
+  ~aligned_allocator() {}
-    {
-    }
 
-    template<class U>
+  pointer allocate(size_type num, const void* /*hint*/ = 0)
-    aligned_allocator( const aligned_allocator<U>& )
+  {
-    {
+    internal::check_size_for_overflow<T>(num);
-    }
+    return static_cast<pointer>( internal::aligned_malloc(num * sizeof(T)) );
+  }
 
-    ~aligned_allocator()
+  void deallocate(pointer p, size_type /*num*/)
-    {
+  {
-    }
+    internal::aligned_free(p);
-
+  }
-    size_type max_size() const
-    {
-        return (std::numeric_limits<size_type>::max)();
-    }
-
-    pointer allocate( size_type num, const void* hint = 0 )
-    {
-        EIGEN_UNUSED_VARIABLE(hint);
-        internal::check_size_for_overflow<T>(num);
-        return static_cast<pointer>( internal::aligned_malloc( num * sizeof(T) ) );
-    }
-
-    void construct( pointer p, const T& value )
-    {
-        ::new( p ) T( value );
-    }
-
-    void destroy( pointer p )
-    {
-        p->~T();
-    }
-
-    void deallocate( pointer p, size_type /*num*/ )
-    {
-        internal::aligned_free( p );
-    }
-
-    bool operator!=(const aligned_allocator<T>& ) const
-    { return false; }
-
-    bool operator==(const aligned_allocator<T>& ) const
-    { return true; }
 };
 
 //---------- Cache sizes ----------

Eigen/src/Core/util/ReenableStupidWarnings.h

@@ -8,7 +8,10 @@
     #pragma warning pop
   #elif defined __clang__
     #pragma clang diagnostic pop
+  #elif defined __GNUC__ && __GNUC__>=6
+    #pragma GCC diagnostic pop
   #endif
+
 #endif
 
 #endif // EIGEN_WARNINGS_DISABLED

Eigen/src/Core/util/StaticAssert.h

@@ -26,7 +26,7 @@
 
 #ifndef EIGEN_NO_STATIC_ASSERT
 
-  #if defined(__GXX_EXPERIMENTAL_CXX0X__) || (defined(_MSC_VER) && (_MSC_VER >= 1600))
+  #if __has_feature(cxx_static_assert) || (defined(__cplusplus) && __cplusplus >= 201103L) || (EIGEN_COMP_MSVC >= 1600)
 
     // if native static_assert is enabled, let's use it
     #define EIGEN_STATIC_ASSERT(X,MSG) static_assert(X,#MSG);

Eigen/src/Core/util/XprHelper.h

@@ -366,17 +366,17 @@ struct dense_xpr_base<Derived, ArrayXpr>
 
 /** \internal Helper base class to add a scalar multiple operator
   * overloads for complex types */
-template<typename Derived,typename Scalar,typename OtherScalar,
+template<typename Derived, typename Scalar, typename OtherScalar, typename BaseType,
          bool EnableIt = !is_same<Scalar,OtherScalar>::value >
-struct special_scalar_op_base : public DenseCoeffsBase<Derived>
+struct special_scalar_op_base : public BaseType
 {
   // dummy operator* so that the
   // "using special_scalar_op_base::operator*" compiles
   void operator*() const;
 };
 
-template<typename Derived,typename Scalar,typename OtherScalar>
+template<typename Derived,typename Scalar,typename OtherScalar, typename BaseType>
-struct special_scalar_op_base<Derived,Scalar,OtherScalar,true>  : public DenseCoeffsBase<Derived>
+struct special_scalar_op_base<Derived,Scalar,OtherScalar,BaseType,true>  : public BaseType
 {
   const CwiseUnaryOp<scalar_multiple2_op<Scalar,OtherScalar>, Derived>
   operator*(const OtherScalar& scalar) const

Eigen/src/Eigenvalues/ComplexEigenSolver.h

@@ -234,6 +234,12 @@ template<typename _MatrixType> class ComplexEigenSolver
     }
 
   protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+    
     EigenvectorType m_eivec;
     EigenvalueType m_eivalues;
     ComplexSchur<MatrixType> m_schur;
@@ -251,6 +257,8 @@ template<typename MatrixType>
 ComplexEigenSolver<MatrixType>& 
 ComplexEigenSolver<MatrixType>::compute(const MatrixType& matrix, bool computeEigenvectors)
 {
+  check_template_parameters();
+  
   // this code is inspired from Jampack
   eigen_assert(matrix.cols() == matrix.rows());
 

Eigen/src/Eigenvalues/ComplexSchur_MKL.h

@@ -45,7 +45,6 @@ ComplexSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >& \
 ComplexSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(const Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW>& matrix, bool computeU) \
 { \
   typedef Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> MatrixType; \
-  typedef MatrixType::Scalar Scalar; \
   typedef MatrixType::RealScalar RealScalar; \
   typedef std::complex<RealScalar> ComplexScalar; \
 \

Eigen/src/Eigenvalues/EigenSolver.h

@@ -298,6 +298,13 @@ template<typename _MatrixType> class EigenSolver
     void doComputeEigenvectors();
 
   protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+      EIGEN_STATIC_ASSERT(!NumTraits<Scalar>::IsComplex, NUMERIC_TYPE_MUST_BE_REAL);
+    }
+    
     MatrixType m_eivec;
     EigenvalueType m_eivalues;
     bool m_isInitialized;
@@ -364,6 +371,8 @@ template<typename MatrixType>
 EigenSolver<MatrixType>& 
 EigenSolver<MatrixType>::compute(const MatrixType& matrix, bool computeEigenvectors)
 {
+  check_template_parameters();
+  
   using std::sqrt;
   using std::abs;
   eigen_assert(matrix.cols() == matrix.rows());

Eigen/src/Eigenvalues/GeneralizedEigenSolver.h

@@ -263,6 +263,13 @@ template<typename _MatrixType> class GeneralizedEigenSolver
     }
 
   protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+      EIGEN_STATIC_ASSERT(!NumTraits<Scalar>::IsComplex, NUMERIC_TYPE_MUST_BE_REAL);
+    }
+    
     MatrixType m_eivec;
     ComplexVectorType m_alphas;
     VectorType m_betas;
@@ -290,6 +297,8 @@ template<typename MatrixType>
 GeneralizedEigenSolver<MatrixType>&
 GeneralizedEigenSolver<MatrixType>::compute(const MatrixType& A, const MatrixType& B, bool computeEigenvectors)
 {
+  check_template_parameters();
+  
   using std::sqrt;
   using std::abs;
   eigen_assert(A.cols() == A.rows() && B.cols() == A.rows() && B.cols() == B.rows());
@@ -318,13 +327,33 @@ GeneralizedEigenSolver<MatrixType>::compute(const MatrixType& A, const MatrixTyp
       }
       else
       {
-        Scalar p = Scalar(0.5) * (m_matS.coeff(i, i) - m_matS.coeff(i+1, i+1));
+        // We need to extract the generalized eigenvalues of the pair of a general 2x2 block S and a triangular 2x2 block T
-        Scalar z = sqrt(abs(p * p + m_matS.coeff(i+1, i) * m_matS.coeff(i, i+1)));
+        // From the eigen decomposition of T = U * E * U^-1,
-        m_alphas.coeffRef(i)   = ComplexScalar(m_matS.coeff(i+1, i+1) + p, z);
+        // we can extract the eigenvalues of (U^-1 * S * U) / E
-        m_alphas.coeffRef(i+1) = ComplexScalar(m_matS.coeff(i+1, i+1) + p, -z);
+        // Here, we can take advantage that E = diag(T), and U = [ 1 T_01 ; 0 T_11-T_00], and U^-1 = [1 -T_11/(T_11-T_00) ; 0 1/(T_11-T_00)].
+        // Then taking beta=T_00*T_11*(T_11-T_00), we can avoid any division, and alpha is the eigenvalues of A = (U^-1 * S * U) * diag(T_11,T_00) * (T_11-T_00):
+
+        // T =  [a b ; 0 c]
+        // S =  [e f ; g h]
+        RealScalar a = m_realQZ.matrixT().coeff(i, i), b = m_realQZ.matrixT().coeff(i, i+1), c = m_realQZ.matrixT().coeff(i+1, i+1);
+        RealScalar e = m_matS.coeff(i, i), f = m_matS.coeff(i, i+1), g = m_matS.coeff(i+1, i), h = m_matS.coeff(i+1, i+1);
+        RealScalar d = c-a;
+        RealScalar gb = g*b;
+        Matrix<RealScalar,2,2> A;
+        A <<  (e*d-gb)*c,  ((e*b+f*d-h*b)*d-gb*b)*a,
+              g*c       ,  (gb+h*d)*a;
+
+        // NOTE, we could also compute the SVD of T's block during the QZ factorization so that the respective T block is guaranteed to be diagonal,
+        // and then we could directly apply the formula below (while taking care of scaling S columns by T11,T00):
+
+        Scalar p = Scalar(0.5) * (A.coeff(i, i) - A.coeff(i+1, i+1));
+        Scalar z = sqrt(abs(p * p + A.coeff(i+1, i) * A.coeff(i, i+1)));
+        m_alphas.coeffRef(i)   = ComplexScalar(A.coeff(i+1, i+1) + p, z);
+        m_alphas.coeffRef(i+1) = ComplexScalar(A.coeff(i+1, i+1) + p, -z);
+
+        m_betas.coeffRef(i)   =
+        m_betas.coeffRef(i+1) = a*c*d;
 
-        m_betas.coeffRef(i)   = m_realQZ.matrixT().coeff(i,i);
-        m_betas.coeffRef(i+1) = m_realQZ.matrixT().coeff(i,i);
         i += 2;
       }
     }

Eigen/src/Eigenvalues/RealQZ.h

@@ -240,10 +240,10 @@ namespace Eigen {
             m_S.coeffRef(i,j) = Scalar(0.0);
             m_S.rightCols(dim-j-1).applyOnTheLeft(i-1,i,G.adjoint());
             m_T.rightCols(dim-i+1).applyOnTheLeft(i-1,i,G.adjoint());
+            // update Q
+            if (m_computeQZ)
+              m_Q.applyOnTheRight(i-1,i,G);
           }
-          // update Q
-          if (m_computeQZ)
-            m_Q.applyOnTheRight(i-1,i,G);
           // kill T(i,i-1)
           if(m_T.coeff(i,i-1)!=Scalar(0))
           {
@@ -251,10 +251,10 @@ namespace Eigen {
             m_T.coeffRef(i,i-1) = Scalar(0.0);
             m_S.applyOnTheRight(i,i-1,G);
             m_T.topRows(i).applyOnTheRight(i,i-1,G);
+            // update Z
+            if (m_computeQZ)
+              m_Z.applyOnTheLeft(i,i-1,G.adjoint());
           }
-          // update Z
-          if (m_computeQZ)
-            m_Z.applyOnTheLeft(i,i-1,G.adjoint());
         }
       }
     }
@@ -313,7 +313,7 @@ namespace Eigen {
       using std::abs;
       using std::sqrt;
       const Index dim=m_S.cols();
-      if (abs(m_S.coeff(i+1,i)==Scalar(0)))
+      if (abs(m_S.coeff(i+1,i))==Scalar(0))
         return;
       Index z = findSmallDiagEntry(i,i+1);
       if (z==i-1)

Eigen/src/Eigenvalues/RealSchur.h

@@ -234,7 +234,7 @@ template<typename _MatrixType> class RealSchur
     typedef Matrix<Scalar,3,1> Vector3s;
 
     Scalar computeNormOfT();
-    Index findSmallSubdiagEntry(Index iu, const Scalar& norm);
+    Index findSmallSubdiagEntry(Index iu);
     void splitOffTwoRows(Index iu, bool computeU, const Scalar& exshift);
     void computeShift(Index iu, Index iter, Scalar& exshift, Vector3s& shiftInfo);
     void initFrancisQRStep(Index il, Index iu, const Vector3s& shiftInfo, Index& im, Vector3s& firstHouseholderVector);
@@ -286,7 +286,7 @@ RealSchur<MatrixType>& RealSchur<MatrixType>::computeFromHessenberg(const HessMa
   {
     while (iu >= 0)
     {
-      Index il = findSmallSubdiagEntry(iu, norm);
+      Index il = findSmallSubdiagEntry(iu);
 
       // Check for convergence
       if (il == iu) // One root found
@@ -343,16 +343,14 @@ inline typename MatrixType::Scalar RealSchur<MatrixType>::computeNormOfT()
 
 /** \internal Look for single small sub-diagonal element and returns its index */
 template<typename MatrixType>
-inline typename MatrixType::Index RealSchur<MatrixType>::findSmallSubdiagEntry(Index iu, const Scalar& norm)
+inline typename MatrixType::Index RealSchur<MatrixType>::findSmallSubdiagEntry(Index iu)
 {
   using std::abs;
   Index res = iu;
   while (res > 0)
   {
     Scalar s = abs(m_matT.coeff(res-1,res-1)) + abs(m_matT.coeff(res,res));
-    if (s == 0.0)
+    if (abs(m_matT.coeff(res,res-1)) <= NumTraits<Scalar>::epsilon() * s)
-      s = norm;
-    if (abs(m_matT.coeff(res,res-1)) < NumTraits<Scalar>::epsilon() * s)
       break;
     res--;
   }
@@ -457,9 +455,7 @@ inline void RealSchur<MatrixType>::initFrancisQRStep(Index il, Index iu, const V
     const Scalar lhs = m_matT.coeff(im,im-1) * (abs(v.coeff(1)) + abs(v.coeff(2)));
     const Scalar rhs = v.coeff(0) * (abs(m_matT.coeff(im-1,im-1)) + abs(Tmm) + abs(m_matT.coeff(im+1,im+1)));
     if (abs(lhs) < NumTraits<Scalar>::epsilon() * rhs)
-    {
       break;
-    }
   }
 }
 

Eigen/src/Eigenvalues/RealSchur_MKL.h

@@ -44,10 +44,6 @@ template<> inline \
 RealSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >& \
 RealSchur<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> >::compute(const Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW>& matrix, bool computeU) \
 { \
-  typedef Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW> MatrixType; \
-  typedef MatrixType::Scalar Scalar; \
-  typedef MatrixType::RealScalar RealScalar; \
-\
   eigen_assert(matrix.cols() == matrix.rows()); \
 \
   lapack_int n = matrix.cols(), sdim, info; \

Eigen/src/Eigenvalues/SelfAdjointEigenSolver.h

@@ -80,6 +80,8 @@ template<typename _MatrixType> class SelfAdjointEigenSolver
     /** \brief Scalar type for matrices of type \p _MatrixType. */
     typedef typename MatrixType::Scalar Scalar;
     typedef typename MatrixType::Index Index;
+    
+    typedef Matrix<Scalar,Size,Size,ColMajor,MaxColsAtCompileTime,MaxColsAtCompileTime> EigenvectorsType;
 
     /** \brief Real scalar type for \p _MatrixType.
       *
@@ -225,7 +227,7 @@ template<typename _MatrixType> class SelfAdjointEigenSolver
       *
       * \sa eigenvalues()
       */
-    const MatrixType& eigenvectors() const
+    const EigenvectorsType& eigenvectors() const
     {
       eigen_assert(m_isInitialized && "SelfAdjointEigenSolver is not initialized.");
       eigen_assert(m_eigenvectorsOk && "The eigenvectors have not been computed together with the eigenvalues.");
@@ -351,7 +353,12 @@ template<typename _MatrixType> class SelfAdjointEigenSolver
     #endif // EIGEN2_SUPPORT
 
   protected:
-    MatrixType m_eivec;
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+    
+    EigenvectorsType m_eivec;
     RealVectorType m_eivalues;
     typename TridiagonalizationType::SubDiagonalType m_subdiag;
     ComputationInfo m_info;
@@ -376,7 +383,7 @@ template<typename _MatrixType> class SelfAdjointEigenSolver
   * "implicit symmetric QR step with Wilkinson shift"
   */
 namespace internal {
-template<int StorageOrder,typename RealScalar, typename Scalar, typename Index>
+template<typename RealScalar, typename Scalar, typename Index>
 static void tridiagonal_qr_step(RealScalar* diag, RealScalar* subdiag, Index start, Index end, Scalar* matrixQ, Index n);
 }
 
@@ -384,6 +391,8 @@ template<typename MatrixType>
 SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType>
 ::compute(const MatrixType& matrix, int options)
 {
+  check_template_parameters();
+  
   using std::abs;
   eigen_assert(matrix.cols() == matrix.rows());
   eigen_assert((options&~(EigVecMask|GenEigMask))==0
@@ -406,7 +415,7 @@ SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType>
 
   // declare some aliases
   RealVectorType& diag = m_eivalues;
-  MatrixType& mat = m_eivec;
+  EigenvectorsType& mat = m_eivec;
 
   // map the matrix coefficients to [-1:1] to avoid over- and underflow.
   mat = matrix.template triangularView<Lower>();
@@ -442,7 +451,7 @@ SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType>
     while (start>0 && m_subdiag[start-1]!=0)
       start--;
 
-    internal::tridiagonal_qr_step<MatrixType::Flags&RowMajorBit ? RowMajor : ColMajor>(diag.data(), m_subdiag.data(), start, end, computeEigenvectors ? m_eivec.data() : (Scalar*)0, n);
+    internal::tridiagonal_qr_step(diag.data(), m_subdiag.data(), start, end, computeEigenvectors ? m_eivec.data() : (Scalar*)0, n);
   }
 
   if (iter <= m_maxIterations * n)
@@ -490,7 +499,13 @@ template<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,3
   typedef typename SolverType::MatrixType MatrixType;
   typedef typename SolverType::RealVectorType VectorType;
   typedef typename SolverType::Scalar Scalar;
+  typedef typename MatrixType::Index Index;
+  typedef typename SolverType::EigenvectorsType EigenvectorsType;
   
+  /** \internal
+   * Computes the roots of the characteristic polynomial of \a m.
+   * For numerical stability m.trace() should be near zero and to avoid over- or underflow m should be normalized.
+   */
   static inline void computeRoots(const MatrixType& m, VectorType& roots)
   {
     using std::sqrt;
@@ -510,158 +525,123 @@ template<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,3
     // Construct the parameters used in classifying the roots of the equation
     // and in solving the equation for the roots in closed form.
     Scalar c2_over_3 = c2*s_inv3;
-    Scalar a_over_3 = (c1 - c2*c2_over_3)*s_inv3;
+    Scalar a_over_3 = (c2*c2_over_3 - c1)*s_inv3;
-    if (a_over_3 > Scalar(0))
+    if(a_over_3<Scalar(0))
       a_over_3 = Scalar(0);
 
     Scalar half_b = Scalar(0.5)*(c0 + c2_over_3*(Scalar(2)*c2_over_3*c2_over_3 - c1));
 
-    Scalar q = half_b*half_b + a_over_3*a_over_3*a_over_3;
+    Scalar q = a_over_3*a_over_3*a_over_3 - half_b*half_b;
-    if (q > Scalar(0))
+    if(q<Scalar(0))
       q = Scalar(0);
 
     // Compute the eigenvalues by solving for the roots of the polynomial.
-    Scalar rho = sqrt(-a_over_3);
+    Scalar rho = sqrt(a_over_3);
-    Scalar theta = atan2(sqrt(-q),half_b)*s_inv3;
+    Scalar theta = atan2(sqrt(q),half_b)*s_inv3;  // since sqrt(q) > 0, atan2 is in [0, pi] and theta is in [0, pi/3]
     Scalar cos_theta = cos(theta);
     Scalar sin_theta = sin(theta);
-    roots(0) = c2_over_3 + Scalar(2)*rho*cos_theta;
+    // roots are already sorted, since cos is monotonically decreasing on [0, pi]
-    roots(1) = c2_over_3 - rho*(cos_theta + s_sqrt3*sin_theta);
+    roots(0) = c2_over_3 - rho*(cos_theta + s_sqrt3*sin_theta); // == 2*rho*cos(theta+2pi/3)
-    roots(2) = c2_over_3 - rho*(cos_theta - s_sqrt3*sin_theta);
+    roots(1) = c2_over_3 - rho*(cos_theta - s_sqrt3*sin_theta); // == 2*rho*cos(theta+ pi/3)
-
+    roots(2) = c2_over_3 + Scalar(2)*rho*cos_theta;
-    // Sort in increasing order.
-    if (roots(0) >= roots(1))
-      std::swap(roots(0),roots(1));
-    if (roots(1) >= roots(2))
-    {
-      std::swap(roots(1),roots(2));
-      if (roots(0) >= roots(1))
-        std::swap(roots(0),roots(1));
-    }
   }
-  
+
+  static inline bool extract_kernel(MatrixType& mat, Ref<VectorType> res, Ref<VectorType> representative)
+  {
+    using std::abs;
+    Index i0;
+    // Find non-zero column i0 (by construction, there must exist a non zero coefficient on the diagonal):
+    mat.diagonal().cwiseAbs().maxCoeff(&i0);
+    // mat.col(i0) is a good candidate for an orthogonal vector to the current eigenvector,
+    // so let's save it:
+    representative = mat.col(i0);
+    Scalar n0, n1;
+    VectorType c0, c1;
+    n0 = (c0 = representative.cross(mat.col((i0+1)%3))).squaredNorm();
+    n1 = (c1 = representative.cross(mat.col((i0+2)%3))).squaredNorm();
+    if(n0>n1) res = c0/std::sqrt(n0);
+    else      res = c1/std::sqrt(n1);
+
+    return true;
+  }
+
   static inline void run(SolverType& solver, const MatrixType& mat, int options)
   {
-    using std::sqrt;
     eigen_assert(mat.cols() == 3 && mat.cols() == mat.rows());
     eigen_assert((options&~(EigVecMask|GenEigMask))==0
             && (options&EigVecMask)!=EigVecMask
             && "invalid option parameter");
     bool computeEigenvectors = (options&ComputeEigenvectors)==ComputeEigenvectors;
     
-    MatrixType& eivecs = solver.m_eivec;
+    EigenvectorsType& eivecs = solver.m_eivec;
     VectorType& eivals = solver.m_eivalues;
   
-    // map the matrix coefficients to [-1:1] to avoid over- and underflow.
+    // Shift the matrix to the mean eigenvalue and map the matrix coefficients to [-1:1] to avoid over- and underflow.
-    Scalar scale = mat.cwiseAbs().maxCoeff();
+    Scalar shift = mat.trace() / Scalar(3);
-    MatrixType scaledMat = mat / scale;
+    // TODO Avoid this copy. Currently it is necessary to suppress bogus values when determining maxCoeff and for computing the eigenvectors later
+    MatrixType scaledMat = mat.template selfadjointView<Lower>();
+    scaledMat.diagonal().array() -= shift;
+    Scalar scale = scaledMat.cwiseAbs().maxCoeff();
+    if(scale > 0) scaledMat /= scale;   // TODO for scale==0 we could save the remaining operations
 
     // compute the eigenvalues
     computeRoots(scaledMat,eivals);
 
-    // compute the eigen vectors
+    // compute the eigenvectors
     if(computeEigenvectors)
     {
-      Scalar safeNorm2 = Eigen::NumTraits<Scalar>::epsilon();
       if((eivals(2)-eivals(0))<=Eigen::NumTraits<Scalar>::epsilon())
       {
+        // All three eigenvalues are numerically the same
         eivecs.setIdentity();
       }
       else
       {
-        scaledMat = scaledMat.template selfadjointView<Lower>();
         MatrixType tmp;
         tmp = scaledMat;
 
+        // Compute the eigenvector of the most distinct eigenvalue
         Scalar d0 = eivals(2) - eivals(1);
         Scalar d1 = eivals(1) - eivals(0);
-        int k =  d0 > d1 ? 2 : 0;
+        Index k(0), l(2);
-        d0 = d0 > d1 ? d0 : d1;
+        if(d0 > d1)
-
-        tmp.diagonal().array () -= eivals(k);
-        VectorType cross;
-        Scalar n;
-        n = (cross = tmp.row(0).cross(tmp.row(1))).squaredNorm();
-
-        if(n>safeNorm2)
         {
-          eivecs.col(k) = cross / sqrt(n);
+          std::swap(k,l);
+          d0 = d1;
+        }
+
+        // Compute the eigenvector of index k
+        {
+          tmp.diagonal().array () -= eivals(k);
+          // By construction, 'tmp' is of rank 2, and its kernel corresponds to the respective eigenvector.
+          extract_kernel(tmp, eivecs.col(k), eivecs.col(l));
+        }
+
+        // Compute eigenvector of index l
+        if(d0<=2*Eigen::NumTraits<Scalar>::epsilon()*d1)
+        {
+          // If d0 is too small, then the two other eigenvalues are numerically the same,
+          // and thus we only have to ortho-normalize the near orthogonal vector we saved above.
+          eivecs.col(l) -= eivecs.col(k).dot(eivecs.col(l))*eivecs.col(l);
+          eivecs.col(l).normalize();
         }
         else
         {
-          n = (cross = tmp.row(0).cross(tmp.row(2))).squaredNorm();
+          tmp = scaledMat;
+          tmp.diagonal().array () -= eivals(l);
 
-          if(n>safeNorm2)
+          VectorType dummy;
-          {
+          extract_kernel(tmp, eivecs.col(l), dummy);
-            eivecs.col(k) = cross / sqrt(n);
-          }
-          else
-          {
-            n = (cross = tmp.row(1).cross(tmp.row(2))).squaredNorm();
-
-            if(n>safeNorm2)
-            {
-              eivecs.col(k) = cross / sqrt(n);
-            }
-            else
-            {
-              // the input matrix and/or the eigenvaues probably contains some inf/NaN,
-              // => exit
-              // scale back to the original size.
-              eivals *= scale;
-
-              solver.m_info = NumericalIssue;
-              solver.m_isInitialized = true;
-              solver.m_eigenvectorsOk = computeEigenvectors;
-              return;
-            }
-          }
         }
 
-        tmp = scaledMat;
+        // Compute last eigenvector from the other two
-        tmp.diagonal().array() -= eivals(1);
+        eivecs.col(1) = eivecs.col(2).cross(eivecs.col(0)).normalized();
-
-        if(d0<=Eigen::NumTraits<Scalar>::epsilon())
-        {
-          eivecs.col(1) = eivecs.col(k).unitOrthogonal();
-        }
-        else
-        {
-          n = (cross = eivecs.col(k).cross(tmp.row(0))).squaredNorm();
-          if(n>safeNorm2)
-          {
-            eivecs.col(1) = cross / sqrt(n);
-          }
-          else
-          {
-            n = (cross = eivecs.col(k).cross(tmp.row(1))).squaredNorm();
-            if(n>safeNorm2)
-              eivecs.col(1) = cross / sqrt(n);
-            else
-            {
-              n = (cross = eivecs.col(k).cross(tmp.row(2))).squaredNorm();
-              if(n>safeNorm2)
-                eivecs.col(1) = cross / sqrt(n);
-              else
-              {
-                // we should never reach this point,
-                // if so the last two eigenvalues are likely to be very close to each other
-                eivecs.col(1) = eivecs.col(k).unitOrthogonal();
-              }
-            }
-          }
-
-          // make sure that eivecs[1] is orthogonal to eivecs[2]
-          // FIXME: this step should not be needed
-          Scalar d = eivecs.col(1).dot(eivecs.col(k));
-          eivecs.col(1) = (eivecs.col(1) - d * eivecs.col(k)).normalized();
-        }
-
-        eivecs.col(k==2 ? 0 : 2) = eivecs.col(k).cross(eivecs.col(1)).normalized();
       }
     }
+
     // Rescale back to the original size.
     eivals *= scale;
+    eivals.array() += shift;
     
     solver.m_info = Success;
     solver.m_isInitialized = true;
@@ -675,11 +655,12 @@ template<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,2
   typedef typename SolverType::MatrixType MatrixType;
   typedef typename SolverType::RealVectorType VectorType;
   typedef typename SolverType::Scalar Scalar;
+  typedef typename SolverType::EigenvectorsType EigenvectorsType;
   
   static inline void computeRoots(const MatrixType& m, VectorType& roots)
   {
     using std::sqrt;
-    const Scalar t0 = Scalar(0.5) * sqrt( numext::abs2(m(0,0)-m(1,1)) + Scalar(4)*m(1,0)*m(1,0));
+    const Scalar t0 = Scalar(0.5) * sqrt( numext::abs2(m(0,0)-m(1,1)) + Scalar(4)*numext::abs2(m(1,0)));
     const Scalar t1 = Scalar(0.5) * (m(0,0) + m(1,1));
     roots(0) = t1 - t0;
     roots(1) = t1 + t0;
@@ -688,13 +669,15 @@ template<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,2
   static inline void run(SolverType& solver, const MatrixType& mat, int options)
   {
     using std::sqrt;
+    using std::abs;
+
     eigen_assert(mat.cols() == 2 && mat.cols() == mat.rows());
     eigen_assert((options&~(EigVecMask|GenEigMask))==0
             && (options&EigVecMask)!=EigVecMask
             && "invalid option parameter");
     bool computeEigenvectors = (options&ComputeEigenvectors)==ComputeEigenvectors;
     
-    MatrixType& eivecs = solver.m_eivec;
+    EigenvectorsType& eivecs = solver.m_eivec;
     VectorType& eivals = solver.m_eivalues;
   
     // map the matrix coefficients to [-1:1] to avoid over- and underflow.
@@ -708,22 +691,29 @@ template<typename SolverType> struct direct_selfadjoint_eigenvalues<SolverType,2
     // compute the eigen vectors
     if(computeEigenvectors)
     {
-      scaledMat.diagonal().array () -= eivals(1);
+      if((eivals(1)-eivals(0))<=abs(eivals(1))*Eigen::NumTraits<Scalar>::epsilon())
-      Scalar a2 = numext::abs2(scaledMat(0,0));
-      Scalar c2 = numext::abs2(scaledMat(1,1));
-      Scalar b2 = numext::abs2(scaledMat(1,0));
-      if(a2>c2)
       {
-        eivecs.col(1) << -scaledMat(1,0), scaledMat(0,0);
+        eivecs.setIdentity();
-        eivecs.col(1) /= sqrt(a2+b2);
       }
       else
       {
-        eivecs.col(1) << -scaledMat(1,1), scaledMat(1,0);
+        scaledMat.diagonal().array () -= eivals(1);
-        eivecs.col(1) /= sqrt(c2+b2);
+        Scalar a2 = numext::abs2(scaledMat(0,0));
-      }
+        Scalar c2 = numext::abs2(scaledMat(1,1));
+        Scalar b2 = numext::abs2(scaledMat(1,0));
+        if(a2>c2)
+        {
+          eivecs.col(1) << -scaledMat(1,0), scaledMat(0,0);
+          eivecs.col(1) /= sqrt(a2+b2);
+        }
+        else
+        {
+          eivecs.col(1) << -scaledMat(1,1), scaledMat(1,0);
+          eivecs.col(1) /= sqrt(c2+b2);
+        }
 
-      eivecs.col(0) << eivecs.col(1).unitOrthogonal();
+        eivecs.col(0) << eivecs.col(1).unitOrthogonal();
+      }
     }
     
     // Rescale back to the original size.
@@ -746,7 +736,7 @@ SelfAdjointEigenSolver<MatrixType>& SelfAdjointEigenSolver<MatrixType>
 }
 
 namespace internal {
-template<int StorageOrder,typename RealScalar, typename Scalar, typename Index>
+template<typename RealScalar, typename Scalar, typename Index>
 static void tridiagonal_qr_step(RealScalar* diag, RealScalar* subdiag, Index start, Index end, Scalar* matrixQ, Index n)
 {
   using std::abs;
@@ -798,8 +788,7 @@ static void tridiagonal_qr_step(RealScalar* diag, RealScalar* subdiag, Index sta
     // apply the givens rotation to the unit matrix Q = Q * G
     if (matrixQ)
     {
-      // FIXME if StorageOrder == RowMajor this operation is not very efficient
+      Map<Matrix<Scalar,Dynamic,Dynamic,ColMajor> > q(matrixQ,n,n);
-      Map<Matrix<Scalar,Dynamic,Dynamic,StorageOrder> > q(matrixQ,n,n);
       q.applyOnTheRight(k,k+1,rot);
     }
   }

Eigen/src/Eigenvalues/Tridiagonalization.h

@@ -367,10 +367,10 @@ void tridiagonalization_inplace(MatrixType& matA, CoeffVectorType& hCoeffs)
     hCoeffs.tail(n-i-1).noalias() = (matA.bottomRightCorner(remainingSize,remainingSize).template selfadjointView<Lower>()
                                   * (conj(h) * matA.col(i).tail(remainingSize)));
 
-    hCoeffs.tail(n-i-1) += (conj(h)*Scalar(-0.5)*(hCoeffs.tail(remainingSize).dot(matA.col(i).tail(remainingSize)))) * matA.col(i).tail(n-i-1);
+    hCoeffs.tail(n-i-1) += (conj(h)*RealScalar(-0.5)*(hCoeffs.tail(remainingSize).dot(matA.col(i).tail(remainingSize)))) * matA.col(i).tail(n-i-1);
 
     matA.bottomRightCorner(remainingSize, remainingSize).template selfadjointView<Lower>()
-      .rankUpdate(matA.col(i).tail(remainingSize), hCoeffs.tail(remainingSize), -1);
+      .rankUpdate(matA.col(i).tail(remainingSize), hCoeffs.tail(remainingSize), Scalar(-1));
 
     matA.col(i).coeffRef(i+1) = beta;
     hCoeffs.coeffRef(i) = h;

Eigen/src/Geometry/AlignedBox.h

@@ -19,10 +19,12 @@ namespace Eigen {
   *
   * \brief An axis aligned box
   *
-  * \param _Scalar the type of the scalar coefficients
+  * \tparam _Scalar the type of the scalar coefficients
-  * \param _AmbientDim the dimension of the ambient space, can be a compile time value or Dynamic.
+  * \tparam _AmbientDim the dimension of the ambient space, can be a compile time value or Dynamic.
   *
   * This class represents an axis aligned box as a pair of the minimal and maximal corners.
+  * \warning The result of most methods is undefined when applied to an empty box. You can check for empty boxes using isEmpty().
+  * \sa alignedboxtypedefs
   */
 template <typename _Scalar, int _AmbientDim>
 class AlignedBox
@@ -40,18 +42,21 @@ EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim)
   /** Define constants to name the corners of a 1D, 2D or 3D axis aligned bounding box */
   enum CornerType
   {
-    /** 1D names */
+    /** 1D names @{ */
     Min=0, Max=1,
+    /** @} */
 
-    /** Added names for 2D */
+    /** Identifier for 2D corner @{ */
     BottomLeft=0, BottomRight=1,
     TopLeft=2, TopRight=3,
+    /** @} */
 
-    /** Added names for 3D */
+    /** Identifier for 3D corner  @{ */
     BottomLeftFloor=0, BottomRightFloor=1,
     TopLeftFloor=2, TopRightFloor=3,
     BottomLeftCeil=4, BottomRightCeil=5,
     TopLeftCeil=6, TopRightCeil=7
+    /** @} */
   };
 
 
@@ -63,34 +68,33 @@ EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim)
   inline explicit AlignedBox(Index _dim) : m_min(_dim), m_max(_dim)
   { setEmpty(); }
 
-  /** Constructs a box with extremities \a _min and \a _max. */
+  /** Constructs a box with extremities \a _min and \a _max.
+   * \warning If either component of \a _min is larger than the same component of \a _max, the constructed box is empty. */
   template<typename OtherVectorType1, typename OtherVectorType2>
   inline AlignedBox(const OtherVectorType1& _min, const OtherVectorType2& _max) : m_min(_min), m_max(_max) {}
 
   /** Constructs a box containing a single point \a p. */
   template<typename Derived>
-  inline explicit AlignedBox(const MatrixBase<Derived>& a_p)
+  inline explicit AlignedBox(const MatrixBase<Derived>& p) : m_min(p), m_max(m_min)
-  {
+  { }
-    typename internal::nested<Derived,2>::type p(a_p.derived());
-    m_min = p;
-    m_max = p;
-  }
 
   ~AlignedBox() {}
 
   /** \returns the dimension in which the box holds */
   inline Index dim() const { return AmbientDimAtCompileTime==Dynamic ? m_min.size() : Index(AmbientDimAtCompileTime); }
 
-  /** \deprecated use isEmpty */
+  /** \deprecated use isEmpty() */
   inline bool isNull() const { return isEmpty(); }
 
-  /** \deprecated use setEmpty */
+  /** \deprecated use setEmpty() */
   inline void setNull() { setEmpty(); }
 
-  /** \returns true if the box is empty. */
+  /** \returns true if the box is empty.
+   * \sa setEmpty */
   inline bool isEmpty() const { return (m_min.array() > m_max.array()).any(); }
 
-  /** Makes \c *this an empty box. */
+  /** Makes \c *this an empty box.
+   * \sa isEmpty */
   inline void setEmpty()
   {
     m_min.setConstant( ScalarTraits::highest() );
@@ -159,7 +163,7 @@ EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim)
    * a uniform distribution */
   inline VectorType sample() const
   {
-    VectorType r;
+    VectorType r(dim());
     for(Index d=0; d<dim(); ++d)
     {
       if(!ScalarTraits::IsInteger)
@@ -175,27 +179,34 @@ EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim)
 
   /** \returns true if the point \a p is inside the box \c *this. */
   template<typename Derived>
-  inline bool contains(const MatrixBase<Derived>& a_p) const
+  inline bool contains(const MatrixBase<Derived>& p) const
   {
-    typename internal::nested<Derived,2>::type p(a_p.derived());
+    typename internal::nested<Derived,2>::type p_n(p.derived());
-    return (m_min.array()<=p.array()).all() && (p.array()<=m_max.array()).all();
+    return (m_min.array()<=p_n.array()).all() && (p_n.array()<=m_max.array()).all();
   }
 
   /** \returns true if the box \a b is entirely inside the box \c *this. */
   inline bool contains(const AlignedBox& b) const
   { return (m_min.array()<=(b.min)().array()).all() && ((b.max)().array()<=m_max.array()).all(); }
 
-  /** Extends \c *this such that it contains the point \a p and returns a reference to \c *this. */
+  /** \returns true if the box \a b is intersecting the box \c *this.
+   * \sa intersection, clamp */
+  inline bool intersects(const AlignedBox& b) const
+  { return (m_min.array()<=(b.max)().array()).all() && ((b.min)().array()<=m_max.array()).all(); }
+
+  /** Extends \c *this such that it contains the point \a p and returns a reference to \c *this.
+   * \sa extend(const AlignedBox&) */
   template<typename Derived>
-  inline AlignedBox& extend(const MatrixBase<Derived>& a_p)
+  inline AlignedBox& extend(const MatrixBase<Derived>& p)
   {
-    typename internal::nested<Derived,2>::type p(a_p.derived());
+    typename internal::nested<Derived,2>::type p_n(p.derived());
-    m_min = m_min.cwiseMin(p);
+    m_min = m_min.cwiseMin(p_n);
-    m_max = m_max.cwiseMax(p);
+    m_max = m_max.cwiseMax(p_n);
     return *this;
   }
 
-  /** Extends \c *this such that it contains the box \a b and returns a reference to \c *this. */
+  /** Extends \c *this such that it contains the box \a b and returns a reference to \c *this.
+   * \sa merged, extend(const MatrixBase&) */
   inline AlignedBox& extend(const AlignedBox& b)
   {
     m_min = m_min.cwiseMin(b.m_min);
@@ -203,7 +214,9 @@ EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim)
     return *this;
   }
 
-  /** Clamps \c *this by the box \a b and returns a reference to \c *this. */
+  /** Clamps \c *this by the box \a b and returns a reference to \c *this.
+   * \note If the boxes don't intersect, the resulting box is empty.
+   * \sa intersection(), intersects() */
   inline AlignedBox& clamp(const AlignedBox& b)
   {
     m_min = m_min.cwiseMax(b.m_min);
@@ -211,11 +224,15 @@ EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim)
     return *this;
   }
 
-  /** Returns an AlignedBox that is the intersection of \a b and \c *this */
+  /** Returns an AlignedBox that is the intersection of \a b and \c *this
+   * \note If the boxes don't intersect, the resulting box is empty.
+   * \sa intersects(), clamp, contains()  */
   inline AlignedBox intersection(const AlignedBox& b) const
   {return AlignedBox(m_min.cwiseMax(b.m_min), m_max.cwiseMin(b.m_max)); }
 
-  /** Returns an AlignedBox that is the union of \a b and \c *this */
+  /** Returns an AlignedBox that is the union of \a b and \c *this.
+   * \note Merging with an empty box may result in a box bigger than \c *this. 
+   * \sa extend(const AlignedBox&) */
   inline AlignedBox merged(const AlignedBox& b) const
   { return AlignedBox(m_min.cwiseMin(b.m_min), m_max.cwiseMax(b.m_max)); }
 
@@ -231,20 +248,20 @@ EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim)
 
   /** \returns the squared distance between the point \a p and the box \c *this,
     * and zero if \a p is inside the box.
-    * \sa exteriorDistance()
+    * \sa exteriorDistance(const MatrixBase&), squaredExteriorDistance(const AlignedBox&)
     */
   template<typename Derived>
-  inline Scalar squaredExteriorDistance(const MatrixBase<Derived>& a_p) const;
+  inline Scalar squaredExteriorDistance(const MatrixBase<Derived>& p) const;
 
   /** \returns the squared distance between the boxes \a b and \c *this,
     * and zero if the boxes intersect.
-    * \sa exteriorDistance()
+    * \sa exteriorDistance(const AlignedBox&), squaredExteriorDistance(const MatrixBase&)
     */
   inline Scalar squaredExteriorDistance(const AlignedBox& b) const;
 
   /** \returns the distance between the point \a p and the box \c *this,
     * and zero if \a p is inside the box.
-    * \sa squaredExteriorDistance()
+    * \sa squaredExteriorDistance(const MatrixBase&), exteriorDistance(const AlignedBox&)
     */
   template<typename Derived>
   inline NonInteger exteriorDistance(const MatrixBase<Derived>& p) const
@@ -252,7 +269,7 @@ EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF_VECTORIZABLE_FIXED_SIZE(_Scalar,_AmbientDim)
 
   /** \returns the distance between the boxes \a b and \c *this,
     * and zero if the boxes intersect.
-    * \sa squaredExteriorDistance()
+    * \sa squaredExteriorDistance(const AlignedBox&), exteriorDistance(const MatrixBase&)
     */
   inline NonInteger exteriorDistance(const AlignedBox& b) const
   { using std::sqrt; return sqrt(NonInteger(squaredExteriorDistance(b))); }

Eigen/src/Geometry/AngleAxis.h

@@ -83,10 +83,17 @@ public:
   template<typename Derived>
   inline explicit AngleAxis(const MatrixBase<Derived>& m) { *this = m; }
 
+  /** \returns the value of the rotation angle in radian */
   Scalar angle() const { return m_angle; }
+  /** \returns a read-write reference to the stored angle in radian */
   Scalar& angle() { return m_angle; }
 
+  /** \returns the rotation axis */
   const Vector3& axis() const { return m_axis; }
+  /** \returns a read-write reference to the stored rotation axis.
+    *
+    * \warning The rotation axis must remain a \b unit vector.
+    */
   Vector3& axis() { return m_axis; }
 
   /** Concatenates two rotations */
@@ -131,7 +138,7 @@ public:
     m_angle = Scalar(other.angle());
   }
 
-  static inline const AngleAxis Identity() { return AngleAxis(0, Vector3::UnitX()); }
+  static inline const AngleAxis Identity() { return AngleAxis(Scalar(0), Vector3::UnitX()); }
 
   /** \returns \c true if \c *this is approximately equal to \a other, within the precision
     * determined by \a prec.
@@ -165,8 +172,8 @@ AngleAxis<Scalar>& AngleAxis<Scalar>::operator=(const QuaternionBase<QuatDerived
   Scalar n2 = q.vec().squaredNorm();
   if (n2 < NumTraits<Scalar>::dummy_precision()*NumTraits<Scalar>::dummy_precision())
   {
-    m_angle = 0;
+    m_angle = Scalar(0);
-    m_axis << 1, 0, 0;
+    m_axis << Scalar(1), Scalar(0), Scalar(0);
   }
   else
   {

Eigen/src/Geometry/Homogeneous.h

@@ -75,11 +75,11 @@ template<typename MatrixType,int _Direction> class Homogeneous
     inline Index rows() const { return m_matrix.rows() + (int(Direction)==Vertical   ? 1 : 0); }
     inline Index cols() const { return m_matrix.cols() + (int(Direction)==Horizontal ? 1 : 0); }
 
-    inline Scalar coeff(Index row, Index col) const
+    inline Scalar coeff(Index row, Index col=0) const
     {
       if(  (int(Direction)==Vertical   && row==m_matrix.rows())
         || (int(Direction)==Horizontal && col==m_matrix.cols()))
-        return 1;
+        return Scalar(1);
       return m_matrix.coeff(row, col);
     }
 

Eigen/src/Geometry/ParametrizedLine.h

@@ -129,7 +129,7 @@ public:
     * determined by \a prec.
     *
     * \sa MatrixBase::isApprox() */
-  bool isApprox(const ParametrizedLine& other, typename NumTraits<Scalar>::Real prec = NumTraits<Scalar>::dummy_precision()) const
+  bool isApprox(const ParametrizedLine& other, const typename NumTraits<Scalar>::Real& prec = NumTraits<Scalar>::dummy_precision()) const
   { return m_origin.isApprox(other.m_origin, prec) && m_direction.isApprox(other.m_direction, prec); }
 
 protected:

Eigen/src/Geometry/Quaternion.h

@@ -102,11 +102,11 @@ public:
   /** \returns a quaternion representing an identity rotation
     * \sa MatrixBase::Identity()
     */
-  static inline Quaternion<Scalar> Identity() { return Quaternion<Scalar>(1, 0, 0, 0); }
+  static inline Quaternion<Scalar> Identity() { return Quaternion<Scalar>(Scalar(1), Scalar(0), Scalar(0), Scalar(0)); }
 
   /** \sa QuaternionBase::Identity(), MatrixBase::setIdentity()
     */
-  inline QuaternionBase& setIdentity() { coeffs() << 0, 0, 0, 1; return *this; }
+  inline QuaternionBase& setIdentity() { coeffs() << Scalar(0), Scalar(0), Scalar(0), Scalar(1); return *this; }
 
   /** \returns the squared norm of the quaternion's coefficients
     * \sa QuaternionBase::norm(), MatrixBase::squaredNorm()
@@ -161,7 +161,7 @@ public:
   { return coeffs().isApprox(other.coeffs(), prec); }
 
 	/** return the result vector of \a v through the rotation*/
-  EIGEN_STRONG_INLINE Vector3 _transformVector(Vector3 v) const;
+  EIGEN_STRONG_INLINE Vector3 _transformVector(const Vector3& v) const;
 
   /** \returns \c *this with scalar type casted to \a NewScalarType
     *
@@ -231,7 +231,7 @@ class Quaternion : public QuaternionBase<Quaternion<_Scalar,_Options> >
 public:
   typedef _Scalar Scalar;
 
-  EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Quaternion)
+  EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Quaternion)
   using Base::operator*=;
 
   typedef typename internal::traits<Quaternion>::Coefficients Coefficients;
@@ -276,7 +276,7 @@ public:
   inline Coefficients& coeffs() { return m_coeffs;}
   inline const Coefficients& coeffs() const { return m_coeffs;}
 
-  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(IsAligned)
+  EIGEN_MAKE_ALIGNED_OPERATOR_NEW_IF(bool(IsAligned))
 
 protected:
   Coefficients m_coeffs;
@@ -341,7 +341,7 @@ class Map<const Quaternion<_Scalar>, _Options >
   public:
     typedef _Scalar Scalar;
     typedef typename internal::traits<Map>::Coefficients Coefficients;
-    EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Map)
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Map)
     using Base::operator*=;
 
     /** Constructs a Mapped Quaternion object from the pointer \a coeffs
@@ -378,7 +378,7 @@ class Map<Quaternion<_Scalar>, _Options >
   public:
     typedef _Scalar Scalar;
     typedef typename internal::traits<Map>::Coefficients Coefficients;
-    EIGEN_INHERIT_ASSIGNMENT_EQUAL_OPERATOR(Map)
+    EIGEN_INHERIT_ASSIGNMENT_OPERATORS(Map)
     using Base::operator*=;
 
     /** Constructs a Mapped Quaternion object from the pointer \a coeffs
@@ -461,7 +461,7 @@ EIGEN_STRONG_INLINE Derived& QuaternionBase<Derived>::operator*= (const Quaterni
   */
 template <class Derived>
 EIGEN_STRONG_INLINE typename QuaternionBase<Derived>::Vector3
-QuaternionBase<Derived>::_transformVector(Vector3 v) const
+QuaternionBase<Derived>::_transformVector(const Vector3& v) const
 {
     // Note that this algorithm comes from the optimization by hand
     // of the conversion to a Matrix followed by a Matrix/Vector product.
@@ -637,7 +637,7 @@ inline Quaternion<typename internal::traits<Derived>::Scalar> QuaternionBase<Der
 {
   // FIXME should this function be called multiplicativeInverse and conjugate() be called inverse() or opposite()  ??
   Scalar n2 = this->squaredNorm();
-  if (n2 > 0)
+  if (n2 > Scalar(0))
     return Quaternion<Scalar>(conjugate().coeffs() / n2);
   else
   {
@@ -667,12 +667,10 @@ template <class OtherDerived>
 inline typename internal::traits<Derived>::Scalar
 QuaternionBase<Derived>::angularDistance(const QuaternionBase<OtherDerived>& other) const
 {
-  using std::acos;
+  using std::atan2;
   using std::abs;
-  Scalar d = abs(this->dot(other));
+  Quaternion<Scalar> d = (*this) * other.conjugate();
-  if (d>=Scalar(1))
+  return Scalar(2) * atan2( d.vec().norm(), abs(d.w()) );
-    return Scalar(0);
-  return Scalar(2) * acos(d);
 }
 
  
@@ -712,7 +710,7 @@ QuaternionBase<Derived>::slerp(const Scalar& t, const QuaternionBase<OtherDerive
     scale0 = sin( ( Scalar(1) - t ) * theta) / sinTheta;
     scale1 = sin( ( t * theta) ) / sinTheta;
   }
-  if(d<0) scale1 = -scale1;
+  if(d<Scalar(0)) scale1 = -scale1;
 
   return Quaternion<Scalar>(scale0 * coeffs() + scale1 * other.coeffs());
 }

Eigen/src/Geometry/Rotation2D.h

@@ -59,7 +59,7 @@ protected:
 public:
 
   /** Construct a 2D counter clock wise rotation from the angle \a a in radian. */
-  explicit inline Rotation2D(const Scalar& a) : m_angle(a) {}
+  inline Rotation2D(const Scalar& a) : m_angle(a) {}
   
   /** Default constructor wihtout initialization. The represented rotation is undefined. */
   Rotation2D() {}

Eigen/src/Geometry/Transform.h

@@ -102,15 +102,15 @@ template<int Mode> struct transform_make_affine;
   *
   * However, unlike a plain matrix, the Transform class provides many features
   * simplifying both its assembly and usage. In particular, it can be composed
-  * with any other transformations (Transform,Translation,RotationBase,Matrix)
+  * with any other transformations (Transform,Translation,RotationBase,DiagonalMatrix)
   * and can be directly used to transform implicit homogeneous vectors. All these
   * operations are handled via the operator*. For the composition of transformations,
   * its principle consists to first convert the right/left hand sides of the product
   * to a compatible (Dim+1)^2 matrix and then perform a pure matrix product.
   * Of course, internally, operator* tries to perform the minimal number of operations
   * according to the nature of each terms. Likewise, when applying the transform
-  * to non homogeneous vectors, the latters are automatically promoted to homogeneous
+  * to points, the latters are automatically promoted to homogeneous vectors
-  * one before doing the matrix product. The convertions to homogeneous representations
+  * before doing the matrix product. The conventions to homogeneous representations
   * are performed as follow:
   *
   * \b Translation t (Dim)x(1):
@@ -124,7 +124,7 @@ template<int Mode> struct transform_make_affine;
   * R & 0\\
   * 0\,...\,0 & 1
   * \end{array} \right) \f$
-  *
+  *<!--
   * \b Linear \b Matrix L (Dim)x(Dim):
   * \f$ \left( \begin{array}{cc}
   * L & 0\\
@@ -136,14 +136,20 @@ template<int Mode> struct transform_make_affine;
   * A\\
   * 0\,...\,0\,1
   * \end{array} \right) \f$
+  *-->
+  * \b Scaling \b DiagonalMatrix S (Dim)x(Dim):
+  * \f$ \left( \begin{array}{cc}
+  * S & 0\\
+  * 0\,...\,0 & 1
+  * \end{array} \right) \f$
   *
-  * \b Column \b vector v (Dim)x(1):
+  * \b Column \b point v (Dim)x(1):
   * \f$ \left( \begin{array}{c}
   * v\\
   * 1
   * \end{array} \right) \f$
   *
-  * \b Set \b of \b column \b vectors V1...Vn (Dim)x(n):
+  * \b Set \b of \b column \b points V1...Vn (Dim)x(n):
   * \f$ \left( \begin{array}{ccc}
   * v_1 & ... & v_n\\
   * 1 & ... & 1
@@ -384,26 +390,39 @@ public:
   /** \returns a writable expression of the translation vector of the transformation */
   inline TranslationPart translation() { return TranslationPart(m_matrix,0,Dim); }
 
-  /** \returns an expression of the product between the transform \c *this and a matrix expression \a other
+  /** \returns an expression of the product between the transform \c *this and a matrix expression \a other.
     *
-    * The right hand side \a other might be either:
+    * The right-hand-side \a other can be either:
-    * \li a vector of size Dim,
     * \li an homogeneous vector of size Dim+1,
-    * \li a set of vectors of size Dim x Dynamic,
+    * \li a set of homogeneous vectors of size Dim+1 x N,
-    * \li a set of homogeneous vectors of size Dim+1 x Dynamic,
-    * \li a linear transformation matrix of size Dim x Dim,
-    * \li an affine transformation matrix of size Dim x Dim+1,
     * \li a transformation matrix of size Dim+1 x Dim+1.
+    *
+    * Moreover, if \c *this represents an affine transformation (i.e., Mode!=Projective), then \a other can also be:
+    * \li a point of size Dim (computes: \code this->linear() * other + this->translation()\endcode),
+    * \li a set of N points as a Dim x N matrix (computes: \code (this->linear() * other).colwise() + this->translation()\endcode),
+    *
+    * In all cases, the return type is a matrix or vector of same sizes as the right-hand-side \a other.
+    *
+    * If you want to interpret \a other as a linear or affine transformation, then first convert it to a Transform<> type,
+    * or do your own cooking.
+    *
+    * Finally, if you want to apply Affine transformations to vectors, then explicitly apply the linear part only:
+    * \code
+    * Affine3f A;
+    * Vector3f v1, v2;
+    * v2 = A.linear() * v1;
+    * \endcode
+    *
     */
   // note: this function is defined here because some compilers cannot find the respective declaration
   template<typename OtherDerived>
-  EIGEN_STRONG_INLINE const typename internal::transform_right_product_impl<Transform, OtherDerived>::ResultType
+  EIGEN_STRONG_INLINE const typename OtherDerived::PlainObject
   operator * (const EigenBase<OtherDerived> &other) const
   { return internal::transform_right_product_impl<Transform, OtherDerived>::run(*this,other.derived()); }
 
   /** \returns the product expression of a transformation matrix \a a times a transform \a b
     *
-    * The left hand side \a other might be either:
+    * The left hand side \a other can be either:
     * \li a linear transformation matrix of size Dim x Dim,
     * \li an affine transformation matrix of size Dim x Dim+1,
     * \li a general transformation matrix of size Dim+1 x Dim+1.

Eigen/src/Geometry/Translation.h

@@ -162,7 +162,7 @@ public:
     * determined by \a prec.
     *
     * \sa MatrixBase::isApprox() */
-  bool isApprox(const Translation& other, typename NumTraits<Scalar>::Real prec = NumTraits<Scalar>::dummy_precision()) const
+  bool isApprox(const Translation& other, const typename NumTraits<Scalar>::Real& prec = NumTraits<Scalar>::dummy_precision()) const
   { return m_coeffs.isApprox(other.m_coeffs, prec); }
 
 };

Eigen/src/Householder/Householder.h

@@ -75,8 +75,9 @@ void MatrixBase<Derived>::makeHouseholder(
   
   RealScalar tailSqNorm = size()==1 ? RealScalar(0) : tail.squaredNorm();
   Scalar c0 = coeff(0);
+  const RealScalar tol = (std::numeric_limits<RealScalar>::min)();
 
-  if(tailSqNorm == RealScalar(0) && numext::imag(c0)==RealScalar(0))
+  if(tailSqNorm <= tol && numext::abs2(numext::imag(c0))<=tol)
   {
     tau = RealScalar(0);
     beta = numext::real(c0);

Eigen/src/Householder/HouseholderSequence.h

@@ -237,8 +237,9 @@ template<typename VectorsType, typename CoeffsType, int Side> class HouseholderS
     {
       workspace.resize(rows());
       Index vecs = m_length;
+      const typename Dest::Scalar *dst_data = internal::extract_data(dst);
       if(    internal::is_same<typename internal::remove_all<VectorsType>::type,Dest>::value
-          && internal::extract_data(dst) == internal::extract_data(m_vectors))
+          && dst_data!=0 && dst_data == internal::extract_data(m_vectors))
       {
         // in-place
         dst.diagonal().setOnes();

Eigen/src/IterativeLinearSolvers/BasicPreconditioners.h

@@ -65,10 +65,10 @@ class DiagonalPreconditioner
       {
         typename MatType::InnerIterator it(mat,j);
         while(it && it.index()!=j) ++it;
-        if(it && it.index()==j)
+        if(it && it.index()==j && it.value()!=Scalar(0))
           m_invdiag(j) = Scalar(1)/it.value();
         else
-          m_invdiag(j) = 0;
+          m_invdiag(j) = Scalar(1);
       }
       m_isInitialized = true;
       return *this;

Eigen/src/IterativeLinearSolvers/BiCGSTAB.h

@@ -151,20 +151,7 @@ struct traits<BiCGSTAB<_MatrixType,_Preconditioner> >
   * \endcode
   * 
   * By default the iterations start with x=0 as an initial guess of the solution.
-  * One can control the start using the solveWithGuess() method. Here is a step by
+  * One can control the start using the solveWithGuess() method.
-  * step execution example starting with a random guess and printing the evolution
-  * of the estimated error:
-  * * \code
-  * x = VectorXd::Random(n);
-  * solver.setMaxIterations(1);
-  * int i = 0;
-  * do {
-  *   x = solver.solveWithGuess(b,x);
-  *   std::cout << i << " : " << solver.error() << std::endl;
-  *   ++i;
-  * } while (solver.info()!=Success && i<100);
-  * \endcode
-  * Note that such a step by step excution is slightly slower.
   * 
   * \sa class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner
   */
@@ -199,7 +186,8 @@ public:
     * this class becomes invalid. Call compute() to update it with the new
     * matrix A, or modify a copy of A.
     */
-  BiCGSTAB(const MatrixType& A) : Base(A) {}
+  template<typename MatrixDerived>
+  explicit BiCGSTAB(const EigenBase<MatrixDerived>& A) : Base(A.derived()) {}
 
   ~BiCGSTAB() {}
   

Eigen/src/IterativeLinearSolvers/ConjugateGradient.h

@@ -112,9 +112,9 @@ struct traits<ConjugateGradient<_MatrixType,_UpLo,_Preconditioner> >
   * This class allows to solve for A.x = b sparse linear problems using a conjugate gradient algorithm.
   * The sparse matrix A must be selfadjoint. The vectors x and b can be either dense or sparse.
   *
-  * \tparam _MatrixType the type of the sparse matrix A, can be a dense or a sparse matrix.
+  * \tparam _MatrixType the type of the matrix A, can be a dense or a sparse matrix.
-  * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower
+  * \tparam _UpLo the triangular part that will be used for the computations. It can be Lower,
-  *               or Upper. Default is Lower.
+  *               Upper, or Lower|Upper in which the full matrix entries will be considered. Default is Lower.
   * \tparam _Preconditioner the type of the preconditioner. Default is DiagonalPreconditioner
   *
   * The maximal number of iterations and tolerance value can be controlled via the setMaxIterations()
@@ -137,21 +137,10 @@ struct traits<ConjugateGradient<_MatrixType,_UpLo,_Preconditioner> >
   * \endcode
   * 
   * By default the iterations start with x=0 as an initial guess of the solution.
-  * One can control the start using the solveWithGuess() method. Here is a step by
+  * One can control the start using the solveWithGuess() method.
-  * step execution example starting with a random guess and printing the evolution
-  * of the estimated error:
-  * * \code
-  * x = VectorXd::Random(n);
-  * cg.setMaxIterations(1);
-  * int i = 0;
-  * do {
-  *   x = cg.solveWithGuess(b,x);
-  *   std::cout << i << " : " << cg.error() << std::endl;
-  *   ++i;
-  * } while (cg.info()!=Success && i<100);
-  * \endcode
-  * Note that such a step by step excution is slightly slower.
   * 
+  * ConjugateGradient can also be used in a matrix-free context, see the following \link MatrixfreeSolverExample example \endlink.
+  *
   * \sa class SimplicialCholesky, DiagonalPreconditioner, IdentityPreconditioner
   */
 template< typename _MatrixType, int _UpLo, typename _Preconditioner>
@@ -189,7 +178,8 @@ public:
     * this class becomes invalid. Call compute() to update it with the new
     * matrix A, or modify a copy of A.
     */
-  ConjugateGradient(const MatrixType& A) : Base(A) {}
+  template<typename MatrixDerived>
+  explicit ConjugateGradient(const EigenBase<MatrixDerived>& A) : Base(A.derived()) {}
 
   ~ConjugateGradient() {}
   
@@ -213,6 +203,10 @@ public:
   template<typename Rhs,typename Dest>
   void _solveWithGuess(const Rhs& b, Dest& x) const
   {
+    typedef typename internal::conditional<UpLo==(Lower|Upper),
+                                           const MatrixType&,
+                                           SparseSelfAdjointView<const MatrixType, UpLo>
+                                          >::type MatrixWrapperType;
     m_iterations = Base::maxIterations();
     m_error = Base::m_tolerance;
 
@@ -222,8 +216,7 @@ public:
       m_error = Base::m_tolerance;
 
       typename Dest::ColXpr xj(x,j);
-      internal::conjugate_gradient(mp_matrix->template selfadjointView<UpLo>(), b.col(j), xj,
+      internal::conjugate_gradient(MatrixWrapperType(*mp_matrix), b.col(j), xj, Base::m_preconditioner, m_iterations, m_error);
-                                   Base::m_preconditioner, m_iterations, m_error);
     }
 
     m_isInitialized = true;
@@ -234,7 +227,7 @@ public:
   template<typename Rhs,typename Dest>
   void _solve(const Rhs& b, Dest& x) const
   {
-    x.setOnes();
+    x.setZero();
     _solveWithGuess(b,x);
   }
 

Eigen/src/IterativeLinearSolvers/IncompleteLUT.h

@@ -150,7 +150,6 @@ class IncompleteLUT : internal::noncopyable
     {
       analyzePattern(amat); 
       factorize(amat);
-      m_isInitialized = m_factorizationIsOk;
       return *this;
     }
 
@@ -160,7 +159,7 @@ class IncompleteLUT : internal::noncopyable
     template<typename Rhs, typename Dest>
     void _solve(const Rhs& b, Dest& x) const
     {
-      x = m_Pinv * b;  
+      x = m_Pinv * b;
       x = m_lu.template triangularView<UnitLower>().solve(x);
       x = m_lu.template triangularView<Upper>().solve(x);
       x = m_P * x; 
@@ -223,18 +222,29 @@ template<typename _MatrixType>
 void IncompleteLUT<Scalar>::analyzePattern(const _MatrixType& amat)
 {
   // Compute the Fill-reducing permutation
+  // Since ILUT does not perform any numerical pivoting,
+  // it is highly preferable to keep the diagonal through symmetric permutations.
+#ifndef EIGEN_MPL2_ONLY
+  // To this end, let's symmetrize the pattern and perform AMD on it.
   SparseMatrix<Scalar,ColMajor, Index> mat1 = amat;
   SparseMatrix<Scalar,ColMajor, Index> mat2 = amat.transpose();
-  // Symmetrize the pattern
   // FIXME for a matrix with nearly symmetric pattern, mat2+mat1 is the appropriate choice.
   //       on the other hand for a really non-symmetric pattern, mat2*mat1 should be prefered...
   SparseMatrix<Scalar,ColMajor, Index> AtA = mat2 + mat1;
-  AtA.prune(keep_diag());
+  AMDOrdering<Index> ordering;
-  internal::minimum_degree_ordering<Scalar, Index>(AtA, m_P);  // Then compute the AMD ordering...
+  ordering(AtA,m_P);
-
+  m_Pinv  = m_P.inverse(); // cache the inverse permutation
-  m_Pinv  = m_P.inverse(); // ... and the inverse permutation
+#else
+  // If AMD is not available, (MPL2-only), then let's use the slower COLAMD routine.
+  SparseMatrix<Scalar,ColMajor, Index> mat1 = amat;
+  COLAMDOrdering<Index> ordering;
+  ordering(mat1,m_Pinv);
+  m_P = m_Pinv.inverse();
+#endif
 
   m_analysisIsOk = true;
+  m_factorizationIsOk = false;
+  m_isInitialized = false;
 }
 
 template <typename Scalar>
@@ -442,6 +452,7 @@ void IncompleteLUT<Scalar>::factorize(const _MatrixType& amat)
   m_lu.makeCompressed();
 
   m_factorizationIsOk = true;
+  m_isInitialized = m_factorizationIsOk;
   m_info = Success;
 }
 

Eigen/src/IterativeLinearSolvers/IterativeSolverBase.h

@@ -49,10 +49,11 @@ public:
     * this class becomes invalid. Call compute() to update it with the new
     * matrix A, or modify a copy of A.
     */
-  IterativeSolverBase(const MatrixType& A)
+  template<typename InputDerived>
+  IterativeSolverBase(const EigenBase<InputDerived>& A)
   {
     init();
-    compute(A);
+    compute(A.derived());
   }
 
   ~IterativeSolverBase() {}
@@ -62,9 +63,11 @@ public:
     * Currently, this function mostly call analyzePattern on the preconditioner. In the future
     * we might, for instance, implement column reodering for faster matrix vector products.
     */
-  Derived& analyzePattern(const MatrixType& A)
+  template<typename InputDerived>
+  Derived& analyzePattern(const EigenBase<InputDerived>& A)
   {
-    m_preconditioner.analyzePattern(A);
+    grabInput(A.derived());
+    m_preconditioner.analyzePattern(*mp_matrix);
     m_isInitialized = true;
     m_analysisIsOk = true;
     m_info = Success;
@@ -80,11 +83,12 @@ public:
     * this class becomes invalid. Call compute() to update it with the new
     * matrix A, or modify a copy of A.
     */
-  Derived& factorize(const MatrixType& A)
+  template<typename InputDerived>
+  Derived& factorize(const EigenBase<InputDerived>& A)
   {
+    grabInput(A.derived());
     eigen_assert(m_analysisIsOk && "You must first call analyzePattern()"); 
-    mp_matrix = &A;
+    m_preconditioner.factorize(*mp_matrix);
-    m_preconditioner.factorize(A);
     m_factorizationIsOk = true;
     m_info = Success;
     return derived();
@@ -100,10 +104,11 @@ public:
     * this class becomes invalid. Call compute() to update it with the new
     * matrix A, or modify a copy of A.
     */
-  Derived& compute(const MatrixType& A)
+  template<typename InputDerived>
+  Derived& compute(const EigenBase<InputDerived>& A)
   {
-    mp_matrix = &A;
+    grabInput(A.derived());
-    m_preconditioner.compute(A);
+    m_preconditioner.compute(*mp_matrix);
     m_isInitialized = true;
     m_analysisIsOk = true;
     m_factorizationIsOk = true;
@@ -212,6 +217,28 @@ public:
   }
 
 protected:
+
+  template<typename InputDerived>
+  void grabInput(const EigenBase<InputDerived>& A)
+  {
+    // we const cast to prevent the creation of a MatrixType temporary by the compiler.
+    grabInput_impl(A.const_cast_derived());
+  }
+
+  template<typename InputDerived>
+  void grabInput_impl(const EigenBase<InputDerived>& A)
+  {
+    m_copyMatrix = A;
+    mp_matrix = &m_copyMatrix;
+  }
+
+  void grabInput_impl(MatrixType& A)
+  {
+    if(MatrixType::RowsAtCompileTime==Dynamic && MatrixType::ColsAtCompileTime==Dynamic)
+      m_copyMatrix.resize(0,0);
+    mp_matrix = &A;
+  }
+
   void init()
   {
     m_isInitialized = false;
@@ -220,6 +247,7 @@ protected:
     m_maxIterations = -1;
     m_tolerance = NumTraits<Scalar>::epsilon();
   }
+  MatrixType m_copyMatrix;
   const MatrixType* mp_matrix;
   Preconditioner m_preconditioner;
 

Eigen/src/LU/FullPivLU.h

@@ -374,6 +374,12 @@ template<typename _MatrixType> class FullPivLU
     inline Index cols() const { return m_lu.cols(); }
 
   protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+    
     MatrixType m_lu;
     PermutationPType m_p;
     PermutationQType m_q;
@@ -418,6 +424,8 @@ FullPivLU<MatrixType>::FullPivLU(const MatrixType& matrix)
 template<typename MatrixType>
 FullPivLU<MatrixType>& FullPivLU<MatrixType>::compute(const MatrixType& matrix)
 {
+  check_template_parameters();
+  
   // the permutations are stored as int indices, so just to be sure:
   eigen_assert(matrix.rows()<=NumTraits<int>::highest() && matrix.cols()<=NumTraits<int>::highest());
   
@@ -680,7 +688,7 @@ struct solve_retval<FullPivLU<_MatrixType>, Rhs>
      */
 
     const Index rows = dec().rows(), cols = dec().cols(),
-              nonzero_pivots = dec().nonzeroPivots();
+              nonzero_pivots = dec().rank();
     eigen_assert(rhs().rows() == rows);
     const Index smalldim = (std::min)(rows, cols);
 

Eigen/src/LU/Inverse.h

@@ -290,7 +290,7 @@ struct inverse_impl : public ReturnByValue<inverse_impl<MatrixType> >
   {
     const int Size = EIGEN_PLAIN_ENUM_MIN(MatrixType::ColsAtCompileTime,Dest::ColsAtCompileTime);
     EIGEN_ONLY_USED_FOR_DEBUG(Size);
-    eigen_assert(( (Size<=1) || (Size>4) || (extract_data(m_matrix)!=extract_data(dst)))
+    eigen_assert(( (Size<=1) || (Size>4) || (extract_data(m_matrix)!=0 && extract_data(m_matrix)!=extract_data(dst)))
               && "Aliasing problem detected in inverse(), you need to do inverse().eval() here.");
 
     compute_inverse<MatrixTypeNestedCleaned, Dest>::run(m_matrix, dst);

Eigen/src/LU/PartialPivLU.h

@@ -171,6 +171,12 @@ template<typename _MatrixType> class PartialPivLU
     inline Index cols() const { return m_lu.cols(); }
 
   protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+    
     MatrixType m_lu;
     PermutationType m_p;
     TranspositionType m_rowsTranspositions;
@@ -386,6 +392,8 @@ void partial_lu_inplace(MatrixType& lu, TranspositionType& row_transpositions, t
 template<typename MatrixType>
 PartialPivLU<MatrixType>& PartialPivLU<MatrixType>::compute(const MatrixType& matrix)
 {
+  check_template_parameters();
+  
   // the row permutation is stored as int indices, so just to be sure:
   eigen_assert(matrix.rows()<NumTraits<int>::highest());
   

Eigen/src/OrderingMethods/Amd.h

@@ -8,7 +8,7 @@
 NOTE: this routine has been adapted from the CSparse library:
 
 Copyright (c) 2006, Timothy A. Davis.
-http://www.cise.ufl.edu/research/sparse/CSparse
+http://www.suitesparse.com
 
 CSparse is free software; you can redistribute it and/or
 modify it under the terms of the GNU Lesser General Public
@@ -137,22 +137,27 @@ void minimum_degree_ordering(SparseMatrix<Scalar,ColMajor,Index>& C, Permutation
     degree[i] = len[i];                 // degree of node i
   }
   mark = internal::cs_wclear<Index>(0, 0, w, n);         /* clear w */
-  elen[n] = -2;                         /* n is a dead element */
-  Cp[n] = -1;                           /* n is a root of assembly tree */
-  w[n] = 0;                             /* n is a dead element */
   
   /* --- Initialize degree lists ------------------------------------------ */
   for(i = 0; i < n; i++)
   {
+    bool has_diag = false;
+    for(p = Cp[i]; p<Cp[i+1]; ++p)
+      if(Ci[p]==i)
+      {
+        has_diag = true;
+        break;
+      }
+   
     d = degree[i];
-    if(d == 0)                         /* node i is empty */
+    if(d == 1 && has_diag)           /* node i is empty */
     {
       elen[i] = -2;                 /* element i is dead */
       nel++;
       Cp[i] = -1;                   /* i is a root of assembly tree */
       w[i] = 0;
     }
-    else if(d > dense)                 /* node i is dense */
+    else if(d > dense || !has_diag)  /* node i is dense or has no structural diagonal element */
     {
       nv[i] = 0;                    /* absorb i into element n */
       elen[i] = -1;                 /* node i is dead */
@@ -168,6 +173,10 @@ void minimum_degree_ordering(SparseMatrix<Scalar,ColMajor,Index>& C, Permutation
     }
   }
   
+  elen[n] = -2;                         /* n is a dead element */
+  Cp[n] = -1;                           /* n is a root of assembly tree */
+  w[n] = 0;                             /* n is a dead element */
+  
   while (nel < n)                         /* while (selecting pivots) do */
   {
     /* --- Select node of minimum approximate degree -------------------- */

Eigen/src/OrderingMethods/Eigen_Colamd.h

@@ -41,12 +41,8 @@
 // 
 //   The colamd/symamd library is available at
 // 
-//       http://www.cise.ufl.edu/research/sparse/colamd/
+//       http://www.suitesparse.com
 
-//   This is the http://www.cise.ufl.edu/research/sparse/colamd/colamd.h
-//   file.  It is required by the colamd.c, colamdmex.c, and symamdmex.c
-//   files, and by any C code that calls the routines whose prototypes are
-//   listed below, or that uses the colamd/symamd definitions listed below.
   
 #ifndef EIGEN_COLAMD_H
 #define EIGEN_COLAMD_H
@@ -102,9 +98,6 @@ namespace internal {
 /* === Definitions ========================================================== */
 /* ========================================================================== */
 
-#define COLAMD_MAX(a,b) (((a) > (b)) ? (a) : (b))
-#define COLAMD_MIN(a,b) (((a) < (b)) ? (a) : (b))
-
 #define ONES_COMPLEMENT(r) (-(r)-1)
 
 /* -------------------------------------------------------------------------- */
@@ -516,7 +509,7 @@ static Index init_rows_cols  /* returns true if OK, or false otherwise */
     Col [col].start = p [col] ;
     Col [col].length = p [col+1] - p [col] ;
 
-    if (Col [col].length < 0)
+    if ((Col [col].length) < 0) // extra parentheses to work-around gcc bug 10200
     {
       /* column pointers must be non-decreasing */
       stats [COLAMD_STATUS] = COLAMD_ERROR_col_length_negative ;
@@ -739,8 +732,8 @@ static void init_scoring
 
   /* === Extract knobs ==================================================== */
 
-  dense_row_count = COLAMD_MAX (0, COLAMD_MIN (knobs [COLAMD_DENSE_ROW] * n_col, n_col)) ;
+  dense_row_count = std::max<Index>(0, (std::min)(Index(knobs [COLAMD_DENSE_ROW] * n_col), n_col)) ;
-  dense_col_count = COLAMD_MAX (0, COLAMD_MIN (knobs [COLAMD_DENSE_COL] * n_row, n_row)) ;
+  dense_col_count = std::max<Index>(0, (std::min)(Index(knobs [COLAMD_DENSE_COL] * n_row), n_row)) ;
   COLAMD_DEBUG1 (("colamd: densecount: %d %d\n", dense_row_count, dense_col_count)) ;
   max_deg = 0 ;
   n_col2 = n_col ;
@@ -804,7 +797,7 @@ static void init_scoring
     else
     {
       /* keep track of max degree of remaining rows */
-      max_deg = COLAMD_MAX (max_deg, deg) ;
+      max_deg = (std::max)(max_deg, deg) ;
     }
   }
   COLAMD_DEBUG1 (("colamd: Dense and null rows killed: %d\n", n_row - n_row2)) ;
@@ -842,7 +835,7 @@ static void init_scoring
       /* add row's external degree */
       score += Row [row].shared1.degree - 1 ;
       /* guard against integer overflow */
-      score = COLAMD_MIN (score, n_col) ;
+      score = (std::min)(score, n_col) ;
     }
     /* determine pruned column length */
     col_length = (Index) (new_cp - &A [Col [c].start]) ;
@@ -914,7 +907,7 @@ static void init_scoring
       head [score] = c ;
 
       /* see if this score is less than current min */
-      min_score = COLAMD_MIN (min_score, score) ;
+      min_score = (std::min)(min_score, score) ;
 
 
     }
@@ -1040,7 +1033,7 @@ static Index find_ordering /* return the number of garbage collections */
 
     /* === Garbage_collection, if necessary ============================= */
 
-    needed_memory = COLAMD_MIN (pivot_col_score, n_col - k) ;
+    needed_memory = (std::min)(pivot_col_score, n_col - k) ;
     if (pfree + needed_memory >= Alen)
     {
       pfree = Eigen::internal::garbage_collection (n_row, n_col, Row, Col, A, &A [pfree]) ;
@@ -1099,7 +1092,7 @@ static Index find_ordering /* return the number of garbage collections */
 
     /* clear tag on pivot column */
     Col [pivot_col].shared1.thickness = pivot_col_thickness ;
-    max_deg = COLAMD_MAX (max_deg, pivot_row_degree) ;
+    max_deg = (std::max)(max_deg, pivot_row_degree) ;
 
 
     /* === Kill all rows used to construct pivot row ==================== */
@@ -1273,7 +1266,7 @@ static Index find_ordering /* return the number of garbage collections */
 	/* add set difference */
 	cur_score += row_mark - tag_mark ;
 	/* integer overflow... */
-	cur_score = COLAMD_MIN (cur_score, n_col) ;
+	cur_score = (std::min)(cur_score, n_col) ;
       }
 
       /* recompute the column's length */
@@ -1386,7 +1379,7 @@ static Index find_ordering /* return the number of garbage collections */
       cur_score -= Col [col].shared1.thickness ;
 
       /* make sure score is less or equal than the max score */
-      cur_score = COLAMD_MIN (cur_score, max_score) ;
+      cur_score = (std::min)(cur_score, max_score) ;
       COLAMD_ASSERT (cur_score >= 0) ;
 
       /* store updated score */
@@ -1409,7 +1402,7 @@ static Index find_ordering /* return the number of garbage collections */
       head [cur_score] = col ;
 
       /* see if this score is less than current min */
-      min_score = COLAMD_MIN (min_score, cur_score) ;
+      min_score = (std::min)(min_score, cur_score) ;
 
     }
 

Eigen/src/PaStiXSupport/PaStiXSupport.h

@@ -10,6 +10,14 @@
 #ifndef EIGEN_PASTIXSUPPORT_H
 #define EIGEN_PASTIXSUPPORT_H
 
+#if defined(DCOMPLEX)
+  #define PASTIX_COMPLEX  COMPLEX
+  #define PASTIX_DCOMPLEX DCOMPLEX
+#else
+  #define PASTIX_COMPLEX  std::complex<float>
+  #define PASTIX_DCOMPLEX std::complex<double>
+#endif
+
 namespace Eigen { 
 
 /** \ingroup PaStiXSupport_Module
@@ -74,14 +82,14 @@ namespace internal
   {
     if (n == 0) { ptr = NULL; idx = NULL; vals = NULL; }
     if (nbrhs == 0) {x = NULL; nbrhs=1;}
-    c_pastix(pastix_data, pastix_comm, n, ptr, idx, reinterpret_cast<COMPLEX*>(vals), perm, invp, reinterpret_cast<COMPLEX*>(x), nbrhs, iparm, dparm); 
+    c_pastix(pastix_data, pastix_comm, n, ptr, idx, reinterpret_cast<PASTIX_COMPLEX*>(vals), perm, invp, reinterpret_cast<PASTIX_COMPLEX*>(x), nbrhs, iparm, dparm);
   }
   
   void eigen_pastix(pastix_data_t **pastix_data, int pastix_comm, int n, int *ptr, int *idx, std::complex<double> *vals, int *perm, int * invp, std::complex<double> *x, int nbrhs, int *iparm, double *dparm)
   {
     if (n == 0) { ptr = NULL; idx = NULL; vals = NULL; }
     if (nbrhs == 0) {x = NULL; nbrhs=1;}
-    z_pastix(pastix_data, pastix_comm, n, ptr, idx, reinterpret_cast<DCOMPLEX*>(vals), perm, invp, reinterpret_cast<DCOMPLEX*>(x), nbrhs, iparm, dparm); 
+    z_pastix(pastix_data, pastix_comm, n, ptr, idx, reinterpret_cast<PASTIX_DCOMPLEX*>(vals), perm, invp, reinterpret_cast<PASTIX_DCOMPLEX*>(x), nbrhs, iparm, dparm);
   }
 
   // Convert the matrix  to Fortran-style Numbering

Eigen/src/PardisoSupport/PardisoSupport.h

@@ -221,11 +221,11 @@ class PardisoImpl
       m_type = type;
       bool symmetric = std::abs(m_type) < 10;
       m_iparm[0] = 1;   // No solver default
-      m_iparm[1] = 3;   // use Metis for the ordering
+      m_iparm[1] = 2;   // use Metis for the ordering
-      m_iparm[2] = 1;   // Numbers of processors, value of OMP_NUM_THREADS
+      m_iparm[2] = 0;   // Reserved. Set to zero. (??Numbers of processors, value of OMP_NUM_THREADS??)
       m_iparm[3] = 0;   // No iterative-direct algorithm
       m_iparm[4] = 0;   // No user fill-in reducing permutation
-      m_iparm[5] = 0;   // Write solution into x
+      m_iparm[5] = 0;   // Write solution into x, b is left unchanged
       m_iparm[6] = 0;   // Not in use
       m_iparm[7] = 2;   // Max numbers of iterative refinement steps
       m_iparm[8] = 0;   // Not in use
@@ -246,7 +246,10 @@ class PardisoImpl
       m_iparm[26] = 0;  // No matrix checker
       m_iparm[27] = (sizeof(RealScalar) == 4) ? 1 : 0;
       m_iparm[34] = 1;  // C indexing
-      m_iparm[59] = 1;  // Automatic switch between In-Core and Out-of-Core modes
+      m_iparm[36] = 0;  // CSR
+      m_iparm[59] = 0;  // 0 - In-Core ; 1 - Automatic switch between In-Core and Out-of-Core modes ; 2 - Out-of-Core
+      
+      memset(m_pt, 0, sizeof(m_pt));
     }
 
   protected:
@@ -384,7 +387,6 @@ bool PardisoImpl<Base>::_solve(const MatrixBase<BDerived> &b, MatrixBase<XDerive
                                                      m_matrix.valuePtr(), m_matrix.outerIndexPtr(), m_matrix.innerIndexPtr(),
                                                      m_perm.data(), nrhs, m_iparm.data(), m_msglvl,
                                                      rhs_ptr, x.derived().data());
-
   return error==0;
 }
 
@@ -397,6 +399,9 @@ bool PardisoImpl<Base>::_solve(const MatrixBase<BDerived> &b, MatrixBase<XDerive
   * using the Intel MKL PARDISO library. The sparse matrix A must be squared and invertible.
   * The vectors or matrices X and B can be either dense or sparse.
   *
+  * By default, it runs in in-core mode. To enable PARDISO's out-of-core feature, set:
+  * \code solver.pardisoParameterArray()[59] = 1; \endcode
+  *
   * \tparam _MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
   *
   * \sa \ref TutorialSparseDirectSolvers
@@ -447,6 +452,9 @@ class PardisoLU : public PardisoImpl< PardisoLU<MatrixType> >
   * using the Intel MKL PARDISO library. The sparse matrix A must be selfajoint and positive definite.
   * The vectors or matrices X and B can be either dense or sparse.
   *
+  * By default, it runs in in-core mode. To enable PARDISO's out-of-core feature, set:
+  * \code solver.pardisoParameterArray()[59] = 1; \endcode
+  *
   * \tparam MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
   * \tparam UpLo can be any bitwise combination of Upper, Lower. The default is Upper, meaning only the upper triangular part has to be used.
   *         Upper|Lower can be used to tell both triangular parts can be used as input.
@@ -507,6 +515,9 @@ class PardisoLLT : public PardisoImpl< PardisoLLT<MatrixType,_UpLo> >
   * For complex matrices, A can also be symmetric only, see the \a Options template parameter.
   * The vectors or matrices X and B can be either dense or sparse.
   *
+  * By default, it runs in in-core mode. To enable PARDISO's out-of-core feature, set:
+  * \code solver.pardisoParameterArray()[59] = 1; \endcode
+  *
   * \tparam MatrixType the type of the sparse matrix A, it must be a SparseMatrix<>
   * \tparam Options can be any bitwise combination of Upper, Lower, and Symmetric. The default is Upper, meaning only the upper triangular part has to be used.
   *         Symmetric can be used for symmetric, non-selfadjoint complex matrices, the default being to assume a selfadjoint matrix.

Eigen/src/QR/ColPivHouseholderQR.h

@@ -384,6 +384,12 @@ template<typename _MatrixType> class ColPivHouseholderQR
     }
 
   protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+    
     MatrixType m_qr;
     HCoeffsType m_hCoeffs;
     PermutationType m_colsPermutation;
@@ -422,6 +428,8 @@ typename MatrixType::RealScalar ColPivHouseholderQR<MatrixType>::logAbsDetermina
 template<typename MatrixType>
 ColPivHouseholderQR<MatrixType>& ColPivHouseholderQR<MatrixType>::compute(const MatrixType& matrix)
 {
+  check_template_parameters();
+  
   using std::abs;
   Index rows = matrix.rows();
   Index cols = matrix.cols();
@@ -463,20 +471,10 @@ ColPivHouseholderQR<MatrixType>& ColPivHouseholderQR<MatrixType>::compute(const
     // we store that back into our table: it can't hurt to correct our table.
     m_colSqNorms.coeffRef(biggest_col_index) = biggest_col_sq_norm;
 
-    // if the current biggest column is smaller than epsilon times the initial biggest column,
+    // Track the number of meaningful pivots but do not stop the decomposition to make
-    // terminate to avoid generating nan/inf values.
+    // sure that the initial matrix is properly reproduced. See bug 941.
-    // Note that here, if we test instead for "biggest == 0", we get a failure every 1000 (or so)
+    if(m_nonzero_pivots==size && biggest_col_sq_norm < threshold_helper * RealScalar(rows-k))
-    // repetitions of the unit test, with the result of solve() filled with large values of the order
-    // of 1/(size*epsilon).
-    if(biggest_col_sq_norm < threshold_helper * RealScalar(rows-k))
-    {
       m_nonzero_pivots = k;
-      m_hCoeffs.tail(size-k).setZero();
-      m_qr.bottomRightCorner(rows-k,cols-k)
-          .template triangularView<StrictlyLower>()
-          .setZero();
-      break;
-    }
 
     // apply the transposition to the columns
     m_colsTranspositions.coeffRef(k) = biggest_col_index;
@@ -505,7 +503,7 @@ ColPivHouseholderQR<MatrixType>& ColPivHouseholderQR<MatrixType>::compute(const
   }
 
   m_colsPermutation.setIdentity(PermIndexType(cols));
-  for(PermIndexType k = 0; k < m_nonzero_pivots; ++k)
+  for(PermIndexType k = 0; k < size/*m_nonzero_pivots*/; ++k)
     m_colsPermutation.applyTranspositionOnTheRight(k, PermIndexType(m_colsTranspositions.coeff(k)));
 
   m_det_pq = (number_of_transpositions%2) ? -1 : 1;
@@ -555,13 +553,15 @@ struct solve_retval<ColPivHouseholderQR<_MatrixType>, Rhs>
 
 } // end namespace internal
 
-/** \returns the matrix Q as a sequence of householder transformations */
+/** \returns the matrix Q as a sequence of householder transformations.
+  * You can extract the meaningful part only by using:
+  * \code qr.householderQ().setLength(qr.nonzeroPivots()) \endcode*/
 template<typename MatrixType>
 typename ColPivHouseholderQR<MatrixType>::HouseholderSequenceType ColPivHouseholderQR<MatrixType>
   ::householderQ() const
 {
   eigen_assert(m_isInitialized && "ColPivHouseholderQR is not initialized.");
-  return HouseholderSequenceType(m_qr, m_hCoeffs.conjugate()).setLength(m_nonzero_pivots);
+  return HouseholderSequenceType(m_qr, m_hCoeffs.conjugate());
 }
 
 /** \return the column-pivoting Householder QR decomposition of \c *this.

Eigen/src/QR/ColPivHouseholderQR_MKL.h

@@ -49,7 +49,6 @@ ColPivHouseholderQR<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynami
 { \
   using std::abs; \
   typedef Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic> MatrixType; \
-  typedef MatrixType::Scalar Scalar; \
   typedef MatrixType::RealScalar RealScalar; \
   Index rows = matrix.rows();\
   Index cols = matrix.cols();\

Eigen/src/QR/FullPivHouseholderQR.h

@@ -368,6 +368,12 @@ template<typename _MatrixType> class FullPivHouseholderQR
     RealScalar maxPivot() const { return m_maxpivot; }
 
   protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+    
     MatrixType m_qr;
     HCoeffsType m_hCoeffs;
     IntDiagSizeVectorType m_rows_transpositions;
@@ -407,6 +413,8 @@ typename MatrixType::RealScalar FullPivHouseholderQR<MatrixType>::logAbsDetermin
 template<typename MatrixType>
 FullPivHouseholderQR<MatrixType>& FullPivHouseholderQR<MatrixType>::compute(const MatrixType& matrix)
 {
+  check_template_parameters();
+  
   using std::abs;
   Index rows = matrix.rows();
   Index cols = matrix.cols();

Eigen/src/QR/HouseholderQR.h

@@ -189,6 +189,12 @@ template<typename _MatrixType> class HouseholderQR
     const HCoeffsType& hCoeffs() const { return m_hCoeffs; }
 
   protected:
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
+    
     MatrixType m_qr;
     HCoeffsType m_hCoeffs;
     RowVectorType m_temp;
@@ -251,56 +257,62 @@ void householder_qr_inplace_unblocked(MatrixQR& mat, HCoeffs& hCoeffs, typename
 }
 
 /** \internal */
-template<typename MatrixQR, typename HCoeffs>
+template<typename MatrixQR, typename HCoeffs,
-void householder_qr_inplace_blocked(MatrixQR& mat, HCoeffs& hCoeffs,
+  typename MatrixQRScalar = typename MatrixQR::Scalar,
-                                       typename MatrixQR::Index maxBlockSize=32,
+  bool InnerStrideIsOne = (MatrixQR::InnerStrideAtCompileTime == 1 && HCoeffs::InnerStrideAtCompileTime == 1)>
-                                       typename MatrixQR::Scalar* tempData = 0)
+struct householder_qr_inplace_blocked
 {
-  typedef typename MatrixQR::Index Index;
+  // This is specialized for MKL-supported Scalar types in HouseholderQR_MKL.h
-  typedef typename MatrixQR::Scalar Scalar;
+  static void run(MatrixQR& mat, HCoeffs& hCoeffs,
-  typedef Block<MatrixQR,Dynamic,Dynamic> BlockType;
+      typename MatrixQR::Index maxBlockSize=32,
-
+      typename MatrixQR::Scalar* tempData = 0)
-  Index rows = mat.rows();
-  Index cols = mat.cols();
-  Index size = (std::min)(rows, cols);
-
-  typedef Matrix<Scalar,Dynamic,1,ColMajor,MatrixQR::MaxColsAtCompileTime,1> TempType;
-  TempType tempVector;
-  if(tempData==0)
   {
-    tempVector.resize(cols);
+    typedef typename MatrixQR::Index Index;
-    tempData = tempVector.data();
+    typedef typename MatrixQR::Scalar Scalar;
-  }
+    typedef Block<MatrixQR,Dynamic,Dynamic> BlockType;
 
-  Index blockSize = (std::min)(maxBlockSize,size);
+    Index rows = mat.rows();
+    Index cols = mat.cols();
+    Index size = (std::min)(rows, cols);
 
-  Index k = 0;
+    typedef Matrix<Scalar,Dynamic,1,ColMajor,MatrixQR::MaxColsAtCompileTime,1> TempType;
-  for (k = 0; k < size; k += blockSize)
+    TempType tempVector;
-  {
+    if(tempData==0)
-    Index bs = (std::min)(size-k,blockSize);  // actual size of the block
-    Index tcols = cols - k - bs;            // trailing columns
-    Index brows = rows-k;                   // rows of the block
-
-    // partition the matrix:
-    //        A00 | A01 | A02
-    // mat  = A10 | A11 | A12
-    //        A20 | A21 | A22
-    // and performs the qr dec of [A11^T A12^T]^T
-    // and update [A21^T A22^T]^T using level 3 operations.
-    // Finally, the algorithm continue on A22
-
-    BlockType A11_21 = mat.block(k,k,brows,bs);
-    Block<HCoeffs,Dynamic,1> hCoeffsSegment = hCoeffs.segment(k,bs);
-
-    householder_qr_inplace_unblocked(A11_21, hCoeffsSegment, tempData);
-
-    if(tcols)
     {
-      BlockType A21_22 = mat.block(k,k+bs,brows,tcols);
+      tempVector.resize(cols);
-      apply_block_householder_on_the_left(A21_22,A11_21,hCoeffsSegment.adjoint());
+      tempData = tempVector.data();
+    }
+
+    Index blockSize = (std::min)(maxBlockSize,size);
+
+    Index k = 0;
+    for (k = 0; k < size; k += blockSize)
+    {
+      Index bs = (std::min)(size-k,blockSize);  // actual size of the block
+      Index tcols = cols - k - bs;            // trailing columns
+      Index brows = rows-k;                   // rows of the block
+
+      // partition the matrix:
+      //        A00 | A01 | A02
+      // mat  = A10 | A11 | A12
+      //        A20 | A21 | A22
+      // and performs the qr dec of [A11^T A12^T]^T
+      // and update [A21^T A22^T]^T using level 3 operations.
+      // Finally, the algorithm continue on A22
+
+      BlockType A11_21 = mat.block(k,k,brows,bs);
+      Block<HCoeffs,Dynamic,1> hCoeffsSegment = hCoeffs.segment(k,bs);
+
+      householder_qr_inplace_unblocked(A11_21, hCoeffsSegment, tempData);
+
+      if(tcols)
+      {
+        BlockType A21_22 = mat.block(k,k+bs,brows,tcols);
+        apply_block_householder_on_the_left(A21_22,A11_21,hCoeffsSegment.adjoint());
+      }
     }
   }
-}
+};
 
 template<typename _MatrixType, typename Rhs>
 struct solve_retval<HouseholderQR<_MatrixType>, Rhs>
@@ -343,6 +355,8 @@ struct solve_retval<HouseholderQR<_MatrixType>, Rhs>
 template<typename MatrixType>
 HouseholderQR<MatrixType>& HouseholderQR<MatrixType>::compute(const MatrixType& matrix)
 {
+  check_template_parameters();
+  
   Index rows = matrix.rows();
   Index cols = matrix.cols();
   Index size = (std::min)(rows,cols);
@@ -352,7 +366,7 @@ HouseholderQR<MatrixType>& HouseholderQR<MatrixType>::compute(const MatrixType&
 
   m_temp.resize(cols);
 
-  internal::householder_qr_inplace_blocked(m_qr, m_hCoeffs, 48, m_temp.data());
+  internal::householder_qr_inplace_blocked<MatrixType, HCoeffsType>::run(m_qr, m_hCoeffs, 48, m_temp.data());
 
   m_isInitialized = true;
   return *this;

Eigen/src/QR/HouseholderQR_MKL.h

@@ -34,28 +34,30 @@
 #ifndef EIGEN_QR_MKL_H
 #define EIGEN_QR_MKL_H
 
-#include "Eigen/src/Core/util/MKL_support.h"
+#include "../Core/util/MKL_support.h"
 
 namespace Eigen { 
 
-namespace internal {
+  namespace internal {
 
-/** \internal Specialization for the data types supported by MKL */
+    /** \internal Specialization for the data types supported by MKL */
 
 #define EIGEN_MKL_QR_NOPIV(EIGTYPE, MKLTYPE, MKLPREFIX) \
 template<typename MatrixQR, typename HCoeffs> \
-void householder_qr_inplace_blocked(MatrixQR& mat, HCoeffs& hCoeffs, \
+struct householder_qr_inplace_blocked<MatrixQR, HCoeffs, EIGTYPE, true> \
-                                       typename MatrixQR::Index maxBlockSize=32, \
-                                       EIGTYPE* tempData = 0) \
 { \
-  lapack_int m = mat.rows(); \
+  static void run(MatrixQR& mat, HCoeffs& hCoeffs, \
-  lapack_int n = mat.cols(); \
+      typename MatrixQR::Index = 32, \
-  lapack_int lda = mat.outerStride(); \
+      typename MatrixQR::Scalar* = 0) \
-  lapack_int matrix_order = (MatrixQR::IsRowMajor) ? LAPACK_ROW_MAJOR : LAPACK_COL_MAJOR; \
+  { \
-  LAPACKE_##MKLPREFIX##geqrf( matrix_order, m, n, (MKLTYPE*)mat.data(), lda, (MKLTYPE*)hCoeffs.data()); \
+    lapack_int m = (lapack_int) mat.rows(); \
-  hCoeffs.adjointInPlace(); \
+    lapack_int n = (lapack_int) mat.cols(); \
-\
+    lapack_int lda = (lapack_int) mat.outerStride(); \
-}
+    lapack_int matrix_order = (MatrixQR::IsRowMajor) ? LAPACK_ROW_MAJOR : LAPACK_COL_MAJOR; \
+    LAPACKE_##MKLPREFIX##geqrf( matrix_order, m, n, (MKLTYPE*)mat.data(), lda, (MKLTYPE*)hCoeffs.data()); \
+    hCoeffs.adjointInPlace(); \
+  } \
+};
 
 EIGEN_MKL_QR_NOPIV(double, double, d)
 EIGEN_MKL_QR_NOPIV(float, float, s)

Eigen/src/SPQRSupport/SuiteSparseQRSupport.h

@@ -47,7 +47,7 @@ namespace Eigen {
  * You can then apply it to a vector.
  * 
  * R is the sparse triangular factor. Use matrixQR() to get it as SparseMatrix.
- * NOTE : The Index type of R is always UF_long. You can get it with SPQR::Index
+ * NOTE : The Index type of R is always SuiteSparse_long. You can get it with SPQR::Index
  * 
  * \tparam _MatrixType The type of the sparse matrix A, must be a column-major SparseMatrix<>
  * NOTE 
@@ -59,24 +59,18 @@ class SPQR
   public:
     typedef typename _MatrixType::Scalar Scalar;
     typedef typename _MatrixType::RealScalar RealScalar;
-    typedef UF_long Index ; 
+    typedef SuiteSparse_long Index ;
     typedef SparseMatrix<Scalar, ColMajor, Index> MatrixType;
     typedef PermutationMatrix<Dynamic, Dynamic> PermutationType;
   public:
     SPQR() 
-      : m_isInitialized(false),
+      : m_isInitialized(false), m_ordering(SPQR_ORDERING_DEFAULT), m_allow_tol(SPQR_DEFAULT_TOL), m_tolerance (NumTraits<Scalar>::epsilon()), m_useDefaultThreshold(true)
-      m_ordering(SPQR_ORDERING_DEFAULT),
-      m_allow_tol(SPQR_DEFAULT_TOL),
-      m_tolerance (NumTraits<Scalar>::epsilon())
     { 
       cholmod_l_start(&m_cc);
     }
     
-    SPQR(const _MatrixType& matrix) 
+    SPQR(const _MatrixType& matrix)
-    : m_isInitialized(false),
+    : m_isInitialized(false), m_ordering(SPQR_ORDERING_DEFAULT), m_allow_tol(SPQR_DEFAULT_TOL), m_tolerance (NumTraits<Scalar>::epsilon()), m_useDefaultThreshold(true)
-      m_ordering(SPQR_ORDERING_DEFAULT),
-      m_allow_tol(SPQR_DEFAULT_TOL),
-      m_tolerance (NumTraits<Scalar>::epsilon())
     {
       cholmod_l_start(&m_cc);
       compute(matrix);
@@ -101,10 +95,26 @@ class SPQR
       if(m_isInitialized) SPQR_free();
 
       MatrixType mat(matrix);
+      
+      /* Compute the default threshold as in MatLab, see:
+       * Tim Davis, "Algorithm 915, SuiteSparseQR: Multifrontal Multithreaded Rank-Revealing
+       * Sparse QR Factorization, ACM Trans. on Math. Soft. 38(1), 2011, Page 8:3 
+       */
+      RealScalar pivotThreshold = m_tolerance;
+      if(m_useDefaultThreshold) 
+      {
+        using std::max;
+        RealScalar max2Norm = 0.0;
+        for (int j = 0; j < mat.cols(); j++) max2Norm = (max)(max2Norm, mat.col(j).norm());
+        if(max2Norm==RealScalar(0))
+          max2Norm = RealScalar(1);
+        pivotThreshold = 20 * (mat.rows() + mat.cols()) * max2Norm * NumTraits<RealScalar>::epsilon();
+      }
+      
       cholmod_sparse A; 
       A = viewAsCholmod(mat);
       Index col = matrix.cols();
-      m_rank = SuiteSparseQR<Scalar>(m_ordering, m_tolerance, col, &A, 
+      m_rank = SuiteSparseQR<Scalar>(m_ordering, pivotThreshold, col, &A, 
                              &m_cR, &m_E, &m_H, &m_HPinv, &m_HTau, &m_cc);
 
       if (!m_cR)
@@ -120,7 +130,7 @@ class SPQR
     /** 
      * Get the number of rows of the input matrix and the Q matrix
      */
-    inline Index rows() const {return m_H->nrow; }
+    inline Index rows() const {return m_cR->nrow; }
     
     /** 
      * Get the number of columns of the input matrix. 
@@ -145,16 +155,25 @@ class SPQR
     {
       eigen_assert(m_isInitialized && " The QR factorization should be computed first, call compute()");
       eigen_assert(b.cols()==1 && "This method is for vectors only");
-      
+
       //Compute Q^T * b
-      typename Dest::PlainObject y;
+      typename Dest::PlainObject y, y2;
       y = matrixQ().transpose() * b;
-        // Solves with the triangular matrix R
+      
+      // Solves with the triangular matrix R
       Index rk = this->rank();
-      y.topRows(rk) = this->matrixR().topLeftCorner(rk, rk).template triangularView<Upper>().solve(y.topRows(rk));
+      y2 = y;
-      y.bottomRows(cols()-rk).setZero();
+      y.resize((std::max)(cols(),Index(y.rows())),y.cols());
+      y.topRows(rk) = this->matrixR().topLeftCorner(rk, rk).template triangularView<Upper>().solve(y2.topRows(rk));
+
       // Apply the column permutation 
-      dest.topRows(cols()) = colsPermutation() * y.topRows(cols());
+      // colsPermutation() performs a copy of the permutation,
+      // so let's apply it manually:
+      for(Index i = 0; i < rk; ++i) dest.row(m_E[i]) = y.row(i);
+      for(Index i = rk; i < cols(); ++i) dest.row(m_E[i]).setZero();
+      
+//       y.bottomRows(y.rows()-rk).setZero();
+//       dest = colsPermutation() * y.topRows(cols());
       
       m_info = Success;
     }
@@ -197,7 +216,11 @@ class SPQR
     /// Set the fill-reducing ordering method to be used
     void setSPQROrdering(int ord) { m_ordering = ord;}
     /// Set the tolerance tol to treat columns with 2-norm < =tol as zero
-    void setPivotThreshold(const RealScalar& tol) { m_tolerance = tol; }
+    void setPivotThreshold(const RealScalar& tol)
+    {
+      m_useDefaultThreshold = false;
+      m_tolerance = tol;
+    }
     
     /** \returns a pointer to the SPQR workspace */
     cholmod_common *cholmodCommon() const { return &m_cc; }
@@ -230,6 +253,7 @@ class SPQR
     mutable cholmod_dense *m_HTau; // The Householder coefficients
     mutable Index m_rank; // The rank of the matrix
     mutable cholmod_common m_cc; // Workspace and parameters
+    bool m_useDefaultThreshold;     // Use default threshold
     template<typename ,typename > friend struct SPQR_QProduct;
 };
 

Eigen/src/SVD/JacobiSVD.h

@@ -359,29 +359,42 @@ struct svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner, false
 {
   typedef JacobiSVD<MatrixType, QRPreconditioner> SVD;
   typedef typename SVD::Index Index;
-  static void run(typename SVD::WorkMatrixType&, SVD&, Index, Index) {}
+  typedef typename MatrixType::RealScalar RealScalar;
+  static bool run(typename SVD::WorkMatrixType&, SVD&, Index, Index, RealScalar&) { return true; }
 };
 
 template<typename MatrixType, int QRPreconditioner>
 struct svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner, true>
 {
   typedef JacobiSVD<MatrixType, QRPreconditioner> SVD;
+  typedef typename SVD::Index Index;
   typedef typename MatrixType::Scalar Scalar;
   typedef typename MatrixType::RealScalar RealScalar;
-  typedef typename SVD::Index Index;
+  static bool run(typename SVD::WorkMatrixType& work_matrix, SVD& svd, Index p, Index q, RealScalar& maxDiagEntry)
-  static void run(typename SVD::WorkMatrixType& work_matrix, SVD& svd, Index p, Index q)
   {
     using std::sqrt;
+    using std::abs;
+    using std::max;
     Scalar z;
     JacobiRotation<Scalar> rot;
     RealScalar n = sqrt(numext::abs2(work_matrix.coeff(p,p)) + numext::abs2(work_matrix.coeff(q,p)));
-    
+
+    const RealScalar considerAsZero = (std::numeric_limits<RealScalar>::min)();
+    const RealScalar precision = NumTraits<Scalar>::epsilon();
+
     if(n==0)
     {
-      z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q);
+      // make sure first column is zero
-      work_matrix.row(p) *= z;
+      work_matrix.coeffRef(p,p) = work_matrix.coeffRef(q,p) = Scalar(0);
-      if(svd.computeU()) svd.m_matrixU.col(p) *= conj(z);
+
-      if(work_matrix.coeff(q,q)!=Scalar(0))
+      if(abs(numext::imag(work_matrix.coeff(p,q)))>considerAsZero)
+      {
+        // work_matrix.coeff(p,q) can be zero if work_matrix.coeff(q,p) is not zero but small enough to underflow when computing n
+        z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q);
+        work_matrix.row(p) *= z;
+        if(svd.computeU()) svd.m_matrixU.col(p) *= conj(z);
+      }
+      if(abs(numext::imag(work_matrix.coeff(q,q)))>considerAsZero)
       {
         z = abs(work_matrix.coeff(q,q)) / work_matrix.coeff(q,q);
         work_matrix.row(q) *= z;
@@ -395,19 +408,25 @@ struct svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner, true>
       rot.s() = work_matrix.coeff(q,p) / n;
       work_matrix.applyOnTheLeft(p,q,rot);
       if(svd.computeU()) svd.m_matrixU.applyOnTheRight(p,q,rot.adjoint());
-      if(work_matrix.coeff(p,q) != Scalar(0))
+      if(abs(numext::imag(work_matrix.coeff(p,q)))>considerAsZero)
       {
-        Scalar z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q);
+        z = abs(work_matrix.coeff(p,q)) / work_matrix.coeff(p,q);
         work_matrix.col(q) *= z;
         if(svd.computeV()) svd.m_matrixV.col(q) *= z;
       }
-      if(work_matrix.coeff(q,q) != Scalar(0))
+      if(abs(numext::imag(work_matrix.coeff(q,q)))>considerAsZero)
       {
         z = abs(work_matrix.coeff(q,q)) / work_matrix.coeff(q,q);
         work_matrix.row(q) *= z;
         if(svd.computeU()) svd.m_matrixU.col(q) *= conj(z);
       }
     }
+
+    // update largest diagonal entry
+    maxDiagEntry = max EIGEN_EMPTY (maxDiagEntry,max EIGEN_EMPTY (abs(work_matrix.coeff(p,p)), abs(work_matrix.coeff(q,q))));
+    // and check whether the 2x2 block is already diagonal
+    RealScalar threshold = max EIGEN_EMPTY (considerAsZero, precision * maxDiagEntry);
+    return abs(work_matrix.coeff(p,q))>threshold || abs(work_matrix.coeff(q,p)) > threshold;
   }
 };
 
@@ -424,22 +443,23 @@ void real_2x2_jacobi_svd(const MatrixType& matrix, Index p, Index q,
   JacobiRotation<RealScalar> rot1;
   RealScalar t = m.coeff(0,0) + m.coeff(1,1);
   RealScalar d = m.coeff(1,0) - m.coeff(0,1);
-  if(t == RealScalar(0))
+  if(d == RealScalar(0))
   {
-    rot1.c() = RealScalar(0);
+    rot1.s() = RealScalar(0);
-    rot1.s() = d > RealScalar(0) ? RealScalar(1) : RealScalar(-1);
+    rot1.c() = RealScalar(1);
   }
   else
   {
-    RealScalar t2d2 = numext::hypot(t,d);
+    // If d!=0, then t/d cannot overflow because the magnitude of the
-    rot1.c() = abs(t)/t2d2;
+    // entries forming d are not too small compared to the ones forming t.
-    rot1.s() = d/t2d2;
+    RealScalar u = t / d;
-    if(t<RealScalar(0))
+    RealScalar tmp = sqrt(RealScalar(1) + numext::abs2(u));
-      rot1.s() = -rot1.s();
+    rot1.s() = RealScalar(1) / tmp;
+    rot1.c() = u / tmp;
   }
   m.applyOnTheLeft(0,1,rot1);
   j_right->makeJacobi(m,0,1);
-  *j_left  = rot1 * j_right->transpose();
+  *j_left = rot1 * j_right->transpose();
 }
 
 } // end namespace internal
@@ -742,6 +762,11 @@ template<typename _MatrixType, int QRPreconditioner> class JacobiSVD
 
   private:
     void allocate(Index rows, Index cols, unsigned int computationOptions);
+    
+    static void check_template_parameters()
+    {
+      EIGEN_STATIC_ASSERT_NON_INTEGER(Scalar);
+    }
 
   protected:
     MatrixUType m_matrixU;
@@ -811,14 +836,17 @@ void JacobiSVD<MatrixType, QRPreconditioner>::allocate(Index rows, Index cols, u
   
   if(m_cols>m_rows)   m_qr_precond_morecols.allocate(*this);
   if(m_rows>m_cols)   m_qr_precond_morerows.allocate(*this);
-  if(m_cols!=m_cols)  m_scaledMatrix.resize(rows,cols);
+  if(m_rows!=m_cols)  m_scaledMatrix.resize(rows,cols);
 }
 
 template<typename MatrixType, int QRPreconditioner>
 JacobiSVD<MatrixType, QRPreconditioner>&
 JacobiSVD<MatrixType, QRPreconditioner>::compute(const MatrixType& matrix, unsigned int computationOptions)
 {
+  check_template_parameters();
+  
   using std::abs;
+  using std::max;
   allocate(matrix.rows(), matrix.cols(), computationOptions);
 
   // currently we stop when we reach precision 2*epsilon as the last bit of precision can require an unreasonable number of iterations,
@@ -850,6 +878,7 @@ JacobiSVD<MatrixType, QRPreconditioner>::compute(const MatrixType& matrix, unsig
   }
 
   /*** step 2. The main Jacobi SVD iteration. ***/
+  RealScalar maxDiagEntry = m_workMatrix.cwiseAbs().diagonal().maxCoeff();
 
   bool finished = false;
   while(!finished)
@@ -865,25 +894,27 @@ JacobiSVD<MatrixType, QRPreconditioner>::compute(const MatrixType& matrix, unsig
         // if this 2x2 sub-matrix is not diagonal already...
         // notice that this comparison will evaluate to false if any NaN is involved, ensuring that NaN's don't
         // keep us iterating forever. Similarly, small denormal numbers are considered zero.
-        using std::max;
+        RealScalar threshold = max EIGEN_EMPTY (considerAsZero, precision * maxDiagEntry);
-        RealScalar threshold = (max)(considerAsZero, precision * (max)(abs(m_workMatrix.coeff(p,p)),
-                                                                       abs(m_workMatrix.coeff(q,q))));
-        // We compare both values to threshold instead of calling max to be robust to NaN (See bug 791)
         if(abs(m_workMatrix.coeff(p,q))>threshold || abs(m_workMatrix.coeff(q,p)) > threshold)
         {
           finished = false;
-
           // perform SVD decomposition of 2x2 sub-matrix corresponding to indices p,q to make it diagonal
-          internal::svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner>::run(m_workMatrix, *this, p, q);
+          // the complex to real operation returns true is the updated 2x2 block is not already diagonal
-          JacobiRotation<RealScalar> j_left, j_right;
+          if(internal::svd_precondition_2x2_block_to_be_real<MatrixType, QRPreconditioner>::run(m_workMatrix, *this, p, q, maxDiagEntry))
-          internal::real_2x2_jacobi_svd(m_workMatrix, p, q, &j_left, &j_right);
+          {
+            JacobiRotation<RealScalar> j_left, j_right;
+            internal::real_2x2_jacobi_svd(m_workMatrix, p, q, &j_left, &j_right);
 
-          // accumulate resulting Jacobi rotations
+            // accumulate resulting Jacobi rotations
-          m_workMatrix.applyOnTheLeft(p,q,j_left);
+            m_workMatrix.applyOnTheLeft(p,q,j_left);
-          if(computeU()) m_matrixU.applyOnTheRight(p,q,j_left.transpose());
+            if(computeU()) m_matrixU.applyOnTheRight(p,q,j_left.transpose());
 
-          m_workMatrix.applyOnTheRight(p,q,j_right);
+            m_workMatrix.applyOnTheRight(p,q,j_right);
-          if(computeV()) m_matrixV.applyOnTheRight(p,q,j_right);
+            if(computeV()) m_matrixV.applyOnTheRight(p,q,j_right);
+
+            // keep track of the largest diagonal coefficient
+            maxDiagEntry = max EIGEN_EMPTY (maxDiagEntry,max EIGEN_EMPTY (abs(m_workMatrix.coeff(p,p)), abs(m_workMatrix.coeff(q,q))));
+          }
         }
       }
     }

Eigen/src/SVD/JacobiSVD_MKL.h

@@ -45,8 +45,8 @@ JacobiSVD<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic>, ColPiv
 JacobiSVD<Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic>, ColPivHouseholderQRPreconditioner>::compute(const Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic>& matrix, unsigned int computationOptions) \
 { \
   typedef Matrix<EIGTYPE, Dynamic, Dynamic, EIGCOLROW, Dynamic, Dynamic> MatrixType; \
-  typedef MatrixType::Scalar Scalar; \
+  /*typedef MatrixType::Scalar Scalar;*/ \
-  typedef MatrixType::RealScalar RealScalar; \
+  /*typedef MatrixType::RealScalar RealScalar;*/ \
   allocate(matrix.rows(), matrix.cols(), computationOptions); \
 \
   /*const RealScalar precision = RealScalar(2) * NumTraits<Scalar>::epsilon();*/ \

Eigen/src/SparseCore/CompressedStorage.h

@@ -102,6 +102,11 @@ class CompressedStorage
     inline size_t allocatedSize() const { return m_allocatedSize; }
     inline void clear() { m_size = 0; }
 
+    const Scalar* valuePtr() const { return m_values; }
+    Scalar* valuePtr() { return m_values; }
+    const Index* indexPtr() const { return m_indices; }
+    Index* indexPtr() { return m_indices; }
+
     inline Scalar& value(size_t i) { return m_values[i]; }
     inline const Scalar& value(size_t i) const { return m_values[i]; }
 
@@ -208,8 +213,10 @@ class CompressedStorage
       Index* newIndices = new Index[size];
       size_t copySize = (std::min)(size, m_size);
       // copy
-      internal::smart_copy(m_values, m_values+copySize, newValues);
+      if (copySize>0) {
-      internal::smart_copy(m_indices, m_indices+copySize, newIndices);
+        internal::smart_copy(m_values, m_values+copySize, newValues);
+        internal::smart_copy(m_indices, m_indices+copySize, newIndices);
+      }
       // delete old stuff
       delete[] m_values;
       delete[] m_indices;

Eigen/src/SparseCore/SparseCwiseBinaryOp.h

@@ -55,10 +55,9 @@ class CwiseBinaryOpImpl<BinaryOp, Lhs, Rhs, Sparse>
     EIGEN_SPARSE_PUBLIC_INTERFACE(Derived)
     CwiseBinaryOpImpl()
     {
-      typedef typename internal::traits<Lhs>::StorageKind LhsStorageKind;
-      typedef typename internal::traits<Rhs>::StorageKind RhsStorageKind;
       EIGEN_STATIC_ASSERT((
-                (!internal::is_same<LhsStorageKind,RhsStorageKind>::value)
+                (!internal::is_same<typename internal::traits<Lhs>::StorageKind,
+                                    typename internal::traits<Rhs>::StorageKind>::value)
             ||  ((Lhs::Flags&RowMajorBit) == (Rhs::Flags&RowMajorBit))),
             THE_STORAGE_ORDER_OF_BOTH_SIDES_MUST_MATCH);
     }
@@ -314,10 +313,10 @@ SparseMatrixBase<Derived>::operator+=(const SparseMatrixBase<OtherDerived>& othe
 
 template<typename Derived>
 template<typename OtherDerived>
-EIGEN_STRONG_INLINE const EIGEN_SPARSE_CWISE_PRODUCT_RETURN_TYPE
+EIGEN_STRONG_INLINE const typename SparseMatrixBase<Derived>::template CwiseProductDenseReturnType<OtherDerived>::Type
 SparseMatrixBase<Derived>::cwiseProduct(const MatrixBase<OtherDerived> &other) const
 {
-  return EIGEN_SPARSE_CWISE_PRODUCT_RETURN_TYPE(derived(), other.derived());
+  return typename CwiseProductDenseReturnType<OtherDerived>::Type(derived(), other.derived());
 }
 
 } // end namespace Eigen